aboutsummaryrefslogtreecommitdiffstats
path: root/fs/ocfs2/dlm
diff options
context:
space:
mode:
Diffstat (limited to 'fs/ocfs2/dlm')
-rw-r--r--fs/ocfs2/dlm/Makefile8
-rw-r--r--fs/ocfs2/dlm/dlmapi.h214
-rw-r--r--fs/ocfs2/dlm/dlmast.c466
-rw-r--r--fs/ocfs2/dlm/dlmcommon.h884
-rw-r--r--fs/ocfs2/dlm/dlmconvert.c530
-rw-r--r--fs/ocfs2/dlm/dlmconvert.h35
-rw-r--r--fs/ocfs2/dlm/dlmdebug.c246
-rw-r--r--fs/ocfs2/dlm/dlmdebug.h30
-rw-r--r--fs/ocfs2/dlm/dlmdomain.c1469
-rw-r--r--fs/ocfs2/dlm/dlmdomain.h36
-rw-r--r--fs/ocfs2/dlm/dlmfs.c640
-rw-r--r--fs/ocfs2/dlm/dlmfsver.c42
-rw-r--r--fs/ocfs2/dlm/dlmfsver.h31
-rw-r--r--fs/ocfs2/dlm/dlmlock.c676
-rw-r--r--fs/ocfs2/dlm/dlmmaster.c2664
-rw-r--r--fs/ocfs2/dlm/dlmrecovery.c2132
-rw-r--r--fs/ocfs2/dlm/dlmthread.c692
-rw-r--r--fs/ocfs2/dlm/dlmunlock.c672
-rw-r--r--fs/ocfs2/dlm/dlmver.c42
-rw-r--r--fs/ocfs2/dlm/dlmver.h31
-rw-r--r--fs/ocfs2/dlm/userdlm.c658
-rw-r--r--fs/ocfs2/dlm/userdlm.h111
22 files changed, 12309 insertions, 0 deletions
diff --git a/fs/ocfs2/dlm/Makefile b/fs/ocfs2/dlm/Makefile
new file mode 100644
index 000000000000..ce3f7c29d270
--- /dev/null
+++ b/fs/ocfs2/dlm/Makefile
@@ -0,0 +1,8 @@
1EXTRA_CFLAGS += -Ifs/ocfs2
2
3obj-$(CONFIG_OCFS2_FS) += ocfs2_dlm.o ocfs2_dlmfs.o
4
5ocfs2_dlm-objs := dlmdomain.o dlmdebug.o dlmthread.o dlmrecovery.o \
6 dlmmaster.o dlmast.o dlmconvert.o dlmlock.o dlmunlock.o dlmver.o
7
8ocfs2_dlmfs-objs := userdlm.o dlmfs.o dlmfsver.o
diff --git a/fs/ocfs2/dlm/dlmapi.h b/fs/ocfs2/dlm/dlmapi.h
new file mode 100644
index 000000000000..53652f51c0e1
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmapi.h
@@ -0,0 +1,214 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmapi.h
5 *
6 * externally exported dlm interfaces
7 *
8 * Copyright (C) 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 *
25 */
26
27#ifndef DLMAPI_H
28#define DLMAPI_H
29
30struct dlm_lock;
31struct dlm_ctxt;
32
33/* NOTE: changes made to this enum should be reflected in dlmdebug.c */
34enum dlm_status {
35 DLM_NORMAL = 0, /* 0: request in progress */
36 DLM_GRANTED, /* 1: request granted */
37 DLM_DENIED, /* 2: request denied */
38 DLM_DENIED_NOLOCKS, /* 3: request denied, out of system resources */
39 DLM_WORKING, /* 4: async request in progress */
40 DLM_BLOCKED, /* 5: lock request blocked */
41 DLM_BLOCKED_ORPHAN, /* 6: lock request blocked by a orphan lock*/
42 DLM_DENIED_GRACE_PERIOD, /* 7: topological change in progress */
43 DLM_SYSERR, /* 8: system error */
44 DLM_NOSUPPORT, /* 9: unsupported */
45 DLM_CANCELGRANT, /* 10: can't cancel convert: already granted */
46 DLM_IVLOCKID, /* 11: bad lockid */
47 DLM_SYNC, /* 12: synchronous request granted */
48 DLM_BADTYPE, /* 13: bad resource type */
49 DLM_BADRESOURCE, /* 14: bad resource handle */
50 DLM_MAXHANDLES, /* 15: no more resource handles */
51 DLM_NOCLINFO, /* 16: can't contact cluster manager */
52 DLM_NOLOCKMGR, /* 17: can't contact lock manager */
53 DLM_NOPURGED, /* 18: can't contact purge daemon */
54 DLM_BADARGS, /* 19: bad api args */
55 DLM_VOID, /* 20: no status */
56 DLM_NOTQUEUED, /* 21: NOQUEUE was specified and request failed */
57 DLM_IVBUFLEN, /* 22: invalid resource name length */
58 DLM_CVTUNGRANT, /* 23: attempted to convert ungranted lock */
59 DLM_BADPARAM, /* 24: invalid lock mode specified */
60 DLM_VALNOTVALID, /* 25: value block has been invalidated */
61 DLM_REJECTED, /* 26: request rejected, unrecognized client */
62 DLM_ABORT, /* 27: blocked lock request cancelled */
63 DLM_CANCEL, /* 28: conversion request cancelled */
64 DLM_IVRESHANDLE, /* 29: invalid resource handle */
65 DLM_DEADLOCK, /* 30: deadlock recovery refused this request */
66 DLM_DENIED_NOASTS, /* 31: failed to allocate AST */
67 DLM_FORWARD, /* 32: request must wait for primary's response */
68 DLM_TIMEOUT, /* 33: timeout value for lock has expired */
69 DLM_IVGROUPID, /* 34: invalid group specification */
70 DLM_VERS_CONFLICT, /* 35: version conflicts prevent request handling */
71 DLM_BAD_DEVICE_PATH, /* 36: Locks device does not exist or path wrong */
72 DLM_NO_DEVICE_PERMISSION, /* 37: Client has insufficient pers for device */
73 DLM_NO_CONTROL_DEVICE, /* 38: Cannot set options on opened device */
74
75 DLM_RECOVERING, /* 39: extension, allows caller to fail a lock
76 request if it is being recovered */
77 DLM_MIGRATING, /* 40: extension, allows caller to fail a lock
78 request if it is being migrated */
79 DLM_MAXSTATS, /* 41: upper limit for return code validation */
80};
81
82/* for pretty-printing dlm_status error messages */
83const char *dlm_errmsg(enum dlm_status err);
84/* for pretty-printing dlm_status error names */
85const char *dlm_errname(enum dlm_status err);
86
87/* Eventually the DLM will use standard errno values, but in the
88 * meantime this lets us track dlm errors as they bubble up. When we
89 * bring its error reporting into line with the rest of the stack,
90 * these can just be replaced with calls to mlog_errno. */
91#define dlm_error(st) do { \
92 if ((st) != DLM_RECOVERING && \
93 (st) != DLM_MIGRATING && \
94 (st) != DLM_FORWARD) \
95 mlog(ML_ERROR, "dlm status = %s\n", dlm_errname((st))); \
96} while (0)
97
98#define DLM_LKSB_UNUSED1 0x01
99#define DLM_LKSB_PUT_LVB 0x02
100#define DLM_LKSB_GET_LVB 0x04
101#define DLM_LKSB_UNUSED2 0x08
102#define DLM_LKSB_UNUSED3 0x10
103#define DLM_LKSB_UNUSED4 0x20
104#define DLM_LKSB_UNUSED5 0x40
105#define DLM_LKSB_UNUSED6 0x80
106
107#define DLM_LVB_LEN 64
108
109/* Callers are only allowed access to the lvb and status members of
110 * this struct. */
111struct dlm_lockstatus {
112 enum dlm_status status;
113 u32 flags;
114 struct dlm_lock *lockid;
115 char lvb[DLM_LVB_LEN];
116};
117
118/* Valid lock modes. */
119#define LKM_IVMODE (-1) /* invalid mode */
120#define LKM_NLMODE 0 /* null lock */
121#define LKM_CRMODE 1 /* concurrent read unsupported */
122#define LKM_CWMODE 2 /* concurrent write unsupported */
123#define LKM_PRMODE 3 /* protected read */
124#define LKM_PWMODE 4 /* protected write unsupported */
125#define LKM_EXMODE 5 /* exclusive */
126#define LKM_MAXMODE 5
127#define LKM_MODEMASK 0xff
128
129/* Flags passed to dlmlock and dlmunlock:
130 * reserved: flags used by the "real" dlm
131 * only a few are supported by this dlm
132 * (U) = unsupported by ocfs2 dlm */
133#define LKM_ORPHAN 0x00000010 /* this lock is orphanable (U) */
134#define LKM_PARENTABLE 0x00000020 /* this lock was orphaned (U) */
135#define LKM_BLOCK 0x00000040 /* blocking lock request (U) */
136#define LKM_LOCAL 0x00000080 /* local lock request */
137#define LKM_VALBLK 0x00000100 /* lock value block request */
138#define LKM_NOQUEUE 0x00000200 /* non blocking request */
139#define LKM_CONVERT 0x00000400 /* conversion request */
140#define LKM_NODLCKWT 0x00000800 /* this lock wont deadlock (U) */
141#define LKM_UNLOCK 0x00001000 /* deallocate this lock */
142#define LKM_CANCEL 0x00002000 /* cancel conversion request */
143#define LKM_DEQALL 0x00004000 /* remove all locks held by proc (U) */
144#define LKM_INVVALBLK 0x00008000 /* invalidate lock value block */
145#define LKM_SYNCSTS 0x00010000 /* return synchronous status if poss (U) */
146#define LKM_TIMEOUT 0x00020000 /* lock request contains timeout (U) */
147#define LKM_SNGLDLCK 0x00040000 /* request can self-deadlock (U) */
148#define LKM_FINDLOCAL 0x00080000 /* find local lock request (U) */
149#define LKM_PROC_OWNED 0x00100000 /* owned by process, not group (U) */
150#define LKM_XID 0x00200000 /* use transaction id for deadlock (U) */
151#define LKM_XID_CONFLICT 0x00400000 /* do not allow lock inheritance (U) */
152#define LKM_FORCE 0x00800000 /* force unlock flag */
153#define LKM_REVVALBLK 0x01000000 /* temporary solution: re-validate
154 lock value block (U) */
155/* unused */
156#define LKM_UNUSED1 0x00000001 /* unused */
157#define LKM_UNUSED2 0x00000002 /* unused */
158#define LKM_UNUSED3 0x00000004 /* unused */
159#define LKM_UNUSED4 0x00000008 /* unused */
160#define LKM_UNUSED5 0x02000000 /* unused */
161#define LKM_UNUSED6 0x04000000 /* unused */
162#define LKM_UNUSED7 0x08000000 /* unused */
163
164/* ocfs2 extensions: internal only
165 * should never be used by caller */
166#define LKM_MIGRATION 0x10000000 /* extension: lockres is to be migrated
167 to another node */
168#define LKM_PUT_LVB 0x20000000 /* extension: lvb is being passed
169 should be applied to lockres */
170#define LKM_GET_LVB 0x40000000 /* extension: lvb should be copied
171 from lockres when lock is granted */
172#define LKM_RECOVERY 0x80000000 /* extension: flag for recovery lock
173 used to avoid recovery rwsem */
174
175
176typedef void (dlm_astlockfunc_t)(void *);
177typedef void (dlm_bastlockfunc_t)(void *, int);
178typedef void (dlm_astunlockfunc_t)(void *, enum dlm_status);
179
180enum dlm_status dlmlock(struct dlm_ctxt *dlm,
181 int mode,
182 struct dlm_lockstatus *lksb,
183 int flags,
184 const char *name,
185 dlm_astlockfunc_t *ast,
186 void *data,
187 dlm_bastlockfunc_t *bast);
188
189enum dlm_status dlmunlock(struct dlm_ctxt *dlm,
190 struct dlm_lockstatus *lksb,
191 int flags,
192 dlm_astunlockfunc_t *unlockast,
193 void *data);
194
195struct dlm_ctxt * dlm_register_domain(const char *domain, u32 key);
196
197void dlm_unregister_domain(struct dlm_ctxt *dlm);
198
199void dlm_print_one_lock(struct dlm_lock *lockid);
200
201typedef void (dlm_eviction_func)(int, void *);
202struct dlm_eviction_cb {
203 struct list_head ec_item;
204 dlm_eviction_func *ec_func;
205 void *ec_data;
206};
207void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
208 dlm_eviction_func *f,
209 void *data);
210void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
211 struct dlm_eviction_cb *cb);
212void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb);
213
214#endif /* DLMAPI_H */
diff --git a/fs/ocfs2/dlm/dlmast.c b/fs/ocfs2/dlm/dlmast.c
new file mode 100644
index 000000000000..8d17d28ef91c
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmast.c
@@ -0,0 +1,466 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmast.c
5 *
6 * AST and BAST functionality for local and remote nodes
7 *
8 * Copyright (C) 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 *
25 */
26
27
28#include <linux/module.h>
29#include <linux/fs.h>
30#include <linux/types.h>
31#include <linux/slab.h>
32#include <linux/highmem.h>
33#include <linux/utsname.h>
34#include <linux/init.h>
35#include <linux/sysctl.h>
36#include <linux/random.h>
37#include <linux/blkdev.h>
38#include <linux/socket.h>
39#include <linux/inet.h>
40#include <linux/spinlock.h>
41
42
43#include "cluster/heartbeat.h"
44#include "cluster/nodemanager.h"
45#include "cluster/tcp.h"
46#include "cluster/endian.h"
47
48#include "dlmapi.h"
49#include "dlmcommon.h"
50
51#define MLOG_MASK_PREFIX ML_DLM
52#include "cluster/masklog.h"
53
54static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
55 struct dlm_lock *lock);
56static int dlm_should_cancel_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
57
58/* Should be called as an ast gets queued to see if the new
59 * lock level will obsolete a pending bast.
60 * For example, if dlm_thread queued a bast for an EX lock that
61 * was blocking another EX, but before sending the bast the
62 * lock owner downconverted to NL, the bast is now obsolete.
63 * Only the ast should be sent.
64 * This is needed because the lock and convert paths can queue
65 * asts out-of-band (not waiting for dlm_thread) in order to
66 * allow for LKM_NOQUEUE to get immediate responses. */
67static int dlm_should_cancel_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
68{
69 assert_spin_locked(&dlm->ast_lock);
70 assert_spin_locked(&lock->spinlock);
71
72 if (lock->ml.highest_blocked == LKM_IVMODE)
73 return 0;
74 BUG_ON(lock->ml.highest_blocked == LKM_NLMODE);
75
76 if (lock->bast_pending &&
77 list_empty(&lock->bast_list))
78 /* old bast already sent, ok */
79 return 0;
80
81 if (lock->ml.type == LKM_EXMODE)
82 /* EX blocks anything left, any bast still valid */
83 return 0;
84 else if (lock->ml.type == LKM_NLMODE)
85 /* NL blocks nothing, no reason to send any bast, cancel it */
86 return 1;
87 else if (lock->ml.highest_blocked != LKM_EXMODE)
88 /* PR only blocks EX */
89 return 1;
90
91 return 0;
92}
93
94static void __dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
95{
96 mlog_entry_void();
97
98 BUG_ON(!dlm);
99 BUG_ON(!lock);
100
101 assert_spin_locked(&dlm->ast_lock);
102 if (!list_empty(&lock->ast_list)) {
103 mlog(ML_ERROR, "ast list not empty!! pending=%d, newlevel=%d\n",
104 lock->ast_pending, lock->ml.type);
105 BUG();
106 }
107 BUG_ON(!list_empty(&lock->ast_list));
108 if (lock->ast_pending)
109 mlog(0, "lock has an ast getting flushed right now\n");
110
111 /* putting lock on list, add a ref */
112 dlm_lock_get(lock);
113 spin_lock(&lock->spinlock);
114
115 /* check to see if this ast obsoletes the bast */
116 if (dlm_should_cancel_bast(dlm, lock)) {
117 struct dlm_lock_resource *res = lock->lockres;
118 mlog(0, "%s: cancelling bast for %.*s\n",
119 dlm->name, res->lockname.len, res->lockname.name);
120 lock->bast_pending = 0;
121 list_del_init(&lock->bast_list);
122 lock->ml.highest_blocked = LKM_IVMODE;
123 /* removing lock from list, remove a ref. guaranteed
124 * this won't be the last ref because of the get above,
125 * so res->spinlock will not be taken here */
126 dlm_lock_put(lock);
127 /* free up the reserved bast that we are cancelling.
128 * guaranteed that this will not be the last reserved
129 * ast because *both* an ast and a bast were reserved
130 * to get to this point. the res->spinlock will not be
131 * taken here */
132 dlm_lockres_release_ast(dlm, res);
133 }
134 list_add_tail(&lock->ast_list, &dlm->pending_asts);
135 lock->ast_pending = 1;
136 spin_unlock(&lock->spinlock);
137}
138
139void dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
140{
141 mlog_entry_void();
142
143 BUG_ON(!dlm);
144 BUG_ON(!lock);
145
146 spin_lock(&dlm->ast_lock);
147 __dlm_queue_ast(dlm, lock);
148 spin_unlock(&dlm->ast_lock);
149}
150
151
152static void __dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
153{
154 mlog_entry_void();
155
156 BUG_ON(!dlm);
157 BUG_ON(!lock);
158 assert_spin_locked(&dlm->ast_lock);
159
160 BUG_ON(!list_empty(&lock->bast_list));
161 if (lock->bast_pending)
162 mlog(0, "lock has a bast getting flushed right now\n");
163
164 /* putting lock on list, add a ref */
165 dlm_lock_get(lock);
166 spin_lock(&lock->spinlock);
167 list_add_tail(&lock->bast_list, &dlm->pending_basts);
168 lock->bast_pending = 1;
169 spin_unlock(&lock->spinlock);
170}
171
172void dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock)
173{
174 mlog_entry_void();
175
176 BUG_ON(!dlm);
177 BUG_ON(!lock);
178
179 spin_lock(&dlm->ast_lock);
180 __dlm_queue_bast(dlm, lock);
181 spin_unlock(&dlm->ast_lock);
182}
183
184static void dlm_update_lvb(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
185 struct dlm_lock *lock)
186{
187 struct dlm_lockstatus *lksb = lock->lksb;
188 BUG_ON(!lksb);
189
190 /* only updates if this node masters the lockres */
191 if (res->owner == dlm->node_num) {
192
193 spin_lock(&res->spinlock);
194 /* check the lksb flags for the direction */
195 if (lksb->flags & DLM_LKSB_GET_LVB) {
196 mlog(0, "getting lvb from lockres for %s node\n",
197 lock->ml.node == dlm->node_num ? "master" :
198 "remote");
199 memcpy(lksb->lvb, res->lvb, DLM_LVB_LEN);
200 } else if (lksb->flags & DLM_LKSB_PUT_LVB) {
201 mlog(0, "setting lvb from lockres for %s node\n",
202 lock->ml.node == dlm->node_num ? "master" :
203 "remote");
204 memcpy(res->lvb, lksb->lvb, DLM_LVB_LEN);
205 }
206 spin_unlock(&res->spinlock);
207 }
208
209 /* reset any lvb flags on the lksb */
210 lksb->flags &= ~(DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB);
211}
212
213void dlm_do_local_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
214 struct dlm_lock *lock)
215{
216 dlm_astlockfunc_t *fn;
217 struct dlm_lockstatus *lksb;
218
219 mlog_entry_void();
220
221 lksb = lock->lksb;
222 fn = lock->ast;
223 BUG_ON(lock->ml.node != dlm->node_num);
224
225 dlm_update_lvb(dlm, res, lock);
226 (*fn)(lock->astdata);
227}
228
229
230int dlm_do_remote_ast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
231 struct dlm_lock *lock)
232{
233 int ret;
234 struct dlm_lockstatus *lksb;
235 int lksbflags;
236
237 mlog_entry_void();
238
239 lksb = lock->lksb;
240 BUG_ON(lock->ml.node == dlm->node_num);
241
242 lksbflags = lksb->flags;
243 dlm_update_lvb(dlm, res, lock);
244
245 /* lock request came from another node
246 * go do the ast over there */
247 ret = dlm_send_proxy_ast(dlm, res, lock, lksbflags);
248 return ret;
249}
250
251void dlm_do_local_bast(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
252 struct dlm_lock *lock, int blocked_type)
253{
254 dlm_bastlockfunc_t *fn = lock->bast;
255
256 mlog_entry_void();
257 BUG_ON(lock->ml.node != dlm->node_num);
258
259 (*fn)(lock->astdata, blocked_type);
260}
261
262
263
264int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data)
265{
266 int ret;
267 unsigned int locklen;
268 struct dlm_ctxt *dlm = data;
269 struct dlm_lock_resource *res = NULL;
270 struct dlm_lock *lock = NULL;
271 struct dlm_proxy_ast *past = (struct dlm_proxy_ast *) msg->buf;
272 char *name;
273 struct list_head *iter, *head=NULL;
274 u64 cookie;
275 u32 flags;
276
277 if (!dlm_grab(dlm)) {
278 dlm_error(DLM_REJECTED);
279 return DLM_REJECTED;
280 }
281
282 mlog_bug_on_msg(!dlm_domain_fully_joined(dlm),
283 "Domain %s not fully joined!\n", dlm->name);
284
285 name = past->name;
286 locklen = past->namelen;
287 cookie = be64_to_cpu(past->cookie);
288 flags = be32_to_cpu(past->flags);
289
290 if (locklen > DLM_LOCKID_NAME_MAX) {
291 ret = DLM_IVBUFLEN;
292 mlog(ML_ERROR, "Invalid name length in proxy ast handler!\n");
293 goto leave;
294 }
295
296 if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) ==
297 (LKM_PUT_LVB|LKM_GET_LVB)) {
298 mlog(ML_ERROR, "both PUT and GET lvb specified\n");
299 ret = DLM_BADARGS;
300 goto leave;
301 }
302
303 mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" :
304 (flags & LKM_GET_LVB ? "get lvb" : "none"));
305
306 mlog(0, "type=%d, blocked_type=%d\n", past->type, past->blocked_type);
307
308 if (past->type != DLM_AST &&
309 past->type != DLM_BAST) {
310 mlog(ML_ERROR, "Unknown ast type! %d, cookie=%"MLFu64", "
311 "name=%.*s\n", past->type, cookie, locklen, name);
312 ret = DLM_IVLOCKID;
313 goto leave;
314 }
315
316 res = dlm_lookup_lockres(dlm, name, locklen);
317 if (!res) {
318 mlog(ML_ERROR, "got %sast for unknown lockres! "
319 "cookie=%"MLFu64", name=%.*s, namelen=%u\n",
320 past->type == DLM_AST ? "" : "b",
321 cookie, locklen, name, locklen);
322 ret = DLM_IVLOCKID;
323 goto leave;
324 }
325
326 /* cannot get a proxy ast message if this node owns it */
327 BUG_ON(res->owner == dlm->node_num);
328
329 mlog(0, "lockres %.*s\n", res->lockname.len, res->lockname.name);
330
331 spin_lock(&res->spinlock);
332 if (res->state & DLM_LOCK_RES_RECOVERING) {
333 mlog(0, "responding with DLM_RECOVERING!\n");
334 ret = DLM_RECOVERING;
335 goto unlock_out;
336 }
337 if (res->state & DLM_LOCK_RES_MIGRATING) {
338 mlog(0, "responding with DLM_MIGRATING!\n");
339 ret = DLM_MIGRATING;
340 goto unlock_out;
341 }
342 /* try convert queue for both ast/bast */
343 head = &res->converting;
344 lock = NULL;
345 list_for_each(iter, head) {
346 lock = list_entry (iter, struct dlm_lock, list);
347 if (be64_to_cpu(lock->ml.cookie) == cookie)
348 goto do_ast;
349 }
350
351 /* if not on convert, try blocked for ast, granted for bast */
352 if (past->type == DLM_AST)
353 head = &res->blocked;
354 else
355 head = &res->granted;
356
357 list_for_each(iter, head) {
358 lock = list_entry (iter, struct dlm_lock, list);
359 if (be64_to_cpu(lock->ml.cookie) == cookie)
360 goto do_ast;
361 }
362
363 mlog(ML_ERROR, "got %sast for unknown lock! cookie=%"MLFu64", "
364 "name=%.*s, namelen=%u\n",
365 past->type == DLM_AST ? "" : "b", cookie, locklen, name, locklen);
366
367 ret = DLM_NORMAL;
368unlock_out:
369 spin_unlock(&res->spinlock);
370 goto leave;
371
372do_ast:
373 ret = DLM_NORMAL;
374 if (past->type == DLM_AST) {
375 /* do not alter lock refcount. switching lists. */
376 list_del_init(&lock->list);
377 list_add_tail(&lock->list, &res->granted);
378 mlog(0, "ast: adding to granted list... type=%d, "
379 "convert_type=%d\n", lock->ml.type, lock->ml.convert_type);
380 if (lock->ml.convert_type != LKM_IVMODE) {
381 lock->ml.type = lock->ml.convert_type;
382 lock->ml.convert_type = LKM_IVMODE;
383 } else {
384 // should already be there....
385 }
386
387 lock->lksb->status = DLM_NORMAL;
388
389 /* if we requested the lvb, fetch it into our lksb now */
390 if (flags & LKM_GET_LVB) {
391 BUG_ON(!(lock->lksb->flags & DLM_LKSB_GET_LVB));
392 memcpy(lock->lksb->lvb, past->lvb, DLM_LVB_LEN);
393 }
394 }
395 spin_unlock(&res->spinlock);
396
397 if (past->type == DLM_AST)
398 dlm_do_local_ast(dlm, res, lock);
399 else
400 dlm_do_local_bast(dlm, res, lock, past->blocked_type);
401
402leave:
403
404 if (res)
405 dlm_lockres_put(res);
406
407 dlm_put(dlm);
408 return ret;
409}
410
411
412
413int dlm_send_proxy_ast_msg(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
414 struct dlm_lock *lock, int msg_type,
415 int blocked_type, int flags)
416{
417 int ret = 0;
418 struct dlm_proxy_ast past;
419 struct kvec vec[2];
420 size_t veclen = 1;
421 int status;
422
423 mlog_entry("res %.*s, to=%u, type=%d, blocked_type=%d\n",
424 res->lockname.len, res->lockname.name, lock->ml.node,
425 msg_type, blocked_type);
426
427 memset(&past, 0, sizeof(struct dlm_proxy_ast));
428 past.node_idx = dlm->node_num;
429 past.type = msg_type;
430 past.blocked_type = blocked_type;
431 past.namelen = res->lockname.len;
432 memcpy(past.name, res->lockname.name, past.namelen);
433 past.cookie = lock->ml.cookie;
434
435 vec[0].iov_len = sizeof(struct dlm_proxy_ast);
436 vec[0].iov_base = &past;
437 if (flags & DLM_LKSB_GET_LVB) {
438 mlog(0, "returning requested LVB data\n");
439 be32_add_cpu(&past.flags, LKM_GET_LVB);
440 vec[1].iov_len = DLM_LVB_LEN;
441 vec[1].iov_base = lock->lksb->lvb;
442 veclen++;
443 }
444
445 ret = o2net_send_message_vec(DLM_PROXY_AST_MSG, dlm->key, vec, veclen,
446 lock->ml.node, &status);
447 if (ret < 0)
448 mlog_errno(ret);
449 else {
450 if (status == DLM_RECOVERING) {
451 mlog(ML_ERROR, "sent AST to node %u, it thinks this "
452 "node is dead!\n", lock->ml.node);
453 BUG();
454 } else if (status == DLM_MIGRATING) {
455 mlog(ML_ERROR, "sent AST to node %u, it returned "
456 "DLM_MIGRATING!\n", lock->ml.node);
457 BUG();
458 } else if (status != DLM_NORMAL) {
459 mlog(ML_ERROR, "AST to node %u returned %d!\n",
460 lock->ml.node, status);
461 /* ignore it */
462 }
463 ret = 0;
464 }
465 return ret;
466}
diff --git a/fs/ocfs2/dlm/dlmcommon.h b/fs/ocfs2/dlm/dlmcommon.h
new file mode 100644
index 000000000000..3fecba0a6023
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmcommon.h
@@ -0,0 +1,884 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmcommon.h
5 *
6 * Copyright (C) 2004 Oracle. All rights reserved.
7 *
8 * This program is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU General Public
10 * License as published by the Free Software Foundation; either
11 * version 2 of the License, or (at your option) any later version.
12 *
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public
19 * License along with this program; if not, write to the
20 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
21 * Boston, MA 021110-1307, USA.
22 *
23 */
24
25#ifndef DLMCOMMON_H
26#define DLMCOMMON_H
27
28#include <linux/kref.h>
29
30#define DLM_HB_NODE_DOWN_PRI (0xf000000)
31#define DLM_HB_NODE_UP_PRI (0x8000000)
32
33#define DLM_LOCKID_NAME_MAX 32
34
35#define DLM_DOMAIN_NAME_MAX_LEN 255
36#define DLM_LOCK_RES_OWNER_UNKNOWN O2NM_MAX_NODES
37#define DLM_THREAD_SHUFFLE_INTERVAL 5 // flush everything every 5 passes
38#define DLM_THREAD_MS 200 // flush at least every 200 ms
39
40#define DLM_HASH_BITS 7
41#define DLM_HASH_SIZE (1 << DLM_HASH_BITS)
42#define DLM_HASH_MASK (DLM_HASH_SIZE - 1)
43
44enum dlm_ast_type {
45 DLM_AST = 0,
46 DLM_BAST,
47 DLM_ASTUNLOCK
48};
49
50
51#define LKM_VALID_FLAGS (LKM_VALBLK | LKM_CONVERT | LKM_UNLOCK | \
52 LKM_CANCEL | LKM_INVVALBLK | LKM_FORCE | \
53 LKM_RECOVERY | LKM_LOCAL | LKM_NOQUEUE)
54
55#define DLM_RECOVERY_LOCK_NAME "$RECOVERY"
56#define DLM_RECOVERY_LOCK_NAME_LEN 9
57
58static inline int dlm_is_recovery_lock(const char *lock_name, int name_len)
59{
60 if (name_len == DLM_RECOVERY_LOCK_NAME_LEN &&
61 memcmp(lock_name, DLM_RECOVERY_LOCK_NAME, name_len)==0)
62 return 1;
63 return 0;
64}
65
66#define DLM_RECO_STATE_ACTIVE 0x0001
67
68struct dlm_recovery_ctxt
69{
70 struct list_head resources;
71 struct list_head received;
72 struct list_head node_data;
73 u8 new_master;
74 u8 dead_node;
75 u16 state;
76 unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
77 wait_queue_head_t event;
78};
79
80enum dlm_ctxt_state {
81 DLM_CTXT_NEW = 0,
82 DLM_CTXT_JOINED,
83 DLM_CTXT_IN_SHUTDOWN,
84 DLM_CTXT_LEAVING,
85};
86
87struct dlm_ctxt
88{
89 struct list_head list;
90 struct list_head *resources;
91 struct list_head dirty_list;
92 struct list_head purge_list;
93 struct list_head pending_asts;
94 struct list_head pending_basts;
95 unsigned int purge_count;
96 spinlock_t spinlock;
97 spinlock_t ast_lock;
98 char *name;
99 u8 node_num;
100 u32 key;
101 u8 joining_node;
102 wait_queue_head_t dlm_join_events;
103 unsigned long live_nodes_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
104 unsigned long domain_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
105 unsigned long recovery_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
106 struct dlm_recovery_ctxt reco;
107 spinlock_t master_lock;
108 struct list_head master_list;
109 struct list_head mle_hb_events;
110
111 /* these give a really vague idea of the system load */
112 atomic_t local_resources;
113 atomic_t remote_resources;
114 atomic_t unknown_resources;
115
116 /* NOTE: Next three are protected by dlm_domain_lock */
117 struct kref dlm_refs;
118 enum dlm_ctxt_state dlm_state;
119 unsigned int num_joins;
120
121 struct o2hb_callback_func dlm_hb_up;
122 struct o2hb_callback_func dlm_hb_down;
123 struct task_struct *dlm_thread_task;
124 struct task_struct *dlm_reco_thread_task;
125 wait_queue_head_t dlm_thread_wq;
126 wait_queue_head_t dlm_reco_thread_wq;
127 wait_queue_head_t ast_wq;
128 wait_queue_head_t migration_wq;
129
130 struct work_struct dispatched_work;
131 struct list_head work_list;
132 spinlock_t work_lock;
133 struct list_head dlm_domain_handlers;
134 struct list_head dlm_eviction_callbacks;
135};
136
137/* these keventd work queue items are for less-frequently
138 * called functions that cannot be directly called from the
139 * net message handlers for some reason, usually because
140 * they need to send net messages of their own. */
141void dlm_dispatch_work(void *data);
142
143struct dlm_lock_resource;
144struct dlm_work_item;
145
146typedef void (dlm_workfunc_t)(struct dlm_work_item *, void *);
147
148struct dlm_request_all_locks_priv
149{
150 u8 reco_master;
151 u8 dead_node;
152};
153
154struct dlm_mig_lockres_priv
155{
156 struct dlm_lock_resource *lockres;
157 u8 real_master;
158};
159
160struct dlm_assert_master_priv
161{
162 struct dlm_lock_resource *lockres;
163 u8 request_from;
164 u32 flags;
165 unsigned ignore_higher:1;
166};
167
168
169struct dlm_work_item
170{
171 struct list_head list;
172 dlm_workfunc_t *func;
173 struct dlm_ctxt *dlm;
174 void *data;
175 union {
176 struct dlm_request_all_locks_priv ral;
177 struct dlm_mig_lockres_priv ml;
178 struct dlm_assert_master_priv am;
179 } u;
180};
181
182static inline void dlm_init_work_item(struct dlm_ctxt *dlm,
183 struct dlm_work_item *i,
184 dlm_workfunc_t *f, void *data)
185{
186 memset(i, 0, sizeof(*i));
187 i->func = f;
188 INIT_LIST_HEAD(&i->list);
189 i->data = data;
190 i->dlm = dlm; /* must have already done a dlm_grab on this! */
191}
192
193
194
195static inline void __dlm_set_joining_node(struct dlm_ctxt *dlm,
196 u8 node)
197{
198 assert_spin_locked(&dlm->spinlock);
199
200 dlm->joining_node = node;
201 wake_up(&dlm->dlm_join_events);
202}
203
204#define DLM_LOCK_RES_UNINITED 0x00000001
205#define DLM_LOCK_RES_RECOVERING 0x00000002
206#define DLM_LOCK_RES_READY 0x00000004
207#define DLM_LOCK_RES_DIRTY 0x00000008
208#define DLM_LOCK_RES_IN_PROGRESS 0x00000010
209#define DLM_LOCK_RES_MIGRATING 0x00000020
210
211#define DLM_PURGE_INTERVAL_MS (8 * 1000)
212
213struct dlm_lock_resource
214{
215 /* WARNING: Please see the comment in dlm_init_lockres before
216 * adding fields here. */
217 struct list_head list;
218 struct kref refs;
219
220 /* please keep these next 3 in this order
221 * some funcs want to iterate over all lists */
222 struct list_head granted;
223 struct list_head converting;
224 struct list_head blocked;
225
226 struct list_head dirty;
227 struct list_head recovering; // dlm_recovery_ctxt.resources list
228
229 /* unused lock resources have their last_used stamped and are
230 * put on a list for the dlm thread to run. */
231 struct list_head purge;
232 unsigned long last_used;
233
234 unsigned migration_pending:1;
235 atomic_t asts_reserved;
236 spinlock_t spinlock;
237 wait_queue_head_t wq;
238 u8 owner; //node which owns the lock resource, or unknown
239 u16 state;
240 struct qstr lockname;
241 char lvb[DLM_LVB_LEN];
242};
243
244struct dlm_migratable_lock
245{
246 __be64 cookie;
247
248 /* these 3 are just padding for the in-memory structure, but
249 * list and flags are actually used when sent over the wire */
250 __be16 pad1;
251 u8 list; // 0=granted, 1=converting, 2=blocked
252 u8 flags;
253
254 s8 type;
255 s8 convert_type;
256 s8 highest_blocked;
257 u8 node;
258}; // 16 bytes
259
260struct dlm_lock
261{
262 struct dlm_migratable_lock ml;
263
264 struct list_head list;
265 struct list_head ast_list;
266 struct list_head bast_list;
267 struct dlm_lock_resource *lockres;
268 spinlock_t spinlock;
269 struct kref lock_refs;
270
271 // ast and bast must be callable while holding a spinlock!
272 dlm_astlockfunc_t *ast;
273 dlm_bastlockfunc_t *bast;
274 void *astdata;
275 struct dlm_lockstatus *lksb;
276 unsigned ast_pending:1,
277 bast_pending:1,
278 convert_pending:1,
279 lock_pending:1,
280 cancel_pending:1,
281 unlock_pending:1,
282 lksb_kernel_allocated:1;
283};
284
285
286#define DLM_LKSB_UNUSED1 0x01
287#define DLM_LKSB_PUT_LVB 0x02
288#define DLM_LKSB_GET_LVB 0x04
289#define DLM_LKSB_UNUSED2 0x08
290#define DLM_LKSB_UNUSED3 0x10
291#define DLM_LKSB_UNUSED4 0x20
292#define DLM_LKSB_UNUSED5 0x40
293#define DLM_LKSB_UNUSED6 0x80
294
295
296enum dlm_lockres_list {
297 DLM_GRANTED_LIST = 0,
298 DLM_CONVERTING_LIST,
299 DLM_BLOCKED_LIST
300};
301
302static inline struct list_head *
303dlm_list_idx_to_ptr(struct dlm_lock_resource *res, enum dlm_lockres_list idx)
304{
305 struct list_head *ret = NULL;
306 if (idx == DLM_GRANTED_LIST)
307 ret = &res->granted;
308 else if (idx == DLM_CONVERTING_LIST)
309 ret = &res->converting;
310 else if (idx == DLM_BLOCKED_LIST)
311 ret = &res->blocked;
312 else
313 BUG();
314 return ret;
315}
316
317
318
319
320struct dlm_node_iter
321{
322 unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
323 int curnode;
324};
325
326
327enum {
328 DLM_MASTER_REQUEST_MSG = 500,
329 DLM_UNUSED_MSG1, /* 501 */
330 DLM_ASSERT_MASTER_MSG, /* 502 */
331 DLM_CREATE_LOCK_MSG, /* 503 */
332 DLM_CONVERT_LOCK_MSG, /* 504 */
333 DLM_PROXY_AST_MSG, /* 505 */
334 DLM_UNLOCK_LOCK_MSG, /* 506 */
335 DLM_UNUSED_MSG2, /* 507 */
336 DLM_MIGRATE_REQUEST_MSG, /* 508 */
337 DLM_MIG_LOCKRES_MSG, /* 509 */
338 DLM_QUERY_JOIN_MSG, /* 510 */
339 DLM_ASSERT_JOINED_MSG, /* 511 */
340 DLM_CANCEL_JOIN_MSG, /* 512 */
341 DLM_EXIT_DOMAIN_MSG, /* 513 */
342 DLM_MASTER_REQUERY_MSG, /* 514 */
343 DLM_LOCK_REQUEST_MSG, /* 515 */
344 DLM_RECO_DATA_DONE_MSG, /* 516 */
345 DLM_BEGIN_RECO_MSG, /* 517 */
346 DLM_FINALIZE_RECO_MSG /* 518 */
347};
348
349struct dlm_reco_node_data
350{
351 int state;
352 u8 node_num;
353 struct list_head list;
354};
355
356enum {
357 DLM_RECO_NODE_DATA_DEAD = -1,
358 DLM_RECO_NODE_DATA_INIT = 0,
359 DLM_RECO_NODE_DATA_REQUESTING,
360 DLM_RECO_NODE_DATA_REQUESTED,
361 DLM_RECO_NODE_DATA_RECEIVING,
362 DLM_RECO_NODE_DATA_DONE,
363 DLM_RECO_NODE_DATA_FINALIZE_SENT,
364};
365
366
367enum {
368 DLM_MASTER_RESP_NO = 0,
369 DLM_MASTER_RESP_YES,
370 DLM_MASTER_RESP_MAYBE,
371 DLM_MASTER_RESP_ERROR
372};
373
374
375struct dlm_master_request
376{
377 u8 node_idx;
378 u8 namelen;
379 __be16 pad1;
380 __be32 flags;
381
382 u8 name[O2NM_MAX_NAME_LEN];
383};
384
385#define DLM_ASSERT_MASTER_MLE_CLEANUP 0x00000001
386#define DLM_ASSERT_MASTER_REQUERY 0x00000002
387#define DLM_ASSERT_MASTER_FINISH_MIGRATION 0x00000004
388struct dlm_assert_master
389{
390 u8 node_idx;
391 u8 namelen;
392 __be16 pad1;
393 __be32 flags;
394
395 u8 name[O2NM_MAX_NAME_LEN];
396};
397
398struct dlm_migrate_request
399{
400 u8 master;
401 u8 new_master;
402 u8 namelen;
403 u8 pad1;
404 __be32 pad2;
405 u8 name[O2NM_MAX_NAME_LEN];
406};
407
408struct dlm_master_requery
409{
410 u8 pad1;
411 u8 pad2;
412 u8 node_idx;
413 u8 namelen;
414 __be32 pad3;
415 u8 name[O2NM_MAX_NAME_LEN];
416};
417
418#define DLM_MRES_RECOVERY 0x01
419#define DLM_MRES_MIGRATION 0x02
420#define DLM_MRES_ALL_DONE 0x04
421
422/*
423 * We would like to get one whole lockres into a single network
424 * message whenever possible. Generally speaking, there will be
425 * at most one dlm_lock on a lockres for each node in the cluster,
426 * plus (infrequently) any additional locks coming in from userdlm.
427 *
428 * struct _dlm_lockres_page
429 * {
430 * dlm_migratable_lockres mres;
431 * dlm_migratable_lock ml[DLM_MAX_MIGRATABLE_LOCKS];
432 * u8 pad[DLM_MIG_LOCKRES_RESERVED];
433 * };
434 *
435 * from ../cluster/tcp.h
436 * NET_MAX_PAYLOAD_BYTES (4096 - sizeof(net_msg))
437 * (roughly 4080 bytes)
438 * and sizeof(dlm_migratable_lockres) = 112 bytes
439 * and sizeof(dlm_migratable_lock) = 16 bytes
440 *
441 * Choosing DLM_MAX_MIGRATABLE_LOCKS=240 and
442 * DLM_MIG_LOCKRES_RESERVED=128 means we have this:
443 *
444 * (DLM_MAX_MIGRATABLE_LOCKS * sizeof(dlm_migratable_lock)) +
445 * sizeof(dlm_migratable_lockres) + DLM_MIG_LOCKRES_RESERVED =
446 * NET_MAX_PAYLOAD_BYTES
447 * (240 * 16) + 112 + 128 = 4080
448 *
449 * So a lockres would need more than 240 locks before it would
450 * use more than one network packet to recover. Not too bad.
451 */
452#define DLM_MAX_MIGRATABLE_LOCKS 240
453
454struct dlm_migratable_lockres
455{
456 u8 master;
457 u8 lockname_len;
458 u8 num_locks; // locks sent in this structure
459 u8 flags;
460 __be32 total_locks; // locks to be sent for this migration cookie
461 __be64 mig_cookie; // cookie for this lockres migration
462 // or zero if not needed
463 // 16 bytes
464 u8 lockname[DLM_LOCKID_NAME_MAX];
465 // 48 bytes
466 u8 lvb[DLM_LVB_LEN];
467 // 112 bytes
468 struct dlm_migratable_lock ml[0]; // 16 bytes each, begins at byte 112
469};
470#define DLM_MIG_LOCKRES_MAX_LEN \
471 (sizeof(struct dlm_migratable_lockres) + \
472 (sizeof(struct dlm_migratable_lock) * \
473 DLM_MAX_MIGRATABLE_LOCKS) )
474
475/* from above, 128 bytes
476 * for some undetermined future use */
477#define DLM_MIG_LOCKRES_RESERVED (NET_MAX_PAYLOAD_BYTES - \
478 DLM_MIG_LOCKRES_MAX_LEN)
479
480struct dlm_create_lock
481{
482 __be64 cookie;
483
484 __be32 flags;
485 u8 pad1;
486 u8 node_idx;
487 s8 requested_type;
488 u8 namelen;
489
490 u8 name[O2NM_MAX_NAME_LEN];
491};
492
493struct dlm_convert_lock
494{
495 __be64 cookie;
496
497 __be32 flags;
498 u8 pad1;
499 u8 node_idx;
500 s8 requested_type;
501 u8 namelen;
502
503 u8 name[O2NM_MAX_NAME_LEN];
504
505 s8 lvb[0];
506};
507#define DLM_CONVERT_LOCK_MAX_LEN (sizeof(struct dlm_convert_lock)+DLM_LVB_LEN)
508
509struct dlm_unlock_lock
510{
511 __be64 cookie;
512
513 __be32 flags;
514 __be16 pad1;
515 u8 node_idx;
516 u8 namelen;
517
518 u8 name[O2NM_MAX_NAME_LEN];
519
520 s8 lvb[0];
521};
522#define DLM_UNLOCK_LOCK_MAX_LEN (sizeof(struct dlm_unlock_lock)+DLM_LVB_LEN)
523
524struct dlm_proxy_ast
525{
526 __be64 cookie;
527
528 __be32 flags;
529 u8 node_idx;
530 u8 type;
531 u8 blocked_type;
532 u8 namelen;
533
534 u8 name[O2NM_MAX_NAME_LEN];
535
536 s8 lvb[0];
537};
538#define DLM_PROXY_AST_MAX_LEN (sizeof(struct dlm_proxy_ast)+DLM_LVB_LEN)
539
540#define DLM_MOD_KEY (0x666c6172)
541enum dlm_query_join_response {
542 JOIN_DISALLOW = 0,
543 JOIN_OK,
544 JOIN_OK_NO_MAP,
545};
546
547struct dlm_lock_request
548{
549 u8 node_idx;
550 u8 dead_node;
551 __be16 pad1;
552 __be32 pad2;
553};
554
555struct dlm_reco_data_done
556{
557 u8 node_idx;
558 u8 dead_node;
559 __be16 pad1;
560 __be32 pad2;
561
562 /* unused for now */
563 /* eventually we can use this to attempt
564 * lvb recovery based on each node's info */
565 u8 reco_lvb[DLM_LVB_LEN];
566};
567
568struct dlm_begin_reco
569{
570 u8 node_idx;
571 u8 dead_node;
572 __be16 pad1;
573 __be32 pad2;
574};
575
576
577struct dlm_query_join_request
578{
579 u8 node_idx;
580 u8 pad1[2];
581 u8 name_len;
582 u8 domain[O2NM_MAX_NAME_LEN];
583};
584
585struct dlm_assert_joined
586{
587 u8 node_idx;
588 u8 pad1[2];
589 u8 name_len;
590 u8 domain[O2NM_MAX_NAME_LEN];
591};
592
593struct dlm_cancel_join
594{
595 u8 node_idx;
596 u8 pad1[2];
597 u8 name_len;
598 u8 domain[O2NM_MAX_NAME_LEN];
599};
600
601struct dlm_exit_domain
602{
603 u8 node_idx;
604 u8 pad1[3];
605};
606
607struct dlm_finalize_reco
608{
609 u8 node_idx;
610 u8 dead_node;
611 __be16 pad1;
612 __be32 pad2;
613};
614
615static inline enum dlm_status
616__dlm_lockres_state_to_status(struct dlm_lock_resource *res)
617{
618 enum dlm_status status = DLM_NORMAL;
619
620 assert_spin_locked(&res->spinlock);
621
622 if (res->state & DLM_LOCK_RES_RECOVERING)
623 status = DLM_RECOVERING;
624 else if (res->state & DLM_LOCK_RES_MIGRATING)
625 status = DLM_MIGRATING;
626 else if (res->state & DLM_LOCK_RES_IN_PROGRESS)
627 status = DLM_FORWARD;
628
629 return status;
630}
631
632struct dlm_lock * dlm_new_lock(int type, u8 node, u64 cookie,
633 struct dlm_lockstatus *lksb);
634void dlm_lock_get(struct dlm_lock *lock);
635void dlm_lock_put(struct dlm_lock *lock);
636
637void dlm_lock_attach_lockres(struct dlm_lock *lock,
638 struct dlm_lock_resource *res);
639
640int dlm_create_lock_handler(struct o2net_msg *msg, u32 len, void *data);
641int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data);
642int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data);
643
644void dlm_revert_pending_convert(struct dlm_lock_resource *res,
645 struct dlm_lock *lock);
646void dlm_revert_pending_lock(struct dlm_lock_resource *res,
647 struct dlm_lock *lock);
648
649int dlm_unlock_lock_handler(struct o2net_msg *msg, u32 len, void *data);
650void dlm_commit_pending_cancel(struct dlm_lock_resource *res,
651 struct dlm_lock *lock);
652void dlm_commit_pending_unlock(struct dlm_lock_resource *res,
653 struct dlm_lock *lock);
654
655int dlm_launch_thread(struct dlm_ctxt *dlm);
656void dlm_complete_thread(struct dlm_ctxt *dlm);
657int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
658void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
659void dlm_wait_for_recovery(struct dlm_ctxt *dlm);
660
661void dlm_put(struct dlm_ctxt *dlm);
662struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm);
663int dlm_domain_fully_joined(struct dlm_ctxt *dlm);
664
665void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
666 struct dlm_lock_resource *res);
667void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
668 struct dlm_lock_resource *res);
669void dlm_purge_lockres(struct dlm_ctxt *dlm,
670 struct dlm_lock_resource *lockres);
671void dlm_lockres_get(struct dlm_lock_resource *res);
672void dlm_lockres_put(struct dlm_lock_resource *res);
673void __dlm_unhash_lockres(struct dlm_lock_resource *res);
674void __dlm_insert_lockres(struct dlm_ctxt *dlm,
675 struct dlm_lock_resource *res);
676struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
677 const char *name,
678 unsigned int len);
679struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
680 const char *name,
681 unsigned int len);
682
683int dlm_is_host_down(int errno);
684void dlm_change_lockres_owner(struct dlm_ctxt *dlm,
685 struct dlm_lock_resource *res,
686 u8 owner);
687struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
688 const char *lockid,
689 int flags);
690struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
691 const char *name,
692 unsigned int namelen);
693
694void dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
695void dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
696void dlm_do_local_ast(struct dlm_ctxt *dlm,
697 struct dlm_lock_resource *res,
698 struct dlm_lock *lock);
699int dlm_do_remote_ast(struct dlm_ctxt *dlm,
700 struct dlm_lock_resource *res,
701 struct dlm_lock *lock);
702void dlm_do_local_bast(struct dlm_ctxt *dlm,
703 struct dlm_lock_resource *res,
704 struct dlm_lock *lock,
705 int blocked_type);
706int dlm_send_proxy_ast_msg(struct dlm_ctxt *dlm,
707 struct dlm_lock_resource *res,
708 struct dlm_lock *lock,
709 int msg_type,
710 int blocked_type, int flags);
711static inline int dlm_send_proxy_bast(struct dlm_ctxt *dlm,
712 struct dlm_lock_resource *res,
713 struct dlm_lock *lock,
714 int blocked_type)
715{
716 return dlm_send_proxy_ast_msg(dlm, res, lock, DLM_BAST,
717 blocked_type, 0);
718}
719
720static inline int dlm_send_proxy_ast(struct dlm_ctxt *dlm,
721 struct dlm_lock_resource *res,
722 struct dlm_lock *lock,
723 int flags)
724{
725 return dlm_send_proxy_ast_msg(dlm, res, lock, DLM_AST,
726 0, flags);
727}
728
729void dlm_print_one_lock_resource(struct dlm_lock_resource *res);
730void __dlm_print_one_lock_resource(struct dlm_lock_resource *res);
731
732u8 dlm_nm_this_node(struct dlm_ctxt *dlm);
733void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
734void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
735
736
737int dlm_nm_init(struct dlm_ctxt *dlm);
738int dlm_heartbeat_init(struct dlm_ctxt *dlm);
739void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data);
740void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data);
741
742int dlm_lockres_is_dirty(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
743int dlm_migrate_lockres(struct dlm_ctxt *dlm,
744 struct dlm_lock_resource *res,
745 u8 target);
746int dlm_finish_migration(struct dlm_ctxt *dlm,
747 struct dlm_lock_resource *res,
748 u8 old_master);
749void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
750 struct dlm_lock_resource *res);
751void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res);
752
753int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data);
754int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data);
755int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data);
756int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data);
757int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data);
758int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data);
759int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data);
760int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data);
761int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data);
762
763int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
764 struct dlm_lock_resource *res,
765 int ignore_higher,
766 u8 request_from,
767 u32 flags);
768
769
770int dlm_send_one_lockres(struct dlm_ctxt *dlm,
771 struct dlm_lock_resource *res,
772 struct dlm_migratable_lockres *mres,
773 u8 send_to,
774 u8 flags);
775void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
776 struct dlm_lock_resource *res);
777
778/* will exit holding res->spinlock, but may drop in function */
779void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags);
780void __dlm_wait_on_lockres_flags_set(struct dlm_lock_resource *res, int flags);
781
782/* will exit holding res->spinlock, but may drop in function */
783static inline void __dlm_wait_on_lockres(struct dlm_lock_resource *res)
784{
785 __dlm_wait_on_lockres_flags(res, (DLM_LOCK_RES_IN_PROGRESS|
786 DLM_LOCK_RES_RECOVERING|
787 DLM_LOCK_RES_MIGRATING));
788}
789
790
791int dlm_init_mle_cache(void);
792void dlm_destroy_mle_cache(void);
793void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up);
794void dlm_clean_master_list(struct dlm_ctxt *dlm,
795 u8 dead_node);
796int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock);
797
798
799static inline const char * dlm_lock_mode_name(int mode)
800{
801 switch (mode) {
802 case LKM_EXMODE:
803 return "EX";
804 case LKM_PRMODE:
805 return "PR";
806 case LKM_NLMODE:
807 return "NL";
808 }
809 return "UNKNOWN";
810}
811
812
813static inline int dlm_lock_compatible(int existing, int request)
814{
815 /* NO_LOCK compatible with all */
816 if (request == LKM_NLMODE ||
817 existing == LKM_NLMODE)
818 return 1;
819
820 /* EX incompatible with all non-NO_LOCK */
821 if (request == LKM_EXMODE)
822 return 0;
823
824 /* request must be PR, which is compatible with PR */
825 if (existing == LKM_PRMODE)
826 return 1;
827
828 return 0;
829}
830
831static inline int dlm_lock_on_list(struct list_head *head,
832 struct dlm_lock *lock)
833{
834 struct list_head *iter;
835 struct dlm_lock *tmplock;
836
837 list_for_each(iter, head) {
838 tmplock = list_entry(iter, struct dlm_lock, list);
839 if (tmplock == lock)
840 return 1;
841 }
842 return 0;
843}
844
845
846static inline enum dlm_status dlm_err_to_dlm_status(int err)
847{
848 enum dlm_status ret;
849 if (err == -ENOMEM)
850 ret = DLM_SYSERR;
851 else if (err == -ETIMEDOUT || o2net_link_down(err, NULL))
852 ret = DLM_NOLOCKMGR;
853 else if (err == -EINVAL)
854 ret = DLM_BADPARAM;
855 else if (err == -ENAMETOOLONG)
856 ret = DLM_IVBUFLEN;
857 else
858 ret = DLM_BADARGS;
859 return ret;
860}
861
862
863static inline void dlm_node_iter_init(unsigned long *map,
864 struct dlm_node_iter *iter)
865{
866 memcpy(iter->node_map, map, sizeof(iter->node_map));
867 iter->curnode = -1;
868}
869
870static inline int dlm_node_iter_next(struct dlm_node_iter *iter)
871{
872 int bit;
873 bit = find_next_bit(iter->node_map, O2NM_MAX_NODES, iter->curnode+1);
874 if (bit >= O2NM_MAX_NODES) {
875 iter->curnode = O2NM_MAX_NODES;
876 return -ENOENT;
877 }
878 iter->curnode = bit;
879 return bit;
880}
881
882
883
884#endif /* DLMCOMMON_H */
diff --git a/fs/ocfs2/dlm/dlmconvert.c b/fs/ocfs2/dlm/dlmconvert.c
new file mode 100644
index 000000000000..6001b22a997d
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmconvert.c
@@ -0,0 +1,530 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmconvert.c
5 *
6 * underlying calls for lock conversion
7 *
8 * Copyright (C) 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 *
25 */
26
27
28#include <linux/module.h>
29#include <linux/fs.h>
30#include <linux/types.h>
31#include <linux/slab.h>
32#include <linux/highmem.h>
33#include <linux/utsname.h>
34#include <linux/init.h>
35#include <linux/sysctl.h>
36#include <linux/random.h>
37#include <linux/blkdev.h>
38#include <linux/socket.h>
39#include <linux/inet.h>
40#include <linux/spinlock.h>
41
42
43#include "cluster/heartbeat.h"
44#include "cluster/nodemanager.h"
45#include "cluster/tcp.h"
46
47#include "dlmapi.h"
48#include "dlmcommon.h"
49
50#include "dlmconvert.h"
51
52#define MLOG_MASK_PREFIX ML_DLM
53#include "cluster/masklog.h"
54
55/* NOTE: __dlmconvert_master is the only function in here that
56 * needs a spinlock held on entry (res->spinlock) and it is the
57 * only one that holds a lock on exit (res->spinlock).
58 * All other functions in here need no locks and drop all of
59 * the locks that they acquire. */
60static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm,
61 struct dlm_lock_resource *res,
62 struct dlm_lock *lock, int flags,
63 int type, int *call_ast,
64 int *kick_thread);
65static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm,
66 struct dlm_lock_resource *res,
67 struct dlm_lock *lock, int flags, int type);
68
69/*
70 * this is only called directly by dlmlock(), and only when the
71 * local node is the owner of the lockres
72 * locking:
73 * caller needs: none
74 * taken: takes and drops res->spinlock
75 * held on exit: none
76 * returns: see __dlmconvert_master
77 */
78enum dlm_status dlmconvert_master(struct dlm_ctxt *dlm,
79 struct dlm_lock_resource *res,
80 struct dlm_lock *lock, int flags, int type)
81{
82 int call_ast = 0, kick_thread = 0;
83 enum dlm_status status;
84
85 spin_lock(&res->spinlock);
86 /* we are not in a network handler, this is fine */
87 __dlm_wait_on_lockres(res);
88 __dlm_lockres_reserve_ast(res);
89 res->state |= DLM_LOCK_RES_IN_PROGRESS;
90
91 status = __dlmconvert_master(dlm, res, lock, flags, type,
92 &call_ast, &kick_thread);
93
94 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
95 spin_unlock(&res->spinlock);
96 wake_up(&res->wq);
97 if (status != DLM_NORMAL && status != DLM_NOTQUEUED)
98 dlm_error(status);
99
100 /* either queue the ast or release it */
101 if (call_ast)
102 dlm_queue_ast(dlm, lock);
103 else
104 dlm_lockres_release_ast(dlm, res);
105
106 if (kick_thread)
107 dlm_kick_thread(dlm, res);
108
109 return status;
110}
111
112/* performs lock conversion at the lockres master site
113 * locking:
114 * caller needs: res->spinlock
115 * taken: takes and drops lock->spinlock
116 * held on exit: res->spinlock
117 * returns: DLM_NORMAL, DLM_NOTQUEUED, DLM_DENIED
118 * call_ast: whether ast should be called for this lock
119 * kick_thread: whether dlm_kick_thread should be called
120 */
121static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm,
122 struct dlm_lock_resource *res,
123 struct dlm_lock *lock, int flags,
124 int type, int *call_ast,
125 int *kick_thread)
126{
127 enum dlm_status status = DLM_NORMAL;
128 struct list_head *iter;
129 struct dlm_lock *tmplock=NULL;
130
131 assert_spin_locked(&res->spinlock);
132
133 mlog_entry("type=%d, convert_type=%d, new convert_type=%d\n",
134 lock->ml.type, lock->ml.convert_type, type);
135
136 spin_lock(&lock->spinlock);
137
138 /* already converting? */
139 if (lock->ml.convert_type != LKM_IVMODE) {
140 mlog(ML_ERROR, "attempted to convert a lock with a lock "
141 "conversion pending\n");
142 status = DLM_DENIED;
143 goto unlock_exit;
144 }
145
146 /* must be on grant queue to convert */
147 if (!dlm_lock_on_list(&res->granted, lock)) {
148 mlog(ML_ERROR, "attempted to convert a lock not on grant "
149 "queue\n");
150 status = DLM_DENIED;
151 goto unlock_exit;
152 }
153
154 if (flags & LKM_VALBLK) {
155 switch (lock->ml.type) {
156 case LKM_EXMODE:
157 /* EX + LKM_VALBLK + convert == set lvb */
158 mlog(0, "will set lvb: converting %s->%s\n",
159 dlm_lock_mode_name(lock->ml.type),
160 dlm_lock_mode_name(type));
161 lock->lksb->flags |= DLM_LKSB_PUT_LVB;
162 break;
163 case LKM_PRMODE:
164 case LKM_NLMODE:
165 /* refetch if new level is not NL */
166 if (type > LKM_NLMODE) {
167 mlog(0, "will fetch new value into "
168 "lvb: converting %s->%s\n",
169 dlm_lock_mode_name(lock->ml.type),
170 dlm_lock_mode_name(type));
171 lock->lksb->flags |= DLM_LKSB_GET_LVB;
172 } else {
173 mlog(0, "will NOT fetch new value "
174 "into lvb: converting %s->%s\n",
175 dlm_lock_mode_name(lock->ml.type),
176 dlm_lock_mode_name(type));
177 flags &= ~(LKM_VALBLK);
178 }
179 break;
180 }
181 }
182
183
184 /* in-place downconvert? */
185 if (type <= lock->ml.type)
186 goto grant;
187
188 /* upconvert from here on */
189 status = DLM_NORMAL;
190 list_for_each(iter, &res->granted) {
191 tmplock = list_entry(iter, struct dlm_lock, list);
192 if (tmplock == lock)
193 continue;
194 if (!dlm_lock_compatible(tmplock->ml.type, type))
195 goto switch_queues;
196 }
197
198 list_for_each(iter, &res->converting) {
199 tmplock = list_entry(iter, struct dlm_lock, list);
200 if (!dlm_lock_compatible(tmplock->ml.type, type))
201 goto switch_queues;
202 /* existing conversion requests take precedence */
203 if (!dlm_lock_compatible(tmplock->ml.convert_type, type))
204 goto switch_queues;
205 }
206
207 /* fall thru to grant */
208
209grant:
210 mlog(0, "res %.*s, granting %s lock\n", res->lockname.len,
211 res->lockname.name, dlm_lock_mode_name(type));
212 /* immediately grant the new lock type */
213 lock->lksb->status = DLM_NORMAL;
214 if (lock->ml.node == dlm->node_num)
215 mlog(0, "doing in-place convert for nonlocal lock\n");
216 lock->ml.type = type;
217 status = DLM_NORMAL;
218 *call_ast = 1;
219 goto unlock_exit;
220
221switch_queues:
222 if (flags & LKM_NOQUEUE) {
223 mlog(0, "failed to convert NOQUEUE lock %.*s from "
224 "%d to %d...\n", res->lockname.len, res->lockname.name,
225 lock->ml.type, type);
226 status = DLM_NOTQUEUED;
227 goto unlock_exit;
228 }
229 mlog(0, "res %.*s, queueing...\n", res->lockname.len,
230 res->lockname.name);
231
232 lock->ml.convert_type = type;
233 /* do not alter lock refcount. switching lists. */
234 list_del_init(&lock->list);
235 list_add_tail(&lock->list, &res->converting);
236
237unlock_exit:
238 spin_unlock(&lock->spinlock);
239 if (status == DLM_DENIED) {
240 __dlm_print_one_lock_resource(res);
241 }
242 if (status == DLM_NORMAL)
243 *kick_thread = 1;
244 return status;
245}
246
247void dlm_revert_pending_convert(struct dlm_lock_resource *res,
248 struct dlm_lock *lock)
249{
250 /* do not alter lock refcount. switching lists. */
251 list_del_init(&lock->list);
252 list_add_tail(&lock->list, &res->granted);
253 lock->ml.convert_type = LKM_IVMODE;
254 lock->lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB);
255}
256
257/* messages the master site to do lock conversion
258 * locking:
259 * caller needs: none
260 * taken: takes and drops res->spinlock, uses DLM_LOCK_RES_IN_PROGRESS
261 * held on exit: none
262 * returns: DLM_NORMAL, DLM_RECOVERING, status from remote node
263 */
264enum dlm_status dlmconvert_remote(struct dlm_ctxt *dlm,
265 struct dlm_lock_resource *res,
266 struct dlm_lock *lock, int flags, int type)
267{
268 enum dlm_status status;
269
270 mlog(0, "type=%d, convert_type=%d, busy=%d\n", lock->ml.type,
271 lock->ml.convert_type, res->state & DLM_LOCK_RES_IN_PROGRESS);
272
273 spin_lock(&res->spinlock);
274 if (res->state & DLM_LOCK_RES_RECOVERING) {
275 mlog(0, "bailing out early since res is RECOVERING "
276 "on secondary queue\n");
277 /* __dlm_print_one_lock_resource(res); */
278 status = DLM_RECOVERING;
279 goto bail;
280 }
281 /* will exit this call with spinlock held */
282 __dlm_wait_on_lockres(res);
283
284 if (lock->ml.convert_type != LKM_IVMODE) {
285 __dlm_print_one_lock_resource(res);
286 mlog(ML_ERROR, "converting a remote lock that is already "
287 "converting! (cookie=%"MLFu64", conv=%d)\n",
288 lock->ml.cookie, lock->ml.convert_type);
289 status = DLM_DENIED;
290 goto bail;
291 }
292 res->state |= DLM_LOCK_RES_IN_PROGRESS;
293 /* move lock to local convert queue */
294 /* do not alter lock refcount. switching lists. */
295 list_del_init(&lock->list);
296 list_add_tail(&lock->list, &res->converting);
297 lock->convert_pending = 1;
298 lock->ml.convert_type = type;
299
300 if (flags & LKM_VALBLK) {
301 if (lock->ml.type == LKM_EXMODE) {
302 flags |= LKM_PUT_LVB;
303 lock->lksb->flags |= DLM_LKSB_PUT_LVB;
304 } else {
305 if (lock->ml.convert_type == LKM_NLMODE)
306 flags &= ~LKM_VALBLK;
307 else {
308 flags |= LKM_GET_LVB;
309 lock->lksb->flags |= DLM_LKSB_GET_LVB;
310 }
311 }
312 }
313 spin_unlock(&res->spinlock);
314
315 /* no locks held here.
316 * need to wait for a reply as to whether it got queued or not. */
317 status = dlm_send_remote_convert_request(dlm, res, lock, flags, type);
318
319 spin_lock(&res->spinlock);
320 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
321 lock->convert_pending = 0;
322 /* if it failed, move it back to granted queue */
323 if (status != DLM_NORMAL) {
324 if (status != DLM_NOTQUEUED)
325 dlm_error(status);
326 dlm_revert_pending_convert(res, lock);
327 }
328bail:
329 spin_unlock(&res->spinlock);
330
331 /* TODO: should this be a wake_one? */
332 /* wake up any IN_PROGRESS waiters */
333 wake_up(&res->wq);
334
335 return status;
336}
337
338/* sends DLM_CONVERT_LOCK_MSG to master site
339 * locking:
340 * caller needs: none
341 * taken: none
342 * held on exit: none
343 * returns: DLM_NOLOCKMGR, status from remote node
344 */
345static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm,
346 struct dlm_lock_resource *res,
347 struct dlm_lock *lock, int flags, int type)
348{
349 struct dlm_convert_lock convert;
350 int tmpret;
351 enum dlm_status ret;
352 int status = 0;
353 struct kvec vec[2];
354 size_t veclen = 1;
355
356 mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
357
358 memset(&convert, 0, sizeof(struct dlm_convert_lock));
359 convert.node_idx = dlm->node_num;
360 convert.requested_type = type;
361 convert.cookie = lock->ml.cookie;
362 convert.namelen = res->lockname.len;
363 convert.flags = cpu_to_be32(flags);
364 memcpy(convert.name, res->lockname.name, convert.namelen);
365
366 vec[0].iov_len = sizeof(struct dlm_convert_lock);
367 vec[0].iov_base = &convert;
368
369 if (flags & LKM_PUT_LVB) {
370 /* extra data to send if we are updating lvb */
371 vec[1].iov_len = DLM_LVB_LEN;
372 vec[1].iov_base = lock->lksb->lvb;
373 veclen++;
374 }
375
376 tmpret = o2net_send_message_vec(DLM_CONVERT_LOCK_MSG, dlm->key,
377 vec, veclen, res->owner, &status);
378 if (tmpret >= 0) {
379 // successfully sent and received
380 ret = status; // this is already a dlm_status
381 if (ret == DLM_RECOVERING) {
382 mlog(0, "node %u returned DLM_RECOVERING from convert "
383 "message!\n", res->owner);
384 } else if (ret == DLM_MIGRATING) {
385 mlog(0, "node %u returned DLM_MIGRATING from convert "
386 "message!\n", res->owner);
387 } else if (ret == DLM_FORWARD) {
388 mlog(0, "node %u returned DLM_FORWARD from convert "
389 "message!\n", res->owner);
390 } else if (ret != DLM_NORMAL && ret != DLM_NOTQUEUED)
391 dlm_error(ret);
392 } else {
393 mlog_errno(tmpret);
394 if (dlm_is_host_down(tmpret)) {
395 ret = DLM_RECOVERING;
396 mlog(0, "node %u died so returning DLM_RECOVERING "
397 "from convert message!\n", res->owner);
398 } else {
399 ret = dlm_err_to_dlm_status(tmpret);
400 }
401 }
402
403 return ret;
404}
405
406/* handler for DLM_CONVERT_LOCK_MSG on master site
407 * locking:
408 * caller needs: none
409 * taken: takes and drop res->spinlock
410 * held on exit: none
411 * returns: DLM_NORMAL, DLM_IVLOCKID, DLM_BADARGS,
412 * status from __dlmconvert_master
413 */
414int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data)
415{
416 struct dlm_ctxt *dlm = data;
417 struct dlm_convert_lock *cnv = (struct dlm_convert_lock *)msg->buf;
418 struct dlm_lock_resource *res = NULL;
419 struct list_head *iter;
420 struct dlm_lock *lock = NULL;
421 struct dlm_lockstatus *lksb;
422 enum dlm_status status = DLM_NORMAL;
423 u32 flags;
424 int call_ast = 0, kick_thread = 0;
425
426 if (!dlm_grab(dlm)) {
427 dlm_error(DLM_REJECTED);
428 return DLM_REJECTED;
429 }
430
431 mlog_bug_on_msg(!dlm_domain_fully_joined(dlm),
432 "Domain %s not fully joined!\n", dlm->name);
433
434 if (cnv->namelen > DLM_LOCKID_NAME_MAX) {
435 status = DLM_IVBUFLEN;
436 dlm_error(status);
437 goto leave;
438 }
439
440 flags = be32_to_cpu(cnv->flags);
441
442 if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) ==
443 (LKM_PUT_LVB|LKM_GET_LVB)) {
444 mlog(ML_ERROR, "both PUT and GET lvb specified\n");
445 status = DLM_BADARGS;
446 goto leave;
447 }
448
449 mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" :
450 (flags & LKM_GET_LVB ? "get lvb" : "none"));
451
452 status = DLM_IVLOCKID;
453 res = dlm_lookup_lockres(dlm, cnv->name, cnv->namelen);
454 if (!res) {
455 dlm_error(status);
456 goto leave;
457 }
458
459 spin_lock(&res->spinlock);
460 list_for_each(iter, &res->granted) {
461 lock = list_entry(iter, struct dlm_lock, list);
462 if (lock->ml.cookie == cnv->cookie &&
463 lock->ml.node == cnv->node_idx) {
464 dlm_lock_get(lock);
465 break;
466 }
467 lock = NULL;
468 }
469 spin_unlock(&res->spinlock);
470 if (!lock) {
471 status = DLM_IVLOCKID;
472 dlm_error(status);
473 goto leave;
474 }
475
476 /* found the lock */
477 lksb = lock->lksb;
478
479 /* see if caller needed to get/put lvb */
480 if (flags & LKM_PUT_LVB) {
481 BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
482 lksb->flags |= DLM_LKSB_PUT_LVB;
483 memcpy(&lksb->lvb[0], &cnv->lvb[0], DLM_LVB_LEN);
484 } else if (flags & LKM_GET_LVB) {
485 BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
486 lksb->flags |= DLM_LKSB_GET_LVB;
487 }
488
489 spin_lock(&res->spinlock);
490 status = __dlm_lockres_state_to_status(res);
491 if (status == DLM_NORMAL) {
492 __dlm_lockres_reserve_ast(res);
493 res->state |= DLM_LOCK_RES_IN_PROGRESS;
494 status = __dlmconvert_master(dlm, res, lock, flags,
495 cnv->requested_type,
496 &call_ast, &kick_thread);
497 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
498 }
499 spin_unlock(&res->spinlock);
500
501 if (status != DLM_NORMAL) {
502 if (status != DLM_NOTQUEUED)
503 dlm_error(status);
504 lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB);
505 }
506
507leave:
508 if (!lock)
509 mlog(ML_ERROR, "did not find lock to convert on grant queue! "
510 "cookie=%"MLFu64"\n",
511 cnv->cookie);
512 else
513 dlm_lock_put(lock);
514
515 /* either queue the ast or release it */
516 if (call_ast)
517 dlm_queue_ast(dlm, lock);
518 else
519 dlm_lockres_release_ast(dlm, res);
520
521 if (kick_thread)
522 dlm_kick_thread(dlm, res);
523
524 if (res)
525 dlm_lockres_put(res);
526
527 dlm_put(dlm);
528
529 return status;
530}
diff --git a/fs/ocfs2/dlm/dlmconvert.h b/fs/ocfs2/dlm/dlmconvert.h
new file mode 100644
index 000000000000..b2e3677df878
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmconvert.h
@@ -0,0 +1,35 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmconvert.h
5 *
6 * Copyright (C) 2004 Oracle. All rights reserved.
7 *
8 * This program is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU General Public
10 * License as published by the Free Software Foundation; either
11 * version 2 of the License, or (at your option) any later version.
12 *
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public
19 * License along with this program; if not, write to the
20 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
21 * Boston, MA 021110-1307, USA.
22 *
23 */
24
25#ifndef DLMCONVERT_H
26#define DLMCONVERT_H
27
28enum dlm_status dlmconvert_master(struct dlm_ctxt *dlm,
29 struct dlm_lock_resource *res,
30 struct dlm_lock *lock, int flags, int type);
31enum dlm_status dlmconvert_remote(struct dlm_ctxt *dlm,
32 struct dlm_lock_resource *res,
33 struct dlm_lock *lock, int flags, int type);
34
35#endif
diff --git a/fs/ocfs2/dlm/dlmdebug.c b/fs/ocfs2/dlm/dlmdebug.c
new file mode 100644
index 000000000000..f339fe27975a
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmdebug.c
@@ -0,0 +1,246 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmdebug.c
5 *
6 * debug functionality for the dlm
7 *
8 * Copyright (C) 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 *
25 */
26
27#include <linux/types.h>
28#include <linux/slab.h>
29#include <linux/highmem.h>
30#include <linux/utsname.h>
31#include <linux/sysctl.h>
32#include <linux/spinlock.h>
33
34#include "cluster/heartbeat.h"
35#include "cluster/nodemanager.h"
36#include "cluster/tcp.h"
37
38#include "dlmapi.h"
39#include "dlmcommon.h"
40#include "dlmdebug.h"
41
42#include "dlmdomain.h"
43#include "dlmdebug.h"
44
45#define MLOG_MASK_PREFIX ML_DLM
46#include "cluster/masklog.h"
47
48void dlm_print_one_lock_resource(struct dlm_lock_resource *res)
49{
50 mlog(ML_NOTICE, "lockres: %.*s, owner=%u, state=%u\n",
51 res->lockname.len, res->lockname.name,
52 res->owner, res->state);
53 spin_lock(&res->spinlock);
54 __dlm_print_one_lock_resource(res);
55 spin_unlock(&res->spinlock);
56}
57
58void __dlm_print_one_lock_resource(struct dlm_lock_resource *res)
59{
60 struct list_head *iter2;
61 struct dlm_lock *lock;
62
63 assert_spin_locked(&res->spinlock);
64
65 mlog(ML_NOTICE, "lockres: %.*s, owner=%u, state=%u\n",
66 res->lockname.len, res->lockname.name,
67 res->owner, res->state);
68 mlog(ML_NOTICE, " last used: %lu, on purge list: %s\n",
69 res->last_used, list_empty(&res->purge) ? "no" : "yes");
70 mlog(ML_NOTICE, " granted queue: \n");
71 list_for_each(iter2, &res->granted) {
72 lock = list_entry(iter2, struct dlm_lock, list);
73 spin_lock(&lock->spinlock);
74 mlog(ML_NOTICE, " type=%d, conv=%d, node=%u, "
75 "cookie=%"MLFu64", ast=(empty=%c,pend=%c), bast=(empty=%c,pend=%c)\n",
76 lock->ml.type, lock->ml.convert_type, lock->ml.node, lock->ml.cookie,
77 list_empty(&lock->ast_list) ? 'y' : 'n',
78 lock->ast_pending ? 'y' : 'n',
79 list_empty(&lock->bast_list) ? 'y' : 'n',
80 lock->bast_pending ? 'y' : 'n');
81 spin_unlock(&lock->spinlock);
82 }
83 mlog(ML_NOTICE, " converting queue: \n");
84 list_for_each(iter2, &res->converting) {
85 lock = list_entry(iter2, struct dlm_lock, list);
86 spin_lock(&lock->spinlock);
87 mlog(ML_NOTICE, " type=%d, conv=%d, node=%u, "
88 "cookie=%"MLFu64", ast=(empty=%c,pend=%c), bast=(empty=%c,pend=%c)\n",
89 lock->ml.type, lock->ml.convert_type, lock->ml.node, lock->ml.cookie,
90 list_empty(&lock->ast_list) ? 'y' : 'n',
91 lock->ast_pending ? 'y' : 'n',
92 list_empty(&lock->bast_list) ? 'y' : 'n',
93 lock->bast_pending ? 'y' : 'n');
94 spin_unlock(&lock->spinlock);
95 }
96 mlog(ML_NOTICE, " blocked queue: \n");
97 list_for_each(iter2, &res->blocked) {
98 lock = list_entry(iter2, struct dlm_lock, list);
99 spin_lock(&lock->spinlock);
100 mlog(ML_NOTICE, " type=%d, conv=%d, node=%u, "
101 "cookie=%"MLFu64", ast=(empty=%c,pend=%c), bast=(empty=%c,pend=%c)\n",
102 lock->ml.type, lock->ml.convert_type, lock->ml.node, lock->ml.cookie,
103 list_empty(&lock->ast_list) ? 'y' : 'n',
104 lock->ast_pending ? 'y' : 'n',
105 list_empty(&lock->bast_list) ? 'y' : 'n',
106 lock->bast_pending ? 'y' : 'n');
107 spin_unlock(&lock->spinlock);
108 }
109}
110
111void dlm_print_one_lock(struct dlm_lock *lockid)
112{
113 dlm_print_one_lock_resource(lockid->lockres);
114}
115EXPORT_SYMBOL_GPL(dlm_print_one_lock);
116
117void dlm_dump_lock_resources(struct dlm_ctxt *dlm)
118{
119 struct dlm_lock_resource *res;
120 struct list_head *iter;
121 struct list_head *bucket;
122 int i;
123
124 mlog(ML_NOTICE, "struct dlm_ctxt: %s, node=%u, key=%u\n",
125 dlm->name, dlm->node_num, dlm->key);
126 if (!dlm || !dlm->name) {
127 mlog(ML_ERROR, "dlm=%p\n", dlm);
128 return;
129 }
130
131 spin_lock(&dlm->spinlock);
132 for (i=0; i<DLM_HASH_SIZE; i++) {
133 bucket = &(dlm->resources[i]);
134 list_for_each(iter, bucket) {
135 res = list_entry(iter, struct dlm_lock_resource, list);
136 dlm_print_one_lock_resource(res);
137 }
138 }
139 spin_unlock(&dlm->spinlock);
140}
141
142static const char *dlm_errnames[] = {
143 [DLM_NORMAL] = "DLM_NORMAL",
144 [DLM_GRANTED] = "DLM_GRANTED",
145 [DLM_DENIED] = "DLM_DENIED",
146 [DLM_DENIED_NOLOCKS] = "DLM_DENIED_NOLOCKS",
147 [DLM_WORKING] = "DLM_WORKING",
148 [DLM_BLOCKED] = "DLM_BLOCKED",
149 [DLM_BLOCKED_ORPHAN] = "DLM_BLOCKED_ORPHAN",
150 [DLM_DENIED_GRACE_PERIOD] = "DLM_DENIED_GRACE_PERIOD",
151 [DLM_SYSERR] = "DLM_SYSERR",
152 [DLM_NOSUPPORT] = "DLM_NOSUPPORT",
153 [DLM_CANCELGRANT] = "DLM_CANCELGRANT",
154 [DLM_IVLOCKID] = "DLM_IVLOCKID",
155 [DLM_SYNC] = "DLM_SYNC",
156 [DLM_BADTYPE] = "DLM_BADTYPE",
157 [DLM_BADRESOURCE] = "DLM_BADRESOURCE",
158 [DLM_MAXHANDLES] = "DLM_MAXHANDLES",
159 [DLM_NOCLINFO] = "DLM_NOCLINFO",
160 [DLM_NOLOCKMGR] = "DLM_NOLOCKMGR",
161 [DLM_NOPURGED] = "DLM_NOPURGED",
162 [DLM_BADARGS] = "DLM_BADARGS",
163 [DLM_VOID] = "DLM_VOID",
164 [DLM_NOTQUEUED] = "DLM_NOTQUEUED",
165 [DLM_IVBUFLEN] = "DLM_IVBUFLEN",
166 [DLM_CVTUNGRANT] = "DLM_CVTUNGRANT",
167 [DLM_BADPARAM] = "DLM_BADPARAM",
168 [DLM_VALNOTVALID] = "DLM_VALNOTVALID",
169 [DLM_REJECTED] = "DLM_REJECTED",
170 [DLM_ABORT] = "DLM_ABORT",
171 [DLM_CANCEL] = "DLM_CANCEL",
172 [DLM_IVRESHANDLE] = "DLM_IVRESHANDLE",
173 [DLM_DEADLOCK] = "DLM_DEADLOCK",
174 [DLM_DENIED_NOASTS] = "DLM_DENIED_NOASTS",
175 [DLM_FORWARD] = "DLM_FORWARD",
176 [DLM_TIMEOUT] = "DLM_TIMEOUT",
177 [DLM_IVGROUPID] = "DLM_IVGROUPID",
178 [DLM_VERS_CONFLICT] = "DLM_VERS_CONFLICT",
179 [DLM_BAD_DEVICE_PATH] = "DLM_BAD_DEVICE_PATH",
180 [DLM_NO_DEVICE_PERMISSION] = "DLM_NO_DEVICE_PERMISSION",
181 [DLM_NO_CONTROL_DEVICE ] = "DLM_NO_CONTROL_DEVICE ",
182 [DLM_RECOVERING] = "DLM_RECOVERING",
183 [DLM_MIGRATING] = "DLM_MIGRATING",
184 [DLM_MAXSTATS] = "DLM_MAXSTATS",
185};
186
187static const char *dlm_errmsgs[] = {
188 [DLM_NORMAL] = "request in progress",
189 [DLM_GRANTED] = "request granted",
190 [DLM_DENIED] = "request denied",
191 [DLM_DENIED_NOLOCKS] = "request denied, out of system resources",
192 [DLM_WORKING] = "async request in progress",
193 [DLM_BLOCKED] = "lock request blocked",
194 [DLM_BLOCKED_ORPHAN] = "lock request blocked by a orphan lock",
195 [DLM_DENIED_GRACE_PERIOD] = "topological change in progress",
196 [DLM_SYSERR] = "system error",
197 [DLM_NOSUPPORT] = "unsupported",
198 [DLM_CANCELGRANT] = "can't cancel convert: already granted",
199 [DLM_IVLOCKID] = "bad lockid",
200 [DLM_SYNC] = "synchronous request granted",
201 [DLM_BADTYPE] = "bad resource type",
202 [DLM_BADRESOURCE] = "bad resource handle",
203 [DLM_MAXHANDLES] = "no more resource handles",
204 [DLM_NOCLINFO] = "can't contact cluster manager",
205 [DLM_NOLOCKMGR] = "can't contact lock manager",
206 [DLM_NOPURGED] = "can't contact purge daemon",
207 [DLM_BADARGS] = "bad api args",
208 [DLM_VOID] = "no status",
209 [DLM_NOTQUEUED] = "NOQUEUE was specified and request failed",
210 [DLM_IVBUFLEN] = "invalid resource name length",
211 [DLM_CVTUNGRANT] = "attempted to convert ungranted lock",
212 [DLM_BADPARAM] = "invalid lock mode specified",
213 [DLM_VALNOTVALID] = "value block has been invalidated",
214 [DLM_REJECTED] = "request rejected, unrecognized client",
215 [DLM_ABORT] = "blocked lock request cancelled",
216 [DLM_CANCEL] = "conversion request cancelled",
217 [DLM_IVRESHANDLE] = "invalid resource handle",
218 [DLM_DEADLOCK] = "deadlock recovery refused this request",
219 [DLM_DENIED_NOASTS] = "failed to allocate AST",
220 [DLM_FORWARD] = "request must wait for primary's response",
221 [DLM_TIMEOUT] = "timeout value for lock has expired",
222 [DLM_IVGROUPID] = "invalid group specification",
223 [DLM_VERS_CONFLICT] = "version conflicts prevent request handling",
224 [DLM_BAD_DEVICE_PATH] = "Locks device does not exist or path wrong",
225 [DLM_NO_DEVICE_PERMISSION] = "Client has insufficient perms for device",
226 [DLM_NO_CONTROL_DEVICE] = "Cannot set options on opened device ",
227 [DLM_RECOVERING] = "lock resource being recovered",
228 [DLM_MIGRATING] = "lock resource being migrated",
229 [DLM_MAXSTATS] = "invalid error number",
230};
231
232const char *dlm_errmsg(enum dlm_status err)
233{
234 if (err >= DLM_MAXSTATS || err < 0)
235 return dlm_errmsgs[DLM_MAXSTATS];
236 return dlm_errmsgs[err];
237}
238EXPORT_SYMBOL_GPL(dlm_errmsg);
239
240const char *dlm_errname(enum dlm_status err)
241{
242 if (err >= DLM_MAXSTATS || err < 0)
243 return dlm_errnames[DLM_MAXSTATS];
244 return dlm_errnames[err];
245}
246EXPORT_SYMBOL_GPL(dlm_errname);
diff --git a/fs/ocfs2/dlm/dlmdebug.h b/fs/ocfs2/dlm/dlmdebug.h
new file mode 100644
index 000000000000..6858510c3ccd
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmdebug.h
@@ -0,0 +1,30 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmdebug.h
5 *
6 * Copyright (C) 2004 Oracle. All rights reserved.
7 *
8 * This program is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU General Public
10 * License as published by the Free Software Foundation; either
11 * version 2 of the License, or (at your option) any later version.
12 *
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public
19 * License along with this program; if not, write to the
20 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
21 * Boston, MA 021110-1307, USA.
22 *
23 */
24
25#ifndef DLMDEBUG_H
26#define DLMDEBUG_H
27
28void dlm_dump_lock_resources(struct dlm_ctxt *dlm);
29
30#endif
diff --git a/fs/ocfs2/dlm/dlmdomain.c b/fs/ocfs2/dlm/dlmdomain.c
new file mode 100644
index 000000000000..da3c22045f89
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmdomain.c
@@ -0,0 +1,1469 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmdomain.c
5 *
6 * defines domain join / leave apis
7 *
8 * Copyright (C) 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 *
25 */
26
27#include <linux/module.h>
28#include <linux/types.h>
29#include <linux/slab.h>
30#include <linux/highmem.h>
31#include <linux/utsname.h>
32#include <linux/init.h>
33#include <linux/spinlock.h>
34#include <linux/delay.h>
35#include <linux/err.h>
36
37#include "cluster/heartbeat.h"
38#include "cluster/nodemanager.h"
39#include "cluster/tcp.h"
40
41#include "dlmapi.h"
42#include "dlmcommon.h"
43
44#include "dlmdebug.h"
45#include "dlmdomain.h"
46
47#include "dlmver.h"
48
49#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
50#include "cluster/masklog.h"
51
52/*
53 *
54 * spinlock lock ordering: if multiple locks are needed, obey this ordering:
55 * dlm_domain_lock
56 * struct dlm_ctxt->spinlock
57 * struct dlm_lock_resource->spinlock
58 * struct dlm_ctxt->master_lock
59 * struct dlm_ctxt->ast_lock
60 * dlm_master_list_entry->spinlock
61 * dlm_lock->spinlock
62 *
63 */
64
65spinlock_t dlm_domain_lock = SPIN_LOCK_UNLOCKED;
66LIST_HEAD(dlm_domains);
67static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);
68
69#define DLM_DOMAIN_BACKOFF_MS 200
70
71static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data);
72static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data);
73static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data);
74static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data);
75
76static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
77
78void __dlm_unhash_lockres(struct dlm_lock_resource *lockres)
79{
80 list_del_init(&lockres->list);
81 dlm_lockres_put(lockres);
82}
83
84void __dlm_insert_lockres(struct dlm_ctxt *dlm,
85 struct dlm_lock_resource *res)
86{
87 struct list_head *bucket;
88 struct qstr *q;
89
90 assert_spin_locked(&dlm->spinlock);
91
92 q = &res->lockname;
93 q->hash = full_name_hash(q->name, q->len);
94 bucket = &(dlm->resources[q->hash & DLM_HASH_MASK]);
95
96 /* get a reference for our hashtable */
97 dlm_lockres_get(res);
98
99 list_add_tail(&res->list, bucket);
100}
101
102struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
103 const char *name,
104 unsigned int len)
105{
106 unsigned int hash;
107 struct list_head *iter;
108 struct dlm_lock_resource *tmpres=NULL;
109 struct list_head *bucket;
110
111 mlog_entry("%.*s\n", len, name);
112
113 assert_spin_locked(&dlm->spinlock);
114
115 hash = full_name_hash(name, len);
116
117 bucket = &(dlm->resources[hash & DLM_HASH_MASK]);
118
119 /* check for pre-existing lock */
120 list_for_each(iter, bucket) {
121 tmpres = list_entry(iter, struct dlm_lock_resource, list);
122 if (tmpres->lockname.len == len &&
123 memcmp(tmpres->lockname.name, name, len) == 0) {
124 dlm_lockres_get(tmpres);
125 break;
126 }
127
128 tmpres = NULL;
129 }
130 return tmpres;
131}
132
133struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
134 const char *name,
135 unsigned int len)
136{
137 struct dlm_lock_resource *res;
138
139 spin_lock(&dlm->spinlock);
140 res = __dlm_lookup_lockres(dlm, name, len);
141 spin_unlock(&dlm->spinlock);
142 return res;
143}
144
145static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
146{
147 struct dlm_ctxt *tmp = NULL;
148 struct list_head *iter;
149
150 assert_spin_locked(&dlm_domain_lock);
151
152 /* tmp->name here is always NULL terminated,
153 * but domain may not be! */
154 list_for_each(iter, &dlm_domains) {
155 tmp = list_entry (iter, struct dlm_ctxt, list);
156 if (strlen(tmp->name) == len &&
157 memcmp(tmp->name, domain, len)==0)
158 break;
159 tmp = NULL;
160 }
161
162 return tmp;
163}
164
165/* For null terminated domain strings ONLY */
166static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
167{
168 assert_spin_locked(&dlm_domain_lock);
169
170 return __dlm_lookup_domain_full(domain, strlen(domain));
171}
172
173
174/* returns true on one of two conditions:
175 * 1) the domain does not exist
176 * 2) the domain exists and it's state is "joined" */
177static int dlm_wait_on_domain_helper(const char *domain)
178{
179 int ret = 0;
180 struct dlm_ctxt *tmp = NULL;
181
182 spin_lock(&dlm_domain_lock);
183
184 tmp = __dlm_lookup_domain(domain);
185 if (!tmp)
186 ret = 1;
187 else if (tmp->dlm_state == DLM_CTXT_JOINED)
188 ret = 1;
189
190 spin_unlock(&dlm_domain_lock);
191 return ret;
192}
193
194static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
195{
196 if (dlm->resources)
197 free_page((unsigned long) dlm->resources);
198
199 if (dlm->name)
200 kfree(dlm->name);
201
202 kfree(dlm);
203}
204
205/* A little strange - this function will be called while holding
206 * dlm_domain_lock and is expected to be holding it on the way out. We
207 * will however drop and reacquire it multiple times */
208static void dlm_ctxt_release(struct kref *kref)
209{
210 struct dlm_ctxt *dlm;
211
212 dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
213
214 BUG_ON(dlm->num_joins);
215 BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
216
217 /* we may still be in the list if we hit an error during join. */
218 list_del_init(&dlm->list);
219
220 spin_unlock(&dlm_domain_lock);
221
222 mlog(0, "freeing memory from domain %s\n", dlm->name);
223
224 wake_up(&dlm_domain_events);
225
226 dlm_free_ctxt_mem(dlm);
227
228 spin_lock(&dlm_domain_lock);
229}
230
231void dlm_put(struct dlm_ctxt *dlm)
232{
233 spin_lock(&dlm_domain_lock);
234 kref_put(&dlm->dlm_refs, dlm_ctxt_release);
235 spin_unlock(&dlm_domain_lock);
236}
237
238static void __dlm_get(struct dlm_ctxt *dlm)
239{
240 kref_get(&dlm->dlm_refs);
241}
242
243/* given a questionable reference to a dlm object, gets a reference if
244 * it can find it in the list, otherwise returns NULL in which case
245 * you shouldn't trust your pointer. */
246struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
247{
248 struct list_head *iter;
249 struct dlm_ctxt *target = NULL;
250
251 spin_lock(&dlm_domain_lock);
252
253 list_for_each(iter, &dlm_domains) {
254 target = list_entry (iter, struct dlm_ctxt, list);
255
256 if (target == dlm) {
257 __dlm_get(target);
258 break;
259 }
260
261 target = NULL;
262 }
263
264 spin_unlock(&dlm_domain_lock);
265
266 return target;
267}
268
269int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
270{
271 int ret;
272
273 spin_lock(&dlm_domain_lock);
274 ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
275 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
276 spin_unlock(&dlm_domain_lock);
277
278 return ret;
279}
280
281static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
282{
283 dlm_unregister_domain_handlers(dlm);
284 dlm_complete_thread(dlm);
285 dlm_complete_recovery_thread(dlm);
286
287 /* We've left the domain. Now we can take ourselves out of the
288 * list and allow the kref stuff to help us free the
289 * memory. */
290 spin_lock(&dlm_domain_lock);
291 list_del_init(&dlm->list);
292 spin_unlock(&dlm_domain_lock);
293
294 /* Wake up anyone waiting for us to remove this domain */
295 wake_up(&dlm_domain_events);
296}
297
298static void dlm_migrate_all_locks(struct dlm_ctxt *dlm)
299{
300 int i;
301 struct dlm_lock_resource *res;
302
303 mlog(0, "Migrating locks from domain %s\n", dlm->name);
304restart:
305 spin_lock(&dlm->spinlock);
306 for (i=0; i<DLM_HASH_SIZE; i++) {
307 while (!list_empty(&dlm->resources[i])) {
308 res = list_entry(dlm->resources[i].next,
309 struct dlm_lock_resource, list);
310 /* need reference when manually grabbing lockres */
311 dlm_lockres_get(res);
312 /* this should unhash the lockres
313 * and exit with dlm->spinlock */
314 mlog(0, "purging res=%p\n", res);
315 if (dlm_lockres_is_dirty(dlm, res)) {
316 /* HACK! this should absolutely go.
317 * need to figure out why some empty
318 * lockreses are still marked dirty */
319 mlog(ML_ERROR, "lockres %.*s dirty!\n",
320 res->lockname.len, res->lockname.name);
321
322 spin_unlock(&dlm->spinlock);
323 dlm_kick_thread(dlm, res);
324 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
325 dlm_lockres_put(res);
326 goto restart;
327 }
328 dlm_purge_lockres(dlm, res);
329 dlm_lockres_put(res);
330 }
331 }
332 spin_unlock(&dlm->spinlock);
333
334 mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
335}
336
337static int dlm_no_joining_node(struct dlm_ctxt *dlm)
338{
339 int ret;
340
341 spin_lock(&dlm->spinlock);
342 ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
343 spin_unlock(&dlm->spinlock);
344
345 return ret;
346}
347
348static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
349{
350 /* Yikes, a double spinlock! I need domain_lock for the dlm
351 * state and the dlm spinlock for join state... Sorry! */
352again:
353 spin_lock(&dlm_domain_lock);
354 spin_lock(&dlm->spinlock);
355
356 if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
357 mlog(0, "Node %d is joining, we wait on it.\n",
358 dlm->joining_node);
359 spin_unlock(&dlm->spinlock);
360 spin_unlock(&dlm_domain_lock);
361
362 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
363 goto again;
364 }
365
366 dlm->dlm_state = DLM_CTXT_LEAVING;
367 spin_unlock(&dlm->spinlock);
368 spin_unlock(&dlm_domain_lock);
369}
370
371static void __dlm_print_nodes(struct dlm_ctxt *dlm)
372{
373 int node = -1;
374
375 assert_spin_locked(&dlm->spinlock);
376
377 mlog(ML_NOTICE, "Nodes in my domain (\"%s\"):\n", dlm->name);
378
379 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
380 node + 1)) < O2NM_MAX_NODES) {
381 mlog(ML_NOTICE, " node %d\n", node);
382 }
383}
384
385static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data)
386{
387 struct dlm_ctxt *dlm = data;
388 unsigned int node;
389 struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
390
391 mlog_entry("%p %u %p", msg, len, data);
392
393 if (!dlm_grab(dlm))
394 return 0;
395
396 node = exit_msg->node_idx;
397
398 mlog(0, "Node %u leaves domain %s\n", node, dlm->name);
399
400 spin_lock(&dlm->spinlock);
401 clear_bit(node, dlm->domain_map);
402 __dlm_print_nodes(dlm);
403
404 /* notify anything attached to the heartbeat events */
405 dlm_hb_event_notify_attached(dlm, node, 0);
406
407 spin_unlock(&dlm->spinlock);
408
409 dlm_put(dlm);
410
411 return 0;
412}
413
414static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm,
415 unsigned int node)
416{
417 int status;
418 struct dlm_exit_domain leave_msg;
419
420 mlog(0, "Asking node %u if we can leave the domain %s me = %u\n",
421 node, dlm->name, dlm->node_num);
422
423 memset(&leave_msg, 0, sizeof(leave_msg));
424 leave_msg.node_idx = dlm->node_num;
425
426 status = o2net_send_message(DLM_EXIT_DOMAIN_MSG, dlm->key,
427 &leave_msg, sizeof(leave_msg), node,
428 NULL);
429
430 mlog(0, "status return %d from o2net_send_message\n", status);
431
432 return status;
433}
434
435
436static void dlm_leave_domain(struct dlm_ctxt *dlm)
437{
438 int node, clear_node, status;
439
440 /* At this point we've migrated away all our locks and won't
441 * accept mastership of new ones. The dlm is responsible for
442 * almost nothing now. We make sure not to confuse any joining
443 * nodes and then commence shutdown procedure. */
444
445 spin_lock(&dlm->spinlock);
446 /* Clear ourselves from the domain map */
447 clear_bit(dlm->node_num, dlm->domain_map);
448 while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
449 0)) < O2NM_MAX_NODES) {
450 /* Drop the dlm spinlock. This is safe wrt the domain_map.
451 * -nodes cannot be added now as the
452 * query_join_handlers knows to respond with OK_NO_MAP
453 * -we catch the right network errors if a node is
454 * removed from the map while we're sending him the
455 * exit message. */
456 spin_unlock(&dlm->spinlock);
457
458 clear_node = 1;
459
460 status = dlm_send_one_domain_exit(dlm, node);
461 if (status < 0 &&
462 status != -ENOPROTOOPT &&
463 status != -ENOTCONN) {
464 mlog(ML_NOTICE, "Error %d sending domain exit message "
465 "to node %d\n", status, node);
466
467 /* Not sure what to do here but lets sleep for
468 * a bit in case this was a transient
469 * error... */
470 msleep(DLM_DOMAIN_BACKOFF_MS);
471 clear_node = 0;
472 }
473
474 spin_lock(&dlm->spinlock);
475 /* If we're not clearing the node bit then we intend
476 * to loop back around to try again. */
477 if (clear_node)
478 clear_bit(node, dlm->domain_map);
479 }
480 spin_unlock(&dlm->spinlock);
481}
482
483int dlm_joined(struct dlm_ctxt *dlm)
484{
485 int ret = 0;
486
487 spin_lock(&dlm_domain_lock);
488
489 if (dlm->dlm_state == DLM_CTXT_JOINED)
490 ret = 1;
491
492 spin_unlock(&dlm_domain_lock);
493
494 return ret;
495}
496
497int dlm_shutting_down(struct dlm_ctxt *dlm)
498{
499 int ret = 0;
500
501 spin_lock(&dlm_domain_lock);
502
503 if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
504 ret = 1;
505
506 spin_unlock(&dlm_domain_lock);
507
508 return ret;
509}
510
511void dlm_unregister_domain(struct dlm_ctxt *dlm)
512{
513 int leave = 0;
514
515 spin_lock(&dlm_domain_lock);
516 BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
517 BUG_ON(!dlm->num_joins);
518
519 dlm->num_joins--;
520 if (!dlm->num_joins) {
521 /* We mark it "in shutdown" now so new register
522 * requests wait until we've completely left the
523 * domain. Don't use DLM_CTXT_LEAVING yet as we still
524 * want new domain joins to communicate with us at
525 * least until we've completed migration of our
526 * resources. */
527 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
528 leave = 1;
529 }
530 spin_unlock(&dlm_domain_lock);
531
532 if (leave) {
533 mlog(0, "shutting down domain %s\n", dlm->name);
534
535 /* We changed dlm state, notify the thread */
536 dlm_kick_thread(dlm, NULL);
537
538 dlm_migrate_all_locks(dlm);
539 dlm_mark_domain_leaving(dlm);
540 dlm_leave_domain(dlm);
541 dlm_complete_dlm_shutdown(dlm);
542 }
543 dlm_put(dlm);
544}
545EXPORT_SYMBOL_GPL(dlm_unregister_domain);
546
547static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data)
548{
549 struct dlm_query_join_request *query;
550 enum dlm_query_join_response response;
551 struct dlm_ctxt *dlm = NULL;
552
553 query = (struct dlm_query_join_request *) msg->buf;
554
555 mlog(0, "node %u wants to join domain %s\n", query->node_idx,
556 query->domain);
557
558 /*
559 * If heartbeat doesn't consider the node live, tell it
560 * to back off and try again. This gives heartbeat a chance
561 * to catch up.
562 */
563 if (!o2hb_check_node_heartbeating(query->node_idx)) {
564 mlog(0, "node %u is not in our live map yet\n",
565 query->node_idx);
566
567 response = JOIN_DISALLOW;
568 goto respond;
569 }
570
571 response = JOIN_OK_NO_MAP;
572
573 spin_lock(&dlm_domain_lock);
574 dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
575 /* Once the dlm ctxt is marked as leaving then we don't want
576 * to be put in someone's domain map. */
577 if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
578 spin_lock(&dlm->spinlock);
579
580 if (dlm->dlm_state == DLM_CTXT_NEW &&
581 dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
582 /*If this is a brand new context and we
583 * haven't started our join process yet, then
584 * the other node won the race. */
585 response = JOIN_OK_NO_MAP;
586 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
587 /* Disallow parallel joins. */
588 response = JOIN_DISALLOW;
589 } else {
590 /* Alright we're fully a part of this domain
591 * so we keep some state as to who's joining
592 * and indicate to him that needs to be fixed
593 * up. */
594 response = JOIN_OK;
595 __dlm_set_joining_node(dlm, query->node_idx);
596 }
597
598 spin_unlock(&dlm->spinlock);
599 }
600 spin_unlock(&dlm_domain_lock);
601
602respond:
603 mlog(0, "We respond with %u\n", response);
604
605 return response;
606}
607
608static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data)
609{
610 struct dlm_assert_joined *assert;
611 struct dlm_ctxt *dlm = NULL;
612
613 assert = (struct dlm_assert_joined *) msg->buf;
614
615 mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
616 assert->domain);
617
618 spin_lock(&dlm_domain_lock);
619 dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
620 /* XXX should we consider no dlm ctxt an error? */
621 if (dlm) {
622 spin_lock(&dlm->spinlock);
623
624 /* Alright, this node has officially joined our
625 * domain. Set him in the map and clean up our
626 * leftover join state. */
627 BUG_ON(dlm->joining_node != assert->node_idx);
628 set_bit(assert->node_idx, dlm->domain_map);
629 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
630
631 __dlm_print_nodes(dlm);
632
633 /* notify anything attached to the heartbeat events */
634 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
635
636 spin_unlock(&dlm->spinlock);
637 }
638 spin_unlock(&dlm_domain_lock);
639
640 return 0;
641}
642
643static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data)
644{
645 struct dlm_cancel_join *cancel;
646 struct dlm_ctxt *dlm = NULL;
647
648 cancel = (struct dlm_cancel_join *) msg->buf;
649
650 mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
651 cancel->domain);
652
653 spin_lock(&dlm_domain_lock);
654 dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
655
656 if (dlm) {
657 spin_lock(&dlm->spinlock);
658
659 /* Yikes, this guy wants to cancel his join. No
660 * problem, we simply cleanup our join state. */
661 BUG_ON(dlm->joining_node != cancel->node_idx);
662 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
663
664 spin_unlock(&dlm->spinlock);
665 }
666 spin_unlock(&dlm_domain_lock);
667
668 return 0;
669}
670
671static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
672 unsigned int node)
673{
674 int status;
675 struct dlm_cancel_join cancel_msg;
676
677 memset(&cancel_msg, 0, sizeof(cancel_msg));
678 cancel_msg.node_idx = dlm->node_num;
679 cancel_msg.name_len = strlen(dlm->name);
680 memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
681
682 status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
683 &cancel_msg, sizeof(cancel_msg), node,
684 NULL);
685 if (status < 0) {
686 mlog_errno(status);
687 goto bail;
688 }
689
690bail:
691 return status;
692}
693
694/* map_size should be in bytes. */
695static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
696 unsigned long *node_map,
697 unsigned int map_size)
698{
699 int status, tmpstat;
700 unsigned int node;
701
702 if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
703 sizeof(unsigned long))) {
704 mlog(ML_ERROR,
705 "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
706 map_size, BITS_TO_LONGS(O2NM_MAX_NODES));
707 return -EINVAL;
708 }
709
710 status = 0;
711 node = -1;
712 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
713 node + 1)) < O2NM_MAX_NODES) {
714 if (node == dlm->node_num)
715 continue;
716
717 tmpstat = dlm_send_one_join_cancel(dlm, node);
718 if (tmpstat) {
719 mlog(ML_ERROR, "Error return %d cancelling join on "
720 "node %d\n", tmpstat, node);
721 if (!status)
722 status = tmpstat;
723 }
724 }
725
726 if (status)
727 mlog_errno(status);
728 return status;
729}
730
731static int dlm_request_join(struct dlm_ctxt *dlm,
732 int node,
733 enum dlm_query_join_response *response)
734{
735 int status, retval;
736 struct dlm_query_join_request join_msg;
737
738 mlog(0, "querying node %d\n", node);
739
740 memset(&join_msg, 0, sizeof(join_msg));
741 join_msg.node_idx = dlm->node_num;
742 join_msg.name_len = strlen(dlm->name);
743 memcpy(join_msg.domain, dlm->name, join_msg.name_len);
744
745 status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
746 sizeof(join_msg), node, &retval);
747 if (status < 0 && status != -ENOPROTOOPT) {
748 mlog_errno(status);
749 goto bail;
750 }
751
752 /* -ENOPROTOOPT from the net code means the other side isn't
753 listening for our message type -- that's fine, it means
754 his dlm isn't up, so we can consider him a 'yes' but not
755 joined into the domain. */
756 if (status == -ENOPROTOOPT) {
757 status = 0;
758 *response = JOIN_OK_NO_MAP;
759 } else if (retval == JOIN_DISALLOW ||
760 retval == JOIN_OK ||
761 retval == JOIN_OK_NO_MAP) {
762 *response = retval;
763 } else {
764 status = -EINVAL;
765 mlog(ML_ERROR, "invalid response %d from node %u\n", retval,
766 node);
767 }
768
769 mlog(0, "status %d, node %d response is %d\n", status, node,
770 *response);
771
772bail:
773 return status;
774}
775
776static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
777 unsigned int node)
778{
779 int status;
780 struct dlm_assert_joined assert_msg;
781
782 mlog(0, "Sending join assert to node %u\n", node);
783
784 memset(&assert_msg, 0, sizeof(assert_msg));
785 assert_msg.node_idx = dlm->node_num;
786 assert_msg.name_len = strlen(dlm->name);
787 memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
788
789 status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
790 &assert_msg, sizeof(assert_msg), node,
791 NULL);
792 if (status < 0)
793 mlog_errno(status);
794
795 return status;
796}
797
798static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
799 unsigned long *node_map)
800{
801 int status, node, live;
802
803 status = 0;
804 node = -1;
805 while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
806 node + 1)) < O2NM_MAX_NODES) {
807 if (node == dlm->node_num)
808 continue;
809
810 do {
811 /* It is very important that this message be
812 * received so we spin until either the node
813 * has died or it gets the message. */
814 status = dlm_send_one_join_assert(dlm, node);
815
816 spin_lock(&dlm->spinlock);
817 live = test_bit(node, dlm->live_nodes_map);
818 spin_unlock(&dlm->spinlock);
819
820 if (status) {
821 mlog(ML_ERROR, "Error return %d asserting "
822 "join on node %d\n", status, node);
823
824 /* give us some time between errors... */
825 if (live)
826 msleep(DLM_DOMAIN_BACKOFF_MS);
827 }
828 } while (status && live);
829 }
830}
831
832struct domain_join_ctxt {
833 unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
834 unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
835};
836
837static int dlm_should_restart_join(struct dlm_ctxt *dlm,
838 struct domain_join_ctxt *ctxt,
839 enum dlm_query_join_response response)
840{
841 int ret;
842
843 if (response == JOIN_DISALLOW) {
844 mlog(0, "Latest response of disallow -- should restart\n");
845 return 1;
846 }
847
848 spin_lock(&dlm->spinlock);
849 /* For now, we restart the process if the node maps have
850 * changed at all */
851 ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
852 sizeof(dlm->live_nodes_map));
853 spin_unlock(&dlm->spinlock);
854
855 if (ret)
856 mlog(0, "Node maps changed -- should restart\n");
857
858 return ret;
859}
860
861static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
862{
863 int status = 0, tmpstat, node;
864 struct domain_join_ctxt *ctxt;
865 enum dlm_query_join_response response;
866
867 mlog_entry("%p", dlm);
868
869 ctxt = kcalloc(1, sizeof(*ctxt), GFP_KERNEL);
870 if (!ctxt) {
871 status = -ENOMEM;
872 mlog_errno(status);
873 goto bail;
874 }
875
876 /* group sem locking should work for us here -- we're already
877 * registered for heartbeat events so filling this should be
878 * atomic wrt getting those handlers called. */
879 o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
880
881 spin_lock(&dlm->spinlock);
882 memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
883
884 __dlm_set_joining_node(dlm, dlm->node_num);
885
886 spin_unlock(&dlm->spinlock);
887
888 node = -1;
889 while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
890 node + 1)) < O2NM_MAX_NODES) {
891 if (node == dlm->node_num)
892 continue;
893
894 status = dlm_request_join(dlm, node, &response);
895 if (status < 0) {
896 mlog_errno(status);
897 goto bail;
898 }
899
900 /* Ok, either we got a response or the node doesn't have a
901 * dlm up. */
902 if (response == JOIN_OK)
903 set_bit(node, ctxt->yes_resp_map);
904
905 if (dlm_should_restart_join(dlm, ctxt, response)) {
906 status = -EAGAIN;
907 goto bail;
908 }
909 }
910
911 mlog(0, "Yay, done querying nodes!\n");
912
913 /* Yay, everyone agree's we can join the domain. My domain is
914 * comprised of all nodes who were put in the
915 * yes_resp_map. Copy that into our domain map and send a join
916 * assert message to clean up everyone elses state. */
917 spin_lock(&dlm->spinlock);
918 memcpy(dlm->domain_map, ctxt->yes_resp_map,
919 sizeof(ctxt->yes_resp_map));
920 set_bit(dlm->node_num, dlm->domain_map);
921 spin_unlock(&dlm->spinlock);
922
923 dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
924
925 /* Joined state *must* be set before the joining node
926 * information, otherwise the query_join handler may read no
927 * current joiner but a state of NEW and tell joining nodes
928 * we're not in the domain. */
929 spin_lock(&dlm_domain_lock);
930 dlm->dlm_state = DLM_CTXT_JOINED;
931 dlm->num_joins++;
932 spin_unlock(&dlm_domain_lock);
933
934bail:
935 spin_lock(&dlm->spinlock);
936 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
937 if (!status)
938 __dlm_print_nodes(dlm);
939 spin_unlock(&dlm->spinlock);
940
941 if (ctxt) {
942 /* Do we need to send a cancel message to any nodes? */
943 if (status < 0) {
944 tmpstat = dlm_send_join_cancels(dlm,
945 ctxt->yes_resp_map,
946 sizeof(ctxt->yes_resp_map));
947 if (tmpstat < 0)
948 mlog_errno(tmpstat);
949 }
950 kfree(ctxt);
951 }
952
953 mlog(0, "returning %d\n", status);
954 return status;
955}
956
957static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
958{
959 o2hb_unregister_callback(&dlm->dlm_hb_up);
960 o2hb_unregister_callback(&dlm->dlm_hb_down);
961 o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
962}
963
964static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
965{
966 int status;
967
968 mlog(0, "registering handlers.\n");
969
970 o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
971 dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
972 status = o2hb_register_callback(&dlm->dlm_hb_down);
973 if (status)
974 goto bail;
975
976 o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
977 dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
978 status = o2hb_register_callback(&dlm->dlm_hb_up);
979 if (status)
980 goto bail;
981
982 status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
983 sizeof(struct dlm_master_request),
984 dlm_master_request_handler,
985 dlm, &dlm->dlm_domain_handlers);
986 if (status)
987 goto bail;
988
989 status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
990 sizeof(struct dlm_assert_master),
991 dlm_assert_master_handler,
992 dlm, &dlm->dlm_domain_handlers);
993 if (status)
994 goto bail;
995
996 status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
997 sizeof(struct dlm_create_lock),
998 dlm_create_lock_handler,
999 dlm, &dlm->dlm_domain_handlers);
1000 if (status)
1001 goto bail;
1002
1003 status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1004 DLM_CONVERT_LOCK_MAX_LEN,
1005 dlm_convert_lock_handler,
1006 dlm, &dlm->dlm_domain_handlers);
1007 if (status)
1008 goto bail;
1009
1010 status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1011 DLM_UNLOCK_LOCK_MAX_LEN,
1012 dlm_unlock_lock_handler,
1013 dlm, &dlm->dlm_domain_handlers);
1014 if (status)
1015 goto bail;
1016
1017 status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1018 DLM_PROXY_AST_MAX_LEN,
1019 dlm_proxy_ast_handler,
1020 dlm, &dlm->dlm_domain_handlers);
1021 if (status)
1022 goto bail;
1023
1024 status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1025 sizeof(struct dlm_exit_domain),
1026 dlm_exit_domain_handler,
1027 dlm, &dlm->dlm_domain_handlers);
1028 if (status)
1029 goto bail;
1030
1031 status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1032 sizeof(struct dlm_migrate_request),
1033 dlm_migrate_request_handler,
1034 dlm, &dlm->dlm_domain_handlers);
1035 if (status)
1036 goto bail;
1037
1038 status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1039 DLM_MIG_LOCKRES_MAX_LEN,
1040 dlm_mig_lockres_handler,
1041 dlm, &dlm->dlm_domain_handlers);
1042 if (status)
1043 goto bail;
1044
1045 status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1046 sizeof(struct dlm_master_requery),
1047 dlm_master_requery_handler,
1048 dlm, &dlm->dlm_domain_handlers);
1049 if (status)
1050 goto bail;
1051
1052 status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1053 sizeof(struct dlm_lock_request),
1054 dlm_request_all_locks_handler,
1055 dlm, &dlm->dlm_domain_handlers);
1056 if (status)
1057 goto bail;
1058
1059 status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1060 sizeof(struct dlm_reco_data_done),
1061 dlm_reco_data_done_handler,
1062 dlm, &dlm->dlm_domain_handlers);
1063 if (status)
1064 goto bail;
1065
1066 status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1067 sizeof(struct dlm_begin_reco),
1068 dlm_begin_reco_handler,
1069 dlm, &dlm->dlm_domain_handlers);
1070 if (status)
1071 goto bail;
1072
1073 status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1074 sizeof(struct dlm_finalize_reco),
1075 dlm_finalize_reco_handler,
1076 dlm, &dlm->dlm_domain_handlers);
1077 if (status)
1078 goto bail;
1079
1080bail:
1081 if (status)
1082 dlm_unregister_domain_handlers(dlm);
1083
1084 return status;
1085}
1086
1087static int dlm_join_domain(struct dlm_ctxt *dlm)
1088{
1089 int status;
1090
1091 BUG_ON(!dlm);
1092
1093 mlog(0, "Join domain %s\n", dlm->name);
1094
1095 status = dlm_register_domain_handlers(dlm);
1096 if (status) {
1097 mlog_errno(status);
1098 goto bail;
1099 }
1100
1101 status = dlm_launch_thread(dlm);
1102 if (status < 0) {
1103 mlog_errno(status);
1104 goto bail;
1105 }
1106
1107 status = dlm_launch_recovery_thread(dlm);
1108 if (status < 0) {
1109 mlog_errno(status);
1110 goto bail;
1111 }
1112
1113 do {
1114 unsigned int backoff;
1115 status = dlm_try_to_join_domain(dlm);
1116
1117 /* If we're racing another node to the join, then we
1118 * need to back off temporarily and let them
1119 * complete. */
1120 if (status == -EAGAIN) {
1121 if (signal_pending(current)) {
1122 status = -ERESTARTSYS;
1123 goto bail;
1124 }
1125
1126 /*
1127 * <chip> After you!
1128 * <dale> No, after you!
1129 * <chip> I insist!
1130 * <dale> But you first!
1131 * ...
1132 */
1133 backoff = (unsigned int)(jiffies & 0x3);
1134 backoff *= DLM_DOMAIN_BACKOFF_MS;
1135 mlog(0, "backoff %d\n", backoff);
1136 msleep(backoff);
1137 }
1138 } while (status == -EAGAIN);
1139
1140 if (status < 0) {
1141 mlog_errno(status);
1142 goto bail;
1143 }
1144
1145 status = 0;
1146bail:
1147 wake_up(&dlm_domain_events);
1148
1149 if (status) {
1150 dlm_unregister_domain_handlers(dlm);
1151 dlm_complete_thread(dlm);
1152 dlm_complete_recovery_thread(dlm);
1153 }
1154
1155 return status;
1156}
1157
1158static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1159 u32 key)
1160{
1161 int i;
1162 struct dlm_ctxt *dlm = NULL;
1163
1164 dlm = kcalloc(1, sizeof(*dlm), GFP_KERNEL);
1165 if (!dlm) {
1166 mlog_errno(-ENOMEM);
1167 goto leave;
1168 }
1169
1170 dlm->name = kmalloc(strlen(domain) + 1, GFP_KERNEL);
1171 if (dlm->name == NULL) {
1172 mlog_errno(-ENOMEM);
1173 kfree(dlm);
1174 dlm = NULL;
1175 goto leave;
1176 }
1177
1178 dlm->resources = (struct list_head *) __get_free_page(GFP_KERNEL);
1179 if (!dlm->resources) {
1180 mlog_errno(-ENOMEM);
1181 kfree(dlm->name);
1182 kfree(dlm);
1183 dlm = NULL;
1184 goto leave;
1185 }
1186 memset(dlm->resources, 0, PAGE_SIZE);
1187
1188 for (i=0; i<DLM_HASH_SIZE; i++)
1189 INIT_LIST_HEAD(&dlm->resources[i]);
1190
1191 strcpy(dlm->name, domain);
1192 dlm->key = key;
1193 dlm->node_num = o2nm_this_node();
1194
1195 spin_lock_init(&dlm->spinlock);
1196 spin_lock_init(&dlm->master_lock);
1197 spin_lock_init(&dlm->ast_lock);
1198 INIT_LIST_HEAD(&dlm->list);
1199 INIT_LIST_HEAD(&dlm->dirty_list);
1200 INIT_LIST_HEAD(&dlm->reco.resources);
1201 INIT_LIST_HEAD(&dlm->reco.received);
1202 INIT_LIST_HEAD(&dlm->reco.node_data);
1203 INIT_LIST_HEAD(&dlm->purge_list);
1204 INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
1205 dlm->reco.state = 0;
1206
1207 INIT_LIST_HEAD(&dlm->pending_asts);
1208 INIT_LIST_HEAD(&dlm->pending_basts);
1209
1210 mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
1211 dlm->recovery_map, &(dlm->recovery_map[0]));
1212
1213 memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
1214 memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
1215 memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
1216
1217 dlm->dlm_thread_task = NULL;
1218 dlm->dlm_reco_thread_task = NULL;
1219 init_waitqueue_head(&dlm->dlm_thread_wq);
1220 init_waitqueue_head(&dlm->dlm_reco_thread_wq);
1221 init_waitqueue_head(&dlm->reco.event);
1222 init_waitqueue_head(&dlm->ast_wq);
1223 init_waitqueue_head(&dlm->migration_wq);
1224 INIT_LIST_HEAD(&dlm->master_list);
1225 INIT_LIST_HEAD(&dlm->mle_hb_events);
1226
1227 dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
1228 init_waitqueue_head(&dlm->dlm_join_events);
1229
1230 dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
1231 dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
1232 atomic_set(&dlm->local_resources, 0);
1233 atomic_set(&dlm->remote_resources, 0);
1234 atomic_set(&dlm->unknown_resources, 0);
1235
1236 spin_lock_init(&dlm->work_lock);
1237 INIT_LIST_HEAD(&dlm->work_list);
1238 INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work, dlm);
1239
1240 kref_init(&dlm->dlm_refs);
1241 dlm->dlm_state = DLM_CTXT_NEW;
1242
1243 INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
1244
1245 mlog(0, "context init: refcount %u\n",
1246 atomic_read(&dlm->dlm_refs.refcount));
1247
1248leave:
1249 return dlm;
1250}
1251
1252/*
1253 * dlm_register_domain: one-time setup per "domain"
1254 */
1255struct dlm_ctxt * dlm_register_domain(const char *domain,
1256 u32 key)
1257{
1258 int ret;
1259 struct dlm_ctxt *dlm = NULL;
1260 struct dlm_ctxt *new_ctxt = NULL;
1261
1262 if (strlen(domain) > O2NM_MAX_NAME_LEN) {
1263 ret = -ENAMETOOLONG;
1264 mlog(ML_ERROR, "domain name length too long\n");
1265 goto leave;
1266 }
1267
1268 if (!o2hb_check_local_node_heartbeating()) {
1269 mlog(ML_ERROR, "the local node has not been configured, or is "
1270 "not heartbeating\n");
1271 ret = -EPROTO;
1272 goto leave;
1273 }
1274
1275 mlog(0, "register called for domain \"%s\"\n", domain);
1276
1277retry:
1278 dlm = NULL;
1279 if (signal_pending(current)) {
1280 ret = -ERESTARTSYS;
1281 mlog_errno(ret);
1282 goto leave;
1283 }
1284
1285 spin_lock(&dlm_domain_lock);
1286
1287 dlm = __dlm_lookup_domain(domain);
1288 if (dlm) {
1289 if (dlm->dlm_state != DLM_CTXT_JOINED) {
1290 spin_unlock(&dlm_domain_lock);
1291
1292 mlog(0, "This ctxt is not joined yet!\n");
1293 wait_event_interruptible(dlm_domain_events,
1294 dlm_wait_on_domain_helper(
1295 domain));
1296 goto retry;
1297 }
1298
1299 __dlm_get(dlm);
1300 dlm->num_joins++;
1301
1302 spin_unlock(&dlm_domain_lock);
1303
1304 ret = 0;
1305 goto leave;
1306 }
1307
1308 /* doesn't exist */
1309 if (!new_ctxt) {
1310 spin_unlock(&dlm_domain_lock);
1311
1312 new_ctxt = dlm_alloc_ctxt(domain, key);
1313 if (new_ctxt)
1314 goto retry;
1315
1316 ret = -ENOMEM;
1317 mlog_errno(ret);
1318 goto leave;
1319 }
1320
1321 /* a little variable switch-a-roo here... */
1322 dlm = new_ctxt;
1323 new_ctxt = NULL;
1324
1325 /* add the new domain */
1326 list_add_tail(&dlm->list, &dlm_domains);
1327 spin_unlock(&dlm_domain_lock);
1328
1329 ret = dlm_join_domain(dlm);
1330 if (ret) {
1331 mlog_errno(ret);
1332 dlm_put(dlm);
1333 goto leave;
1334 }
1335
1336 ret = 0;
1337leave:
1338 if (new_ctxt)
1339 dlm_free_ctxt_mem(new_ctxt);
1340
1341 if (ret < 0)
1342 dlm = ERR_PTR(ret);
1343
1344 return dlm;
1345}
1346EXPORT_SYMBOL_GPL(dlm_register_domain);
1347
1348static LIST_HEAD(dlm_join_handlers);
1349
1350static void dlm_unregister_net_handlers(void)
1351{
1352 o2net_unregister_handler_list(&dlm_join_handlers);
1353}
1354
1355static int dlm_register_net_handlers(void)
1356{
1357 int status = 0;
1358
1359 status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1360 sizeof(struct dlm_query_join_request),
1361 dlm_query_join_handler,
1362 NULL, &dlm_join_handlers);
1363 if (status)
1364 goto bail;
1365
1366 status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1367 sizeof(struct dlm_assert_joined),
1368 dlm_assert_joined_handler,
1369 NULL, &dlm_join_handlers);
1370 if (status)
1371 goto bail;
1372
1373 status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1374 sizeof(struct dlm_cancel_join),
1375 dlm_cancel_join_handler,
1376 NULL, &dlm_join_handlers);
1377
1378bail:
1379 if (status < 0)
1380 dlm_unregister_net_handlers();
1381
1382 return status;
1383}
1384
1385/* Domain eviction callback handling.
1386 *
1387 * The file system requires notification of node death *before* the
1388 * dlm completes it's recovery work, otherwise it may be able to
1389 * acquire locks on resources requiring recovery. Since the dlm can
1390 * evict a node from it's domain *before* heartbeat fires, a similar
1391 * mechanism is required. */
1392
1393/* Eviction is not expected to happen often, so a per-domain lock is
1394 * not necessary. Eviction callbacks are allowed to sleep for short
1395 * periods of time. */
1396static DECLARE_RWSEM(dlm_callback_sem);
1397
1398void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
1399 int node_num)
1400{
1401 struct list_head *iter;
1402 struct dlm_eviction_cb *cb;
1403
1404 down_read(&dlm_callback_sem);
1405 list_for_each(iter, &dlm->dlm_eviction_callbacks) {
1406 cb = list_entry(iter, struct dlm_eviction_cb, ec_item);
1407
1408 cb->ec_func(node_num, cb->ec_data);
1409 }
1410 up_read(&dlm_callback_sem);
1411}
1412
1413void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
1414 dlm_eviction_func *f,
1415 void *data)
1416{
1417 INIT_LIST_HEAD(&cb->ec_item);
1418 cb->ec_func = f;
1419 cb->ec_data = data;
1420}
1421EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
1422
1423void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
1424 struct dlm_eviction_cb *cb)
1425{
1426 down_write(&dlm_callback_sem);
1427 list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
1428 up_write(&dlm_callback_sem);
1429}
1430EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
1431
1432void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
1433{
1434 down_write(&dlm_callback_sem);
1435 list_del_init(&cb->ec_item);
1436 up_write(&dlm_callback_sem);
1437}
1438EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
1439
1440static int __init dlm_init(void)
1441{
1442 int status;
1443
1444 dlm_print_version();
1445
1446 status = dlm_init_mle_cache();
1447 if (status)
1448 return -1;
1449
1450 status = dlm_register_net_handlers();
1451 if (status) {
1452 dlm_destroy_mle_cache();
1453 return -1;
1454 }
1455
1456 return 0;
1457}
1458
1459static void __exit dlm_exit (void)
1460{
1461 dlm_unregister_net_handlers();
1462 dlm_destroy_mle_cache();
1463}
1464
1465MODULE_AUTHOR("Oracle");
1466MODULE_LICENSE("GPL");
1467
1468module_init(dlm_init);
1469module_exit(dlm_exit);
diff --git a/fs/ocfs2/dlm/dlmdomain.h b/fs/ocfs2/dlm/dlmdomain.h
new file mode 100644
index 000000000000..2f7f60bfeb3b
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmdomain.h
@@ -0,0 +1,36 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmdomain.h
5 *
6 * Copyright (C) 2004 Oracle. All rights reserved.
7 *
8 * This program is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU General Public
10 * License as published by the Free Software Foundation; either
11 * version 2 of the License, or (at your option) any later version.
12 *
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public
19 * License along with this program; if not, write to the
20 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
21 * Boston, MA 021110-1307, USA.
22 *
23 */
24
25#ifndef DLMDOMAIN_H
26#define DLMDOMAIN_H
27
28extern spinlock_t dlm_domain_lock;
29extern struct list_head dlm_domains;
30
31int dlm_joined(struct dlm_ctxt *dlm);
32int dlm_shutting_down(struct dlm_ctxt *dlm);
33void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
34 int node_num);
35
36#endif
diff --git a/fs/ocfs2/dlm/dlmfs.c b/fs/ocfs2/dlm/dlmfs.c
new file mode 100644
index 000000000000..dd2d24dc25e0
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmfs.c
@@ -0,0 +1,640 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmfs.c
5 *
6 * Code which implements the kernel side of a minimal userspace
7 * interface to our DLM. This file handles the virtual file system
8 * used for communication with userspace. Credit should go to ramfs,
9 * which was a template for the fs side of this module.
10 *
11 * Copyright (C) 2003, 2004 Oracle. All rights reserved.
12 *
13 * This program is free software; you can redistribute it and/or
14 * modify it under the terms of the GNU General Public
15 * License as published by the Free Software Foundation; either
16 * version 2 of the License, or (at your option) any later version.
17 *
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
21 * General Public License for more details.
22 *
23 * You should have received a copy of the GNU General Public
24 * License along with this program; if not, write to the
25 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
26 * Boston, MA 021110-1307, USA.
27 */
28
29/* Simple VFS hooks based on: */
30/*
31 * Resizable simple ram filesystem for Linux.
32 *
33 * Copyright (C) 2000 Linus Torvalds.
34 * 2000 Transmeta Corp.
35 */
36
37#include <linux/module.h>
38#include <linux/fs.h>
39#include <linux/pagemap.h>
40#include <linux/types.h>
41#include <linux/slab.h>
42#include <linux/highmem.h>
43#include <linux/init.h>
44#include <linux/string.h>
45#include <linux/smp_lock.h>
46#include <linux/backing-dev.h>
47
48#include <asm/uaccess.h>
49
50
51#include "cluster/nodemanager.h"
52#include "cluster/heartbeat.h"
53#include "cluster/tcp.h"
54
55#include "dlmapi.h"
56
57#include "userdlm.h"
58
59#include "dlmfsver.h"
60
61#define MLOG_MASK_PREFIX ML_DLMFS
62#include "cluster/masklog.h"
63
64static struct super_operations dlmfs_ops;
65static struct file_operations dlmfs_file_operations;
66static struct inode_operations dlmfs_dir_inode_operations;
67static struct inode_operations dlmfs_root_inode_operations;
68static struct inode_operations dlmfs_file_inode_operations;
69static kmem_cache_t *dlmfs_inode_cache;
70
71struct workqueue_struct *user_dlm_worker;
72
73/*
74 * decodes a set of open flags into a valid lock level and a set of flags.
75 * returns < 0 if we have invalid flags
76 * flags which mean something to us:
77 * O_RDONLY -> PRMODE level
78 * O_WRONLY -> EXMODE level
79 *
80 * O_NONBLOCK -> LKM_NOQUEUE
81 */
82static int dlmfs_decode_open_flags(int open_flags,
83 int *level,
84 int *flags)
85{
86 if (open_flags & (O_WRONLY|O_RDWR))
87 *level = LKM_EXMODE;
88 else
89 *level = LKM_PRMODE;
90
91 *flags = 0;
92 if (open_flags & O_NONBLOCK)
93 *flags |= LKM_NOQUEUE;
94
95 return 0;
96}
97
98static int dlmfs_file_open(struct inode *inode,
99 struct file *file)
100{
101 int status, level, flags;
102 struct dlmfs_filp_private *fp = NULL;
103 struct dlmfs_inode_private *ip;
104
105 if (S_ISDIR(inode->i_mode))
106 BUG();
107
108 mlog(0, "open called on inode %lu, flags 0x%x\n", inode->i_ino,
109 file->f_flags);
110
111 status = dlmfs_decode_open_flags(file->f_flags, &level, &flags);
112 if (status < 0)
113 goto bail;
114
115 /* We don't want to honor O_APPEND at read/write time as it
116 * doesn't make sense for LVB writes. */
117 file->f_flags &= ~O_APPEND;
118
119 fp = kmalloc(sizeof(*fp), GFP_KERNEL);
120 if (!fp) {
121 status = -ENOMEM;
122 goto bail;
123 }
124 fp->fp_lock_level = level;
125
126 ip = DLMFS_I(inode);
127
128 status = user_dlm_cluster_lock(&ip->ip_lockres, level, flags);
129 if (status < 0) {
130 /* this is a strange error to return here but I want
131 * to be able userspace to be able to distinguish a
132 * valid lock request from one that simply couldn't be
133 * granted. */
134 if (flags & LKM_NOQUEUE && status == -EAGAIN)
135 status = -ETXTBSY;
136 kfree(fp);
137 goto bail;
138 }
139
140 file->private_data = fp;
141bail:
142 return status;
143}
144
145static int dlmfs_file_release(struct inode *inode,
146 struct file *file)
147{
148 int level, status;
149 struct dlmfs_inode_private *ip = DLMFS_I(inode);
150 struct dlmfs_filp_private *fp =
151 (struct dlmfs_filp_private *) file->private_data;
152
153 if (S_ISDIR(inode->i_mode))
154 BUG();
155
156 mlog(0, "close called on inode %lu\n", inode->i_ino);
157
158 status = 0;
159 if (fp) {
160 level = fp->fp_lock_level;
161 if (level != LKM_IVMODE)
162 user_dlm_cluster_unlock(&ip->ip_lockres, level);
163
164 kfree(fp);
165 file->private_data = NULL;
166 }
167
168 return 0;
169}
170
171static ssize_t dlmfs_file_read(struct file *filp,
172 char __user *buf,
173 size_t count,
174 loff_t *ppos)
175{
176 int bytes_left;
177 ssize_t readlen;
178 char *lvb_buf;
179 struct inode *inode = filp->f_dentry->d_inode;
180
181 mlog(0, "inode %lu, count = %zu, *ppos = %llu\n",
182 inode->i_ino, count, *ppos);
183
184 if (*ppos >= i_size_read(inode))
185 return 0;
186
187 if (!count)
188 return 0;
189
190 if (!access_ok(VERIFY_WRITE, buf, count))
191 return -EFAULT;
192
193 /* don't read past the lvb */
194 if ((count + *ppos) > i_size_read(inode))
195 readlen = i_size_read(inode) - *ppos;
196 else
197 readlen = count - *ppos;
198
199 lvb_buf = kmalloc(readlen, GFP_KERNEL);
200 if (!lvb_buf)
201 return -ENOMEM;
202
203 user_dlm_read_lvb(inode, lvb_buf, readlen);
204 bytes_left = __copy_to_user(buf, lvb_buf, readlen);
205 readlen -= bytes_left;
206
207 kfree(lvb_buf);
208
209 *ppos = *ppos + readlen;
210
211 mlog(0, "read %zd bytes\n", readlen);
212 return readlen;
213}
214
215static ssize_t dlmfs_file_write(struct file *filp,
216 const char __user *buf,
217 size_t count,
218 loff_t *ppos)
219{
220 int bytes_left;
221 ssize_t writelen;
222 char *lvb_buf;
223 struct inode *inode = filp->f_dentry->d_inode;
224
225 mlog(0, "inode %lu, count = %zu, *ppos = %llu\n",
226 inode->i_ino, count, *ppos);
227
228 if (*ppos >= i_size_read(inode))
229 return -ENOSPC;
230
231 if (!count)
232 return 0;
233
234 if (!access_ok(VERIFY_READ, buf, count))
235 return -EFAULT;
236
237 /* don't write past the lvb */
238 if ((count + *ppos) > i_size_read(inode))
239 writelen = i_size_read(inode) - *ppos;
240 else
241 writelen = count - *ppos;
242
243 lvb_buf = kmalloc(writelen, GFP_KERNEL);
244 if (!lvb_buf)
245 return -ENOMEM;
246
247 bytes_left = copy_from_user(lvb_buf, buf, writelen);
248 writelen -= bytes_left;
249 if (writelen)
250 user_dlm_write_lvb(inode, lvb_buf, writelen);
251
252 kfree(lvb_buf);
253
254 *ppos = *ppos + writelen;
255 mlog(0, "wrote %zd bytes\n", writelen);
256 return writelen;
257}
258
259static void dlmfs_init_once(void *foo,
260 kmem_cache_t *cachep,
261 unsigned long flags)
262{
263 struct dlmfs_inode_private *ip =
264 (struct dlmfs_inode_private *) foo;
265
266 if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) ==
267 SLAB_CTOR_CONSTRUCTOR) {
268 ip->ip_dlm = NULL;
269 ip->ip_parent = NULL;
270
271 inode_init_once(&ip->ip_vfs_inode);
272 }
273}
274
275static struct inode *dlmfs_alloc_inode(struct super_block *sb)
276{
277 struct dlmfs_inode_private *ip;
278
279 ip = kmem_cache_alloc(dlmfs_inode_cache, SLAB_NOFS);
280 if (!ip)
281 return NULL;
282
283 return &ip->ip_vfs_inode;
284}
285
286static void dlmfs_destroy_inode(struct inode *inode)
287{
288 kmem_cache_free(dlmfs_inode_cache, DLMFS_I(inode));
289}
290
291static void dlmfs_clear_inode(struct inode *inode)
292{
293 int status;
294 struct dlmfs_inode_private *ip;
295
296 if (!inode)
297 return;
298
299 mlog(0, "inode %lu\n", inode->i_ino);
300
301 ip = DLMFS_I(inode);
302
303 if (S_ISREG(inode->i_mode)) {
304 status = user_dlm_destroy_lock(&ip->ip_lockres);
305 if (status < 0)
306 mlog_errno(status);
307 iput(ip->ip_parent);
308 goto clear_fields;
309 }
310
311 mlog(0, "we're a directory, ip->ip_dlm = 0x%p\n", ip->ip_dlm);
312 /* we must be a directory. If required, lets unregister the
313 * dlm context now. */
314 if (ip->ip_dlm)
315 user_dlm_unregister_context(ip->ip_dlm);
316clear_fields:
317 ip->ip_parent = NULL;
318 ip->ip_dlm = NULL;
319}
320
321static struct backing_dev_info dlmfs_backing_dev_info = {
322 .ra_pages = 0, /* No readahead */
323 .capabilities = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
324};
325
326static struct inode *dlmfs_get_root_inode(struct super_block *sb)
327{
328 struct inode *inode = new_inode(sb);
329 int mode = S_IFDIR | 0755;
330 struct dlmfs_inode_private *ip;
331
332 if (inode) {
333 ip = DLMFS_I(inode);
334
335 inode->i_mode = mode;
336 inode->i_uid = current->fsuid;
337 inode->i_gid = current->fsgid;
338 inode->i_blksize = PAGE_CACHE_SIZE;
339 inode->i_blocks = 0;
340 inode->i_mapping->backing_dev_info = &dlmfs_backing_dev_info;
341 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
342 inode->i_nlink++;
343
344 inode->i_fop = &simple_dir_operations;
345 inode->i_op = &dlmfs_root_inode_operations;
346 }
347
348 return inode;
349}
350
351static struct inode *dlmfs_get_inode(struct inode *parent,
352 struct dentry *dentry,
353 int mode)
354{
355 struct super_block *sb = parent->i_sb;
356 struct inode * inode = new_inode(sb);
357 struct dlmfs_inode_private *ip;
358
359 if (!inode)
360 return NULL;
361
362 inode->i_mode = mode;
363 inode->i_uid = current->fsuid;
364 inode->i_gid = current->fsgid;
365 inode->i_blksize = PAGE_CACHE_SIZE;
366 inode->i_blocks = 0;
367 inode->i_mapping->backing_dev_info = &dlmfs_backing_dev_info;
368 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
369
370 ip = DLMFS_I(inode);
371 ip->ip_dlm = DLMFS_I(parent)->ip_dlm;
372
373 switch (mode & S_IFMT) {
374 default:
375 /* for now we don't support anything other than
376 * directories and regular files. */
377 BUG();
378 break;
379 case S_IFREG:
380 inode->i_op = &dlmfs_file_inode_operations;
381 inode->i_fop = &dlmfs_file_operations;
382
383 i_size_write(inode, DLM_LVB_LEN);
384
385 user_dlm_lock_res_init(&ip->ip_lockres, dentry);
386
387 /* released at clear_inode time, this insures that we
388 * get to drop the dlm reference on each lock *before*
389 * we call the unregister code for releasing parent
390 * directories. */
391 ip->ip_parent = igrab(parent);
392 BUG_ON(!ip->ip_parent);
393 break;
394 case S_IFDIR:
395 inode->i_op = &dlmfs_dir_inode_operations;
396 inode->i_fop = &simple_dir_operations;
397
398 /* directory inodes start off with i_nlink ==
399 * 2 (for "." entry) */
400 inode->i_nlink++;
401 break;
402 }
403
404 if (parent->i_mode & S_ISGID) {
405 inode->i_gid = parent->i_gid;
406 if (S_ISDIR(mode))
407 inode->i_mode |= S_ISGID;
408 }
409
410 return inode;
411}
412
413/*
414 * File creation. Allocate an inode, and we're done..
415 */
416/* SMP-safe */
417static int dlmfs_mkdir(struct inode * dir,
418 struct dentry * dentry,
419 int mode)
420{
421 int status;
422 struct inode *inode = NULL;
423 struct qstr *domain = &dentry->d_name;
424 struct dlmfs_inode_private *ip;
425 struct dlm_ctxt *dlm;
426
427 mlog(0, "mkdir %.*s\n", domain->len, domain->name);
428
429 /* verify that we have a proper domain */
430 if (domain->len >= O2NM_MAX_NAME_LEN) {
431 status = -EINVAL;
432 mlog(ML_ERROR, "invalid domain name for directory.\n");
433 goto bail;
434 }
435
436 inode = dlmfs_get_inode(dir, dentry, mode | S_IFDIR);
437 if (!inode) {
438 status = -ENOMEM;
439 mlog_errno(status);
440 goto bail;
441 }
442
443 ip = DLMFS_I(inode);
444
445 dlm = user_dlm_register_context(domain);
446 if (IS_ERR(dlm)) {
447 status = PTR_ERR(dlm);
448 mlog(ML_ERROR, "Error %d could not register domain \"%.*s\"\n",
449 status, domain->len, domain->name);
450 goto bail;
451 }
452 ip->ip_dlm = dlm;
453
454 dir->i_nlink++;
455 d_instantiate(dentry, inode);
456 dget(dentry); /* Extra count - pin the dentry in core */
457
458 status = 0;
459bail:
460 if (status < 0)
461 iput(inode);
462 return status;
463}
464
465static int dlmfs_create(struct inode *dir,
466 struct dentry *dentry,
467 int mode,
468 struct nameidata *nd)
469{
470 int status = 0;
471 struct inode *inode;
472 struct qstr *name = &dentry->d_name;
473
474 mlog(0, "create %.*s\n", name->len, name->name);
475
476 /* verify name is valid and doesn't contain any dlm reserved
477 * characters */
478 if (name->len >= USER_DLM_LOCK_ID_MAX_LEN ||
479 name->name[0] == '$') {
480 status = -EINVAL;
481 mlog(ML_ERROR, "invalid lock name, %.*s\n", name->len,
482 name->name);
483 goto bail;
484 }
485
486 inode = dlmfs_get_inode(dir, dentry, mode | S_IFREG);
487 if (!inode) {
488 status = -ENOMEM;
489 mlog_errno(status);
490 goto bail;
491 }
492
493 d_instantiate(dentry, inode);
494 dget(dentry); /* Extra count - pin the dentry in core */
495bail:
496 return status;
497}
498
499static int dlmfs_unlink(struct inode *dir,
500 struct dentry *dentry)
501{
502 int status;
503 struct inode *inode = dentry->d_inode;
504
505 mlog(0, "unlink inode %lu\n", inode->i_ino);
506
507 /* if there are no current holders, or none that are waiting
508 * to acquire a lock, this basically destroys our lockres. */
509 status = user_dlm_destroy_lock(&DLMFS_I(inode)->ip_lockres);
510 if (status < 0) {
511 mlog(ML_ERROR, "unlink %.*s, error %d from destroy\n",
512 dentry->d_name.len, dentry->d_name.name, status);
513 goto bail;
514 }
515 status = simple_unlink(dir, dentry);
516bail:
517 return status;
518}
519
520static int dlmfs_fill_super(struct super_block * sb,
521 void * data,
522 int silent)
523{
524 struct inode * inode;
525 struct dentry * root;
526
527 sb->s_maxbytes = MAX_LFS_FILESIZE;
528 sb->s_blocksize = PAGE_CACHE_SIZE;
529 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
530 sb->s_magic = DLMFS_MAGIC;
531 sb->s_op = &dlmfs_ops;
532 inode = dlmfs_get_root_inode(sb);
533 if (!inode)
534 return -ENOMEM;
535
536 root = d_alloc_root(inode);
537 if (!root) {
538 iput(inode);
539 return -ENOMEM;
540 }
541 sb->s_root = root;
542 return 0;
543}
544
545static struct file_operations dlmfs_file_operations = {
546 .open = dlmfs_file_open,
547 .release = dlmfs_file_release,
548 .read = dlmfs_file_read,
549 .write = dlmfs_file_write,
550};
551
552static struct inode_operations dlmfs_dir_inode_operations = {
553 .create = dlmfs_create,
554 .lookup = simple_lookup,
555 .unlink = dlmfs_unlink,
556};
557
558/* this way we can restrict mkdir to only the toplevel of the fs. */
559static struct inode_operations dlmfs_root_inode_operations = {
560 .lookup = simple_lookup,
561 .mkdir = dlmfs_mkdir,
562 .rmdir = simple_rmdir,
563};
564
565static struct super_operations dlmfs_ops = {
566 .statfs = simple_statfs,
567 .alloc_inode = dlmfs_alloc_inode,
568 .destroy_inode = dlmfs_destroy_inode,
569 .clear_inode = dlmfs_clear_inode,
570 .drop_inode = generic_delete_inode,
571};
572
573static struct inode_operations dlmfs_file_inode_operations = {
574 .getattr = simple_getattr,
575};
576
577static struct super_block *dlmfs_get_sb(struct file_system_type *fs_type,
578 int flags, const char *dev_name, void *data)
579{
580 return get_sb_nodev(fs_type, flags, data, dlmfs_fill_super);
581}
582
583static struct file_system_type dlmfs_fs_type = {
584 .owner = THIS_MODULE,
585 .name = "ocfs2_dlmfs",
586 .get_sb = dlmfs_get_sb,
587 .kill_sb = kill_litter_super,
588};
589
590static int __init init_dlmfs_fs(void)
591{
592 int status;
593 int cleanup_inode = 0, cleanup_worker = 0;
594
595 dlmfs_print_version();
596
597 dlmfs_inode_cache = kmem_cache_create("dlmfs_inode_cache",
598 sizeof(struct dlmfs_inode_private),
599 0, SLAB_HWCACHE_ALIGN|SLAB_RECLAIM_ACCOUNT,
600 dlmfs_init_once, NULL);
601 if (!dlmfs_inode_cache)
602 return -ENOMEM;
603 cleanup_inode = 1;
604
605 user_dlm_worker = create_singlethread_workqueue("user_dlm");
606 if (!user_dlm_worker) {
607 status = -ENOMEM;
608 goto bail;
609 }
610 cleanup_worker = 1;
611
612 status = register_filesystem(&dlmfs_fs_type);
613bail:
614 if (status) {
615 if (cleanup_inode)
616 kmem_cache_destroy(dlmfs_inode_cache);
617 if (cleanup_worker)
618 destroy_workqueue(user_dlm_worker);
619 } else
620 printk("OCFS2 User DLM kernel interface loaded\n");
621 return status;
622}
623
624static void __exit exit_dlmfs_fs(void)
625{
626 unregister_filesystem(&dlmfs_fs_type);
627
628 flush_workqueue(user_dlm_worker);
629 destroy_workqueue(user_dlm_worker);
630
631 if (kmem_cache_destroy(dlmfs_inode_cache))
632 printk(KERN_INFO "dlmfs_inode_cache: not all structures "
633 "were freed\n");
634}
635
636MODULE_AUTHOR("Oracle");
637MODULE_LICENSE("GPL");
638
639module_init(init_dlmfs_fs)
640module_exit(exit_dlmfs_fs)
diff --git a/fs/ocfs2/dlm/dlmfsver.c b/fs/ocfs2/dlm/dlmfsver.c
new file mode 100644
index 000000000000..d2be3ad841f9
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmfsver.c
@@ -0,0 +1,42 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmfsver.c
5 *
6 * version string
7 *
8 * Copyright (C) 2002, 2005 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 */
25
26#include <linux/module.h>
27#include <linux/kernel.h>
28
29#include "dlmfsver.h"
30
31#define DLM_BUILD_VERSION "1.3.3"
32
33#define VERSION_STR "OCFS2 DLMFS " DLM_BUILD_VERSION
34
35void dlmfs_print_version(void)
36{
37 printk(KERN_INFO "%s\n", VERSION_STR);
38}
39
40MODULE_DESCRIPTION(VERSION_STR);
41
42MODULE_VERSION(DLM_BUILD_VERSION);
diff --git a/fs/ocfs2/dlm/dlmfsver.h b/fs/ocfs2/dlm/dlmfsver.h
new file mode 100644
index 000000000000..f35eadbed25c
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmfsver.h
@@ -0,0 +1,31 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmver.h
5 *
6 * Function prototypes
7 *
8 * Copyright (C) 2005 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 */
25
26#ifndef DLMFS_VER_H
27#define DLMFS_VER_H
28
29void dlmfs_print_version(void);
30
31#endif /* DLMFS_VER_H */
diff --git a/fs/ocfs2/dlm/dlmlock.c b/fs/ocfs2/dlm/dlmlock.c
new file mode 100644
index 000000000000..d1a0038557a3
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmlock.c
@@ -0,0 +1,676 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmlock.c
5 *
6 * underlying calls for lock creation
7 *
8 * Copyright (C) 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 *
25 */
26
27
28#include <linux/module.h>
29#include <linux/fs.h>
30#include <linux/types.h>
31#include <linux/slab.h>
32#include <linux/highmem.h>
33#include <linux/utsname.h>
34#include <linux/init.h>
35#include <linux/sysctl.h>
36#include <linux/random.h>
37#include <linux/blkdev.h>
38#include <linux/socket.h>
39#include <linux/inet.h>
40#include <linux/spinlock.h>
41#include <linux/delay.h>
42
43
44#include "cluster/heartbeat.h"
45#include "cluster/nodemanager.h"
46#include "cluster/tcp.h"
47
48#include "dlmapi.h"
49#include "dlmcommon.h"
50
51#include "dlmconvert.h"
52
53#define MLOG_MASK_PREFIX ML_DLM
54#include "cluster/masklog.h"
55
56static spinlock_t dlm_cookie_lock = SPIN_LOCK_UNLOCKED;
57static u64 dlm_next_cookie = 1;
58
59static enum dlm_status dlm_send_remote_lock_request(struct dlm_ctxt *dlm,
60 struct dlm_lock_resource *res,
61 struct dlm_lock *lock, int flags);
62static void dlm_init_lock(struct dlm_lock *newlock, int type,
63 u8 node, u64 cookie);
64static void dlm_lock_release(struct kref *kref);
65static void dlm_lock_detach_lockres(struct dlm_lock *lock);
66
67/* Tell us whether we can grant a new lock request.
68 * locking:
69 * caller needs: res->spinlock
70 * taken: none
71 * held on exit: none
72 * returns: 1 if the lock can be granted, 0 otherwise.
73 */
74static int dlm_can_grant_new_lock(struct dlm_lock_resource *res,
75 struct dlm_lock *lock)
76{
77 struct list_head *iter;
78 struct dlm_lock *tmplock;
79
80 list_for_each(iter, &res->granted) {
81 tmplock = list_entry(iter, struct dlm_lock, list);
82
83 if (!dlm_lock_compatible(tmplock->ml.type, lock->ml.type))
84 return 0;
85 }
86
87 list_for_each(iter, &res->converting) {
88 tmplock = list_entry(iter, struct dlm_lock, list);
89
90 if (!dlm_lock_compatible(tmplock->ml.type, lock->ml.type))
91 return 0;
92 }
93
94 return 1;
95}
96
97/* performs lock creation at the lockres master site
98 * locking:
99 * caller needs: none
100 * taken: takes and drops res->spinlock
101 * held on exit: none
102 * returns: DLM_NORMAL, DLM_NOTQUEUED
103 */
104static enum dlm_status dlmlock_master(struct dlm_ctxt *dlm,
105 struct dlm_lock_resource *res,
106 struct dlm_lock *lock, int flags)
107{
108 int call_ast = 0, kick_thread = 0;
109 enum dlm_status status = DLM_NORMAL;
110
111 mlog_entry("type=%d\n", lock->ml.type);
112
113 spin_lock(&res->spinlock);
114 /* if called from dlm_create_lock_handler, need to
115 * ensure it will not sleep in dlm_wait_on_lockres */
116 status = __dlm_lockres_state_to_status(res);
117 if (status != DLM_NORMAL &&
118 lock->ml.node != dlm->node_num) {
119 /* erf. state changed after lock was dropped. */
120 spin_unlock(&res->spinlock);
121 dlm_error(status);
122 return status;
123 }
124 __dlm_wait_on_lockres(res);
125 __dlm_lockres_reserve_ast(res);
126
127 if (dlm_can_grant_new_lock(res, lock)) {
128 mlog(0, "I can grant this lock right away\n");
129 /* got it right away */
130 lock->lksb->status = DLM_NORMAL;
131 status = DLM_NORMAL;
132 dlm_lock_get(lock);
133 list_add_tail(&lock->list, &res->granted);
134
135 /* for the recovery lock, we can't allow the ast
136 * to be queued since the dlmthread is already
137 * frozen. but the recovery lock is always locked
138 * with LKM_NOQUEUE so we do not need the ast in
139 * this special case */
140 if (!dlm_is_recovery_lock(res->lockname.name,
141 res->lockname.len)) {
142 kick_thread = 1;
143 call_ast = 1;
144 }
145 } else {
146 /* for NOQUEUE request, unless we get the
147 * lock right away, return DLM_NOTQUEUED */
148 if (flags & LKM_NOQUEUE)
149 status = DLM_NOTQUEUED;
150 else {
151 dlm_lock_get(lock);
152 list_add_tail(&lock->list, &res->blocked);
153 kick_thread = 1;
154 }
155 }
156
157 spin_unlock(&res->spinlock);
158 wake_up(&res->wq);
159
160 /* either queue the ast or release it */
161 if (call_ast)
162 dlm_queue_ast(dlm, lock);
163 else
164 dlm_lockres_release_ast(dlm, res);
165
166 dlm_lockres_calc_usage(dlm, res);
167 if (kick_thread)
168 dlm_kick_thread(dlm, res);
169
170 return status;
171}
172
173void dlm_revert_pending_lock(struct dlm_lock_resource *res,
174 struct dlm_lock *lock)
175{
176 /* remove from local queue if it failed */
177 list_del_init(&lock->list);
178 lock->lksb->flags &= ~DLM_LKSB_GET_LVB;
179}
180
181
182/*
183 * locking:
184 * caller needs: none
185 * taken: takes and drops res->spinlock
186 * held on exit: none
187 * returns: DLM_DENIED, DLM_RECOVERING, or net status
188 */
189static enum dlm_status dlmlock_remote(struct dlm_ctxt *dlm,
190 struct dlm_lock_resource *res,
191 struct dlm_lock *lock, int flags)
192{
193 enum dlm_status status = DLM_DENIED;
194
195 mlog_entry("type=%d\n", lock->ml.type);
196 mlog(0, "lockres %.*s, flags = 0x%x\n", res->lockname.len,
197 res->lockname.name, flags);
198
199 spin_lock(&res->spinlock);
200
201 /* will exit this call with spinlock held */
202 __dlm_wait_on_lockres(res);
203 res->state |= DLM_LOCK_RES_IN_PROGRESS;
204
205 /* add lock to local (secondary) queue */
206 dlm_lock_get(lock);
207 list_add_tail(&lock->list, &res->blocked);
208 lock->lock_pending = 1;
209 spin_unlock(&res->spinlock);
210
211 /* spec seems to say that you will get DLM_NORMAL when the lock
212 * has been queued, meaning we need to wait for a reply here. */
213 status = dlm_send_remote_lock_request(dlm, res, lock, flags);
214
215 spin_lock(&res->spinlock);
216 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
217 lock->lock_pending = 0;
218 if (status != DLM_NORMAL) {
219 if (status != DLM_NOTQUEUED)
220 dlm_error(status);
221 dlm_revert_pending_lock(res, lock);
222 dlm_lock_put(lock);
223 }
224 spin_unlock(&res->spinlock);
225
226 dlm_lockres_calc_usage(dlm, res);
227
228 wake_up(&res->wq);
229 return status;
230}
231
232
233/* for remote lock creation.
234 * locking:
235 * caller needs: none, but need res->state & DLM_LOCK_RES_IN_PROGRESS
236 * taken: none
237 * held on exit: none
238 * returns: DLM_NOLOCKMGR, or net status
239 */
240static enum dlm_status dlm_send_remote_lock_request(struct dlm_ctxt *dlm,
241 struct dlm_lock_resource *res,
242 struct dlm_lock *lock, int flags)
243{
244 struct dlm_create_lock create;
245 int tmpret, status = 0;
246 enum dlm_status ret;
247
248 mlog_entry_void();
249
250 memset(&create, 0, sizeof(create));
251 create.node_idx = dlm->node_num;
252 create.requested_type = lock->ml.type;
253 create.cookie = lock->ml.cookie;
254 create.namelen = res->lockname.len;
255 create.flags = cpu_to_be32(flags);
256 memcpy(create.name, res->lockname.name, create.namelen);
257
258 tmpret = o2net_send_message(DLM_CREATE_LOCK_MSG, dlm->key, &create,
259 sizeof(create), res->owner, &status);
260 if (tmpret >= 0) {
261 // successfully sent and received
262 ret = status; // this is already a dlm_status
263 } else {
264 mlog_errno(tmpret);
265 if (dlm_is_host_down(tmpret)) {
266 ret = DLM_RECOVERING;
267 mlog(0, "node %u died so returning DLM_RECOVERING "
268 "from lock message!\n", res->owner);
269 } else {
270 ret = dlm_err_to_dlm_status(tmpret);
271 }
272 }
273
274 return ret;
275}
276
277void dlm_lock_get(struct dlm_lock *lock)
278{
279 kref_get(&lock->lock_refs);
280}
281
282void dlm_lock_put(struct dlm_lock *lock)
283{
284 kref_put(&lock->lock_refs, dlm_lock_release);
285}
286
287static void dlm_lock_release(struct kref *kref)
288{
289 struct dlm_lock *lock;
290
291 lock = container_of(kref, struct dlm_lock, lock_refs);
292
293 BUG_ON(!list_empty(&lock->list));
294 BUG_ON(!list_empty(&lock->ast_list));
295 BUG_ON(!list_empty(&lock->bast_list));
296 BUG_ON(lock->ast_pending);
297 BUG_ON(lock->bast_pending);
298
299 dlm_lock_detach_lockres(lock);
300
301 if (lock->lksb_kernel_allocated) {
302 mlog(0, "freeing kernel-allocated lksb\n");
303 kfree(lock->lksb);
304 }
305 kfree(lock);
306}
307
308/* associate a lock with it's lockres, getting a ref on the lockres */
309void dlm_lock_attach_lockres(struct dlm_lock *lock,
310 struct dlm_lock_resource *res)
311{
312 dlm_lockres_get(res);
313 lock->lockres = res;
314}
315
316/* drop ref on lockres, if there is still one associated with lock */
317static void dlm_lock_detach_lockres(struct dlm_lock *lock)
318{
319 struct dlm_lock_resource *res;
320
321 res = lock->lockres;
322 if (res) {
323 lock->lockres = NULL;
324 mlog(0, "removing lock's lockres reference\n");
325 dlm_lockres_put(res);
326 }
327}
328
329static void dlm_init_lock(struct dlm_lock *newlock, int type,
330 u8 node, u64 cookie)
331{
332 INIT_LIST_HEAD(&newlock->list);
333 INIT_LIST_HEAD(&newlock->ast_list);
334 INIT_LIST_HEAD(&newlock->bast_list);
335 spin_lock_init(&newlock->spinlock);
336 newlock->ml.type = type;
337 newlock->ml.convert_type = LKM_IVMODE;
338 newlock->ml.highest_blocked = LKM_IVMODE;
339 newlock->ml.node = node;
340 newlock->ml.pad1 = 0;
341 newlock->ml.list = 0;
342 newlock->ml.flags = 0;
343 newlock->ast = NULL;
344 newlock->bast = NULL;
345 newlock->astdata = NULL;
346 newlock->ml.cookie = cpu_to_be64(cookie);
347 newlock->ast_pending = 0;
348 newlock->bast_pending = 0;
349 newlock->convert_pending = 0;
350 newlock->lock_pending = 0;
351 newlock->unlock_pending = 0;
352 newlock->cancel_pending = 0;
353 newlock->lksb_kernel_allocated = 0;
354
355 kref_init(&newlock->lock_refs);
356}
357
358struct dlm_lock * dlm_new_lock(int type, u8 node, u64 cookie,
359 struct dlm_lockstatus *lksb)
360{
361 struct dlm_lock *lock;
362 int kernel_allocated = 0;
363
364 lock = kcalloc(1, sizeof(*lock), GFP_KERNEL);
365 if (!lock)
366 return NULL;
367
368 if (!lksb) {
369 /* zero memory only if kernel-allocated */
370 lksb = kcalloc(1, sizeof(*lksb), GFP_KERNEL);
371 if (!lksb) {
372 kfree(lock);
373 return NULL;
374 }
375 kernel_allocated = 1;
376 }
377
378 dlm_init_lock(lock, type, node, cookie);
379 if (kernel_allocated)
380 lock->lksb_kernel_allocated = 1;
381 lock->lksb = lksb;
382 lksb->lockid = lock;
383 return lock;
384}
385
386/* handler for lock creation net message
387 * locking:
388 * caller needs: none
389 * taken: takes and drops res->spinlock
390 * held on exit: none
391 * returns: DLM_NORMAL, DLM_SYSERR, DLM_IVLOCKID, DLM_NOTQUEUED
392 */
393int dlm_create_lock_handler(struct o2net_msg *msg, u32 len, void *data)
394{
395 struct dlm_ctxt *dlm = data;
396 struct dlm_create_lock *create = (struct dlm_create_lock *)msg->buf;
397 struct dlm_lock_resource *res = NULL;
398 struct dlm_lock *newlock = NULL;
399 struct dlm_lockstatus *lksb = NULL;
400 enum dlm_status status = DLM_NORMAL;
401 char *name;
402 unsigned int namelen;
403
404 BUG_ON(!dlm);
405
406 mlog_entry_void();
407
408 if (!dlm_grab(dlm))
409 return DLM_REJECTED;
410
411 mlog_bug_on_msg(!dlm_domain_fully_joined(dlm),
412 "Domain %s not fully joined!\n", dlm->name);
413
414 name = create->name;
415 namelen = create->namelen;
416
417 status = DLM_IVBUFLEN;
418 if (namelen > DLM_LOCKID_NAME_MAX) {
419 dlm_error(status);
420 goto leave;
421 }
422
423 status = DLM_SYSERR;
424 newlock = dlm_new_lock(create->requested_type,
425 create->node_idx,
426 be64_to_cpu(create->cookie), NULL);
427 if (!newlock) {
428 dlm_error(status);
429 goto leave;
430 }
431
432 lksb = newlock->lksb;
433
434 if (be32_to_cpu(create->flags) & LKM_GET_LVB) {
435 lksb->flags |= DLM_LKSB_GET_LVB;
436 mlog(0, "set DLM_LKSB_GET_LVB flag\n");
437 }
438
439 status = DLM_IVLOCKID;
440 res = dlm_lookup_lockres(dlm, name, namelen);
441 if (!res) {
442 dlm_error(status);
443 goto leave;
444 }
445
446 spin_lock(&res->spinlock);
447 status = __dlm_lockres_state_to_status(res);
448 spin_unlock(&res->spinlock);
449
450 if (status != DLM_NORMAL) {
451 mlog(0, "lockres recovering/migrating/in-progress\n");
452 goto leave;
453 }
454
455 dlm_lock_attach_lockres(newlock, res);
456
457 status = dlmlock_master(dlm, res, newlock, be32_to_cpu(create->flags));
458leave:
459 if (status != DLM_NORMAL)
460 if (newlock)
461 dlm_lock_put(newlock);
462
463 if (res)
464 dlm_lockres_put(res);
465
466 dlm_put(dlm);
467
468 return status;
469}
470
471
472/* fetch next node-local (u8 nodenum + u56 cookie) into u64 */
473static inline void dlm_get_next_cookie(u8 node_num, u64 *cookie)
474{
475 u64 tmpnode = node_num;
476
477 /* shift single byte of node num into top 8 bits */
478 tmpnode <<= 56;
479
480 spin_lock(&dlm_cookie_lock);
481 *cookie = (dlm_next_cookie | tmpnode);
482 if (++dlm_next_cookie & 0xff00000000000000ull) {
483 mlog(0, "This node's cookie will now wrap!\n");
484 dlm_next_cookie = 1;
485 }
486 spin_unlock(&dlm_cookie_lock);
487}
488
489enum dlm_status dlmlock(struct dlm_ctxt *dlm, int mode,
490 struct dlm_lockstatus *lksb, int flags,
491 const char *name, dlm_astlockfunc_t *ast, void *data,
492 dlm_bastlockfunc_t *bast)
493{
494 enum dlm_status status;
495 struct dlm_lock_resource *res = NULL;
496 struct dlm_lock *lock = NULL;
497 int convert = 0, recovery = 0;
498
499 /* yes this function is a mess.
500 * TODO: clean this up. lots of common code in the
501 * lock and convert paths, especially in the retry blocks */
502 if (!lksb) {
503 dlm_error(DLM_BADARGS);
504 return DLM_BADARGS;
505 }
506
507 status = DLM_BADPARAM;
508 if (mode != LKM_EXMODE && mode != LKM_PRMODE && mode != LKM_NLMODE) {
509 dlm_error(status);
510 goto error;
511 }
512
513 if (flags & ~LKM_VALID_FLAGS) {
514 dlm_error(status);
515 goto error;
516 }
517
518 convert = (flags & LKM_CONVERT);
519 recovery = (flags & LKM_RECOVERY);
520
521 if (recovery &&
522 (!dlm_is_recovery_lock(name, strlen(name)) || convert) ) {
523 dlm_error(status);
524 goto error;
525 }
526 if (convert && (flags & LKM_LOCAL)) {
527 mlog(ML_ERROR, "strange LOCAL convert request!\n");
528 goto error;
529 }
530
531 if (convert) {
532 /* CONVERT request */
533
534 /* if converting, must pass in a valid dlm_lock */
535 lock = lksb->lockid;
536 if (!lock) {
537 mlog(ML_ERROR, "NULL lock pointer in convert "
538 "request\n");
539 goto error;
540 }
541
542 res = lock->lockres;
543 if (!res) {
544 mlog(ML_ERROR, "NULL lockres pointer in convert "
545 "request\n");
546 goto error;
547 }
548 dlm_lockres_get(res);
549
550 /* XXX: for ocfs2 purposes, the ast/bast/astdata/lksb are
551 * static after the original lock call. convert requests will
552 * ensure that everything is the same, or return DLM_BADARGS.
553 * this means that DLM_DENIED_NOASTS will never be returned.
554 */
555 if (lock->lksb != lksb || lock->ast != ast ||
556 lock->bast != bast || lock->astdata != data) {
557 status = DLM_BADARGS;
558 mlog(ML_ERROR, "new args: lksb=%p, ast=%p, bast=%p, "
559 "astdata=%p\n", lksb, ast, bast, data);
560 mlog(ML_ERROR, "orig args: lksb=%p, ast=%p, bast=%p, "
561 "astdata=%p\n", lock->lksb, lock->ast,
562 lock->bast, lock->astdata);
563 goto error;
564 }
565retry_convert:
566 dlm_wait_for_recovery(dlm);
567
568 if (res->owner == dlm->node_num)
569 status = dlmconvert_master(dlm, res, lock, flags, mode);
570 else
571 status = dlmconvert_remote(dlm, res, lock, flags, mode);
572 if (status == DLM_RECOVERING || status == DLM_MIGRATING ||
573 status == DLM_FORWARD) {
574 /* for now, see how this works without sleeping
575 * and just retry right away. I suspect the reco
576 * or migration will complete fast enough that
577 * no waiting will be necessary */
578 mlog(0, "retrying convert with migration/recovery/"
579 "in-progress\n");
580 msleep(100);
581 goto retry_convert;
582 }
583 } else {
584 u64 tmpcookie;
585
586 /* LOCK request */
587 status = DLM_BADARGS;
588 if (!name) {
589 dlm_error(status);
590 goto error;
591 }
592
593 status = DLM_IVBUFLEN;
594 if (strlen(name) > DLM_LOCKID_NAME_MAX || strlen(name) < 1) {
595 dlm_error(status);
596 goto error;
597 }
598
599 dlm_get_next_cookie(dlm->node_num, &tmpcookie);
600 lock = dlm_new_lock(mode, dlm->node_num, tmpcookie, lksb);
601 if (!lock) {
602 dlm_error(status);
603 goto error;
604 }
605
606 if (!recovery)
607 dlm_wait_for_recovery(dlm);
608
609 /* find or create the lock resource */
610 res = dlm_get_lock_resource(dlm, name, flags);
611 if (!res) {
612 status = DLM_IVLOCKID;
613 dlm_error(status);
614 goto error;
615 }
616
617 mlog(0, "type=%d, flags = 0x%x\n", mode, flags);
618 mlog(0, "creating lock: lock=%p res=%p\n", lock, res);
619
620 dlm_lock_attach_lockres(lock, res);
621 lock->ast = ast;
622 lock->bast = bast;
623 lock->astdata = data;
624
625retry_lock:
626 if (flags & LKM_VALBLK) {
627 mlog(0, "LKM_VALBLK passed by caller\n");
628
629 /* LVB requests for non PR, PW or EX locks are
630 * ignored. */
631 if (mode < LKM_PRMODE)
632 flags &= ~LKM_VALBLK;
633 else {
634 flags |= LKM_GET_LVB;
635 lock->lksb->flags |= DLM_LKSB_GET_LVB;
636 }
637 }
638
639 if (res->owner == dlm->node_num)
640 status = dlmlock_master(dlm, res, lock, flags);
641 else
642 status = dlmlock_remote(dlm, res, lock, flags);
643
644 if (status == DLM_RECOVERING || status == DLM_MIGRATING ||
645 status == DLM_FORWARD) {
646 mlog(0, "retrying lock with migration/"
647 "recovery/in progress\n");
648 msleep(100);
649 dlm_wait_for_recovery(dlm);
650 goto retry_lock;
651 }
652
653 if (status != DLM_NORMAL) {
654 lock->lksb->flags &= ~DLM_LKSB_GET_LVB;
655 if (status != DLM_NOTQUEUED)
656 dlm_error(status);
657 goto error;
658 }
659 }
660
661error:
662 if (status != DLM_NORMAL) {
663 if (lock && !convert)
664 dlm_lock_put(lock);
665 // this is kind of unnecessary
666 lksb->status = status;
667 }
668
669 /* put lockres ref from the convert path
670 * or from dlm_get_lock_resource */
671 if (res)
672 dlm_lockres_put(res);
673
674 return status;
675}
676EXPORT_SYMBOL_GPL(dlmlock);
diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
new file mode 100644
index 000000000000..27e984f7e4cd
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmmaster.c
@@ -0,0 +1,2664 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmmod.c
5 *
6 * standalone DLM module
7 *
8 * Copyright (C) 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 *
25 */
26
27
28#include <linux/module.h>
29#include <linux/fs.h>
30#include <linux/types.h>
31#include <linux/slab.h>
32#include <linux/highmem.h>
33#include <linux/utsname.h>
34#include <linux/init.h>
35#include <linux/sysctl.h>
36#include <linux/random.h>
37#include <linux/blkdev.h>
38#include <linux/socket.h>
39#include <linux/inet.h>
40#include <linux/spinlock.h>
41#include <linux/delay.h>
42
43
44#include "cluster/heartbeat.h"
45#include "cluster/nodemanager.h"
46#include "cluster/tcp.h"
47
48#include "dlmapi.h"
49#include "dlmcommon.h"
50#include "dlmdebug.h"
51#include "dlmdomain.h"
52
53#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_MASTER)
54#include "cluster/masklog.h"
55
56enum dlm_mle_type {
57 DLM_MLE_BLOCK,
58 DLM_MLE_MASTER,
59 DLM_MLE_MIGRATION
60};
61
62struct dlm_lock_name
63{
64 u8 len;
65 u8 name[DLM_LOCKID_NAME_MAX];
66};
67
68struct dlm_master_list_entry
69{
70 struct list_head list;
71 struct list_head hb_events;
72 struct dlm_ctxt *dlm;
73 spinlock_t spinlock;
74 wait_queue_head_t wq;
75 atomic_t woken;
76 struct kref mle_refs;
77 unsigned long maybe_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
78 unsigned long vote_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
79 unsigned long response_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
80 unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
81 u8 master;
82 u8 new_master;
83 enum dlm_mle_type type;
84 struct o2hb_callback_func mle_hb_up;
85 struct o2hb_callback_func mle_hb_down;
86 union {
87 struct dlm_lock_resource *res;
88 struct dlm_lock_name name;
89 } u;
90};
91
92static void dlm_mle_node_down(struct dlm_ctxt *dlm,
93 struct dlm_master_list_entry *mle,
94 struct o2nm_node *node,
95 int idx);
96static void dlm_mle_node_up(struct dlm_ctxt *dlm,
97 struct dlm_master_list_entry *mle,
98 struct o2nm_node *node,
99 int idx);
100
101static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
102static int dlm_do_assert_master(struct dlm_ctxt *dlm, const char *lockname,
103 unsigned int namelen, void *nodemap,
104 u32 flags);
105
106static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
107 struct dlm_master_list_entry *mle,
108 const char *name,
109 unsigned int namelen)
110{
111 struct dlm_lock_resource *res;
112
113 if (dlm != mle->dlm)
114 return 0;
115
116 if (mle->type == DLM_MLE_BLOCK ||
117 mle->type == DLM_MLE_MIGRATION) {
118 if (namelen != mle->u.name.len ||
119 memcmp(name, mle->u.name.name, namelen)!=0)
120 return 0;
121 } else {
122 res = mle->u.res;
123 if (namelen != res->lockname.len ||
124 memcmp(res->lockname.name, name, namelen) != 0)
125 return 0;
126 }
127 return 1;
128}
129
130#if 0
131/* Code here is included but defined out as it aids debugging */
132
133void dlm_print_one_mle(struct dlm_master_list_entry *mle)
134{
135 int i = 0, refs;
136 char *type;
137 char attached;
138 u8 master;
139 unsigned int namelen;
140 const char *name;
141 struct kref *k;
142
143 k = &mle->mle_refs;
144 if (mle->type == DLM_MLE_BLOCK)
145 type = "BLK";
146 else if (mle->type == DLM_MLE_MASTER)
147 type = "MAS";
148 else
149 type = "MIG";
150 refs = atomic_read(&k->refcount);
151 master = mle->master;
152 attached = (list_empty(&mle->hb_events) ? 'N' : 'Y');
153
154 if (mle->type != DLM_MLE_MASTER) {
155 namelen = mle->u.name.len;
156 name = mle->u.name.name;
157 } else {
158 namelen = mle->u.res->lockname.len;
159 name = mle->u.res->lockname.name;
160 }
161
162 mlog(ML_NOTICE, " #%3d: %3s %3d %3u %3u %c (%d)%.*s\n",
163 i, type, refs, master, mle->new_master, attached,
164 namelen, namelen, name);
165}
166
167static void dlm_dump_mles(struct dlm_ctxt *dlm)
168{
169 struct dlm_master_list_entry *mle;
170 struct list_head *iter;
171
172 mlog(ML_NOTICE, "dumping all mles for domain %s:\n", dlm->name);
173 mlog(ML_NOTICE, " ####: type refs owner new events? lockname nodemap votemap respmap maybemap\n");
174 spin_lock(&dlm->master_lock);
175 list_for_each(iter, &dlm->master_list) {
176 mle = list_entry(iter, struct dlm_master_list_entry, list);
177 dlm_print_one_mle(mle);
178 }
179 spin_unlock(&dlm->master_lock);
180}
181
182int dlm_dump_all_mles(const char __user *data, unsigned int len)
183{
184 struct list_head *iter;
185 struct dlm_ctxt *dlm;
186
187 spin_lock(&dlm_domain_lock);
188 list_for_each(iter, &dlm_domains) {
189 dlm = list_entry (iter, struct dlm_ctxt, list);
190 mlog(ML_NOTICE, "found dlm: %p, name=%s\n", dlm, dlm->name);
191 dlm_dump_mles(dlm);
192 }
193 spin_unlock(&dlm_domain_lock);
194 return len;
195}
196EXPORT_SYMBOL_GPL(dlm_dump_all_mles);
197
198#endif /* 0 */
199
200
201static kmem_cache_t *dlm_mle_cache = NULL;
202
203
204static void dlm_mle_release(struct kref *kref);
205static void dlm_init_mle(struct dlm_master_list_entry *mle,
206 enum dlm_mle_type type,
207 struct dlm_ctxt *dlm,
208 struct dlm_lock_resource *res,
209 const char *name,
210 unsigned int namelen);
211static void dlm_put_mle(struct dlm_master_list_entry *mle);
212static void __dlm_put_mle(struct dlm_master_list_entry *mle);
213static int dlm_find_mle(struct dlm_ctxt *dlm,
214 struct dlm_master_list_entry **mle,
215 char *name, unsigned int namelen);
216
217static int dlm_do_master_request(struct dlm_master_list_entry *mle, int to);
218
219
220static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
221 struct dlm_lock_resource *res,
222 struct dlm_master_list_entry *mle,
223 int *blocked);
224static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
225 struct dlm_lock_resource *res,
226 struct dlm_master_list_entry *mle,
227 int blocked);
228static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
229 struct dlm_lock_resource *res,
230 struct dlm_master_list_entry *mle,
231 struct dlm_master_list_entry **oldmle,
232 const char *name, unsigned int namelen,
233 u8 new_master, u8 master);
234
235static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
236 struct dlm_lock_resource *res);
237static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
238 struct dlm_lock_resource *res);
239static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
240 struct dlm_lock_resource *res,
241 u8 target);
242
243
244int dlm_is_host_down(int errno)
245{
246 switch (errno) {
247 case -EBADF:
248 case -ECONNREFUSED:
249 case -ENOTCONN:
250 case -ECONNRESET:
251 case -EPIPE:
252 case -EHOSTDOWN:
253 case -EHOSTUNREACH:
254 case -ETIMEDOUT:
255 case -ECONNABORTED:
256 case -ENETDOWN:
257 case -ENETUNREACH:
258 case -ENETRESET:
259 case -ESHUTDOWN:
260 case -ENOPROTOOPT:
261 case -EINVAL: /* if returned from our tcp code,
262 this means there is no socket */
263 return 1;
264 }
265 return 0;
266}
267
268
269/*
270 * MASTER LIST FUNCTIONS
271 */
272
273
274/*
275 * regarding master list entries and heartbeat callbacks:
276 *
277 * in order to avoid sleeping and allocation that occurs in
278 * heartbeat, master list entries are simply attached to the
279 * dlm's established heartbeat callbacks. the mle is attached
280 * when it is created, and since the dlm->spinlock is held at
281 * that time, any heartbeat event will be properly discovered
282 * by the mle. the mle needs to be detached from the
283 * dlm->mle_hb_events list as soon as heartbeat events are no
284 * longer useful to the mle, and before the mle is freed.
285 *
286 * as a general rule, heartbeat events are no longer needed by
287 * the mle once an "answer" regarding the lock master has been
288 * received.
289 */
290static inline void __dlm_mle_attach_hb_events(struct dlm_ctxt *dlm,
291 struct dlm_master_list_entry *mle)
292{
293 assert_spin_locked(&dlm->spinlock);
294
295 list_add_tail(&mle->hb_events, &dlm->mle_hb_events);
296}
297
298
299static inline void __dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
300 struct dlm_master_list_entry *mle)
301{
302 if (!list_empty(&mle->hb_events))
303 list_del_init(&mle->hb_events);
304}
305
306
307static inline void dlm_mle_detach_hb_events(struct dlm_ctxt *dlm,
308 struct dlm_master_list_entry *mle)
309{
310 spin_lock(&dlm->spinlock);
311 __dlm_mle_detach_hb_events(dlm, mle);
312 spin_unlock(&dlm->spinlock);
313}
314
315/* remove from list and free */
316static void __dlm_put_mle(struct dlm_master_list_entry *mle)
317{
318 struct dlm_ctxt *dlm;
319 dlm = mle->dlm;
320
321 assert_spin_locked(&dlm->spinlock);
322 assert_spin_locked(&dlm->master_lock);
323 BUG_ON(!atomic_read(&mle->mle_refs.refcount));
324
325 kref_put(&mle->mle_refs, dlm_mle_release);
326}
327
328
329/* must not have any spinlocks coming in */
330static void dlm_put_mle(struct dlm_master_list_entry *mle)
331{
332 struct dlm_ctxt *dlm;
333 dlm = mle->dlm;
334
335 spin_lock(&dlm->spinlock);
336 spin_lock(&dlm->master_lock);
337 __dlm_put_mle(mle);
338 spin_unlock(&dlm->master_lock);
339 spin_unlock(&dlm->spinlock);
340}
341
342static inline void dlm_get_mle(struct dlm_master_list_entry *mle)
343{
344 kref_get(&mle->mle_refs);
345}
346
347static void dlm_init_mle(struct dlm_master_list_entry *mle,
348 enum dlm_mle_type type,
349 struct dlm_ctxt *dlm,
350 struct dlm_lock_resource *res,
351 const char *name,
352 unsigned int namelen)
353{
354 assert_spin_locked(&dlm->spinlock);
355
356 mle->dlm = dlm;
357 mle->type = type;
358 INIT_LIST_HEAD(&mle->list);
359 INIT_LIST_HEAD(&mle->hb_events);
360 memset(mle->maybe_map, 0, sizeof(mle->maybe_map));
361 spin_lock_init(&mle->spinlock);
362 init_waitqueue_head(&mle->wq);
363 atomic_set(&mle->woken, 0);
364 kref_init(&mle->mle_refs);
365 memset(mle->response_map, 0, sizeof(mle->response_map));
366 mle->master = O2NM_MAX_NODES;
367 mle->new_master = O2NM_MAX_NODES;
368
369 if (mle->type == DLM_MLE_MASTER) {
370 BUG_ON(!res);
371 mle->u.res = res;
372 } else if (mle->type == DLM_MLE_BLOCK) {
373 BUG_ON(!name);
374 memcpy(mle->u.name.name, name, namelen);
375 mle->u.name.len = namelen;
376 } else /* DLM_MLE_MIGRATION */ {
377 BUG_ON(!name);
378 memcpy(mle->u.name.name, name, namelen);
379 mle->u.name.len = namelen;
380 }
381
382 /* copy off the node_map and register hb callbacks on our copy */
383 memcpy(mle->node_map, dlm->domain_map, sizeof(mle->node_map));
384 memcpy(mle->vote_map, dlm->domain_map, sizeof(mle->vote_map));
385 clear_bit(dlm->node_num, mle->vote_map);
386 clear_bit(dlm->node_num, mle->node_map);
387
388 /* attach the mle to the domain node up/down events */
389 __dlm_mle_attach_hb_events(dlm, mle);
390}
391
392
393/* returns 1 if found, 0 if not */
394static int dlm_find_mle(struct dlm_ctxt *dlm,
395 struct dlm_master_list_entry **mle,
396 char *name, unsigned int namelen)
397{
398 struct dlm_master_list_entry *tmpmle;
399 struct list_head *iter;
400
401 assert_spin_locked(&dlm->master_lock);
402
403 list_for_each(iter, &dlm->master_list) {
404 tmpmle = list_entry(iter, struct dlm_master_list_entry, list);
405 if (!dlm_mle_equal(dlm, tmpmle, name, namelen))
406 continue;
407 dlm_get_mle(tmpmle);
408 *mle = tmpmle;
409 return 1;
410 }
411 return 0;
412}
413
414void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up)
415{
416 struct dlm_master_list_entry *mle;
417 struct list_head *iter;
418
419 assert_spin_locked(&dlm->spinlock);
420
421 list_for_each(iter, &dlm->mle_hb_events) {
422 mle = list_entry(iter, struct dlm_master_list_entry,
423 hb_events);
424 if (node_up)
425 dlm_mle_node_up(dlm, mle, NULL, idx);
426 else
427 dlm_mle_node_down(dlm, mle, NULL, idx);
428 }
429}
430
431static void dlm_mle_node_down(struct dlm_ctxt *dlm,
432 struct dlm_master_list_entry *mle,
433 struct o2nm_node *node, int idx)
434{
435 spin_lock(&mle->spinlock);
436
437 if (!test_bit(idx, mle->node_map))
438 mlog(0, "node %u already removed from nodemap!\n", idx);
439 else
440 clear_bit(idx, mle->node_map);
441
442 spin_unlock(&mle->spinlock);
443}
444
445static void dlm_mle_node_up(struct dlm_ctxt *dlm,
446 struct dlm_master_list_entry *mle,
447 struct o2nm_node *node, int idx)
448{
449 spin_lock(&mle->spinlock);
450
451 if (test_bit(idx, mle->node_map))
452 mlog(0, "node %u already in node map!\n", idx);
453 else
454 set_bit(idx, mle->node_map);
455
456 spin_unlock(&mle->spinlock);
457}
458
459
460int dlm_init_mle_cache(void)
461{
462 dlm_mle_cache = kmem_cache_create("dlm_mle_cache",
463 sizeof(struct dlm_master_list_entry),
464 0, SLAB_HWCACHE_ALIGN,
465 NULL, NULL);
466 if (dlm_mle_cache == NULL)
467 return -ENOMEM;
468 return 0;
469}
470
471void dlm_destroy_mle_cache(void)
472{
473 if (dlm_mle_cache)
474 kmem_cache_destroy(dlm_mle_cache);
475}
476
477static void dlm_mle_release(struct kref *kref)
478{
479 struct dlm_master_list_entry *mle;
480 struct dlm_ctxt *dlm;
481
482 mlog_entry_void();
483
484 mle = container_of(kref, struct dlm_master_list_entry, mle_refs);
485 dlm = mle->dlm;
486
487 if (mle->type != DLM_MLE_MASTER) {
488 mlog(0, "calling mle_release for %.*s, type %d\n",
489 mle->u.name.len, mle->u.name.name, mle->type);
490 } else {
491 mlog(0, "calling mle_release for %.*s, type %d\n",
492 mle->u.res->lockname.len,
493 mle->u.res->lockname.name, mle->type);
494 }
495 assert_spin_locked(&dlm->spinlock);
496 assert_spin_locked(&dlm->master_lock);
497
498 /* remove from list if not already */
499 if (!list_empty(&mle->list))
500 list_del_init(&mle->list);
501
502 /* detach the mle from the domain node up/down events */
503 __dlm_mle_detach_hb_events(dlm, mle);
504
505 /* NOTE: kfree under spinlock here.
506 * if this is bad, we can move this to a freelist. */
507 kmem_cache_free(dlm_mle_cache, mle);
508}
509
510
511/*
512 * LOCK RESOURCE FUNCTIONS
513 */
514
515static void dlm_set_lockres_owner(struct dlm_ctxt *dlm,
516 struct dlm_lock_resource *res,
517 u8 owner)
518{
519 assert_spin_locked(&res->spinlock);
520
521 mlog_entry("%.*s, %u\n", res->lockname.len, res->lockname.name, owner);
522
523 if (owner == dlm->node_num)
524 atomic_inc(&dlm->local_resources);
525 else if (owner == DLM_LOCK_RES_OWNER_UNKNOWN)
526 atomic_inc(&dlm->unknown_resources);
527 else
528 atomic_inc(&dlm->remote_resources);
529
530 res->owner = owner;
531}
532
533void dlm_change_lockres_owner(struct dlm_ctxt *dlm,
534 struct dlm_lock_resource *res, u8 owner)
535{
536 assert_spin_locked(&res->spinlock);
537
538 if (owner == res->owner)
539 return;
540
541 if (res->owner == dlm->node_num)
542 atomic_dec(&dlm->local_resources);
543 else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN)
544 atomic_dec(&dlm->unknown_resources);
545 else
546 atomic_dec(&dlm->remote_resources);
547
548 dlm_set_lockres_owner(dlm, res, owner);
549}
550
551
552static void dlm_lockres_release(struct kref *kref)
553{
554 struct dlm_lock_resource *res;
555
556 res = container_of(kref, struct dlm_lock_resource, refs);
557
558 /* This should not happen -- all lockres' have a name
559 * associated with them at init time. */
560 BUG_ON(!res->lockname.name);
561
562 mlog(0, "destroying lockres %.*s\n", res->lockname.len,
563 res->lockname.name);
564
565 /* By the time we're ready to blow this guy away, we shouldn't
566 * be on any lists. */
567 BUG_ON(!list_empty(&res->list));
568 BUG_ON(!list_empty(&res->granted));
569 BUG_ON(!list_empty(&res->converting));
570 BUG_ON(!list_empty(&res->blocked));
571 BUG_ON(!list_empty(&res->dirty));
572 BUG_ON(!list_empty(&res->recovering));
573 BUG_ON(!list_empty(&res->purge));
574
575 kfree(res->lockname.name);
576
577 kfree(res);
578}
579
580void dlm_lockres_get(struct dlm_lock_resource *res)
581{
582 kref_get(&res->refs);
583}
584
585void dlm_lockres_put(struct dlm_lock_resource *res)
586{
587 kref_put(&res->refs, dlm_lockres_release);
588}
589
590static void dlm_init_lockres(struct dlm_ctxt *dlm,
591 struct dlm_lock_resource *res,
592 const char *name, unsigned int namelen)
593{
594 char *qname;
595
596 /* If we memset here, we lose our reference to the kmalloc'd
597 * res->lockname.name, so be sure to init every field
598 * correctly! */
599
600 qname = (char *) res->lockname.name;
601 memcpy(qname, name, namelen);
602
603 res->lockname.len = namelen;
604 res->lockname.hash = full_name_hash(name, namelen);
605
606 init_waitqueue_head(&res->wq);
607 spin_lock_init(&res->spinlock);
608 INIT_LIST_HEAD(&res->list);
609 INIT_LIST_HEAD(&res->granted);
610 INIT_LIST_HEAD(&res->converting);
611 INIT_LIST_HEAD(&res->blocked);
612 INIT_LIST_HEAD(&res->dirty);
613 INIT_LIST_HEAD(&res->recovering);
614 INIT_LIST_HEAD(&res->purge);
615 atomic_set(&res->asts_reserved, 0);
616 res->migration_pending = 0;
617
618 kref_init(&res->refs);
619
620 /* just for consistency */
621 spin_lock(&res->spinlock);
622 dlm_set_lockres_owner(dlm, res, DLM_LOCK_RES_OWNER_UNKNOWN);
623 spin_unlock(&res->spinlock);
624
625 res->state = DLM_LOCK_RES_IN_PROGRESS;
626
627 res->last_used = 0;
628
629 memset(res->lvb, 0, DLM_LVB_LEN);
630}
631
632struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
633 const char *name,
634 unsigned int namelen)
635{
636 struct dlm_lock_resource *res;
637
638 res = kmalloc(sizeof(struct dlm_lock_resource), GFP_KERNEL);
639 if (!res)
640 return NULL;
641
642 res->lockname.name = kmalloc(namelen, GFP_KERNEL);
643 if (!res->lockname.name) {
644 kfree(res);
645 return NULL;
646 }
647
648 dlm_init_lockres(dlm, res, name, namelen);
649 return res;
650}
651
652/*
653 * lookup a lock resource by name.
654 * may already exist in the hashtable.
655 * lockid is null terminated
656 *
657 * if not, allocate enough for the lockres and for
658 * the temporary structure used in doing the mastering.
659 *
660 * also, do a lookup in the dlm->master_list to see
661 * if another node has begun mastering the same lock.
662 * if so, there should be a block entry in there
663 * for this name, and we should *not* attempt to master
664 * the lock here. need to wait around for that node
665 * to assert_master (or die).
666 *
667 */
668struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
669 const char *lockid,
670 int flags)
671{
672 struct dlm_lock_resource *tmpres=NULL, *res=NULL;
673 struct dlm_master_list_entry *mle = NULL;
674 struct dlm_master_list_entry *alloc_mle = NULL;
675 int blocked = 0;
676 int ret, nodenum;
677 struct dlm_node_iter iter;
678 unsigned int namelen;
679 int tries = 0;
680
681 BUG_ON(!lockid);
682
683 namelen = strlen(lockid);
684
685 mlog(0, "get lockres %s (len %d)\n", lockid, namelen);
686
687lookup:
688 spin_lock(&dlm->spinlock);
689 tmpres = __dlm_lookup_lockres(dlm, lockid, namelen);
690 if (tmpres) {
691 spin_unlock(&dlm->spinlock);
692 mlog(0, "found in hash!\n");
693 if (res)
694 dlm_lockres_put(res);
695 res = tmpres;
696 goto leave;
697 }
698
699 if (!res) {
700 spin_unlock(&dlm->spinlock);
701 mlog(0, "allocating a new resource\n");
702 /* nothing found and we need to allocate one. */
703 alloc_mle = (struct dlm_master_list_entry *)
704 kmem_cache_alloc(dlm_mle_cache, GFP_KERNEL);
705 if (!alloc_mle)
706 goto leave;
707 res = dlm_new_lockres(dlm, lockid, namelen);
708 if (!res)
709 goto leave;
710 goto lookup;
711 }
712
713 mlog(0, "no lockres found, allocated our own: %p\n", res);
714
715 if (flags & LKM_LOCAL) {
716 /* caller knows it's safe to assume it's not mastered elsewhere
717 * DONE! return right away */
718 spin_lock(&res->spinlock);
719 dlm_change_lockres_owner(dlm, res, dlm->node_num);
720 __dlm_insert_lockres(dlm, res);
721 spin_unlock(&res->spinlock);
722 spin_unlock(&dlm->spinlock);
723 /* lockres still marked IN_PROGRESS */
724 goto wake_waiters;
725 }
726
727 /* check master list to see if another node has started mastering it */
728 spin_lock(&dlm->master_lock);
729
730 /* if we found a block, wait for lock to be mastered by another node */
731 blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
732 if (blocked) {
733 if (mle->type == DLM_MLE_MASTER) {
734 mlog(ML_ERROR, "master entry for nonexistent lock!\n");
735 BUG();
736 } else if (mle->type == DLM_MLE_MIGRATION) {
737 /* migration is in progress! */
738 /* the good news is that we now know the
739 * "current" master (mle->master). */
740
741 spin_unlock(&dlm->master_lock);
742 assert_spin_locked(&dlm->spinlock);
743
744 /* set the lockres owner and hash it */
745 spin_lock(&res->spinlock);
746 dlm_set_lockres_owner(dlm, res, mle->master);
747 __dlm_insert_lockres(dlm, res);
748 spin_unlock(&res->spinlock);
749 spin_unlock(&dlm->spinlock);
750
751 /* master is known, detach */
752 dlm_mle_detach_hb_events(dlm, mle);
753 dlm_put_mle(mle);
754 mle = NULL;
755 goto wake_waiters;
756 }
757 } else {
758 /* go ahead and try to master lock on this node */
759 mle = alloc_mle;
760 /* make sure this does not get freed below */
761 alloc_mle = NULL;
762 dlm_init_mle(mle, DLM_MLE_MASTER, dlm, res, NULL, 0);
763 set_bit(dlm->node_num, mle->maybe_map);
764 list_add(&mle->list, &dlm->master_list);
765 }
766
767 /* at this point there is either a DLM_MLE_BLOCK or a
768 * DLM_MLE_MASTER on the master list, so it's safe to add the
769 * lockres to the hashtable. anyone who finds the lock will
770 * still have to wait on the IN_PROGRESS. */
771
772 /* finally add the lockres to its hash bucket */
773 __dlm_insert_lockres(dlm, res);
774 /* get an extra ref on the mle in case this is a BLOCK
775 * if so, the creator of the BLOCK may try to put the last
776 * ref at this time in the assert master handler, so we
777 * need an extra one to keep from a bad ptr deref. */
778 dlm_get_mle(mle);
779 spin_unlock(&dlm->master_lock);
780 spin_unlock(&dlm->spinlock);
781
782 /* must wait for lock to be mastered elsewhere */
783 if (blocked)
784 goto wait;
785
786redo_request:
787 ret = -EINVAL;
788 dlm_node_iter_init(mle->vote_map, &iter);
789 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
790 ret = dlm_do_master_request(mle, nodenum);
791 if (ret < 0)
792 mlog_errno(ret);
793 if (mle->master != O2NM_MAX_NODES) {
794 /* found a master ! */
795 break;
796 }
797 }
798
799wait:
800 /* keep going until the response map includes all nodes */
801 ret = dlm_wait_for_lock_mastery(dlm, res, mle, &blocked);
802 if (ret < 0) {
803 mlog(0, "%s:%.*s: node map changed, redo the "
804 "master request now, blocked=%d\n",
805 dlm->name, res->lockname.len,
806 res->lockname.name, blocked);
807 if (++tries > 20) {
808 mlog(ML_ERROR, "%s:%.*s: spinning on "
809 "dlm_wait_for_lock_mastery, blocked=%d\n",
810 dlm->name, res->lockname.len,
811 res->lockname.name, blocked);
812 dlm_print_one_lock_resource(res);
813 /* dlm_print_one_mle(mle); */
814 tries = 0;
815 }
816 goto redo_request;
817 }
818
819 mlog(0, "lockres mastered by %u\n", res->owner);
820 /* make sure we never continue without this */
821 BUG_ON(res->owner == O2NM_MAX_NODES);
822
823 /* master is known, detach if not already detached */
824 dlm_mle_detach_hb_events(dlm, mle);
825 dlm_put_mle(mle);
826 /* put the extra ref */
827 dlm_put_mle(mle);
828
829wake_waiters:
830 spin_lock(&res->spinlock);
831 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
832 spin_unlock(&res->spinlock);
833 wake_up(&res->wq);
834
835leave:
836 /* need to free the unused mle */
837 if (alloc_mle)
838 kmem_cache_free(dlm_mle_cache, alloc_mle);
839
840 return res;
841}
842
843
844#define DLM_MASTERY_TIMEOUT_MS 5000
845
846static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
847 struct dlm_lock_resource *res,
848 struct dlm_master_list_entry *mle,
849 int *blocked)
850{
851 u8 m;
852 int ret, bit;
853 int map_changed, voting_done;
854 int assert, sleep;
855
856recheck:
857 ret = 0;
858 assert = 0;
859
860 /* check if another node has already become the owner */
861 spin_lock(&res->spinlock);
862 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
863 spin_unlock(&res->spinlock);
864 goto leave;
865 }
866 spin_unlock(&res->spinlock);
867
868 spin_lock(&mle->spinlock);
869 m = mle->master;
870 map_changed = (memcmp(mle->vote_map, mle->node_map,
871 sizeof(mle->vote_map)) != 0);
872 voting_done = (memcmp(mle->vote_map, mle->response_map,
873 sizeof(mle->vote_map)) == 0);
874
875 /* restart if we hit any errors */
876 if (map_changed) {
877 int b;
878 mlog(0, "%s: %.*s: node map changed, restarting\n",
879 dlm->name, res->lockname.len, res->lockname.name);
880 ret = dlm_restart_lock_mastery(dlm, res, mle, *blocked);
881 b = (mle->type == DLM_MLE_BLOCK);
882 if ((*blocked && !b) || (!*blocked && b)) {
883 mlog(0, "%s:%.*s: status change: old=%d new=%d\n",
884 dlm->name, res->lockname.len, res->lockname.name,
885 *blocked, b);
886 *blocked = b;
887 }
888 spin_unlock(&mle->spinlock);
889 if (ret < 0) {
890 mlog_errno(ret);
891 goto leave;
892 }
893 mlog(0, "%s:%.*s: restart lock mastery succeeded, "
894 "rechecking now\n", dlm->name, res->lockname.len,
895 res->lockname.name);
896 goto recheck;
897 }
898
899 if (m != O2NM_MAX_NODES) {
900 /* another node has done an assert!
901 * all done! */
902 sleep = 0;
903 } else {
904 sleep = 1;
905 /* have all nodes responded? */
906 if (voting_done && !*blocked) {
907 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
908 if (dlm->node_num <= bit) {
909 /* my node number is lowest.
910 * now tell other nodes that I am
911 * mastering this. */
912 mle->master = dlm->node_num;
913 assert = 1;
914 sleep = 0;
915 }
916 /* if voting is done, but we have not received
917 * an assert master yet, we must sleep */
918 }
919 }
920
921 spin_unlock(&mle->spinlock);
922
923 /* sleep if we haven't finished voting yet */
924 if (sleep) {
925 unsigned long timeo = msecs_to_jiffies(DLM_MASTERY_TIMEOUT_MS);
926
927 /*
928 if (atomic_read(&mle->mle_refs.refcount) < 2)
929 mlog(ML_ERROR, "mle (%p) refs=%d, name=%.*s\n", mle,
930 atomic_read(&mle->mle_refs.refcount),
931 res->lockname.len, res->lockname.name);
932 */
933 atomic_set(&mle->woken, 0);
934 (void)wait_event_timeout(mle->wq,
935 (atomic_read(&mle->woken) == 1),
936 timeo);
937 if (res->owner == O2NM_MAX_NODES) {
938 mlog(0, "waiting again\n");
939 goto recheck;
940 }
941 mlog(0, "done waiting, master is %u\n", res->owner);
942 ret = 0;
943 goto leave;
944 }
945
946 ret = 0; /* done */
947 if (assert) {
948 m = dlm->node_num;
949 mlog(0, "about to master %.*s here, this=%u\n",
950 res->lockname.len, res->lockname.name, m);
951 ret = dlm_do_assert_master(dlm, res->lockname.name,
952 res->lockname.len, mle->vote_map, 0);
953 if (ret) {
954 /* This is a failure in the network path,
955 * not in the response to the assert_master
956 * (any nonzero response is a BUG on this node).
957 * Most likely a socket just got disconnected
958 * due to node death. */
959 mlog_errno(ret);
960 }
961 /* no longer need to restart lock mastery.
962 * all living nodes have been contacted. */
963 ret = 0;
964 }
965
966 /* set the lockres owner */
967 spin_lock(&res->spinlock);
968 dlm_change_lockres_owner(dlm, res, m);
969 spin_unlock(&res->spinlock);
970
971leave:
972 return ret;
973}
974
975struct dlm_bitmap_diff_iter
976{
977 int curnode;
978 unsigned long *orig_bm;
979 unsigned long *cur_bm;
980 unsigned long diff_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
981};
982
983enum dlm_node_state_change
984{
985 NODE_DOWN = -1,
986 NODE_NO_CHANGE = 0,
987 NODE_UP
988};
989
990static void dlm_bitmap_diff_iter_init(struct dlm_bitmap_diff_iter *iter,
991 unsigned long *orig_bm,
992 unsigned long *cur_bm)
993{
994 unsigned long p1, p2;
995 int i;
996
997 iter->curnode = -1;
998 iter->orig_bm = orig_bm;
999 iter->cur_bm = cur_bm;
1000
1001 for (i = 0; i < BITS_TO_LONGS(O2NM_MAX_NODES); i++) {
1002 p1 = *(iter->orig_bm + i);
1003 p2 = *(iter->cur_bm + i);
1004 iter->diff_bm[i] = (p1 & ~p2) | (p2 & ~p1);
1005 }
1006}
1007
1008static int dlm_bitmap_diff_iter_next(struct dlm_bitmap_diff_iter *iter,
1009 enum dlm_node_state_change *state)
1010{
1011 int bit;
1012
1013 if (iter->curnode >= O2NM_MAX_NODES)
1014 return -ENOENT;
1015
1016 bit = find_next_bit(iter->diff_bm, O2NM_MAX_NODES,
1017 iter->curnode+1);
1018 if (bit >= O2NM_MAX_NODES) {
1019 iter->curnode = O2NM_MAX_NODES;
1020 return -ENOENT;
1021 }
1022
1023 /* if it was there in the original then this node died */
1024 if (test_bit(bit, iter->orig_bm))
1025 *state = NODE_DOWN;
1026 else
1027 *state = NODE_UP;
1028
1029 iter->curnode = bit;
1030 return bit;
1031}
1032
1033
1034static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
1035 struct dlm_lock_resource *res,
1036 struct dlm_master_list_entry *mle,
1037 int blocked)
1038{
1039 struct dlm_bitmap_diff_iter bdi;
1040 enum dlm_node_state_change sc;
1041 int node;
1042 int ret = 0;
1043
1044 mlog(0, "something happened such that the "
1045 "master process may need to be restarted!\n");
1046
1047 assert_spin_locked(&mle->spinlock);
1048
1049 dlm_bitmap_diff_iter_init(&bdi, mle->vote_map, mle->node_map);
1050 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1051 while (node >= 0) {
1052 if (sc == NODE_UP) {
1053 /* a node came up. easy. might not even need
1054 * to talk to it if its node number is higher
1055 * or if we are already blocked. */
1056 mlog(0, "node up! %d\n", node);
1057 if (blocked)
1058 goto next;
1059
1060 if (node > dlm->node_num) {
1061 mlog(0, "node > this node. skipping.\n");
1062 goto next;
1063 }
1064
1065 /* redo the master request, but only for the new node */
1066 mlog(0, "sending request to new node\n");
1067 clear_bit(node, mle->response_map);
1068 set_bit(node, mle->vote_map);
1069 } else {
1070 mlog(ML_ERROR, "node down! %d\n", node);
1071
1072 /* if the node wasn't involved in mastery skip it,
1073 * but clear it out from the maps so that it will
1074 * not affect mastery of this lockres */
1075 clear_bit(node, mle->response_map);
1076 clear_bit(node, mle->vote_map);
1077 if (!test_bit(node, mle->maybe_map))
1078 goto next;
1079
1080 /* if we're already blocked on lock mastery, and the
1081 * dead node wasn't the expected master, or there is
1082 * another node in the maybe_map, keep waiting */
1083 if (blocked) {
1084 int lowest = find_next_bit(mle->maybe_map,
1085 O2NM_MAX_NODES, 0);
1086
1087 /* act like it was never there */
1088 clear_bit(node, mle->maybe_map);
1089
1090 if (node != lowest)
1091 goto next;
1092
1093 mlog(ML_ERROR, "expected master %u died while "
1094 "this node was blocked waiting on it!\n",
1095 node);
1096 lowest = find_next_bit(mle->maybe_map,
1097 O2NM_MAX_NODES,
1098 lowest+1);
1099 if (lowest < O2NM_MAX_NODES) {
1100 mlog(0, "still blocked. waiting "
1101 "on %u now\n", lowest);
1102 goto next;
1103 }
1104
1105 /* mle is an MLE_BLOCK, but there is now
1106 * nothing left to block on. we need to return
1107 * all the way back out and try again with
1108 * an MLE_MASTER. dlm_do_local_recovery_cleanup
1109 * has already run, so the mle refcount is ok */
1110 mlog(0, "no longer blocking. we can "
1111 "try to master this here\n");
1112 mle->type = DLM_MLE_MASTER;
1113 memset(mle->maybe_map, 0,
1114 sizeof(mle->maybe_map));
1115 memset(mle->response_map, 0,
1116 sizeof(mle->maybe_map));
1117 memcpy(mle->vote_map, mle->node_map,
1118 sizeof(mle->node_map));
1119 mle->u.res = res;
1120 set_bit(dlm->node_num, mle->maybe_map);
1121
1122 ret = -EAGAIN;
1123 goto next;
1124 }
1125
1126 clear_bit(node, mle->maybe_map);
1127 if (node > dlm->node_num)
1128 goto next;
1129
1130 mlog(0, "dead node in map!\n");
1131 /* yuck. go back and re-contact all nodes
1132 * in the vote_map, removing this node. */
1133 memset(mle->response_map, 0,
1134 sizeof(mle->response_map));
1135 }
1136 ret = -EAGAIN;
1137next:
1138 node = dlm_bitmap_diff_iter_next(&bdi, &sc);
1139 }
1140 return ret;
1141}
1142
1143
1144/*
1145 * DLM_MASTER_REQUEST_MSG
1146 *
1147 * returns: 0 on success,
1148 * -errno on a network error
1149 *
1150 * on error, the caller should assume the target node is "dead"
1151 *
1152 */
1153
1154static int dlm_do_master_request(struct dlm_master_list_entry *mle, int to)
1155{
1156 struct dlm_ctxt *dlm = mle->dlm;
1157 struct dlm_master_request request;
1158 int ret, response=0, resend;
1159
1160 memset(&request, 0, sizeof(request));
1161 request.node_idx = dlm->node_num;
1162
1163 BUG_ON(mle->type == DLM_MLE_MIGRATION);
1164
1165 if (mle->type != DLM_MLE_MASTER) {
1166 request.namelen = mle->u.name.len;
1167 memcpy(request.name, mle->u.name.name, request.namelen);
1168 } else {
1169 request.namelen = mle->u.res->lockname.len;
1170 memcpy(request.name, mle->u.res->lockname.name,
1171 request.namelen);
1172 }
1173
1174again:
1175 ret = o2net_send_message(DLM_MASTER_REQUEST_MSG, dlm->key, &request,
1176 sizeof(request), to, &response);
1177 if (ret < 0) {
1178 if (ret == -ESRCH) {
1179 /* should never happen */
1180 mlog(ML_ERROR, "TCP stack not ready!\n");
1181 BUG();
1182 } else if (ret == -EINVAL) {
1183 mlog(ML_ERROR, "bad args passed to o2net!\n");
1184 BUG();
1185 } else if (ret == -ENOMEM) {
1186 mlog(ML_ERROR, "out of memory while trying to send "
1187 "network message! retrying\n");
1188 /* this is totally crude */
1189 msleep(50);
1190 goto again;
1191 } else if (!dlm_is_host_down(ret)) {
1192 /* not a network error. bad. */
1193 mlog_errno(ret);
1194 mlog(ML_ERROR, "unhandled error!");
1195 BUG();
1196 }
1197 /* all other errors should be network errors,
1198 * and likely indicate node death */
1199 mlog(ML_ERROR, "link to %d went down!\n", to);
1200 goto out;
1201 }
1202
1203 ret = 0;
1204 resend = 0;
1205 spin_lock(&mle->spinlock);
1206 switch (response) {
1207 case DLM_MASTER_RESP_YES:
1208 set_bit(to, mle->response_map);
1209 mlog(0, "node %u is the master, response=YES\n", to);
1210 mle->master = to;
1211 break;
1212 case DLM_MASTER_RESP_NO:
1213 mlog(0, "node %u not master, response=NO\n", to);
1214 set_bit(to, mle->response_map);
1215 break;
1216 case DLM_MASTER_RESP_MAYBE:
1217 mlog(0, "node %u not master, response=MAYBE\n", to);
1218 set_bit(to, mle->response_map);
1219 set_bit(to, mle->maybe_map);
1220 break;
1221 case DLM_MASTER_RESP_ERROR:
1222 mlog(0, "node %u hit an error, resending\n", to);
1223 resend = 1;
1224 response = 0;
1225 break;
1226 default:
1227 mlog(ML_ERROR, "bad response! %u\n", response);
1228 BUG();
1229 }
1230 spin_unlock(&mle->spinlock);
1231 if (resend) {
1232 /* this is also totally crude */
1233 msleep(50);
1234 goto again;
1235 }
1236
1237out:
1238 return ret;
1239}
1240
1241/*
1242 * locks that can be taken here:
1243 * dlm->spinlock
1244 * res->spinlock
1245 * mle->spinlock
1246 * dlm->master_list
1247 *
1248 * if possible, TRIM THIS DOWN!!!
1249 */
1250int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data)
1251{
1252 u8 response = DLM_MASTER_RESP_MAYBE;
1253 struct dlm_ctxt *dlm = data;
1254 struct dlm_lock_resource *res;
1255 struct dlm_master_request *request = (struct dlm_master_request *) msg->buf;
1256 struct dlm_master_list_entry *mle = NULL, *tmpmle = NULL;
1257 char *name;
1258 unsigned int namelen;
1259 int found, ret;
1260 int set_maybe;
1261
1262 if (!dlm_grab(dlm))
1263 return DLM_MASTER_RESP_NO;
1264
1265 if (!dlm_domain_fully_joined(dlm)) {
1266 response = DLM_MASTER_RESP_NO;
1267 goto send_response;
1268 }
1269
1270 name = request->name;
1271 namelen = request->namelen;
1272
1273 if (namelen > DLM_LOCKID_NAME_MAX) {
1274 response = DLM_IVBUFLEN;
1275 goto send_response;
1276 }
1277
1278way_up_top:
1279 spin_lock(&dlm->spinlock);
1280 res = __dlm_lookup_lockres(dlm, name, namelen);
1281 if (res) {
1282 spin_unlock(&dlm->spinlock);
1283
1284 /* take care of the easy cases up front */
1285 spin_lock(&res->spinlock);
1286 if (res->state & DLM_LOCK_RES_RECOVERING) {
1287 spin_unlock(&res->spinlock);
1288 mlog(0, "returning DLM_MASTER_RESP_ERROR since res is "
1289 "being recovered\n");
1290 response = DLM_MASTER_RESP_ERROR;
1291 if (mle)
1292 kmem_cache_free(dlm_mle_cache, mle);
1293 goto send_response;
1294 }
1295
1296 if (res->owner == dlm->node_num) {
1297 u32 flags = DLM_ASSERT_MASTER_MLE_CLEANUP;
1298 spin_unlock(&res->spinlock);
1299 // mlog(0, "this node is the master\n");
1300 response = DLM_MASTER_RESP_YES;
1301 if (mle)
1302 kmem_cache_free(dlm_mle_cache, mle);
1303
1304 /* this node is the owner.
1305 * there is some extra work that needs to
1306 * happen now. the requesting node has
1307 * caused all nodes up to this one to
1308 * create mles. this node now needs to
1309 * go back and clean those up. */
1310 mlog(0, "%u is the owner of %.*s, cleaning everyone else\n",
1311 dlm->node_num, res->lockname.len, res->lockname.name);
1312 ret = dlm_dispatch_assert_master(dlm, res, 1,
1313 request->node_idx,
1314 flags);
1315 if (ret < 0) {
1316 mlog(ML_ERROR, "failed to dispatch assert "
1317 "master work\n");
1318 response = DLM_MASTER_RESP_ERROR;
1319 }
1320 goto send_response;
1321 } else if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1322 spin_unlock(&res->spinlock);
1323 // mlog(0, "node %u is the master\n", res->owner);
1324 response = DLM_MASTER_RESP_NO;
1325 if (mle)
1326 kmem_cache_free(dlm_mle_cache, mle);
1327 goto send_response;
1328 }
1329
1330 /* ok, there is no owner. either this node is
1331 * being blocked, or it is actively trying to
1332 * master this lock. */
1333 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1334 mlog(ML_ERROR, "lock with no owner should be "
1335 "in-progress!\n");
1336 BUG();
1337 }
1338
1339 // mlog(0, "lockres is in progress...\n");
1340 spin_lock(&dlm->master_lock);
1341 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1342 if (!found) {
1343 mlog(ML_ERROR, "no mle found for this lock!\n");
1344 BUG();
1345 }
1346 set_maybe = 1;
1347 spin_lock(&tmpmle->spinlock);
1348 if (tmpmle->type == DLM_MLE_BLOCK) {
1349 // mlog(0, "this node is waiting for "
1350 // "lockres to be mastered\n");
1351 response = DLM_MASTER_RESP_NO;
1352 } else if (tmpmle->type == DLM_MLE_MIGRATION) {
1353 mlog(0, "node %u is master, but trying to migrate to "
1354 "node %u.\n", tmpmle->master, tmpmle->new_master);
1355 if (tmpmle->master == dlm->node_num) {
1356 response = DLM_MASTER_RESP_YES;
1357 mlog(ML_ERROR, "no owner on lockres, but this "
1358 "node is trying to migrate it to %u?!\n",
1359 tmpmle->new_master);
1360 BUG();
1361 } else {
1362 /* the real master can respond on its own */
1363 response = DLM_MASTER_RESP_NO;
1364 }
1365 } else if (tmpmle->master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1366 set_maybe = 0;
1367 if (tmpmle->master == dlm->node_num)
1368 response = DLM_MASTER_RESP_YES;
1369 else
1370 response = DLM_MASTER_RESP_NO;
1371 } else {
1372 // mlog(0, "this node is attempting to "
1373 // "master lockres\n");
1374 response = DLM_MASTER_RESP_MAYBE;
1375 }
1376 if (set_maybe)
1377 set_bit(request->node_idx, tmpmle->maybe_map);
1378 spin_unlock(&tmpmle->spinlock);
1379
1380 spin_unlock(&dlm->master_lock);
1381 spin_unlock(&res->spinlock);
1382
1383 /* keep the mle attached to heartbeat events */
1384 dlm_put_mle(tmpmle);
1385 if (mle)
1386 kmem_cache_free(dlm_mle_cache, mle);
1387 goto send_response;
1388 }
1389
1390 /*
1391 * lockres doesn't exist on this node
1392 * if there is an MLE_BLOCK, return NO
1393 * if there is an MLE_MASTER, return MAYBE
1394 * otherwise, add an MLE_BLOCK, return NO
1395 */
1396 spin_lock(&dlm->master_lock);
1397 found = dlm_find_mle(dlm, &tmpmle, name, namelen);
1398 if (!found) {
1399 /* this lockid has never been seen on this node yet */
1400 // mlog(0, "no mle found\n");
1401 if (!mle) {
1402 spin_unlock(&dlm->master_lock);
1403 spin_unlock(&dlm->spinlock);
1404
1405 mle = (struct dlm_master_list_entry *)
1406 kmem_cache_alloc(dlm_mle_cache, GFP_KERNEL);
1407 if (!mle) {
1408 // bad bad bad... this sucks.
1409 response = DLM_MASTER_RESP_ERROR;
1410 goto send_response;
1411 }
1412 spin_lock(&dlm->spinlock);
1413 dlm_init_mle(mle, DLM_MLE_BLOCK, dlm, NULL,
1414 name, namelen);
1415 spin_unlock(&dlm->spinlock);
1416 goto way_up_top;
1417 }
1418
1419 // mlog(0, "this is second time thru, already allocated, "
1420 // "add the block.\n");
1421 set_bit(request->node_idx, mle->maybe_map);
1422 list_add(&mle->list, &dlm->master_list);
1423 response = DLM_MASTER_RESP_NO;
1424 } else {
1425 // mlog(0, "mle was found\n");
1426 set_maybe = 1;
1427 spin_lock(&tmpmle->spinlock);
1428 if (tmpmle->type == DLM_MLE_BLOCK)
1429 response = DLM_MASTER_RESP_NO;
1430 else if (tmpmle->type == DLM_MLE_MIGRATION) {
1431 mlog(0, "migration mle was found (%u->%u)\n",
1432 tmpmle->master, tmpmle->new_master);
1433 if (tmpmle->master == dlm->node_num) {
1434 mlog(ML_ERROR, "no lockres, but migration mle "
1435 "says that this node is master!\n");
1436 BUG();
1437 }
1438 /* real master can respond on its own */
1439 response = DLM_MASTER_RESP_NO;
1440 } else {
1441 if (tmpmle->master == dlm->node_num) {
1442 response = DLM_MASTER_RESP_YES;
1443 set_maybe = 0;
1444 } else
1445 response = DLM_MASTER_RESP_MAYBE;
1446 }
1447 if (set_maybe)
1448 set_bit(request->node_idx, tmpmle->maybe_map);
1449 spin_unlock(&tmpmle->spinlock);
1450 }
1451 spin_unlock(&dlm->master_lock);
1452 spin_unlock(&dlm->spinlock);
1453
1454 if (found) {
1455 /* keep the mle attached to heartbeat events */
1456 dlm_put_mle(tmpmle);
1457 }
1458send_response:
1459 dlm_put(dlm);
1460 return response;
1461}
1462
1463/*
1464 * DLM_ASSERT_MASTER_MSG
1465 */
1466
1467
1468/*
1469 * NOTE: this can be used for debugging
1470 * can periodically run all locks owned by this node
1471 * and re-assert across the cluster...
1472 */
1473static int dlm_do_assert_master(struct dlm_ctxt *dlm, const char *lockname,
1474 unsigned int namelen, void *nodemap,
1475 u32 flags)
1476{
1477 struct dlm_assert_master assert;
1478 int to, tmpret;
1479 struct dlm_node_iter iter;
1480 int ret = 0;
1481
1482 BUG_ON(namelen > O2NM_MAX_NAME_LEN);
1483
1484 /* note that if this nodemap is empty, it returns 0 */
1485 dlm_node_iter_init(nodemap, &iter);
1486 while ((to = dlm_node_iter_next(&iter)) >= 0) {
1487 int r = 0;
1488 mlog(0, "sending assert master to %d (%.*s)\n", to,
1489 namelen, lockname);
1490 memset(&assert, 0, sizeof(assert));
1491 assert.node_idx = dlm->node_num;
1492 assert.namelen = namelen;
1493 memcpy(assert.name, lockname, namelen);
1494 assert.flags = cpu_to_be32(flags);
1495
1496 tmpret = o2net_send_message(DLM_ASSERT_MASTER_MSG, dlm->key,
1497 &assert, sizeof(assert), to, &r);
1498 if (tmpret < 0) {
1499 mlog(ML_ERROR, "assert_master returned %d!\n", tmpret);
1500 if (!dlm_is_host_down(tmpret)) {
1501 mlog(ML_ERROR, "unhandled error!\n");
1502 BUG();
1503 }
1504 /* a node died. finish out the rest of the nodes. */
1505 mlog(ML_ERROR, "link to %d went down!\n", to);
1506 /* any nonzero status return will do */
1507 ret = tmpret;
1508 } else if (r < 0) {
1509 /* ok, something horribly messed. kill thyself. */
1510 mlog(ML_ERROR,"during assert master of %.*s to %u, "
1511 "got %d.\n", namelen, lockname, to, r);
1512 dlm_dump_lock_resources(dlm);
1513 BUG();
1514 }
1515 }
1516
1517 return ret;
1518}
1519
1520/*
1521 * locks that can be taken here:
1522 * dlm->spinlock
1523 * res->spinlock
1524 * mle->spinlock
1525 * dlm->master_list
1526 *
1527 * if possible, TRIM THIS DOWN!!!
1528 */
1529int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data)
1530{
1531 struct dlm_ctxt *dlm = data;
1532 struct dlm_master_list_entry *mle = NULL;
1533 struct dlm_assert_master *assert = (struct dlm_assert_master *)msg->buf;
1534 struct dlm_lock_resource *res = NULL;
1535 char *name;
1536 unsigned int namelen;
1537 u32 flags;
1538
1539 if (!dlm_grab(dlm))
1540 return 0;
1541
1542 name = assert->name;
1543 namelen = assert->namelen;
1544 flags = be32_to_cpu(assert->flags);
1545
1546 if (namelen > DLM_LOCKID_NAME_MAX) {
1547 mlog(ML_ERROR, "Invalid name length!");
1548 goto done;
1549 }
1550
1551 spin_lock(&dlm->spinlock);
1552
1553 if (flags)
1554 mlog(0, "assert_master with flags: %u\n", flags);
1555
1556 /* find the MLE */
1557 spin_lock(&dlm->master_lock);
1558 if (!dlm_find_mle(dlm, &mle, name, namelen)) {
1559 /* not an error, could be master just re-asserting */
1560 mlog(0, "just got an assert_master from %u, but no "
1561 "MLE for it! (%.*s)\n", assert->node_idx,
1562 namelen, name);
1563 } else {
1564 int bit = find_next_bit (mle->maybe_map, O2NM_MAX_NODES, 0);
1565 if (bit >= O2NM_MAX_NODES) {
1566 /* not necessarily an error, though less likely.
1567 * could be master just re-asserting. */
1568 mlog(ML_ERROR, "no bits set in the maybe_map, but %u "
1569 "is asserting! (%.*s)\n", assert->node_idx,
1570 namelen, name);
1571 } else if (bit != assert->node_idx) {
1572 if (flags & DLM_ASSERT_MASTER_MLE_CLEANUP) {
1573 mlog(0, "master %u was found, %u should "
1574 "back off\n", assert->node_idx, bit);
1575 } else {
1576 /* with the fix for bug 569, a higher node
1577 * number winning the mastery will respond
1578 * YES to mastery requests, but this node
1579 * had no way of knowing. let it pass. */
1580 mlog(ML_ERROR, "%u is the lowest node, "
1581 "%u is asserting. (%.*s) %u must "
1582 "have begun after %u won.\n", bit,
1583 assert->node_idx, namelen, name, bit,
1584 assert->node_idx);
1585 }
1586 }
1587 }
1588 spin_unlock(&dlm->master_lock);
1589
1590 /* ok everything checks out with the MLE
1591 * now check to see if there is a lockres */
1592 res = __dlm_lookup_lockres(dlm, name, namelen);
1593 if (res) {
1594 spin_lock(&res->spinlock);
1595 if (res->state & DLM_LOCK_RES_RECOVERING) {
1596 mlog(ML_ERROR, "%u asserting but %.*s is "
1597 "RECOVERING!\n", assert->node_idx, namelen, name);
1598 goto kill;
1599 }
1600 if (!mle) {
1601 if (res->owner != assert->node_idx) {
1602 mlog(ML_ERROR, "assert_master from "
1603 "%u, but current owner is "
1604 "%u! (%.*s)\n",
1605 assert->node_idx, res->owner,
1606 namelen, name);
1607 goto kill;
1608 }
1609 } else if (mle->type != DLM_MLE_MIGRATION) {
1610 if (res->owner != DLM_LOCK_RES_OWNER_UNKNOWN) {
1611 /* owner is just re-asserting */
1612 if (res->owner == assert->node_idx) {
1613 mlog(0, "owner %u re-asserting on "
1614 "lock %.*s\n", assert->node_idx,
1615 namelen, name);
1616 goto ok;
1617 }
1618 mlog(ML_ERROR, "got assert_master from "
1619 "node %u, but %u is the owner! "
1620 "(%.*s)\n", assert->node_idx,
1621 res->owner, namelen, name);
1622 goto kill;
1623 }
1624 if (!(res->state & DLM_LOCK_RES_IN_PROGRESS)) {
1625 mlog(ML_ERROR, "got assert from %u, but lock "
1626 "with no owner should be "
1627 "in-progress! (%.*s)\n",
1628 assert->node_idx,
1629 namelen, name);
1630 goto kill;
1631 }
1632 } else /* mle->type == DLM_MLE_MIGRATION */ {
1633 /* should only be getting an assert from new master */
1634 if (assert->node_idx != mle->new_master) {
1635 mlog(ML_ERROR, "got assert from %u, but "
1636 "new master is %u, and old master "
1637 "was %u (%.*s)\n",
1638 assert->node_idx, mle->new_master,
1639 mle->master, namelen, name);
1640 goto kill;
1641 }
1642
1643 }
1644ok:
1645 spin_unlock(&res->spinlock);
1646 }
1647 spin_unlock(&dlm->spinlock);
1648
1649 // mlog(0, "woo! got an assert_master from node %u!\n",
1650 // assert->node_idx);
1651 if (mle) {
1652 int extra_ref;
1653
1654 spin_lock(&mle->spinlock);
1655 extra_ref = !!(mle->type == DLM_MLE_BLOCK
1656 || mle->type == DLM_MLE_MIGRATION);
1657 mle->master = assert->node_idx;
1658 atomic_set(&mle->woken, 1);
1659 wake_up(&mle->wq);
1660 spin_unlock(&mle->spinlock);
1661
1662 if (mle->type == DLM_MLE_MIGRATION && res) {
1663 mlog(0, "finishing off migration of lockres %.*s, "
1664 "from %u to %u\n",
1665 res->lockname.len, res->lockname.name,
1666 dlm->node_num, mle->new_master);
1667 spin_lock(&res->spinlock);
1668 res->state &= ~DLM_LOCK_RES_MIGRATING;
1669 dlm_change_lockres_owner(dlm, res, mle->new_master);
1670 BUG_ON(res->state & DLM_LOCK_RES_DIRTY);
1671 spin_unlock(&res->spinlock);
1672 }
1673 /* master is known, detach if not already detached */
1674 dlm_mle_detach_hb_events(dlm, mle);
1675 dlm_put_mle(mle);
1676
1677 if (extra_ref) {
1678 /* the assert master message now balances the extra
1679 * ref given by the master / migration request message.
1680 * if this is the last put, it will be removed
1681 * from the list. */
1682 dlm_put_mle(mle);
1683 }
1684 }
1685
1686done:
1687 if (res)
1688 dlm_lockres_put(res);
1689 dlm_put(dlm);
1690 return 0;
1691
1692kill:
1693 /* kill the caller! */
1694 spin_unlock(&res->spinlock);
1695 spin_unlock(&dlm->spinlock);
1696 dlm_lockres_put(res);
1697 mlog(ML_ERROR, "Bad message received from another node. Dumping state "
1698 "and killing the other node now! This node is OK and can continue.\n");
1699 dlm_dump_lock_resources(dlm);
1700 dlm_put(dlm);
1701 return -EINVAL;
1702}
1703
1704int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
1705 struct dlm_lock_resource *res,
1706 int ignore_higher, u8 request_from, u32 flags)
1707{
1708 struct dlm_work_item *item;
1709 item = kcalloc(1, sizeof(*item), GFP_KERNEL);
1710 if (!item)
1711 return -ENOMEM;
1712
1713
1714 /* queue up work for dlm_assert_master_worker */
1715 dlm_grab(dlm); /* get an extra ref for the work item */
1716 dlm_init_work_item(dlm, item, dlm_assert_master_worker, NULL);
1717 item->u.am.lockres = res; /* already have a ref */
1718 /* can optionally ignore node numbers higher than this node */
1719 item->u.am.ignore_higher = ignore_higher;
1720 item->u.am.request_from = request_from;
1721 item->u.am.flags = flags;
1722
1723 spin_lock(&dlm->work_lock);
1724 list_add_tail(&item->list, &dlm->work_list);
1725 spin_unlock(&dlm->work_lock);
1726
1727 schedule_work(&dlm->dispatched_work);
1728 return 0;
1729}
1730
1731static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
1732{
1733 struct dlm_ctxt *dlm = data;
1734 int ret = 0;
1735 struct dlm_lock_resource *res;
1736 unsigned long nodemap[BITS_TO_LONGS(O2NM_MAX_NODES)];
1737 int ignore_higher;
1738 int bit;
1739 u8 request_from;
1740 u32 flags;
1741
1742 dlm = item->dlm;
1743 res = item->u.am.lockres;
1744 ignore_higher = item->u.am.ignore_higher;
1745 request_from = item->u.am.request_from;
1746 flags = item->u.am.flags;
1747
1748 spin_lock(&dlm->spinlock);
1749 memcpy(nodemap, dlm->domain_map, sizeof(nodemap));
1750 spin_unlock(&dlm->spinlock);
1751
1752 clear_bit(dlm->node_num, nodemap);
1753 if (ignore_higher) {
1754 /* if is this just to clear up mles for nodes below
1755 * this node, do not send the message to the original
1756 * caller or any node number higher than this */
1757 clear_bit(request_from, nodemap);
1758 bit = dlm->node_num;
1759 while (1) {
1760 bit = find_next_bit(nodemap, O2NM_MAX_NODES,
1761 bit+1);
1762 if (bit >= O2NM_MAX_NODES)
1763 break;
1764 clear_bit(bit, nodemap);
1765 }
1766 }
1767
1768 /* this call now finishes out the nodemap
1769 * even if one or more nodes die */
1770 mlog(0, "worker about to master %.*s here, this=%u\n",
1771 res->lockname.len, res->lockname.name, dlm->node_num);
1772 ret = dlm_do_assert_master(dlm, res->lockname.name,
1773 res->lockname.len,
1774 nodemap, flags);
1775 if (ret < 0) {
1776 /* no need to restart, we are done */
1777 mlog_errno(ret);
1778 }
1779
1780 dlm_lockres_put(res);
1781
1782 mlog(0, "finished with dlm_assert_master_worker\n");
1783}
1784
1785
1786/*
1787 * DLM_MIGRATE_LOCKRES
1788 */
1789
1790
1791int dlm_migrate_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1792 u8 target)
1793{
1794 struct dlm_master_list_entry *mle = NULL;
1795 struct dlm_master_list_entry *oldmle = NULL;
1796 struct dlm_migratable_lockres *mres = NULL;
1797 int ret = -EINVAL;
1798 const char *name;
1799 unsigned int namelen;
1800 int mle_added = 0;
1801 struct list_head *queue, *iter;
1802 int i;
1803 struct dlm_lock *lock;
1804 int empty = 1;
1805
1806 if (!dlm_grab(dlm))
1807 return -EINVAL;
1808
1809 name = res->lockname.name;
1810 namelen = res->lockname.len;
1811
1812 mlog(0, "migrating %.*s to %u\n", namelen, name, target);
1813
1814 /*
1815 * ensure this lockres is a proper candidate for migration
1816 */
1817 spin_lock(&res->spinlock);
1818 if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
1819 mlog(0, "cannot migrate lockres with unknown owner!\n");
1820 spin_unlock(&res->spinlock);
1821 goto leave;
1822 }
1823 if (res->owner != dlm->node_num) {
1824 mlog(0, "cannot migrate lockres this node doesn't own!\n");
1825 spin_unlock(&res->spinlock);
1826 goto leave;
1827 }
1828 mlog(0, "checking queues...\n");
1829 queue = &res->granted;
1830 for (i=0; i<3; i++) {
1831 list_for_each(iter, queue) {
1832 lock = list_entry (iter, struct dlm_lock, list);
1833 empty = 0;
1834 if (lock->ml.node == dlm->node_num) {
1835 mlog(0, "found a lock owned by this node "
1836 "still on the %s queue! will not "
1837 "migrate this lockres\n",
1838 i==0 ? "granted" :
1839 (i==1 ? "converting" : "blocked"));
1840 spin_unlock(&res->spinlock);
1841 ret = -ENOTEMPTY;
1842 goto leave;
1843 }
1844 }
1845 queue++;
1846 }
1847 mlog(0, "all locks on this lockres are nonlocal. continuing\n");
1848 spin_unlock(&res->spinlock);
1849
1850 /* no work to do */
1851 if (empty) {
1852 mlog(0, "no locks were found on this lockres! done!\n");
1853 ret = 0;
1854 goto leave;
1855 }
1856
1857 /*
1858 * preallocate up front
1859 * if this fails, abort
1860 */
1861
1862 ret = -ENOMEM;
1863 mres = (struct dlm_migratable_lockres *) __get_free_page(GFP_KERNEL);
1864 if (!mres) {
1865 mlog_errno(ret);
1866 goto leave;
1867 }
1868
1869 mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
1870 GFP_KERNEL);
1871 if (!mle) {
1872 mlog_errno(ret);
1873 goto leave;
1874 }
1875 ret = 0;
1876
1877 /*
1878 * find a node to migrate the lockres to
1879 */
1880
1881 mlog(0, "picking a migration node\n");
1882 spin_lock(&dlm->spinlock);
1883 /* pick a new node */
1884 if (!test_bit(target, dlm->domain_map) ||
1885 target >= O2NM_MAX_NODES) {
1886 target = dlm_pick_migration_target(dlm, res);
1887 }
1888 mlog(0, "node %u chosen for migration\n", target);
1889
1890 if (target >= O2NM_MAX_NODES ||
1891 !test_bit(target, dlm->domain_map)) {
1892 /* target chosen is not alive */
1893 ret = -EINVAL;
1894 }
1895
1896 if (ret) {
1897 spin_unlock(&dlm->spinlock);
1898 goto fail;
1899 }
1900
1901 mlog(0, "continuing with target = %u\n", target);
1902
1903 /*
1904 * clear any existing master requests and
1905 * add the migration mle to the list
1906 */
1907 spin_lock(&dlm->master_lock);
1908 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
1909 namelen, target, dlm->node_num);
1910 spin_unlock(&dlm->master_lock);
1911 spin_unlock(&dlm->spinlock);
1912
1913 if (ret == -EEXIST) {
1914 mlog(0, "another process is already migrating it\n");
1915 goto fail;
1916 }
1917 mle_added = 1;
1918
1919 /*
1920 * set the MIGRATING flag and flush asts
1921 * if we fail after this we need to re-dirty the lockres
1922 */
1923 if (dlm_mark_lockres_migrating(dlm, res, target) < 0) {
1924 mlog(ML_ERROR, "tried to migrate %.*s to %u, but "
1925 "the target went down.\n", res->lockname.len,
1926 res->lockname.name, target);
1927 spin_lock(&res->spinlock);
1928 res->state &= ~DLM_LOCK_RES_MIGRATING;
1929 spin_unlock(&res->spinlock);
1930 ret = -EINVAL;
1931 }
1932
1933fail:
1934 if (oldmle) {
1935 /* master is known, detach if not already detached */
1936 dlm_mle_detach_hb_events(dlm, oldmle);
1937 dlm_put_mle(oldmle);
1938 }
1939
1940 if (ret < 0) {
1941 if (mle_added) {
1942 dlm_mle_detach_hb_events(dlm, mle);
1943 dlm_put_mle(mle);
1944 } else if (mle) {
1945 kmem_cache_free(dlm_mle_cache, mle);
1946 }
1947 goto leave;
1948 }
1949
1950 /*
1951 * at this point, we have a migration target, an mle
1952 * in the master list, and the MIGRATING flag set on
1953 * the lockres
1954 */
1955
1956
1957 /* get an extra reference on the mle.
1958 * otherwise the assert_master from the new
1959 * master will destroy this.
1960 * also, make sure that all callers of dlm_get_mle
1961 * take both dlm->spinlock and dlm->master_lock */
1962 spin_lock(&dlm->spinlock);
1963 spin_lock(&dlm->master_lock);
1964 dlm_get_mle(mle);
1965 spin_unlock(&dlm->master_lock);
1966 spin_unlock(&dlm->spinlock);
1967
1968 /* notify new node and send all lock state */
1969 /* call send_one_lockres with migration flag.
1970 * this serves as notice to the target node that a
1971 * migration is starting. */
1972 ret = dlm_send_one_lockres(dlm, res, mres, target,
1973 DLM_MRES_MIGRATION);
1974
1975 if (ret < 0) {
1976 mlog(0, "migration to node %u failed with %d\n",
1977 target, ret);
1978 /* migration failed, detach and clean up mle */
1979 dlm_mle_detach_hb_events(dlm, mle);
1980 dlm_put_mle(mle);
1981 dlm_put_mle(mle);
1982 goto leave;
1983 }
1984
1985 /* at this point, the target sends a message to all nodes,
1986 * (using dlm_do_migrate_request). this node is skipped since
1987 * we had to put an mle in the list to begin the process. this
1988 * node now waits for target to do an assert master. this node
1989 * will be the last one notified, ensuring that the migration
1990 * is complete everywhere. if the target dies while this is
1991 * going on, some nodes could potentially see the target as the
1992 * master, so it is important that my recovery finds the migration
1993 * mle and sets the master to UNKNONWN. */
1994
1995
1996 /* wait for new node to assert master */
1997 while (1) {
1998 ret = wait_event_interruptible_timeout(mle->wq,
1999 (atomic_read(&mle->woken) == 1),
2000 msecs_to_jiffies(5000));
2001
2002 if (ret >= 0) {
2003 if (atomic_read(&mle->woken) == 1 ||
2004 res->owner == target)
2005 break;
2006
2007 mlog(0, "timed out during migration\n");
2008 }
2009 if (ret == -ERESTARTSYS) {
2010 /* migration failed, detach and clean up mle */
2011 dlm_mle_detach_hb_events(dlm, mle);
2012 dlm_put_mle(mle);
2013 dlm_put_mle(mle);
2014 goto leave;
2015 }
2016 /* TODO: if node died: stop, clean up, return error */
2017 }
2018
2019 /* all done, set the owner, clear the flag */
2020 spin_lock(&res->spinlock);
2021 dlm_set_lockres_owner(dlm, res, target);
2022 res->state &= ~DLM_LOCK_RES_MIGRATING;
2023 dlm_remove_nonlocal_locks(dlm, res);
2024 spin_unlock(&res->spinlock);
2025 wake_up(&res->wq);
2026
2027 /* master is known, detach if not already detached */
2028 dlm_mle_detach_hb_events(dlm, mle);
2029 dlm_put_mle(mle);
2030 ret = 0;
2031
2032 dlm_lockres_calc_usage(dlm, res);
2033
2034leave:
2035 /* re-dirty the lockres if we failed */
2036 if (ret < 0)
2037 dlm_kick_thread(dlm, res);
2038
2039 /* TODO: cleanup */
2040 if (mres)
2041 free_page((unsigned long)mres);
2042
2043 dlm_put(dlm);
2044
2045 mlog(0, "returning %d\n", ret);
2046 return ret;
2047}
2048EXPORT_SYMBOL_GPL(dlm_migrate_lockres);
2049
2050int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
2051{
2052 int ret;
2053 spin_lock(&dlm->ast_lock);
2054 spin_lock(&lock->spinlock);
2055 ret = (list_empty(&lock->bast_list) && !lock->bast_pending);
2056 spin_unlock(&lock->spinlock);
2057 spin_unlock(&dlm->ast_lock);
2058 return ret;
2059}
2060
2061static int dlm_migration_can_proceed(struct dlm_ctxt *dlm,
2062 struct dlm_lock_resource *res,
2063 u8 mig_target)
2064{
2065 int can_proceed;
2066 spin_lock(&res->spinlock);
2067 can_proceed = !!(res->state & DLM_LOCK_RES_MIGRATING);
2068 spin_unlock(&res->spinlock);
2069
2070 /* target has died, so make the caller break out of the
2071 * wait_event, but caller must recheck the domain_map */
2072 spin_lock(&dlm->spinlock);
2073 if (!test_bit(mig_target, dlm->domain_map))
2074 can_proceed = 1;
2075 spin_unlock(&dlm->spinlock);
2076 return can_proceed;
2077}
2078
2079int dlm_lockres_is_dirty(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
2080{
2081 int ret;
2082 spin_lock(&res->spinlock);
2083 ret = !!(res->state & DLM_LOCK_RES_DIRTY);
2084 spin_unlock(&res->spinlock);
2085 return ret;
2086}
2087
2088
2089static int dlm_mark_lockres_migrating(struct dlm_ctxt *dlm,
2090 struct dlm_lock_resource *res,
2091 u8 target)
2092{
2093 int ret = 0;
2094
2095 mlog(0, "dlm_mark_lockres_migrating: %.*s, from %u to %u\n",
2096 res->lockname.len, res->lockname.name, dlm->node_num,
2097 target);
2098 /* need to set MIGRATING flag on lockres. this is done by
2099 * ensuring that all asts have been flushed for this lockres. */
2100 spin_lock(&res->spinlock);
2101 BUG_ON(res->migration_pending);
2102 res->migration_pending = 1;
2103 /* strategy is to reserve an extra ast then release
2104 * it below, letting the release do all of the work */
2105 __dlm_lockres_reserve_ast(res);
2106 spin_unlock(&res->spinlock);
2107
2108 /* now flush all the pending asts.. hang out for a bit */
2109 dlm_kick_thread(dlm, res);
2110 wait_event(dlm->ast_wq, !dlm_lockres_is_dirty(dlm, res));
2111 dlm_lockres_release_ast(dlm, res);
2112
2113 mlog(0, "about to wait on migration_wq, dirty=%s\n",
2114 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
2115 /* if the extra ref we just put was the final one, this
2116 * will pass thru immediately. otherwise, we need to wait
2117 * for the last ast to finish. */
2118again:
2119 ret = wait_event_interruptible_timeout(dlm->migration_wq,
2120 dlm_migration_can_proceed(dlm, res, target),
2121 msecs_to_jiffies(1000));
2122 if (ret < 0) {
2123 mlog(0, "woken again: migrating? %s, dead? %s\n",
2124 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2125 test_bit(target, dlm->domain_map) ? "no":"yes");
2126 } else {
2127 mlog(0, "all is well: migrating? %s, dead? %s\n",
2128 res->state & DLM_LOCK_RES_MIGRATING ? "yes":"no",
2129 test_bit(target, dlm->domain_map) ? "no":"yes");
2130 }
2131 if (!dlm_migration_can_proceed(dlm, res, target)) {
2132 mlog(0, "trying again...\n");
2133 goto again;
2134 }
2135
2136 /* did the target go down or die? */
2137 spin_lock(&dlm->spinlock);
2138 if (!test_bit(target, dlm->domain_map)) {
2139 mlog(ML_ERROR, "aha. migration target %u just went down\n",
2140 target);
2141 ret = -EHOSTDOWN;
2142 }
2143 spin_unlock(&dlm->spinlock);
2144
2145 /*
2146 * at this point:
2147 *
2148 * o the DLM_LOCK_RES_MIGRATING flag is set
2149 * o there are no pending asts on this lockres
2150 * o all processes trying to reserve an ast on this
2151 * lockres must wait for the MIGRATING flag to clear
2152 */
2153 return ret;
2154}
2155
2156/* last step in the migration process.
2157 * original master calls this to free all of the dlm_lock
2158 * structures that used to be for other nodes. */
2159static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
2160 struct dlm_lock_resource *res)
2161{
2162 struct list_head *iter, *iter2;
2163 struct list_head *queue = &res->granted;
2164 int i;
2165 struct dlm_lock *lock;
2166
2167 assert_spin_locked(&res->spinlock);
2168
2169 BUG_ON(res->owner == dlm->node_num);
2170
2171 for (i=0; i<3; i++) {
2172 list_for_each_safe(iter, iter2, queue) {
2173 lock = list_entry (iter, struct dlm_lock, list);
2174 if (lock->ml.node != dlm->node_num) {
2175 mlog(0, "putting lock for node %u\n",
2176 lock->ml.node);
2177 /* be extra careful */
2178 BUG_ON(!list_empty(&lock->ast_list));
2179 BUG_ON(!list_empty(&lock->bast_list));
2180 BUG_ON(lock->ast_pending);
2181 BUG_ON(lock->bast_pending);
2182 list_del_init(&lock->list);
2183 dlm_lock_put(lock);
2184 }
2185 }
2186 queue++;
2187 }
2188}
2189
2190/* for now this is not too intelligent. we will
2191 * need stats to make this do the right thing.
2192 * this just finds the first lock on one of the
2193 * queues and uses that node as the target. */
2194static u8 dlm_pick_migration_target(struct dlm_ctxt *dlm,
2195 struct dlm_lock_resource *res)
2196{
2197 int i;
2198 struct list_head *queue = &res->granted;
2199 struct list_head *iter;
2200 struct dlm_lock *lock;
2201 int nodenum;
2202
2203 assert_spin_locked(&dlm->spinlock);
2204
2205 spin_lock(&res->spinlock);
2206 for (i=0; i<3; i++) {
2207 list_for_each(iter, queue) {
2208 /* up to the caller to make sure this node
2209 * is alive */
2210 lock = list_entry (iter, struct dlm_lock, list);
2211 if (lock->ml.node != dlm->node_num) {
2212 spin_unlock(&res->spinlock);
2213 return lock->ml.node;
2214 }
2215 }
2216 queue++;
2217 }
2218 spin_unlock(&res->spinlock);
2219 mlog(0, "have not found a suitable target yet! checking domain map\n");
2220
2221 /* ok now we're getting desperate. pick anyone alive. */
2222 nodenum = -1;
2223 while (1) {
2224 nodenum = find_next_bit(dlm->domain_map,
2225 O2NM_MAX_NODES, nodenum+1);
2226 mlog(0, "found %d in domain map\n", nodenum);
2227 if (nodenum >= O2NM_MAX_NODES)
2228 break;
2229 if (nodenum != dlm->node_num) {
2230 mlog(0, "picking %d\n", nodenum);
2231 return nodenum;
2232 }
2233 }
2234
2235 mlog(0, "giving up. no master to migrate to\n");
2236 return DLM_LOCK_RES_OWNER_UNKNOWN;
2237}
2238
2239
2240
2241/* this is called by the new master once all lockres
2242 * data has been received */
2243static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
2244 struct dlm_lock_resource *res,
2245 u8 master, u8 new_master,
2246 struct dlm_node_iter *iter)
2247{
2248 struct dlm_migrate_request migrate;
2249 int ret, status = 0;
2250 int nodenum;
2251
2252 memset(&migrate, 0, sizeof(migrate));
2253 migrate.namelen = res->lockname.len;
2254 memcpy(migrate.name, res->lockname.name, migrate.namelen);
2255 migrate.new_master = new_master;
2256 migrate.master = master;
2257
2258 ret = 0;
2259
2260 /* send message to all nodes, except the master and myself */
2261 while ((nodenum = dlm_node_iter_next(iter)) >= 0) {
2262 if (nodenum == master ||
2263 nodenum == new_master)
2264 continue;
2265
2266 ret = o2net_send_message(DLM_MIGRATE_REQUEST_MSG, dlm->key,
2267 &migrate, sizeof(migrate), nodenum,
2268 &status);
2269 if (ret < 0)
2270 mlog_errno(ret);
2271 else if (status < 0) {
2272 mlog(0, "migrate request (node %u) returned %d!\n",
2273 nodenum, status);
2274 ret = status;
2275 }
2276 }
2277
2278 if (ret < 0)
2279 mlog_errno(ret);
2280
2281 mlog(0, "returning ret=%d\n", ret);
2282 return ret;
2283}
2284
2285
2286/* if there is an existing mle for this lockres, we now know who the master is.
2287 * (the one who sent us *this* message) we can clear it up right away.
2288 * since the process that put the mle on the list still has a reference to it,
2289 * we can unhash it now, set the master and wake the process. as a result,
2290 * we will have no mle in the list to start with. now we can add an mle for
2291 * the migration and this should be the only one found for those scanning the
2292 * list. */
2293int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data)
2294{
2295 struct dlm_ctxt *dlm = data;
2296 struct dlm_lock_resource *res = NULL;
2297 struct dlm_migrate_request *migrate = (struct dlm_migrate_request *) msg->buf;
2298 struct dlm_master_list_entry *mle = NULL, *oldmle = NULL;
2299 const char *name;
2300 unsigned int namelen;
2301 int ret = 0;
2302
2303 if (!dlm_grab(dlm))
2304 return -EINVAL;
2305
2306 name = migrate->name;
2307 namelen = migrate->namelen;
2308
2309 /* preallocate.. if this fails, abort */
2310 mle = (struct dlm_master_list_entry *) kmem_cache_alloc(dlm_mle_cache,
2311 GFP_KERNEL);
2312
2313 if (!mle) {
2314 ret = -ENOMEM;
2315 goto leave;
2316 }
2317
2318 /* check for pre-existing lock */
2319 spin_lock(&dlm->spinlock);
2320 res = __dlm_lookup_lockres(dlm, name, namelen);
2321 spin_lock(&dlm->master_lock);
2322
2323 if (res) {
2324 spin_lock(&res->spinlock);
2325 if (res->state & DLM_LOCK_RES_RECOVERING) {
2326 /* if all is working ok, this can only mean that we got
2327 * a migrate request from a node that we now see as
2328 * dead. what can we do here? drop it to the floor? */
2329 spin_unlock(&res->spinlock);
2330 mlog(ML_ERROR, "Got a migrate request, but the "
2331 "lockres is marked as recovering!");
2332 kmem_cache_free(dlm_mle_cache, mle);
2333 ret = -EINVAL; /* need a better solution */
2334 goto unlock;
2335 }
2336 res->state |= DLM_LOCK_RES_MIGRATING;
2337 spin_unlock(&res->spinlock);
2338 }
2339
2340 /* ignore status. only nonzero status would BUG. */
2341 ret = dlm_add_migration_mle(dlm, res, mle, &oldmle,
2342 name, namelen,
2343 migrate->new_master,
2344 migrate->master);
2345
2346unlock:
2347 spin_unlock(&dlm->master_lock);
2348 spin_unlock(&dlm->spinlock);
2349
2350 if (oldmle) {
2351 /* master is known, detach if not already detached */
2352 dlm_mle_detach_hb_events(dlm, oldmle);
2353 dlm_put_mle(oldmle);
2354 }
2355
2356 if (res)
2357 dlm_lockres_put(res);
2358leave:
2359 dlm_put(dlm);
2360 return ret;
2361}
2362
2363/* must be holding dlm->spinlock and dlm->master_lock
2364 * when adding a migration mle, we can clear any other mles
2365 * in the master list because we know with certainty that
2366 * the master is "master". so we remove any old mle from
2367 * the list after setting it's master field, and then add
2368 * the new migration mle. this way we can hold with the rule
2369 * of having only one mle for a given lock name at all times. */
2370static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
2371 struct dlm_lock_resource *res,
2372 struct dlm_master_list_entry *mle,
2373 struct dlm_master_list_entry **oldmle,
2374 const char *name, unsigned int namelen,
2375 u8 new_master, u8 master)
2376{
2377 int found;
2378 int ret = 0;
2379
2380 *oldmle = NULL;
2381
2382 mlog_entry_void();
2383
2384 assert_spin_locked(&dlm->spinlock);
2385 assert_spin_locked(&dlm->master_lock);
2386
2387 /* caller is responsible for any ref taken here on oldmle */
2388 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
2389 if (found) {
2390 struct dlm_master_list_entry *tmp = *oldmle;
2391 spin_lock(&tmp->spinlock);
2392 if (tmp->type == DLM_MLE_MIGRATION) {
2393 if (master == dlm->node_num) {
2394 /* ah another process raced me to it */
2395 mlog(0, "tried to migrate %.*s, but some "
2396 "process beat me to it\n",
2397 namelen, name);
2398 ret = -EEXIST;
2399 } else {
2400 /* bad. 2 NODES are trying to migrate! */
2401 mlog(ML_ERROR, "migration error mle: "
2402 "master=%u new_master=%u // request: "
2403 "master=%u new_master=%u // "
2404 "lockres=%.*s\n",
2405 tmp->master, tmp->new_master,
2406 master, new_master,
2407 namelen, name);
2408 BUG();
2409 }
2410 } else {
2411 /* this is essentially what assert_master does */
2412 tmp->master = master;
2413 atomic_set(&tmp->woken, 1);
2414 wake_up(&tmp->wq);
2415 /* remove it from the list so that only one
2416 * mle will be found */
2417 list_del_init(&tmp->list);
2418 }
2419 spin_unlock(&tmp->spinlock);
2420 }
2421
2422 /* now add a migration mle to the tail of the list */
2423 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
2424 mle->new_master = new_master;
2425 mle->master = master;
2426 /* do this for consistency with other mle types */
2427 set_bit(new_master, mle->maybe_map);
2428 list_add(&mle->list, &dlm->master_list);
2429
2430 return ret;
2431}
2432
2433
2434void dlm_clean_master_list(struct dlm_ctxt *dlm, u8 dead_node)
2435{
2436 struct list_head *iter, *iter2;
2437 struct dlm_master_list_entry *mle;
2438 struct dlm_lock_resource *res;
2439
2440 mlog_entry("dlm=%s, dead node=%u\n", dlm->name, dead_node);
2441top:
2442 assert_spin_locked(&dlm->spinlock);
2443
2444 /* clean the master list */
2445 spin_lock(&dlm->master_lock);
2446 list_for_each_safe(iter, iter2, &dlm->master_list) {
2447 mle = list_entry(iter, struct dlm_master_list_entry, list);
2448
2449 BUG_ON(mle->type != DLM_MLE_BLOCK &&
2450 mle->type != DLM_MLE_MASTER &&
2451 mle->type != DLM_MLE_MIGRATION);
2452
2453 /* MASTER mles are initiated locally. the waiting
2454 * process will notice the node map change
2455 * shortly. let that happen as normal. */
2456 if (mle->type == DLM_MLE_MASTER)
2457 continue;
2458
2459
2460 /* BLOCK mles are initiated by other nodes.
2461 * need to clean up if the dead node would have
2462 * been the master. */
2463 if (mle->type == DLM_MLE_BLOCK) {
2464 int bit;
2465
2466 spin_lock(&mle->spinlock);
2467 bit = find_next_bit(mle->maybe_map, O2NM_MAX_NODES, 0);
2468 if (bit != dead_node) {
2469 mlog(0, "mle found, but dead node %u would "
2470 "not have been master\n", dead_node);
2471 spin_unlock(&mle->spinlock);
2472 } else {
2473 /* must drop the refcount by one since the
2474 * assert_master will never arrive. this
2475 * may result in the mle being unlinked and
2476 * freed, but there may still be a process
2477 * waiting in the dlmlock path which is fine. */
2478 mlog(ML_ERROR, "node %u was expected master\n",
2479 dead_node);
2480 atomic_set(&mle->woken, 1);
2481 spin_unlock(&mle->spinlock);
2482 wake_up(&mle->wq);
2483 /* final put will take care of list removal */
2484 __dlm_put_mle(mle);
2485 }
2486 continue;
2487 }
2488
2489 /* everything else is a MIGRATION mle */
2490
2491 /* the rule for MIGRATION mles is that the master
2492 * becomes UNKNOWN if *either* the original or
2493 * the new master dies. all UNKNOWN lockreses
2494 * are sent to whichever node becomes the recovery
2495 * master. the new master is responsible for
2496 * determining if there is still a master for
2497 * this lockres, or if he needs to take over
2498 * mastery. either way, this node should expect
2499 * another message to resolve this. */
2500 if (mle->master != dead_node &&
2501 mle->new_master != dead_node)
2502 continue;
2503
2504 /* if we have reached this point, this mle needs to
2505 * be removed from the list and freed. */
2506
2507 /* remove from the list early. NOTE: unlinking
2508 * list_head while in list_for_each_safe */
2509 spin_lock(&mle->spinlock);
2510 list_del_init(&mle->list);
2511 atomic_set(&mle->woken, 1);
2512 spin_unlock(&mle->spinlock);
2513 wake_up(&mle->wq);
2514
2515 mlog(0, "node %u died during migration from "
2516 "%u to %u!\n", dead_node,
2517 mle->master, mle->new_master);
2518 /* if there is a lockres associated with this
2519 * mle, find it and set its owner to UNKNOWN */
2520 res = __dlm_lookup_lockres(dlm, mle->u.name.name,
2521 mle->u.name.len);
2522 if (res) {
2523 /* unfortunately if we hit this rare case, our
2524 * lock ordering is messed. we need to drop
2525 * the master lock so that we can take the
2526 * lockres lock, meaning that we will have to
2527 * restart from the head of list. */
2528 spin_unlock(&dlm->master_lock);
2529
2530 /* move lockres onto recovery list */
2531 spin_lock(&res->spinlock);
2532 dlm_set_lockres_owner(dlm, res,
2533 DLM_LOCK_RES_OWNER_UNKNOWN);
2534 dlm_move_lockres_to_recovery_list(dlm, res);
2535 spin_unlock(&res->spinlock);
2536 dlm_lockres_put(res);
2537
2538 /* dump the mle */
2539 spin_lock(&dlm->master_lock);
2540 __dlm_put_mle(mle);
2541 spin_unlock(&dlm->master_lock);
2542
2543 /* restart */
2544 goto top;
2545 }
2546
2547 /* this may be the last reference */
2548 __dlm_put_mle(mle);
2549 }
2550 spin_unlock(&dlm->master_lock);
2551}
2552
2553
2554int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
2555 u8 old_master)
2556{
2557 struct dlm_node_iter iter;
2558 int ret = 0;
2559
2560 spin_lock(&dlm->spinlock);
2561 dlm_node_iter_init(dlm->domain_map, &iter);
2562 clear_bit(old_master, iter.node_map);
2563 clear_bit(dlm->node_num, iter.node_map);
2564 spin_unlock(&dlm->spinlock);
2565
2566 mlog(0, "now time to do a migrate request to other nodes\n");
2567 ret = dlm_do_migrate_request(dlm, res, old_master,
2568 dlm->node_num, &iter);
2569 if (ret < 0) {
2570 mlog_errno(ret);
2571 goto leave;
2572 }
2573
2574 mlog(0, "doing assert master of %.*s to all except the original node\n",
2575 res->lockname.len, res->lockname.name);
2576 /* this call now finishes out the nodemap
2577 * even if one or more nodes die */
2578 ret = dlm_do_assert_master(dlm, res->lockname.name,
2579 res->lockname.len, iter.node_map,
2580 DLM_ASSERT_MASTER_FINISH_MIGRATION);
2581 if (ret < 0) {
2582 /* no longer need to retry. all living nodes contacted. */
2583 mlog_errno(ret);
2584 ret = 0;
2585 }
2586
2587 memset(iter.node_map, 0, sizeof(iter.node_map));
2588 set_bit(old_master, iter.node_map);
2589 mlog(0, "doing assert master of %.*s back to %u\n",
2590 res->lockname.len, res->lockname.name, old_master);
2591 ret = dlm_do_assert_master(dlm, res->lockname.name,
2592 res->lockname.len, iter.node_map,
2593 DLM_ASSERT_MASTER_FINISH_MIGRATION);
2594 if (ret < 0) {
2595 mlog(0, "assert master to original master failed "
2596 "with %d.\n", ret);
2597 /* the only nonzero status here would be because of
2598 * a dead original node. we're done. */
2599 ret = 0;
2600 }
2601
2602 /* all done, set the owner, clear the flag */
2603 spin_lock(&res->spinlock);
2604 dlm_set_lockres_owner(dlm, res, dlm->node_num);
2605 res->state &= ~DLM_LOCK_RES_MIGRATING;
2606 spin_unlock(&res->spinlock);
2607 /* re-dirty it on the new master */
2608 dlm_kick_thread(dlm, res);
2609 wake_up(&res->wq);
2610leave:
2611 return ret;
2612}
2613
2614/*
2615 * LOCKRES AST REFCOUNT
2616 * this is integral to migration
2617 */
2618
2619/* for future intent to call an ast, reserve one ahead of time.
2620 * this should be called only after waiting on the lockres
2621 * with dlm_wait_on_lockres, and while still holding the
2622 * spinlock after the call. */
2623void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res)
2624{
2625 assert_spin_locked(&res->spinlock);
2626 if (res->state & DLM_LOCK_RES_MIGRATING) {
2627 __dlm_print_one_lock_resource(res);
2628 }
2629 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
2630
2631 atomic_inc(&res->asts_reserved);
2632}
2633
2634/*
2635 * used to drop the reserved ast, either because it went unused,
2636 * or because the ast/bast was actually called.
2637 *
2638 * also, if there is a pending migration on this lockres,
2639 * and this was the last pending ast on the lockres,
2640 * atomically set the MIGRATING flag before we drop the lock.
2641 * this is how we ensure that migration can proceed with no
2642 * asts in progress. note that it is ok if the state of the
2643 * queues is such that a lock should be granted in the future
2644 * or that a bast should be fired, because the new master will
2645 * shuffle the lists on this lockres as soon as it is migrated.
2646 */
2647void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
2648 struct dlm_lock_resource *res)
2649{
2650 if (!atomic_dec_and_lock(&res->asts_reserved, &res->spinlock))
2651 return;
2652
2653 if (!res->migration_pending) {
2654 spin_unlock(&res->spinlock);
2655 return;
2656 }
2657
2658 BUG_ON(res->state & DLM_LOCK_RES_MIGRATING);
2659 res->migration_pending = 0;
2660 res->state |= DLM_LOCK_RES_MIGRATING;
2661 spin_unlock(&res->spinlock);
2662 wake_up(&res->wq);
2663 wake_up(&dlm->migration_wq);
2664}
diff --git a/fs/ocfs2/dlm/dlmrecovery.c b/fs/ocfs2/dlm/dlmrecovery.c
new file mode 100644
index 000000000000..0c8eb1093f00
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmrecovery.c
@@ -0,0 +1,2132 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmrecovery.c
5 *
6 * recovery stuff
7 *
8 * Copyright (C) 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 *
25 */
26
27
28#include <linux/module.h>
29#include <linux/fs.h>
30#include <linux/types.h>
31#include <linux/slab.h>
32#include <linux/highmem.h>
33#include <linux/utsname.h>
34#include <linux/init.h>
35#include <linux/sysctl.h>
36#include <linux/random.h>
37#include <linux/blkdev.h>
38#include <linux/socket.h>
39#include <linux/inet.h>
40#include <linux/timer.h>
41#include <linux/kthread.h>
42
43
44#include "cluster/heartbeat.h"
45#include "cluster/nodemanager.h"
46#include "cluster/tcp.h"
47
48#include "dlmapi.h"
49#include "dlmcommon.h"
50#include "dlmdomain.h"
51
52#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_RECOVERY)
53#include "cluster/masklog.h"
54
55static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node);
56
57static int dlm_recovery_thread(void *data);
58void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
59int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
60static void dlm_kick_recovery_thread(struct dlm_ctxt *dlm);
61static int dlm_do_recovery(struct dlm_ctxt *dlm);
62
63static int dlm_pick_recovery_master(struct dlm_ctxt *dlm);
64static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node);
65static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
66static int dlm_request_all_locks(struct dlm_ctxt *dlm,
67 u8 request_from, u8 dead_node);
68static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node);
69
70static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res);
71static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
72 const char *lockname, int namelen,
73 int total_locks, u64 cookie,
74 u8 flags, u8 master);
75static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
76 struct dlm_migratable_lockres *mres,
77 u8 send_to,
78 struct dlm_lock_resource *res,
79 int total_locks);
80static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
81 struct dlm_lock_resource *res,
82 u8 *real_master);
83static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
84 struct dlm_lock_resource *res,
85 struct dlm_migratable_lockres *mres);
86static int dlm_do_master_requery(struct dlm_ctxt *dlm,
87 struct dlm_lock_resource *res,
88 u8 nodenum, u8 *real_master);
89static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm);
90static int dlm_send_all_done_msg(struct dlm_ctxt *dlm,
91 u8 dead_node, u8 send_to);
92static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node);
93static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
94 struct list_head *list, u8 dead_node);
95static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
96 u8 dead_node, u8 new_master);
97static void dlm_reco_ast(void *astdata);
98static void dlm_reco_bast(void *astdata, int blocked_type);
99static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st);
100static void dlm_request_all_locks_worker(struct dlm_work_item *item,
101 void *data);
102static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data);
103
104static u64 dlm_get_next_mig_cookie(void);
105
106static spinlock_t dlm_reco_state_lock = SPIN_LOCK_UNLOCKED;
107static spinlock_t dlm_mig_cookie_lock = SPIN_LOCK_UNLOCKED;
108static u64 dlm_mig_cookie = 1;
109
110static u64 dlm_get_next_mig_cookie(void)
111{
112 u64 c;
113 spin_lock(&dlm_mig_cookie_lock);
114 c = dlm_mig_cookie;
115 if (dlm_mig_cookie == (~0ULL))
116 dlm_mig_cookie = 1;
117 else
118 dlm_mig_cookie++;
119 spin_unlock(&dlm_mig_cookie_lock);
120 return c;
121}
122
123static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
124{
125 spin_lock(&dlm->spinlock);
126 clear_bit(dlm->reco.dead_node, dlm->recovery_map);
127 dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
128 dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
129 spin_unlock(&dlm->spinlock);
130}
131
132/* Worker function used during recovery. */
133void dlm_dispatch_work(void *data)
134{
135 struct dlm_ctxt *dlm = (struct dlm_ctxt *)data;
136 LIST_HEAD(tmp_list);
137 struct list_head *iter, *iter2;
138 struct dlm_work_item *item;
139 dlm_workfunc_t *workfunc;
140
141 spin_lock(&dlm->work_lock);
142 list_splice_init(&dlm->work_list, &tmp_list);
143 spin_unlock(&dlm->work_lock);
144
145 list_for_each_safe(iter, iter2, &tmp_list) {
146 item = list_entry(iter, struct dlm_work_item, list);
147 workfunc = item->func;
148 list_del_init(&item->list);
149
150 /* already have ref on dlm to avoid having
151 * it disappear. just double-check. */
152 BUG_ON(item->dlm != dlm);
153
154 /* this is allowed to sleep and
155 * call network stuff */
156 workfunc(item, item->data);
157
158 dlm_put(dlm);
159 kfree(item);
160 }
161}
162
163/*
164 * RECOVERY THREAD
165 */
166
167static void dlm_kick_recovery_thread(struct dlm_ctxt *dlm)
168{
169 /* wake the recovery thread
170 * this will wake the reco thread in one of three places
171 * 1) sleeping with no recovery happening
172 * 2) sleeping with recovery mastered elsewhere
173 * 3) recovery mastered here, waiting on reco data */
174
175 wake_up(&dlm->dlm_reco_thread_wq);
176}
177
178/* Launch the recovery thread */
179int dlm_launch_recovery_thread(struct dlm_ctxt *dlm)
180{
181 mlog(0, "starting dlm recovery thread...\n");
182
183 dlm->dlm_reco_thread_task = kthread_run(dlm_recovery_thread, dlm,
184 "dlm_reco_thread");
185 if (IS_ERR(dlm->dlm_reco_thread_task)) {
186 mlog_errno(PTR_ERR(dlm->dlm_reco_thread_task));
187 dlm->dlm_reco_thread_task = NULL;
188 return -EINVAL;
189 }
190
191 return 0;
192}
193
194void dlm_complete_recovery_thread(struct dlm_ctxt *dlm)
195{
196 if (dlm->dlm_reco_thread_task) {
197 mlog(0, "waiting for dlm recovery thread to exit\n");
198 kthread_stop(dlm->dlm_reco_thread_task);
199 dlm->dlm_reco_thread_task = NULL;
200 }
201}
202
203
204
205/*
206 * this is lame, but here's how recovery works...
207 * 1) all recovery threads cluster wide will work on recovering
208 * ONE node at a time
209 * 2) negotiate who will take over all the locks for the dead node.
210 * thats right... ALL the locks.
211 * 3) once a new master is chosen, everyone scans all locks
212 * and moves aside those mastered by the dead guy
213 * 4) each of these locks should be locked until recovery is done
214 * 5) the new master collects up all of secondary lock queue info
215 * one lock at a time, forcing each node to communicate back
216 * before continuing
217 * 6) each secondary lock queue responds with the full known lock info
218 * 7) once the new master has run all its locks, it sends a ALLDONE!
219 * message to everyone
220 * 8) upon receiving this message, the secondary queue node unlocks
221 * and responds to the ALLDONE
222 * 9) once the new master gets responses from everyone, he unlocks
223 * everything and recovery for this dead node is done
224 *10) go back to 2) while there are still dead nodes
225 *
226 */
227
228
229#define DLM_RECO_THREAD_TIMEOUT_MS (5 * 1000)
230
231static int dlm_recovery_thread(void *data)
232{
233 int status;
234 struct dlm_ctxt *dlm = data;
235 unsigned long timeout = msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS);
236
237 mlog(0, "dlm thread running for %s...\n", dlm->name);
238
239 while (!kthread_should_stop()) {
240 if (dlm_joined(dlm)) {
241 status = dlm_do_recovery(dlm);
242 if (status == -EAGAIN) {
243 /* do not sleep, recheck immediately. */
244 continue;
245 }
246 if (status < 0)
247 mlog_errno(status);
248 }
249
250 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
251 kthread_should_stop(),
252 timeout);
253 }
254
255 mlog(0, "quitting DLM recovery thread\n");
256 return 0;
257}
258
259/* callers of the top-level api calls (dlmlock/dlmunlock) should
260 * block on the dlm->reco.event when recovery is in progress.
261 * the dlm recovery thread will set this state when it begins
262 * recovering a dead node (as the new master or not) and clear
263 * the state and wake as soon as all affected lock resources have
264 * been marked with the RECOVERY flag */
265static int dlm_in_recovery(struct dlm_ctxt *dlm)
266{
267 int in_recovery;
268 spin_lock(&dlm->spinlock);
269 in_recovery = !!(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
270 spin_unlock(&dlm->spinlock);
271 return in_recovery;
272}
273
274
275void dlm_wait_for_recovery(struct dlm_ctxt *dlm)
276{
277 wait_event(dlm->reco.event, !dlm_in_recovery(dlm));
278}
279
280static void dlm_begin_recovery(struct dlm_ctxt *dlm)
281{
282 spin_lock(&dlm->spinlock);
283 BUG_ON(dlm->reco.state & DLM_RECO_STATE_ACTIVE);
284 dlm->reco.state |= DLM_RECO_STATE_ACTIVE;
285 spin_unlock(&dlm->spinlock);
286}
287
288static void dlm_end_recovery(struct dlm_ctxt *dlm)
289{
290 spin_lock(&dlm->spinlock);
291 BUG_ON(!(dlm->reco.state & DLM_RECO_STATE_ACTIVE));
292 dlm->reco.state &= ~DLM_RECO_STATE_ACTIVE;
293 spin_unlock(&dlm->spinlock);
294 wake_up(&dlm->reco.event);
295}
296
297static int dlm_do_recovery(struct dlm_ctxt *dlm)
298{
299 int status = 0;
300
301 spin_lock(&dlm->spinlock);
302
303 /* check to see if the new master has died */
304 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM &&
305 test_bit(dlm->reco.new_master, dlm->recovery_map)) {
306 mlog(0, "new master %u died while recovering %u!\n",
307 dlm->reco.new_master, dlm->reco.dead_node);
308 /* unset the new_master, leave dead_node */
309 dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
310 }
311
312 /* select a target to recover */
313 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
314 int bit;
315
316 bit = find_next_bit (dlm->recovery_map, O2NM_MAX_NODES+1, 0);
317 if (bit >= O2NM_MAX_NODES || bit < 0)
318 dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
319 else
320 dlm->reco.dead_node = bit;
321 } else if (!test_bit(dlm->reco.dead_node, dlm->recovery_map)) {
322 /* BUG? */
323 mlog(ML_ERROR, "dead_node %u no longer in recovery map!\n",
324 dlm->reco.dead_node);
325 dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
326 }
327
328 if (dlm->reco.dead_node == O2NM_INVALID_NODE_NUM) {
329 // mlog(0, "nothing to recover! sleeping now!\n");
330 spin_unlock(&dlm->spinlock);
331 /* return to main thread loop and sleep. */
332 return 0;
333 }
334 mlog(0, "recovery thread found node %u in the recovery map!\n",
335 dlm->reco.dead_node);
336 spin_unlock(&dlm->spinlock);
337
338 /* take write barrier */
339 /* (stops the list reshuffling thread, proxy ast handling) */
340 dlm_begin_recovery(dlm);
341
342 if (dlm->reco.new_master == dlm->node_num)
343 goto master_here;
344
345 if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
346 /* choose a new master */
347 if (!dlm_pick_recovery_master(dlm)) {
348 /* already notified everyone. go. */
349 dlm->reco.new_master = dlm->node_num;
350 goto master_here;
351 }
352 mlog(0, "another node will master this recovery session.\n");
353 }
354 mlog(0, "dlm=%s, new_master=%u, this node=%u, dead_node=%u\n",
355 dlm->name, dlm->reco.new_master,
356 dlm->node_num, dlm->reco.dead_node);
357
358 /* it is safe to start everything back up here
359 * because all of the dead node's lock resources
360 * have been marked as in-recovery */
361 dlm_end_recovery(dlm);
362
363 /* sleep out in main dlm_recovery_thread loop. */
364 return 0;
365
366master_here:
367 mlog(0, "mastering recovery of %s:%u here(this=%u)!\n",
368 dlm->name, dlm->reco.dead_node, dlm->node_num);
369
370 status = dlm_remaster_locks(dlm, dlm->reco.dead_node);
371 if (status < 0) {
372 mlog(ML_ERROR, "error %d remastering locks for node %u, "
373 "retrying.\n", status, dlm->reco.dead_node);
374 } else {
375 /* success! see if any other nodes need recovery */
376 dlm_reset_recovery(dlm);
377 }
378 dlm_end_recovery(dlm);
379
380 /* continue and look for another dead node */
381 return -EAGAIN;
382}
383
384static int dlm_remaster_locks(struct dlm_ctxt *dlm, u8 dead_node)
385{
386 int status = 0;
387 struct dlm_reco_node_data *ndata;
388 struct list_head *iter;
389 int all_nodes_done;
390 int destroy = 0;
391 int pass = 0;
392
393 status = dlm_init_recovery_area(dlm, dead_node);
394 if (status < 0)
395 goto leave;
396
397 /* safe to access the node data list without a lock, since this
398 * process is the only one to change the list */
399 list_for_each(iter, &dlm->reco.node_data) {
400 ndata = list_entry (iter, struct dlm_reco_node_data, list);
401 BUG_ON(ndata->state != DLM_RECO_NODE_DATA_INIT);
402 ndata->state = DLM_RECO_NODE_DATA_REQUESTING;
403
404 mlog(0, "requesting lock info from node %u\n",
405 ndata->node_num);
406
407 if (ndata->node_num == dlm->node_num) {
408 ndata->state = DLM_RECO_NODE_DATA_DONE;
409 continue;
410 }
411
412 status = dlm_request_all_locks(dlm, ndata->node_num, dead_node);
413 if (status < 0) {
414 mlog_errno(status);
415 if (dlm_is_host_down(status))
416 ndata->state = DLM_RECO_NODE_DATA_DEAD;
417 else {
418 destroy = 1;
419 goto leave;
420 }
421 }
422
423 switch (ndata->state) {
424 case DLM_RECO_NODE_DATA_INIT:
425 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
426 case DLM_RECO_NODE_DATA_REQUESTED:
427 BUG();
428 break;
429 case DLM_RECO_NODE_DATA_DEAD:
430 mlog(0, "node %u died after requesting "
431 "recovery info for node %u\n",
432 ndata->node_num, dead_node);
433 // start all over
434 destroy = 1;
435 status = -EAGAIN;
436 goto leave;
437 case DLM_RECO_NODE_DATA_REQUESTING:
438 ndata->state = DLM_RECO_NODE_DATA_REQUESTED;
439 mlog(0, "now receiving recovery data from "
440 "node %u for dead node %u\n",
441 ndata->node_num, dead_node);
442 break;
443 case DLM_RECO_NODE_DATA_RECEIVING:
444 mlog(0, "already receiving recovery data from "
445 "node %u for dead node %u\n",
446 ndata->node_num, dead_node);
447 break;
448 case DLM_RECO_NODE_DATA_DONE:
449 mlog(0, "already DONE receiving recovery data "
450 "from node %u for dead node %u\n",
451 ndata->node_num, dead_node);
452 break;
453 }
454 }
455
456 mlog(0, "done requesting all lock info\n");
457
458 /* nodes should be sending reco data now
459 * just need to wait */
460
461 while (1) {
462 /* check all the nodes now to see if we are
463 * done, or if anyone died */
464 all_nodes_done = 1;
465 spin_lock(&dlm_reco_state_lock);
466 list_for_each(iter, &dlm->reco.node_data) {
467 ndata = list_entry (iter, struct dlm_reco_node_data, list);
468
469 mlog(0, "checking recovery state of node %u\n",
470 ndata->node_num);
471 switch (ndata->state) {
472 case DLM_RECO_NODE_DATA_INIT:
473 case DLM_RECO_NODE_DATA_REQUESTING:
474 mlog(ML_ERROR, "bad ndata state for "
475 "node %u: state=%d\n",
476 ndata->node_num, ndata->state);
477 BUG();
478 break;
479 case DLM_RECO_NODE_DATA_DEAD:
480 mlog(0, "node %u died after "
481 "requesting recovery info for "
482 "node %u\n", ndata->node_num,
483 dead_node);
484 spin_unlock(&dlm_reco_state_lock);
485 // start all over
486 destroy = 1;
487 status = -EAGAIN;
488 goto leave;
489 case DLM_RECO_NODE_DATA_RECEIVING:
490 case DLM_RECO_NODE_DATA_REQUESTED:
491 all_nodes_done = 0;
492 break;
493 case DLM_RECO_NODE_DATA_DONE:
494 break;
495 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
496 break;
497 }
498 }
499 spin_unlock(&dlm_reco_state_lock);
500
501 mlog(0, "pass #%d, all_nodes_done?: %s\n", ++pass,
502 all_nodes_done?"yes":"no");
503 if (all_nodes_done) {
504 int ret;
505
506 /* all nodes are now in DLM_RECO_NODE_DATA_DONE state
507 * just send a finalize message to everyone and
508 * clean up */
509 mlog(0, "all nodes are done! send finalize\n");
510 ret = dlm_send_finalize_reco_message(dlm);
511 if (ret < 0)
512 mlog_errno(ret);
513
514 spin_lock(&dlm->spinlock);
515 dlm_finish_local_lockres_recovery(dlm, dead_node,
516 dlm->node_num);
517 spin_unlock(&dlm->spinlock);
518 mlog(0, "should be done with recovery!\n");
519
520 mlog(0, "finishing recovery of %s at %lu, "
521 "dead=%u, this=%u, new=%u\n", dlm->name,
522 jiffies, dlm->reco.dead_node,
523 dlm->node_num, dlm->reco.new_master);
524 destroy = 1;
525 status = ret;
526 /* rescan everything marked dirty along the way */
527 dlm_kick_thread(dlm, NULL);
528 break;
529 }
530 /* wait to be signalled, with periodic timeout
531 * to check for node death */
532 wait_event_interruptible_timeout(dlm->dlm_reco_thread_wq,
533 kthread_should_stop(),
534 msecs_to_jiffies(DLM_RECO_THREAD_TIMEOUT_MS));
535
536 }
537
538leave:
539 if (destroy)
540 dlm_destroy_recovery_area(dlm, dead_node);
541
542 mlog_exit(status);
543 return status;
544}
545
546static int dlm_init_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
547{
548 int num=0;
549 struct dlm_reco_node_data *ndata;
550
551 spin_lock(&dlm->spinlock);
552 memcpy(dlm->reco.node_map, dlm->domain_map, sizeof(dlm->domain_map));
553 /* nodes can only be removed (by dying) after dropping
554 * this lock, and death will be trapped later, so this should do */
555 spin_unlock(&dlm->spinlock);
556
557 while (1) {
558 num = find_next_bit (dlm->reco.node_map, O2NM_MAX_NODES, num);
559 if (num >= O2NM_MAX_NODES) {
560 break;
561 }
562 BUG_ON(num == dead_node);
563
564 ndata = kcalloc(1, sizeof(*ndata), GFP_KERNEL);
565 if (!ndata) {
566 dlm_destroy_recovery_area(dlm, dead_node);
567 return -ENOMEM;
568 }
569 ndata->node_num = num;
570 ndata->state = DLM_RECO_NODE_DATA_INIT;
571 spin_lock(&dlm_reco_state_lock);
572 list_add_tail(&ndata->list, &dlm->reco.node_data);
573 spin_unlock(&dlm_reco_state_lock);
574 num++;
575 }
576
577 return 0;
578}
579
580static void dlm_destroy_recovery_area(struct dlm_ctxt *dlm, u8 dead_node)
581{
582 struct list_head *iter, *iter2;
583 struct dlm_reco_node_data *ndata;
584 LIST_HEAD(tmplist);
585
586 spin_lock(&dlm_reco_state_lock);
587 list_splice_init(&dlm->reco.node_data, &tmplist);
588 spin_unlock(&dlm_reco_state_lock);
589
590 list_for_each_safe(iter, iter2, &tmplist) {
591 ndata = list_entry (iter, struct dlm_reco_node_data, list);
592 list_del_init(&ndata->list);
593 kfree(ndata);
594 }
595}
596
597static int dlm_request_all_locks(struct dlm_ctxt *dlm, u8 request_from,
598 u8 dead_node)
599{
600 struct dlm_lock_request lr;
601 enum dlm_status ret;
602
603 mlog(0, "\n");
604
605
606 mlog(0, "dlm_request_all_locks: dead node is %u, sending request "
607 "to %u\n", dead_node, request_from);
608
609 memset(&lr, 0, sizeof(lr));
610 lr.node_idx = dlm->node_num;
611 lr.dead_node = dead_node;
612
613 // send message
614 ret = DLM_NOLOCKMGR;
615 ret = o2net_send_message(DLM_LOCK_REQUEST_MSG, dlm->key,
616 &lr, sizeof(lr), request_from, NULL);
617
618 /* negative status is handled by caller */
619 if (ret < 0)
620 mlog_errno(ret);
621
622 // return from here, then
623 // sleep until all received or error
624 return ret;
625
626}
627
628int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data)
629{
630 struct dlm_ctxt *dlm = data;
631 struct dlm_lock_request *lr = (struct dlm_lock_request *)msg->buf;
632 char *buf = NULL;
633 struct dlm_work_item *item = NULL;
634
635 if (!dlm_grab(dlm))
636 return -EINVAL;
637
638 BUG_ON(lr->dead_node != dlm->reco.dead_node);
639
640 item = kcalloc(1, sizeof(*item), GFP_KERNEL);
641 if (!item) {
642 dlm_put(dlm);
643 return -ENOMEM;
644 }
645
646 /* this will get freed by dlm_request_all_locks_worker */
647 buf = (char *) __get_free_page(GFP_KERNEL);
648 if (!buf) {
649 kfree(item);
650 dlm_put(dlm);
651 return -ENOMEM;
652 }
653
654 /* queue up work for dlm_request_all_locks_worker */
655 dlm_grab(dlm); /* get an extra ref for the work item */
656 dlm_init_work_item(dlm, item, dlm_request_all_locks_worker, buf);
657 item->u.ral.reco_master = lr->node_idx;
658 item->u.ral.dead_node = lr->dead_node;
659 spin_lock(&dlm->work_lock);
660 list_add_tail(&item->list, &dlm->work_list);
661 spin_unlock(&dlm->work_lock);
662 schedule_work(&dlm->dispatched_work);
663
664 dlm_put(dlm);
665 return 0;
666}
667
668static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
669{
670 struct dlm_migratable_lockres *mres;
671 struct dlm_lock_resource *res;
672 struct dlm_ctxt *dlm;
673 LIST_HEAD(resources);
674 struct list_head *iter;
675 int ret;
676 u8 dead_node, reco_master;
677
678 dlm = item->dlm;
679 dead_node = item->u.ral.dead_node;
680 reco_master = item->u.ral.reco_master;
681 BUG_ON(dead_node != dlm->reco.dead_node);
682 BUG_ON(reco_master != dlm->reco.new_master);
683
684 mres = (struct dlm_migratable_lockres *)data;
685
686 /* lock resources should have already been moved to the
687 * dlm->reco.resources list. now move items from that list
688 * to a temp list if the dead owner matches. note that the
689 * whole cluster recovers only one node at a time, so we
690 * can safely move UNKNOWN lock resources for each recovery
691 * session. */
692 dlm_move_reco_locks_to_list(dlm, &resources, dead_node);
693
694 /* now we can begin blasting lockreses without the dlm lock */
695 list_for_each(iter, &resources) {
696 res = list_entry (iter, struct dlm_lock_resource, recovering);
697 ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
698 DLM_MRES_RECOVERY);
699 if (ret < 0)
700 mlog_errno(ret);
701 }
702
703 /* move the resources back to the list */
704 spin_lock(&dlm->spinlock);
705 list_splice_init(&resources, &dlm->reco.resources);
706 spin_unlock(&dlm->spinlock);
707
708 ret = dlm_send_all_done_msg(dlm, dead_node, reco_master);
709 if (ret < 0)
710 mlog_errno(ret);
711
712 free_page((unsigned long)data);
713}
714
715
716static int dlm_send_all_done_msg(struct dlm_ctxt *dlm, u8 dead_node, u8 send_to)
717{
718 int ret, tmpret;
719 struct dlm_reco_data_done done_msg;
720
721 memset(&done_msg, 0, sizeof(done_msg));
722 done_msg.node_idx = dlm->node_num;
723 done_msg.dead_node = dead_node;
724 mlog(0, "sending DATA DONE message to %u, "
725 "my node=%u, dead node=%u\n", send_to, done_msg.node_idx,
726 done_msg.dead_node);
727
728 ret = o2net_send_message(DLM_RECO_DATA_DONE_MSG, dlm->key, &done_msg,
729 sizeof(done_msg), send_to, &tmpret);
730 /* negative status is ignored by the caller */
731 if (ret >= 0)
732 ret = tmpret;
733 return ret;
734}
735
736
737int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data)
738{
739 struct dlm_ctxt *dlm = data;
740 struct dlm_reco_data_done *done = (struct dlm_reco_data_done *)msg->buf;
741 struct list_head *iter;
742 struct dlm_reco_node_data *ndata = NULL;
743 int ret = -EINVAL;
744
745 if (!dlm_grab(dlm))
746 return -EINVAL;
747
748 mlog(0, "got DATA DONE: dead_node=%u, reco.dead_node=%u, "
749 "node_idx=%u, this node=%u\n", done->dead_node,
750 dlm->reco.dead_node, done->node_idx, dlm->node_num);
751 BUG_ON(done->dead_node != dlm->reco.dead_node);
752
753 spin_lock(&dlm_reco_state_lock);
754 list_for_each(iter, &dlm->reco.node_data) {
755 ndata = list_entry (iter, struct dlm_reco_node_data, list);
756 if (ndata->node_num != done->node_idx)
757 continue;
758
759 switch (ndata->state) {
760 case DLM_RECO_NODE_DATA_INIT:
761 case DLM_RECO_NODE_DATA_DEAD:
762 case DLM_RECO_NODE_DATA_DONE:
763 case DLM_RECO_NODE_DATA_FINALIZE_SENT:
764 mlog(ML_ERROR, "bad ndata state for node %u:"
765 " state=%d\n", ndata->node_num,
766 ndata->state);
767 BUG();
768 break;
769 case DLM_RECO_NODE_DATA_RECEIVING:
770 case DLM_RECO_NODE_DATA_REQUESTED:
771 case DLM_RECO_NODE_DATA_REQUESTING:
772 mlog(0, "node %u is DONE sending "
773 "recovery data!\n",
774 ndata->node_num);
775
776 ndata->state = DLM_RECO_NODE_DATA_DONE;
777 ret = 0;
778 break;
779 }
780 }
781 spin_unlock(&dlm_reco_state_lock);
782
783 /* wake the recovery thread, some node is done */
784 if (!ret)
785 dlm_kick_recovery_thread(dlm);
786
787 if (ret < 0)
788 mlog(ML_ERROR, "failed to find recovery node data for node "
789 "%u\n", done->node_idx);
790 dlm_put(dlm);
791
792 mlog(0, "leaving reco data done handler, ret=%d\n", ret);
793 return ret;
794}
795
796static void dlm_move_reco_locks_to_list(struct dlm_ctxt *dlm,
797 struct list_head *list,
798 u8 dead_node)
799{
800 struct dlm_lock_resource *res;
801 struct list_head *iter, *iter2;
802
803 spin_lock(&dlm->spinlock);
804 list_for_each_safe(iter, iter2, &dlm->reco.resources) {
805 res = list_entry (iter, struct dlm_lock_resource, recovering);
806 if (dlm_is_recovery_lock(res->lockname.name,
807 res->lockname.len))
808 continue;
809 if (res->owner == dead_node) {
810 mlog(0, "found lockres owned by dead node while "
811 "doing recovery for node %u. sending it.\n",
812 dead_node);
813 list_del_init(&res->recovering);
814 list_add_tail(&res->recovering, list);
815 } else if (res->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
816 mlog(0, "found UNKNOWN owner while doing recovery "
817 "for node %u. sending it.\n", dead_node);
818 list_del_init(&res->recovering);
819 list_add_tail(&res->recovering, list);
820 }
821 }
822 spin_unlock(&dlm->spinlock);
823}
824
825static inline int dlm_num_locks_in_lockres(struct dlm_lock_resource *res)
826{
827 int total_locks = 0;
828 struct list_head *iter, *queue = &res->granted;
829 int i;
830
831 for (i=0; i<3; i++) {
832 list_for_each(iter, queue)
833 total_locks++;
834 queue++;
835 }
836 return total_locks;
837}
838
839
840static int dlm_send_mig_lockres_msg(struct dlm_ctxt *dlm,
841 struct dlm_migratable_lockres *mres,
842 u8 send_to,
843 struct dlm_lock_resource *res,
844 int total_locks)
845{
846 u64 mig_cookie = be64_to_cpu(mres->mig_cookie);
847 int mres_total_locks = be32_to_cpu(mres->total_locks);
848 int sz, ret = 0, status = 0;
849 u8 orig_flags = mres->flags,
850 orig_master = mres->master;
851
852 BUG_ON(mres->num_locks > DLM_MAX_MIGRATABLE_LOCKS);
853 if (!mres->num_locks)
854 return 0;
855
856 sz = sizeof(struct dlm_migratable_lockres) +
857 (mres->num_locks * sizeof(struct dlm_migratable_lock));
858
859 /* add an all-done flag if we reached the last lock */
860 orig_flags = mres->flags;
861 BUG_ON(total_locks > mres_total_locks);
862 if (total_locks == mres_total_locks)
863 mres->flags |= DLM_MRES_ALL_DONE;
864
865 /* send it */
866 ret = o2net_send_message(DLM_MIG_LOCKRES_MSG, dlm->key, mres,
867 sz, send_to, &status);
868 if (ret < 0) {
869 /* XXX: negative status is not handled.
870 * this will end up killing this node. */
871 mlog_errno(ret);
872 } else {
873 /* might get an -ENOMEM back here */
874 ret = status;
875 if (ret < 0) {
876 mlog_errno(ret);
877
878 if (ret == -EFAULT) {
879 mlog(ML_ERROR, "node %u told me to kill "
880 "myself!\n", send_to);
881 BUG();
882 }
883 }
884 }
885
886 /* zero and reinit the message buffer */
887 dlm_init_migratable_lockres(mres, res->lockname.name,
888 res->lockname.len, mres_total_locks,
889 mig_cookie, orig_flags, orig_master);
890 return ret;
891}
892
893static void dlm_init_migratable_lockres(struct dlm_migratable_lockres *mres,
894 const char *lockname, int namelen,
895 int total_locks, u64 cookie,
896 u8 flags, u8 master)
897{
898 /* mres here is one full page */
899 memset(mres, 0, PAGE_SIZE);
900 mres->lockname_len = namelen;
901 memcpy(mres->lockname, lockname, namelen);
902 mres->num_locks = 0;
903 mres->total_locks = cpu_to_be32(total_locks);
904 mres->mig_cookie = cpu_to_be64(cookie);
905 mres->flags = flags;
906 mres->master = master;
907}
908
909
910/* returns 1 if this lock fills the network structure,
911 * 0 otherwise */
912static int dlm_add_lock_to_array(struct dlm_lock *lock,
913 struct dlm_migratable_lockres *mres, int queue)
914{
915 struct dlm_migratable_lock *ml;
916 int lock_num = mres->num_locks;
917
918 ml = &(mres->ml[lock_num]);
919 ml->cookie = lock->ml.cookie;
920 ml->type = lock->ml.type;
921 ml->convert_type = lock->ml.convert_type;
922 ml->highest_blocked = lock->ml.highest_blocked;
923 ml->list = queue;
924 if (lock->lksb) {
925 ml->flags = lock->lksb->flags;
926 /* send our current lvb */
927 if (ml->type == LKM_EXMODE ||
928 ml->type == LKM_PRMODE) {
929 /* if it is already set, this had better be a PR
930 * and it has to match */
931 if (mres->lvb[0] && (ml->type == LKM_EXMODE ||
932 memcmp(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN))) {
933 mlog(ML_ERROR, "mismatched lvbs!\n");
934 __dlm_print_one_lock_resource(lock->lockres);
935 BUG();
936 }
937 memcpy(mres->lvb, lock->lksb->lvb, DLM_LVB_LEN);
938 }
939 }
940 ml->node = lock->ml.node;
941 mres->num_locks++;
942 /* we reached the max, send this network message */
943 if (mres->num_locks == DLM_MAX_MIGRATABLE_LOCKS)
944 return 1;
945 return 0;
946}
947
948
949int dlm_send_one_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
950 struct dlm_migratable_lockres *mres,
951 u8 send_to, u8 flags)
952{
953 struct list_head *queue, *iter;
954 int total_locks, i;
955 u64 mig_cookie = 0;
956 struct dlm_lock *lock;
957 int ret = 0;
958
959 BUG_ON(!(flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
960
961 mlog(0, "sending to %u\n", send_to);
962
963 total_locks = dlm_num_locks_in_lockres(res);
964 if (total_locks > DLM_MAX_MIGRATABLE_LOCKS) {
965 /* rare, but possible */
966 mlog(0, "argh. lockres has %d locks. this will "
967 "require more than one network packet to "
968 "migrate\n", total_locks);
969 mig_cookie = dlm_get_next_mig_cookie();
970 }
971
972 dlm_init_migratable_lockres(mres, res->lockname.name,
973 res->lockname.len, total_locks,
974 mig_cookie, flags, res->owner);
975
976 total_locks = 0;
977 for (i=DLM_GRANTED_LIST; i<=DLM_BLOCKED_LIST; i++) {
978 queue = dlm_list_idx_to_ptr(res, i);
979 list_for_each(iter, queue) {
980 lock = list_entry (iter, struct dlm_lock, list);
981
982 /* add another lock. */
983 total_locks++;
984 if (!dlm_add_lock_to_array(lock, mres, i))
985 continue;
986
987 /* this filled the lock message,
988 * we must send it immediately. */
989 ret = dlm_send_mig_lockres_msg(dlm, mres, send_to,
990 res, total_locks);
991 if (ret < 0) {
992 // TODO
993 mlog(ML_ERROR, "dlm_send_mig_lockres_msg "
994 "returned %d, TODO\n", ret);
995 BUG();
996 }
997 }
998 }
999 /* flush any remaining locks */
1000 ret = dlm_send_mig_lockres_msg(dlm, mres, send_to, res, total_locks);
1001 if (ret < 0) {
1002 // TODO
1003 mlog(ML_ERROR, "dlm_send_mig_lockres_msg returned %d, "
1004 "TODO\n", ret);
1005 BUG();
1006 }
1007 return ret;
1008}
1009
1010
1011
1012/*
1013 * this message will contain no more than one page worth of
1014 * recovery data, and it will work on only one lockres.
1015 * there may be many locks in this page, and we may need to wait
1016 * for additional packets to complete all the locks (rare, but
1017 * possible).
1018 */
1019/*
1020 * NOTE: the allocation error cases here are scary
1021 * we really cannot afford to fail an alloc in recovery
1022 * do we spin? returning an error only delays the problem really
1023 */
1024
1025int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data)
1026{
1027 struct dlm_ctxt *dlm = data;
1028 struct dlm_migratable_lockres *mres =
1029 (struct dlm_migratable_lockres *)msg->buf;
1030 int ret = 0;
1031 u8 real_master;
1032 char *buf = NULL;
1033 struct dlm_work_item *item = NULL;
1034 struct dlm_lock_resource *res = NULL;
1035
1036 if (!dlm_grab(dlm))
1037 return -EINVAL;
1038
1039 BUG_ON(!(mres->flags & (DLM_MRES_RECOVERY|DLM_MRES_MIGRATION)));
1040
1041 real_master = mres->master;
1042 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1043 /* cannot migrate a lockres with no master */
1044 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1045 }
1046
1047 mlog(0, "%s message received from node %u\n",
1048 (mres->flags & DLM_MRES_RECOVERY) ?
1049 "recovery" : "migration", mres->master);
1050 if (mres->flags & DLM_MRES_ALL_DONE)
1051 mlog(0, "all done flag. all lockres data received!\n");
1052
1053 ret = -ENOMEM;
1054 buf = kmalloc(be16_to_cpu(msg->data_len), GFP_KERNEL);
1055 item = kcalloc(1, sizeof(*item), GFP_KERNEL);
1056 if (!buf || !item)
1057 goto leave;
1058
1059 /* lookup the lock to see if we have a secondary queue for this
1060 * already... just add the locks in and this will have its owner
1061 * and RECOVERY flag changed when it completes. */
1062 res = dlm_lookup_lockres(dlm, mres->lockname, mres->lockname_len);
1063 if (res) {
1064 /* this will get a ref on res */
1065 /* mark it as recovering/migrating and hash it */
1066 spin_lock(&res->spinlock);
1067 if (mres->flags & DLM_MRES_RECOVERY) {
1068 res->state |= DLM_LOCK_RES_RECOVERING;
1069 } else {
1070 if (res->state & DLM_LOCK_RES_MIGRATING) {
1071 /* this is at least the second
1072 * lockres message */
1073 mlog(0, "lock %.*s is already migrating\n",
1074 mres->lockname_len,
1075 mres->lockname);
1076 } else if (res->state & DLM_LOCK_RES_RECOVERING) {
1077 /* caller should BUG */
1078 mlog(ML_ERROR, "node is attempting to migrate "
1079 "lock %.*s, but marked as recovering!\n",
1080 mres->lockname_len, mres->lockname);
1081 ret = -EFAULT;
1082 spin_unlock(&res->spinlock);
1083 goto leave;
1084 }
1085 res->state |= DLM_LOCK_RES_MIGRATING;
1086 }
1087 spin_unlock(&res->spinlock);
1088 } else {
1089 /* need to allocate, just like if it was
1090 * mastered here normally */
1091 res = dlm_new_lockres(dlm, mres->lockname, mres->lockname_len);
1092 if (!res)
1093 goto leave;
1094
1095 /* to match the ref that we would have gotten if
1096 * dlm_lookup_lockres had succeeded */
1097 dlm_lockres_get(res);
1098
1099 /* mark it as recovering/migrating and hash it */
1100 if (mres->flags & DLM_MRES_RECOVERY)
1101 res->state |= DLM_LOCK_RES_RECOVERING;
1102 else
1103 res->state |= DLM_LOCK_RES_MIGRATING;
1104
1105 spin_lock(&dlm->spinlock);
1106 __dlm_insert_lockres(dlm, res);
1107 spin_unlock(&dlm->spinlock);
1108
1109 /* now that the new lockres is inserted,
1110 * make it usable by other processes */
1111 spin_lock(&res->spinlock);
1112 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
1113 spin_unlock(&res->spinlock);
1114
1115 /* add an extra ref for just-allocated lockres
1116 * otherwise the lockres will be purged immediately */
1117 dlm_lockres_get(res);
1118
1119 }
1120
1121 /* at this point we have allocated everything we need,
1122 * and we have a hashed lockres with an extra ref and
1123 * the proper res->state flags. */
1124 ret = 0;
1125 if (mres->master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1126 /* migration cannot have an unknown master */
1127 BUG_ON(!(mres->flags & DLM_MRES_RECOVERY));
1128 mlog(0, "recovery has passed me a lockres with an "
1129 "unknown owner.. will need to requery: "
1130 "%.*s\n", mres->lockname_len, mres->lockname);
1131 } else {
1132 spin_lock(&res->spinlock);
1133 dlm_change_lockres_owner(dlm, res, dlm->node_num);
1134 spin_unlock(&res->spinlock);
1135 }
1136
1137 /* queue up work for dlm_mig_lockres_worker */
1138 dlm_grab(dlm); /* get an extra ref for the work item */
1139 memcpy(buf, msg->buf, be16_to_cpu(msg->data_len)); /* copy the whole message */
1140 dlm_init_work_item(dlm, item, dlm_mig_lockres_worker, buf);
1141 item->u.ml.lockres = res; /* already have a ref */
1142 item->u.ml.real_master = real_master;
1143 spin_lock(&dlm->work_lock);
1144 list_add_tail(&item->list, &dlm->work_list);
1145 spin_unlock(&dlm->work_lock);
1146 schedule_work(&dlm->dispatched_work);
1147
1148leave:
1149 dlm_put(dlm);
1150 if (ret < 0) {
1151 if (buf)
1152 kfree(buf);
1153 if (item)
1154 kfree(item);
1155 }
1156
1157 mlog_exit(ret);
1158 return ret;
1159}
1160
1161
1162static void dlm_mig_lockres_worker(struct dlm_work_item *item, void *data)
1163{
1164 struct dlm_ctxt *dlm = data;
1165 struct dlm_migratable_lockres *mres;
1166 int ret = 0;
1167 struct dlm_lock_resource *res;
1168 u8 real_master;
1169
1170 dlm = item->dlm;
1171 mres = (struct dlm_migratable_lockres *)data;
1172
1173 res = item->u.ml.lockres;
1174 real_master = item->u.ml.real_master;
1175
1176 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1177 /* this case is super-rare. only occurs if
1178 * node death happens during migration. */
1179again:
1180 ret = dlm_lockres_master_requery(dlm, res, &real_master);
1181 if (ret < 0) {
1182 mlog(0, "dlm_lockres_master_requery failure: %d\n",
1183 ret);
1184 goto again;
1185 }
1186 if (real_master == DLM_LOCK_RES_OWNER_UNKNOWN) {
1187 mlog(0, "lockres %.*s not claimed. "
1188 "this node will take it.\n",
1189 res->lockname.len, res->lockname.name);
1190 } else {
1191 mlog(0, "master needs to respond to sender "
1192 "that node %u still owns %.*s\n",
1193 real_master, res->lockname.len,
1194 res->lockname.name);
1195 /* cannot touch this lockres */
1196 goto leave;
1197 }
1198 }
1199
1200 ret = dlm_process_recovery_data(dlm, res, mres);
1201 if (ret < 0)
1202 mlog(0, "dlm_process_recovery_data returned %d\n", ret);
1203 else
1204 mlog(0, "dlm_process_recovery_data succeeded\n");
1205
1206 if ((mres->flags & (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) ==
1207 (DLM_MRES_MIGRATION|DLM_MRES_ALL_DONE)) {
1208 ret = dlm_finish_migration(dlm, res, mres->master);
1209 if (ret < 0)
1210 mlog_errno(ret);
1211 }
1212
1213leave:
1214 kfree(data);
1215 mlog_exit(ret);
1216}
1217
1218
1219
1220static int dlm_lockres_master_requery(struct dlm_ctxt *dlm,
1221 struct dlm_lock_resource *res,
1222 u8 *real_master)
1223{
1224 struct dlm_node_iter iter;
1225 int nodenum;
1226 int ret = 0;
1227
1228 *real_master = DLM_LOCK_RES_OWNER_UNKNOWN;
1229
1230 /* we only reach here if one of the two nodes in a
1231 * migration died while the migration was in progress.
1232 * at this point we need to requery the master. we
1233 * know that the new_master got as far as creating
1234 * an mle on at least one node, but we do not know
1235 * if any nodes had actually cleared the mle and set
1236 * the master to the new_master. the old master
1237 * is supposed to set the owner to UNKNOWN in the
1238 * event of a new_master death, so the only possible
1239 * responses that we can get from nodes here are
1240 * that the master is new_master, or that the master
1241 * is UNKNOWN.
1242 * if all nodes come back with UNKNOWN then we know
1243 * the lock needs remastering here.
1244 * if any node comes back with a valid master, check
1245 * to see if that master is the one that we are
1246 * recovering. if so, then the new_master died and
1247 * we need to remaster this lock. if not, then the
1248 * new_master survived and that node will respond to
1249 * other nodes about the owner.
1250 * if there is an owner, this node needs to dump this
1251 * lockres and alert the sender that this lockres
1252 * was rejected. */
1253 spin_lock(&dlm->spinlock);
1254 dlm_node_iter_init(dlm->domain_map, &iter);
1255 spin_unlock(&dlm->spinlock);
1256
1257 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1258 /* do not send to self */
1259 if (nodenum == dlm->node_num)
1260 continue;
1261 ret = dlm_do_master_requery(dlm, res, nodenum, real_master);
1262 if (ret < 0) {
1263 mlog_errno(ret);
1264 BUG();
1265 /* TODO: need to figure a way to restart this */
1266 }
1267 if (*real_master != DLM_LOCK_RES_OWNER_UNKNOWN) {
1268 mlog(0, "lock master is %u\n", *real_master);
1269 break;
1270 }
1271 }
1272 return ret;
1273}
1274
1275
1276static int dlm_do_master_requery(struct dlm_ctxt *dlm,
1277 struct dlm_lock_resource *res,
1278 u8 nodenum, u8 *real_master)
1279{
1280 int ret = -EINVAL;
1281 struct dlm_master_requery req;
1282 int status = DLM_LOCK_RES_OWNER_UNKNOWN;
1283
1284 memset(&req, 0, sizeof(req));
1285 req.node_idx = dlm->node_num;
1286 req.namelen = res->lockname.len;
1287 memcpy(req.name, res->lockname.name, res->lockname.len);
1288
1289 ret = o2net_send_message(DLM_MASTER_REQUERY_MSG, dlm->key,
1290 &req, sizeof(req), nodenum, &status);
1291 /* XXX: negative status not handled properly here. */
1292 if (ret < 0)
1293 mlog_errno(ret);
1294 else {
1295 BUG_ON(status < 0);
1296 BUG_ON(status > DLM_LOCK_RES_OWNER_UNKNOWN);
1297 *real_master = (u8) (status & 0xff);
1298 mlog(0, "node %u responded to master requery with %u\n",
1299 nodenum, *real_master);
1300 ret = 0;
1301 }
1302 return ret;
1303}
1304
1305
1306/* this function cannot error, so unless the sending
1307 * or receiving of the message failed, the owner can
1308 * be trusted */
1309int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data)
1310{
1311 struct dlm_ctxt *dlm = data;
1312 struct dlm_master_requery *req = (struct dlm_master_requery *)msg->buf;
1313 struct dlm_lock_resource *res = NULL;
1314 int master = DLM_LOCK_RES_OWNER_UNKNOWN;
1315 u32 flags = DLM_ASSERT_MASTER_REQUERY;
1316
1317 if (!dlm_grab(dlm)) {
1318 /* since the domain has gone away on this
1319 * node, the proper response is UNKNOWN */
1320 return master;
1321 }
1322
1323 spin_lock(&dlm->spinlock);
1324 res = __dlm_lookup_lockres(dlm, req->name, req->namelen);
1325 if (res) {
1326 spin_lock(&res->spinlock);
1327 master = res->owner;
1328 if (master == dlm->node_num) {
1329 int ret = dlm_dispatch_assert_master(dlm, res,
1330 0, 0, flags);
1331 if (ret < 0) {
1332 mlog_errno(-ENOMEM);
1333 /* retry!? */
1334 BUG();
1335 }
1336 }
1337 spin_unlock(&res->spinlock);
1338 }
1339 spin_unlock(&dlm->spinlock);
1340
1341 dlm_put(dlm);
1342 return master;
1343}
1344
1345static inline struct list_head *
1346dlm_list_num_to_pointer(struct dlm_lock_resource *res, int list_num)
1347{
1348 struct list_head *ret;
1349 BUG_ON(list_num < 0);
1350 BUG_ON(list_num > 2);
1351 ret = &(res->granted);
1352 ret += list_num;
1353 return ret;
1354}
1355/* TODO: do ast flush business
1356 * TODO: do MIGRATING and RECOVERING spinning
1357 */
1358
1359/*
1360* NOTE about in-flight requests during migration:
1361*
1362* Before attempting the migrate, the master has marked the lockres as
1363* MIGRATING and then flushed all of its pending ASTS. So any in-flight
1364* requests either got queued before the MIGRATING flag got set, in which
1365* case the lock data will reflect the change and a return message is on
1366* the way, or the request failed to get in before MIGRATING got set. In
1367* this case, the caller will be told to spin and wait for the MIGRATING
1368* flag to be dropped, then recheck the master.
1369* This holds true for the convert, cancel and unlock cases, and since lvb
1370* updates are tied to these same messages, it applies to lvb updates as
1371* well. For the lock case, there is no way a lock can be on the master
1372* queue and not be on the secondary queue since the lock is always added
1373* locally first. This means that the new target node will never be sent
1374* a lock that he doesn't already have on the list.
1375* In total, this means that the local lock is correct and should not be
1376* updated to match the one sent by the master. Any messages sent back
1377* from the master before the MIGRATING flag will bring the lock properly
1378* up-to-date, and the change will be ordered properly for the waiter.
1379* We will *not* attempt to modify the lock underneath the waiter.
1380*/
1381
1382static int dlm_process_recovery_data(struct dlm_ctxt *dlm,
1383 struct dlm_lock_resource *res,
1384 struct dlm_migratable_lockres *mres)
1385{
1386 struct dlm_migratable_lock *ml;
1387 struct list_head *queue;
1388 struct dlm_lock *newlock = NULL;
1389 struct dlm_lockstatus *lksb = NULL;
1390 int ret = 0;
1391 int i;
1392 struct list_head *iter;
1393 struct dlm_lock *lock = NULL;
1394
1395 mlog(0, "running %d locks for this lockres\n", mres->num_locks);
1396 for (i=0; i<mres->num_locks; i++) {
1397 ml = &(mres->ml[i]);
1398 BUG_ON(ml->highest_blocked != LKM_IVMODE);
1399 newlock = NULL;
1400 lksb = NULL;
1401
1402 queue = dlm_list_num_to_pointer(res, ml->list);
1403
1404 /* if the lock is for the local node it needs to
1405 * be moved to the proper location within the queue.
1406 * do not allocate a new lock structure. */
1407 if (ml->node == dlm->node_num) {
1408 /* MIGRATION ONLY! */
1409 BUG_ON(!(mres->flags & DLM_MRES_MIGRATION));
1410
1411 spin_lock(&res->spinlock);
1412 list_for_each(iter, queue) {
1413 lock = list_entry (iter, struct dlm_lock, list);
1414 if (lock->ml.cookie != ml->cookie)
1415 lock = NULL;
1416 else
1417 break;
1418 }
1419
1420 /* lock is always created locally first, and
1421 * destroyed locally last. it must be on the list */
1422 if (!lock) {
1423 mlog(ML_ERROR, "could not find local lock "
1424 "with cookie %"MLFu64"!\n",
1425 ml->cookie);
1426 BUG();
1427 }
1428 BUG_ON(lock->ml.node != ml->node);
1429
1430 /* see NOTE above about why we do not update
1431 * to match the master here */
1432
1433 /* move the lock to its proper place */
1434 /* do not alter lock refcount. switching lists. */
1435 list_del_init(&lock->list);
1436 list_add_tail(&lock->list, queue);
1437 spin_unlock(&res->spinlock);
1438
1439 mlog(0, "just reordered a local lock!\n");
1440 continue;
1441 }
1442
1443 /* lock is for another node. */
1444 newlock = dlm_new_lock(ml->type, ml->node,
1445 be64_to_cpu(ml->cookie), NULL);
1446 if (!newlock) {
1447 ret = -ENOMEM;
1448 goto leave;
1449 }
1450 lksb = newlock->lksb;
1451 dlm_lock_attach_lockres(newlock, res);
1452
1453 if (ml->convert_type != LKM_IVMODE) {
1454 BUG_ON(queue != &res->converting);
1455 newlock->ml.convert_type = ml->convert_type;
1456 }
1457 lksb->flags |= (ml->flags &
1458 (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
1459
1460 if (mres->lvb[0]) {
1461 if (lksb->flags & DLM_LKSB_PUT_LVB) {
1462 /* other node was trying to update
1463 * lvb when node died. recreate the
1464 * lksb with the updated lvb. */
1465 memcpy(lksb->lvb, mres->lvb, DLM_LVB_LEN);
1466 } else {
1467 /* otherwise, the node is sending its
1468 * most recent valid lvb info */
1469 BUG_ON(ml->type != LKM_EXMODE &&
1470 ml->type != LKM_PRMODE);
1471 if (res->lvb[0] && (ml->type == LKM_EXMODE ||
1472 memcmp(res->lvb, mres->lvb, DLM_LVB_LEN))) {
1473 mlog(ML_ERROR, "received bad lvb!\n");
1474 __dlm_print_one_lock_resource(res);
1475 BUG();
1476 }
1477 memcpy(res->lvb, mres->lvb, DLM_LVB_LEN);
1478 }
1479 }
1480
1481
1482 /* NOTE:
1483 * wrt lock queue ordering and recovery:
1484 * 1. order of locks on granted queue is
1485 * meaningless.
1486 * 2. order of locks on converting queue is
1487 * LOST with the node death. sorry charlie.
1488 * 3. order of locks on the blocked queue is
1489 * also LOST.
1490 * order of locks does not affect integrity, it
1491 * just means that a lock request may get pushed
1492 * back in line as a result of the node death.
1493 * also note that for a given node the lock order
1494 * for its secondary queue locks is preserved
1495 * relative to each other, but clearly *not*
1496 * preserved relative to locks from other nodes.
1497 */
1498 spin_lock(&res->spinlock);
1499 dlm_lock_get(newlock);
1500 list_add_tail(&newlock->list, queue);
1501 spin_unlock(&res->spinlock);
1502 }
1503 mlog(0, "done running all the locks\n");
1504
1505leave:
1506 if (ret < 0) {
1507 mlog_errno(ret);
1508 if (newlock)
1509 dlm_lock_put(newlock);
1510 }
1511
1512 mlog_exit(ret);
1513 return ret;
1514}
1515
1516void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
1517 struct dlm_lock_resource *res)
1518{
1519 int i;
1520 struct list_head *queue, *iter, *iter2;
1521 struct dlm_lock *lock;
1522
1523 res->state |= DLM_LOCK_RES_RECOVERING;
1524 if (!list_empty(&res->recovering))
1525 list_del_init(&res->recovering);
1526 list_add_tail(&res->recovering, &dlm->reco.resources);
1527
1528 /* find any pending locks and put them back on proper list */
1529 for (i=DLM_BLOCKED_LIST; i>=DLM_GRANTED_LIST; i--) {
1530 queue = dlm_list_idx_to_ptr(res, i);
1531 list_for_each_safe(iter, iter2, queue) {
1532 lock = list_entry (iter, struct dlm_lock, list);
1533 dlm_lock_get(lock);
1534 if (lock->convert_pending) {
1535 /* move converting lock back to granted */
1536 BUG_ON(i != DLM_CONVERTING_LIST);
1537 mlog(0, "node died with convert pending "
1538 "on %.*s. move back to granted list.\n",
1539 res->lockname.len, res->lockname.name);
1540 dlm_revert_pending_convert(res, lock);
1541 lock->convert_pending = 0;
1542 } else if (lock->lock_pending) {
1543 /* remove pending lock requests completely */
1544 BUG_ON(i != DLM_BLOCKED_LIST);
1545 mlog(0, "node died with lock pending "
1546 "on %.*s. remove from blocked list and skip.\n",
1547 res->lockname.len, res->lockname.name);
1548 /* lock will be floating until ref in
1549 * dlmlock_remote is freed after the network
1550 * call returns. ok for it to not be on any
1551 * list since no ast can be called
1552 * (the master is dead). */
1553 dlm_revert_pending_lock(res, lock);
1554 lock->lock_pending = 0;
1555 } else if (lock->unlock_pending) {
1556 /* if an unlock was in progress, treat as
1557 * if this had completed successfully
1558 * before sending this lock state to the
1559 * new master. note that the dlm_unlock
1560 * call is still responsible for calling
1561 * the unlockast. that will happen after
1562 * the network call times out. for now,
1563 * just move lists to prepare the new
1564 * recovery master. */
1565 BUG_ON(i != DLM_GRANTED_LIST);
1566 mlog(0, "node died with unlock pending "
1567 "on %.*s. remove from blocked list and skip.\n",
1568 res->lockname.len, res->lockname.name);
1569 dlm_commit_pending_unlock(res, lock);
1570 lock->unlock_pending = 0;
1571 } else if (lock->cancel_pending) {
1572 /* if a cancel was in progress, treat as
1573 * if this had completed successfully
1574 * before sending this lock state to the
1575 * new master */
1576 BUG_ON(i != DLM_CONVERTING_LIST);
1577 mlog(0, "node died with cancel pending "
1578 "on %.*s. move back to granted list.\n",
1579 res->lockname.len, res->lockname.name);
1580 dlm_commit_pending_cancel(res, lock);
1581 lock->cancel_pending = 0;
1582 }
1583 dlm_lock_put(lock);
1584 }
1585 }
1586}
1587
1588
1589
1590/* removes all recovered locks from the recovery list.
1591 * sets the res->owner to the new master.
1592 * unsets the RECOVERY flag and wakes waiters. */
1593static void dlm_finish_local_lockres_recovery(struct dlm_ctxt *dlm,
1594 u8 dead_node, u8 new_master)
1595{
1596 int i;
1597 struct list_head *iter, *iter2, *bucket;
1598 struct dlm_lock_resource *res;
1599
1600 mlog_entry_void();
1601
1602 assert_spin_locked(&dlm->spinlock);
1603
1604 list_for_each_safe(iter, iter2, &dlm->reco.resources) {
1605 res = list_entry (iter, struct dlm_lock_resource, recovering);
1606 if (res->owner == dead_node) {
1607 list_del_init(&res->recovering);
1608 spin_lock(&res->spinlock);
1609 dlm_change_lockres_owner(dlm, res, new_master);
1610 res->state &= ~DLM_LOCK_RES_RECOVERING;
1611 __dlm_dirty_lockres(dlm, res);
1612 spin_unlock(&res->spinlock);
1613 wake_up(&res->wq);
1614 }
1615 }
1616
1617 /* this will become unnecessary eventually, but
1618 * for now we need to run the whole hash, clear
1619 * the RECOVERING state and set the owner
1620 * if necessary */
1621 for (i=0; i<DLM_HASH_SIZE; i++) {
1622 bucket = &(dlm->resources[i]);
1623 list_for_each(iter, bucket) {
1624 res = list_entry (iter, struct dlm_lock_resource, list);
1625 if (res->state & DLM_LOCK_RES_RECOVERING) {
1626 if (res->owner == dead_node) {
1627 mlog(0, "(this=%u) res %.*s owner=%u "
1628 "was not on recovering list, but "
1629 "clearing state anyway\n",
1630 dlm->node_num, res->lockname.len,
1631 res->lockname.name, new_master);
1632 } else if (res->owner == dlm->node_num) {
1633 mlog(0, "(this=%u) res %.*s owner=%u "
1634 "was not on recovering list, "
1635 "owner is THIS node, clearing\n",
1636 dlm->node_num, res->lockname.len,
1637 res->lockname.name, new_master);
1638 } else
1639 continue;
1640
1641 spin_lock(&res->spinlock);
1642 dlm_change_lockres_owner(dlm, res, new_master);
1643 res->state &= ~DLM_LOCK_RES_RECOVERING;
1644 __dlm_dirty_lockres(dlm, res);
1645 spin_unlock(&res->spinlock);
1646 wake_up(&res->wq);
1647 }
1648 }
1649 }
1650}
1651
1652static inline int dlm_lvb_needs_invalidation(struct dlm_lock *lock, int local)
1653{
1654 if (local) {
1655 if (lock->ml.type != LKM_EXMODE &&
1656 lock->ml.type != LKM_PRMODE)
1657 return 1;
1658 } else if (lock->ml.type == LKM_EXMODE)
1659 return 1;
1660 return 0;
1661}
1662
1663static void dlm_revalidate_lvb(struct dlm_ctxt *dlm,
1664 struct dlm_lock_resource *res, u8 dead_node)
1665{
1666 struct list_head *iter, *queue;
1667 struct dlm_lock *lock;
1668 int blank_lvb = 0, local = 0;
1669 int i;
1670 u8 search_node;
1671
1672 assert_spin_locked(&dlm->spinlock);
1673 assert_spin_locked(&res->spinlock);
1674
1675 if (res->owner == dlm->node_num)
1676 /* if this node owned the lockres, and if the dead node
1677 * had an EX when he died, blank out the lvb */
1678 search_node = dead_node;
1679 else {
1680 /* if this is a secondary lockres, and we had no EX or PR
1681 * locks granted, we can no longer trust the lvb */
1682 search_node = dlm->node_num;
1683 local = 1; /* check local state for valid lvb */
1684 }
1685
1686 for (i=DLM_GRANTED_LIST; i<=DLM_CONVERTING_LIST; i++) {
1687 queue = dlm_list_idx_to_ptr(res, i);
1688 list_for_each(iter, queue) {
1689 lock = list_entry (iter, struct dlm_lock, list);
1690 if (lock->ml.node == search_node) {
1691 if (dlm_lvb_needs_invalidation(lock, local)) {
1692 /* zero the lksb lvb and lockres lvb */
1693 blank_lvb = 1;
1694 memset(lock->lksb->lvb, 0, DLM_LVB_LEN);
1695 }
1696 }
1697 }
1698 }
1699
1700 if (blank_lvb) {
1701 mlog(0, "clearing %.*s lvb, dead node %u had EX\n",
1702 res->lockname.len, res->lockname.name, dead_node);
1703 memset(res->lvb, 0, DLM_LVB_LEN);
1704 }
1705}
1706
1707static void dlm_free_dead_locks(struct dlm_ctxt *dlm,
1708 struct dlm_lock_resource *res, u8 dead_node)
1709{
1710 struct list_head *iter, *tmpiter;
1711 struct dlm_lock *lock;
1712
1713 /* this node is the lockres master:
1714 * 1) remove any stale locks for the dead node
1715 * 2) if the dead node had an EX when he died, blank out the lvb
1716 */
1717 assert_spin_locked(&dlm->spinlock);
1718 assert_spin_locked(&res->spinlock);
1719
1720 /* TODO: check pending_asts, pending_basts here */
1721 list_for_each_safe(iter, tmpiter, &res->granted) {
1722 lock = list_entry (iter, struct dlm_lock, list);
1723 if (lock->ml.node == dead_node) {
1724 list_del_init(&lock->list);
1725 dlm_lock_put(lock);
1726 }
1727 }
1728 list_for_each_safe(iter, tmpiter, &res->converting) {
1729 lock = list_entry (iter, struct dlm_lock, list);
1730 if (lock->ml.node == dead_node) {
1731 list_del_init(&lock->list);
1732 dlm_lock_put(lock);
1733 }
1734 }
1735 list_for_each_safe(iter, tmpiter, &res->blocked) {
1736 lock = list_entry (iter, struct dlm_lock, list);
1737 if (lock->ml.node == dead_node) {
1738 list_del_init(&lock->list);
1739 dlm_lock_put(lock);
1740 }
1741 }
1742
1743 /* do not kick thread yet */
1744 __dlm_dirty_lockres(dlm, res);
1745}
1746
1747/* if this node is the recovery master, and there are no
1748 * locks for a given lockres owned by this node that are in
1749 * either PR or EX mode, zero out the lvb before requesting.
1750 *
1751 */
1752
1753
1754static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
1755{
1756 struct list_head *iter;
1757 struct dlm_lock_resource *res;
1758 int i;
1759 struct list_head *bucket;
1760
1761
1762 /* purge any stale mles */
1763 dlm_clean_master_list(dlm, dead_node);
1764
1765 /*
1766 * now clean up all lock resources. there are two rules:
1767 *
1768 * 1) if the dead node was the master, move the lockres
1769 * to the recovering list. set the RECOVERING flag.
1770 * this lockres needs to be cleaned up before it can
1771 * be used further.
1772 *
1773 * 2) if this node was the master, remove all locks from
1774 * each of the lockres queues that were owned by the
1775 * dead node. once recovery finishes, the dlm thread
1776 * can be kicked again to see if any ASTs or BASTs
1777 * need to be fired as a result.
1778 */
1779 for (i=0; i<DLM_HASH_SIZE; i++) {
1780 bucket = &(dlm->resources[i]);
1781 list_for_each(iter, bucket) {
1782 res = list_entry (iter, struct dlm_lock_resource, list);
1783 if (dlm_is_recovery_lock(res->lockname.name,
1784 res->lockname.len))
1785 continue;
1786
1787 spin_lock(&res->spinlock);
1788 /* zero the lvb if necessary */
1789 dlm_revalidate_lvb(dlm, res, dead_node);
1790 if (res->owner == dead_node)
1791 dlm_move_lockres_to_recovery_list(dlm, res);
1792 else if (res->owner == dlm->node_num) {
1793 dlm_free_dead_locks(dlm, res, dead_node);
1794 __dlm_lockres_calc_usage(dlm, res);
1795 }
1796 spin_unlock(&res->spinlock);
1797 }
1798 }
1799
1800}
1801
1802static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
1803{
1804 assert_spin_locked(&dlm->spinlock);
1805
1806 /* check to see if the node is already considered dead */
1807 if (!test_bit(idx, dlm->live_nodes_map)) {
1808 mlog(0, "for domain %s, node %d is already dead. "
1809 "another node likely did recovery already.\n",
1810 dlm->name, idx);
1811 return;
1812 }
1813
1814 /* check to see if we do not care about this node */
1815 if (!test_bit(idx, dlm->domain_map)) {
1816 /* This also catches the case that we get a node down
1817 * but haven't joined the domain yet. */
1818 mlog(0, "node %u already removed from domain!\n", idx);
1819 return;
1820 }
1821
1822 clear_bit(idx, dlm->live_nodes_map);
1823
1824 /* Clean up join state on node death. */
1825 if (dlm->joining_node == idx) {
1826 mlog(0, "Clearing join state for node %u\n", idx);
1827 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1828 }
1829
1830 /* make sure local cleanup occurs before the heartbeat events */
1831 if (!test_bit(idx, dlm->recovery_map))
1832 dlm_do_local_recovery_cleanup(dlm, idx);
1833
1834 /* notify anything attached to the heartbeat events */
1835 dlm_hb_event_notify_attached(dlm, idx, 0);
1836
1837 mlog(0, "node %u being removed from domain map!\n", idx);
1838 clear_bit(idx, dlm->domain_map);
1839 /* wake up migration waiters if a node goes down.
1840 * perhaps later we can genericize this for other waiters. */
1841 wake_up(&dlm->migration_wq);
1842
1843 if (test_bit(idx, dlm->recovery_map))
1844 mlog(0, "domain %s, node %u already added "
1845 "to recovery map!\n", dlm->name, idx);
1846 else
1847 set_bit(idx, dlm->recovery_map);
1848}
1849
1850void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data)
1851{
1852 struct dlm_ctxt *dlm = data;
1853
1854 if (!dlm_grab(dlm))
1855 return;
1856
1857 spin_lock(&dlm->spinlock);
1858 __dlm_hb_node_down(dlm, idx);
1859 spin_unlock(&dlm->spinlock);
1860
1861 dlm_put(dlm);
1862}
1863
1864void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data)
1865{
1866 struct dlm_ctxt *dlm = data;
1867
1868 if (!dlm_grab(dlm))
1869 return;
1870
1871 spin_lock(&dlm->spinlock);
1872
1873 set_bit(idx, dlm->live_nodes_map);
1874
1875 /* notify any mles attached to the heartbeat events */
1876 dlm_hb_event_notify_attached(dlm, idx, 1);
1877
1878 spin_unlock(&dlm->spinlock);
1879
1880 dlm_put(dlm);
1881}
1882
1883static void dlm_reco_ast(void *astdata)
1884{
1885 struct dlm_ctxt *dlm = astdata;
1886 mlog(0, "ast for recovery lock fired!, this=%u, dlm=%s\n",
1887 dlm->node_num, dlm->name);
1888}
1889static void dlm_reco_bast(void *astdata, int blocked_type)
1890{
1891 struct dlm_ctxt *dlm = astdata;
1892 mlog(0, "bast for recovery lock fired!, this=%u, dlm=%s\n",
1893 dlm->node_num, dlm->name);
1894}
1895static void dlm_reco_unlock_ast(void *astdata, enum dlm_status st)
1896{
1897 mlog(0, "unlockast for recovery lock fired!\n");
1898}
1899
1900
1901static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
1902{
1903 enum dlm_status ret;
1904 struct dlm_lockstatus lksb;
1905 int status = -EINVAL;
1906
1907 mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
1908 dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
1909retry:
1910 memset(&lksb, 0, sizeof(lksb));
1911
1912 ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
1913 DLM_RECOVERY_LOCK_NAME, dlm_reco_ast, dlm, dlm_reco_bast);
1914
1915 if (ret == DLM_NORMAL) {
1916 mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
1917 dlm->name, dlm->node_num);
1918 /* I am master, send message to all nodes saying
1919 * that I am beginning a recovery session */
1920 status = dlm_send_begin_reco_message(dlm,
1921 dlm->reco.dead_node);
1922
1923 /* recovery lock is a special case. ast will not get fired,
1924 * so just go ahead and unlock it. */
1925 ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
1926 if (ret != DLM_NORMAL) {
1927 /* this would really suck. this could only happen
1928 * if there was a network error during the unlock
1929 * because of node death. this means the unlock
1930 * is actually "done" and the lock structure is
1931 * even freed. we can continue, but only
1932 * because this specific lock name is special. */
1933 mlog(0, "dlmunlock returned %d\n", ret);
1934 }
1935
1936 if (status < 0) {
1937 mlog(0, "failed to send recovery message. "
1938 "must retry with new node map.\n");
1939 goto retry;
1940 }
1941 } else if (ret == DLM_NOTQUEUED) {
1942 mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
1943 dlm->name, dlm->node_num);
1944 /* another node is master. wait on
1945 * reco.new_master != O2NM_INVALID_NODE_NUM */
1946 status = -EEXIST;
1947 }
1948
1949 return status;
1950}
1951
1952static int dlm_send_begin_reco_message(struct dlm_ctxt *dlm, u8 dead_node)
1953{
1954 struct dlm_begin_reco br;
1955 int ret = 0;
1956 struct dlm_node_iter iter;
1957 int nodenum;
1958 int status;
1959
1960 mlog_entry("%u\n", dead_node);
1961
1962 mlog(0, "dead node is %u\n", dead_node);
1963
1964 spin_lock(&dlm->spinlock);
1965 dlm_node_iter_init(dlm->domain_map, &iter);
1966 spin_unlock(&dlm->spinlock);
1967
1968 clear_bit(dead_node, iter.node_map);
1969
1970 memset(&br, 0, sizeof(br));
1971 br.node_idx = dlm->node_num;
1972 br.dead_node = dead_node;
1973
1974 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
1975 ret = 0;
1976 if (nodenum == dead_node) {
1977 mlog(0, "not sending begin reco to dead node "
1978 "%u\n", dead_node);
1979 continue;
1980 }
1981 if (nodenum == dlm->node_num) {
1982 mlog(0, "not sending begin reco to self\n");
1983 continue;
1984 }
1985
1986 ret = -EINVAL;
1987 mlog(0, "attempting to send begin reco msg to %d\n",
1988 nodenum);
1989 ret = o2net_send_message(DLM_BEGIN_RECO_MSG, dlm->key,
1990 &br, sizeof(br), nodenum, &status);
1991 /* negative status is handled ok by caller here */
1992 if (ret >= 0)
1993 ret = status;
1994 if (ret < 0) {
1995 struct dlm_lock_resource *res;
1996 mlog_errno(ret);
1997 mlog(ML_ERROR, "begin reco of dlm %s to node %u "
1998 " returned %d\n", dlm->name, nodenum, ret);
1999 res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
2000 DLM_RECOVERY_LOCK_NAME_LEN);
2001 if (res) {
2002 dlm_print_one_lock_resource(res);
2003 dlm_lockres_put(res);
2004 } else {
2005 mlog(ML_ERROR, "recovery lock not found\n");
2006 }
2007 break;
2008 }
2009 }
2010
2011 return ret;
2012}
2013
2014int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data)
2015{
2016 struct dlm_ctxt *dlm = data;
2017 struct dlm_begin_reco *br = (struct dlm_begin_reco *)msg->buf;
2018
2019 /* ok to return 0, domain has gone away */
2020 if (!dlm_grab(dlm))
2021 return 0;
2022
2023 mlog(0, "node %u wants to recover node %u\n",
2024 br->node_idx, br->dead_node);
2025
2026 dlm_fire_domain_eviction_callbacks(dlm, br->dead_node);
2027
2028 spin_lock(&dlm->spinlock);
2029 if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
2030 mlog(0, "new_master already set to %u!\n",
2031 dlm->reco.new_master);
2032 }
2033 if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
2034 mlog(0, "dead_node already set to %u!\n",
2035 dlm->reco.dead_node);
2036 }
2037 dlm->reco.new_master = br->node_idx;
2038 dlm->reco.dead_node = br->dead_node;
2039 if (!test_bit(br->dead_node, dlm->recovery_map)) {
2040 mlog(ML_ERROR, "recovery master %u sees %u as dead, but this "
2041 "node has not yet. marking %u as dead\n",
2042 br->node_idx, br->dead_node, br->dead_node);
2043 __dlm_hb_node_down(dlm, br->dead_node);
2044 }
2045 spin_unlock(&dlm->spinlock);
2046
2047 dlm_kick_recovery_thread(dlm);
2048 dlm_put(dlm);
2049 return 0;
2050}
2051
2052static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
2053{
2054 int ret = 0;
2055 struct dlm_finalize_reco fr;
2056 struct dlm_node_iter iter;
2057 int nodenum;
2058 int status;
2059
2060 mlog(0, "finishing recovery for node %s:%u\n",
2061 dlm->name, dlm->reco.dead_node);
2062
2063 spin_lock(&dlm->spinlock);
2064 dlm_node_iter_init(dlm->domain_map, &iter);
2065 spin_unlock(&dlm->spinlock);
2066
2067 memset(&fr, 0, sizeof(fr));
2068 fr.node_idx = dlm->node_num;
2069 fr.dead_node = dlm->reco.dead_node;
2070
2071 while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
2072 if (nodenum == dlm->node_num)
2073 continue;
2074 ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
2075 &fr, sizeof(fr), nodenum, &status);
2076 if (ret >= 0) {
2077 ret = status;
2078 if (dlm_is_host_down(ret)) {
2079 /* this has no effect on this recovery
2080 * session, so set the status to zero to
2081 * finish out the last recovery */
2082 mlog(ML_ERROR, "node %u went down after this "
2083 "node finished recovery.\n", nodenum);
2084 ret = 0;
2085 }
2086 }
2087 if (ret < 0) {
2088 mlog_errno(ret);
2089 break;
2090 }
2091 }
2092
2093 return ret;
2094}
2095
2096int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data)
2097{
2098 struct dlm_ctxt *dlm = data;
2099 struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
2100
2101 /* ok to return 0, domain has gone away */
2102 if (!dlm_grab(dlm))
2103 return 0;
2104
2105 mlog(0, "node %u finalizing recovery of node %u\n",
2106 fr->node_idx, fr->dead_node);
2107
2108 spin_lock(&dlm->spinlock);
2109
2110 if (dlm->reco.new_master != fr->node_idx) {
2111 mlog(ML_ERROR, "node %u sent recovery finalize msg, but node "
2112 "%u is supposed to be the new master, dead=%u\n",
2113 fr->node_idx, dlm->reco.new_master, fr->dead_node);
2114 BUG();
2115 }
2116 if (dlm->reco.dead_node != fr->dead_node) {
2117 mlog(ML_ERROR, "node %u sent recovery finalize msg for dead "
2118 "node %u, but node %u is supposed to be dead\n",
2119 fr->node_idx, fr->dead_node, dlm->reco.dead_node);
2120 BUG();
2121 }
2122
2123 dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
2124
2125 spin_unlock(&dlm->spinlock);
2126
2127 dlm_reset_recovery(dlm);
2128
2129 dlm_kick_recovery_thread(dlm);
2130 dlm_put(dlm);
2131 return 0;
2132}
diff --git a/fs/ocfs2/dlm/dlmthread.c b/fs/ocfs2/dlm/dlmthread.c
new file mode 100644
index 000000000000..5be9d14f12cb
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmthread.c
@@ -0,0 +1,692 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmthread.c
5 *
6 * standalone DLM module
7 *
8 * Copyright (C) 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 *
25 */
26
27
28#include <linux/module.h>
29#include <linux/fs.h>
30#include <linux/types.h>
31#include <linux/slab.h>
32#include <linux/highmem.h>
33#include <linux/utsname.h>
34#include <linux/init.h>
35#include <linux/sysctl.h>
36#include <linux/random.h>
37#include <linux/blkdev.h>
38#include <linux/socket.h>
39#include <linux/inet.h>
40#include <linux/timer.h>
41#include <linux/kthread.h>
42
43
44#include "cluster/heartbeat.h"
45#include "cluster/nodemanager.h"
46#include "cluster/tcp.h"
47
48#include "dlmapi.h"
49#include "dlmcommon.h"
50#include "dlmdomain.h"
51
52#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
53#include "cluster/masklog.h"
54
55static int dlm_thread(void *data);
56
57static void dlm_flush_asts(struct dlm_ctxt *dlm);
58
59#define dlm_lock_is_remote(dlm, lock) ((lock)->ml.node != (dlm)->node_num)
60
61/* will exit holding res->spinlock, but may drop in function */
62/* waits until flags are cleared on res->state */
63void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
64{
65 DECLARE_WAITQUEUE(wait, current);
66
67 assert_spin_locked(&res->spinlock);
68
69 add_wait_queue(&res->wq, &wait);
70repeat:
71 set_current_state(TASK_UNINTERRUPTIBLE);
72 if (res->state & flags) {
73 spin_unlock(&res->spinlock);
74 schedule();
75 spin_lock(&res->spinlock);
76 goto repeat;
77 }
78 remove_wait_queue(&res->wq, &wait);
79 current->state = TASK_RUNNING;
80}
81
82
83static int __dlm_lockres_unused(struct dlm_lock_resource *res)
84{
85 if (list_empty(&res->granted) &&
86 list_empty(&res->converting) &&
87 list_empty(&res->blocked) &&
88 list_empty(&res->dirty))
89 return 1;
90 return 0;
91}
92
93
94/* Call whenever you may have added or deleted something from one of
95 * the lockres queue's. This will figure out whether it belongs on the
96 * unused list or not and does the appropriate thing. */
97void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
98 struct dlm_lock_resource *res)
99{
100 mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
101
102 assert_spin_locked(&dlm->spinlock);
103 assert_spin_locked(&res->spinlock);
104
105 if (__dlm_lockres_unused(res)){
106 if (list_empty(&res->purge)) {
107 mlog(0, "putting lockres %.*s from purge list\n",
108 res->lockname.len, res->lockname.name);
109
110 res->last_used = jiffies;
111 list_add_tail(&res->purge, &dlm->purge_list);
112 dlm->purge_count++;
113 }
114 } else if (!list_empty(&res->purge)) {
115 mlog(0, "removing lockres %.*s from purge list\n",
116 res->lockname.len, res->lockname.name);
117
118 list_del_init(&res->purge);
119 dlm->purge_count--;
120 }
121}
122
123void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
124 struct dlm_lock_resource *res)
125{
126 mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
127 spin_lock(&dlm->spinlock);
128 spin_lock(&res->spinlock);
129
130 __dlm_lockres_calc_usage(dlm, res);
131
132 spin_unlock(&res->spinlock);
133 spin_unlock(&dlm->spinlock);
134}
135
136/* TODO: Eventual API: Called with the dlm spinlock held, may drop it
137 * to do migration, but will re-acquire before exit. */
138void dlm_purge_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *lockres)
139{
140 int master;
141 int ret;
142
143 spin_lock(&lockres->spinlock);
144 master = lockres->owner == dlm->node_num;
145 spin_unlock(&lockres->spinlock);
146
147 mlog(0, "purging lockres %.*s, master = %d\n", lockres->lockname.len,
148 lockres->lockname.name, master);
149
150 /* Non master is the easy case -- no migration required, just
151 * quit. */
152 if (!master)
153 goto finish;
154
155 /* Wheee! Migrate lockres here! */
156 spin_unlock(&dlm->spinlock);
157again:
158
159 ret = dlm_migrate_lockres(dlm, lockres, O2NM_MAX_NODES);
160 if (ret == -ENOTEMPTY) {
161 mlog(ML_ERROR, "lockres %.*s still has local locks!\n",
162 lockres->lockname.len, lockres->lockname.name);
163
164 BUG();
165 } else if (ret < 0) {
166 mlog(ML_NOTICE, "lockres %.*s: migrate failed, retrying\n",
167 lockres->lockname.len, lockres->lockname.name);
168 goto again;
169 }
170
171 spin_lock(&dlm->spinlock);
172
173finish:
174 if (!list_empty(&lockres->purge)) {
175 list_del_init(&lockres->purge);
176 dlm->purge_count--;
177 }
178 __dlm_unhash_lockres(lockres);
179}
180
181static void dlm_run_purge_list(struct dlm_ctxt *dlm,
182 int purge_now)
183{
184 unsigned int run_max, unused;
185 unsigned long purge_jiffies;
186 struct dlm_lock_resource *lockres;
187
188 spin_lock(&dlm->spinlock);
189 run_max = dlm->purge_count;
190
191 while(run_max && !list_empty(&dlm->purge_list)) {
192 run_max--;
193
194 lockres = list_entry(dlm->purge_list.next,
195 struct dlm_lock_resource, purge);
196
197 /* Status of the lockres *might* change so double
198 * check. If the lockres is unused, holding the dlm
199 * spinlock will prevent people from getting and more
200 * refs on it -- there's no need to keep the lockres
201 * spinlock. */
202 spin_lock(&lockres->spinlock);
203 unused = __dlm_lockres_unused(lockres);
204 spin_unlock(&lockres->spinlock);
205
206 if (!unused)
207 continue;
208
209 purge_jiffies = lockres->last_used +
210 msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
211
212 /* Make sure that we want to be processing this guy at
213 * this time. */
214 if (!purge_now && time_after(purge_jiffies, jiffies)) {
215 /* Since resources are added to the purge list
216 * in tail order, we can stop at the first
217 * unpurgable resource -- anyone added after
218 * him will have a greater last_used value */
219 break;
220 }
221
222 list_del_init(&lockres->purge);
223 dlm->purge_count--;
224
225 /* This may drop and reacquire the dlm spinlock if it
226 * has to do migration. */
227 mlog(0, "calling dlm_purge_lockres!\n");
228 dlm_purge_lockres(dlm, lockres);
229 mlog(0, "DONE calling dlm_purge_lockres!\n");
230
231 /* Avoid adding any scheduling latencies */
232 cond_resched_lock(&dlm->spinlock);
233 }
234
235 spin_unlock(&dlm->spinlock);
236}
237
238static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
239 struct dlm_lock_resource *res)
240{
241 struct dlm_lock *lock, *target;
242 struct list_head *iter;
243 struct list_head *head;
244 int can_grant = 1;
245
246 //mlog(0, "res->lockname.len=%d\n", res->lockname.len);
247 //mlog(0, "res->lockname.name=%p\n", res->lockname.name);
248 //mlog(0, "shuffle res %.*s\n", res->lockname.len,
249 // res->lockname.name);
250
251 /* because this function is called with the lockres
252 * spinlock, and because we know that it is not migrating/
253 * recovering/in-progress, it is fine to reserve asts and
254 * basts right before queueing them all throughout */
255 assert_spin_locked(&res->spinlock);
256 BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
257 DLM_LOCK_RES_RECOVERING|
258 DLM_LOCK_RES_IN_PROGRESS)));
259
260converting:
261 if (list_empty(&res->converting))
262 goto blocked;
263 mlog(0, "res %.*s has locks on a convert queue\n", res->lockname.len,
264 res->lockname.name);
265
266 target = list_entry(res->converting.next, struct dlm_lock, list);
267 if (target->ml.convert_type == LKM_IVMODE) {
268 mlog(ML_ERROR, "%.*s: converting a lock with no "
269 "convert_type!\n", res->lockname.len, res->lockname.name);
270 BUG();
271 }
272 head = &res->granted;
273 list_for_each(iter, head) {
274 lock = list_entry(iter, struct dlm_lock, list);
275 if (lock==target)
276 continue;
277 if (!dlm_lock_compatible(lock->ml.type,
278 target->ml.convert_type)) {
279 can_grant = 0;
280 /* queue the BAST if not already */
281 if (lock->ml.highest_blocked == LKM_IVMODE) {
282 __dlm_lockres_reserve_ast(res);
283 dlm_queue_bast(dlm, lock);
284 }
285 /* update the highest_blocked if needed */
286 if (lock->ml.highest_blocked < target->ml.convert_type)
287 lock->ml.highest_blocked =
288 target->ml.convert_type;
289 }
290 }
291 head = &res->converting;
292 list_for_each(iter, head) {
293 lock = list_entry(iter, struct dlm_lock, list);
294 if (lock==target)
295 continue;
296 if (!dlm_lock_compatible(lock->ml.type,
297 target->ml.convert_type)) {
298 can_grant = 0;
299 if (lock->ml.highest_blocked == LKM_IVMODE) {
300 __dlm_lockres_reserve_ast(res);
301 dlm_queue_bast(dlm, lock);
302 }
303 if (lock->ml.highest_blocked < target->ml.convert_type)
304 lock->ml.highest_blocked =
305 target->ml.convert_type;
306 }
307 }
308
309 /* we can convert the lock */
310 if (can_grant) {
311 spin_lock(&target->spinlock);
312 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
313
314 mlog(0, "calling ast for converting lock: %.*s, have: %d, "
315 "granting: %d, node: %u\n", res->lockname.len,
316 res->lockname.name, target->ml.type,
317 target->ml.convert_type, target->ml.node);
318
319 target->ml.type = target->ml.convert_type;
320 target->ml.convert_type = LKM_IVMODE;
321 list_del_init(&target->list);
322 list_add_tail(&target->list, &res->granted);
323
324 BUG_ON(!target->lksb);
325 target->lksb->status = DLM_NORMAL;
326
327 spin_unlock(&target->spinlock);
328
329 __dlm_lockres_reserve_ast(res);
330 dlm_queue_ast(dlm, target);
331 /* go back and check for more */
332 goto converting;
333 }
334
335blocked:
336 if (list_empty(&res->blocked))
337 goto leave;
338 target = list_entry(res->blocked.next, struct dlm_lock, list);
339
340 head = &res->granted;
341 list_for_each(iter, head) {
342 lock = list_entry(iter, struct dlm_lock, list);
343 if (lock==target)
344 continue;
345 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
346 can_grant = 0;
347 if (lock->ml.highest_blocked == LKM_IVMODE) {
348 __dlm_lockres_reserve_ast(res);
349 dlm_queue_bast(dlm, lock);
350 }
351 if (lock->ml.highest_blocked < target->ml.type)
352 lock->ml.highest_blocked = target->ml.type;
353 }
354 }
355
356 head = &res->converting;
357 list_for_each(iter, head) {
358 lock = list_entry(iter, struct dlm_lock, list);
359 if (lock==target)
360 continue;
361 if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
362 can_grant = 0;
363 if (lock->ml.highest_blocked == LKM_IVMODE) {
364 __dlm_lockres_reserve_ast(res);
365 dlm_queue_bast(dlm, lock);
366 }
367 if (lock->ml.highest_blocked < target->ml.type)
368 lock->ml.highest_blocked = target->ml.type;
369 }
370 }
371
372 /* we can grant the blocked lock (only
373 * possible if converting list empty) */
374 if (can_grant) {
375 spin_lock(&target->spinlock);
376 BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
377
378 mlog(0, "calling ast for blocked lock: %.*s, granting: %d, "
379 "node: %u\n", res->lockname.len, res->lockname.name,
380 target->ml.type, target->ml.node);
381
382 // target->ml.type is already correct
383 list_del_init(&target->list);
384 list_add_tail(&target->list, &res->granted);
385
386 BUG_ON(!target->lksb);
387 target->lksb->status = DLM_NORMAL;
388
389 spin_unlock(&target->spinlock);
390
391 __dlm_lockres_reserve_ast(res);
392 dlm_queue_ast(dlm, target);
393 /* go back and check for more */
394 goto converting;
395 }
396
397leave:
398 return;
399}
400
401/* must have NO locks when calling this with res !=NULL * */
402void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
403{
404 mlog_entry("dlm=%p, res=%p\n", dlm, res);
405 if (res) {
406 spin_lock(&dlm->spinlock);
407 spin_lock(&res->spinlock);
408 __dlm_dirty_lockres(dlm, res);
409 spin_unlock(&res->spinlock);
410 spin_unlock(&dlm->spinlock);
411 }
412 wake_up(&dlm->dlm_thread_wq);
413}
414
415void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
416{
417 mlog_entry("dlm=%p, res=%p\n", dlm, res);
418
419 assert_spin_locked(&dlm->spinlock);
420 assert_spin_locked(&res->spinlock);
421
422 /* don't shuffle secondary queues */
423 if ((res->owner == dlm->node_num) &&
424 !(res->state & DLM_LOCK_RES_DIRTY)) {
425 list_add_tail(&res->dirty, &dlm->dirty_list);
426 res->state |= DLM_LOCK_RES_DIRTY;
427 }
428}
429
430
431/* Launch the NM thread for the mounted volume */
432int dlm_launch_thread(struct dlm_ctxt *dlm)
433{
434 mlog(0, "starting dlm thread...\n");
435
436 dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
437 if (IS_ERR(dlm->dlm_thread_task)) {
438 mlog_errno(PTR_ERR(dlm->dlm_thread_task));
439 dlm->dlm_thread_task = NULL;
440 return -EINVAL;
441 }
442
443 return 0;
444}
445
446void dlm_complete_thread(struct dlm_ctxt *dlm)
447{
448 if (dlm->dlm_thread_task) {
449 mlog(ML_KTHREAD, "waiting for dlm thread to exit\n");
450 kthread_stop(dlm->dlm_thread_task);
451 dlm->dlm_thread_task = NULL;
452 }
453}
454
455static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
456{
457 int empty;
458
459 spin_lock(&dlm->spinlock);
460 empty = list_empty(&dlm->dirty_list);
461 spin_unlock(&dlm->spinlock);
462
463 return empty;
464}
465
466static void dlm_flush_asts(struct dlm_ctxt *dlm)
467{
468 int ret;
469 struct dlm_lock *lock;
470 struct dlm_lock_resource *res;
471 u8 hi;
472
473 spin_lock(&dlm->ast_lock);
474 while (!list_empty(&dlm->pending_asts)) {
475 lock = list_entry(dlm->pending_asts.next,
476 struct dlm_lock, ast_list);
477 /* get an extra ref on lock */
478 dlm_lock_get(lock);
479 res = lock->lockres;
480 mlog(0, "delivering an ast for this lockres\n");
481
482 BUG_ON(!lock->ast_pending);
483
484 /* remove from list (including ref) */
485 list_del_init(&lock->ast_list);
486 dlm_lock_put(lock);
487 spin_unlock(&dlm->ast_lock);
488
489 if (lock->ml.node != dlm->node_num) {
490 ret = dlm_do_remote_ast(dlm, res, lock);
491 if (ret < 0)
492 mlog_errno(ret);
493 } else
494 dlm_do_local_ast(dlm, res, lock);
495
496 spin_lock(&dlm->ast_lock);
497
498 /* possible that another ast was queued while
499 * we were delivering the last one */
500 if (!list_empty(&lock->ast_list)) {
501 mlog(0, "aha another ast got queued while "
502 "we were finishing the last one. will "
503 "keep the ast_pending flag set.\n");
504 } else
505 lock->ast_pending = 0;
506
507 /* drop the extra ref.
508 * this may drop it completely. */
509 dlm_lock_put(lock);
510 dlm_lockres_release_ast(dlm, res);
511 }
512
513 while (!list_empty(&dlm->pending_basts)) {
514 lock = list_entry(dlm->pending_basts.next,
515 struct dlm_lock, bast_list);
516 /* get an extra ref on lock */
517 dlm_lock_get(lock);
518 res = lock->lockres;
519
520 BUG_ON(!lock->bast_pending);
521
522 /* get the highest blocked lock, and reset */
523 spin_lock(&lock->spinlock);
524 BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
525 hi = lock->ml.highest_blocked;
526 lock->ml.highest_blocked = LKM_IVMODE;
527 spin_unlock(&lock->spinlock);
528
529 /* remove from list (including ref) */
530 list_del_init(&lock->bast_list);
531 dlm_lock_put(lock);
532 spin_unlock(&dlm->ast_lock);
533
534 mlog(0, "delivering a bast for this lockres "
535 "(blocked = %d\n", hi);
536
537 if (lock->ml.node != dlm->node_num) {
538 ret = dlm_send_proxy_bast(dlm, res, lock, hi);
539 if (ret < 0)
540 mlog_errno(ret);
541 } else
542 dlm_do_local_bast(dlm, res, lock, hi);
543
544 spin_lock(&dlm->ast_lock);
545
546 /* possible that another bast was queued while
547 * we were delivering the last one */
548 if (!list_empty(&lock->bast_list)) {
549 mlog(0, "aha another bast got queued while "
550 "we were finishing the last one. will "
551 "keep the bast_pending flag set.\n");
552 } else
553 lock->bast_pending = 0;
554
555 /* drop the extra ref.
556 * this may drop it completely. */
557 dlm_lock_put(lock);
558 dlm_lockres_release_ast(dlm, res);
559 }
560 wake_up(&dlm->ast_wq);
561 spin_unlock(&dlm->ast_lock);
562}
563
564
565#define DLM_THREAD_TIMEOUT_MS (4 * 1000)
566#define DLM_THREAD_MAX_DIRTY 100
567#define DLM_THREAD_MAX_ASTS 10
568
569static int dlm_thread(void *data)
570{
571 struct dlm_lock_resource *res;
572 struct dlm_ctxt *dlm = data;
573 unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
574
575 mlog(0, "dlm thread running for %s...\n", dlm->name);
576
577 while (!kthread_should_stop()) {
578 int n = DLM_THREAD_MAX_DIRTY;
579
580 /* dlm_shutting_down is very point-in-time, but that
581 * doesn't matter as we'll just loop back around if we
582 * get false on the leading edge of a state
583 * transition. */
584 dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
585
586 /* We really don't want to hold dlm->spinlock while
587 * calling dlm_shuffle_lists on each lockres that
588 * needs to have its queues adjusted and AST/BASTs
589 * run. So let's pull each entry off the dirty_list
590 * and drop dlm->spinlock ASAP. Once off the list,
591 * res->spinlock needs to be taken again to protect
592 * the queues while calling dlm_shuffle_lists. */
593 spin_lock(&dlm->spinlock);
594 while (!list_empty(&dlm->dirty_list)) {
595 int delay = 0;
596 res = list_entry(dlm->dirty_list.next,
597 struct dlm_lock_resource, dirty);
598
599 /* peel a lockres off, remove it from the list,
600 * unset the dirty flag and drop the dlm lock */
601 BUG_ON(!res);
602 dlm_lockres_get(res);
603
604 spin_lock(&res->spinlock);
605 res->state &= ~DLM_LOCK_RES_DIRTY;
606 list_del_init(&res->dirty);
607 spin_unlock(&res->spinlock);
608 spin_unlock(&dlm->spinlock);
609
610 /* lockres can be re-dirtied/re-added to the
611 * dirty_list in this gap, but that is ok */
612
613 spin_lock(&res->spinlock);
614 if (res->owner != dlm->node_num) {
615 __dlm_print_one_lock_resource(res);
616 mlog(ML_ERROR, "inprog:%s, mig:%s, reco:%s, dirty:%s\n",
617 res->state & DLM_LOCK_RES_IN_PROGRESS ? "yes" : "no",
618 res->state & DLM_LOCK_RES_MIGRATING ? "yes" : "no",
619 res->state & DLM_LOCK_RES_RECOVERING ? "yes" : "no",
620 res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
621 }
622 BUG_ON(res->owner != dlm->node_num);
623
624 /* it is now ok to move lockreses in these states
625 * to the dirty list, assuming that they will only be
626 * dirty for a short while. */
627 if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
628 DLM_LOCK_RES_MIGRATING |
629 DLM_LOCK_RES_RECOVERING)) {
630 /* move it to the tail and keep going */
631 spin_unlock(&res->spinlock);
632 mlog(0, "delaying list shuffling for in-"
633 "progress lockres %.*s, state=%d\n",
634 res->lockname.len, res->lockname.name,
635 res->state);
636 delay = 1;
637 goto in_progress;
638 }
639
640 /* at this point the lockres is not migrating/
641 * recovering/in-progress. we have the lockres
642 * spinlock and do NOT have the dlm lock.
643 * safe to reserve/queue asts and run the lists. */
644
645 mlog(0, "calling dlm_shuffle_lists with dlm=%p, "
646 "res=%p\n", dlm, res);
647
648 /* called while holding lockres lock */
649 dlm_shuffle_lists(dlm, res);
650 spin_unlock(&res->spinlock);
651
652 dlm_lockres_calc_usage(dlm, res);
653
654in_progress:
655
656 spin_lock(&dlm->spinlock);
657 /* if the lock was in-progress, stick
658 * it on the back of the list */
659 if (delay) {
660 spin_lock(&res->spinlock);
661 list_add_tail(&res->dirty, &dlm->dirty_list);
662 res->state |= DLM_LOCK_RES_DIRTY;
663 spin_unlock(&res->spinlock);
664 }
665 dlm_lockres_put(res);
666
667 /* unlikely, but we may need to give time to
668 * other tasks */
669 if (!--n) {
670 mlog(0, "throttling dlm_thread\n");
671 break;
672 }
673 }
674
675 spin_unlock(&dlm->spinlock);
676 dlm_flush_asts(dlm);
677
678 /* yield and continue right away if there is more work to do */
679 if (!n) {
680 yield();
681 continue;
682 }
683
684 wait_event_interruptible_timeout(dlm->dlm_thread_wq,
685 !dlm_dirty_list_empty(dlm) ||
686 kthread_should_stop(),
687 timeout);
688 }
689
690 mlog(0, "quitting DLM thread\n");
691 return 0;
692}
diff --git a/fs/ocfs2/dlm/dlmunlock.c b/fs/ocfs2/dlm/dlmunlock.c
new file mode 100644
index 000000000000..cec2ce1cd318
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmunlock.c
@@ -0,0 +1,672 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmunlock.c
5 *
6 * underlying calls for unlocking locks
7 *
8 * Copyright (C) 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 *
25 */
26
27
28#include <linux/module.h>
29#include <linux/fs.h>
30#include <linux/types.h>
31#include <linux/slab.h>
32#include <linux/highmem.h>
33#include <linux/utsname.h>
34#include <linux/init.h>
35#include <linux/sysctl.h>
36#include <linux/random.h>
37#include <linux/blkdev.h>
38#include <linux/socket.h>
39#include <linux/inet.h>
40#include <linux/spinlock.h>
41#include <linux/delay.h>
42
43#include "cluster/heartbeat.h"
44#include "cluster/nodemanager.h"
45#include "cluster/tcp.h"
46
47#include "dlmapi.h"
48#include "dlmcommon.h"
49
50#define MLOG_MASK_PREFIX ML_DLM
51#include "cluster/masklog.h"
52
53#define DLM_UNLOCK_FREE_LOCK 0x00000001
54#define DLM_UNLOCK_CALL_AST 0x00000002
55#define DLM_UNLOCK_REMOVE_LOCK 0x00000004
56#define DLM_UNLOCK_REGRANT_LOCK 0x00000008
57#define DLM_UNLOCK_CLEAR_CONVERT_TYPE 0x00000010
58
59
60static enum dlm_status dlm_get_cancel_actions(struct dlm_ctxt *dlm,
61 struct dlm_lock_resource *res,
62 struct dlm_lock *lock,
63 struct dlm_lockstatus *lksb,
64 int *actions);
65static enum dlm_status dlm_get_unlock_actions(struct dlm_ctxt *dlm,
66 struct dlm_lock_resource *res,
67 struct dlm_lock *lock,
68 struct dlm_lockstatus *lksb,
69 int *actions);
70
71static enum dlm_status dlm_send_remote_unlock_request(struct dlm_ctxt *dlm,
72 struct dlm_lock_resource *res,
73 struct dlm_lock *lock,
74 struct dlm_lockstatus *lksb,
75 int flags,
76 u8 owner);
77
78
79/*
80 * according to the spec:
81 * http://opendlm.sourceforge.net/cvsmirror/opendlm/docs/dlmbook_final.pdf
82 *
83 * flags & LKM_CANCEL != 0: must be converting or blocked
84 * flags & LKM_CANCEL == 0: must be granted
85 *
86 * So to unlock a converting lock, you must first cancel the
87 * convert (passing LKM_CANCEL in flags), then call the unlock
88 * again (with no LKM_CANCEL in flags).
89 */
90
91
92/*
93 * locking:
94 * caller needs: none
95 * taken: res->spinlock and lock->spinlock taken and dropped
96 * held on exit: none
97 * returns: DLM_NORMAL, DLM_NOLOCKMGR, status from network
98 * all callers should have taken an extra ref on lock coming in
99 */
100static enum dlm_status dlmunlock_common(struct dlm_ctxt *dlm,
101 struct dlm_lock_resource *res,
102 struct dlm_lock *lock,
103 struct dlm_lockstatus *lksb,
104 int flags, int *call_ast,
105 int master_node)
106{
107 enum dlm_status status;
108 int actions = 0;
109 int in_use;
110 u8 owner;
111
112 mlog(0, "master_node = %d, valblk = %d\n", master_node,
113 flags & LKM_VALBLK);
114
115 if (master_node)
116 BUG_ON(res->owner != dlm->node_num);
117 else
118 BUG_ON(res->owner == dlm->node_num);
119
120 spin_lock(&dlm->spinlock);
121 /* We want to be sure that we're not freeing a lock
122 * that still has AST's pending... */
123 in_use = !list_empty(&lock->ast_list);
124 spin_unlock(&dlm->spinlock);
125 if (in_use) {
126 mlog(ML_ERROR, "lockres %.*s: Someone is calling dlmunlock "
127 "while waiting for an ast!", res->lockname.len,
128 res->lockname.name);
129 return DLM_BADPARAM;
130 }
131
132 spin_lock(&res->spinlock);
133 if (res->state & DLM_LOCK_RES_IN_PROGRESS) {
134 if (master_node) {
135 mlog(ML_ERROR, "lockres in progress!\n");
136 spin_unlock(&res->spinlock);
137 return DLM_FORWARD;
138 }
139 /* ok for this to sleep if not in a network handler */
140 __dlm_wait_on_lockres(res);
141 res->state |= DLM_LOCK_RES_IN_PROGRESS;
142 }
143 spin_lock(&lock->spinlock);
144
145 if (res->state & DLM_LOCK_RES_RECOVERING) {
146 status = DLM_RECOVERING;
147 goto leave;
148 }
149
150
151 /* see above for what the spec says about
152 * LKM_CANCEL and the lock queue state */
153 if (flags & LKM_CANCEL)
154 status = dlm_get_cancel_actions(dlm, res, lock, lksb, &actions);
155 else
156 status = dlm_get_unlock_actions(dlm, res, lock, lksb, &actions);
157
158 if (status != DLM_NORMAL)
159 goto leave;
160
161 /* By now this has been masked out of cancel requests. */
162 if (flags & LKM_VALBLK) {
163 /* make the final update to the lvb */
164 if (master_node)
165 memcpy(res->lvb, lksb->lvb, DLM_LVB_LEN);
166 else
167 flags |= LKM_PUT_LVB; /* let the send function
168 * handle it. */
169 }
170
171 if (!master_node) {
172 owner = res->owner;
173 /* drop locks and send message */
174 if (flags & LKM_CANCEL)
175 lock->cancel_pending = 1;
176 else
177 lock->unlock_pending = 1;
178 spin_unlock(&lock->spinlock);
179 spin_unlock(&res->spinlock);
180 status = dlm_send_remote_unlock_request(dlm, res, lock, lksb,
181 flags, owner);
182 spin_lock(&res->spinlock);
183 spin_lock(&lock->spinlock);
184 /* if the master told us the lock was already granted,
185 * let the ast handle all of these actions */
186 if (status == DLM_NORMAL &&
187 lksb->status == DLM_CANCELGRANT) {
188 actions &= ~(DLM_UNLOCK_REMOVE_LOCK|
189 DLM_UNLOCK_REGRANT_LOCK|
190 DLM_UNLOCK_CLEAR_CONVERT_TYPE);
191 }
192 if (flags & LKM_CANCEL)
193 lock->cancel_pending = 0;
194 else
195 lock->unlock_pending = 0;
196
197 }
198
199 /* get an extra ref on lock. if we are just switching
200 * lists here, we dont want the lock to go away. */
201 dlm_lock_get(lock);
202
203 if (actions & DLM_UNLOCK_REMOVE_LOCK) {
204 list_del_init(&lock->list);
205 dlm_lock_put(lock);
206 }
207 if (actions & DLM_UNLOCK_REGRANT_LOCK) {
208 dlm_lock_get(lock);
209 list_add_tail(&lock->list, &res->granted);
210 }
211 if (actions & DLM_UNLOCK_CLEAR_CONVERT_TYPE) {
212 mlog(0, "clearing convert_type at %smaster node\n",
213 master_node ? "" : "non-");
214 lock->ml.convert_type = LKM_IVMODE;
215 }
216
217 /* remove the extra ref on lock */
218 dlm_lock_put(lock);
219
220leave:
221 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
222 if (!dlm_lock_on_list(&res->converting, lock))
223 BUG_ON(lock->ml.convert_type != LKM_IVMODE);
224 else
225 BUG_ON(lock->ml.convert_type == LKM_IVMODE);
226 spin_unlock(&lock->spinlock);
227 spin_unlock(&res->spinlock);
228 wake_up(&res->wq);
229
230 /* let the caller's final dlm_lock_put handle the actual kfree */
231 if (actions & DLM_UNLOCK_FREE_LOCK) {
232 /* this should always be coupled with list removal */
233 BUG_ON(!(actions & DLM_UNLOCK_REMOVE_LOCK));
234 mlog(0, "lock %"MLFu64" should be gone now! refs=%d\n",
235 lock->ml.cookie, atomic_read(&lock->lock_refs.refcount)-1);
236 dlm_lock_put(lock);
237 }
238 if (actions & DLM_UNLOCK_CALL_AST)
239 *call_ast = 1;
240
241 /* if cancel or unlock succeeded, lvb work is done */
242 if (status == DLM_NORMAL)
243 lksb->flags &= ~(DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB);
244
245 return status;
246}
247
248void dlm_commit_pending_unlock(struct dlm_lock_resource *res,
249 struct dlm_lock *lock)
250{
251 /* leave DLM_LKSB_PUT_LVB on the lksb so any final
252 * update of the lvb will be sent to the new master */
253 list_del_init(&lock->list);
254}
255
256void dlm_commit_pending_cancel(struct dlm_lock_resource *res,
257 struct dlm_lock *lock)
258{
259 list_del_init(&lock->list);
260 list_add_tail(&lock->list, &res->granted);
261 lock->ml.convert_type = LKM_IVMODE;
262}
263
264
265static inline enum dlm_status dlmunlock_master(struct dlm_ctxt *dlm,
266 struct dlm_lock_resource *res,
267 struct dlm_lock *lock,
268 struct dlm_lockstatus *lksb,
269 int flags,
270 int *call_ast)
271{
272 return dlmunlock_common(dlm, res, lock, lksb, flags, call_ast, 1);
273}
274
275static inline enum dlm_status dlmunlock_remote(struct dlm_ctxt *dlm,
276 struct dlm_lock_resource *res,
277 struct dlm_lock *lock,
278 struct dlm_lockstatus *lksb,
279 int flags, int *call_ast)
280{
281 return dlmunlock_common(dlm, res, lock, lksb, flags, call_ast, 0);
282}
283
284/*
285 * locking:
286 * caller needs: none
287 * taken: none
288 * held on exit: none
289 * returns: DLM_NORMAL, DLM_NOLOCKMGR, status from network
290 */
291static enum dlm_status dlm_send_remote_unlock_request(struct dlm_ctxt *dlm,
292 struct dlm_lock_resource *res,
293 struct dlm_lock *lock,
294 struct dlm_lockstatus *lksb,
295 int flags,
296 u8 owner)
297{
298 struct dlm_unlock_lock unlock;
299 int tmpret;
300 enum dlm_status ret;
301 int status = 0;
302 struct kvec vec[2];
303 size_t veclen = 1;
304
305 mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
306
307 memset(&unlock, 0, sizeof(unlock));
308 unlock.node_idx = dlm->node_num;
309 unlock.flags = cpu_to_be32(flags);
310 unlock.cookie = lock->ml.cookie;
311 unlock.namelen = res->lockname.len;
312 memcpy(unlock.name, res->lockname.name, unlock.namelen);
313
314 vec[0].iov_len = sizeof(struct dlm_unlock_lock);
315 vec[0].iov_base = &unlock;
316
317 if (flags & LKM_PUT_LVB) {
318 /* extra data to send if we are updating lvb */
319 vec[1].iov_len = DLM_LVB_LEN;
320 vec[1].iov_base = lock->lksb->lvb;
321 veclen++;
322 }
323
324 tmpret = o2net_send_message_vec(DLM_UNLOCK_LOCK_MSG, dlm->key,
325 vec, veclen, owner, &status);
326 if (tmpret >= 0) {
327 // successfully sent and received
328 if (status == DLM_CANCELGRANT)
329 ret = DLM_NORMAL;
330 else if (status == DLM_FORWARD) {
331 mlog(0, "master was in-progress. retry\n");
332 ret = DLM_FORWARD;
333 } else
334 ret = status;
335 lksb->status = status;
336 } else {
337 mlog_errno(tmpret);
338 if (dlm_is_host_down(tmpret)) {
339 /* NOTE: this seems strange, but it is what we want.
340 * when the master goes down during a cancel or
341 * unlock, the recovery code completes the operation
342 * as if the master had not died, then passes the
343 * updated state to the recovery master. this thread
344 * just needs to finish out the operation and call
345 * the unlockast. */
346 ret = DLM_NORMAL;
347 } else {
348 /* something bad. this will BUG in ocfs2 */
349 ret = dlm_err_to_dlm_status(tmpret);
350 }
351 lksb->status = ret;
352 }
353
354 return ret;
355}
356
357/*
358 * locking:
359 * caller needs: none
360 * taken: takes and drops res->spinlock
361 * held on exit: none
362 * returns: DLM_NORMAL, DLM_BADARGS, DLM_IVLOCKID,
363 * return value from dlmunlock_master
364 */
365int dlm_unlock_lock_handler(struct o2net_msg *msg, u32 len, void *data)
366{
367 struct dlm_ctxt *dlm = data;
368 struct dlm_unlock_lock *unlock = (struct dlm_unlock_lock *)msg->buf;
369 struct dlm_lock_resource *res = NULL;
370 struct list_head *iter;
371 struct dlm_lock *lock = NULL;
372 enum dlm_status status = DLM_NORMAL;
373 int found = 0, i;
374 struct dlm_lockstatus *lksb = NULL;
375 int ignore;
376 u32 flags;
377 struct list_head *queue;
378
379 flags = be32_to_cpu(unlock->flags);
380
381 if (flags & LKM_GET_LVB) {
382 mlog(ML_ERROR, "bad args! GET_LVB specified on unlock!\n");
383 return DLM_BADARGS;
384 }
385
386 if ((flags & (LKM_PUT_LVB|LKM_CANCEL)) == (LKM_PUT_LVB|LKM_CANCEL)) {
387 mlog(ML_ERROR, "bad args! cannot modify lvb on a CANCEL "
388 "request!\n");
389 return DLM_BADARGS;
390 }
391
392 if (unlock->namelen > DLM_LOCKID_NAME_MAX) {
393 mlog(ML_ERROR, "Invalid name length in unlock handler!\n");
394 return DLM_IVBUFLEN;
395 }
396
397 if (!dlm_grab(dlm))
398 return DLM_REJECTED;
399
400 mlog_bug_on_msg(!dlm_domain_fully_joined(dlm),
401 "Domain %s not fully joined!\n", dlm->name);
402
403 mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" : "none");
404
405 res = dlm_lookup_lockres(dlm, unlock->name, unlock->namelen);
406 if (!res) {
407 /* We assume here that a no lock resource simply means
408 * it was migrated away and destroyed before the other
409 * node could detect it. */
410 mlog(0, "returning DLM_FORWARD -- res no longer exists\n");
411 status = DLM_FORWARD;
412 goto not_found;
413 }
414
415 queue=&res->granted;
416 found = 0;
417 spin_lock(&res->spinlock);
418 if (res->state & DLM_LOCK_RES_RECOVERING) {
419 spin_unlock(&res->spinlock);
420 mlog(0, "returning DLM_RECOVERING\n");
421 status = DLM_RECOVERING;
422 goto leave;
423 }
424
425 if (res->state & DLM_LOCK_RES_MIGRATING) {
426 spin_unlock(&res->spinlock);
427 mlog(0, "returning DLM_MIGRATING\n");
428 status = DLM_MIGRATING;
429 goto leave;
430 }
431
432 if (res->owner != dlm->node_num) {
433 spin_unlock(&res->spinlock);
434 mlog(0, "returning DLM_FORWARD -- not master\n");
435 status = DLM_FORWARD;
436 goto leave;
437 }
438
439 for (i=0; i<3; i++) {
440 list_for_each(iter, queue) {
441 lock = list_entry(iter, struct dlm_lock, list);
442 if (lock->ml.cookie == unlock->cookie &&
443 lock->ml.node == unlock->node_idx) {
444 dlm_lock_get(lock);
445 found = 1;
446 break;
447 }
448 }
449 if (found)
450 break;
451 /* scan granted -> converting -> blocked queues */
452 queue++;
453 }
454 spin_unlock(&res->spinlock);
455 if (!found) {
456 status = DLM_IVLOCKID;
457 goto not_found;
458 }
459
460 /* lock was found on queue */
461 lksb = lock->lksb;
462 /* unlockast only called on originating node */
463 if (flags & LKM_PUT_LVB) {
464 lksb->flags |= DLM_LKSB_PUT_LVB;
465 memcpy(&lksb->lvb[0], &unlock->lvb[0], DLM_LVB_LEN);
466 }
467
468 /* if this is in-progress, propagate the DLM_FORWARD
469 * all the way back out */
470 status = dlmunlock_master(dlm, res, lock, lksb, flags, &ignore);
471 if (status == DLM_FORWARD)
472 mlog(0, "lockres is in progress\n");
473
474 if (flags & LKM_PUT_LVB)
475 lksb->flags &= ~DLM_LKSB_PUT_LVB;
476
477 dlm_lockres_calc_usage(dlm, res);
478 dlm_kick_thread(dlm, res);
479
480not_found:
481 if (!found)
482 mlog(ML_ERROR, "failed to find lock to unlock! "
483 "cookie=%"MLFu64"\n",
484 unlock->cookie);
485 else {
486 /* send the lksb->status back to the other node */
487 status = lksb->status;
488 dlm_lock_put(lock);
489 }
490
491leave:
492 if (res)
493 dlm_lockres_put(res);
494
495 dlm_put(dlm);
496
497 return status;
498}
499
500
501static enum dlm_status dlm_get_cancel_actions(struct dlm_ctxt *dlm,
502 struct dlm_lock_resource *res,
503 struct dlm_lock *lock,
504 struct dlm_lockstatus *lksb,
505 int *actions)
506{
507 enum dlm_status status;
508
509 if (dlm_lock_on_list(&res->blocked, lock)) {
510 /* cancel this outright */
511 lksb->status = DLM_NORMAL;
512 status = DLM_NORMAL;
513 *actions = (DLM_UNLOCK_CALL_AST |
514 DLM_UNLOCK_REMOVE_LOCK);
515 } else if (dlm_lock_on_list(&res->converting, lock)) {
516 /* cancel the request, put back on granted */
517 lksb->status = DLM_NORMAL;
518 status = DLM_NORMAL;
519 *actions = (DLM_UNLOCK_CALL_AST |
520 DLM_UNLOCK_REMOVE_LOCK |
521 DLM_UNLOCK_REGRANT_LOCK |
522 DLM_UNLOCK_CLEAR_CONVERT_TYPE);
523 } else if (dlm_lock_on_list(&res->granted, lock)) {
524 /* too late, already granted. DLM_CANCELGRANT */
525 lksb->status = DLM_CANCELGRANT;
526 status = DLM_NORMAL;
527 *actions = DLM_UNLOCK_CALL_AST;
528 } else {
529 mlog(ML_ERROR, "lock to cancel is not on any list!\n");
530 lksb->status = DLM_IVLOCKID;
531 status = DLM_IVLOCKID;
532 *actions = 0;
533 }
534 return status;
535}
536
537static enum dlm_status dlm_get_unlock_actions(struct dlm_ctxt *dlm,
538 struct dlm_lock_resource *res,
539 struct dlm_lock *lock,
540 struct dlm_lockstatus *lksb,
541 int *actions)
542{
543 enum dlm_status status;
544
545 /* unlock request */
546 if (!dlm_lock_on_list(&res->granted, lock)) {
547 lksb->status = DLM_DENIED;
548 status = DLM_DENIED;
549 dlm_error(status);
550 *actions = 0;
551 } else {
552 /* unlock granted lock */
553 lksb->status = DLM_NORMAL;
554 status = DLM_NORMAL;
555 *actions = (DLM_UNLOCK_FREE_LOCK |
556 DLM_UNLOCK_CALL_AST |
557 DLM_UNLOCK_REMOVE_LOCK);
558 }
559 return status;
560}
561
562/* there seems to be no point in doing this async
563 * since (even for the remote case) there is really
564 * no work to queue up... so just do it and fire the
565 * unlockast by hand when done... */
566enum dlm_status dlmunlock(struct dlm_ctxt *dlm, struct dlm_lockstatus *lksb,
567 int flags, dlm_astunlockfunc_t *unlockast, void *data)
568{
569 enum dlm_status status;
570 struct dlm_lock_resource *res;
571 struct dlm_lock *lock = NULL;
572 int call_ast, is_master;
573
574 mlog_entry_void();
575
576 if (!lksb) {
577 dlm_error(DLM_BADARGS);
578 return DLM_BADARGS;
579 }
580
581 if (flags & ~(LKM_CANCEL | LKM_VALBLK | LKM_INVVALBLK)) {
582 dlm_error(DLM_BADPARAM);
583 return DLM_BADPARAM;
584 }
585
586 if ((flags & (LKM_VALBLK | LKM_CANCEL)) == (LKM_VALBLK | LKM_CANCEL)) {
587 mlog(0, "VALBLK given with CANCEL: ignoring VALBLK\n");
588 flags &= ~LKM_VALBLK;
589 }
590
591 if (!lksb->lockid || !lksb->lockid->lockres) {
592 dlm_error(DLM_BADPARAM);
593 return DLM_BADPARAM;
594 }
595
596 lock = lksb->lockid;
597 BUG_ON(!lock);
598 dlm_lock_get(lock);
599
600 res = lock->lockres;
601 BUG_ON(!res);
602 dlm_lockres_get(res);
603retry:
604 call_ast = 0;
605 /* need to retry up here because owner may have changed */
606 mlog(0, "lock=%p res=%p\n", lock, res);
607
608 spin_lock(&res->spinlock);
609 is_master = (res->owner == dlm->node_num);
610 spin_unlock(&res->spinlock);
611
612 if (is_master) {
613 status = dlmunlock_master(dlm, res, lock, lksb, flags,
614 &call_ast);
615 mlog(0, "done calling dlmunlock_master: returned %d, "
616 "call_ast is %d\n", status, call_ast);
617 } else {
618 status = dlmunlock_remote(dlm, res, lock, lksb, flags,
619 &call_ast);
620 mlog(0, "done calling dlmunlock_remote: returned %d, "
621 "call_ast is %d\n", status, call_ast);
622 }
623
624 if (status == DLM_RECOVERING ||
625 status == DLM_MIGRATING ||
626 status == DLM_FORWARD) {
627 /* We want to go away for a tiny bit to allow recovery
628 * / migration to complete on this resource. I don't
629 * know of any wait queue we could sleep on as this
630 * may be happening on another node. Perhaps the
631 * proper solution is to queue up requests on the
632 * other end? */
633
634 /* do we want to yield(); ?? */
635 msleep(50);
636
637 mlog(0, "retrying unlock due to pending recovery/"
638 "migration/in-progress\n");
639 goto retry;
640 }
641
642 if (call_ast) {
643 mlog(0, "calling unlockast(%p, %d)\n", data, lksb->status);
644 if (is_master) {
645 /* it is possible that there is one last bast
646 * pending. make sure it is flushed, then
647 * call the unlockast.
648 * not an issue if this is a mastered remotely,
649 * since this lock has been removed from the
650 * lockres queues and cannot be found. */
651 dlm_kick_thread(dlm, NULL);
652 wait_event(dlm->ast_wq,
653 dlm_lock_basts_flushed(dlm, lock));
654 }
655 (*unlockast)(data, lksb->status);
656 }
657
658 if (status == DLM_NORMAL) {
659 mlog(0, "kicking the thread\n");
660 dlm_kick_thread(dlm, res);
661 } else
662 dlm_error(status);
663
664 dlm_lockres_calc_usage(dlm, res);
665 dlm_lockres_put(res);
666 dlm_lock_put(lock);
667
668 mlog(0, "returning status=%d!\n", status);
669 return status;
670}
671EXPORT_SYMBOL_GPL(dlmunlock);
672
diff --git a/fs/ocfs2/dlm/dlmver.c b/fs/ocfs2/dlm/dlmver.c
new file mode 100644
index 000000000000..7ef2653f8f41
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmver.c
@@ -0,0 +1,42 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmver.c
5 *
6 * version string
7 *
8 * Copyright (C) 2002, 2005 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 */
25
26#include <linux/module.h>
27#include <linux/kernel.h>
28
29#include "dlmver.h"
30
31#define DLM_BUILD_VERSION "1.3.3"
32
33#define VERSION_STR "OCFS2 DLM " DLM_BUILD_VERSION
34
35void dlm_print_version(void)
36{
37 printk(KERN_INFO "%s\n", VERSION_STR);
38}
39
40MODULE_DESCRIPTION(VERSION_STR);
41
42MODULE_VERSION(DLM_BUILD_VERSION);
diff --git a/fs/ocfs2/dlm/dlmver.h b/fs/ocfs2/dlm/dlmver.h
new file mode 100644
index 000000000000..f674aee77a16
--- /dev/null
+++ b/fs/ocfs2/dlm/dlmver.h
@@ -0,0 +1,31 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dlmfsver.h
5 *
6 * Function prototypes
7 *
8 * Copyright (C) 2005 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 */
25
26#ifndef DLM_VER_H
27#define DLM_VER_H
28
29void dlm_print_version(void);
30
31#endif /* DLM_VER_H */
diff --git a/fs/ocfs2/dlm/userdlm.c b/fs/ocfs2/dlm/userdlm.c
new file mode 100644
index 000000000000..e1fdd288796e
--- /dev/null
+++ b/fs/ocfs2/dlm/userdlm.c
@@ -0,0 +1,658 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * userdlm.c
5 *
6 * Code which implements the kernel side of a minimal userspace
7 * interface to our DLM.
8 *
9 * Many of the functions here are pared down versions of dlmglue.c
10 * functions.
11 *
12 * Copyright (C) 2003, 2004 Oracle. All rights reserved.
13 *
14 * This program is free software; you can redistribute it and/or
15 * modify it under the terms of the GNU General Public
16 * License as published by the Free Software Foundation; either
17 * version 2 of the License, or (at your option) any later version.
18 *
19 * This program is distributed in the hope that it will be useful,
20 * but WITHOUT ANY WARRANTY; without even the implied warranty of
21 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
22 * General Public License for more details.
23 *
24 * You should have received a copy of the GNU General Public
25 * License along with this program; if not, write to the
26 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
27 * Boston, MA 021110-1307, USA.
28 */
29
30#include <asm/signal.h>
31
32#include <linux/module.h>
33#include <linux/fs.h>
34#include <linux/types.h>
35#include <linux/crc32.h>
36
37
38#include "cluster/nodemanager.h"
39#include "cluster/heartbeat.h"
40#include "cluster/tcp.h"
41
42#include "dlmapi.h"
43
44#include "userdlm.h"
45
46#define MLOG_MASK_PREFIX ML_DLMFS
47#include "cluster/masklog.h"
48
49static inline int user_check_wait_flag(struct user_lock_res *lockres,
50 int flag)
51{
52 int ret;
53
54 spin_lock(&lockres->l_lock);
55 ret = lockres->l_flags & flag;
56 spin_unlock(&lockres->l_lock);
57
58 return ret;
59}
60
61static inline void user_wait_on_busy_lock(struct user_lock_res *lockres)
62
63{
64 wait_event(lockres->l_event,
65 !user_check_wait_flag(lockres, USER_LOCK_BUSY));
66}
67
68static inline void user_wait_on_blocked_lock(struct user_lock_res *lockres)
69
70{
71 wait_event(lockres->l_event,
72 !user_check_wait_flag(lockres, USER_LOCK_BLOCKED));
73}
74
75/* I heart container_of... */
76static inline struct dlm_ctxt *
77dlm_ctxt_from_user_lockres(struct user_lock_res *lockres)
78{
79 struct dlmfs_inode_private *ip;
80
81 ip = container_of(lockres,
82 struct dlmfs_inode_private,
83 ip_lockres);
84 return ip->ip_dlm;
85}
86
87static struct inode *
88user_dlm_inode_from_user_lockres(struct user_lock_res *lockres)
89{
90 struct dlmfs_inode_private *ip;
91
92 ip = container_of(lockres,
93 struct dlmfs_inode_private,
94 ip_lockres);
95 return &ip->ip_vfs_inode;
96}
97
98static inline void user_recover_from_dlm_error(struct user_lock_res *lockres)
99{
100 spin_lock(&lockres->l_lock);
101 lockres->l_flags &= ~USER_LOCK_BUSY;
102 spin_unlock(&lockres->l_lock);
103}
104
105#define user_log_dlm_error(_func, _stat, _lockres) do { \
106 mlog(ML_ERROR, "Dlm error \"%s\" while calling %s on " \
107 "resource %s: %s\n", dlm_errname(_stat), _func, \
108 _lockres->l_name, dlm_errmsg(_stat)); \
109} while (0)
110
111/* WARNING: This function lives in a world where the only three lock
112 * levels are EX, PR, and NL. It *will* have to be adjusted when more
113 * lock types are added. */
114static inline int user_highest_compat_lock_level(int level)
115{
116 int new_level = LKM_EXMODE;
117
118 if (level == LKM_EXMODE)
119 new_level = LKM_NLMODE;
120 else if (level == LKM_PRMODE)
121 new_level = LKM_PRMODE;
122 return new_level;
123}
124
125static void user_ast(void *opaque)
126{
127 struct user_lock_res *lockres = opaque;
128 struct dlm_lockstatus *lksb;
129
130 mlog(0, "AST fired for lockres %s\n", lockres->l_name);
131
132 spin_lock(&lockres->l_lock);
133
134 lksb = &(lockres->l_lksb);
135 if (lksb->status != DLM_NORMAL) {
136 mlog(ML_ERROR, "lksb status value of %u on lockres %s\n",
137 lksb->status, lockres->l_name);
138 spin_unlock(&lockres->l_lock);
139 return;
140 }
141
142 /* we're downconverting. */
143 if (lockres->l_requested < lockres->l_level) {
144 if (lockres->l_requested <=
145 user_highest_compat_lock_level(lockres->l_blocking)) {
146 lockres->l_blocking = LKM_NLMODE;
147 lockres->l_flags &= ~USER_LOCK_BLOCKED;
148 }
149 }
150
151 lockres->l_level = lockres->l_requested;
152 lockres->l_requested = LKM_IVMODE;
153 lockres->l_flags |= USER_LOCK_ATTACHED;
154 lockres->l_flags &= ~USER_LOCK_BUSY;
155
156 spin_unlock(&lockres->l_lock);
157
158 wake_up(&lockres->l_event);
159}
160
161static inline void user_dlm_grab_inode_ref(struct user_lock_res *lockres)
162{
163 struct inode *inode;
164 inode = user_dlm_inode_from_user_lockres(lockres);
165 if (!igrab(inode))
166 BUG();
167}
168
169static void user_dlm_unblock_lock(void *opaque);
170
171static void __user_dlm_queue_lockres(struct user_lock_res *lockres)
172{
173 if (!(lockres->l_flags & USER_LOCK_QUEUED)) {
174 user_dlm_grab_inode_ref(lockres);
175
176 INIT_WORK(&lockres->l_work, user_dlm_unblock_lock,
177 lockres);
178
179 queue_work(user_dlm_worker, &lockres->l_work);
180 lockres->l_flags |= USER_LOCK_QUEUED;
181 }
182}
183
184static void __user_dlm_cond_queue_lockres(struct user_lock_res *lockres)
185{
186 int queue = 0;
187
188 if (!(lockres->l_flags & USER_LOCK_BLOCKED))
189 return;
190
191 switch (lockres->l_blocking) {
192 case LKM_EXMODE:
193 if (!lockres->l_ex_holders && !lockres->l_ro_holders)
194 queue = 1;
195 break;
196 case LKM_PRMODE:
197 if (!lockres->l_ex_holders)
198 queue = 1;
199 break;
200 default:
201 BUG();
202 }
203
204 if (queue)
205 __user_dlm_queue_lockres(lockres);
206}
207
208static void user_bast(void *opaque, int level)
209{
210 struct user_lock_res *lockres = opaque;
211
212 mlog(0, "Blocking AST fired for lockres %s. Blocking level %d\n",
213 lockres->l_name, level);
214
215 spin_lock(&lockres->l_lock);
216 lockres->l_flags |= USER_LOCK_BLOCKED;
217 if (level > lockres->l_blocking)
218 lockres->l_blocking = level;
219
220 __user_dlm_queue_lockres(lockres);
221 spin_unlock(&lockres->l_lock);
222
223 wake_up(&lockres->l_event);
224}
225
226static void user_unlock_ast(void *opaque, enum dlm_status status)
227{
228 struct user_lock_res *lockres = opaque;
229
230 mlog(0, "UNLOCK AST called on lock %s\n", lockres->l_name);
231
232 if (status != DLM_NORMAL)
233 mlog(ML_ERROR, "Dlm returns status %d\n", status);
234
235 spin_lock(&lockres->l_lock);
236 if (lockres->l_flags & USER_LOCK_IN_TEARDOWN)
237 lockres->l_level = LKM_IVMODE;
238 else {
239 lockres->l_requested = LKM_IVMODE; /* cancel an
240 * upconvert
241 * request. */
242 lockres->l_flags &= ~USER_LOCK_IN_CANCEL;
243 /* we want the unblock thread to look at it again
244 * now. */
245 __user_dlm_queue_lockres(lockres);
246 }
247
248 lockres->l_flags &= ~USER_LOCK_BUSY;
249 spin_unlock(&lockres->l_lock);
250
251 wake_up(&lockres->l_event);
252}
253
254static inline void user_dlm_drop_inode_ref(struct user_lock_res *lockres)
255{
256 struct inode *inode;
257 inode = user_dlm_inode_from_user_lockres(lockres);
258 iput(inode);
259}
260
261static void user_dlm_unblock_lock(void *opaque)
262{
263 int new_level, status;
264 struct user_lock_res *lockres = (struct user_lock_res *) opaque;
265 struct dlm_ctxt *dlm = dlm_ctxt_from_user_lockres(lockres);
266
267 mlog(0, "processing lockres %s\n", lockres->l_name);
268
269 spin_lock(&lockres->l_lock);
270
271 BUG_ON(!(lockres->l_flags & USER_LOCK_BLOCKED));
272 BUG_ON(!(lockres->l_flags & USER_LOCK_QUEUED));
273
274 /* notice that we don't clear USER_LOCK_BLOCKED here. That's
275 * for user_ast to do. */
276 lockres->l_flags &= ~USER_LOCK_QUEUED;
277
278 if (lockres->l_flags & USER_LOCK_IN_TEARDOWN) {
279 mlog(0, "lock is in teardown so we do nothing\n");
280 spin_unlock(&lockres->l_lock);
281 goto drop_ref;
282 }
283
284 if (lockres->l_flags & USER_LOCK_BUSY) {
285 mlog(0, "BUSY flag detected...\n");
286 if (lockres->l_flags & USER_LOCK_IN_CANCEL) {
287 spin_unlock(&lockres->l_lock);
288 goto drop_ref;
289 }
290
291 lockres->l_flags |= USER_LOCK_IN_CANCEL;
292 spin_unlock(&lockres->l_lock);
293
294 status = dlmunlock(dlm,
295 &lockres->l_lksb,
296 LKM_CANCEL,
297 user_unlock_ast,
298 lockres);
299 if (status == DLM_CANCELGRANT) {
300 /* If we got this, then the ast was fired
301 * before we could cancel. We cleanup our
302 * state, and restart the function. */
303 spin_lock(&lockres->l_lock);
304 lockres->l_flags &= ~USER_LOCK_IN_CANCEL;
305 spin_unlock(&lockres->l_lock);
306 } else if (status != DLM_NORMAL)
307 user_log_dlm_error("dlmunlock", status, lockres);
308 goto drop_ref;
309 }
310
311 /* If there are still incompat holders, we can exit safely
312 * without worrying about re-queueing this lock as that will
313 * happen on the last call to user_cluster_unlock. */
314 if ((lockres->l_blocking == LKM_EXMODE)
315 && (lockres->l_ex_holders || lockres->l_ro_holders)) {
316 spin_unlock(&lockres->l_lock);
317 mlog(0, "can't downconvert for ex: ro = %u, ex = %u\n",
318 lockres->l_ro_holders, lockres->l_ex_holders);
319 goto drop_ref;
320 }
321
322 if ((lockres->l_blocking == LKM_PRMODE)
323 && lockres->l_ex_holders) {
324 spin_unlock(&lockres->l_lock);
325 mlog(0, "can't downconvert for pr: ex = %u\n",
326 lockres->l_ex_holders);
327 goto drop_ref;
328 }
329
330 /* yay, we can downconvert now. */
331 new_level = user_highest_compat_lock_level(lockres->l_blocking);
332 lockres->l_requested = new_level;
333 lockres->l_flags |= USER_LOCK_BUSY;
334 mlog(0, "Downconvert lock from %d to %d\n",
335 lockres->l_level, new_level);
336 spin_unlock(&lockres->l_lock);
337
338 /* need lock downconvert request now... */
339 status = dlmlock(dlm,
340 new_level,
341 &lockres->l_lksb,
342 LKM_CONVERT|LKM_VALBLK,
343 lockres->l_name,
344 user_ast,
345 lockres,
346 user_bast);
347 if (status != DLM_NORMAL) {
348 user_log_dlm_error("dlmlock", status, lockres);
349 user_recover_from_dlm_error(lockres);
350 }
351
352drop_ref:
353 user_dlm_drop_inode_ref(lockres);
354}
355
356static inline void user_dlm_inc_holders(struct user_lock_res *lockres,
357 int level)
358{
359 switch(level) {
360 case LKM_EXMODE:
361 lockres->l_ex_holders++;
362 break;
363 case LKM_PRMODE:
364 lockres->l_ro_holders++;
365 break;
366 default:
367 BUG();
368 }
369}
370
371/* predict what lock level we'll be dropping down to on behalf
372 * of another node, and return true if the currently wanted
373 * level will be compatible with it. */
374static inline int
375user_may_continue_on_blocked_lock(struct user_lock_res *lockres,
376 int wanted)
377{
378 BUG_ON(!(lockres->l_flags & USER_LOCK_BLOCKED));
379
380 return wanted <= user_highest_compat_lock_level(lockres->l_blocking);
381}
382
383int user_dlm_cluster_lock(struct user_lock_res *lockres,
384 int level,
385 int lkm_flags)
386{
387 int status, local_flags;
388 struct dlm_ctxt *dlm = dlm_ctxt_from_user_lockres(lockres);
389
390 if (level != LKM_EXMODE &&
391 level != LKM_PRMODE) {
392 mlog(ML_ERROR, "lockres %s: invalid request!\n",
393 lockres->l_name);
394 status = -EINVAL;
395 goto bail;
396 }
397
398 mlog(0, "lockres %s: asking for %s lock, passed flags = 0x%x\n",
399 lockres->l_name,
400 (level == LKM_EXMODE) ? "LKM_EXMODE" : "LKM_PRMODE",
401 lkm_flags);
402
403again:
404 if (signal_pending(current)) {
405 status = -ERESTARTSYS;
406 goto bail;
407 }
408
409 spin_lock(&lockres->l_lock);
410
411 /* We only compare against the currently granted level
412 * here. If the lock is blocked waiting on a downconvert,
413 * we'll get caught below. */
414 if ((lockres->l_flags & USER_LOCK_BUSY) &&
415 (level > lockres->l_level)) {
416 /* is someone sitting in dlm_lock? If so, wait on
417 * them. */
418 spin_unlock(&lockres->l_lock);
419
420 user_wait_on_busy_lock(lockres);
421 goto again;
422 }
423
424 if ((lockres->l_flags & USER_LOCK_BLOCKED) &&
425 (!user_may_continue_on_blocked_lock(lockres, level))) {
426 /* is the lock is currently blocked on behalf of
427 * another node */
428 spin_unlock(&lockres->l_lock);
429
430 user_wait_on_blocked_lock(lockres);
431 goto again;
432 }
433
434 if (level > lockres->l_level) {
435 local_flags = lkm_flags | LKM_VALBLK;
436 if (lockres->l_level != LKM_IVMODE)
437 local_flags |= LKM_CONVERT;
438
439 lockres->l_requested = level;
440 lockres->l_flags |= USER_LOCK_BUSY;
441 spin_unlock(&lockres->l_lock);
442
443 BUG_ON(level == LKM_IVMODE);
444 BUG_ON(level == LKM_NLMODE);
445
446 mlog(0, "lock %s, get lock from %d to level = %d\n",
447 lockres->l_name, lockres->l_level, level);
448
449 /* call dlm_lock to upgrade lock now */
450 status = dlmlock(dlm,
451 level,
452 &lockres->l_lksb,
453 local_flags,
454 lockres->l_name,
455 user_ast,
456 lockres,
457 user_bast);
458 if (status != DLM_NORMAL) {
459 if ((lkm_flags & LKM_NOQUEUE) &&
460 (status == DLM_NOTQUEUED))
461 status = -EAGAIN;
462 else {
463 user_log_dlm_error("dlmlock", status, lockres);
464 status = -EINVAL;
465 }
466 user_recover_from_dlm_error(lockres);
467 goto bail;
468 }
469
470 mlog(0, "lock %s, successfull return from dlmlock\n",
471 lockres->l_name);
472
473 user_wait_on_busy_lock(lockres);
474 goto again;
475 }
476
477 user_dlm_inc_holders(lockres, level);
478 spin_unlock(&lockres->l_lock);
479
480 mlog(0, "lockres %s: Got %s lock!\n", lockres->l_name,
481 (level == LKM_EXMODE) ? "LKM_EXMODE" : "LKM_PRMODE");
482
483 status = 0;
484bail:
485 return status;
486}
487
488static inline void user_dlm_dec_holders(struct user_lock_res *lockres,
489 int level)
490{
491 switch(level) {
492 case LKM_EXMODE:
493 BUG_ON(!lockres->l_ex_holders);
494 lockres->l_ex_holders--;
495 break;
496 case LKM_PRMODE:
497 BUG_ON(!lockres->l_ro_holders);
498 lockres->l_ro_holders--;
499 break;
500 default:
501 BUG();
502 }
503}
504
505void user_dlm_cluster_unlock(struct user_lock_res *lockres,
506 int level)
507{
508 if (level != LKM_EXMODE &&
509 level != LKM_PRMODE) {
510 mlog(ML_ERROR, "lockres %s: invalid request!\n", lockres->l_name);
511 return;
512 }
513
514 mlog(0, "lockres %s: dropping %s lock\n", lockres->l_name,
515 (level == LKM_EXMODE) ? "LKM_EXMODE" : "LKM_PRMODE");
516
517 spin_lock(&lockres->l_lock);
518 user_dlm_dec_holders(lockres, level);
519 __user_dlm_cond_queue_lockres(lockres);
520 spin_unlock(&lockres->l_lock);
521}
522
523void user_dlm_write_lvb(struct inode *inode,
524 const char *val,
525 unsigned int len)
526{
527 struct user_lock_res *lockres = &DLMFS_I(inode)->ip_lockres;
528 char *lvb = lockres->l_lksb.lvb;
529
530 BUG_ON(len > DLM_LVB_LEN);
531
532 spin_lock(&lockres->l_lock);
533
534 BUG_ON(lockres->l_level < LKM_EXMODE);
535 memcpy(lvb, val, len);
536
537 spin_unlock(&lockres->l_lock);
538}
539
540void user_dlm_read_lvb(struct inode *inode,
541 char *val,
542 unsigned int len)
543{
544 struct user_lock_res *lockres = &DLMFS_I(inode)->ip_lockres;
545 char *lvb = lockres->l_lksb.lvb;
546
547 BUG_ON(len > DLM_LVB_LEN);
548
549 spin_lock(&lockres->l_lock);
550
551 BUG_ON(lockres->l_level < LKM_PRMODE);
552 memcpy(val, lvb, len);
553
554 spin_unlock(&lockres->l_lock);
555}
556
557void user_dlm_lock_res_init(struct user_lock_res *lockres,
558 struct dentry *dentry)
559{
560 memset(lockres, 0, sizeof(*lockres));
561
562 spin_lock_init(&lockres->l_lock);
563 init_waitqueue_head(&lockres->l_event);
564 lockres->l_level = LKM_IVMODE;
565 lockres->l_requested = LKM_IVMODE;
566 lockres->l_blocking = LKM_IVMODE;
567
568 /* should have been checked before getting here. */
569 BUG_ON(dentry->d_name.len >= USER_DLM_LOCK_ID_MAX_LEN);
570
571 memcpy(lockres->l_name,
572 dentry->d_name.name,
573 dentry->d_name.len);
574}
575
576int user_dlm_destroy_lock(struct user_lock_res *lockres)
577{
578 int status = -EBUSY;
579 struct dlm_ctxt *dlm = dlm_ctxt_from_user_lockres(lockres);
580
581 mlog(0, "asked to destroy %s\n", lockres->l_name);
582
583 spin_lock(&lockres->l_lock);
584 while (lockres->l_flags & USER_LOCK_BUSY) {
585 spin_unlock(&lockres->l_lock);
586
587 mlog(0, "lock %s is busy\n", lockres->l_name);
588
589 user_wait_on_busy_lock(lockres);
590
591 spin_lock(&lockres->l_lock);
592 }
593
594 if (lockres->l_ro_holders || lockres->l_ex_holders) {
595 spin_unlock(&lockres->l_lock);
596 mlog(0, "lock %s has holders\n", lockres->l_name);
597 goto bail;
598 }
599
600 status = 0;
601 if (!(lockres->l_flags & USER_LOCK_ATTACHED)) {
602 spin_unlock(&lockres->l_lock);
603 mlog(0, "lock %s is not attached\n", lockres->l_name);
604 goto bail;
605 }
606
607 lockres->l_flags &= ~USER_LOCK_ATTACHED;
608 lockres->l_flags |= USER_LOCK_BUSY;
609 lockres->l_flags |= USER_LOCK_IN_TEARDOWN;
610 spin_unlock(&lockres->l_lock);
611
612 mlog(0, "unlocking lockres %s\n", lockres->l_name);
613 status = dlmunlock(dlm,
614 &lockres->l_lksb,
615 LKM_VALBLK,
616 user_unlock_ast,
617 lockres);
618 if (status != DLM_NORMAL) {
619 user_log_dlm_error("dlmunlock", status, lockres);
620 status = -EINVAL;
621 goto bail;
622 }
623
624 user_wait_on_busy_lock(lockres);
625
626 status = 0;
627bail:
628 return status;
629}
630
631struct dlm_ctxt *user_dlm_register_context(struct qstr *name)
632{
633 struct dlm_ctxt *dlm;
634 u32 dlm_key;
635 char *domain;
636
637 domain = kmalloc(name->len + 1, GFP_KERNEL);
638 if (!domain) {
639 mlog_errno(-ENOMEM);
640 return ERR_PTR(-ENOMEM);
641 }
642
643 dlm_key = crc32_le(0, name->name, name->len);
644
645 snprintf(domain, name->len + 1, "%.*s", name->len, name->name);
646
647 dlm = dlm_register_domain(domain, dlm_key);
648 if (IS_ERR(dlm))
649 mlog_errno(PTR_ERR(dlm));
650
651 kfree(domain);
652 return dlm;
653}
654
655void user_dlm_unregister_context(struct dlm_ctxt *dlm)
656{
657 dlm_unregister_domain(dlm);
658}
diff --git a/fs/ocfs2/dlm/userdlm.h b/fs/ocfs2/dlm/userdlm.h
new file mode 100644
index 000000000000..04178bc40b76
--- /dev/null
+++ b/fs/ocfs2/dlm/userdlm.h
@@ -0,0 +1,111 @@
1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * userdlm.h
5 *
6 * Userspace dlm defines
7 *
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
24 */
25
26
27#ifndef USERDLM_H
28#define USERDLM_H
29
30#include <linux/module.h>
31#include <linux/fs.h>
32#include <linux/types.h>
33#include <linux/workqueue.h>
34
35/* user_lock_res->l_flags flags. */
36#define USER_LOCK_ATTACHED (0x00000001) /* have we initialized
37 * the lvb */
38#define USER_LOCK_BUSY (0x00000002) /* we are currently in
39 * dlm_lock */
40#define USER_LOCK_BLOCKED (0x00000004) /* blocked waiting to
41 * downconvert*/
42#define USER_LOCK_IN_TEARDOWN (0x00000008) /* we're currently
43 * destroying this
44 * lock. */
45#define USER_LOCK_QUEUED (0x00000010) /* lock is on the
46 * workqueue */
47#define USER_LOCK_IN_CANCEL (0x00000020)
48
49struct user_lock_res {
50 spinlock_t l_lock;
51
52 int l_flags;
53
54#define USER_DLM_LOCK_ID_MAX_LEN 32
55 char l_name[USER_DLM_LOCK_ID_MAX_LEN];
56 int l_level;
57 unsigned int l_ro_holders;
58 unsigned int l_ex_holders;
59 struct dlm_lockstatus l_lksb;
60
61 int l_requested;
62 int l_blocking;
63
64 wait_queue_head_t l_event;
65
66 struct work_struct l_work;
67};
68
69extern struct workqueue_struct *user_dlm_worker;
70
71void user_dlm_lock_res_init(struct user_lock_res *lockres,
72 struct dentry *dentry);
73int user_dlm_destroy_lock(struct user_lock_res *lockres);
74int user_dlm_cluster_lock(struct user_lock_res *lockres,
75 int level,
76 int lkm_flags);
77void user_dlm_cluster_unlock(struct user_lock_res *lockres,
78 int level);
79void user_dlm_write_lvb(struct inode *inode,
80 const char *val,
81 unsigned int len);
82void user_dlm_read_lvb(struct inode *inode,
83 char *val,
84 unsigned int len);
85struct dlm_ctxt *user_dlm_register_context(struct qstr *name);
86void user_dlm_unregister_context(struct dlm_ctxt *dlm);
87
88struct dlmfs_inode_private {
89 struct dlm_ctxt *ip_dlm;
90
91 struct user_lock_res ip_lockres; /* unused for directories. */
92 struct inode *ip_parent;
93
94 struct inode ip_vfs_inode;
95};
96
97static inline struct dlmfs_inode_private *
98DLMFS_I(struct inode *inode)
99{
100 return container_of(inode,
101 struct dlmfs_inode_private,
102 ip_vfs_inode);
103}
104
105struct dlmfs_filp_private {
106 int fp_lock_level;
107};
108
109#define DLMFS_MAGIC 0x76a9f425
110
111#endif /* USERDLM_H */
generated by cgit v1.2.2 (git 2.25.0) at 2024-11-23 15:00:41 -0500