aboutsummaryrefslogtreecommitdiffstats
path: root/net/rds
diff options
context:
space:
mode:
authorAndy Grover <andy.grover@oracle.com>2009-02-24 10:30:30 -0500
committerDavid S. Miller <davem@davemloft.net>2009-02-27 02:39:30 -0500
commitec16227e14141e4fd7ae76354c09dadfe2449d9e (patch)
tree9a20de7d9a0688d721517c2b38d6be83956c7ca4 /net/rds
parenteff5f53bef75c0864a5da06bb688939092b848dc (diff)
RDS/IB: Infiniband transport
Registers as an RDS transport and an IB client, and uses IB CM API to allocate ids, queue pairs, and the rest of that fun stuff. Signed-off-by: Andy Grover <andy.grover@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'net/rds')
-rw-r--r--net/rds/ib.c323
-rw-r--r--net/rds/ib.h367
-rw-r--r--net/rds/ib_cm.c726
3 files changed, 1416 insertions, 0 deletions
diff --git a/net/rds/ib.c b/net/rds/ib.c
new file mode 100644
index 000000000000..06a7b798d9a7
--- /dev/null
+++ b/net/rds/ib.c
@@ -0,0 +1,323 @@
1/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/in.h>
35#include <linux/if.h>
36#include <linux/netdevice.h>
37#include <linux/inetdevice.h>
38#include <linux/if_arp.h>
39#include <linux/delay.h>
40
41#include "rds.h"
42#include "ib.h"
43
44unsigned int fmr_pool_size = RDS_FMR_POOL_SIZE;
45unsigned int fmr_message_size = RDS_FMR_SIZE + 1; /* +1 allows for unaligned MRs */
46
47module_param(fmr_pool_size, int, 0444);
48MODULE_PARM_DESC(fmr_pool_size, " Max number of fmr per HCA");
49module_param(fmr_message_size, int, 0444);
50MODULE_PARM_DESC(fmr_message_size, " Max size of a RDMA transfer");
51
52struct list_head rds_ib_devices;
53
54DEFINE_SPINLOCK(ib_nodev_conns_lock);
55LIST_HEAD(ib_nodev_conns);
56
57void rds_ib_add_one(struct ib_device *device)
58{
59 struct rds_ib_device *rds_ibdev;
60 struct ib_device_attr *dev_attr;
61
62 /* Only handle IB (no iWARP) devices */
63 if (device->node_type != RDMA_NODE_IB_CA)
64 return;
65
66 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
67 if (!dev_attr)
68 return;
69
70 if (ib_query_device(device, dev_attr)) {
71 rdsdebug("Query device failed for %s\n", device->name);
72 goto free_attr;
73 }
74
75 rds_ibdev = kmalloc(sizeof *rds_ibdev, GFP_KERNEL);
76 if (!rds_ibdev)
77 goto free_attr;
78
79 spin_lock_init(&rds_ibdev->spinlock);
80
81 rds_ibdev->max_wrs = dev_attr->max_qp_wr;
82 rds_ibdev->max_sge = min(dev_attr->max_sge, RDS_IB_MAX_SGE);
83
84 rds_ibdev->fmr_page_shift = max(9, ffs(dev_attr->page_size_cap) - 1);
85 rds_ibdev->fmr_page_size = 1 << rds_ibdev->fmr_page_shift;
86 rds_ibdev->fmr_page_mask = ~((u64) rds_ibdev->fmr_page_size - 1);
87 rds_ibdev->fmr_max_remaps = dev_attr->max_map_per_fmr?: 32;
88 rds_ibdev->max_fmrs = dev_attr->max_fmr ?
89 min_t(unsigned int, dev_attr->max_fmr, fmr_pool_size) :
90 fmr_pool_size;
91
92 rds_ibdev->dev = device;
93 rds_ibdev->pd = ib_alloc_pd(device);
94 if (IS_ERR(rds_ibdev->pd))
95 goto free_dev;
96
97 rds_ibdev->mr = ib_get_dma_mr(rds_ibdev->pd,
98 IB_ACCESS_LOCAL_WRITE);
99 if (IS_ERR(rds_ibdev->mr))
100 goto err_pd;
101
102 rds_ibdev->mr_pool = rds_ib_create_mr_pool(rds_ibdev);
103 if (IS_ERR(rds_ibdev->mr_pool)) {
104 rds_ibdev->mr_pool = NULL;
105 goto err_mr;
106 }
107
108 INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
109 INIT_LIST_HEAD(&rds_ibdev->conn_list);
110 list_add_tail(&rds_ibdev->list, &rds_ib_devices);
111
112 ib_set_client_data(device, &rds_ib_client, rds_ibdev);
113
114 goto free_attr;
115
116err_mr:
117 ib_dereg_mr(rds_ibdev->mr);
118err_pd:
119 ib_dealloc_pd(rds_ibdev->pd);
120free_dev:
121 kfree(rds_ibdev);
122free_attr:
123 kfree(dev_attr);
124}
125
126void rds_ib_remove_one(struct ib_device *device)
127{
128 struct rds_ib_device *rds_ibdev;
129 struct rds_ib_ipaddr *i_ipaddr, *i_next;
130
131 rds_ibdev = ib_get_client_data(device, &rds_ib_client);
132 if (!rds_ibdev)
133 return;
134
135 list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
136 list_del(&i_ipaddr->list);
137 kfree(i_ipaddr);
138 }
139
140 rds_ib_remove_conns(rds_ibdev);
141
142 if (rds_ibdev->mr_pool)
143 rds_ib_destroy_mr_pool(rds_ibdev->mr_pool);
144
145 ib_dereg_mr(rds_ibdev->mr);
146
147 while (ib_dealloc_pd(rds_ibdev->pd)) {
148 rdsdebug("Failed to dealloc pd %p\n", rds_ibdev->pd);
149 msleep(1);
150 }
151
152 list_del(&rds_ibdev->list);
153 kfree(rds_ibdev);
154}
155
156struct ib_client rds_ib_client = {
157 .name = "rds_ib",
158 .add = rds_ib_add_one,
159 .remove = rds_ib_remove_one
160};
161
162static int rds_ib_conn_info_visitor(struct rds_connection *conn,
163 void *buffer)
164{
165 struct rds_info_rdma_connection *iinfo = buffer;
166 struct rds_ib_connection *ic;
167
168 /* We will only ever look at IB transports */
169 if (conn->c_trans != &rds_ib_transport)
170 return 0;
171
172 iinfo->src_addr = conn->c_laddr;
173 iinfo->dst_addr = conn->c_faddr;
174
175 memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
176 memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
177 if (rds_conn_state(conn) == RDS_CONN_UP) {
178 struct rds_ib_device *rds_ibdev;
179 struct rdma_dev_addr *dev_addr;
180
181 ic = conn->c_transport_data;
182 dev_addr = &ic->i_cm_id->route.addr.dev_addr;
183
184 ib_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid);
185 ib_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid);
186
187 rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
188 iinfo->max_send_wr = ic->i_send_ring.w_nr;
189 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
190 iinfo->max_send_sge = rds_ibdev->max_sge;
191 rds_ib_get_mr_info(rds_ibdev, iinfo);
192 }
193 return 1;
194}
195
196static void rds_ib_ic_info(struct socket *sock, unsigned int len,
197 struct rds_info_iterator *iter,
198 struct rds_info_lengths *lens)
199{
200 rds_for_each_conn_info(sock, len, iter, lens,
201 rds_ib_conn_info_visitor,
202 sizeof(struct rds_info_rdma_connection));
203}
204
205
206/*
207 * Early RDS/IB was built to only bind to an address if there is an IPoIB
208 * device with that address set.
209 *
210 * If it were me, I'd advocate for something more flexible. Sending and
211 * receiving should be device-agnostic. Transports would try and maintain
212 * connections between peers who have messages queued. Userspace would be
213 * allowed to influence which paths have priority. We could call userspace
214 * asserting this policy "routing".
215 */
216static int rds_ib_laddr_check(__be32 addr)
217{
218 int ret;
219 struct rdma_cm_id *cm_id;
220 struct sockaddr_in sin;
221
222 /* Create a CMA ID and try to bind it. This catches both
223 * IB and iWARP capable NICs.
224 */
225 cm_id = rdma_create_id(NULL, NULL, RDMA_PS_TCP);
226 if (!cm_id)
227 return -EADDRNOTAVAIL;
228
229 memset(&sin, 0, sizeof(sin));
230 sin.sin_family = AF_INET;
231 sin.sin_addr.s_addr = addr;
232
233 /* rdma_bind_addr will only succeed for IB & iWARP devices */
234 ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
235 /* due to this, we will claim to support iWARP devices unless we
236 check node_type. */
237 if (ret || cm_id->device->node_type != RDMA_NODE_IB_CA)
238 ret = -EADDRNOTAVAIL;
239
240 rdsdebug("addr %pI4 ret %d node type %d\n",
241 &addr, ret,
242 cm_id->device ? cm_id->device->node_type : -1);
243
244 rdma_destroy_id(cm_id);
245
246 return ret;
247}
248
249void rds_ib_exit(void)
250{
251 rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
252 rds_ib_remove_nodev_conns();
253 ib_unregister_client(&rds_ib_client);
254 rds_ib_sysctl_exit();
255 rds_ib_recv_exit();
256 rds_trans_unregister(&rds_ib_transport);
257}
258
259struct rds_transport rds_ib_transport = {
260 .laddr_check = rds_ib_laddr_check,
261 .xmit_complete = rds_ib_xmit_complete,
262 .xmit = rds_ib_xmit,
263 .xmit_cong_map = NULL,
264 .xmit_rdma = rds_ib_xmit_rdma,
265 .recv = rds_ib_recv,
266 .conn_alloc = rds_ib_conn_alloc,
267 .conn_free = rds_ib_conn_free,
268 .conn_connect = rds_ib_conn_connect,
269 .conn_shutdown = rds_ib_conn_shutdown,
270 .inc_copy_to_user = rds_ib_inc_copy_to_user,
271 .inc_purge = rds_ib_inc_purge,
272 .inc_free = rds_ib_inc_free,
273 .cm_initiate_connect = rds_ib_cm_initiate_connect,
274 .cm_handle_connect = rds_ib_cm_handle_connect,
275 .cm_connect_complete = rds_ib_cm_connect_complete,
276 .stats_info_copy = rds_ib_stats_info_copy,
277 .exit = rds_ib_exit,
278 .get_mr = rds_ib_get_mr,
279 .sync_mr = rds_ib_sync_mr,
280 .free_mr = rds_ib_free_mr,
281 .flush_mrs = rds_ib_flush_mrs,
282 .t_owner = THIS_MODULE,
283 .t_name = "infiniband",
284};
285
286int __init rds_ib_init(void)
287{
288 int ret;
289
290 INIT_LIST_HEAD(&rds_ib_devices);
291
292 ret = ib_register_client(&rds_ib_client);
293 if (ret)
294 goto out;
295
296 ret = rds_ib_sysctl_init();
297 if (ret)
298 goto out_ibreg;
299
300 ret = rds_ib_recv_init();
301 if (ret)
302 goto out_sysctl;
303
304 ret = rds_trans_register(&rds_ib_transport);
305 if (ret)
306 goto out_recv;
307
308 rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
309
310 goto out;
311
312out_recv:
313 rds_ib_recv_exit();
314out_sysctl:
315 rds_ib_sysctl_exit();
316out_ibreg:
317 ib_unregister_client(&rds_ib_client);
318out:
319 return ret;
320}
321
322MODULE_LICENSE("GPL");
323
diff --git a/net/rds/ib.h b/net/rds/ib.h
new file mode 100644
index 000000000000..8be563a1363a
--- /dev/null
+++ b/net/rds/ib.h
@@ -0,0 +1,367 @@
1#ifndef _RDS_IB_H
2#define _RDS_IB_H
3
4#include <rdma/ib_verbs.h>
5#include <rdma/rdma_cm.h>
6#include "rds.h"
7#include "rdma_transport.h"
8
9#define RDS_FMR_SIZE 256
10#define RDS_FMR_POOL_SIZE 4096
11
12#define RDS_IB_MAX_SGE 8
13#define RDS_IB_RECV_SGE 2
14
15#define RDS_IB_DEFAULT_RECV_WR 1024
16#define RDS_IB_DEFAULT_SEND_WR 256
17
18#define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */
19
20extern struct list_head rds_ib_devices;
21
22/*
23 * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
24 * try and minimize the amount of memory tied up both the device and
25 * socket receive queues.
26 */
27/* page offset of the final full frag that fits in the page */
28#define RDS_PAGE_LAST_OFF (((PAGE_SIZE / RDS_FRAG_SIZE) - 1) * RDS_FRAG_SIZE)
29struct rds_page_frag {
30 struct list_head f_item;
31 struct page *f_page;
32 unsigned long f_offset;
33 dma_addr_t f_mapped;
34};
35
36struct rds_ib_incoming {
37 struct list_head ii_frags;
38 struct rds_incoming ii_inc;
39};
40
41struct rds_ib_connect_private {
42 /* Add new fields at the end, and don't permute existing fields. */
43 __be32 dp_saddr;
44 __be32 dp_daddr;
45 u8 dp_protocol_major;
46 u8 dp_protocol_minor;
47 __be16 dp_protocol_minor_mask; /* bitmask */
48 __be32 dp_reserved1;
49 __be64 dp_ack_seq;
50 __be32 dp_credit; /* non-zero enables flow ctl */
51};
52
53struct rds_ib_send_work {
54 struct rds_message *s_rm;
55 struct rds_rdma_op *s_op;
56 struct ib_send_wr s_wr;
57 struct ib_sge s_sge[RDS_IB_MAX_SGE];
58 unsigned long s_queued;
59};
60
61struct rds_ib_recv_work {
62 struct rds_ib_incoming *r_ibinc;
63 struct rds_page_frag *r_frag;
64 struct ib_recv_wr r_wr;
65 struct ib_sge r_sge[2];
66};
67
68struct rds_ib_work_ring {
69 u32 w_nr;
70 u32 w_alloc_ptr;
71 u32 w_alloc_ctr;
72 u32 w_free_ptr;
73 atomic_t w_free_ctr;
74};
75
76struct rds_ib_device;
77
78struct rds_ib_connection {
79
80 struct list_head ib_node;
81 struct rds_ib_device *rds_ibdev;
82 struct rds_connection *conn;
83
84 /* alphabet soup, IBTA style */
85 struct rdma_cm_id *i_cm_id;
86 struct ib_pd *i_pd;
87 struct ib_mr *i_mr;
88 struct ib_cq *i_send_cq;
89 struct ib_cq *i_recv_cq;
90
91 /* tx */
92 struct rds_ib_work_ring i_send_ring;
93 struct rds_message *i_rm;
94 struct rds_header *i_send_hdrs;
95 u64 i_send_hdrs_dma;
96 struct rds_ib_send_work *i_sends;
97
98 /* rx */
99 struct mutex i_recv_mutex;
100 struct rds_ib_work_ring i_recv_ring;
101 struct rds_ib_incoming *i_ibinc;
102 u32 i_recv_data_rem;
103 struct rds_header *i_recv_hdrs;
104 u64 i_recv_hdrs_dma;
105 struct rds_ib_recv_work *i_recvs;
106 struct rds_page_frag i_frag;
107 u64 i_ack_recv; /* last ACK received */
108
109 /* sending acks */
110 unsigned long i_ack_flags;
111 u64 i_ack_next; /* next ACK to send */
112 struct rds_header *i_ack;
113 struct ib_send_wr i_ack_wr;
114 struct ib_sge i_ack_sge;
115 u64 i_ack_dma;
116 unsigned long i_ack_queued;
117
118 /* Flow control related information
119 *
120 * Our algorithm uses a pair variables that we need to access
121 * atomically - one for the send credits, and one posted
122 * recv credits we need to transfer to remote.
123 * Rather than protect them using a slow spinlock, we put both into
124 * a single atomic_t and update it using cmpxchg
125 */
126 atomic_t i_credits;
127
128 /* Protocol version specific information */
129 unsigned int i_flowctl:1; /* enable/disable flow ctl */
130
131 /* Batched completions */
132 unsigned int i_unsignaled_wrs;
133 long i_unsignaled_bytes;
134};
135
136/* This assumes that atomic_t is at least 32 bits */
137#define IB_GET_SEND_CREDITS(v) ((v) & 0xffff)
138#define IB_GET_POST_CREDITS(v) ((v) >> 16)
139#define IB_SET_SEND_CREDITS(v) ((v) & 0xffff)
140#define IB_SET_POST_CREDITS(v) ((v) << 16)
141
142struct rds_ib_ipaddr {
143 struct list_head list;
144 __be32 ipaddr;
145};
146
147struct rds_ib_device {
148 struct list_head list;
149 struct list_head ipaddr_list;
150 struct list_head conn_list;
151 struct ib_device *dev;
152 struct ib_pd *pd;
153 struct ib_mr *mr;
154 struct rds_ib_mr_pool *mr_pool;
155 int fmr_page_shift;
156 int fmr_page_size;
157 u64 fmr_page_mask;
158 unsigned int fmr_max_remaps;
159 unsigned int max_fmrs;
160 int max_sge;
161 unsigned int max_wrs;
162 spinlock_t spinlock; /* protect the above */
163};
164
165/* bits for i_ack_flags */
166#define IB_ACK_IN_FLIGHT 0
167#define IB_ACK_REQUESTED 1
168
169/* Magic WR_ID for ACKs */
170#define RDS_IB_ACK_WR_ID (~(u64) 0)
171
172struct rds_ib_statistics {
173 uint64_t s_ib_connect_raced;
174 uint64_t s_ib_listen_closed_stale;
175 uint64_t s_ib_tx_cq_call;
176 uint64_t s_ib_tx_cq_event;
177 uint64_t s_ib_tx_ring_full;
178 uint64_t s_ib_tx_throttle;
179 uint64_t s_ib_tx_sg_mapping_failure;
180 uint64_t s_ib_tx_stalled;
181 uint64_t s_ib_tx_credit_updates;
182 uint64_t s_ib_rx_cq_call;
183 uint64_t s_ib_rx_cq_event;
184 uint64_t s_ib_rx_ring_empty;
185 uint64_t s_ib_rx_refill_from_cq;
186 uint64_t s_ib_rx_refill_from_thread;
187 uint64_t s_ib_rx_alloc_limit;
188 uint64_t s_ib_rx_credit_updates;
189 uint64_t s_ib_ack_sent;
190 uint64_t s_ib_ack_send_failure;
191 uint64_t s_ib_ack_send_delayed;
192 uint64_t s_ib_ack_send_piggybacked;
193 uint64_t s_ib_ack_received;
194 uint64_t s_ib_rdma_mr_alloc;
195 uint64_t s_ib_rdma_mr_free;
196 uint64_t s_ib_rdma_mr_used;
197 uint64_t s_ib_rdma_mr_pool_flush;
198 uint64_t s_ib_rdma_mr_pool_wait;
199 uint64_t s_ib_rdma_mr_pool_depleted;
200};
201
202extern struct workqueue_struct *rds_ib_wq;
203
204/*
205 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
206 * doesn't define it.
207 */
208static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev,
209 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
210{
211 unsigned int i;
212
213 for (i = 0; i < sg_dma_len; ++i) {
214 ib_dma_sync_single_for_cpu(dev,
215 ib_sg_dma_address(dev, &sg[i]),
216 ib_sg_dma_len(dev, &sg[i]),
217 direction);
218 }
219}
220#define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu
221
222static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
223 struct scatterlist *sg, unsigned int sg_dma_len, int direction)
224{
225 unsigned int i;
226
227 for (i = 0; i < sg_dma_len; ++i) {
228 ib_dma_sync_single_for_device(dev,
229 ib_sg_dma_address(dev, &sg[i]),
230 ib_sg_dma_len(dev, &sg[i]),
231 direction);
232 }
233}
234#define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device
235
236
237/* ib.c */
238extern struct rds_transport rds_ib_transport;
239extern void rds_ib_add_one(struct ib_device *device);
240extern void rds_ib_remove_one(struct ib_device *device);
241extern struct ib_client rds_ib_client;
242
243extern unsigned int fmr_pool_size;
244extern unsigned int fmr_message_size;
245
246extern spinlock_t ib_nodev_conns_lock;
247extern struct list_head ib_nodev_conns;
248
249/* ib_cm.c */
250int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp);
251void rds_ib_conn_free(void *arg);
252int rds_ib_conn_connect(struct rds_connection *conn);
253void rds_ib_conn_shutdown(struct rds_connection *conn);
254void rds_ib_state_change(struct sock *sk);
255int __init rds_ib_listen_init(void);
256void rds_ib_listen_stop(void);
257void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
258int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
259 struct rdma_cm_event *event);
260int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id);
261void rds_ib_cm_connect_complete(struct rds_connection *conn,
262 struct rdma_cm_event *event);
263
264
265#define rds_ib_conn_error(conn, fmt...) \
266 __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
267
268/* ib_rdma.c */
269int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
270int rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
271void rds_ib_remove_nodev_conns(void);
272void rds_ib_remove_conns(struct rds_ib_device *rds_ibdev);
273struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *);
274void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo);
275void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
276void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
277 struct rds_sock *rs, u32 *key_ret);
278void rds_ib_sync_mr(void *trans_private, int dir);
279void rds_ib_free_mr(void *trans_private, int invalidate);
280void rds_ib_flush_mrs(void);
281
282/* ib_recv.c */
283int __init rds_ib_recv_init(void);
284void rds_ib_recv_exit(void);
285int rds_ib_recv(struct rds_connection *conn);
286int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp,
287 gfp_t page_gfp, int prefill);
288void rds_ib_inc_purge(struct rds_incoming *inc);
289void rds_ib_inc_free(struct rds_incoming *inc);
290int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov,
291 size_t size);
292void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context);
293void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
294void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
295void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
296void rds_ib_attempt_ack(struct rds_ib_connection *ic);
297void rds_ib_ack_send_complete(struct rds_ib_connection *ic);
298u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic);
299
300/* ib_ring.c */
301void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr);
302void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr);
303u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos);
304void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val);
305void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val);
306int rds_ib_ring_empty(struct rds_ib_work_ring *ring);
307int rds_ib_ring_low(struct rds_ib_work_ring *ring);
308u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring);
309u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
310extern wait_queue_head_t rds_ib_ring_empty_wait;
311
312/* ib_send.c */
313void rds_ib_xmit_complete(struct rds_connection *conn);
314int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
315 unsigned int hdr_off, unsigned int sg, unsigned int off);
316void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context);
317void rds_ib_send_init_ring(struct rds_ib_connection *ic);
318void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
319int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op);
320void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
321void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
322int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
323 u32 *adv_credits, int need_posted);
324
325/* ib_stats.c */
326DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats);
327#define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
328unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
329 unsigned int avail);
330
331/* ib_sysctl.c */
332int __init rds_ib_sysctl_init(void);
333void rds_ib_sysctl_exit(void);
334extern unsigned long rds_ib_sysctl_max_send_wr;
335extern unsigned long rds_ib_sysctl_max_recv_wr;
336extern unsigned long rds_ib_sysctl_max_unsig_wrs;
337extern unsigned long rds_ib_sysctl_max_unsig_bytes;
338extern unsigned long rds_ib_sysctl_max_recv_allocation;
339extern unsigned int rds_ib_sysctl_flow_control;
340extern ctl_table rds_ib_sysctl_table[];
341
342/*
343 * Helper functions for getting/setting the header and data SGEs in
344 * RDS packets (not RDMA)
345 */
346static inline struct ib_sge *
347rds_ib_header_sge(struct rds_ib_connection *ic, struct ib_sge *sge)
348{
349 return &sge[0];
350}
351
352static inline struct ib_sge *
353rds_ib_data_sge(struct rds_ib_connection *ic, struct ib_sge *sge)
354{
355 return &sge[1];
356}
357
358static inline void rds_ib_set_64bit(u64 *ptr, u64 val)
359{
360#if BITS_PER_LONG == 64
361 *ptr = val;
362#else
363 set_64bit(ptr, val);
364#endif
365}
366
367#endif
diff --git a/net/rds/ib_cm.c b/net/rds/ib_cm.c
new file mode 100644
index 000000000000..0532237bd128
--- /dev/null
+++ b/net/rds/ib_cm.c
@@ -0,0 +1,726 @@
1/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/in.h>
35#include <linux/vmalloc.h>
36
37#include "rds.h"
38#include "ib.h"
39
40/*
41 * Set the selected protocol version
42 */
43static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
44{
45 conn->c_version = version;
46}
47
48/*
49 * Set up flow control
50 */
51static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
52{
53 struct rds_ib_connection *ic = conn->c_transport_data;
54
55 if (rds_ib_sysctl_flow_control && credits != 0) {
56 /* We're doing flow control */
57 ic->i_flowctl = 1;
58 rds_ib_send_add_credits(conn, credits);
59 } else {
60 ic->i_flowctl = 0;
61 }
62}
63
64/*
65 * Tune RNR behavior. Without flow control, we use a rather
66 * low timeout, but not the absolute minimum - this should
67 * be tunable.
68 *
69 * We already set the RNR retry count to 7 (which is the
70 * smallest infinite number :-) above.
71 * If flow control is off, we want to change this back to 0
72 * so that we learn quickly when our credit accounting is
73 * buggy.
74 *
75 * Caller passes in a qp_attr pointer - don't waste stack spacv
76 * by allocation this twice.
77 */
78static void
79rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
80{
81 int ret;
82
83 attr->min_rnr_timer = IB_RNR_TIMER_000_32;
84 ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
85 if (ret)
86 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
87}
88
89/*
90 * Connection established.
91 * We get here for both outgoing and incoming connection.
92 */
93void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
94{
95 const struct rds_ib_connect_private *dp = NULL;
96 struct rds_ib_connection *ic = conn->c_transport_data;
97 struct rds_ib_device *rds_ibdev;
98 struct ib_qp_attr qp_attr;
99 int err;
100
101 if (event->param.conn.private_data_len) {
102 dp = event->param.conn.private_data;
103
104 rds_ib_set_protocol(conn,
105 RDS_PROTOCOL(dp->dp_protocol_major,
106 dp->dp_protocol_minor));
107 rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
108 }
109
110 printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
111 &conn->c_laddr,
112 RDS_PROTOCOL_MAJOR(conn->c_version),
113 RDS_PROTOCOL_MINOR(conn->c_version),
114 ic->i_flowctl ? ", flow control" : "");
115
116 /* Tune RNR behavior */
117 rds_ib_tune_rnr(ic, &qp_attr);
118
119 qp_attr.qp_state = IB_QPS_RTS;
120 err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
121 if (err)
122 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
123
124 /* update ib_device with this local ipaddr & conn */
125 rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
126 err = rds_ib_update_ipaddr(rds_ibdev, conn->c_laddr);
127 if (err)
128 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", err);
129 err = rds_ib_add_conn(rds_ibdev, conn);
130 if (err)
131 printk(KERN_ERR "rds_ib_add_conn failed (%d)\n", err);
132
133 /* If the peer gave us the last packet it saw, process this as if
134 * we had received a regular ACK. */
135 if (dp && dp->dp_ack_seq)
136 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
137
138 rds_connect_complete(conn);
139}
140
141static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
142 struct rdma_conn_param *conn_param,
143 struct rds_ib_connect_private *dp,
144 u32 protocol_version)
145{
146 memset(conn_param, 0, sizeof(struct rdma_conn_param));
147 /* XXX tune these? */
148 conn_param->responder_resources = 1;
149 conn_param->initiator_depth = 1;
150 conn_param->retry_count = 7;
151 conn_param->rnr_retry_count = 7;
152
153 if (dp) {
154 struct rds_ib_connection *ic = conn->c_transport_data;
155
156 memset(dp, 0, sizeof(*dp));
157 dp->dp_saddr = conn->c_laddr;
158 dp->dp_daddr = conn->c_faddr;
159 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
160 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
161 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
162 dp->dp_ack_seq = rds_ib_piggyb_ack(ic);
163
164 /* Advertise flow control */
165 if (ic->i_flowctl) {
166 unsigned int credits;
167
168 credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
169 dp->dp_credit = cpu_to_be32(credits);
170 atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
171 }
172
173 conn_param->private_data = dp;
174 conn_param->private_data_len = sizeof(*dp);
175 }
176}
177
178static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
179{
180 rdsdebug("event %u data %p\n", event->event, data);
181}
182
183static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
184{
185 struct rds_connection *conn = data;
186 struct rds_ib_connection *ic = conn->c_transport_data;
187
188 rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event);
189
190 switch (event->event) {
191 case IB_EVENT_COMM_EST:
192 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
193 break;
194 default:
195 printk(KERN_WARNING "RDS/ib: unhandled QP event %u "
196 "on connection to %pI4\n", event->event,
197 &conn->c_faddr);
198 break;
199 }
200}
201
202/*
203 * This needs to be very careful to not leave IS_ERR pointers around for
204 * cleanup to trip over.
205 */
206static int rds_ib_setup_qp(struct rds_connection *conn)
207{
208 struct rds_ib_connection *ic = conn->c_transport_data;
209 struct ib_device *dev = ic->i_cm_id->device;
210 struct ib_qp_init_attr attr;
211 struct rds_ib_device *rds_ibdev;
212 int ret;
213
214 /* rds_ib_add_one creates a rds_ib_device object per IB device,
215 * and allocates a protection domain, memory range and FMR pool
216 * for each. If that fails for any reason, it will not register
217 * the rds_ibdev at all.
218 */
219 rds_ibdev = ib_get_client_data(dev, &rds_ib_client);
220 if (rds_ibdev == NULL) {
221 if (printk_ratelimit())
222 printk(KERN_NOTICE "RDS/IB: No client_data for device %s\n",
223 dev->name);
224 return -EOPNOTSUPP;
225 }
226
227 if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
228 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
229 if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
230 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
231
232 /* Protection domain and memory range */
233 ic->i_pd = rds_ibdev->pd;
234 ic->i_mr = rds_ibdev->mr;
235
236 ic->i_send_cq = ib_create_cq(dev, rds_ib_send_cq_comp_handler,
237 rds_ib_cq_event_handler, conn,
238 ic->i_send_ring.w_nr + 1, 0);
239 if (IS_ERR(ic->i_send_cq)) {
240 ret = PTR_ERR(ic->i_send_cq);
241 ic->i_send_cq = NULL;
242 rdsdebug("ib_create_cq send failed: %d\n", ret);
243 goto out;
244 }
245
246 ic->i_recv_cq = ib_create_cq(dev, rds_ib_recv_cq_comp_handler,
247 rds_ib_cq_event_handler, conn,
248 ic->i_recv_ring.w_nr, 0);
249 if (IS_ERR(ic->i_recv_cq)) {
250 ret = PTR_ERR(ic->i_recv_cq);
251 ic->i_recv_cq = NULL;
252 rdsdebug("ib_create_cq recv failed: %d\n", ret);
253 goto out;
254 }
255
256 ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
257 if (ret) {
258 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
259 goto out;
260 }
261
262 ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
263 if (ret) {
264 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
265 goto out;
266 }
267
268 /* XXX negotiate max send/recv with remote? */
269 memset(&attr, 0, sizeof(attr));
270 attr.event_handler = rds_ib_qp_event_handler;
271 attr.qp_context = conn;
272 /* + 1 to allow for the single ack message */
273 attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1;
274 attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
275 attr.cap.max_send_sge = rds_ibdev->max_sge;
276 attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
277 attr.sq_sig_type = IB_SIGNAL_REQ_WR;
278 attr.qp_type = IB_QPT_RC;
279 attr.send_cq = ic->i_send_cq;
280 attr.recv_cq = ic->i_recv_cq;
281
282 /*
283 * XXX this can fail if max_*_wr is too large? Are we supposed
284 * to back off until we get a value that the hardware can support?
285 */
286 ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
287 if (ret) {
288 rdsdebug("rdma_create_qp failed: %d\n", ret);
289 goto out;
290 }
291
292 ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
293 ic->i_send_ring.w_nr *
294 sizeof(struct rds_header),
295 &ic->i_send_hdrs_dma, GFP_KERNEL);
296 if (ic->i_send_hdrs == NULL) {
297 ret = -ENOMEM;
298 rdsdebug("ib_dma_alloc_coherent send failed\n");
299 goto out;
300 }
301
302 ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
303 ic->i_recv_ring.w_nr *
304 sizeof(struct rds_header),
305 &ic->i_recv_hdrs_dma, GFP_KERNEL);
306 if (ic->i_recv_hdrs == NULL) {
307 ret = -ENOMEM;
308 rdsdebug("ib_dma_alloc_coherent recv failed\n");
309 goto out;
310 }
311
312 ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
313 &ic->i_ack_dma, GFP_KERNEL);
314 if (ic->i_ack == NULL) {
315 ret = -ENOMEM;
316 rdsdebug("ib_dma_alloc_coherent ack failed\n");
317 goto out;
318 }
319
320 ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work));
321 if (ic->i_sends == NULL) {
322 ret = -ENOMEM;
323 rdsdebug("send allocation failed\n");
324 goto out;
325 }
326 rds_ib_send_init_ring(ic);
327
328 ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work));
329 if (ic->i_recvs == NULL) {
330 ret = -ENOMEM;
331 rdsdebug("recv allocation failed\n");
332 goto out;
333 }
334
335 rds_ib_recv_init_ring(ic);
336 rds_ib_recv_init_ack(ic);
337
338 /* Post receive buffers - as a side effect, this will update
339 * the posted credit count. */
340 rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 1);
341
342 rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
343 ic->i_send_cq, ic->i_recv_cq);
344
345out:
346 return ret;
347}
348
349static u32 rds_ib_protocol_compatible(const struct rds_ib_connect_private *dp)
350{
351 u16 common;
352 u32 version = 0;
353
354 /* rdma_cm private data is odd - when there is any private data in the
355 * request, we will be given a pretty large buffer without telling us the
356 * original size. The only way to tell the difference is by looking at
357 * the contents, which are initialized to zero.
358 * If the protocol version fields aren't set, this is a connection attempt
359 * from an older version. This could could be 3.0 or 2.0 - we can't tell.
360 * We really should have changed this for OFED 1.3 :-( */
361 if (dp->dp_protocol_major == 0)
362 return RDS_PROTOCOL_3_0;
363
364 common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
365 if (dp->dp_protocol_major == 3 && common) {
366 version = RDS_PROTOCOL_3_0;
367 while ((common >>= 1) != 0)
368 version++;
369 } else if (printk_ratelimit()) {
370 printk(KERN_NOTICE "RDS: Connection from %pI4 using "
371 "incompatible protocol version %u.%u\n",
372 &dp->dp_saddr,
373 dp->dp_protocol_major,
374 dp->dp_protocol_minor);
375 }
376 return version;
377}
378
379int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
380 struct rdma_cm_event *event)
381{
382 __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
383 __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
384 const struct rds_ib_connect_private *dp = event->param.conn.private_data;
385 struct rds_ib_connect_private dp_rep;
386 struct rds_connection *conn = NULL;
387 struct rds_ib_connection *ic = NULL;
388 struct rdma_conn_param conn_param;
389 u32 version;
390 int err, destroy = 1;
391
392 /* Check whether the remote protocol version matches ours. */
393 version = rds_ib_protocol_compatible(dp);
394 if (!version)
395 goto out;
396
397 rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
398 "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
399 RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
400 (unsigned long long)be64_to_cpu(lguid),
401 (unsigned long long)be64_to_cpu(fguid));
402
403 conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport,
404 GFP_KERNEL);
405 if (IS_ERR(conn)) {
406 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
407 conn = NULL;
408 goto out;
409 }
410
411 /*
412 * The connection request may occur while the
413 * previous connection exist, e.g. in case of failover.
414 * But as connections may be initiated simultaneously
415 * by both hosts, we have a random backoff mechanism -
416 * see the comment above rds_queue_reconnect()
417 */
418 mutex_lock(&conn->c_cm_lock);
419 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
420 if (rds_conn_state(conn) == RDS_CONN_UP) {
421 rdsdebug("incoming connect while connecting\n");
422 rds_conn_drop(conn);
423 rds_ib_stats_inc(s_ib_listen_closed_stale);
424 } else
425 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
426 /* Wait and see - our connect may still be succeeding */
427 rds_ib_stats_inc(s_ib_connect_raced);
428 }
429 mutex_unlock(&conn->c_cm_lock);
430 goto out;
431 }
432
433 ic = conn->c_transport_data;
434
435 rds_ib_set_protocol(conn, version);
436 rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
437
438 /* If the peer gave us the last packet it saw, process this as if
439 * we had received a regular ACK. */
440 if (dp->dp_ack_seq)
441 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
442
443 BUG_ON(cm_id->context);
444 BUG_ON(ic->i_cm_id);
445
446 ic->i_cm_id = cm_id;
447 cm_id->context = conn;
448
449 /* We got halfway through setting up the ib_connection, if we
450 * fail now, we have to take the long route out of this mess. */
451 destroy = 0;
452
453 err = rds_ib_setup_qp(conn);
454 if (err) {
455 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
456 goto out;
457 }
458
459 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version);
460
461 /* rdma_accept() calls rdma_reject() internally if it fails */
462 err = rdma_accept(cm_id, &conn_param);
463 mutex_unlock(&conn->c_cm_lock);
464 if (err) {
465 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
466 goto out;
467 }
468
469 return 0;
470
471out:
472 rdma_reject(cm_id, NULL, 0);
473 return destroy;
474}
475
476
477int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
478{
479 struct rds_connection *conn = cm_id->context;
480 struct rds_ib_connection *ic = conn->c_transport_data;
481 struct rdma_conn_param conn_param;
482 struct rds_ib_connect_private dp;
483 int ret;
484
485 /* If the peer doesn't do protocol negotiation, we must
486 * default to RDSv3.0 */
487 rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
488 ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */
489
490 ret = rds_ib_setup_qp(conn);
491 if (ret) {
492 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
493 goto out;
494 }
495
496 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION);
497
498 ret = rdma_connect(cm_id, &conn_param);
499 if (ret)
500 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
501
502out:
503 /* Beware - returning non-zero tells the rdma_cm to destroy
504 * the cm_id. We should certainly not do it as long as we still
505 * "own" the cm_id. */
506 if (ret) {
507 if (ic->i_cm_id == cm_id)
508 ret = 0;
509 }
510 return ret;
511}
512
513int rds_ib_conn_connect(struct rds_connection *conn)
514{
515 struct rds_ib_connection *ic = conn->c_transport_data;
516 struct sockaddr_in src, dest;
517 int ret;
518
519 /* XXX I wonder what affect the port space has */
520 /* delegate cm event handler to rdma_transport */
521 ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
522 RDMA_PS_TCP);
523 if (IS_ERR(ic->i_cm_id)) {
524 ret = PTR_ERR(ic->i_cm_id);
525 ic->i_cm_id = NULL;
526 rdsdebug("rdma_create_id() failed: %d\n", ret);
527 goto out;
528 }
529
530 rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
531
532 src.sin_family = AF_INET;
533 src.sin_addr.s_addr = (__force u32)conn->c_laddr;
534 src.sin_port = (__force u16)htons(0);
535
536 dest.sin_family = AF_INET;
537 dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
538 dest.sin_port = (__force u16)htons(RDS_PORT);
539
540 ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
541 (struct sockaddr *)&dest,
542 RDS_RDMA_RESOLVE_TIMEOUT_MS);
543 if (ret) {
544 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
545 ret);
546 rdma_destroy_id(ic->i_cm_id);
547 ic->i_cm_id = NULL;
548 }
549
550out:
551 return ret;
552}
553
554/*
555 * This is so careful about only cleaning up resources that were built up
556 * so that it can be called at any point during startup. In fact it
557 * can be called multiple times for a given connection.
558 */
559void rds_ib_conn_shutdown(struct rds_connection *conn)
560{
561 struct rds_ib_connection *ic = conn->c_transport_data;
562 int err = 0;
563
564 rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
565 ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
566 ic->i_cm_id ? ic->i_cm_id->qp : NULL);
567
568 if (ic->i_cm_id) {
569 struct ib_device *dev = ic->i_cm_id->device;
570
571 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
572 err = rdma_disconnect(ic->i_cm_id);
573 if (err) {
574 /* Actually this may happen quite frequently, when
575 * an outgoing connect raced with an incoming connect.
576 */
577 rdsdebug("failed to disconnect, cm: %p err %d\n",
578 ic->i_cm_id, err);
579 }
580
581 wait_event(rds_ib_ring_empty_wait,
582 rds_ib_ring_empty(&ic->i_send_ring) &&
583 rds_ib_ring_empty(&ic->i_recv_ring));
584
585 if (ic->i_send_hdrs)
586 ib_dma_free_coherent(dev,
587 ic->i_send_ring.w_nr *
588 sizeof(struct rds_header),
589 ic->i_send_hdrs,
590 ic->i_send_hdrs_dma);
591
592 if (ic->i_recv_hdrs)
593 ib_dma_free_coherent(dev,
594 ic->i_recv_ring.w_nr *
595 sizeof(struct rds_header),
596 ic->i_recv_hdrs,
597 ic->i_recv_hdrs_dma);
598
599 if (ic->i_ack)
600 ib_dma_free_coherent(dev, sizeof(struct rds_header),
601 ic->i_ack, ic->i_ack_dma);
602
603 if (ic->i_sends)
604 rds_ib_send_clear_ring(ic);
605 if (ic->i_recvs)
606 rds_ib_recv_clear_ring(ic);
607
608 if (ic->i_cm_id->qp)
609 rdma_destroy_qp(ic->i_cm_id);
610 if (ic->i_send_cq)
611 ib_destroy_cq(ic->i_send_cq);
612 if (ic->i_recv_cq)
613 ib_destroy_cq(ic->i_recv_cq);
614 rdma_destroy_id(ic->i_cm_id);
615
616 /*
617 * Move connection back to the nodev list.
618 */
619 if (ic->rds_ibdev) {
620
621 spin_lock_irq(&ic->rds_ibdev->spinlock);
622 BUG_ON(list_empty(&ic->ib_node));
623 list_del(&ic->ib_node);
624 spin_unlock_irq(&ic->rds_ibdev->spinlock);
625
626 spin_lock_irq(&ib_nodev_conns_lock);
627 list_add_tail(&ic->ib_node, &ib_nodev_conns);
628 spin_unlock_irq(&ib_nodev_conns_lock);
629 ic->rds_ibdev = NULL;
630 }
631
632 ic->i_cm_id = NULL;
633 ic->i_pd = NULL;
634 ic->i_mr = NULL;
635 ic->i_send_cq = NULL;
636 ic->i_recv_cq = NULL;
637 ic->i_send_hdrs = NULL;
638 ic->i_recv_hdrs = NULL;
639 ic->i_ack = NULL;
640 }
641 BUG_ON(ic->rds_ibdev);
642
643 /* Clear pending transmit */
644 if (ic->i_rm) {
645 rds_message_put(ic->i_rm);
646 ic->i_rm = NULL;
647 }
648
649 /* Clear the ACK state */
650 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
651 rds_ib_set_64bit(&ic->i_ack_next, 0);
652 ic->i_ack_recv = 0;
653
654 /* Clear flow control state */
655 ic->i_flowctl = 0;
656 atomic_set(&ic->i_credits, 0);
657
658 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
659 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
660
661 if (ic->i_ibinc) {
662 rds_inc_put(&ic->i_ibinc->ii_inc);
663 ic->i_ibinc = NULL;
664 }
665
666 vfree(ic->i_sends);
667 ic->i_sends = NULL;
668 vfree(ic->i_recvs);
669 ic->i_recvs = NULL;
670}
671
672int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
673{
674 struct rds_ib_connection *ic;
675 unsigned long flags;
676
677 /* XXX too lazy? */
678 ic = kzalloc(sizeof(struct rds_ib_connection), GFP_KERNEL);
679 if (ic == NULL)
680 return -ENOMEM;
681
682 INIT_LIST_HEAD(&ic->ib_node);
683 mutex_init(&ic->i_recv_mutex);
684
685 /*
686 * rds_ib_conn_shutdown() waits for these to be emptied so they
687 * must be initialized before it can be called.
688 */
689 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
690 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
691
692 ic->conn = conn;
693 conn->c_transport_data = ic;
694
695 spin_lock_irqsave(&ib_nodev_conns_lock, flags);
696 list_add_tail(&ic->ib_node, &ib_nodev_conns);
697 spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
698
699
700 rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
701 return 0;
702}
703
704void rds_ib_conn_free(void *arg)
705{
706 struct rds_ib_connection *ic = arg;
707 rdsdebug("ic %p\n", ic);
708 list_del(&ic->ib_node);
709 kfree(ic);
710}
711
712
713/*
714 * An error occurred on the connection
715 */
716void
717__rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
718{
719 va_list ap;
720
721 rds_conn_drop(conn);
722
723 va_start(ap, fmt);
724 vprintk(fmt, ap);
725 va_end(ap);
726}
generated by cgit v1.2.2 (git 2.25.0) at 2024-09-21 23:31:40 -0400