aboutsummaryrefslogtreecommitdiffstats
path: root/fs/ceph/caps.c
diff options
context:
space:
mode:
Diffstat (limited to 'fs/ceph/caps.c')
-rw-r--r--fs/ceph/caps.c2955
1 files changed, 2955 insertions, 0 deletions
diff --git a/fs/ceph/caps.c b/fs/ceph/caps.c
new file mode 100644
index 000000000000..aa2239fa9a3b
--- /dev/null
+++ b/fs/ceph/caps.c
@@ -0,0 +1,2955 @@
1#include "ceph_debug.h"
2
3#include <linux/fs.h>
4#include <linux/kernel.h>
5#include <linux/sched.h>
6#include <linux/slab.h>
7#include <linux/vmalloc.h>
8#include <linux/wait.h>
9#include <linux/writeback.h>
10
11#include "super.h"
12#include "decode.h"
13#include "messenger.h"
14
15/*
16 * Capability management
17 *
18 * The Ceph metadata servers control client access to inode metadata
19 * and file data by issuing capabilities, granting clients permission
20 * to read and/or write both inode field and file data to OSDs
21 * (storage nodes). Each capability consists of a set of bits
22 * indicating which operations are allowed.
23 *
24 * If the client holds a *_SHARED cap, the client has a coherent value
25 * that can be safely read from the cached inode.
26 *
27 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
28 * client is allowed to change inode attributes (e.g., file size,
29 * mtime), note its dirty state in the ceph_cap, and asynchronously
30 * flush that metadata change to the MDS.
31 *
32 * In the event of a conflicting operation (perhaps by another
33 * client), the MDS will revoke the conflicting client capabilities.
34 *
35 * In order for a client to cache an inode, it must hold a capability
36 * with at least one MDS server. When inodes are released, release
37 * notifications are batched and periodically sent en masse to the MDS
38 * cluster to release server state.
39 */
40
41
42/*
43 * Generate readable cap strings for debugging output.
44 */
45#define MAX_CAP_STR 20
46static char cap_str[MAX_CAP_STR][40];
47static DEFINE_SPINLOCK(cap_str_lock);
48static int last_cap_str;
49
50static char *gcap_string(char *s, int c)
51{
52 if (c & CEPH_CAP_GSHARED)
53 *s++ = 's';
54 if (c & CEPH_CAP_GEXCL)
55 *s++ = 'x';
56 if (c & CEPH_CAP_GCACHE)
57 *s++ = 'c';
58 if (c & CEPH_CAP_GRD)
59 *s++ = 'r';
60 if (c & CEPH_CAP_GWR)
61 *s++ = 'w';
62 if (c & CEPH_CAP_GBUFFER)
63 *s++ = 'b';
64 if (c & CEPH_CAP_GLAZYIO)
65 *s++ = 'l';
66 return s;
67}
68
69const char *ceph_cap_string(int caps)
70{
71 int i;
72 char *s;
73 int c;
74
75 spin_lock(&cap_str_lock);
76 i = last_cap_str++;
77 if (last_cap_str == MAX_CAP_STR)
78 last_cap_str = 0;
79 spin_unlock(&cap_str_lock);
80
81 s = cap_str[i];
82
83 if (caps & CEPH_CAP_PIN)
84 *s++ = 'p';
85
86 c = (caps >> CEPH_CAP_SAUTH) & 3;
87 if (c) {
88 *s++ = 'A';
89 s = gcap_string(s, c);
90 }
91
92 c = (caps >> CEPH_CAP_SLINK) & 3;
93 if (c) {
94 *s++ = 'L';
95 s = gcap_string(s, c);
96 }
97
98 c = (caps >> CEPH_CAP_SXATTR) & 3;
99 if (c) {
100 *s++ = 'X';
101 s = gcap_string(s, c);
102 }
103
104 c = caps >> CEPH_CAP_SFILE;
105 if (c) {
106 *s++ = 'F';
107 s = gcap_string(s, c);
108 }
109
110 if (s == cap_str[i])
111 *s++ = '-';
112 *s = 0;
113 return cap_str[i];
114}
115
116/*
117 * Cap reservations
118 *
119 * Maintain a global pool of preallocated struct ceph_caps, referenced
120 * by struct ceph_caps_reservations. This ensures that we preallocate
121 * memory needed to successfully process an MDS response. (If an MDS
122 * sends us cap information and we fail to process it, we will have
123 * problems due to the client and MDS being out of sync.)
124 *
125 * Reservations are 'owned' by a ceph_cap_reservation context.
126 */
127static spinlock_t caps_list_lock;
128static struct list_head caps_list; /* unused (reserved or unreserved) */
129static int caps_total_count; /* total caps allocated */
130static int caps_use_count; /* in use */
131static int caps_reserve_count; /* unused, reserved */
132static int caps_avail_count; /* unused, unreserved */
133static int caps_min_count; /* keep at least this many (unreserved) */
134
135void __init ceph_caps_init(void)
136{
137 INIT_LIST_HEAD(&caps_list);
138 spin_lock_init(&caps_list_lock);
139}
140
141void ceph_caps_finalize(void)
142{
143 struct ceph_cap *cap;
144
145 spin_lock(&caps_list_lock);
146 while (!list_empty(&caps_list)) {
147 cap = list_first_entry(&caps_list, struct ceph_cap, caps_item);
148 list_del(&cap->caps_item);
149 kmem_cache_free(ceph_cap_cachep, cap);
150 }
151 caps_total_count = 0;
152 caps_avail_count = 0;
153 caps_use_count = 0;
154 caps_reserve_count = 0;
155 caps_min_count = 0;
156 spin_unlock(&caps_list_lock);
157}
158
159void ceph_adjust_min_caps(int delta)
160{
161 spin_lock(&caps_list_lock);
162 caps_min_count += delta;
163 BUG_ON(caps_min_count < 0);
164 spin_unlock(&caps_list_lock);
165}
166
167int ceph_reserve_caps(struct ceph_cap_reservation *ctx, int need)
168{
169 int i;
170 struct ceph_cap *cap;
171 int have;
172 int alloc = 0;
173 LIST_HEAD(newcaps);
174 int ret = 0;
175
176 dout("reserve caps ctx=%p need=%d\n", ctx, need);
177
178 /* first reserve any caps that are already allocated */
179 spin_lock(&caps_list_lock);
180 if (caps_avail_count >= need)
181 have = need;
182 else
183 have = caps_avail_count;
184 caps_avail_count -= have;
185 caps_reserve_count += have;
186 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
187 caps_avail_count);
188 spin_unlock(&caps_list_lock);
189
190 for (i = have; i < need; i++) {
191 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
192 if (!cap) {
193 ret = -ENOMEM;
194 goto out_alloc_count;
195 }
196 list_add(&cap->caps_item, &newcaps);
197 alloc++;
198 }
199 BUG_ON(have + alloc != need);
200
201 spin_lock(&caps_list_lock);
202 caps_total_count += alloc;
203 caps_reserve_count += alloc;
204 list_splice(&newcaps, &caps_list);
205
206 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
207 caps_avail_count);
208 spin_unlock(&caps_list_lock);
209
210 ctx->count = need;
211 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
212 ctx, caps_total_count, caps_use_count, caps_reserve_count,
213 caps_avail_count);
214 return 0;
215
216out_alloc_count:
217 /* we didn't manage to reserve as much as we needed */
218 pr_warning("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
219 ctx, need, have);
220 return ret;
221}
222
223int ceph_unreserve_caps(struct ceph_cap_reservation *ctx)
224{
225 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
226 if (ctx->count) {
227 spin_lock(&caps_list_lock);
228 BUG_ON(caps_reserve_count < ctx->count);
229 caps_reserve_count -= ctx->count;
230 caps_avail_count += ctx->count;
231 ctx->count = 0;
232 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
233 caps_total_count, caps_use_count, caps_reserve_count,
234 caps_avail_count);
235 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
236 caps_avail_count);
237 spin_unlock(&caps_list_lock);
238 }
239 return 0;
240}
241
242static struct ceph_cap *get_cap(struct ceph_cap_reservation *ctx)
243{
244 struct ceph_cap *cap = NULL;
245
246 /* temporary, until we do something about cap import/export */
247 if (!ctx)
248 return kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
249
250 spin_lock(&caps_list_lock);
251 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
252 ctx, ctx->count, caps_total_count, caps_use_count,
253 caps_reserve_count, caps_avail_count);
254 BUG_ON(!ctx->count);
255 BUG_ON(ctx->count > caps_reserve_count);
256 BUG_ON(list_empty(&caps_list));
257
258 ctx->count--;
259 caps_reserve_count--;
260 caps_use_count++;
261
262 cap = list_first_entry(&caps_list, struct ceph_cap, caps_item);
263 list_del(&cap->caps_item);
264
265 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
266 caps_avail_count);
267 spin_unlock(&caps_list_lock);
268 return cap;
269}
270
271void ceph_put_cap(struct ceph_cap *cap)
272{
273 spin_lock(&caps_list_lock);
274 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
275 cap, caps_total_count, caps_use_count,
276 caps_reserve_count, caps_avail_count);
277 caps_use_count--;
278 /*
279 * Keep some preallocated caps around (ceph_min_count), to
280 * avoid lots of free/alloc churn.
281 */
282 if (caps_avail_count >= caps_reserve_count + caps_min_count) {
283 caps_total_count--;
284 kmem_cache_free(ceph_cap_cachep, cap);
285 } else {
286 caps_avail_count++;
287 list_add(&cap->caps_item, &caps_list);
288 }
289
290 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
291 caps_avail_count);
292 spin_unlock(&caps_list_lock);
293}
294
295void ceph_reservation_status(struct ceph_client *client,
296 int *total, int *avail, int *used, int *reserved,
297 int *min)
298{
299 if (total)
300 *total = caps_total_count;
301 if (avail)
302 *avail = caps_avail_count;
303 if (used)
304 *used = caps_use_count;
305 if (reserved)
306 *reserved = caps_reserve_count;
307 if (min)
308 *min = caps_min_count;
309}
310
311/*
312 * Find ceph_cap for given mds, if any.
313 *
314 * Called with i_lock held.
315 */
316static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
317{
318 struct ceph_cap *cap;
319 struct rb_node *n = ci->i_caps.rb_node;
320
321 while (n) {
322 cap = rb_entry(n, struct ceph_cap, ci_node);
323 if (mds < cap->mds)
324 n = n->rb_left;
325 else if (mds > cap->mds)
326 n = n->rb_right;
327 else
328 return cap;
329 }
330 return NULL;
331}
332
333/*
334 * Return id of any MDS with a cap, preferably FILE_WR|WRBUFFER|EXCL, else
335 * -1.
336 */
337static int __ceph_get_cap_mds(struct ceph_inode_info *ci, u32 *mseq)
338{
339 struct ceph_cap *cap;
340 int mds = -1;
341 struct rb_node *p;
342
343 /* prefer mds with WR|WRBUFFER|EXCL caps */
344 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
345 cap = rb_entry(p, struct ceph_cap, ci_node);
346 mds = cap->mds;
347 if (mseq)
348 *mseq = cap->mseq;
349 if (cap->issued & (CEPH_CAP_FILE_WR |
350 CEPH_CAP_FILE_BUFFER |
351 CEPH_CAP_FILE_EXCL))
352 break;
353 }
354 return mds;
355}
356
357int ceph_get_cap_mds(struct inode *inode)
358{
359 int mds;
360 spin_lock(&inode->i_lock);
361 mds = __ceph_get_cap_mds(ceph_inode(inode), NULL);
362 spin_unlock(&inode->i_lock);
363 return mds;
364}
365
366/*
367 * Called under i_lock.
368 */
369static void __insert_cap_node(struct ceph_inode_info *ci,
370 struct ceph_cap *new)
371{
372 struct rb_node **p = &ci->i_caps.rb_node;
373 struct rb_node *parent = NULL;
374 struct ceph_cap *cap = NULL;
375
376 while (*p) {
377 parent = *p;
378 cap = rb_entry(parent, struct ceph_cap, ci_node);
379 if (new->mds < cap->mds)
380 p = &(*p)->rb_left;
381 else if (new->mds > cap->mds)
382 p = &(*p)->rb_right;
383 else
384 BUG();
385 }
386
387 rb_link_node(&new->ci_node, parent, p);
388 rb_insert_color(&new->ci_node, &ci->i_caps);
389}
390
391/*
392 * (re)set cap hold timeouts, which control the delayed release
393 * of unused caps back to the MDS. Should be called on cap use.
394 */
395static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
396 struct ceph_inode_info *ci)
397{
398 struct ceph_mount_args *ma = mdsc->client->mount_args;
399
400 ci->i_hold_caps_min = round_jiffies(jiffies +
401 ma->caps_wanted_delay_min * HZ);
402 ci->i_hold_caps_max = round_jiffies(jiffies +
403 ma->caps_wanted_delay_max * HZ);
404 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
405 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
406}
407
408/*
409 * (Re)queue cap at the end of the delayed cap release list.
410 *
411 * If I_FLUSH is set, leave the inode at the front of the list.
412 *
413 * Caller holds i_lock
414 * -> we take mdsc->cap_delay_lock
415 */
416static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
417 struct ceph_inode_info *ci)
418{
419 __cap_set_timeouts(mdsc, ci);
420 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
421 ci->i_ceph_flags, ci->i_hold_caps_max);
422 if (!mdsc->stopping) {
423 spin_lock(&mdsc->cap_delay_lock);
424 if (!list_empty(&ci->i_cap_delay_list)) {
425 if (ci->i_ceph_flags & CEPH_I_FLUSH)
426 goto no_change;
427 list_del_init(&ci->i_cap_delay_list);
428 }
429 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
430no_change:
431 spin_unlock(&mdsc->cap_delay_lock);
432 }
433}
434
435/*
436 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
437 * indicating we should send a cap message to flush dirty metadata
438 * asap, and move to the front of the delayed cap list.
439 */
440static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
441 struct ceph_inode_info *ci)
442{
443 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
444 spin_lock(&mdsc->cap_delay_lock);
445 ci->i_ceph_flags |= CEPH_I_FLUSH;
446 if (!list_empty(&ci->i_cap_delay_list))
447 list_del_init(&ci->i_cap_delay_list);
448 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
449 spin_unlock(&mdsc->cap_delay_lock);
450}
451
452/*
453 * Cancel delayed work on cap.
454 *
455 * Caller must hold i_lock.
456 */
457static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
458 struct ceph_inode_info *ci)
459{
460 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
461 if (list_empty(&ci->i_cap_delay_list))
462 return;
463 spin_lock(&mdsc->cap_delay_lock);
464 list_del_init(&ci->i_cap_delay_list);
465 spin_unlock(&mdsc->cap_delay_lock);
466}
467
468/*
469 * Common issue checks for add_cap, handle_cap_grant.
470 */
471static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
472 unsigned issued)
473{
474 unsigned had = __ceph_caps_issued(ci, NULL);
475
476 /*
477 * Each time we receive FILE_CACHE anew, we increment
478 * i_rdcache_gen.
479 */
480 if ((issued & CEPH_CAP_FILE_CACHE) &&
481 (had & CEPH_CAP_FILE_CACHE) == 0)
482 ci->i_rdcache_gen++;
483
484 /*
485 * if we are newly issued FILE_SHARED, clear I_COMPLETE; we
486 * don't know what happened to this directory while we didn't
487 * have the cap.
488 */
489 if ((issued & CEPH_CAP_FILE_SHARED) &&
490 (had & CEPH_CAP_FILE_SHARED) == 0) {
491 ci->i_shared_gen++;
492 if (S_ISDIR(ci->vfs_inode.i_mode)) {
493 dout(" marking %p NOT complete\n", &ci->vfs_inode);
494 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
495 }
496 }
497}
498
499/*
500 * Add a capability under the given MDS session.
501 *
502 * Caller should hold session snap_rwsem (read) and s_mutex.
503 *
504 * @fmode is the open file mode, if we are opening a file, otherwise
505 * it is < 0. (This is so we can atomically add the cap and add an
506 * open file reference to it.)
507 */
508int ceph_add_cap(struct inode *inode,
509 struct ceph_mds_session *session, u64 cap_id,
510 int fmode, unsigned issued, unsigned wanted,
511 unsigned seq, unsigned mseq, u64 realmino, int flags,
512 struct ceph_cap_reservation *caps_reservation)
513{
514 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
515 struct ceph_inode_info *ci = ceph_inode(inode);
516 struct ceph_cap *new_cap = NULL;
517 struct ceph_cap *cap;
518 int mds = session->s_mds;
519 int actual_wanted;
520
521 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
522 session->s_mds, cap_id, ceph_cap_string(issued), seq);
523
524 /*
525 * If we are opening the file, include file mode wanted bits
526 * in wanted.
527 */
528 if (fmode >= 0)
529 wanted |= ceph_caps_for_mode(fmode);
530
531retry:
532 spin_lock(&inode->i_lock);
533 cap = __get_cap_for_mds(ci, mds);
534 if (!cap) {
535 if (new_cap) {
536 cap = new_cap;
537 new_cap = NULL;
538 } else {
539 spin_unlock(&inode->i_lock);
540 new_cap = get_cap(caps_reservation);
541 if (new_cap == NULL)
542 return -ENOMEM;
543 goto retry;
544 }
545
546 cap->issued = 0;
547 cap->implemented = 0;
548 cap->mds = mds;
549 cap->mds_wanted = 0;
550
551 cap->ci = ci;
552 __insert_cap_node(ci, cap);
553
554 /* clear out old exporting info? (i.e. on cap import) */
555 if (ci->i_cap_exporting_mds == mds) {
556 ci->i_cap_exporting_issued = 0;
557 ci->i_cap_exporting_mseq = 0;
558 ci->i_cap_exporting_mds = -1;
559 }
560
561 /* add to session cap list */
562 cap->session = session;
563 spin_lock(&session->s_cap_lock);
564 list_add_tail(&cap->session_caps, &session->s_caps);
565 session->s_nr_caps++;
566 spin_unlock(&session->s_cap_lock);
567 }
568
569 if (!ci->i_snap_realm) {
570 /*
571 * add this inode to the appropriate snap realm
572 */
573 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
574 realmino);
575 if (realm) {
576 ceph_get_snap_realm(mdsc, realm);
577 spin_lock(&realm->inodes_with_caps_lock);
578 ci->i_snap_realm = realm;
579 list_add(&ci->i_snap_realm_item,
580 &realm->inodes_with_caps);
581 spin_unlock(&realm->inodes_with_caps_lock);
582 } else {
583 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
584 realmino);
585 }
586 }
587
588 __check_cap_issue(ci, cap, issued);
589
590 /*
591 * If we are issued caps we don't want, or the mds' wanted
592 * value appears to be off, queue a check so we'll release
593 * later and/or update the mds wanted value.
594 */
595 actual_wanted = __ceph_caps_wanted(ci);
596 if ((wanted & ~actual_wanted) ||
597 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
598 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
599 ceph_cap_string(issued), ceph_cap_string(wanted),
600 ceph_cap_string(actual_wanted));
601 __cap_delay_requeue(mdsc, ci);
602 }
603
604 if (flags & CEPH_CAP_FLAG_AUTH)
605 ci->i_auth_cap = cap;
606 else if (ci->i_auth_cap == cap)
607 ci->i_auth_cap = NULL;
608
609 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
610 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
611 ceph_cap_string(issued|cap->issued), seq, mds);
612 cap->cap_id = cap_id;
613 cap->issued = issued;
614 cap->implemented |= issued;
615 cap->mds_wanted |= wanted;
616 cap->seq = seq;
617 cap->issue_seq = seq;
618 cap->mseq = mseq;
619 cap->cap_gen = session->s_cap_gen;
620
621 if (fmode >= 0)
622 __ceph_get_fmode(ci, fmode);
623 spin_unlock(&inode->i_lock);
624 wake_up(&ci->i_cap_wq);
625 return 0;
626}
627
628/*
629 * Return true if cap has not timed out and belongs to the current
630 * generation of the MDS session (i.e. has not gone 'stale' due to
631 * us losing touch with the mds).
632 */
633static int __cap_is_valid(struct ceph_cap *cap)
634{
635 unsigned long ttl;
636 u32 gen;
637
638 spin_lock(&cap->session->s_cap_lock);
639 gen = cap->session->s_cap_gen;
640 ttl = cap->session->s_cap_ttl;
641 spin_unlock(&cap->session->s_cap_lock);
642
643 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
644 dout("__cap_is_valid %p cap %p issued %s "
645 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
646 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
647 return 0;
648 }
649
650 return 1;
651}
652
653/*
654 * Return set of valid cap bits issued to us. Note that caps time
655 * out, and may be invalidated in bulk if the client session times out
656 * and session->s_cap_gen is bumped.
657 */
658int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
659{
660 int have = ci->i_snap_caps | ci->i_cap_exporting_issued;
661 struct ceph_cap *cap;
662 struct rb_node *p;
663
664 if (implemented)
665 *implemented = 0;
666 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
667 cap = rb_entry(p, struct ceph_cap, ci_node);
668 if (!__cap_is_valid(cap))
669 continue;
670 dout("__ceph_caps_issued %p cap %p issued %s\n",
671 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
672 have |= cap->issued;
673 if (implemented)
674 *implemented |= cap->implemented;
675 }
676 return have;
677}
678
679/*
680 * Get cap bits issued by caps other than @ocap
681 */
682int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
683{
684 int have = ci->i_snap_caps;
685 struct ceph_cap *cap;
686 struct rb_node *p;
687
688 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
689 cap = rb_entry(p, struct ceph_cap, ci_node);
690 if (cap == ocap)
691 continue;
692 if (!__cap_is_valid(cap))
693 continue;
694 have |= cap->issued;
695 }
696 return have;
697}
698
699/*
700 * Move a cap to the end of the LRU (oldest caps at list head, newest
701 * at list tail).
702 */
703static void __touch_cap(struct ceph_cap *cap)
704{
705 struct ceph_mds_session *s = cap->session;
706
707 spin_lock(&s->s_cap_lock);
708 if (s->s_cap_iterator == NULL) {
709 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
710 s->s_mds);
711 list_move_tail(&cap->session_caps, &s->s_caps);
712 } else {
713 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
714 &cap->ci->vfs_inode, cap, s->s_mds);
715 }
716 spin_unlock(&s->s_cap_lock);
717}
718
719/*
720 * Check if we hold the given mask. If so, move the cap(s) to the
721 * front of their respective LRUs. (This is the preferred way for
722 * callers to check for caps they want.)
723 */
724int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
725{
726 struct ceph_cap *cap;
727 struct rb_node *p;
728 int have = ci->i_snap_caps;
729
730 if ((have & mask) == mask) {
731 dout("__ceph_caps_issued_mask %p snap issued %s"
732 " (mask %s)\n", &ci->vfs_inode,
733 ceph_cap_string(have),
734 ceph_cap_string(mask));
735 return 1;
736 }
737
738 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
739 cap = rb_entry(p, struct ceph_cap, ci_node);
740 if (!__cap_is_valid(cap))
741 continue;
742 if ((cap->issued & mask) == mask) {
743 dout("__ceph_caps_issued_mask %p cap %p issued %s"
744 " (mask %s)\n", &ci->vfs_inode, cap,
745 ceph_cap_string(cap->issued),
746 ceph_cap_string(mask));
747 if (touch)
748 __touch_cap(cap);
749 return 1;
750 }
751
752 /* does a combination of caps satisfy mask? */
753 have |= cap->issued;
754 if ((have & mask) == mask) {
755 dout("__ceph_caps_issued_mask %p combo issued %s"
756 " (mask %s)\n", &ci->vfs_inode,
757 ceph_cap_string(cap->issued),
758 ceph_cap_string(mask));
759 if (touch) {
760 struct rb_node *q;
761
762 /* touch this + preceeding caps */
763 __touch_cap(cap);
764 for (q = rb_first(&ci->i_caps); q != p;
765 q = rb_next(q)) {
766 cap = rb_entry(q, struct ceph_cap,
767 ci_node);
768 if (!__cap_is_valid(cap))
769 continue;
770 __touch_cap(cap);
771 }
772 }
773 return 1;
774 }
775 }
776
777 return 0;
778}
779
780/*
781 * Return true if mask caps are currently being revoked by an MDS.
782 */
783int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
784{
785 struct inode *inode = &ci->vfs_inode;
786 struct ceph_cap *cap;
787 struct rb_node *p;
788 int ret = 0;
789
790 spin_lock(&inode->i_lock);
791 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
792 cap = rb_entry(p, struct ceph_cap, ci_node);
793 if (__cap_is_valid(cap) &&
794 (cap->implemented & ~cap->issued & mask)) {
795 ret = 1;
796 break;
797 }
798 }
799 spin_unlock(&inode->i_lock);
800 dout("ceph_caps_revoking %p %s = %d\n", inode,
801 ceph_cap_string(mask), ret);
802 return ret;
803}
804
805int __ceph_caps_used(struct ceph_inode_info *ci)
806{
807 int used = 0;
808 if (ci->i_pin_ref)
809 used |= CEPH_CAP_PIN;
810 if (ci->i_rd_ref)
811 used |= CEPH_CAP_FILE_RD;
812 if (ci->i_rdcache_ref || ci->i_rdcache_gen)
813 used |= CEPH_CAP_FILE_CACHE;
814 if (ci->i_wr_ref)
815 used |= CEPH_CAP_FILE_WR;
816 if (ci->i_wrbuffer_ref)
817 used |= CEPH_CAP_FILE_BUFFER;
818 return used;
819}
820
821/*
822 * wanted, by virtue of open file modes
823 */
824int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
825{
826 int want = 0;
827 int mode;
828 for (mode = 0; mode < 4; mode++)
829 if (ci->i_nr_by_mode[mode])
830 want |= ceph_caps_for_mode(mode);
831 return want;
832}
833
834/*
835 * Return caps we have registered with the MDS(s) as 'wanted'.
836 */
837int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
838{
839 struct ceph_cap *cap;
840 struct rb_node *p;
841 int mds_wanted = 0;
842
843 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
844 cap = rb_entry(p, struct ceph_cap, ci_node);
845 if (!__cap_is_valid(cap))
846 continue;
847 mds_wanted |= cap->mds_wanted;
848 }
849 return mds_wanted;
850}
851
852/*
853 * called under i_lock
854 */
855static int __ceph_is_any_caps(struct ceph_inode_info *ci)
856{
857 return !RB_EMPTY_ROOT(&ci->i_caps) || ci->i_cap_exporting_mds >= 0;
858}
859
860/*
861 * caller should hold i_lock.
862 * caller will not hold session s_mutex if called from destroy_inode.
863 */
864void __ceph_remove_cap(struct ceph_cap *cap)
865{
866 struct ceph_mds_session *session = cap->session;
867 struct ceph_inode_info *ci = cap->ci;
868 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc;
869
870 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
871
872 /* remove from inode list */
873 rb_erase(&cap->ci_node, &ci->i_caps);
874 cap->ci = NULL;
875 if (ci->i_auth_cap == cap)
876 ci->i_auth_cap = NULL;
877
878 /* remove from session list */
879 spin_lock(&session->s_cap_lock);
880 if (session->s_cap_iterator == cap) {
881 /* not yet, we are iterating over this very cap */
882 dout("__ceph_remove_cap delaying %p removal from session %p\n",
883 cap, cap->session);
884 } else {
885 list_del_init(&cap->session_caps);
886 session->s_nr_caps--;
887 cap->session = NULL;
888 }
889 spin_unlock(&session->s_cap_lock);
890
891 if (cap->session == NULL)
892 ceph_put_cap(cap);
893
894 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) {
895 struct ceph_snap_realm *realm = ci->i_snap_realm;
896 spin_lock(&realm->inodes_with_caps_lock);
897 list_del_init(&ci->i_snap_realm_item);
898 ci->i_snap_realm_counter++;
899 ci->i_snap_realm = NULL;
900 spin_unlock(&realm->inodes_with_caps_lock);
901 ceph_put_snap_realm(mdsc, realm);
902 }
903 if (!__ceph_is_any_real_caps(ci))
904 __cap_delay_cancel(mdsc, ci);
905}
906
907/*
908 * Build and send a cap message to the given MDS.
909 *
910 * Caller should be holding s_mutex.
911 */
912static int send_cap_msg(struct ceph_mds_session *session,
913 u64 ino, u64 cid, int op,
914 int caps, int wanted, int dirty,
915 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq,
916 u64 size, u64 max_size,
917 struct timespec *mtime, struct timespec *atime,
918 u64 time_warp_seq,
919 uid_t uid, gid_t gid, mode_t mode,
920 u64 xattr_version,
921 struct ceph_buffer *xattrs_buf,
922 u64 follows)
923{
924 struct ceph_mds_caps *fc;
925 struct ceph_msg *msg;
926
927 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
928 " seq %u/%u mseq %u follows %lld size %llu/%llu"
929 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
930 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
931 ceph_cap_string(dirty),
932 seq, issue_seq, mseq, follows, size, max_size,
933 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
934
935 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc), 0, 0, NULL);
936 if (IS_ERR(msg))
937 return PTR_ERR(msg);
938
939 msg->hdr.tid = cpu_to_le64(flush_tid);
940
941 fc = msg->front.iov_base;
942 memset(fc, 0, sizeof(*fc));
943
944 fc->cap_id = cpu_to_le64(cid);
945 fc->op = cpu_to_le32(op);
946 fc->seq = cpu_to_le32(seq);
947 fc->issue_seq = cpu_to_le32(issue_seq);
948 fc->migrate_seq = cpu_to_le32(mseq);
949 fc->caps = cpu_to_le32(caps);
950 fc->wanted = cpu_to_le32(wanted);
951 fc->dirty = cpu_to_le32(dirty);
952 fc->ino = cpu_to_le64(ino);
953 fc->snap_follows = cpu_to_le64(follows);
954
955 fc->size = cpu_to_le64(size);
956 fc->max_size = cpu_to_le64(max_size);
957 if (mtime)
958 ceph_encode_timespec(&fc->mtime, mtime);
959 if (atime)
960 ceph_encode_timespec(&fc->atime, atime);
961 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
962
963 fc->uid = cpu_to_le32(uid);
964 fc->gid = cpu_to_le32(gid);
965 fc->mode = cpu_to_le32(mode);
966
967 fc->xattr_version = cpu_to_le64(xattr_version);
968 if (xattrs_buf) {
969 msg->middle = ceph_buffer_get(xattrs_buf);
970 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
971 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
972 }
973
974 ceph_con_send(&session->s_con, msg);
975 return 0;
976}
977
978/*
979 * Queue cap releases when an inode is dropped from our cache. Since
980 * inode is about to be destroyed, there is no need for i_lock.
981 */
982void ceph_queue_caps_release(struct inode *inode)
983{
984 struct ceph_inode_info *ci = ceph_inode(inode);
985 struct rb_node *p;
986
987 p = rb_first(&ci->i_caps);
988 while (p) {
989 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
990 struct ceph_mds_session *session = cap->session;
991 struct ceph_msg *msg;
992 struct ceph_mds_cap_release *head;
993 struct ceph_mds_cap_item *item;
994
995 spin_lock(&session->s_cap_lock);
996 BUG_ON(!session->s_num_cap_releases);
997 msg = list_first_entry(&session->s_cap_releases,
998 struct ceph_msg, list_head);
999
1000 dout(" adding %p release to mds%d msg %p (%d left)\n",
1001 inode, session->s_mds, msg, session->s_num_cap_releases);
1002
1003 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE);
1004 head = msg->front.iov_base;
1005 head->num = cpu_to_le32(le32_to_cpu(head->num) + 1);
1006 item = msg->front.iov_base + msg->front.iov_len;
1007 item->ino = cpu_to_le64(ceph_ino(inode));
1008 item->cap_id = cpu_to_le64(cap->cap_id);
1009 item->migrate_seq = cpu_to_le32(cap->mseq);
1010 item->seq = cpu_to_le32(cap->issue_seq);
1011
1012 session->s_num_cap_releases--;
1013
1014 msg->front.iov_len += sizeof(*item);
1015 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1016 dout(" release msg %p full\n", msg);
1017 list_move_tail(&msg->list_head,
1018 &session->s_cap_releases_done);
1019 } else {
1020 dout(" release msg %p at %d/%d (%d)\n", msg,
1021 (int)le32_to_cpu(head->num),
1022 (int)CEPH_CAPS_PER_RELEASE,
1023 (int)msg->front.iov_len);
1024 }
1025 spin_unlock(&session->s_cap_lock);
1026 p = rb_next(p);
1027 __ceph_remove_cap(cap);
1028 }
1029}
1030
1031/*
1032 * Send a cap msg on the given inode. Update our caps state, then
1033 * drop i_lock and send the message.
1034 *
1035 * Make note of max_size reported/requested from mds, revoked caps
1036 * that have now been implemented.
1037 *
1038 * Make half-hearted attempt ot to invalidate page cache if we are
1039 * dropping RDCACHE. Note that this will leave behind locked pages
1040 * that we'll then need to deal with elsewhere.
1041 *
1042 * Return non-zero if delayed release, or we experienced an error
1043 * such that the caller should requeue + retry later.
1044 *
1045 * called with i_lock, then drops it.
1046 * caller should hold snap_rwsem (read), s_mutex.
1047 */
1048static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1049 int op, int used, int want, int retain, int flushing,
1050 unsigned *pflush_tid)
1051 __releases(cap->ci->vfs_inode->i_lock)
1052{
1053 struct ceph_inode_info *ci = cap->ci;
1054 struct inode *inode = &ci->vfs_inode;
1055 u64 cap_id = cap->cap_id;
1056 int held, revoking, dropping, keep;
1057 u64 seq, issue_seq, mseq, time_warp_seq, follows;
1058 u64 size, max_size;
1059 struct timespec mtime, atime;
1060 int wake = 0;
1061 mode_t mode;
1062 uid_t uid;
1063 gid_t gid;
1064 struct ceph_mds_session *session;
1065 u64 xattr_version = 0;
1066 int delayed = 0;
1067 u64 flush_tid = 0;
1068 int i;
1069 int ret;
1070
1071 held = cap->issued | cap->implemented;
1072 revoking = cap->implemented & ~cap->issued;
1073 retain &= ~revoking;
1074 dropping = cap->issued & ~retain;
1075
1076 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1077 inode, cap, cap->session,
1078 ceph_cap_string(held), ceph_cap_string(held & retain),
1079 ceph_cap_string(revoking));
1080 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1081
1082 session = cap->session;
1083
1084 /* don't release wanted unless we've waited a bit. */
1085 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1086 time_before(jiffies, ci->i_hold_caps_min)) {
1087 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1088 ceph_cap_string(cap->issued),
1089 ceph_cap_string(cap->issued & retain),
1090 ceph_cap_string(cap->mds_wanted),
1091 ceph_cap_string(want));
1092 want |= cap->mds_wanted;
1093 retain |= cap->issued;
1094 delayed = 1;
1095 }
1096 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1097
1098 cap->issued &= retain; /* drop bits we don't want */
1099 if (cap->implemented & ~cap->issued) {
1100 /*
1101 * Wake up any waiters on wanted -> needed transition.
1102 * This is due to the weird transition from buffered
1103 * to sync IO... we need to flush dirty pages _before_
1104 * allowing sync writes to avoid reordering.
1105 */
1106 wake = 1;
1107 }
1108 cap->implemented &= cap->issued | used;
1109 cap->mds_wanted = want;
1110
1111 if (flushing) {
1112 /*
1113 * assign a tid for flush operations so we can avoid
1114 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark
1115 * clean type races. track latest tid for every bit
1116 * so we can handle flush AxFw, flush Fw, and have the
1117 * first ack clean Ax.
1118 */
1119 flush_tid = ++ci->i_cap_flush_last_tid;
1120 if (pflush_tid)
1121 *pflush_tid = flush_tid;
1122 dout(" cap_flush_tid %d\n", (int)flush_tid);
1123 for (i = 0; i < CEPH_CAP_BITS; i++)
1124 if (flushing & (1 << i))
1125 ci->i_cap_flush_tid[i] = flush_tid;
1126 }
1127
1128 keep = cap->implemented;
1129 seq = cap->seq;
1130 issue_seq = cap->issue_seq;
1131 mseq = cap->mseq;
1132 size = inode->i_size;
1133 ci->i_reported_size = size;
1134 max_size = ci->i_wanted_max_size;
1135 ci->i_requested_max_size = max_size;
1136 mtime = inode->i_mtime;
1137 atime = inode->i_atime;
1138 time_warp_seq = ci->i_time_warp_seq;
1139 follows = ci->i_snap_realm->cached_context->seq;
1140 uid = inode->i_uid;
1141 gid = inode->i_gid;
1142 mode = inode->i_mode;
1143
1144 if (dropping & CEPH_CAP_XATTR_EXCL) {
1145 __ceph_build_xattrs_blob(ci);
1146 xattr_version = ci->i_xattrs.version + 1;
1147 }
1148
1149 spin_unlock(&inode->i_lock);
1150
1151 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1152 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq,
1153 size, max_size, &mtime, &atime, time_warp_seq,
1154 uid, gid, mode,
1155 xattr_version,
1156 (flushing & CEPH_CAP_XATTR_EXCL) ? ci->i_xattrs.blob : NULL,
1157 follows);
1158 if (ret < 0) {
1159 dout("error sending cap msg, must requeue %p\n", inode);
1160 delayed = 1;
1161 }
1162
1163 if (wake)
1164 wake_up(&ci->i_cap_wq);
1165
1166 return delayed;
1167}
1168
1169/*
1170 * When a snapshot is taken, clients accumulate dirty metadata on
1171 * inodes with capabilities in ceph_cap_snaps to describe the file
1172 * state at the time the snapshot was taken. This must be flushed
1173 * asynchronously back to the MDS once sync writes complete and dirty
1174 * data is written out.
1175 *
1176 * Called under i_lock. Takes s_mutex as needed.
1177 */
1178void __ceph_flush_snaps(struct ceph_inode_info *ci,
1179 struct ceph_mds_session **psession)
1180{
1181 struct inode *inode = &ci->vfs_inode;
1182 int mds;
1183 struct ceph_cap_snap *capsnap;
1184 u32 mseq;
1185 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
1186 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1187 session->s_mutex */
1188 u64 next_follows = 0; /* keep track of how far we've gotten through the
1189 i_cap_snaps list, and skip these entries next time
1190 around to avoid an infinite loop */
1191
1192 if (psession)
1193 session = *psession;
1194
1195 dout("__flush_snaps %p\n", inode);
1196retry:
1197 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1198 /* avoid an infiniute loop after retry */
1199 if (capsnap->follows < next_follows)
1200 continue;
1201 /*
1202 * we need to wait for sync writes to complete and for dirty
1203 * pages to be written out.
1204 */
1205 if (capsnap->dirty_pages || capsnap->writing)
1206 continue;
1207
1208 /*
1209 * if cap writeback already occurred, we should have dropped
1210 * the capsnap in ceph_put_wrbuffer_cap_refs.
1211 */
1212 BUG_ON(capsnap->dirty == 0);
1213
1214 /* pick mds, take s_mutex */
1215 mds = __ceph_get_cap_mds(ci, &mseq);
1216 if (session && session->s_mds != mds) {
1217 dout("oops, wrong session %p mutex\n", session);
1218 mutex_unlock(&session->s_mutex);
1219 ceph_put_mds_session(session);
1220 session = NULL;
1221 }
1222 if (!session) {
1223 spin_unlock(&inode->i_lock);
1224 mutex_lock(&mdsc->mutex);
1225 session = __ceph_lookup_mds_session(mdsc, mds);
1226 mutex_unlock(&mdsc->mutex);
1227 if (session) {
1228 dout("inverting session/ino locks on %p\n",
1229 session);
1230 mutex_lock(&session->s_mutex);
1231 }
1232 /*
1233 * if session == NULL, we raced against a cap
1234 * deletion. retry, and we'll get a better
1235 * @mds value next time.
1236 */
1237 spin_lock(&inode->i_lock);
1238 goto retry;
1239 }
1240
1241 capsnap->flush_tid = ++ci->i_cap_flush_last_tid;
1242 atomic_inc(&capsnap->nref);
1243 if (!list_empty(&capsnap->flushing_item))
1244 list_del_init(&capsnap->flushing_item);
1245 list_add_tail(&capsnap->flushing_item,
1246 &session->s_cap_snaps_flushing);
1247 spin_unlock(&inode->i_lock);
1248
1249 dout("flush_snaps %p cap_snap %p follows %lld size %llu\n",
1250 inode, capsnap, next_follows, capsnap->size);
1251 send_cap_msg(session, ceph_vino(inode).ino, 0,
1252 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1253 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq,
1254 capsnap->size, 0,
1255 &capsnap->mtime, &capsnap->atime,
1256 capsnap->time_warp_seq,
1257 capsnap->uid, capsnap->gid, capsnap->mode,
1258 0, NULL,
1259 capsnap->follows);
1260
1261 next_follows = capsnap->follows + 1;
1262 ceph_put_cap_snap(capsnap);
1263
1264 spin_lock(&inode->i_lock);
1265 goto retry;
1266 }
1267
1268 /* we flushed them all; remove this inode from the queue */
1269 spin_lock(&mdsc->snap_flush_lock);
1270 list_del_init(&ci->i_snap_flush_item);
1271 spin_unlock(&mdsc->snap_flush_lock);
1272
1273 if (psession)
1274 *psession = session;
1275 else if (session) {
1276 mutex_unlock(&session->s_mutex);
1277 ceph_put_mds_session(session);
1278 }
1279}
1280
1281static void ceph_flush_snaps(struct ceph_inode_info *ci)
1282{
1283 struct inode *inode = &ci->vfs_inode;
1284
1285 spin_lock(&inode->i_lock);
1286 __ceph_flush_snaps(ci, NULL);
1287 spin_unlock(&inode->i_lock);
1288}
1289
1290/*
1291 * Mark caps dirty. If inode is newly dirty, add to the global dirty
1292 * list.
1293 */
1294void __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1295{
1296 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc;
1297 struct inode *inode = &ci->vfs_inode;
1298 int was = ci->i_dirty_caps;
1299 int dirty = 0;
1300
1301 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1302 ceph_cap_string(mask), ceph_cap_string(was),
1303 ceph_cap_string(was | mask));
1304 ci->i_dirty_caps |= mask;
1305 if (was == 0) {
1306 dout(" inode %p now dirty\n", &ci->vfs_inode);
1307 BUG_ON(!list_empty(&ci->i_dirty_item));
1308 spin_lock(&mdsc->cap_dirty_lock);
1309 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1310 spin_unlock(&mdsc->cap_dirty_lock);
1311 if (ci->i_flushing_caps == 0) {
1312 igrab(inode);
1313 dirty |= I_DIRTY_SYNC;
1314 }
1315 }
1316 BUG_ON(list_empty(&ci->i_dirty_item));
1317 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1318 (mask & CEPH_CAP_FILE_BUFFER))
1319 dirty |= I_DIRTY_DATASYNC;
1320 if (dirty)
1321 __mark_inode_dirty(inode, dirty);
1322 __cap_delay_requeue(mdsc, ci);
1323}
1324
1325/*
1326 * Add dirty inode to the flushing list. Assigned a seq number so we
1327 * can wait for caps to flush without starving.
1328 *
1329 * Called under i_lock.
1330 */
1331static int __mark_caps_flushing(struct inode *inode,
1332 struct ceph_mds_session *session)
1333{
1334 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1335 struct ceph_inode_info *ci = ceph_inode(inode);
1336 int flushing;
1337
1338 BUG_ON(ci->i_dirty_caps == 0);
1339 BUG_ON(list_empty(&ci->i_dirty_item));
1340
1341 flushing = ci->i_dirty_caps;
1342 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1343 ceph_cap_string(flushing),
1344 ceph_cap_string(ci->i_flushing_caps),
1345 ceph_cap_string(ci->i_flushing_caps | flushing));
1346 ci->i_flushing_caps |= flushing;
1347 ci->i_dirty_caps = 0;
1348 dout(" inode %p now !dirty\n", inode);
1349
1350 spin_lock(&mdsc->cap_dirty_lock);
1351 list_del_init(&ci->i_dirty_item);
1352
1353 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq;
1354 if (list_empty(&ci->i_flushing_item)) {
1355 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1356 mdsc->num_cap_flushing++;
1357 dout(" inode %p now flushing seq %lld\n", inode,
1358 ci->i_cap_flush_seq);
1359 } else {
1360 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1361 dout(" inode %p now flushing (more) seq %lld\n", inode,
1362 ci->i_cap_flush_seq);
1363 }
1364 spin_unlock(&mdsc->cap_dirty_lock);
1365
1366 return flushing;
1367}
1368
1369/*
1370 * try to invalidate mapping pages without blocking.
1371 */
1372static int mapping_is_empty(struct address_space *mapping)
1373{
1374 struct page *page = find_get_page(mapping, 0);
1375
1376 if (!page)
1377 return 1;
1378
1379 put_page(page);
1380 return 0;
1381}
1382
1383static int try_nonblocking_invalidate(struct inode *inode)
1384{
1385 struct ceph_inode_info *ci = ceph_inode(inode);
1386 u32 invalidating_gen = ci->i_rdcache_gen;
1387
1388 spin_unlock(&inode->i_lock);
1389 invalidate_mapping_pages(&inode->i_data, 0, -1);
1390 spin_lock(&inode->i_lock);
1391
1392 if (mapping_is_empty(&inode->i_data) &&
1393 invalidating_gen == ci->i_rdcache_gen) {
1394 /* success. */
1395 dout("try_nonblocking_invalidate %p success\n", inode);
1396 ci->i_rdcache_gen = 0;
1397 ci->i_rdcache_revoking = 0;
1398 return 0;
1399 }
1400 dout("try_nonblocking_invalidate %p failed\n", inode);
1401 return -1;
1402}
1403
1404/*
1405 * Swiss army knife function to examine currently used and wanted
1406 * versus held caps. Release, flush, ack revoked caps to mds as
1407 * appropriate.
1408 *
1409 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1410 * cap release further.
1411 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1412 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1413 * further delay.
1414 */
1415void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1416 struct ceph_mds_session *session)
1417 __releases(session->s_mutex)
1418{
1419 struct ceph_client *client = ceph_inode_to_client(&ci->vfs_inode);
1420 struct ceph_mds_client *mdsc = &client->mdsc;
1421 struct inode *inode = &ci->vfs_inode;
1422 struct ceph_cap *cap;
1423 int file_wanted, used;
1424 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1425 int issued, implemented, want, retain, revoking, flushing = 0;
1426 int mds = -1; /* keep track of how far we've gone through i_caps list
1427 to avoid an infinite loop on retry */
1428 struct rb_node *p;
1429 int tried_invalidate = 0;
1430 int delayed = 0, sent = 0, force_requeue = 0, num;
1431 int queue_invalidate = 0;
1432 int is_delayed = flags & CHECK_CAPS_NODELAY;
1433
1434 /* if we are unmounting, flush any unused caps immediately. */
1435 if (mdsc->stopping)
1436 is_delayed = 1;
1437
1438 spin_lock(&inode->i_lock);
1439
1440 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1441 flags |= CHECK_CAPS_FLUSH;
1442
1443 /* flush snaps first time around only */
1444 if (!list_empty(&ci->i_cap_snaps))
1445 __ceph_flush_snaps(ci, &session);
1446 goto retry_locked;
1447retry:
1448 spin_lock(&inode->i_lock);
1449retry_locked:
1450 file_wanted = __ceph_caps_file_wanted(ci);
1451 used = __ceph_caps_used(ci);
1452 want = file_wanted | used;
1453 issued = __ceph_caps_issued(ci, &implemented);
1454 revoking = implemented & ~issued;
1455
1456 retain = want | CEPH_CAP_PIN;
1457 if (!mdsc->stopping && inode->i_nlink > 0) {
1458 if (want) {
1459 retain |= CEPH_CAP_ANY; /* be greedy */
1460 } else {
1461 retain |= CEPH_CAP_ANY_SHARED;
1462 /*
1463 * keep RD only if we didn't have the file open RW,
1464 * because then the mds would revoke it anyway to
1465 * journal max_size=0.
1466 */
1467 if (ci->i_max_size == 0)
1468 retain |= CEPH_CAP_ANY_RD;
1469 }
1470 }
1471
1472 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1473 " issued %s revoking %s retain %s %s%s%s\n", inode,
1474 ceph_cap_string(file_wanted),
1475 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1476 ceph_cap_string(ci->i_flushing_caps),
1477 ceph_cap_string(issued), ceph_cap_string(revoking),
1478 ceph_cap_string(retain),
1479 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1480 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1481 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1482
1483 /*
1484 * If we no longer need to hold onto old our caps, and we may
1485 * have cached pages, but don't want them, then try to invalidate.
1486 * If we fail, it's because pages are locked.... try again later.
1487 */
1488 if ((!is_delayed || mdsc->stopping) &&
1489 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */
1490 ci->i_rdcache_gen && /* may have cached pages */
1491 (file_wanted == 0 || /* no open files */
1492 (revoking & CEPH_CAP_FILE_CACHE)) && /* or revoking cache */
1493 !tried_invalidate) {
1494 dout("check_caps trying to invalidate on %p\n", inode);
1495 if (try_nonblocking_invalidate(inode) < 0) {
1496 if (revoking & CEPH_CAP_FILE_CACHE) {
1497 dout("check_caps queuing invalidate\n");
1498 queue_invalidate = 1;
1499 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1500 } else {
1501 dout("check_caps failed to invalidate pages\n");
1502 /* we failed to invalidate pages. check these
1503 caps again later. */
1504 force_requeue = 1;
1505 __cap_set_timeouts(mdsc, ci);
1506 }
1507 }
1508 tried_invalidate = 1;
1509 goto retry_locked;
1510 }
1511
1512 num = 0;
1513 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1514 cap = rb_entry(p, struct ceph_cap, ci_node);
1515 num++;
1516
1517 /* avoid looping forever */
1518 if (mds >= cap->mds ||
1519 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1520 continue;
1521
1522 /* NOTE: no side-effects allowed, until we take s_mutex */
1523
1524 revoking = cap->implemented & ~cap->issued;
1525 if (revoking)
1526 dout(" mds%d revoking %s\n", cap->mds,
1527 ceph_cap_string(revoking));
1528
1529 if (cap == ci->i_auth_cap &&
1530 (cap->issued & CEPH_CAP_FILE_WR)) {
1531 /* request larger max_size from MDS? */
1532 if (ci->i_wanted_max_size > ci->i_max_size &&
1533 ci->i_wanted_max_size > ci->i_requested_max_size) {
1534 dout("requesting new max_size\n");
1535 goto ack;
1536 }
1537
1538 /* approaching file_max? */
1539 if ((inode->i_size << 1) >= ci->i_max_size &&
1540 (ci->i_reported_size << 1) < ci->i_max_size) {
1541 dout("i_size approaching max_size\n");
1542 goto ack;
1543 }
1544 }
1545 /* flush anything dirty? */
1546 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1547 ci->i_dirty_caps) {
1548 dout("flushing dirty caps\n");
1549 goto ack;
1550 }
1551
1552 /* completed revocation? going down and there are no caps? */
1553 if (revoking && (revoking & used) == 0) {
1554 dout("completed revocation of %s\n",
1555 ceph_cap_string(cap->implemented & ~cap->issued));
1556 goto ack;
1557 }
1558
1559 /* want more caps from mds? */
1560 if (want & ~(cap->mds_wanted | cap->issued))
1561 goto ack;
1562
1563 /* things we might delay */
1564 if ((cap->issued & ~retain) == 0 &&
1565 cap->mds_wanted == want)
1566 continue; /* nope, all good */
1567
1568 if (is_delayed)
1569 goto ack;
1570
1571 /* delay? */
1572 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1573 time_before(jiffies, ci->i_hold_caps_max)) {
1574 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1575 ceph_cap_string(cap->issued),
1576 ceph_cap_string(cap->issued & retain),
1577 ceph_cap_string(cap->mds_wanted),
1578 ceph_cap_string(want));
1579 delayed++;
1580 continue;
1581 }
1582
1583ack:
1584 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1585 dout(" skipping %p I_NOFLUSH set\n", inode);
1586 continue;
1587 }
1588
1589 if (session && session != cap->session) {
1590 dout("oops, wrong session %p mutex\n", session);
1591 mutex_unlock(&session->s_mutex);
1592 session = NULL;
1593 }
1594 if (!session) {
1595 session = cap->session;
1596 if (mutex_trylock(&session->s_mutex) == 0) {
1597 dout("inverting session/ino locks on %p\n",
1598 session);
1599 spin_unlock(&inode->i_lock);
1600 if (took_snap_rwsem) {
1601 up_read(&mdsc->snap_rwsem);
1602 took_snap_rwsem = 0;
1603 }
1604 mutex_lock(&session->s_mutex);
1605 goto retry;
1606 }
1607 }
1608 /* take snap_rwsem after session mutex */
1609 if (!took_snap_rwsem) {
1610 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1611 dout("inverting snap/in locks on %p\n",
1612 inode);
1613 spin_unlock(&inode->i_lock);
1614 down_read(&mdsc->snap_rwsem);
1615 took_snap_rwsem = 1;
1616 goto retry;
1617 }
1618 took_snap_rwsem = 1;
1619 }
1620
1621 if (cap == ci->i_auth_cap && ci->i_dirty_caps)
1622 flushing = __mark_caps_flushing(inode, session);
1623
1624 mds = cap->mds; /* remember mds, so we don't repeat */
1625 sent++;
1626
1627 /* __send_cap drops i_lock */
1628 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, used, want,
1629 retain, flushing, NULL);
1630 goto retry; /* retake i_lock and restart our cap scan. */
1631 }
1632
1633 /*
1634 * Reschedule delayed caps release if we delayed anything,
1635 * otherwise cancel.
1636 */
1637 if (delayed && is_delayed)
1638 force_requeue = 1; /* __send_cap delayed release; requeue */
1639 if (!delayed && !is_delayed)
1640 __cap_delay_cancel(mdsc, ci);
1641 else if (!is_delayed || force_requeue)
1642 __cap_delay_requeue(mdsc, ci);
1643
1644 spin_unlock(&inode->i_lock);
1645
1646 if (queue_invalidate)
1647 ceph_queue_invalidate(inode);
1648
1649 if (session)
1650 mutex_unlock(&session->s_mutex);
1651 if (took_snap_rwsem)
1652 up_read(&mdsc->snap_rwsem);
1653}
1654
1655/*
1656 * Try to flush dirty caps back to the auth mds.
1657 */
1658static int try_flush_caps(struct inode *inode, struct ceph_mds_session *session,
1659 unsigned *flush_tid)
1660{
1661 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1662 struct ceph_inode_info *ci = ceph_inode(inode);
1663 int unlock_session = session ? 0 : 1;
1664 int flushing = 0;
1665
1666retry:
1667 spin_lock(&inode->i_lock);
1668 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1669 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1670 goto out;
1671 }
1672 if (ci->i_dirty_caps && ci->i_auth_cap) {
1673 struct ceph_cap *cap = ci->i_auth_cap;
1674 int used = __ceph_caps_used(ci);
1675 int want = __ceph_caps_wanted(ci);
1676 int delayed;
1677
1678 if (!session) {
1679 spin_unlock(&inode->i_lock);
1680 session = cap->session;
1681 mutex_lock(&session->s_mutex);
1682 goto retry;
1683 }
1684 BUG_ON(session != cap->session);
1685 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1686 goto out;
1687
1688 flushing = __mark_caps_flushing(inode, session);
1689
1690 /* __send_cap drops i_lock */
1691 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1692 cap->issued | cap->implemented, flushing,
1693 flush_tid);
1694 if (!delayed)
1695 goto out_unlocked;
1696
1697 spin_lock(&inode->i_lock);
1698 __cap_delay_requeue(mdsc, ci);
1699 }
1700out:
1701 spin_unlock(&inode->i_lock);
1702out_unlocked:
1703 if (session && unlock_session)
1704 mutex_unlock(&session->s_mutex);
1705 return flushing;
1706}
1707
1708/*
1709 * Return true if we've flushed caps through the given flush_tid.
1710 */
1711static int caps_are_flushed(struct inode *inode, unsigned tid)
1712{
1713 struct ceph_inode_info *ci = ceph_inode(inode);
1714 int dirty, i, ret = 1;
1715
1716 spin_lock(&inode->i_lock);
1717 dirty = __ceph_caps_dirty(ci);
1718 for (i = 0; i < CEPH_CAP_BITS; i++)
1719 if ((ci->i_flushing_caps & (1 << i)) &&
1720 ci->i_cap_flush_tid[i] <= tid) {
1721 /* still flushing this bit */
1722 ret = 0;
1723 break;
1724 }
1725 spin_unlock(&inode->i_lock);
1726 return ret;
1727}
1728
1729/*
1730 * Wait on any unsafe replies for the given inode. First wait on the
1731 * newest request, and make that the upper bound. Then, if there are
1732 * more requests, keep waiting on the oldest as long as it is still older
1733 * than the original request.
1734 */
1735static void sync_write_wait(struct inode *inode)
1736{
1737 struct ceph_inode_info *ci = ceph_inode(inode);
1738 struct list_head *head = &ci->i_unsafe_writes;
1739 struct ceph_osd_request *req;
1740 u64 last_tid;
1741
1742 spin_lock(&ci->i_unsafe_lock);
1743 if (list_empty(head))
1744 goto out;
1745
1746 /* set upper bound as _last_ entry in chain */
1747 req = list_entry(head->prev, struct ceph_osd_request,
1748 r_unsafe_item);
1749 last_tid = req->r_tid;
1750
1751 do {
1752 ceph_osdc_get_request(req);
1753 spin_unlock(&ci->i_unsafe_lock);
1754 dout("sync_write_wait on tid %llu (until %llu)\n",
1755 req->r_tid, last_tid);
1756 wait_for_completion(&req->r_safe_completion);
1757 spin_lock(&ci->i_unsafe_lock);
1758 ceph_osdc_put_request(req);
1759
1760 /*
1761 * from here on look at first entry in chain, since we
1762 * only want to wait for anything older than last_tid
1763 */
1764 if (list_empty(head))
1765 break;
1766 req = list_entry(head->next, struct ceph_osd_request,
1767 r_unsafe_item);
1768 } while (req->r_tid < last_tid);
1769out:
1770 spin_unlock(&ci->i_unsafe_lock);
1771}
1772
1773int ceph_fsync(struct file *file, struct dentry *dentry, int datasync)
1774{
1775 struct inode *inode = dentry->d_inode;
1776 struct ceph_inode_info *ci = ceph_inode(inode);
1777 unsigned flush_tid;
1778 int ret;
1779 int dirty;
1780
1781 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
1782 sync_write_wait(inode);
1783
1784 ret = filemap_write_and_wait(inode->i_mapping);
1785 if (ret < 0)
1786 return ret;
1787
1788 dirty = try_flush_caps(inode, NULL, &flush_tid);
1789 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1790
1791 /*
1792 * only wait on non-file metadata writeback (the mds
1793 * can recover size and mtime, so we don't need to
1794 * wait for that)
1795 */
1796 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1797 dout("fsync waiting for flush_tid %u\n", flush_tid);
1798 ret = wait_event_interruptible(ci->i_cap_wq,
1799 caps_are_flushed(inode, flush_tid));
1800 }
1801
1802 dout("fsync %p%s done\n", inode, datasync ? " datasync" : "");
1803 return ret;
1804}
1805
1806/*
1807 * Flush any dirty caps back to the mds. If we aren't asked to wait,
1808 * queue inode for flush but don't do so immediately, because we can
1809 * get by with fewer MDS messages if we wait for data writeback to
1810 * complete first.
1811 */
1812int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
1813{
1814 struct ceph_inode_info *ci = ceph_inode(inode);
1815 unsigned flush_tid;
1816 int err = 0;
1817 int dirty;
1818 int wait = wbc->sync_mode == WB_SYNC_ALL;
1819
1820 dout("write_inode %p wait=%d\n", inode, wait);
1821 if (wait) {
1822 dirty = try_flush_caps(inode, NULL, &flush_tid);
1823 if (dirty)
1824 err = wait_event_interruptible(ci->i_cap_wq,
1825 caps_are_flushed(inode, flush_tid));
1826 } else {
1827 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1828
1829 spin_lock(&inode->i_lock);
1830 if (__ceph_caps_dirty(ci))
1831 __cap_delay_requeue_front(mdsc, ci);
1832 spin_unlock(&inode->i_lock);
1833 }
1834 return err;
1835}
1836
1837/*
1838 * After a recovering MDS goes active, we need to resend any caps
1839 * we were flushing.
1840 *
1841 * Caller holds session->s_mutex.
1842 */
1843static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
1844 struct ceph_mds_session *session)
1845{
1846 struct ceph_cap_snap *capsnap;
1847
1848 dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
1849 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
1850 flushing_item) {
1851 struct ceph_inode_info *ci = capsnap->ci;
1852 struct inode *inode = &ci->vfs_inode;
1853 struct ceph_cap *cap;
1854
1855 spin_lock(&inode->i_lock);
1856 cap = ci->i_auth_cap;
1857 if (cap && cap->session == session) {
1858 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
1859 cap, capsnap);
1860 __ceph_flush_snaps(ci, &session);
1861 } else {
1862 pr_err("%p auth cap %p not mds%d ???\n", inode,
1863 cap, session->s_mds);
1864 spin_unlock(&inode->i_lock);
1865 }
1866 }
1867}
1868
1869void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
1870 struct ceph_mds_session *session)
1871{
1872 struct ceph_inode_info *ci;
1873
1874 kick_flushing_capsnaps(mdsc, session);
1875
1876 dout("kick_flushing_caps mds%d\n", session->s_mds);
1877 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
1878 struct inode *inode = &ci->vfs_inode;
1879 struct ceph_cap *cap;
1880 int delayed = 0;
1881
1882 spin_lock(&inode->i_lock);
1883 cap = ci->i_auth_cap;
1884 if (cap && cap->session == session) {
1885 dout("kick_flushing_caps %p cap %p %s\n", inode,
1886 cap, ceph_cap_string(ci->i_flushing_caps));
1887 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1888 __ceph_caps_used(ci),
1889 __ceph_caps_wanted(ci),
1890 cap->issued | cap->implemented,
1891 ci->i_flushing_caps, NULL);
1892 if (delayed) {
1893 spin_lock(&inode->i_lock);
1894 __cap_delay_requeue(mdsc, ci);
1895 spin_unlock(&inode->i_lock);
1896 }
1897 } else {
1898 pr_err("%p auth cap %p not mds%d ???\n", inode,
1899 cap, session->s_mds);
1900 spin_unlock(&inode->i_lock);
1901 }
1902 }
1903}
1904
1905
1906/*
1907 * Take references to capabilities we hold, so that we don't release
1908 * them to the MDS prematurely.
1909 *
1910 * Protected by i_lock.
1911 */
1912static void __take_cap_refs(struct ceph_inode_info *ci, int got)
1913{
1914 if (got & CEPH_CAP_PIN)
1915 ci->i_pin_ref++;
1916 if (got & CEPH_CAP_FILE_RD)
1917 ci->i_rd_ref++;
1918 if (got & CEPH_CAP_FILE_CACHE)
1919 ci->i_rdcache_ref++;
1920 if (got & CEPH_CAP_FILE_WR)
1921 ci->i_wr_ref++;
1922 if (got & CEPH_CAP_FILE_BUFFER) {
1923 if (ci->i_wrbuffer_ref == 0)
1924 igrab(&ci->vfs_inode);
1925 ci->i_wrbuffer_ref++;
1926 dout("__take_cap_refs %p wrbuffer %d -> %d (?)\n",
1927 &ci->vfs_inode, ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref);
1928 }
1929}
1930
1931/*
1932 * Try to grab cap references. Specify those refs we @want, and the
1933 * minimal set we @need. Also include the larger offset we are writing
1934 * to (when applicable), and check against max_size here as well.
1935 * Note that caller is responsible for ensuring max_size increases are
1936 * requested from the MDS.
1937 */
1938static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
1939 int *got, loff_t endoff, int *check_max, int *err)
1940{
1941 struct inode *inode = &ci->vfs_inode;
1942 int ret = 0;
1943 int have, implemented;
1944 int file_wanted;
1945
1946 dout("get_cap_refs %p need %s want %s\n", inode,
1947 ceph_cap_string(need), ceph_cap_string(want));
1948 spin_lock(&inode->i_lock);
1949
1950 /* make sure file is actually open */
1951 file_wanted = __ceph_caps_file_wanted(ci);
1952 if ((file_wanted & need) == 0) {
1953 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
1954 ceph_cap_string(need), ceph_cap_string(file_wanted));
1955 *err = -EBADF;
1956 ret = 1;
1957 goto out;
1958 }
1959
1960 if (need & CEPH_CAP_FILE_WR) {
1961 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
1962 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
1963 inode, endoff, ci->i_max_size);
1964 if (endoff > ci->i_wanted_max_size) {
1965 *check_max = 1;
1966 ret = 1;
1967 }
1968 goto out;
1969 }
1970 /*
1971 * If a sync write is in progress, we must wait, so that we
1972 * can get a final snapshot value for size+mtime.
1973 */
1974 if (__ceph_have_pending_cap_snap(ci)) {
1975 dout("get_cap_refs %p cap_snap_pending\n", inode);
1976 goto out;
1977 }
1978 }
1979 have = __ceph_caps_issued(ci, &implemented);
1980
1981 /*
1982 * disallow writes while a truncate is pending
1983 */
1984 if (ci->i_truncate_pending)
1985 have &= ~CEPH_CAP_FILE_WR;
1986
1987 if ((have & need) == need) {
1988 /*
1989 * Look at (implemented & ~have & not) so that we keep waiting
1990 * on transition from wanted -> needed caps. This is needed
1991 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
1992 * going before a prior buffered writeback happens.
1993 */
1994 int not = want & ~(have & need);
1995 int revoking = implemented & ~have;
1996 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
1997 inode, ceph_cap_string(have), ceph_cap_string(not),
1998 ceph_cap_string(revoking));
1999 if ((revoking & not) == 0) {
2000 *got = need | (have & want);
2001 __take_cap_refs(ci, *got);
2002 ret = 1;
2003 }
2004 } else {
2005 dout("get_cap_refs %p have %s needed %s\n", inode,
2006 ceph_cap_string(have), ceph_cap_string(need));
2007 }
2008out:
2009 spin_unlock(&inode->i_lock);
2010 dout("get_cap_refs %p ret %d got %s\n", inode,
2011 ret, ceph_cap_string(*got));
2012 return ret;
2013}
2014
2015/*
2016 * Check the offset we are writing up to against our current
2017 * max_size. If necessary, tell the MDS we want to write to
2018 * a larger offset.
2019 */
2020static void check_max_size(struct inode *inode, loff_t endoff)
2021{
2022 struct ceph_inode_info *ci = ceph_inode(inode);
2023 int check = 0;
2024
2025 /* do we need to explicitly request a larger max_size? */
2026 spin_lock(&inode->i_lock);
2027 if ((endoff >= ci->i_max_size ||
2028 endoff > (inode->i_size << 1)) &&
2029 endoff > ci->i_wanted_max_size) {
2030 dout("write %p at large endoff %llu, req max_size\n",
2031 inode, endoff);
2032 ci->i_wanted_max_size = endoff;
2033 check = 1;
2034 }
2035 spin_unlock(&inode->i_lock);
2036 if (check)
2037 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2038}
2039
2040/*
2041 * Wait for caps, and take cap references. If we can't get a WR cap
2042 * due to a small max_size, make sure we check_max_size (and possibly
2043 * ask the mds) so we don't get hung up indefinitely.
2044 */
2045int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, int *got,
2046 loff_t endoff)
2047{
2048 int check_max, ret, err;
2049
2050retry:
2051 if (endoff > 0)
2052 check_max_size(&ci->vfs_inode, endoff);
2053 check_max = 0;
2054 err = 0;
2055 ret = wait_event_interruptible(ci->i_cap_wq,
2056 try_get_cap_refs(ci, need, want,
2057 got, endoff,
2058 &check_max, &err));
2059 if (err)
2060 ret = err;
2061 if (check_max)
2062 goto retry;
2063 return ret;
2064}
2065
2066/*
2067 * Take cap refs. Caller must already know we hold at least one ref
2068 * on the caps in question or we don't know this is safe.
2069 */
2070void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2071{
2072 spin_lock(&ci->vfs_inode.i_lock);
2073 __take_cap_refs(ci, caps);
2074 spin_unlock(&ci->vfs_inode.i_lock);
2075}
2076
2077/*
2078 * Release cap refs.
2079 *
2080 * If we released the last ref on any given cap, call ceph_check_caps
2081 * to release (or schedule a release).
2082 *
2083 * If we are releasing a WR cap (from a sync write), finalize any affected
2084 * cap_snap, and wake up any waiters.
2085 */
2086void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2087{
2088 struct inode *inode = &ci->vfs_inode;
2089 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2090 struct ceph_cap_snap *capsnap;
2091
2092 spin_lock(&inode->i_lock);
2093 if (had & CEPH_CAP_PIN)
2094 --ci->i_pin_ref;
2095 if (had & CEPH_CAP_FILE_RD)
2096 if (--ci->i_rd_ref == 0)
2097 last++;
2098 if (had & CEPH_CAP_FILE_CACHE)
2099 if (--ci->i_rdcache_ref == 0)
2100 last++;
2101 if (had & CEPH_CAP_FILE_BUFFER) {
2102 if (--ci->i_wrbuffer_ref == 0) {
2103 last++;
2104 put++;
2105 }
2106 dout("put_cap_refs %p wrbuffer %d -> %d (?)\n",
2107 inode, ci->i_wrbuffer_ref+1, ci->i_wrbuffer_ref);
2108 }
2109 if (had & CEPH_CAP_FILE_WR)
2110 if (--ci->i_wr_ref == 0) {
2111 last++;
2112 if (!list_empty(&ci->i_cap_snaps)) {
2113 capsnap = list_first_entry(&ci->i_cap_snaps,
2114 struct ceph_cap_snap,
2115 ci_item);
2116 if (capsnap->writing) {
2117 capsnap->writing = 0;
2118 flushsnaps =
2119 __ceph_finish_cap_snap(ci,
2120 capsnap);
2121 wake = 1;
2122 }
2123 }
2124 }
2125 spin_unlock(&inode->i_lock);
2126
2127 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2128 last ? " last" : "", put ? " put" : "");
2129
2130 if (last && !flushsnaps)
2131 ceph_check_caps(ci, 0, NULL);
2132 else if (flushsnaps)
2133 ceph_flush_snaps(ci);
2134 if (wake)
2135 wake_up(&ci->i_cap_wq);
2136 if (put)
2137 iput(inode);
2138}
2139
2140/*
2141 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2142 * context. Adjust per-snap dirty page accounting as appropriate.
2143 * Once all dirty data for a cap_snap is flushed, flush snapped file
2144 * metadata back to the MDS. If we dropped the last ref, call
2145 * ceph_check_caps.
2146 */
2147void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2148 struct ceph_snap_context *snapc)
2149{
2150 struct inode *inode = &ci->vfs_inode;
2151 int last = 0;
2152 int complete_capsnap = 0;
2153 int drop_capsnap = 0;
2154 int found = 0;
2155 struct ceph_cap_snap *capsnap = NULL;
2156
2157 spin_lock(&inode->i_lock);
2158 ci->i_wrbuffer_ref -= nr;
2159 last = !ci->i_wrbuffer_ref;
2160
2161 if (ci->i_head_snapc == snapc) {
2162 ci->i_wrbuffer_ref_head -= nr;
2163 if (!ci->i_wrbuffer_ref_head) {
2164 ceph_put_snap_context(ci->i_head_snapc);
2165 ci->i_head_snapc = NULL;
2166 }
2167 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2168 inode,
2169 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2170 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2171 last ? " LAST" : "");
2172 } else {
2173 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2174 if (capsnap->context == snapc) {
2175 found = 1;
2176 break;
2177 }
2178 }
2179 BUG_ON(!found);
2180 capsnap->dirty_pages -= nr;
2181 if (capsnap->dirty_pages == 0) {
2182 complete_capsnap = 1;
2183 if (capsnap->dirty == 0)
2184 /* cap writeback completed before we created
2185 * the cap_snap; no FLUSHSNAP is needed */
2186 drop_capsnap = 1;
2187 }
2188 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2189 " snap %lld %d/%d -> %d/%d %s%s%s\n",
2190 inode, capsnap, capsnap->context->seq,
2191 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2192 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2193 last ? " (wrbuffer last)" : "",
2194 complete_capsnap ? " (complete capsnap)" : "",
2195 drop_capsnap ? " (drop capsnap)" : "");
2196 if (drop_capsnap) {
2197 ceph_put_snap_context(capsnap->context);
2198 list_del(&capsnap->ci_item);
2199 list_del(&capsnap->flushing_item);
2200 ceph_put_cap_snap(capsnap);
2201 }
2202 }
2203
2204 spin_unlock(&inode->i_lock);
2205
2206 if (last) {
2207 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2208 iput(inode);
2209 } else if (complete_capsnap) {
2210 ceph_flush_snaps(ci);
2211 wake_up(&ci->i_cap_wq);
2212 }
2213 if (drop_capsnap)
2214 iput(inode);
2215}
2216
2217/*
2218 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2219 * actually be a revocation if it specifies a smaller cap set.)
2220 *
2221 * caller holds s_mutex and i_lock, we drop both.
2222 *
2223 * return value:
2224 * 0 - ok
2225 * 1 - check_caps on auth cap only (writeback)
2226 * 2 - check_caps (ack revoke)
2227 */
2228static void handle_cap_grant(struct inode *inode, struct ceph_mds_caps *grant,
2229 struct ceph_mds_session *session,
2230 struct ceph_cap *cap,
2231 struct ceph_buffer *xattr_buf)
2232 __releases(inode->i_lock)
2233 __releases(session->s_mutex)
2234{
2235 struct ceph_inode_info *ci = ceph_inode(inode);
2236 int mds = session->s_mds;
2237 int seq = le32_to_cpu(grant->seq);
2238 int newcaps = le32_to_cpu(grant->caps);
2239 int issued, implemented, used, wanted, dirty;
2240 u64 size = le64_to_cpu(grant->size);
2241 u64 max_size = le64_to_cpu(grant->max_size);
2242 struct timespec mtime, atime, ctime;
2243 int check_caps = 0;
2244 int wake = 0;
2245 int writeback = 0;
2246 int revoked_rdcache = 0;
2247 int queue_invalidate = 0;
2248
2249 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2250 inode, cap, mds, seq, ceph_cap_string(newcaps));
2251 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2252 inode->i_size);
2253
2254 /*
2255 * If CACHE is being revoked, and we have no dirty buffers,
2256 * try to invalidate (once). (If there are dirty buffers, we
2257 * will invalidate _after_ writeback.)
2258 */
2259 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2260 !ci->i_wrbuffer_ref) {
2261 if (try_nonblocking_invalidate(inode) == 0) {
2262 revoked_rdcache = 1;
2263 } else {
2264 /* there were locked pages.. invalidate later
2265 in a separate thread. */
2266 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2267 queue_invalidate = 1;
2268 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2269 }
2270 }
2271 }
2272
2273 /* side effects now are allowed */
2274
2275 issued = __ceph_caps_issued(ci, &implemented);
2276 issued |= implemented | __ceph_caps_dirty(ci);
2277
2278 cap->cap_gen = session->s_cap_gen;
2279
2280 __check_cap_issue(ci, cap, newcaps);
2281
2282 if ((issued & CEPH_CAP_AUTH_EXCL) == 0) {
2283 inode->i_mode = le32_to_cpu(grant->mode);
2284 inode->i_uid = le32_to_cpu(grant->uid);
2285 inode->i_gid = le32_to_cpu(grant->gid);
2286 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2287 inode->i_uid, inode->i_gid);
2288 }
2289
2290 if ((issued & CEPH_CAP_LINK_EXCL) == 0)
2291 inode->i_nlink = le32_to_cpu(grant->nlink);
2292
2293 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2294 int len = le32_to_cpu(grant->xattr_len);
2295 u64 version = le64_to_cpu(grant->xattr_version);
2296
2297 if (version > ci->i_xattrs.version) {
2298 dout(" got new xattrs v%llu on %p len %d\n",
2299 version, inode, len);
2300 if (ci->i_xattrs.blob)
2301 ceph_buffer_put(ci->i_xattrs.blob);
2302 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2303 ci->i_xattrs.version = version;
2304 }
2305 }
2306
2307 /* size/ctime/mtime/atime? */
2308 ceph_fill_file_size(inode, issued,
2309 le32_to_cpu(grant->truncate_seq),
2310 le64_to_cpu(grant->truncate_size), size);
2311 ceph_decode_timespec(&mtime, &grant->mtime);
2312 ceph_decode_timespec(&atime, &grant->atime);
2313 ceph_decode_timespec(&ctime, &grant->ctime);
2314 ceph_fill_file_time(inode, issued,
2315 le32_to_cpu(grant->time_warp_seq), &ctime, &mtime,
2316 &atime);
2317
2318 /* max size increase? */
2319 if (max_size != ci->i_max_size) {
2320 dout("max_size %lld -> %llu\n", ci->i_max_size, max_size);
2321 ci->i_max_size = max_size;
2322 if (max_size >= ci->i_wanted_max_size) {
2323 ci->i_wanted_max_size = 0; /* reset */
2324 ci->i_requested_max_size = 0;
2325 }
2326 wake = 1;
2327 }
2328
2329 /* check cap bits */
2330 wanted = __ceph_caps_wanted(ci);
2331 used = __ceph_caps_used(ci);
2332 dirty = __ceph_caps_dirty(ci);
2333 dout(" my wanted = %s, used = %s, dirty %s\n",
2334 ceph_cap_string(wanted),
2335 ceph_cap_string(used),
2336 ceph_cap_string(dirty));
2337 if (wanted != le32_to_cpu(grant->wanted)) {
2338 dout("mds wanted %s -> %s\n",
2339 ceph_cap_string(le32_to_cpu(grant->wanted)),
2340 ceph_cap_string(wanted));
2341 grant->wanted = cpu_to_le32(wanted);
2342 }
2343
2344 cap->seq = seq;
2345
2346 /* file layout may have changed */
2347 ci->i_layout = grant->layout;
2348
2349 /* revocation, grant, or no-op? */
2350 if (cap->issued & ~newcaps) {
2351 dout("revocation: %s -> %s\n", ceph_cap_string(cap->issued),
2352 ceph_cap_string(newcaps));
2353 if ((used & ~newcaps) & CEPH_CAP_FILE_BUFFER)
2354 writeback = 1; /* will delay ack */
2355 else if (dirty & ~newcaps)
2356 check_caps = 1; /* initiate writeback in check_caps */
2357 else if (((used & ~newcaps) & CEPH_CAP_FILE_CACHE) == 0 ||
2358 revoked_rdcache)
2359 check_caps = 2; /* send revoke ack in check_caps */
2360 cap->issued = newcaps;
2361 cap->implemented |= newcaps;
2362 } else if (cap->issued == newcaps) {
2363 dout("caps unchanged: %s -> %s\n",
2364 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2365 } else {
2366 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2367 ceph_cap_string(newcaps));
2368 cap->issued = newcaps;
2369 cap->implemented |= newcaps; /* add bits only, to
2370 * avoid stepping on a
2371 * pending revocation */
2372 wake = 1;
2373 }
2374 BUG_ON(cap->issued & ~cap->implemented);
2375
2376 spin_unlock(&inode->i_lock);
2377 if (writeback)
2378 /*
2379 * queue inode for writeback: we can't actually call
2380 * filemap_write_and_wait, etc. from message handler
2381 * context.
2382 */
2383 ceph_queue_writeback(inode);
2384 if (queue_invalidate)
2385 ceph_queue_invalidate(inode);
2386 if (wake)
2387 wake_up(&ci->i_cap_wq);
2388
2389 if (check_caps == 1)
2390 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2391 session);
2392 else if (check_caps == 2)
2393 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
2394 else
2395 mutex_unlock(&session->s_mutex);
2396}
2397
2398/*
2399 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2400 * MDS has been safely committed.
2401 */
2402static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2403 struct ceph_mds_caps *m,
2404 struct ceph_mds_session *session,
2405 struct ceph_cap *cap)
2406 __releases(inode->i_lock)
2407{
2408 struct ceph_inode_info *ci = ceph_inode(inode);
2409 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
2410 unsigned seq = le32_to_cpu(m->seq);
2411 int dirty = le32_to_cpu(m->dirty);
2412 int cleaned = 0;
2413 int drop = 0;
2414 int i;
2415
2416 for (i = 0; i < CEPH_CAP_BITS; i++)
2417 if ((dirty & (1 << i)) &&
2418 flush_tid == ci->i_cap_flush_tid[i])
2419 cleaned |= 1 << i;
2420
2421 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2422 " flushing %s -> %s\n",
2423 inode, session->s_mds, seq, ceph_cap_string(dirty),
2424 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2425 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2426
2427 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned))
2428 goto out;
2429
2430 ci->i_flushing_caps &= ~cleaned;
2431
2432 spin_lock(&mdsc->cap_dirty_lock);
2433 if (ci->i_flushing_caps == 0) {
2434 list_del_init(&ci->i_flushing_item);
2435 if (!list_empty(&session->s_cap_flushing))
2436 dout(" mds%d still flushing cap on %p\n",
2437 session->s_mds,
2438 &list_entry(session->s_cap_flushing.next,
2439 struct ceph_inode_info,
2440 i_flushing_item)->vfs_inode);
2441 mdsc->num_cap_flushing--;
2442 wake_up(&mdsc->cap_flushing_wq);
2443 dout(" inode %p now !flushing\n", inode);
2444
2445 if (ci->i_dirty_caps == 0) {
2446 dout(" inode %p now clean\n", inode);
2447 BUG_ON(!list_empty(&ci->i_dirty_item));
2448 drop = 1;
2449 } else {
2450 BUG_ON(list_empty(&ci->i_dirty_item));
2451 }
2452 }
2453 spin_unlock(&mdsc->cap_dirty_lock);
2454 wake_up(&ci->i_cap_wq);
2455
2456out:
2457 spin_unlock(&inode->i_lock);
2458 if (drop)
2459 iput(inode);
2460}
2461
2462/*
2463 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
2464 * throw away our cap_snap.
2465 *
2466 * Caller hold s_mutex.
2467 */
2468static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
2469 struct ceph_mds_caps *m,
2470 struct ceph_mds_session *session)
2471{
2472 struct ceph_inode_info *ci = ceph_inode(inode);
2473 u64 follows = le64_to_cpu(m->snap_follows);
2474 struct ceph_cap_snap *capsnap;
2475 int drop = 0;
2476
2477 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
2478 inode, ci, session->s_mds, follows);
2479
2480 spin_lock(&inode->i_lock);
2481 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2482 if (capsnap->follows == follows) {
2483 if (capsnap->flush_tid != flush_tid) {
2484 dout(" cap_snap %p follows %lld tid %lld !="
2485 " %lld\n", capsnap, follows,
2486 flush_tid, capsnap->flush_tid);
2487 break;
2488 }
2489 WARN_ON(capsnap->dirty_pages || capsnap->writing);
2490 dout(" removing %p cap_snap %p follows %lld\n",
2491 inode, capsnap, follows);
2492 ceph_put_snap_context(capsnap->context);
2493 list_del(&capsnap->ci_item);
2494 list_del(&capsnap->flushing_item);
2495 ceph_put_cap_snap(capsnap);
2496 drop = 1;
2497 break;
2498 } else {
2499 dout(" skipping cap_snap %p follows %lld\n",
2500 capsnap, capsnap->follows);
2501 }
2502 }
2503 spin_unlock(&inode->i_lock);
2504 if (drop)
2505 iput(inode);
2506}
2507
2508/*
2509 * Handle TRUNC from MDS, indicating file truncation.
2510 *
2511 * caller hold s_mutex.
2512 */
2513static void handle_cap_trunc(struct inode *inode,
2514 struct ceph_mds_caps *trunc,
2515 struct ceph_mds_session *session)
2516 __releases(inode->i_lock)
2517{
2518 struct ceph_inode_info *ci = ceph_inode(inode);
2519 int mds = session->s_mds;
2520 int seq = le32_to_cpu(trunc->seq);
2521 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
2522 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
2523 u64 size = le64_to_cpu(trunc->size);
2524 int implemented = 0;
2525 int dirty = __ceph_caps_dirty(ci);
2526 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
2527 int queue_trunc = 0;
2528
2529 issued |= implemented | dirty;
2530
2531 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
2532 inode, mds, seq, truncate_size, truncate_seq);
2533 queue_trunc = ceph_fill_file_size(inode, issued,
2534 truncate_seq, truncate_size, size);
2535 spin_unlock(&inode->i_lock);
2536
2537 if (queue_trunc)
2538 ceph_queue_vmtruncate(inode);
2539}
2540
2541/*
2542 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
2543 * different one. If we are the most recent migration we've seen (as
2544 * indicated by mseq), make note of the migrating cap bits for the
2545 * duration (until we see the corresponding IMPORT).
2546 *
2547 * caller holds s_mutex
2548 */
2549static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
2550 struct ceph_mds_session *session)
2551{
2552 struct ceph_inode_info *ci = ceph_inode(inode);
2553 int mds = session->s_mds;
2554 unsigned mseq = le32_to_cpu(ex->migrate_seq);
2555 struct ceph_cap *cap = NULL, *t;
2556 struct rb_node *p;
2557 int remember = 1;
2558
2559 dout("handle_cap_export inode %p ci %p mds%d mseq %d\n",
2560 inode, ci, mds, mseq);
2561
2562 spin_lock(&inode->i_lock);
2563
2564 /* make sure we haven't seen a higher mseq */
2565 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2566 t = rb_entry(p, struct ceph_cap, ci_node);
2567 if (ceph_seq_cmp(t->mseq, mseq) > 0) {
2568 dout(" higher mseq on cap from mds%d\n",
2569 t->session->s_mds);
2570 remember = 0;
2571 }
2572 if (t->session->s_mds == mds)
2573 cap = t;
2574 }
2575
2576 if (cap) {
2577 if (remember) {
2578 /* make note */
2579 ci->i_cap_exporting_mds = mds;
2580 ci->i_cap_exporting_mseq = mseq;
2581 ci->i_cap_exporting_issued = cap->issued;
2582 }
2583 __ceph_remove_cap(cap);
2584 }
2585 /* else, we already released it */
2586
2587 spin_unlock(&inode->i_lock);
2588}
2589
2590/*
2591 * Handle cap IMPORT. If there are temp bits from an older EXPORT,
2592 * clean them up.
2593 *
2594 * caller holds s_mutex.
2595 */
2596static void handle_cap_import(struct ceph_mds_client *mdsc,
2597 struct inode *inode, struct ceph_mds_caps *im,
2598 struct ceph_mds_session *session,
2599 void *snaptrace, int snaptrace_len)
2600{
2601 struct ceph_inode_info *ci = ceph_inode(inode);
2602 int mds = session->s_mds;
2603 unsigned issued = le32_to_cpu(im->caps);
2604 unsigned wanted = le32_to_cpu(im->wanted);
2605 unsigned seq = le32_to_cpu(im->seq);
2606 unsigned mseq = le32_to_cpu(im->migrate_seq);
2607 u64 realmino = le64_to_cpu(im->realm);
2608 u64 cap_id = le64_to_cpu(im->cap_id);
2609
2610 if (ci->i_cap_exporting_mds >= 0 &&
2611 ceph_seq_cmp(ci->i_cap_exporting_mseq, mseq) < 0) {
2612 dout("handle_cap_import inode %p ci %p mds%d mseq %d"
2613 " - cleared exporting from mds%d\n",
2614 inode, ci, mds, mseq,
2615 ci->i_cap_exporting_mds);
2616 ci->i_cap_exporting_issued = 0;
2617 ci->i_cap_exporting_mseq = 0;
2618 ci->i_cap_exporting_mds = -1;
2619 } else {
2620 dout("handle_cap_import inode %p ci %p mds%d mseq %d\n",
2621 inode, ci, mds, mseq);
2622 }
2623
2624 down_write(&mdsc->snap_rwsem);
2625 ceph_update_snap_trace(mdsc, snaptrace, snaptrace+snaptrace_len,
2626 false);
2627 downgrade_write(&mdsc->snap_rwsem);
2628 ceph_add_cap(inode, session, cap_id, -1,
2629 issued, wanted, seq, mseq, realmino, CEPH_CAP_FLAG_AUTH,
2630 NULL /* no caps context */);
2631 try_flush_caps(inode, session, NULL);
2632 up_read(&mdsc->snap_rwsem);
2633}
2634
2635/*
2636 * Handle a caps message from the MDS.
2637 *
2638 * Identify the appropriate session, inode, and call the right handler
2639 * based on the cap op.
2640 */
2641void ceph_handle_caps(struct ceph_mds_session *session,
2642 struct ceph_msg *msg)
2643{
2644 struct ceph_mds_client *mdsc = session->s_mdsc;
2645 struct super_block *sb = mdsc->client->sb;
2646 struct inode *inode;
2647 struct ceph_cap *cap;
2648 struct ceph_mds_caps *h;
2649 int mds = session->s_mds;
2650 int op;
2651 u32 seq;
2652 struct ceph_vino vino;
2653 u64 cap_id;
2654 u64 size, max_size;
2655 u64 tid;
2656 void *snaptrace;
2657
2658 dout("handle_caps from mds%d\n", mds);
2659
2660 /* decode */
2661 tid = le64_to_cpu(msg->hdr.tid);
2662 if (msg->front.iov_len < sizeof(*h))
2663 goto bad;
2664 h = msg->front.iov_base;
2665 snaptrace = h + 1;
2666 op = le32_to_cpu(h->op);
2667 vino.ino = le64_to_cpu(h->ino);
2668 vino.snap = CEPH_NOSNAP;
2669 cap_id = le64_to_cpu(h->cap_id);
2670 seq = le32_to_cpu(h->seq);
2671 size = le64_to_cpu(h->size);
2672 max_size = le64_to_cpu(h->max_size);
2673
2674 mutex_lock(&session->s_mutex);
2675 session->s_seq++;
2676 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
2677 (unsigned)seq);
2678
2679 /* lookup ino */
2680 inode = ceph_find_inode(sb, vino);
2681 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
2682 vino.snap, inode);
2683 if (!inode) {
2684 dout(" i don't have ino %llx\n", vino.ino);
2685 goto done;
2686 }
2687
2688 /* these will work even if we don't have a cap yet */
2689 switch (op) {
2690 case CEPH_CAP_OP_FLUSHSNAP_ACK:
2691 handle_cap_flushsnap_ack(inode, tid, h, session);
2692 goto done;
2693
2694 case CEPH_CAP_OP_EXPORT:
2695 handle_cap_export(inode, h, session);
2696 goto done;
2697
2698 case CEPH_CAP_OP_IMPORT:
2699 handle_cap_import(mdsc, inode, h, session,
2700 snaptrace, le32_to_cpu(h->snap_trace_len));
2701 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_NODELAY,
2702 session);
2703 goto done_unlocked;
2704 }
2705
2706 /* the rest require a cap */
2707 spin_lock(&inode->i_lock);
2708 cap = __get_cap_for_mds(ceph_inode(inode), mds);
2709 if (!cap) {
2710 dout("no cap on %p ino %llx.%llx from mds%d, releasing\n",
2711 inode, ceph_ino(inode), ceph_snap(inode), mds);
2712 spin_unlock(&inode->i_lock);
2713 goto done;
2714 }
2715
2716 /* note that each of these drops i_lock for us */
2717 switch (op) {
2718 case CEPH_CAP_OP_REVOKE:
2719 case CEPH_CAP_OP_GRANT:
2720 handle_cap_grant(inode, h, session, cap, msg->middle);
2721 goto done_unlocked;
2722
2723 case CEPH_CAP_OP_FLUSH_ACK:
2724 handle_cap_flush_ack(inode, tid, h, session, cap);
2725 break;
2726
2727 case CEPH_CAP_OP_TRUNC:
2728 handle_cap_trunc(inode, h, session);
2729 break;
2730
2731 default:
2732 spin_unlock(&inode->i_lock);
2733 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
2734 ceph_cap_op_name(op));
2735 }
2736
2737done:
2738 mutex_unlock(&session->s_mutex);
2739done_unlocked:
2740 if (inode)
2741 iput(inode);
2742 return;
2743
2744bad:
2745 pr_err("ceph_handle_caps: corrupt message\n");
2746 ceph_msg_dump(msg);
2747 return;
2748}
2749
2750/*
2751 * Delayed work handler to process end of delayed cap release LRU list.
2752 */
2753void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
2754{
2755 struct ceph_inode_info *ci;
2756 int flags = CHECK_CAPS_NODELAY;
2757
2758 dout("check_delayed_caps\n");
2759 while (1) {
2760 spin_lock(&mdsc->cap_delay_lock);
2761 if (list_empty(&mdsc->cap_delay_list))
2762 break;
2763 ci = list_first_entry(&mdsc->cap_delay_list,
2764 struct ceph_inode_info,
2765 i_cap_delay_list);
2766 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
2767 time_before(jiffies, ci->i_hold_caps_max))
2768 break;
2769 list_del_init(&ci->i_cap_delay_list);
2770 spin_unlock(&mdsc->cap_delay_lock);
2771 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
2772 ceph_check_caps(ci, flags, NULL);
2773 }
2774 spin_unlock(&mdsc->cap_delay_lock);
2775}
2776
2777/*
2778 * Flush all dirty caps to the mds
2779 */
2780void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
2781{
2782 struct ceph_inode_info *ci, *nci = NULL;
2783 struct inode *inode, *ninode = NULL;
2784 struct list_head *p, *n;
2785
2786 dout("flush_dirty_caps\n");
2787 spin_lock(&mdsc->cap_dirty_lock);
2788 list_for_each_safe(p, n, &mdsc->cap_dirty) {
2789 if (nci) {
2790 ci = nci;
2791 inode = ninode;
2792 ci->i_ceph_flags &= ~CEPH_I_NOFLUSH;
2793 dout("flush_dirty_caps inode %p (was next inode)\n",
2794 inode);
2795 } else {
2796 ci = list_entry(p, struct ceph_inode_info,
2797 i_dirty_item);
2798 inode = igrab(&ci->vfs_inode);
2799 BUG_ON(!inode);
2800 dout("flush_dirty_caps inode %p\n", inode);
2801 }
2802 if (n != &mdsc->cap_dirty) {
2803 nci = list_entry(n, struct ceph_inode_info,
2804 i_dirty_item);
2805 ninode = igrab(&nci->vfs_inode);
2806 BUG_ON(!ninode);
2807 nci->i_ceph_flags |= CEPH_I_NOFLUSH;
2808 dout("flush_dirty_caps next inode %p, noflush\n",
2809 ninode);
2810 } else {
2811 nci = NULL;
2812 ninode = NULL;
2813 }
2814 spin_unlock(&mdsc->cap_dirty_lock);
2815 if (inode) {
2816 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH,
2817 NULL);
2818 iput(inode);
2819 }
2820 spin_lock(&mdsc->cap_dirty_lock);
2821 }
2822 spin_unlock(&mdsc->cap_dirty_lock);
2823}
2824
2825/*
2826 * Drop open file reference. If we were the last open file,
2827 * we may need to release capabilities to the MDS (or schedule
2828 * their delayed release).
2829 */
2830void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
2831{
2832 struct inode *inode = &ci->vfs_inode;
2833 int last = 0;
2834
2835 spin_lock(&inode->i_lock);
2836 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
2837 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
2838 BUG_ON(ci->i_nr_by_mode[fmode] == 0);
2839 if (--ci->i_nr_by_mode[fmode] == 0)
2840 last++;
2841 spin_unlock(&inode->i_lock);
2842
2843 if (last && ci->i_vino.snap == CEPH_NOSNAP)
2844 ceph_check_caps(ci, 0, NULL);
2845}
2846
2847/*
2848 * Helpers for embedding cap and dentry lease releases into mds
2849 * requests.
2850 *
2851 * @force is used by dentry_release (below) to force inclusion of a
2852 * record for the directory inode, even when there aren't any caps to
2853 * drop.
2854 */
2855int ceph_encode_inode_release(void **p, struct inode *inode,
2856 int mds, int drop, int unless, int force)
2857{
2858 struct ceph_inode_info *ci = ceph_inode(inode);
2859 struct ceph_cap *cap;
2860 struct ceph_mds_request_release *rel = *p;
2861 int ret = 0;
2862 int used = 0;
2863
2864 spin_lock(&inode->i_lock);
2865 used = __ceph_caps_used(ci);
2866
2867 dout("encode_inode_release %p mds%d used %s drop %s unless %s\n", inode,
2868 mds, ceph_cap_string(used), ceph_cap_string(drop),
2869 ceph_cap_string(unless));
2870
2871 /* only drop unused caps */
2872 drop &= ~used;
2873
2874 cap = __get_cap_for_mds(ci, mds);
2875 if (cap && __cap_is_valid(cap)) {
2876 if (force ||
2877 ((cap->issued & drop) &&
2878 (cap->issued & unless) == 0)) {
2879 if ((cap->issued & drop) &&
2880 (cap->issued & unless) == 0) {
2881 dout("encode_inode_release %p cap %p %s -> "
2882 "%s\n", inode, cap,
2883 ceph_cap_string(cap->issued),
2884 ceph_cap_string(cap->issued & ~drop));
2885 cap->issued &= ~drop;
2886 cap->implemented &= ~drop;
2887 if (ci->i_ceph_flags & CEPH_I_NODELAY) {
2888 int wanted = __ceph_caps_wanted(ci);
2889 dout(" wanted %s -> %s (act %s)\n",
2890 ceph_cap_string(cap->mds_wanted),
2891 ceph_cap_string(cap->mds_wanted &
2892 ~wanted),
2893 ceph_cap_string(wanted));
2894 cap->mds_wanted &= wanted;
2895 }
2896 } else {
2897 dout("encode_inode_release %p cap %p %s"
2898 " (force)\n", inode, cap,
2899 ceph_cap_string(cap->issued));
2900 }
2901
2902 rel->ino = cpu_to_le64(ceph_ino(inode));
2903 rel->cap_id = cpu_to_le64(cap->cap_id);
2904 rel->seq = cpu_to_le32(cap->seq);
2905 rel->issue_seq = cpu_to_le32(cap->issue_seq),
2906 rel->mseq = cpu_to_le32(cap->mseq);
2907 rel->caps = cpu_to_le32(cap->issued);
2908 rel->wanted = cpu_to_le32(cap->mds_wanted);
2909 rel->dname_len = 0;
2910 rel->dname_seq = 0;
2911 *p += sizeof(*rel);
2912 ret = 1;
2913 } else {
2914 dout("encode_inode_release %p cap %p %s\n",
2915 inode, cap, ceph_cap_string(cap->issued));
2916 }
2917 }
2918 spin_unlock(&inode->i_lock);
2919 return ret;
2920}
2921
2922int ceph_encode_dentry_release(void **p, struct dentry *dentry,
2923 int mds, int drop, int unless)
2924{
2925 struct inode *dir = dentry->d_parent->d_inode;
2926 struct ceph_mds_request_release *rel = *p;
2927 struct ceph_dentry_info *di = ceph_dentry(dentry);
2928 int force = 0;
2929 int ret;
2930
2931 /*
2932 * force an record for the directory caps if we have a dentry lease.
2933 * this is racy (can't take i_lock and d_lock together), but it
2934 * doesn't have to be perfect; the mds will revoke anything we don't
2935 * release.
2936 */
2937 spin_lock(&dentry->d_lock);
2938 if (di->lease_session && di->lease_session->s_mds == mds)
2939 force = 1;
2940 spin_unlock(&dentry->d_lock);
2941
2942 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
2943
2944 spin_lock(&dentry->d_lock);
2945 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
2946 dout("encode_dentry_release %p mds%d seq %d\n",
2947 dentry, mds, (int)di->lease_seq);
2948 rel->dname_len = cpu_to_le32(dentry->d_name.len);
2949 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
2950 *p += dentry->d_name.len;
2951 rel->dname_seq = cpu_to_le32(di->lease_seq);
2952 }
2953 spin_unlock(&dentry->d_lock);
2954 return ret;
2955}
generated by cgit v1.2.2 (git 2.25.0) at 2024-11-08 16:49:29 -0500