Merge branch 'for-linus' of git://github.com/dtor/input
[linux-btrfs-devel.git] / fs / ceph / caps.c
blob8d74ad7ba556624c4eb80e1f598ac908e5356d04
1 #include <linux/ceph/ceph_debug.h>
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>
11 #include "super.h"
12 #include "mds_client.h"
13 #include <linux/ceph/decode.h>
14 #include <linux/ceph/messenger.h>
17 * Capability management
19 * The Ceph metadata servers control client access to inode metadata
20 * and file data by issuing capabilities, granting clients permission
21 * to read and/or write both inode field and file data to OSDs
22 * (storage nodes). Each capability consists of a set of bits
23 * indicating which operations are allowed.
25 * If the client holds a *_SHARED cap, the client has a coherent value
26 * that can be safely read from the cached inode.
28 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
29 * client is allowed to change inode attributes (e.g., file size,
30 * mtime), note its dirty state in the ceph_cap, and asynchronously
31 * flush that metadata change to the MDS.
33 * In the event of a conflicting operation (perhaps by another
34 * client), the MDS will revoke the conflicting client capabilities.
36 * In order for a client to cache an inode, it must hold a capability
37 * with at least one MDS server. When inodes are released, release
38 * notifications are batched and periodically sent en masse to the MDS
39 * cluster to release server state.
44 * Generate readable cap strings for debugging output.
46 #define MAX_CAP_STR 20
47 static char cap_str[MAX_CAP_STR][40];
48 static DEFINE_SPINLOCK(cap_str_lock);
49 static int last_cap_str;
51 static char *gcap_string(char *s, int c)
53 if (c & CEPH_CAP_GSHARED)
54 *s++ = 's';
55 if (c & CEPH_CAP_GEXCL)
56 *s++ = 'x';
57 if (c & CEPH_CAP_GCACHE)
58 *s++ = 'c';
59 if (c & CEPH_CAP_GRD)
60 *s++ = 'r';
61 if (c & CEPH_CAP_GWR)
62 *s++ = 'w';
63 if (c & CEPH_CAP_GBUFFER)
64 *s++ = 'b';
65 if (c & CEPH_CAP_GLAZYIO)
66 *s++ = 'l';
67 return s;
70 const char *ceph_cap_string(int caps)
72 int i;
73 char *s;
74 int c;
76 spin_lock(&cap_str_lock);
77 i = last_cap_str++;
78 if (last_cap_str == MAX_CAP_STR)
79 last_cap_str = 0;
80 spin_unlock(&cap_str_lock);
82 s = cap_str[i];
84 if (caps & CEPH_CAP_PIN)
85 *s++ = 'p';
87 c = (caps >> CEPH_CAP_SAUTH) & 3;
88 if (c) {
89 *s++ = 'A';
90 s = gcap_string(s, c);
93 c = (caps >> CEPH_CAP_SLINK) & 3;
94 if (c) {
95 *s++ = 'L';
96 s = gcap_string(s, c);
99 c = (caps >> CEPH_CAP_SXATTR) & 3;
100 if (c) {
101 *s++ = 'X';
102 s = gcap_string(s, c);
105 c = caps >> CEPH_CAP_SFILE;
106 if (c) {
107 *s++ = 'F';
108 s = gcap_string(s, c);
111 if (s == cap_str[i])
112 *s++ = '-';
113 *s = 0;
114 return cap_str[i];
117 void ceph_caps_init(struct ceph_mds_client *mdsc)
119 INIT_LIST_HEAD(&mdsc->caps_list);
120 spin_lock_init(&mdsc->caps_list_lock);
123 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
125 struct ceph_cap *cap;
127 spin_lock(&mdsc->caps_list_lock);
128 while (!list_empty(&mdsc->caps_list)) {
129 cap = list_first_entry(&mdsc->caps_list,
130 struct ceph_cap, caps_item);
131 list_del(&cap->caps_item);
132 kmem_cache_free(ceph_cap_cachep, cap);
134 mdsc->caps_total_count = 0;
135 mdsc->caps_avail_count = 0;
136 mdsc->caps_use_count = 0;
137 mdsc->caps_reserve_count = 0;
138 mdsc->caps_min_count = 0;
139 spin_unlock(&mdsc->caps_list_lock);
142 void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta)
144 spin_lock(&mdsc->caps_list_lock);
145 mdsc->caps_min_count += delta;
146 BUG_ON(mdsc->caps_min_count < 0);
147 spin_unlock(&mdsc->caps_list_lock);
150 int ceph_reserve_caps(struct ceph_mds_client *mdsc,
151 struct ceph_cap_reservation *ctx, int need)
153 int i;
154 struct ceph_cap *cap;
155 int have;
156 int alloc = 0;
157 LIST_HEAD(newcaps);
158 int ret = 0;
160 dout("reserve caps ctx=%p need=%d\n", ctx, need);
162 /* first reserve any caps that are already allocated */
163 spin_lock(&mdsc->caps_list_lock);
164 if (mdsc->caps_avail_count >= need)
165 have = need;
166 else
167 have = mdsc->caps_avail_count;
168 mdsc->caps_avail_count -= have;
169 mdsc->caps_reserve_count += have;
170 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
171 mdsc->caps_reserve_count +
172 mdsc->caps_avail_count);
173 spin_unlock(&mdsc->caps_list_lock);
175 for (i = have; i < need; i++) {
176 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
177 if (!cap) {
178 ret = -ENOMEM;
179 goto out_alloc_count;
181 list_add(&cap->caps_item, &newcaps);
182 alloc++;
184 BUG_ON(have + alloc != need);
186 spin_lock(&mdsc->caps_list_lock);
187 mdsc->caps_total_count += alloc;
188 mdsc->caps_reserve_count += alloc;
189 list_splice(&newcaps, &mdsc->caps_list);
191 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
192 mdsc->caps_reserve_count +
193 mdsc->caps_avail_count);
194 spin_unlock(&mdsc->caps_list_lock);
196 ctx->count = need;
197 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
198 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
199 mdsc->caps_reserve_count, mdsc->caps_avail_count);
200 return 0;
202 out_alloc_count:
203 /* we didn't manage to reserve as much as we needed */
204 pr_warning("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
205 ctx, need, have);
206 return ret;
209 int ceph_unreserve_caps(struct ceph_mds_client *mdsc,
210 struct ceph_cap_reservation *ctx)
212 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
213 if (ctx->count) {
214 spin_lock(&mdsc->caps_list_lock);
215 BUG_ON(mdsc->caps_reserve_count < ctx->count);
216 mdsc->caps_reserve_count -= ctx->count;
217 mdsc->caps_avail_count += ctx->count;
218 ctx->count = 0;
219 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
220 mdsc->caps_total_count, mdsc->caps_use_count,
221 mdsc->caps_reserve_count, mdsc->caps_avail_count);
222 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
223 mdsc->caps_reserve_count +
224 mdsc->caps_avail_count);
225 spin_unlock(&mdsc->caps_list_lock);
227 return 0;
230 static struct ceph_cap *get_cap(struct ceph_mds_client *mdsc,
231 struct ceph_cap_reservation *ctx)
233 struct ceph_cap *cap = NULL;
235 /* temporary, until we do something about cap import/export */
236 if (!ctx) {
237 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
238 if (cap) {
239 mdsc->caps_use_count++;
240 mdsc->caps_total_count++;
242 return cap;
245 spin_lock(&mdsc->caps_list_lock);
246 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
247 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
248 mdsc->caps_reserve_count, mdsc->caps_avail_count);
249 BUG_ON(!ctx->count);
250 BUG_ON(ctx->count > mdsc->caps_reserve_count);
251 BUG_ON(list_empty(&mdsc->caps_list));
253 ctx->count--;
254 mdsc->caps_reserve_count--;
255 mdsc->caps_use_count++;
257 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
258 list_del(&cap->caps_item);
260 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
261 mdsc->caps_reserve_count + mdsc->caps_avail_count);
262 spin_unlock(&mdsc->caps_list_lock);
263 return cap;
266 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
268 spin_lock(&mdsc->caps_list_lock);
269 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
270 cap, mdsc->caps_total_count, mdsc->caps_use_count,
271 mdsc->caps_reserve_count, mdsc->caps_avail_count);
272 mdsc->caps_use_count--;
274 * Keep some preallocated caps around (ceph_min_count), to
275 * avoid lots of free/alloc churn.
277 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
278 mdsc->caps_min_count) {
279 mdsc->caps_total_count--;
280 kmem_cache_free(ceph_cap_cachep, cap);
281 } else {
282 mdsc->caps_avail_count++;
283 list_add(&cap->caps_item, &mdsc->caps_list);
286 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
287 mdsc->caps_reserve_count + mdsc->caps_avail_count);
288 spin_unlock(&mdsc->caps_list_lock);
291 void ceph_reservation_status(struct ceph_fs_client *fsc,
292 int *total, int *avail, int *used, int *reserved,
293 int *min)
295 struct ceph_mds_client *mdsc = fsc->mdsc;
297 if (total)
298 *total = mdsc->caps_total_count;
299 if (avail)
300 *avail = mdsc->caps_avail_count;
301 if (used)
302 *used = mdsc->caps_use_count;
303 if (reserved)
304 *reserved = mdsc->caps_reserve_count;
305 if (min)
306 *min = mdsc->caps_min_count;
310 * Find ceph_cap for given mds, if any.
312 * Called with i_lock held.
314 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
316 struct ceph_cap *cap;
317 struct rb_node *n = ci->i_caps.rb_node;
319 while (n) {
320 cap = rb_entry(n, struct ceph_cap, ci_node);
321 if (mds < cap->mds)
322 n = n->rb_left;
323 else if (mds > cap->mds)
324 n = n->rb_right;
325 else
326 return cap;
328 return NULL;
331 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
333 struct ceph_cap *cap;
335 spin_lock(&ci->vfs_inode.i_lock);
336 cap = __get_cap_for_mds(ci, mds);
337 spin_unlock(&ci->vfs_inode.i_lock);
338 return cap;
342 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
344 static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
346 struct ceph_cap *cap;
347 int mds = -1;
348 struct rb_node *p;
350 /* prefer mds with WR|BUFFER|EXCL caps */
351 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
352 cap = rb_entry(p, struct ceph_cap, ci_node);
353 mds = cap->mds;
354 if (cap->issued & (CEPH_CAP_FILE_WR |
355 CEPH_CAP_FILE_BUFFER |
356 CEPH_CAP_FILE_EXCL))
357 break;
359 return mds;
362 int ceph_get_cap_mds(struct inode *inode)
364 int mds;
365 spin_lock(&inode->i_lock);
366 mds = __ceph_get_cap_mds(ceph_inode(inode));
367 spin_unlock(&inode->i_lock);
368 return mds;
372 * Called under i_lock.
374 static void __insert_cap_node(struct ceph_inode_info *ci,
375 struct ceph_cap *new)
377 struct rb_node **p = &ci->i_caps.rb_node;
378 struct rb_node *parent = NULL;
379 struct ceph_cap *cap = NULL;
381 while (*p) {
382 parent = *p;
383 cap = rb_entry(parent, struct ceph_cap, ci_node);
384 if (new->mds < cap->mds)
385 p = &(*p)->rb_left;
386 else if (new->mds > cap->mds)
387 p = &(*p)->rb_right;
388 else
389 BUG();
392 rb_link_node(&new->ci_node, parent, p);
393 rb_insert_color(&new->ci_node, &ci->i_caps);
397 * (re)set cap hold timeouts, which control the delayed release
398 * of unused caps back to the MDS. Should be called on cap use.
400 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
401 struct ceph_inode_info *ci)
403 struct ceph_mount_options *ma = mdsc->fsc->mount_options;
405 ci->i_hold_caps_min = round_jiffies(jiffies +
406 ma->caps_wanted_delay_min * HZ);
407 ci->i_hold_caps_max = round_jiffies(jiffies +
408 ma->caps_wanted_delay_max * HZ);
409 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
410 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
414 * (Re)queue cap at the end of the delayed cap release list.
416 * If I_FLUSH is set, leave the inode at the front of the list.
418 * Caller holds i_lock
419 * -> we take mdsc->cap_delay_lock
421 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
422 struct ceph_inode_info *ci)
424 __cap_set_timeouts(mdsc, ci);
425 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
426 ci->i_ceph_flags, ci->i_hold_caps_max);
427 if (!mdsc->stopping) {
428 spin_lock(&mdsc->cap_delay_lock);
429 if (!list_empty(&ci->i_cap_delay_list)) {
430 if (ci->i_ceph_flags & CEPH_I_FLUSH)
431 goto no_change;
432 list_del_init(&ci->i_cap_delay_list);
434 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
435 no_change:
436 spin_unlock(&mdsc->cap_delay_lock);
441 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
442 * indicating we should send a cap message to flush dirty metadata
443 * asap, and move to the front of the delayed cap list.
445 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
446 struct ceph_inode_info *ci)
448 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
449 spin_lock(&mdsc->cap_delay_lock);
450 ci->i_ceph_flags |= CEPH_I_FLUSH;
451 if (!list_empty(&ci->i_cap_delay_list))
452 list_del_init(&ci->i_cap_delay_list);
453 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
454 spin_unlock(&mdsc->cap_delay_lock);
458 * Cancel delayed work on cap.
460 * Caller must hold i_lock.
462 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
463 struct ceph_inode_info *ci)
465 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
466 if (list_empty(&ci->i_cap_delay_list))
467 return;
468 spin_lock(&mdsc->cap_delay_lock);
469 list_del_init(&ci->i_cap_delay_list);
470 spin_unlock(&mdsc->cap_delay_lock);
474 * Common issue checks for add_cap, handle_cap_grant.
476 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
477 unsigned issued)
479 unsigned had = __ceph_caps_issued(ci, NULL);
482 * Each time we receive FILE_CACHE anew, we increment
483 * i_rdcache_gen.
485 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
486 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
487 ci->i_rdcache_gen++;
490 * if we are newly issued FILE_SHARED, clear I_COMPLETE; we
491 * don't know what happened to this directory while we didn't
492 * have the cap.
494 if ((issued & CEPH_CAP_FILE_SHARED) &&
495 (had & CEPH_CAP_FILE_SHARED) == 0) {
496 ci->i_shared_gen++;
497 if (S_ISDIR(ci->vfs_inode.i_mode)) {
498 dout(" marking %p NOT complete\n", &ci->vfs_inode);
499 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
505 * Add a capability under the given MDS session.
507 * Caller should hold session snap_rwsem (read) and s_mutex.
509 * @fmode is the open file mode, if we are opening a file, otherwise
510 * it is < 0. (This is so we can atomically add the cap and add an
511 * open file reference to it.)
513 int ceph_add_cap(struct inode *inode,
514 struct ceph_mds_session *session, u64 cap_id,
515 int fmode, unsigned issued, unsigned wanted,
516 unsigned seq, unsigned mseq, u64 realmino, int flags,
517 struct ceph_cap_reservation *caps_reservation)
519 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
520 struct ceph_inode_info *ci = ceph_inode(inode);
521 struct ceph_cap *new_cap = NULL;
522 struct ceph_cap *cap;
523 int mds = session->s_mds;
524 int actual_wanted;
526 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
527 session->s_mds, cap_id, ceph_cap_string(issued), seq);
530 * If we are opening the file, include file mode wanted bits
531 * in wanted.
533 if (fmode >= 0)
534 wanted |= ceph_caps_for_mode(fmode);
536 retry:
537 spin_lock(&inode->i_lock);
538 cap = __get_cap_for_mds(ci, mds);
539 if (!cap) {
540 if (new_cap) {
541 cap = new_cap;
542 new_cap = NULL;
543 } else {
544 spin_unlock(&inode->i_lock);
545 new_cap = get_cap(mdsc, caps_reservation);
546 if (new_cap == NULL)
547 return -ENOMEM;
548 goto retry;
551 cap->issued = 0;
552 cap->implemented = 0;
553 cap->mds = mds;
554 cap->mds_wanted = 0;
556 cap->ci = ci;
557 __insert_cap_node(ci, cap);
559 /* clear out old exporting info? (i.e. on cap import) */
560 if (ci->i_cap_exporting_mds == mds) {
561 ci->i_cap_exporting_issued = 0;
562 ci->i_cap_exporting_mseq = 0;
563 ci->i_cap_exporting_mds = -1;
566 /* add to session cap list */
567 cap->session = session;
568 spin_lock(&session->s_cap_lock);
569 list_add_tail(&cap->session_caps, &session->s_caps);
570 session->s_nr_caps++;
571 spin_unlock(&session->s_cap_lock);
572 } else if (new_cap)
573 ceph_put_cap(mdsc, new_cap);
575 if (!ci->i_snap_realm) {
577 * add this inode to the appropriate snap realm
579 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
580 realmino);
581 if (realm) {
582 ceph_get_snap_realm(mdsc, realm);
583 spin_lock(&realm->inodes_with_caps_lock);
584 ci->i_snap_realm = realm;
585 list_add(&ci->i_snap_realm_item,
586 &realm->inodes_with_caps);
587 spin_unlock(&realm->inodes_with_caps_lock);
588 } else {
589 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
590 realmino);
591 WARN_ON(!realm);
595 __check_cap_issue(ci, cap, issued);
598 * If we are issued caps we don't want, or the mds' wanted
599 * value appears to be off, queue a check so we'll release
600 * later and/or update the mds wanted value.
602 actual_wanted = __ceph_caps_wanted(ci);
603 if ((wanted & ~actual_wanted) ||
604 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
605 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
606 ceph_cap_string(issued), ceph_cap_string(wanted),
607 ceph_cap_string(actual_wanted));
608 __cap_delay_requeue(mdsc, ci);
611 if (flags & CEPH_CAP_FLAG_AUTH)
612 ci->i_auth_cap = cap;
613 else if (ci->i_auth_cap == cap)
614 ci->i_auth_cap = NULL;
616 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
617 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
618 ceph_cap_string(issued|cap->issued), seq, mds);
619 cap->cap_id = cap_id;
620 cap->issued = issued;
621 cap->implemented |= issued;
622 cap->mds_wanted |= wanted;
623 cap->seq = seq;
624 cap->issue_seq = seq;
625 cap->mseq = mseq;
626 cap->cap_gen = session->s_cap_gen;
628 if (fmode >= 0)
629 __ceph_get_fmode(ci, fmode);
630 spin_unlock(&inode->i_lock);
631 wake_up_all(&ci->i_cap_wq);
632 return 0;
636 * Return true if cap has not timed out and belongs to the current
637 * generation of the MDS session (i.e. has not gone 'stale' due to
638 * us losing touch with the mds).
640 static int __cap_is_valid(struct ceph_cap *cap)
642 unsigned long ttl;
643 u32 gen;
645 spin_lock(&cap->session->s_cap_lock);
646 gen = cap->session->s_cap_gen;
647 ttl = cap->session->s_cap_ttl;
648 spin_unlock(&cap->session->s_cap_lock);
650 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
651 dout("__cap_is_valid %p cap %p issued %s "
652 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
653 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
654 return 0;
657 return 1;
661 * Return set of valid cap bits issued to us. Note that caps time
662 * out, and may be invalidated in bulk if the client session times out
663 * and session->s_cap_gen is bumped.
665 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
667 int have = ci->i_snap_caps | ci->i_cap_exporting_issued;
668 struct ceph_cap *cap;
669 struct rb_node *p;
671 if (implemented)
672 *implemented = 0;
673 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
674 cap = rb_entry(p, struct ceph_cap, ci_node);
675 if (!__cap_is_valid(cap))
676 continue;
677 dout("__ceph_caps_issued %p cap %p issued %s\n",
678 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
679 have |= cap->issued;
680 if (implemented)
681 *implemented |= cap->implemented;
683 return have;
687 * Get cap bits issued by caps other than @ocap
689 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
691 int have = ci->i_snap_caps;
692 struct ceph_cap *cap;
693 struct rb_node *p;
695 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
696 cap = rb_entry(p, struct ceph_cap, ci_node);
697 if (cap == ocap)
698 continue;
699 if (!__cap_is_valid(cap))
700 continue;
701 have |= cap->issued;
703 return have;
707 * Move a cap to the end of the LRU (oldest caps at list head, newest
708 * at list tail).
710 static void __touch_cap(struct ceph_cap *cap)
712 struct ceph_mds_session *s = cap->session;
714 spin_lock(&s->s_cap_lock);
715 if (s->s_cap_iterator == NULL) {
716 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
717 s->s_mds);
718 list_move_tail(&cap->session_caps, &s->s_caps);
719 } else {
720 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
721 &cap->ci->vfs_inode, cap, s->s_mds);
723 spin_unlock(&s->s_cap_lock);
727 * Check if we hold the given mask. If so, move the cap(s) to the
728 * front of their respective LRUs. (This is the preferred way for
729 * callers to check for caps they want.)
731 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
733 struct ceph_cap *cap;
734 struct rb_node *p;
735 int have = ci->i_snap_caps;
737 if ((have & mask) == mask) {
738 dout("__ceph_caps_issued_mask %p snap issued %s"
739 " (mask %s)\n", &ci->vfs_inode,
740 ceph_cap_string(have),
741 ceph_cap_string(mask));
742 return 1;
745 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
746 cap = rb_entry(p, struct ceph_cap, ci_node);
747 if (!__cap_is_valid(cap))
748 continue;
749 if ((cap->issued & mask) == mask) {
750 dout("__ceph_caps_issued_mask %p cap %p issued %s"
751 " (mask %s)\n", &ci->vfs_inode, cap,
752 ceph_cap_string(cap->issued),
753 ceph_cap_string(mask));
754 if (touch)
755 __touch_cap(cap);
756 return 1;
759 /* does a combination of caps satisfy mask? */
760 have |= cap->issued;
761 if ((have & mask) == mask) {
762 dout("__ceph_caps_issued_mask %p combo issued %s"
763 " (mask %s)\n", &ci->vfs_inode,
764 ceph_cap_string(cap->issued),
765 ceph_cap_string(mask));
766 if (touch) {
767 struct rb_node *q;
769 /* touch this + preceding caps */
770 __touch_cap(cap);
771 for (q = rb_first(&ci->i_caps); q != p;
772 q = rb_next(q)) {
773 cap = rb_entry(q, struct ceph_cap,
774 ci_node);
775 if (!__cap_is_valid(cap))
776 continue;
777 __touch_cap(cap);
780 return 1;
784 return 0;
788 * Return true if mask caps are currently being revoked by an MDS.
790 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
792 struct inode *inode = &ci->vfs_inode;
793 struct ceph_cap *cap;
794 struct rb_node *p;
795 int ret = 0;
797 spin_lock(&inode->i_lock);
798 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
799 cap = rb_entry(p, struct ceph_cap, ci_node);
800 if (__cap_is_valid(cap) &&
801 (cap->implemented & ~cap->issued & mask)) {
802 ret = 1;
803 break;
806 spin_unlock(&inode->i_lock);
807 dout("ceph_caps_revoking %p %s = %d\n", inode,
808 ceph_cap_string(mask), ret);
809 return ret;
812 int __ceph_caps_used(struct ceph_inode_info *ci)
814 int used = 0;
815 if (ci->i_pin_ref)
816 used |= CEPH_CAP_PIN;
817 if (ci->i_rd_ref)
818 used |= CEPH_CAP_FILE_RD;
819 if (ci->i_rdcache_ref || ci->vfs_inode.i_data.nrpages)
820 used |= CEPH_CAP_FILE_CACHE;
821 if (ci->i_wr_ref)
822 used |= CEPH_CAP_FILE_WR;
823 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
824 used |= CEPH_CAP_FILE_BUFFER;
825 return used;
829 * wanted, by virtue of open file modes
831 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
833 int want = 0;
834 int mode;
835 for (mode = 0; mode < CEPH_FILE_MODE_NUM; mode++)
836 if (ci->i_nr_by_mode[mode])
837 want |= ceph_caps_for_mode(mode);
838 return want;
842 * Return caps we have registered with the MDS(s) as 'wanted'.
844 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
846 struct ceph_cap *cap;
847 struct rb_node *p;
848 int mds_wanted = 0;
850 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
851 cap = rb_entry(p, struct ceph_cap, ci_node);
852 if (!__cap_is_valid(cap))
853 continue;
854 mds_wanted |= cap->mds_wanted;
856 return mds_wanted;
860 * called under i_lock
862 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
864 return !RB_EMPTY_ROOT(&ci->i_caps) || ci->i_cap_exporting_mds >= 0;
868 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
870 * caller should hold i_lock.
871 * caller will not hold session s_mutex if called from destroy_inode.
873 void __ceph_remove_cap(struct ceph_cap *cap)
875 struct ceph_mds_session *session = cap->session;
876 struct ceph_inode_info *ci = cap->ci;
877 struct ceph_mds_client *mdsc =
878 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
879 int removed = 0;
881 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
883 /* remove from session list */
884 spin_lock(&session->s_cap_lock);
885 if (session->s_cap_iterator == cap) {
886 /* not yet, we are iterating over this very cap */
887 dout("__ceph_remove_cap delaying %p removal from session %p\n",
888 cap, cap->session);
889 } else {
890 list_del_init(&cap->session_caps);
891 session->s_nr_caps--;
892 cap->session = NULL;
893 removed = 1;
895 /* protect backpointer with s_cap_lock: see iterate_session_caps */
896 cap->ci = NULL;
897 spin_unlock(&session->s_cap_lock);
899 /* remove from inode list */
900 rb_erase(&cap->ci_node, &ci->i_caps);
901 if (ci->i_auth_cap == cap)
902 ci->i_auth_cap = NULL;
904 if (removed)
905 ceph_put_cap(mdsc, cap);
907 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) {
908 struct ceph_snap_realm *realm = ci->i_snap_realm;
909 spin_lock(&realm->inodes_with_caps_lock);
910 list_del_init(&ci->i_snap_realm_item);
911 ci->i_snap_realm_counter++;
912 ci->i_snap_realm = NULL;
913 spin_unlock(&realm->inodes_with_caps_lock);
914 ceph_put_snap_realm(mdsc, realm);
916 if (!__ceph_is_any_real_caps(ci))
917 __cap_delay_cancel(mdsc, ci);
921 * Build and send a cap message to the given MDS.
923 * Caller should be holding s_mutex.
925 static int send_cap_msg(struct ceph_mds_session *session,
926 u64 ino, u64 cid, int op,
927 int caps, int wanted, int dirty,
928 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq,
929 u64 size, u64 max_size,
930 struct timespec *mtime, struct timespec *atime,
931 u64 time_warp_seq,
932 uid_t uid, gid_t gid, mode_t mode,
933 u64 xattr_version,
934 struct ceph_buffer *xattrs_buf,
935 u64 follows)
937 struct ceph_mds_caps *fc;
938 struct ceph_msg *msg;
940 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
941 " seq %u/%u mseq %u follows %lld size %llu/%llu"
942 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
943 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
944 ceph_cap_string(dirty),
945 seq, issue_seq, mseq, follows, size, max_size,
946 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
948 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc), GFP_NOFS);
949 if (!msg)
950 return -ENOMEM;
952 msg->hdr.tid = cpu_to_le64(flush_tid);
954 fc = msg->front.iov_base;
955 memset(fc, 0, sizeof(*fc));
957 fc->cap_id = cpu_to_le64(cid);
958 fc->op = cpu_to_le32(op);
959 fc->seq = cpu_to_le32(seq);
960 fc->issue_seq = cpu_to_le32(issue_seq);
961 fc->migrate_seq = cpu_to_le32(mseq);
962 fc->caps = cpu_to_le32(caps);
963 fc->wanted = cpu_to_le32(wanted);
964 fc->dirty = cpu_to_le32(dirty);
965 fc->ino = cpu_to_le64(ino);
966 fc->snap_follows = cpu_to_le64(follows);
968 fc->size = cpu_to_le64(size);
969 fc->max_size = cpu_to_le64(max_size);
970 if (mtime)
971 ceph_encode_timespec(&fc->mtime, mtime);
972 if (atime)
973 ceph_encode_timespec(&fc->atime, atime);
974 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
976 fc->uid = cpu_to_le32(uid);
977 fc->gid = cpu_to_le32(gid);
978 fc->mode = cpu_to_le32(mode);
980 fc->xattr_version = cpu_to_le64(xattr_version);
981 if (xattrs_buf) {
982 msg->middle = ceph_buffer_get(xattrs_buf);
983 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
984 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
987 ceph_con_send(&session->s_con, msg);
988 return 0;
991 static void __queue_cap_release(struct ceph_mds_session *session,
992 u64 ino, u64 cap_id, u32 migrate_seq,
993 u32 issue_seq)
995 struct ceph_msg *msg;
996 struct ceph_mds_cap_release *head;
997 struct ceph_mds_cap_item *item;
999 spin_lock(&session->s_cap_lock);
1000 BUG_ON(!session->s_num_cap_releases);
1001 msg = list_first_entry(&session->s_cap_releases,
1002 struct ceph_msg, list_head);
1004 dout(" adding %llx release to mds%d msg %p (%d left)\n",
1005 ino, session->s_mds, msg, session->s_num_cap_releases);
1007 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE);
1008 head = msg->front.iov_base;
1009 head->num = cpu_to_le32(le32_to_cpu(head->num) + 1);
1010 item = msg->front.iov_base + msg->front.iov_len;
1011 item->ino = cpu_to_le64(ino);
1012 item->cap_id = cpu_to_le64(cap_id);
1013 item->migrate_seq = cpu_to_le32(migrate_seq);
1014 item->seq = cpu_to_le32(issue_seq);
1016 session->s_num_cap_releases--;
1018 msg->front.iov_len += sizeof(*item);
1019 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1020 dout(" release msg %p full\n", msg);
1021 list_move_tail(&msg->list_head, &session->s_cap_releases_done);
1022 } else {
1023 dout(" release msg %p at %d/%d (%d)\n", msg,
1024 (int)le32_to_cpu(head->num),
1025 (int)CEPH_CAPS_PER_RELEASE,
1026 (int)msg->front.iov_len);
1028 spin_unlock(&session->s_cap_lock);
1032 * Queue cap releases when an inode is dropped from our cache. Since
1033 * inode is about to be destroyed, there is no need for i_lock.
1035 void ceph_queue_caps_release(struct inode *inode)
1037 struct ceph_inode_info *ci = ceph_inode(inode);
1038 struct rb_node *p;
1040 p = rb_first(&ci->i_caps);
1041 while (p) {
1042 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1043 struct ceph_mds_session *session = cap->session;
1045 __queue_cap_release(session, ceph_ino(inode), cap->cap_id,
1046 cap->mseq, cap->issue_seq);
1047 p = rb_next(p);
1048 __ceph_remove_cap(cap);
1053 * Send a cap msg on the given inode. Update our caps state, then
1054 * drop i_lock and send the message.
1056 * Make note of max_size reported/requested from mds, revoked caps
1057 * that have now been implemented.
1059 * Make half-hearted attempt ot to invalidate page cache if we are
1060 * dropping RDCACHE. Note that this will leave behind locked pages
1061 * that we'll then need to deal with elsewhere.
1063 * Return non-zero if delayed release, or we experienced an error
1064 * such that the caller should requeue + retry later.
1066 * called with i_lock, then drops it.
1067 * caller should hold snap_rwsem (read), s_mutex.
1069 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1070 int op, int used, int want, int retain, int flushing,
1071 unsigned *pflush_tid)
1072 __releases(cap->ci->vfs_inode->i_lock)
1074 struct ceph_inode_info *ci = cap->ci;
1075 struct inode *inode = &ci->vfs_inode;
1076 u64 cap_id = cap->cap_id;
1077 int held, revoking, dropping, keep;
1078 u64 seq, issue_seq, mseq, time_warp_seq, follows;
1079 u64 size, max_size;
1080 struct timespec mtime, atime;
1081 int wake = 0;
1082 mode_t mode;
1083 uid_t uid;
1084 gid_t gid;
1085 struct ceph_mds_session *session;
1086 u64 xattr_version = 0;
1087 struct ceph_buffer *xattr_blob = NULL;
1088 int delayed = 0;
1089 u64 flush_tid = 0;
1090 int i;
1091 int ret;
1093 held = cap->issued | cap->implemented;
1094 revoking = cap->implemented & ~cap->issued;
1095 retain &= ~revoking;
1096 dropping = cap->issued & ~retain;
1098 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1099 inode, cap, cap->session,
1100 ceph_cap_string(held), ceph_cap_string(held & retain),
1101 ceph_cap_string(revoking));
1102 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1104 session = cap->session;
1106 /* don't release wanted unless we've waited a bit. */
1107 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1108 time_before(jiffies, ci->i_hold_caps_min)) {
1109 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1110 ceph_cap_string(cap->issued),
1111 ceph_cap_string(cap->issued & retain),
1112 ceph_cap_string(cap->mds_wanted),
1113 ceph_cap_string(want));
1114 want |= cap->mds_wanted;
1115 retain |= cap->issued;
1116 delayed = 1;
1118 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1120 cap->issued &= retain; /* drop bits we don't want */
1121 if (cap->implemented & ~cap->issued) {
1123 * Wake up any waiters on wanted -> needed transition.
1124 * This is due to the weird transition from buffered
1125 * to sync IO... we need to flush dirty pages _before_
1126 * allowing sync writes to avoid reordering.
1128 wake = 1;
1130 cap->implemented &= cap->issued | used;
1131 cap->mds_wanted = want;
1133 if (flushing) {
1135 * assign a tid for flush operations so we can avoid
1136 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark
1137 * clean type races. track latest tid for every bit
1138 * so we can handle flush AxFw, flush Fw, and have the
1139 * first ack clean Ax.
1141 flush_tid = ++ci->i_cap_flush_last_tid;
1142 if (pflush_tid)
1143 *pflush_tid = flush_tid;
1144 dout(" cap_flush_tid %d\n", (int)flush_tid);
1145 for (i = 0; i < CEPH_CAP_BITS; i++)
1146 if (flushing & (1 << i))
1147 ci->i_cap_flush_tid[i] = flush_tid;
1149 follows = ci->i_head_snapc->seq;
1150 } else {
1151 follows = 0;
1154 keep = cap->implemented;
1155 seq = cap->seq;
1156 issue_seq = cap->issue_seq;
1157 mseq = cap->mseq;
1158 size = inode->i_size;
1159 ci->i_reported_size = size;
1160 max_size = ci->i_wanted_max_size;
1161 ci->i_requested_max_size = max_size;
1162 mtime = inode->i_mtime;
1163 atime = inode->i_atime;
1164 time_warp_seq = ci->i_time_warp_seq;
1165 uid = inode->i_uid;
1166 gid = inode->i_gid;
1167 mode = inode->i_mode;
1169 if (flushing & CEPH_CAP_XATTR_EXCL) {
1170 __ceph_build_xattrs_blob(ci);
1171 xattr_blob = ci->i_xattrs.blob;
1172 xattr_version = ci->i_xattrs.version;
1175 spin_unlock(&inode->i_lock);
1177 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1178 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq,
1179 size, max_size, &mtime, &atime, time_warp_seq,
1180 uid, gid, mode, xattr_version, xattr_blob,
1181 follows);
1182 if (ret < 0) {
1183 dout("error sending cap msg, must requeue %p\n", inode);
1184 delayed = 1;
1187 if (wake)
1188 wake_up_all(&ci->i_cap_wq);
1190 return delayed;
1194 * When a snapshot is taken, clients accumulate dirty metadata on
1195 * inodes with capabilities in ceph_cap_snaps to describe the file
1196 * state at the time the snapshot was taken. This must be flushed
1197 * asynchronously back to the MDS once sync writes complete and dirty
1198 * data is written out.
1200 * Unless @again is true, skip cap_snaps that were already sent to
1201 * the MDS (i.e., during this session).
1203 * Called under i_lock. Takes s_mutex as needed.
1205 void __ceph_flush_snaps(struct ceph_inode_info *ci,
1206 struct ceph_mds_session **psession,
1207 int again)
1208 __releases(ci->vfs_inode->i_lock)
1209 __acquires(ci->vfs_inode->i_lock)
1211 struct inode *inode = &ci->vfs_inode;
1212 int mds;
1213 struct ceph_cap_snap *capsnap;
1214 u32 mseq;
1215 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1216 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1217 session->s_mutex */
1218 u64 next_follows = 0; /* keep track of how far we've gotten through the
1219 i_cap_snaps list, and skip these entries next time
1220 around to avoid an infinite loop */
1222 if (psession)
1223 session = *psession;
1225 dout("__flush_snaps %p\n", inode);
1226 retry:
1227 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1228 /* avoid an infiniute loop after retry */
1229 if (capsnap->follows < next_follows)
1230 continue;
1232 * we need to wait for sync writes to complete and for dirty
1233 * pages to be written out.
1235 if (capsnap->dirty_pages || capsnap->writing)
1236 break;
1239 * if cap writeback already occurred, we should have dropped
1240 * the capsnap in ceph_put_wrbuffer_cap_refs.
1242 BUG_ON(capsnap->dirty == 0);
1244 /* pick mds, take s_mutex */
1245 if (ci->i_auth_cap == NULL) {
1246 dout("no auth cap (migrating?), doing nothing\n");
1247 goto out;
1250 /* only flush each capsnap once */
1251 if (!again && !list_empty(&capsnap->flushing_item)) {
1252 dout("already flushed %p, skipping\n", capsnap);
1253 continue;
1256 mds = ci->i_auth_cap->session->s_mds;
1257 mseq = ci->i_auth_cap->mseq;
1259 if (session && session->s_mds != mds) {
1260 dout("oops, wrong session %p mutex\n", session);
1261 mutex_unlock(&session->s_mutex);
1262 ceph_put_mds_session(session);
1263 session = NULL;
1265 if (!session) {
1266 spin_unlock(&inode->i_lock);
1267 mutex_lock(&mdsc->mutex);
1268 session = __ceph_lookup_mds_session(mdsc, mds);
1269 mutex_unlock(&mdsc->mutex);
1270 if (session) {
1271 dout("inverting session/ino locks on %p\n",
1272 session);
1273 mutex_lock(&session->s_mutex);
1276 * if session == NULL, we raced against a cap
1277 * deletion or migration. retry, and we'll
1278 * get a better @mds value next time.
1280 spin_lock(&inode->i_lock);
1281 goto retry;
1284 capsnap->flush_tid = ++ci->i_cap_flush_last_tid;
1285 atomic_inc(&capsnap->nref);
1286 if (!list_empty(&capsnap->flushing_item))
1287 list_del_init(&capsnap->flushing_item);
1288 list_add_tail(&capsnap->flushing_item,
1289 &session->s_cap_snaps_flushing);
1290 spin_unlock(&inode->i_lock);
1292 dout("flush_snaps %p cap_snap %p follows %lld tid %llu\n",
1293 inode, capsnap, capsnap->follows, capsnap->flush_tid);
1294 send_cap_msg(session, ceph_vino(inode).ino, 0,
1295 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1296 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq,
1297 capsnap->size, 0,
1298 &capsnap->mtime, &capsnap->atime,
1299 capsnap->time_warp_seq,
1300 capsnap->uid, capsnap->gid, capsnap->mode,
1301 capsnap->xattr_version, capsnap->xattr_blob,
1302 capsnap->follows);
1304 next_follows = capsnap->follows + 1;
1305 ceph_put_cap_snap(capsnap);
1307 spin_lock(&inode->i_lock);
1308 goto retry;
1311 /* we flushed them all; remove this inode from the queue */
1312 spin_lock(&mdsc->snap_flush_lock);
1313 list_del_init(&ci->i_snap_flush_item);
1314 spin_unlock(&mdsc->snap_flush_lock);
1316 out:
1317 if (psession)
1318 *psession = session;
1319 else if (session) {
1320 mutex_unlock(&session->s_mutex);
1321 ceph_put_mds_session(session);
1325 static void ceph_flush_snaps(struct ceph_inode_info *ci)
1327 struct inode *inode = &ci->vfs_inode;
1329 spin_lock(&inode->i_lock);
1330 __ceph_flush_snaps(ci, NULL, 0);
1331 spin_unlock(&inode->i_lock);
1335 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1336 * Caller is then responsible for calling __mark_inode_dirty with the
1337 * returned flags value.
1339 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1341 struct ceph_mds_client *mdsc =
1342 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1343 struct inode *inode = &ci->vfs_inode;
1344 int was = ci->i_dirty_caps;
1345 int dirty = 0;
1347 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1348 ceph_cap_string(mask), ceph_cap_string(was),
1349 ceph_cap_string(was | mask));
1350 ci->i_dirty_caps |= mask;
1351 if (was == 0) {
1352 if (!ci->i_head_snapc)
1353 ci->i_head_snapc = ceph_get_snap_context(
1354 ci->i_snap_realm->cached_context);
1355 dout(" inode %p now dirty snapc %p\n", &ci->vfs_inode,
1356 ci->i_head_snapc);
1357 BUG_ON(!list_empty(&ci->i_dirty_item));
1358 spin_lock(&mdsc->cap_dirty_lock);
1359 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1360 spin_unlock(&mdsc->cap_dirty_lock);
1361 if (ci->i_flushing_caps == 0) {
1362 ihold(inode);
1363 dirty |= I_DIRTY_SYNC;
1366 BUG_ON(list_empty(&ci->i_dirty_item));
1367 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1368 (mask & CEPH_CAP_FILE_BUFFER))
1369 dirty |= I_DIRTY_DATASYNC;
1370 __cap_delay_requeue(mdsc, ci);
1371 return dirty;
1375 * Add dirty inode to the flushing list. Assigned a seq number so we
1376 * can wait for caps to flush without starving.
1378 * Called under i_lock.
1380 static int __mark_caps_flushing(struct inode *inode,
1381 struct ceph_mds_session *session)
1383 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1384 struct ceph_inode_info *ci = ceph_inode(inode);
1385 int flushing;
1387 BUG_ON(ci->i_dirty_caps == 0);
1388 BUG_ON(list_empty(&ci->i_dirty_item));
1390 flushing = ci->i_dirty_caps;
1391 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1392 ceph_cap_string(flushing),
1393 ceph_cap_string(ci->i_flushing_caps),
1394 ceph_cap_string(ci->i_flushing_caps | flushing));
1395 ci->i_flushing_caps |= flushing;
1396 ci->i_dirty_caps = 0;
1397 dout(" inode %p now !dirty\n", inode);
1399 spin_lock(&mdsc->cap_dirty_lock);
1400 list_del_init(&ci->i_dirty_item);
1402 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq;
1403 if (list_empty(&ci->i_flushing_item)) {
1404 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1405 mdsc->num_cap_flushing++;
1406 dout(" inode %p now flushing seq %lld\n", inode,
1407 ci->i_cap_flush_seq);
1408 } else {
1409 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1410 dout(" inode %p now flushing (more) seq %lld\n", inode,
1411 ci->i_cap_flush_seq);
1413 spin_unlock(&mdsc->cap_dirty_lock);
1415 return flushing;
1419 * try to invalidate mapping pages without blocking.
1421 static int try_nonblocking_invalidate(struct inode *inode)
1423 struct ceph_inode_info *ci = ceph_inode(inode);
1424 u32 invalidating_gen = ci->i_rdcache_gen;
1426 spin_unlock(&inode->i_lock);
1427 invalidate_mapping_pages(&inode->i_data, 0, -1);
1428 spin_lock(&inode->i_lock);
1430 if (inode->i_data.nrpages == 0 &&
1431 invalidating_gen == ci->i_rdcache_gen) {
1432 /* success. */
1433 dout("try_nonblocking_invalidate %p success\n", inode);
1434 /* save any racing async invalidate some trouble */
1435 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1436 return 0;
1438 dout("try_nonblocking_invalidate %p failed\n", inode);
1439 return -1;
1443 * Swiss army knife function to examine currently used and wanted
1444 * versus held caps. Release, flush, ack revoked caps to mds as
1445 * appropriate.
1447 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1448 * cap release further.
1449 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1450 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1451 * further delay.
1453 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1454 struct ceph_mds_session *session)
1456 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1457 struct ceph_mds_client *mdsc = fsc->mdsc;
1458 struct inode *inode = &ci->vfs_inode;
1459 struct ceph_cap *cap;
1460 int file_wanted, used;
1461 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1462 int issued, implemented, want, retain, revoking, flushing = 0;
1463 int mds = -1; /* keep track of how far we've gone through i_caps list
1464 to avoid an infinite loop on retry */
1465 struct rb_node *p;
1466 int tried_invalidate = 0;
1467 int delayed = 0, sent = 0, force_requeue = 0, num;
1468 int queue_invalidate = 0;
1469 int is_delayed = flags & CHECK_CAPS_NODELAY;
1471 /* if we are unmounting, flush any unused caps immediately. */
1472 if (mdsc->stopping)
1473 is_delayed = 1;
1475 spin_lock(&inode->i_lock);
1477 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1478 flags |= CHECK_CAPS_FLUSH;
1480 /* flush snaps first time around only */
1481 if (!list_empty(&ci->i_cap_snaps))
1482 __ceph_flush_snaps(ci, &session, 0);
1483 goto retry_locked;
1484 retry:
1485 spin_lock(&inode->i_lock);
1486 retry_locked:
1487 file_wanted = __ceph_caps_file_wanted(ci);
1488 used = __ceph_caps_used(ci);
1489 want = file_wanted | used;
1490 issued = __ceph_caps_issued(ci, &implemented);
1491 revoking = implemented & ~issued;
1493 retain = want | CEPH_CAP_PIN;
1494 if (!mdsc->stopping && inode->i_nlink > 0) {
1495 if (want) {
1496 retain |= CEPH_CAP_ANY; /* be greedy */
1497 } else {
1498 retain |= CEPH_CAP_ANY_SHARED;
1500 * keep RD only if we didn't have the file open RW,
1501 * because then the mds would revoke it anyway to
1502 * journal max_size=0.
1504 if (ci->i_max_size == 0)
1505 retain |= CEPH_CAP_ANY_RD;
1509 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1510 " issued %s revoking %s retain %s %s%s%s\n", inode,
1511 ceph_cap_string(file_wanted),
1512 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1513 ceph_cap_string(ci->i_flushing_caps),
1514 ceph_cap_string(issued), ceph_cap_string(revoking),
1515 ceph_cap_string(retain),
1516 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1517 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1518 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1521 * If we no longer need to hold onto old our caps, and we may
1522 * have cached pages, but don't want them, then try to invalidate.
1523 * If we fail, it's because pages are locked.... try again later.
1525 if ((!is_delayed || mdsc->stopping) &&
1526 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */
1527 inode->i_data.nrpages && /* have cached pages */
1528 (file_wanted == 0 || /* no open files */
1529 (revoking & (CEPH_CAP_FILE_CACHE|
1530 CEPH_CAP_FILE_LAZYIO))) && /* or revoking cache */
1531 !tried_invalidate) {
1532 dout("check_caps trying to invalidate on %p\n", inode);
1533 if (try_nonblocking_invalidate(inode) < 0) {
1534 if (revoking & (CEPH_CAP_FILE_CACHE|
1535 CEPH_CAP_FILE_LAZYIO)) {
1536 dout("check_caps queuing invalidate\n");
1537 queue_invalidate = 1;
1538 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1539 } else {
1540 dout("check_caps failed to invalidate pages\n");
1541 /* we failed to invalidate pages. check these
1542 caps again later. */
1543 force_requeue = 1;
1544 __cap_set_timeouts(mdsc, ci);
1547 tried_invalidate = 1;
1548 goto retry_locked;
1551 num = 0;
1552 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1553 cap = rb_entry(p, struct ceph_cap, ci_node);
1554 num++;
1556 /* avoid looping forever */
1557 if (mds >= cap->mds ||
1558 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1559 continue;
1561 /* NOTE: no side-effects allowed, until we take s_mutex */
1563 revoking = cap->implemented & ~cap->issued;
1564 dout(" mds%d cap %p issued %s implemented %s revoking %s\n",
1565 cap->mds, cap, ceph_cap_string(cap->issued),
1566 ceph_cap_string(cap->implemented),
1567 ceph_cap_string(revoking));
1569 if (cap == ci->i_auth_cap &&
1570 (cap->issued & CEPH_CAP_FILE_WR)) {
1571 /* request larger max_size from MDS? */
1572 if (ci->i_wanted_max_size > ci->i_max_size &&
1573 ci->i_wanted_max_size > ci->i_requested_max_size) {
1574 dout("requesting new max_size\n");
1575 goto ack;
1578 /* approaching file_max? */
1579 if ((inode->i_size << 1) >= ci->i_max_size &&
1580 (ci->i_reported_size << 1) < ci->i_max_size) {
1581 dout("i_size approaching max_size\n");
1582 goto ack;
1585 /* flush anything dirty? */
1586 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1587 ci->i_dirty_caps) {
1588 dout("flushing dirty caps\n");
1589 goto ack;
1592 /* completed revocation? going down and there are no caps? */
1593 if (revoking && (revoking & used) == 0) {
1594 dout("completed revocation of %s\n",
1595 ceph_cap_string(cap->implemented & ~cap->issued));
1596 goto ack;
1599 /* want more caps from mds? */
1600 if (want & ~(cap->mds_wanted | cap->issued))
1601 goto ack;
1603 /* things we might delay */
1604 if ((cap->issued & ~retain) == 0 &&
1605 cap->mds_wanted == want)
1606 continue; /* nope, all good */
1608 if (is_delayed)
1609 goto ack;
1611 /* delay? */
1612 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1613 time_before(jiffies, ci->i_hold_caps_max)) {
1614 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1615 ceph_cap_string(cap->issued),
1616 ceph_cap_string(cap->issued & retain),
1617 ceph_cap_string(cap->mds_wanted),
1618 ceph_cap_string(want));
1619 delayed++;
1620 continue;
1623 ack:
1624 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1625 dout(" skipping %p I_NOFLUSH set\n", inode);
1626 continue;
1629 if (session && session != cap->session) {
1630 dout("oops, wrong session %p mutex\n", session);
1631 mutex_unlock(&session->s_mutex);
1632 session = NULL;
1634 if (!session) {
1635 session = cap->session;
1636 if (mutex_trylock(&session->s_mutex) == 0) {
1637 dout("inverting session/ino locks on %p\n",
1638 session);
1639 spin_unlock(&inode->i_lock);
1640 if (took_snap_rwsem) {
1641 up_read(&mdsc->snap_rwsem);
1642 took_snap_rwsem = 0;
1644 mutex_lock(&session->s_mutex);
1645 goto retry;
1648 /* take snap_rwsem after session mutex */
1649 if (!took_snap_rwsem) {
1650 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1651 dout("inverting snap/in locks on %p\n",
1652 inode);
1653 spin_unlock(&inode->i_lock);
1654 down_read(&mdsc->snap_rwsem);
1655 took_snap_rwsem = 1;
1656 goto retry;
1658 took_snap_rwsem = 1;
1661 if (cap == ci->i_auth_cap && ci->i_dirty_caps)
1662 flushing = __mark_caps_flushing(inode, session);
1663 else
1664 flushing = 0;
1666 mds = cap->mds; /* remember mds, so we don't repeat */
1667 sent++;
1669 /* __send_cap drops i_lock */
1670 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, used, want,
1671 retain, flushing, NULL);
1672 goto retry; /* retake i_lock and restart our cap scan. */
1676 * Reschedule delayed caps release if we delayed anything,
1677 * otherwise cancel.
1679 if (delayed && is_delayed)
1680 force_requeue = 1; /* __send_cap delayed release; requeue */
1681 if (!delayed && !is_delayed)
1682 __cap_delay_cancel(mdsc, ci);
1683 else if (!is_delayed || force_requeue)
1684 __cap_delay_requeue(mdsc, ci);
1686 spin_unlock(&inode->i_lock);
1688 if (queue_invalidate)
1689 ceph_queue_invalidate(inode);
1691 if (session)
1692 mutex_unlock(&session->s_mutex);
1693 if (took_snap_rwsem)
1694 up_read(&mdsc->snap_rwsem);
1698 * Try to flush dirty caps back to the auth mds.
1700 static int try_flush_caps(struct inode *inode, struct ceph_mds_session *session,
1701 unsigned *flush_tid)
1703 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1704 struct ceph_inode_info *ci = ceph_inode(inode);
1705 int unlock_session = session ? 0 : 1;
1706 int flushing = 0;
1708 retry:
1709 spin_lock(&inode->i_lock);
1710 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1711 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1712 goto out;
1714 if (ci->i_dirty_caps && ci->i_auth_cap) {
1715 struct ceph_cap *cap = ci->i_auth_cap;
1716 int used = __ceph_caps_used(ci);
1717 int want = __ceph_caps_wanted(ci);
1718 int delayed;
1720 if (!session) {
1721 spin_unlock(&inode->i_lock);
1722 session = cap->session;
1723 mutex_lock(&session->s_mutex);
1724 goto retry;
1726 BUG_ON(session != cap->session);
1727 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1728 goto out;
1730 flushing = __mark_caps_flushing(inode, session);
1732 /* __send_cap drops i_lock */
1733 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1734 cap->issued | cap->implemented, flushing,
1735 flush_tid);
1736 if (!delayed)
1737 goto out_unlocked;
1739 spin_lock(&inode->i_lock);
1740 __cap_delay_requeue(mdsc, ci);
1742 out:
1743 spin_unlock(&inode->i_lock);
1744 out_unlocked:
1745 if (session && unlock_session)
1746 mutex_unlock(&session->s_mutex);
1747 return flushing;
1751 * Return true if we've flushed caps through the given flush_tid.
1753 static int caps_are_flushed(struct inode *inode, unsigned tid)
1755 struct ceph_inode_info *ci = ceph_inode(inode);
1756 int i, ret = 1;
1758 spin_lock(&inode->i_lock);
1759 for (i = 0; i < CEPH_CAP_BITS; i++)
1760 if ((ci->i_flushing_caps & (1 << i)) &&
1761 ci->i_cap_flush_tid[i] <= tid) {
1762 /* still flushing this bit */
1763 ret = 0;
1764 break;
1766 spin_unlock(&inode->i_lock);
1767 return ret;
1771 * Wait on any unsafe replies for the given inode. First wait on the
1772 * newest request, and make that the upper bound. Then, if there are
1773 * more requests, keep waiting on the oldest as long as it is still older
1774 * than the original request.
1776 static void sync_write_wait(struct inode *inode)
1778 struct ceph_inode_info *ci = ceph_inode(inode);
1779 struct list_head *head = &ci->i_unsafe_writes;
1780 struct ceph_osd_request *req;
1781 u64 last_tid;
1783 spin_lock(&ci->i_unsafe_lock);
1784 if (list_empty(head))
1785 goto out;
1787 /* set upper bound as _last_ entry in chain */
1788 req = list_entry(head->prev, struct ceph_osd_request,
1789 r_unsafe_item);
1790 last_tid = req->r_tid;
1792 do {
1793 ceph_osdc_get_request(req);
1794 spin_unlock(&ci->i_unsafe_lock);
1795 dout("sync_write_wait on tid %llu (until %llu)\n",
1796 req->r_tid, last_tid);
1797 wait_for_completion(&req->r_safe_completion);
1798 spin_lock(&ci->i_unsafe_lock);
1799 ceph_osdc_put_request(req);
1802 * from here on look at first entry in chain, since we
1803 * only want to wait for anything older than last_tid
1805 if (list_empty(head))
1806 break;
1807 req = list_entry(head->next, struct ceph_osd_request,
1808 r_unsafe_item);
1809 } while (req->r_tid < last_tid);
1810 out:
1811 spin_unlock(&ci->i_unsafe_lock);
1814 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1816 struct inode *inode = file->f_mapping->host;
1817 struct ceph_inode_info *ci = ceph_inode(inode);
1818 unsigned flush_tid;
1819 int ret;
1820 int dirty;
1822 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
1823 sync_write_wait(inode);
1825 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1826 if (ret < 0)
1827 return ret;
1828 mutex_lock(&inode->i_mutex);
1830 dirty = try_flush_caps(inode, NULL, &flush_tid);
1831 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1834 * only wait on non-file metadata writeback (the mds
1835 * can recover size and mtime, so we don't need to
1836 * wait for that)
1838 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1839 dout("fsync waiting for flush_tid %u\n", flush_tid);
1840 ret = wait_event_interruptible(ci->i_cap_wq,
1841 caps_are_flushed(inode, flush_tid));
1844 dout("fsync %p%s done\n", inode, datasync ? " datasync" : "");
1845 mutex_unlock(&inode->i_mutex);
1846 return ret;
1850 * Flush any dirty caps back to the mds. If we aren't asked to wait,
1851 * queue inode for flush but don't do so immediately, because we can
1852 * get by with fewer MDS messages if we wait for data writeback to
1853 * complete first.
1855 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
1857 struct ceph_inode_info *ci = ceph_inode(inode);
1858 unsigned flush_tid;
1859 int err = 0;
1860 int dirty;
1861 int wait = wbc->sync_mode == WB_SYNC_ALL;
1863 dout("write_inode %p wait=%d\n", inode, wait);
1864 if (wait) {
1865 dirty = try_flush_caps(inode, NULL, &flush_tid);
1866 if (dirty)
1867 err = wait_event_interruptible(ci->i_cap_wq,
1868 caps_are_flushed(inode, flush_tid));
1869 } else {
1870 struct ceph_mds_client *mdsc =
1871 ceph_sb_to_client(inode->i_sb)->mdsc;
1873 spin_lock(&inode->i_lock);
1874 if (__ceph_caps_dirty(ci))
1875 __cap_delay_requeue_front(mdsc, ci);
1876 spin_unlock(&inode->i_lock);
1878 return err;
1882 * After a recovering MDS goes active, we need to resend any caps
1883 * we were flushing.
1885 * Caller holds session->s_mutex.
1887 static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
1888 struct ceph_mds_session *session)
1890 struct ceph_cap_snap *capsnap;
1892 dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
1893 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
1894 flushing_item) {
1895 struct ceph_inode_info *ci = capsnap->ci;
1896 struct inode *inode = &ci->vfs_inode;
1897 struct ceph_cap *cap;
1899 spin_lock(&inode->i_lock);
1900 cap = ci->i_auth_cap;
1901 if (cap && cap->session == session) {
1902 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
1903 cap, capsnap);
1904 __ceph_flush_snaps(ci, &session, 1);
1905 } else {
1906 pr_err("%p auth cap %p not mds%d ???\n", inode,
1907 cap, session->s_mds);
1909 spin_unlock(&inode->i_lock);
1913 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
1914 struct ceph_mds_session *session)
1916 struct ceph_inode_info *ci;
1918 kick_flushing_capsnaps(mdsc, session);
1920 dout("kick_flushing_caps mds%d\n", session->s_mds);
1921 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
1922 struct inode *inode = &ci->vfs_inode;
1923 struct ceph_cap *cap;
1924 int delayed = 0;
1926 spin_lock(&inode->i_lock);
1927 cap = ci->i_auth_cap;
1928 if (cap && cap->session == session) {
1929 dout("kick_flushing_caps %p cap %p %s\n", inode,
1930 cap, ceph_cap_string(ci->i_flushing_caps));
1931 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1932 __ceph_caps_used(ci),
1933 __ceph_caps_wanted(ci),
1934 cap->issued | cap->implemented,
1935 ci->i_flushing_caps, NULL);
1936 if (delayed) {
1937 spin_lock(&inode->i_lock);
1938 __cap_delay_requeue(mdsc, ci);
1939 spin_unlock(&inode->i_lock);
1941 } else {
1942 pr_err("%p auth cap %p not mds%d ???\n", inode,
1943 cap, session->s_mds);
1944 spin_unlock(&inode->i_lock);
1949 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
1950 struct ceph_mds_session *session,
1951 struct inode *inode)
1953 struct ceph_inode_info *ci = ceph_inode(inode);
1954 struct ceph_cap *cap;
1955 int delayed = 0;
1957 spin_lock(&inode->i_lock);
1958 cap = ci->i_auth_cap;
1959 dout("kick_flushing_inode_caps %p flushing %s flush_seq %lld\n", inode,
1960 ceph_cap_string(ci->i_flushing_caps), ci->i_cap_flush_seq);
1961 __ceph_flush_snaps(ci, &session, 1);
1962 if (ci->i_flushing_caps) {
1963 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1964 __ceph_caps_used(ci),
1965 __ceph_caps_wanted(ci),
1966 cap->issued | cap->implemented,
1967 ci->i_flushing_caps, NULL);
1968 if (delayed) {
1969 spin_lock(&inode->i_lock);
1970 __cap_delay_requeue(mdsc, ci);
1971 spin_unlock(&inode->i_lock);
1973 } else {
1974 spin_unlock(&inode->i_lock);
1980 * Take references to capabilities we hold, so that we don't release
1981 * them to the MDS prematurely.
1983 * Protected by i_lock.
1985 static void __take_cap_refs(struct ceph_inode_info *ci, int got)
1987 if (got & CEPH_CAP_PIN)
1988 ci->i_pin_ref++;
1989 if (got & CEPH_CAP_FILE_RD)
1990 ci->i_rd_ref++;
1991 if (got & CEPH_CAP_FILE_CACHE)
1992 ci->i_rdcache_ref++;
1993 if (got & CEPH_CAP_FILE_WR)
1994 ci->i_wr_ref++;
1995 if (got & CEPH_CAP_FILE_BUFFER) {
1996 if (ci->i_wb_ref == 0)
1997 ihold(&ci->vfs_inode);
1998 ci->i_wb_ref++;
1999 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2000 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2005 * Try to grab cap references. Specify those refs we @want, and the
2006 * minimal set we @need. Also include the larger offset we are writing
2007 * to (when applicable), and check against max_size here as well.
2008 * Note that caller is responsible for ensuring max_size increases are
2009 * requested from the MDS.
2011 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2012 int *got, loff_t endoff, int *check_max, int *err)
2014 struct inode *inode = &ci->vfs_inode;
2015 int ret = 0;
2016 int have, implemented;
2017 int file_wanted;
2019 dout("get_cap_refs %p need %s want %s\n", inode,
2020 ceph_cap_string(need), ceph_cap_string(want));
2021 spin_lock(&inode->i_lock);
2023 /* make sure file is actually open */
2024 file_wanted = __ceph_caps_file_wanted(ci);
2025 if ((file_wanted & need) == 0) {
2026 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2027 ceph_cap_string(need), ceph_cap_string(file_wanted));
2028 *err = -EBADF;
2029 ret = 1;
2030 goto out;
2033 if (need & CEPH_CAP_FILE_WR) {
2034 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2035 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2036 inode, endoff, ci->i_max_size);
2037 if (endoff > ci->i_wanted_max_size) {
2038 *check_max = 1;
2039 ret = 1;
2041 goto out;
2044 * If a sync write is in progress, we must wait, so that we
2045 * can get a final snapshot value for size+mtime.
2047 if (__ceph_have_pending_cap_snap(ci)) {
2048 dout("get_cap_refs %p cap_snap_pending\n", inode);
2049 goto out;
2052 have = __ceph_caps_issued(ci, &implemented);
2055 * disallow writes while a truncate is pending
2057 if (ci->i_truncate_pending)
2058 have &= ~CEPH_CAP_FILE_WR;
2060 if ((have & need) == need) {
2062 * Look at (implemented & ~have & not) so that we keep waiting
2063 * on transition from wanted -> needed caps. This is needed
2064 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2065 * going before a prior buffered writeback happens.
2067 int not = want & ~(have & need);
2068 int revoking = implemented & ~have;
2069 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2070 inode, ceph_cap_string(have), ceph_cap_string(not),
2071 ceph_cap_string(revoking));
2072 if ((revoking & not) == 0) {
2073 *got = need | (have & want);
2074 __take_cap_refs(ci, *got);
2075 ret = 1;
2077 } else {
2078 dout("get_cap_refs %p have %s needed %s\n", inode,
2079 ceph_cap_string(have), ceph_cap_string(need));
2081 out:
2082 spin_unlock(&inode->i_lock);
2083 dout("get_cap_refs %p ret %d got %s\n", inode,
2084 ret, ceph_cap_string(*got));
2085 return ret;
2089 * Check the offset we are writing up to against our current
2090 * max_size. If necessary, tell the MDS we want to write to
2091 * a larger offset.
2093 static void check_max_size(struct inode *inode, loff_t endoff)
2095 struct ceph_inode_info *ci = ceph_inode(inode);
2096 int check = 0;
2098 /* do we need to explicitly request a larger max_size? */
2099 spin_lock(&inode->i_lock);
2100 if ((endoff >= ci->i_max_size ||
2101 endoff > (inode->i_size << 1)) &&
2102 endoff > ci->i_wanted_max_size) {
2103 dout("write %p at large endoff %llu, req max_size\n",
2104 inode, endoff);
2105 ci->i_wanted_max_size = endoff;
2106 check = 1;
2108 spin_unlock(&inode->i_lock);
2109 if (check)
2110 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2114 * Wait for caps, and take cap references. If we can't get a WR cap
2115 * due to a small max_size, make sure we check_max_size (and possibly
2116 * ask the mds) so we don't get hung up indefinitely.
2118 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, int *got,
2119 loff_t endoff)
2121 int check_max, ret, err;
2123 retry:
2124 if (endoff > 0)
2125 check_max_size(&ci->vfs_inode, endoff);
2126 check_max = 0;
2127 err = 0;
2128 ret = wait_event_interruptible(ci->i_cap_wq,
2129 try_get_cap_refs(ci, need, want,
2130 got, endoff,
2131 &check_max, &err));
2132 if (err)
2133 ret = err;
2134 if (check_max)
2135 goto retry;
2136 return ret;
2140 * Take cap refs. Caller must already know we hold at least one ref
2141 * on the caps in question or we don't know this is safe.
2143 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2145 spin_lock(&ci->vfs_inode.i_lock);
2146 __take_cap_refs(ci, caps);
2147 spin_unlock(&ci->vfs_inode.i_lock);
2151 * Release cap refs.
2153 * If we released the last ref on any given cap, call ceph_check_caps
2154 * to release (or schedule a release).
2156 * If we are releasing a WR cap (from a sync write), finalize any affected
2157 * cap_snap, and wake up any waiters.
2159 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2161 struct inode *inode = &ci->vfs_inode;
2162 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2163 struct ceph_cap_snap *capsnap;
2165 spin_lock(&inode->i_lock);
2166 if (had & CEPH_CAP_PIN)
2167 --ci->i_pin_ref;
2168 if (had & CEPH_CAP_FILE_RD)
2169 if (--ci->i_rd_ref == 0)
2170 last++;
2171 if (had & CEPH_CAP_FILE_CACHE)
2172 if (--ci->i_rdcache_ref == 0)
2173 last++;
2174 if (had & CEPH_CAP_FILE_BUFFER) {
2175 if (--ci->i_wb_ref == 0) {
2176 last++;
2177 put++;
2179 dout("put_cap_refs %p wb %d -> %d (?)\n",
2180 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2182 if (had & CEPH_CAP_FILE_WR)
2183 if (--ci->i_wr_ref == 0) {
2184 last++;
2185 if (!list_empty(&ci->i_cap_snaps)) {
2186 capsnap = list_first_entry(&ci->i_cap_snaps,
2187 struct ceph_cap_snap,
2188 ci_item);
2189 if (capsnap->writing) {
2190 capsnap->writing = 0;
2191 flushsnaps =
2192 __ceph_finish_cap_snap(ci,
2193 capsnap);
2194 wake = 1;
2198 spin_unlock(&inode->i_lock);
2200 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2201 last ? " last" : "", put ? " put" : "");
2203 if (last && !flushsnaps)
2204 ceph_check_caps(ci, 0, NULL);
2205 else if (flushsnaps)
2206 ceph_flush_snaps(ci);
2207 if (wake)
2208 wake_up_all(&ci->i_cap_wq);
2209 if (put)
2210 iput(inode);
2214 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2215 * context. Adjust per-snap dirty page accounting as appropriate.
2216 * Once all dirty data for a cap_snap is flushed, flush snapped file
2217 * metadata back to the MDS. If we dropped the last ref, call
2218 * ceph_check_caps.
2220 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2221 struct ceph_snap_context *snapc)
2223 struct inode *inode = &ci->vfs_inode;
2224 int last = 0;
2225 int complete_capsnap = 0;
2226 int drop_capsnap = 0;
2227 int found = 0;
2228 struct ceph_cap_snap *capsnap = NULL;
2230 spin_lock(&inode->i_lock);
2231 ci->i_wrbuffer_ref -= nr;
2232 last = !ci->i_wrbuffer_ref;
2234 if (ci->i_head_snapc == snapc) {
2235 ci->i_wrbuffer_ref_head -= nr;
2236 if (ci->i_wrbuffer_ref_head == 0 &&
2237 ci->i_dirty_caps == 0 && ci->i_flushing_caps == 0) {
2238 BUG_ON(!ci->i_head_snapc);
2239 ceph_put_snap_context(ci->i_head_snapc);
2240 ci->i_head_snapc = NULL;
2242 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2243 inode,
2244 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2245 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2246 last ? " LAST" : "");
2247 } else {
2248 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2249 if (capsnap->context == snapc) {
2250 found = 1;
2251 break;
2254 BUG_ON(!found);
2255 capsnap->dirty_pages -= nr;
2256 if (capsnap->dirty_pages == 0) {
2257 complete_capsnap = 1;
2258 if (capsnap->dirty == 0)
2259 /* cap writeback completed before we created
2260 * the cap_snap; no FLUSHSNAP is needed */
2261 drop_capsnap = 1;
2263 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2264 " snap %lld %d/%d -> %d/%d %s%s%s\n",
2265 inode, capsnap, capsnap->context->seq,
2266 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2267 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2268 last ? " (wrbuffer last)" : "",
2269 complete_capsnap ? " (complete capsnap)" : "",
2270 drop_capsnap ? " (drop capsnap)" : "");
2271 if (drop_capsnap) {
2272 ceph_put_snap_context(capsnap->context);
2273 list_del(&capsnap->ci_item);
2274 list_del(&capsnap->flushing_item);
2275 ceph_put_cap_snap(capsnap);
2279 spin_unlock(&inode->i_lock);
2281 if (last) {
2282 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2283 iput(inode);
2284 } else if (complete_capsnap) {
2285 ceph_flush_snaps(ci);
2286 wake_up_all(&ci->i_cap_wq);
2288 if (drop_capsnap)
2289 iput(inode);
2293 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2294 * actually be a revocation if it specifies a smaller cap set.)
2296 * caller holds s_mutex and i_lock, we drop both.
2298 * return value:
2299 * 0 - ok
2300 * 1 - check_caps on auth cap only (writeback)
2301 * 2 - check_caps (ack revoke)
2303 static void handle_cap_grant(struct inode *inode, struct ceph_mds_caps *grant,
2304 struct ceph_mds_session *session,
2305 struct ceph_cap *cap,
2306 struct ceph_buffer *xattr_buf)
2307 __releases(inode->i_lock)
2309 struct ceph_inode_info *ci = ceph_inode(inode);
2310 int mds = session->s_mds;
2311 int seq = le32_to_cpu(grant->seq);
2312 int newcaps = le32_to_cpu(grant->caps);
2313 int issued, implemented, used, wanted, dirty;
2314 u64 size = le64_to_cpu(grant->size);
2315 u64 max_size = le64_to_cpu(grant->max_size);
2316 struct timespec mtime, atime, ctime;
2317 int check_caps = 0;
2318 int wake = 0;
2319 int writeback = 0;
2320 int revoked_rdcache = 0;
2321 int queue_invalidate = 0;
2323 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2324 inode, cap, mds, seq, ceph_cap_string(newcaps));
2325 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2326 inode->i_size);
2329 * If CACHE is being revoked, and we have no dirty buffers,
2330 * try to invalidate (once). (If there are dirty buffers, we
2331 * will invalidate _after_ writeback.)
2333 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2334 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2335 !ci->i_wrbuffer_ref) {
2336 if (try_nonblocking_invalidate(inode) == 0) {
2337 revoked_rdcache = 1;
2338 } else {
2339 /* there were locked pages.. invalidate later
2340 in a separate thread. */
2341 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2342 queue_invalidate = 1;
2343 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2348 /* side effects now are allowed */
2350 issued = __ceph_caps_issued(ci, &implemented);
2351 issued |= implemented | __ceph_caps_dirty(ci);
2353 cap->cap_gen = session->s_cap_gen;
2355 __check_cap_issue(ci, cap, newcaps);
2357 if ((issued & CEPH_CAP_AUTH_EXCL) == 0) {
2358 inode->i_mode = le32_to_cpu(grant->mode);
2359 inode->i_uid = le32_to_cpu(grant->uid);
2360 inode->i_gid = le32_to_cpu(grant->gid);
2361 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2362 inode->i_uid, inode->i_gid);
2365 if ((issued & CEPH_CAP_LINK_EXCL) == 0)
2366 inode->i_nlink = le32_to_cpu(grant->nlink);
2368 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2369 int len = le32_to_cpu(grant->xattr_len);
2370 u64 version = le64_to_cpu(grant->xattr_version);
2372 if (version > ci->i_xattrs.version) {
2373 dout(" got new xattrs v%llu on %p len %d\n",
2374 version, inode, len);
2375 if (ci->i_xattrs.blob)
2376 ceph_buffer_put(ci->i_xattrs.blob);
2377 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2378 ci->i_xattrs.version = version;
2382 /* size/ctime/mtime/atime? */
2383 ceph_fill_file_size(inode, issued,
2384 le32_to_cpu(grant->truncate_seq),
2385 le64_to_cpu(grant->truncate_size), size);
2386 ceph_decode_timespec(&mtime, &grant->mtime);
2387 ceph_decode_timespec(&atime, &grant->atime);
2388 ceph_decode_timespec(&ctime, &grant->ctime);
2389 ceph_fill_file_time(inode, issued,
2390 le32_to_cpu(grant->time_warp_seq), &ctime, &mtime,
2391 &atime);
2393 /* max size increase? */
2394 if (max_size != ci->i_max_size) {
2395 dout("max_size %lld -> %llu\n", ci->i_max_size, max_size);
2396 ci->i_max_size = max_size;
2397 if (max_size >= ci->i_wanted_max_size) {
2398 ci->i_wanted_max_size = 0; /* reset */
2399 ci->i_requested_max_size = 0;
2401 wake = 1;
2404 /* check cap bits */
2405 wanted = __ceph_caps_wanted(ci);
2406 used = __ceph_caps_used(ci);
2407 dirty = __ceph_caps_dirty(ci);
2408 dout(" my wanted = %s, used = %s, dirty %s\n",
2409 ceph_cap_string(wanted),
2410 ceph_cap_string(used),
2411 ceph_cap_string(dirty));
2412 if (wanted != le32_to_cpu(grant->wanted)) {
2413 dout("mds wanted %s -> %s\n",
2414 ceph_cap_string(le32_to_cpu(grant->wanted)),
2415 ceph_cap_string(wanted));
2416 grant->wanted = cpu_to_le32(wanted);
2419 cap->seq = seq;
2421 /* file layout may have changed */
2422 ci->i_layout = grant->layout;
2424 /* revocation, grant, or no-op? */
2425 if (cap->issued & ~newcaps) {
2426 int revoking = cap->issued & ~newcaps;
2428 dout("revocation: %s -> %s (revoking %s)\n",
2429 ceph_cap_string(cap->issued),
2430 ceph_cap_string(newcaps),
2431 ceph_cap_string(revoking));
2432 if (revoking & used & CEPH_CAP_FILE_BUFFER)
2433 writeback = 1; /* initiate writeback; will delay ack */
2434 else if (revoking == CEPH_CAP_FILE_CACHE &&
2435 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2436 queue_invalidate)
2437 ; /* do nothing yet, invalidation will be queued */
2438 else if (cap == ci->i_auth_cap)
2439 check_caps = 1; /* check auth cap only */
2440 else
2441 check_caps = 2; /* check all caps */
2442 cap->issued = newcaps;
2443 cap->implemented |= newcaps;
2444 } else if (cap->issued == newcaps) {
2445 dout("caps unchanged: %s -> %s\n",
2446 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2447 } else {
2448 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2449 ceph_cap_string(newcaps));
2450 cap->issued = newcaps;
2451 cap->implemented |= newcaps; /* add bits only, to
2452 * avoid stepping on a
2453 * pending revocation */
2454 wake = 1;
2456 BUG_ON(cap->issued & ~cap->implemented);
2458 spin_unlock(&inode->i_lock);
2459 if (writeback)
2461 * queue inode for writeback: we can't actually call
2462 * filemap_write_and_wait, etc. from message handler
2463 * context.
2465 ceph_queue_writeback(inode);
2466 if (queue_invalidate)
2467 ceph_queue_invalidate(inode);
2468 if (wake)
2469 wake_up_all(&ci->i_cap_wq);
2471 if (check_caps == 1)
2472 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2473 session);
2474 else if (check_caps == 2)
2475 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
2476 else
2477 mutex_unlock(&session->s_mutex);
2481 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2482 * MDS has been safely committed.
2484 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2485 struct ceph_mds_caps *m,
2486 struct ceph_mds_session *session,
2487 struct ceph_cap *cap)
2488 __releases(inode->i_lock)
2490 struct ceph_inode_info *ci = ceph_inode(inode);
2491 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2492 unsigned seq = le32_to_cpu(m->seq);
2493 int dirty = le32_to_cpu(m->dirty);
2494 int cleaned = 0;
2495 int drop = 0;
2496 int i;
2498 for (i = 0; i < CEPH_CAP_BITS; i++)
2499 if ((dirty & (1 << i)) &&
2500 flush_tid == ci->i_cap_flush_tid[i])
2501 cleaned |= 1 << i;
2503 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2504 " flushing %s -> %s\n",
2505 inode, session->s_mds, seq, ceph_cap_string(dirty),
2506 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2507 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2509 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned))
2510 goto out;
2512 ci->i_flushing_caps &= ~cleaned;
2514 spin_lock(&mdsc->cap_dirty_lock);
2515 if (ci->i_flushing_caps == 0) {
2516 list_del_init(&ci->i_flushing_item);
2517 if (!list_empty(&session->s_cap_flushing))
2518 dout(" mds%d still flushing cap on %p\n",
2519 session->s_mds,
2520 &list_entry(session->s_cap_flushing.next,
2521 struct ceph_inode_info,
2522 i_flushing_item)->vfs_inode);
2523 mdsc->num_cap_flushing--;
2524 wake_up_all(&mdsc->cap_flushing_wq);
2525 dout(" inode %p now !flushing\n", inode);
2527 if (ci->i_dirty_caps == 0) {
2528 dout(" inode %p now clean\n", inode);
2529 BUG_ON(!list_empty(&ci->i_dirty_item));
2530 drop = 1;
2531 if (ci->i_wrbuffer_ref_head == 0) {
2532 BUG_ON(!ci->i_head_snapc);
2533 ceph_put_snap_context(ci->i_head_snapc);
2534 ci->i_head_snapc = NULL;
2536 } else {
2537 BUG_ON(list_empty(&ci->i_dirty_item));
2540 spin_unlock(&mdsc->cap_dirty_lock);
2541 wake_up_all(&ci->i_cap_wq);
2543 out:
2544 spin_unlock(&inode->i_lock);
2545 if (drop)
2546 iput(inode);
2550 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
2551 * throw away our cap_snap.
2553 * Caller hold s_mutex.
2555 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
2556 struct ceph_mds_caps *m,
2557 struct ceph_mds_session *session)
2559 struct ceph_inode_info *ci = ceph_inode(inode);
2560 u64 follows = le64_to_cpu(m->snap_follows);
2561 struct ceph_cap_snap *capsnap;
2562 int drop = 0;
2564 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
2565 inode, ci, session->s_mds, follows);
2567 spin_lock(&inode->i_lock);
2568 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2569 if (capsnap->follows == follows) {
2570 if (capsnap->flush_tid != flush_tid) {
2571 dout(" cap_snap %p follows %lld tid %lld !="
2572 " %lld\n", capsnap, follows,
2573 flush_tid, capsnap->flush_tid);
2574 break;
2576 WARN_ON(capsnap->dirty_pages || capsnap->writing);
2577 dout(" removing %p cap_snap %p follows %lld\n",
2578 inode, capsnap, follows);
2579 ceph_put_snap_context(capsnap->context);
2580 list_del(&capsnap->ci_item);
2581 list_del(&capsnap->flushing_item);
2582 ceph_put_cap_snap(capsnap);
2583 drop = 1;
2584 break;
2585 } else {
2586 dout(" skipping cap_snap %p follows %lld\n",
2587 capsnap, capsnap->follows);
2590 spin_unlock(&inode->i_lock);
2591 if (drop)
2592 iput(inode);
2596 * Handle TRUNC from MDS, indicating file truncation.
2598 * caller hold s_mutex.
2600 static void handle_cap_trunc(struct inode *inode,
2601 struct ceph_mds_caps *trunc,
2602 struct ceph_mds_session *session)
2603 __releases(inode->i_lock)
2605 struct ceph_inode_info *ci = ceph_inode(inode);
2606 int mds = session->s_mds;
2607 int seq = le32_to_cpu(trunc->seq);
2608 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
2609 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
2610 u64 size = le64_to_cpu(trunc->size);
2611 int implemented = 0;
2612 int dirty = __ceph_caps_dirty(ci);
2613 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
2614 int queue_trunc = 0;
2616 issued |= implemented | dirty;
2618 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
2619 inode, mds, seq, truncate_size, truncate_seq);
2620 queue_trunc = ceph_fill_file_size(inode, issued,
2621 truncate_seq, truncate_size, size);
2622 spin_unlock(&inode->i_lock);
2624 if (queue_trunc)
2625 ceph_queue_vmtruncate(inode);
2629 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
2630 * different one. If we are the most recent migration we've seen (as
2631 * indicated by mseq), make note of the migrating cap bits for the
2632 * duration (until we see the corresponding IMPORT).
2634 * caller holds s_mutex
2636 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
2637 struct ceph_mds_session *session,
2638 int *open_target_sessions)
2640 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2641 struct ceph_inode_info *ci = ceph_inode(inode);
2642 int mds = session->s_mds;
2643 unsigned mseq = le32_to_cpu(ex->migrate_seq);
2644 struct ceph_cap *cap = NULL, *t;
2645 struct rb_node *p;
2646 int remember = 1;
2648 dout("handle_cap_export inode %p ci %p mds%d mseq %d\n",
2649 inode, ci, mds, mseq);
2651 spin_lock(&inode->i_lock);
2653 /* make sure we haven't seen a higher mseq */
2654 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2655 t = rb_entry(p, struct ceph_cap, ci_node);
2656 if (ceph_seq_cmp(t->mseq, mseq) > 0) {
2657 dout(" higher mseq on cap from mds%d\n",
2658 t->session->s_mds);
2659 remember = 0;
2661 if (t->session->s_mds == mds)
2662 cap = t;
2665 if (cap) {
2666 if (remember) {
2667 /* make note */
2668 ci->i_cap_exporting_mds = mds;
2669 ci->i_cap_exporting_mseq = mseq;
2670 ci->i_cap_exporting_issued = cap->issued;
2673 * make sure we have open sessions with all possible
2674 * export targets, so that we get the matching IMPORT
2676 *open_target_sessions = 1;
2679 * we can't flush dirty caps that we've seen the
2680 * EXPORT but no IMPORT for
2682 spin_lock(&mdsc->cap_dirty_lock);
2683 if (!list_empty(&ci->i_dirty_item)) {
2684 dout(" moving %p to cap_dirty_migrating\n",
2685 inode);
2686 list_move(&ci->i_dirty_item,
2687 &mdsc->cap_dirty_migrating);
2689 spin_unlock(&mdsc->cap_dirty_lock);
2691 __ceph_remove_cap(cap);
2693 /* else, we already released it */
2695 spin_unlock(&inode->i_lock);
2699 * Handle cap IMPORT. If there are temp bits from an older EXPORT,
2700 * clean them up.
2702 * caller holds s_mutex.
2704 static void handle_cap_import(struct ceph_mds_client *mdsc,
2705 struct inode *inode, struct ceph_mds_caps *im,
2706 struct ceph_mds_session *session,
2707 void *snaptrace, int snaptrace_len)
2709 struct ceph_inode_info *ci = ceph_inode(inode);
2710 int mds = session->s_mds;
2711 unsigned issued = le32_to_cpu(im->caps);
2712 unsigned wanted = le32_to_cpu(im->wanted);
2713 unsigned seq = le32_to_cpu(im->seq);
2714 unsigned mseq = le32_to_cpu(im->migrate_seq);
2715 u64 realmino = le64_to_cpu(im->realm);
2716 u64 cap_id = le64_to_cpu(im->cap_id);
2718 if (ci->i_cap_exporting_mds >= 0 &&
2719 ceph_seq_cmp(ci->i_cap_exporting_mseq, mseq) < 0) {
2720 dout("handle_cap_import inode %p ci %p mds%d mseq %d"
2721 " - cleared exporting from mds%d\n",
2722 inode, ci, mds, mseq,
2723 ci->i_cap_exporting_mds);
2724 ci->i_cap_exporting_issued = 0;
2725 ci->i_cap_exporting_mseq = 0;
2726 ci->i_cap_exporting_mds = -1;
2728 spin_lock(&mdsc->cap_dirty_lock);
2729 if (!list_empty(&ci->i_dirty_item)) {
2730 dout(" moving %p back to cap_dirty\n", inode);
2731 list_move(&ci->i_dirty_item, &mdsc->cap_dirty);
2733 spin_unlock(&mdsc->cap_dirty_lock);
2734 } else {
2735 dout("handle_cap_import inode %p ci %p mds%d mseq %d\n",
2736 inode, ci, mds, mseq);
2739 down_write(&mdsc->snap_rwsem);
2740 ceph_update_snap_trace(mdsc, snaptrace, snaptrace+snaptrace_len,
2741 false);
2742 downgrade_write(&mdsc->snap_rwsem);
2743 ceph_add_cap(inode, session, cap_id, -1,
2744 issued, wanted, seq, mseq, realmino, CEPH_CAP_FLAG_AUTH,
2745 NULL /* no caps context */);
2746 kick_flushing_inode_caps(mdsc, session, inode);
2747 up_read(&mdsc->snap_rwsem);
2749 /* make sure we re-request max_size, if necessary */
2750 spin_lock(&inode->i_lock);
2751 ci->i_requested_max_size = 0;
2752 spin_unlock(&inode->i_lock);
2756 * Handle a caps message from the MDS.
2758 * Identify the appropriate session, inode, and call the right handler
2759 * based on the cap op.
2761 void ceph_handle_caps(struct ceph_mds_session *session,
2762 struct ceph_msg *msg)
2764 struct ceph_mds_client *mdsc = session->s_mdsc;
2765 struct super_block *sb = mdsc->fsc->sb;
2766 struct inode *inode;
2767 struct ceph_cap *cap;
2768 struct ceph_mds_caps *h;
2769 int mds = session->s_mds;
2770 int op;
2771 u32 seq, mseq;
2772 struct ceph_vino vino;
2773 u64 cap_id;
2774 u64 size, max_size;
2775 u64 tid;
2776 void *snaptrace;
2777 size_t snaptrace_len;
2778 void *flock;
2779 u32 flock_len;
2780 int open_target_sessions = 0;
2782 dout("handle_caps from mds%d\n", mds);
2784 /* decode */
2785 tid = le64_to_cpu(msg->hdr.tid);
2786 if (msg->front.iov_len < sizeof(*h))
2787 goto bad;
2788 h = msg->front.iov_base;
2789 op = le32_to_cpu(h->op);
2790 vino.ino = le64_to_cpu(h->ino);
2791 vino.snap = CEPH_NOSNAP;
2792 cap_id = le64_to_cpu(h->cap_id);
2793 seq = le32_to_cpu(h->seq);
2794 mseq = le32_to_cpu(h->migrate_seq);
2795 size = le64_to_cpu(h->size);
2796 max_size = le64_to_cpu(h->max_size);
2798 snaptrace = h + 1;
2799 snaptrace_len = le32_to_cpu(h->snap_trace_len);
2801 if (le16_to_cpu(msg->hdr.version) >= 2) {
2802 void *p, *end;
2804 p = snaptrace + snaptrace_len;
2805 end = msg->front.iov_base + msg->front.iov_len;
2806 ceph_decode_32_safe(&p, end, flock_len, bad);
2807 flock = p;
2808 } else {
2809 flock = NULL;
2810 flock_len = 0;
2813 mutex_lock(&session->s_mutex);
2814 session->s_seq++;
2815 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
2816 (unsigned)seq);
2818 /* lookup ino */
2819 inode = ceph_find_inode(sb, vino);
2820 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
2821 vino.snap, inode);
2822 if (!inode) {
2823 dout(" i don't have ino %llx\n", vino.ino);
2825 if (op == CEPH_CAP_OP_IMPORT)
2826 __queue_cap_release(session, vino.ino, cap_id,
2827 mseq, seq);
2828 goto flush_cap_releases;
2831 /* these will work even if we don't have a cap yet */
2832 switch (op) {
2833 case CEPH_CAP_OP_FLUSHSNAP_ACK:
2834 handle_cap_flushsnap_ack(inode, tid, h, session);
2835 goto done;
2837 case CEPH_CAP_OP_EXPORT:
2838 handle_cap_export(inode, h, session, &open_target_sessions);
2839 goto done;
2841 case CEPH_CAP_OP_IMPORT:
2842 handle_cap_import(mdsc, inode, h, session,
2843 snaptrace, snaptrace_len);
2844 ceph_check_caps(ceph_inode(inode), 0, session);
2845 goto done_unlocked;
2848 /* the rest require a cap */
2849 spin_lock(&inode->i_lock);
2850 cap = __get_cap_for_mds(ceph_inode(inode), mds);
2851 if (!cap) {
2852 dout(" no cap on %p ino %llx.%llx from mds%d\n",
2853 inode, ceph_ino(inode), ceph_snap(inode), mds);
2854 spin_unlock(&inode->i_lock);
2855 goto flush_cap_releases;
2858 /* note that each of these drops i_lock for us */
2859 switch (op) {
2860 case CEPH_CAP_OP_REVOKE:
2861 case CEPH_CAP_OP_GRANT:
2862 handle_cap_grant(inode, h, session, cap, msg->middle);
2863 goto done_unlocked;
2865 case CEPH_CAP_OP_FLUSH_ACK:
2866 handle_cap_flush_ack(inode, tid, h, session, cap);
2867 break;
2869 case CEPH_CAP_OP_TRUNC:
2870 handle_cap_trunc(inode, h, session);
2871 break;
2873 default:
2874 spin_unlock(&inode->i_lock);
2875 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
2876 ceph_cap_op_name(op));
2879 goto done;
2881 flush_cap_releases:
2883 * send any full release message to try to move things
2884 * along for the mds (who clearly thinks we still have this
2885 * cap).
2887 ceph_add_cap_releases(mdsc, session);
2888 ceph_send_cap_releases(mdsc, session);
2890 done:
2891 mutex_unlock(&session->s_mutex);
2892 done_unlocked:
2893 if (inode)
2894 iput(inode);
2895 if (open_target_sessions)
2896 ceph_mdsc_open_export_target_sessions(mdsc, session);
2897 return;
2899 bad:
2900 pr_err("ceph_handle_caps: corrupt message\n");
2901 ceph_msg_dump(msg);
2902 return;
2906 * Delayed work handler to process end of delayed cap release LRU list.
2908 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
2910 struct ceph_inode_info *ci;
2911 int flags = CHECK_CAPS_NODELAY;
2913 dout("check_delayed_caps\n");
2914 while (1) {
2915 spin_lock(&mdsc->cap_delay_lock);
2916 if (list_empty(&mdsc->cap_delay_list))
2917 break;
2918 ci = list_first_entry(&mdsc->cap_delay_list,
2919 struct ceph_inode_info,
2920 i_cap_delay_list);
2921 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
2922 time_before(jiffies, ci->i_hold_caps_max))
2923 break;
2924 list_del_init(&ci->i_cap_delay_list);
2925 spin_unlock(&mdsc->cap_delay_lock);
2926 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
2927 ceph_check_caps(ci, flags, NULL);
2929 spin_unlock(&mdsc->cap_delay_lock);
2933 * Flush all dirty caps to the mds
2935 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
2937 struct ceph_inode_info *ci;
2938 struct inode *inode;
2940 dout("flush_dirty_caps\n");
2941 spin_lock(&mdsc->cap_dirty_lock);
2942 while (!list_empty(&mdsc->cap_dirty)) {
2943 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
2944 i_dirty_item);
2945 inode = &ci->vfs_inode;
2946 ihold(inode);
2947 dout("flush_dirty_caps %p\n", inode);
2948 spin_unlock(&mdsc->cap_dirty_lock);
2949 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
2950 iput(inode);
2951 spin_lock(&mdsc->cap_dirty_lock);
2953 spin_unlock(&mdsc->cap_dirty_lock);
2954 dout("flush_dirty_caps done\n");
2958 * Drop open file reference. If we were the last open file,
2959 * we may need to release capabilities to the MDS (or schedule
2960 * their delayed release).
2962 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
2964 struct inode *inode = &ci->vfs_inode;
2965 int last = 0;
2967 spin_lock(&inode->i_lock);
2968 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
2969 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
2970 BUG_ON(ci->i_nr_by_mode[fmode] == 0);
2971 if (--ci->i_nr_by_mode[fmode] == 0)
2972 last++;
2973 spin_unlock(&inode->i_lock);
2975 if (last && ci->i_vino.snap == CEPH_NOSNAP)
2976 ceph_check_caps(ci, 0, NULL);
2980 * Helpers for embedding cap and dentry lease releases into mds
2981 * requests.
2983 * @force is used by dentry_release (below) to force inclusion of a
2984 * record for the directory inode, even when there aren't any caps to
2985 * drop.
2987 int ceph_encode_inode_release(void **p, struct inode *inode,
2988 int mds, int drop, int unless, int force)
2990 struct ceph_inode_info *ci = ceph_inode(inode);
2991 struct ceph_cap *cap;
2992 struct ceph_mds_request_release *rel = *p;
2993 int used, dirty;
2994 int ret = 0;
2996 spin_lock(&inode->i_lock);
2997 used = __ceph_caps_used(ci);
2998 dirty = __ceph_caps_dirty(ci);
3000 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3001 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3002 ceph_cap_string(unless));
3004 /* only drop unused, clean caps */
3005 drop &= ~(used | dirty);
3007 cap = __get_cap_for_mds(ci, mds);
3008 if (cap && __cap_is_valid(cap)) {
3009 if (force ||
3010 ((cap->issued & drop) &&
3011 (cap->issued & unless) == 0)) {
3012 if ((cap->issued & drop) &&
3013 (cap->issued & unless) == 0) {
3014 dout("encode_inode_release %p cap %p %s -> "
3015 "%s\n", inode, cap,
3016 ceph_cap_string(cap->issued),
3017 ceph_cap_string(cap->issued & ~drop));
3018 cap->issued &= ~drop;
3019 cap->implemented &= ~drop;
3020 if (ci->i_ceph_flags & CEPH_I_NODELAY) {
3021 int wanted = __ceph_caps_wanted(ci);
3022 dout(" wanted %s -> %s (act %s)\n",
3023 ceph_cap_string(cap->mds_wanted),
3024 ceph_cap_string(cap->mds_wanted &
3025 ~wanted),
3026 ceph_cap_string(wanted));
3027 cap->mds_wanted &= wanted;
3029 } else {
3030 dout("encode_inode_release %p cap %p %s"
3031 " (force)\n", inode, cap,
3032 ceph_cap_string(cap->issued));
3035 rel->ino = cpu_to_le64(ceph_ino(inode));
3036 rel->cap_id = cpu_to_le64(cap->cap_id);
3037 rel->seq = cpu_to_le32(cap->seq);
3038 rel->issue_seq = cpu_to_le32(cap->issue_seq),
3039 rel->mseq = cpu_to_le32(cap->mseq);
3040 rel->caps = cpu_to_le32(cap->issued);
3041 rel->wanted = cpu_to_le32(cap->mds_wanted);
3042 rel->dname_len = 0;
3043 rel->dname_seq = 0;
3044 *p += sizeof(*rel);
3045 ret = 1;
3046 } else {
3047 dout("encode_inode_release %p cap %p %s\n",
3048 inode, cap, ceph_cap_string(cap->issued));
3051 spin_unlock(&inode->i_lock);
3052 return ret;
3055 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3056 int mds, int drop, int unless)
3058 struct inode *dir = dentry->d_parent->d_inode;
3059 struct ceph_mds_request_release *rel = *p;
3060 struct ceph_dentry_info *di = ceph_dentry(dentry);
3061 int force = 0;
3062 int ret;
3065 * force an record for the directory caps if we have a dentry lease.
3066 * this is racy (can't take i_lock and d_lock together), but it
3067 * doesn't have to be perfect; the mds will revoke anything we don't
3068 * release.
3070 spin_lock(&dentry->d_lock);
3071 if (di->lease_session && di->lease_session->s_mds == mds)
3072 force = 1;
3073 spin_unlock(&dentry->d_lock);
3075 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3077 spin_lock(&dentry->d_lock);
3078 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3079 dout("encode_dentry_release %p mds%d seq %d\n",
3080 dentry, mds, (int)di->lease_seq);
3081 rel->dname_len = cpu_to_le32(dentry->d_name.len);
3082 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3083 *p += dentry->d_name.len;
3084 rel->dname_seq = cpu_to_le32(di->lease_seq);
3085 __ceph_mdsc_drop_dentry_lease(dentry);
3087 spin_unlock(&dentry->d_lock);
3088 return ret;