1 #include "ceph_debug.h"
4 #include <linux/kernel.h>
5 #include <linux/sched.h>
6 #include <linux/vmalloc.h>
7 #include <linux/wait.h>
11 #include "messenger.h"
14 * Capability management
16 * The Ceph metadata servers control client access to inode metadata
17 * and file data by issuing capabilities, granting clients permission
18 * to read and/or write both inode field and file data to OSDs
19 * (storage nodes). Each capability consists of a set of bits
20 * indicating which operations are allowed.
22 * If the client holds a *_SHARED cap, the client has a coherent value
23 * that can be safely read from the cached inode.
25 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
26 * client is allowed to change inode attributes (e.g., file size,
27 * mtime), note its dirty state in the ceph_cap, and asynchronously
28 * flush that metadata change to the MDS.
30 * In the event of a conflicting operation (perhaps by another
31 * client), the MDS will revoke the conflicting client capabilities.
33 * In order for a client to cache an inode, it must hold a capability
34 * with at least one MDS server. When inodes are released, release
35 * notifications are batched and periodically sent en masse to the MDS
36 * cluster to release server state.
41 * Generate readable cap strings for debugging output.
43 #define MAX_CAP_STR 20
44 static char cap_str[MAX_CAP_STR][40];
45 static DEFINE_SPINLOCK(cap_str_lock);
46 static int last_cap_str;
48 static char *gcap_string(char *s, int c)
50 if (c & CEPH_CAP_GSHARED)
52 if (c & CEPH_CAP_GEXCL)
54 if (c & CEPH_CAP_GCACHE)
60 if (c & CEPH_CAP_GBUFFER)
62 if (c & CEPH_CAP_GLAZYIO)
67 const char *ceph_cap_string(int caps)
73 spin_lock(&cap_str_lock);
75 if (last_cap_str == MAX_CAP_STR)
77 spin_unlock(&cap_str_lock);
81 if (caps & CEPH_CAP_PIN)
84 c = (caps >> CEPH_CAP_SAUTH) & 3;
87 s = gcap_string(s, c);
90 c = (caps >> CEPH_CAP_SLINK) & 3;
93 s = gcap_string(s, c);
96 c = (caps >> CEPH_CAP_SXATTR) & 3;
99 s = gcap_string(s, c);
102 c = caps >> CEPH_CAP_SFILE;
105 s = gcap_string(s, c);
117 * Maintain a global pool of preallocated struct ceph_caps, referenced
118 * by struct ceph_caps_reservations. This ensures that we preallocate
119 * memory needed to successfully process an MDS response. (If an MDS
120 * sends us cap information and we fail to process it, we will have
121 * problems due to the client and MDS being out of sync.)
123 * Reservations are 'owned' by a ceph_cap_reservation context.
125 static spinlock_t caps_list_lock;
126 static struct list_head caps_list; /* unused (reserved or unreserved) */
127 static int caps_total_count; /* total caps allocated */
128 static int caps_use_count; /* in use */
129 static int caps_reserve_count; /* unused, reserved */
130 static int caps_avail_count; /* unused, unreserved */
132 void __init ceph_caps_init(void)
134 INIT_LIST_HEAD(&caps_list);
135 spin_lock_init(&caps_list_lock);
138 void ceph_caps_finalize(void)
140 struct ceph_cap *cap;
142 spin_lock(&caps_list_lock);
143 while (!list_empty(&caps_list)) {
144 cap = list_first_entry(&caps_list, struct ceph_cap, caps_item);
145 list_del(&cap->caps_item);
146 kmem_cache_free(ceph_cap_cachep, cap);
148 caps_total_count = 0;
149 caps_avail_count = 0;
151 caps_reserve_count = 0;
152 spin_unlock(&caps_list_lock);
155 int ceph_reserve_caps(struct ceph_cap_reservation *ctx, int need)
158 struct ceph_cap *cap;
164 dout("reserve caps ctx=%p need=%d\n", ctx, need);
166 /* first reserve any caps that are already allocated */
167 spin_lock(&caps_list_lock);
168 if (caps_avail_count >= need)
171 have = caps_avail_count;
172 caps_avail_count -= have;
173 caps_reserve_count += have;
174 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
176 spin_unlock(&caps_list_lock);
178 for (i = have; i < need; i++) {
179 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
182 goto out_alloc_count;
184 list_add(&cap->caps_item, &newcaps);
187 BUG_ON(have + alloc != need);
189 spin_lock(&caps_list_lock);
190 caps_total_count += alloc;
191 caps_reserve_count += alloc;
192 list_splice(&newcaps, &caps_list);
194 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
196 spin_unlock(&caps_list_lock);
199 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
200 ctx, caps_total_count, caps_use_count, caps_reserve_count,
205 /* we didn't manage to reserve as much as we needed */
206 pr_warning("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
211 int ceph_unreserve_caps(struct ceph_cap_reservation *ctx)
213 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
215 spin_lock(&caps_list_lock);
216 BUG_ON(caps_reserve_count < ctx->count);
217 caps_reserve_count -= ctx->count;
218 caps_avail_count += ctx->count;
220 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
221 caps_total_count, caps_use_count, caps_reserve_count,
223 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
225 spin_unlock(&caps_list_lock);
230 static struct ceph_cap *get_cap(struct ceph_cap_reservation *ctx)
232 struct ceph_cap *cap = NULL;
234 /* temporary, until we do something about cap import/export */
236 return kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
238 spin_lock(&caps_list_lock);
239 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
240 ctx, ctx->count, caps_total_count, caps_use_count,
241 caps_reserve_count, caps_avail_count);
243 BUG_ON(ctx->count > caps_reserve_count);
244 BUG_ON(list_empty(&caps_list));
247 caps_reserve_count--;
250 cap = list_first_entry(&caps_list, struct ceph_cap, caps_item);
251 list_del(&cap->caps_item);
253 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
255 spin_unlock(&caps_list_lock);
259 static void put_cap(struct ceph_cap *cap,
260 struct ceph_cap_reservation *ctx)
262 spin_lock(&caps_list_lock);
263 dout("put_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
264 ctx, ctx ? ctx->count : 0, caps_total_count, caps_use_count,
265 caps_reserve_count, caps_avail_count);
268 * Keep some preallocated caps around, at least enough to do a
269 * readdir (which needs to preallocate lots of them), to avoid
270 * lots of free/alloc churn.
272 if (caps_avail_count >= caps_reserve_count +
273 ceph_client(cap->ci->vfs_inode.i_sb)->mount_args->max_readdir) {
275 kmem_cache_free(ceph_cap_cachep, cap);
279 caps_reserve_count++;
283 list_add(&cap->caps_item, &caps_list);
286 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
288 spin_unlock(&caps_list_lock);
291 void ceph_reservation_status(struct ceph_client *client,
292 int *total, int *avail, int *used, int *reserved)
295 *total = caps_total_count;
297 *avail = caps_avail_count;
299 *used = caps_use_count;
301 *reserved = caps_reserve_count;
305 * Find ceph_cap for given mds, if any.
307 * Called with i_lock held.
309 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
311 struct ceph_cap *cap;
312 struct rb_node *n = ci->i_caps.rb_node;
315 cap = rb_entry(n, struct ceph_cap, ci_node);
318 else if (mds > cap->mds)
327 * Return id of any MDS with a cap, preferably FILE_WR|WRBUFFER|EXCL, else
330 static int __ceph_get_cap_mds(struct ceph_inode_info *ci, u32 *mseq)
332 struct ceph_cap *cap;
336 /* prefer mds with WR|WRBUFFER|EXCL caps */
337 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
338 cap = rb_entry(p, struct ceph_cap, ci_node);
342 if (cap->issued & (CEPH_CAP_FILE_WR |
343 CEPH_CAP_FILE_BUFFER |
350 int ceph_get_cap_mds(struct inode *inode)
353 spin_lock(&inode->i_lock);
354 mds = __ceph_get_cap_mds(ceph_inode(inode), NULL);
355 spin_unlock(&inode->i_lock);
360 * Called under i_lock.
362 static void __insert_cap_node(struct ceph_inode_info *ci,
363 struct ceph_cap *new)
365 struct rb_node **p = &ci->i_caps.rb_node;
366 struct rb_node *parent = NULL;
367 struct ceph_cap *cap = NULL;
371 cap = rb_entry(parent, struct ceph_cap, ci_node);
372 if (new->mds < cap->mds)
374 else if (new->mds > cap->mds)
380 rb_link_node(&new->ci_node, parent, p);
381 rb_insert_color(&new->ci_node, &ci->i_caps);
385 * (re)set cap hold timeouts, which control the delayed release
386 * of unused caps back to the MDS. Should be called on cap use.
388 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
389 struct ceph_inode_info *ci)
391 struct ceph_mount_args *ma = mdsc->client->mount_args;
393 ci->i_hold_caps_min = round_jiffies(jiffies +
394 ma->caps_wanted_delay_min * HZ);
395 ci->i_hold_caps_max = round_jiffies(jiffies +
396 ma->caps_wanted_delay_max * HZ);
397 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
398 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
402 * (Re)queue cap at the end of the delayed cap release list.
404 * If I_FLUSH is set, leave the inode at the front of the list.
406 * Caller holds i_lock
407 * -> we take mdsc->cap_delay_lock
409 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
410 struct ceph_inode_info *ci)
412 __cap_set_timeouts(mdsc, ci);
413 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
414 ci->i_ceph_flags, ci->i_hold_caps_max);
415 if (!mdsc->stopping) {
416 spin_lock(&mdsc->cap_delay_lock);
417 if (!list_empty(&ci->i_cap_delay_list)) {
418 if (ci->i_ceph_flags & CEPH_I_FLUSH)
420 list_del_init(&ci->i_cap_delay_list);
422 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
424 spin_unlock(&mdsc->cap_delay_lock);
429 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
430 * indicating we should send a cap message to flush dirty metadata
431 * asap, and move to the front of the delayed cap list.
433 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
434 struct ceph_inode_info *ci)
436 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
437 spin_lock(&mdsc->cap_delay_lock);
438 ci->i_ceph_flags |= CEPH_I_FLUSH;
439 if (!list_empty(&ci->i_cap_delay_list))
440 list_del_init(&ci->i_cap_delay_list);
441 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
442 spin_unlock(&mdsc->cap_delay_lock);
446 * Cancel delayed work on cap.
448 * Caller must hold i_lock.
450 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
451 struct ceph_inode_info *ci)
453 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
454 if (list_empty(&ci->i_cap_delay_list))
456 spin_lock(&mdsc->cap_delay_lock);
457 list_del_init(&ci->i_cap_delay_list);
458 spin_unlock(&mdsc->cap_delay_lock);
462 * Common issue checks for add_cap, handle_cap_grant.
464 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
467 unsigned had = __ceph_caps_issued(ci, NULL);
470 * Each time we receive FILE_CACHE anew, we increment
473 if ((issued & CEPH_CAP_FILE_CACHE) &&
474 (had & CEPH_CAP_FILE_CACHE) == 0)
478 * if we are newly issued FILE_SHARED, clear I_COMPLETE; we
479 * don't know what happened to this directory while we didn't
482 if ((issued & CEPH_CAP_FILE_SHARED) &&
483 (had & CEPH_CAP_FILE_SHARED) == 0) {
485 if (S_ISDIR(ci->vfs_inode.i_mode)) {
486 dout(" marking %p NOT complete\n", &ci->vfs_inode);
487 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
493 * Add a capability under the given MDS session.
495 * Caller should hold session snap_rwsem (read) and s_mutex.
497 * @fmode is the open file mode, if we are opening a file, otherwise
498 * it is < 0. (This is so we can atomically add the cap and add an
499 * open file reference to it.)
501 int ceph_add_cap(struct inode *inode,
502 struct ceph_mds_session *session, u64 cap_id,
503 int fmode, unsigned issued, unsigned wanted,
504 unsigned seq, unsigned mseq, u64 realmino, int flags,
505 struct ceph_cap_reservation *caps_reservation)
507 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
508 struct ceph_inode_info *ci = ceph_inode(inode);
509 struct ceph_cap *new_cap = NULL;
510 struct ceph_cap *cap;
511 int mds = session->s_mds;
514 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
515 session->s_mds, cap_id, ceph_cap_string(issued), seq);
518 * If we are opening the file, include file mode wanted bits
522 wanted |= ceph_caps_for_mode(fmode);
525 spin_lock(&inode->i_lock);
526 cap = __get_cap_for_mds(ci, mds);
532 spin_unlock(&inode->i_lock);
533 new_cap = get_cap(caps_reservation);
540 cap->implemented = 0;
545 __insert_cap_node(ci, cap);
547 /* clear out old exporting info? (i.e. on cap import) */
548 if (ci->i_cap_exporting_mds == mds) {
549 ci->i_cap_exporting_issued = 0;
550 ci->i_cap_exporting_mseq = 0;
551 ci->i_cap_exporting_mds = -1;
554 /* add to session cap list */
555 cap->session = session;
556 spin_lock(&session->s_cap_lock);
557 list_add_tail(&cap->session_caps, &session->s_caps);
558 session->s_nr_caps++;
559 spin_unlock(&session->s_cap_lock);
562 if (!ci->i_snap_realm) {
564 * add this inode to the appropriate snap realm
566 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
569 ceph_get_snap_realm(mdsc, realm);
570 spin_lock(&realm->inodes_with_caps_lock);
571 ci->i_snap_realm = realm;
572 list_add(&ci->i_snap_realm_item,
573 &realm->inodes_with_caps);
574 spin_unlock(&realm->inodes_with_caps_lock);
576 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
581 __check_cap_issue(ci, cap, issued);
584 * If we are issued caps we don't want, or the mds' wanted
585 * value appears to be off, queue a check so we'll release
586 * later and/or update the mds wanted value.
588 actual_wanted = __ceph_caps_wanted(ci);
589 if ((wanted & ~actual_wanted) ||
590 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
591 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
592 ceph_cap_string(issued), ceph_cap_string(wanted),
593 ceph_cap_string(actual_wanted));
594 __cap_delay_requeue(mdsc, ci);
597 if (flags & CEPH_CAP_FLAG_AUTH)
598 ci->i_auth_cap = cap;
599 else if (ci->i_auth_cap == cap)
600 ci->i_auth_cap = NULL;
602 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
603 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
604 ceph_cap_string(issued|cap->issued), seq, mds);
605 cap->cap_id = cap_id;
606 cap->issued = issued;
607 cap->implemented |= issued;
608 cap->mds_wanted |= wanted;
610 cap->issue_seq = seq;
612 cap->cap_gen = session->s_cap_gen;
615 __ceph_get_fmode(ci, fmode);
616 spin_unlock(&inode->i_lock);
617 wake_up(&ci->i_cap_wq);
622 * Return true if cap has not timed out and belongs to the current
623 * generation of the MDS session (i.e. has not gone 'stale' due to
624 * us losing touch with the mds).
626 static int __cap_is_valid(struct ceph_cap *cap)
631 spin_lock(&cap->session->s_cap_lock);
632 gen = cap->session->s_cap_gen;
633 ttl = cap->session->s_cap_ttl;
634 spin_unlock(&cap->session->s_cap_lock);
636 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
637 dout("__cap_is_valid %p cap %p issued %s "
638 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
639 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
647 * Return set of valid cap bits issued to us. Note that caps time
648 * out, and may be invalidated in bulk if the client session times out
649 * and session->s_cap_gen is bumped.
651 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
653 int have = ci->i_snap_caps;
654 struct ceph_cap *cap;
659 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
660 cap = rb_entry(p, struct ceph_cap, ci_node);
661 if (!__cap_is_valid(cap))
663 dout("__ceph_caps_issued %p cap %p issued %s\n",
664 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
667 *implemented |= cap->implemented;
673 * Get cap bits issued by caps other than @ocap
675 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
677 int have = ci->i_snap_caps;
678 struct ceph_cap *cap;
681 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
682 cap = rb_entry(p, struct ceph_cap, ci_node);
685 if (!__cap_is_valid(cap))
693 * Move a cap to the end of the LRU (oldest caps at list head, newest
696 static void __touch_cap(struct ceph_cap *cap)
698 struct ceph_mds_session *s = cap->session;
700 spin_lock(&s->s_cap_lock);
701 if (!s->s_iterating_caps) {
702 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
704 list_move_tail(&cap->session_caps, &s->s_caps);
706 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
707 &cap->ci->vfs_inode, cap, s->s_mds);
709 spin_unlock(&s->s_cap_lock);
713 * Check if we hold the given mask. If so, move the cap(s) to the
714 * front of their respective LRUs. (This is the preferred way for
715 * callers to check for caps they want.)
717 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
719 struct ceph_cap *cap;
721 int have = ci->i_snap_caps;
723 if ((have & mask) == mask) {
724 dout("__ceph_caps_issued_mask %p snap issued %s"
725 " (mask %s)\n", &ci->vfs_inode,
726 ceph_cap_string(have),
727 ceph_cap_string(mask));
731 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
732 cap = rb_entry(p, struct ceph_cap, ci_node);
733 if (!__cap_is_valid(cap))
735 if ((cap->issued & mask) == mask) {
736 dout("__ceph_caps_issued_mask %p cap %p issued %s"
737 " (mask %s)\n", &ci->vfs_inode, cap,
738 ceph_cap_string(cap->issued),
739 ceph_cap_string(mask));
745 /* does a combination of caps satisfy mask? */
747 if ((have & mask) == mask) {
748 dout("__ceph_caps_issued_mask %p combo issued %s"
749 " (mask %s)\n", &ci->vfs_inode,
750 ceph_cap_string(cap->issued),
751 ceph_cap_string(mask));
755 /* touch this + preceeding caps */
757 for (q = rb_first(&ci->i_caps); q != p;
759 cap = rb_entry(q, struct ceph_cap,
761 if (!__cap_is_valid(cap))
774 * Return true if mask caps are currently being revoked by an MDS.
776 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
778 struct inode *inode = &ci->vfs_inode;
779 struct ceph_cap *cap;
783 spin_lock(&inode->i_lock);
784 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
785 cap = rb_entry(p, struct ceph_cap, ci_node);
786 if (__cap_is_valid(cap) &&
787 (cap->implemented & ~cap->issued & mask)) {
792 spin_unlock(&inode->i_lock);
793 dout("ceph_caps_revoking %p %s = %d\n", inode,
794 ceph_cap_string(mask), ret);
798 int __ceph_caps_used(struct ceph_inode_info *ci)
802 used |= CEPH_CAP_PIN;
804 used |= CEPH_CAP_FILE_RD;
805 if (ci->i_rdcache_ref || ci->i_rdcache_gen)
806 used |= CEPH_CAP_FILE_CACHE;
808 used |= CEPH_CAP_FILE_WR;
809 if (ci->i_wrbuffer_ref)
810 used |= CEPH_CAP_FILE_BUFFER;
815 * wanted, by virtue of open file modes
817 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
821 for (mode = 0; mode < 4; mode++)
822 if (ci->i_nr_by_mode[mode])
823 want |= ceph_caps_for_mode(mode);
828 * Return caps we have registered with the MDS(s) as 'wanted'.
830 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
832 struct ceph_cap *cap;
836 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
837 cap = rb_entry(p, struct ceph_cap, ci_node);
838 if (!__cap_is_valid(cap))
840 mds_wanted |= cap->mds_wanted;
846 * called under i_lock
848 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
850 return !RB_EMPTY_ROOT(&ci->i_caps) || ci->i_cap_exporting_mds >= 0;
854 * caller should hold i_lock, and session s_mutex.
855 * returns true if this is the last cap. if so, caller should iput.
857 void __ceph_remove_cap(struct ceph_cap *cap,
858 struct ceph_cap_reservation *ctx)
860 struct ceph_mds_session *session = cap->session;
861 struct ceph_inode_info *ci = cap->ci;
862 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc;
864 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
866 /* remove from session list */
867 spin_lock(&session->s_cap_lock);
868 list_del_init(&cap->session_caps);
869 session->s_nr_caps--;
870 spin_unlock(&session->s_cap_lock);
872 /* remove from inode list */
873 rb_erase(&cap->ci_node, &ci->i_caps);
875 if (ci->i_auth_cap == cap)
876 ci->i_auth_cap = NULL;
880 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) {
881 struct ceph_snap_realm *realm = ci->i_snap_realm;
882 spin_lock(&realm->inodes_with_caps_lock);
883 list_del_init(&ci->i_snap_realm_item);
884 ci->i_snap_realm_counter++;
885 ci->i_snap_realm = NULL;
886 spin_unlock(&realm->inodes_with_caps_lock);
887 ceph_put_snap_realm(mdsc, realm);
889 if (!__ceph_is_any_real_caps(ci))
890 __cap_delay_cancel(mdsc, ci);
894 * Build and send a cap message to the given MDS.
896 * Caller should be holding s_mutex.
898 static int send_cap_msg(struct ceph_mds_session *session,
899 u64 ino, u64 cid, int op,
900 int caps, int wanted, int dirty,
901 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq,
902 u64 size, u64 max_size,
903 struct timespec *mtime, struct timespec *atime,
905 uid_t uid, gid_t gid, mode_t mode,
907 struct ceph_buffer *xattrs_buf,
910 struct ceph_mds_caps *fc;
911 struct ceph_msg *msg;
913 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
914 " seq %u/%u mseq %u follows %lld size %llu/%llu"
915 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
916 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
917 ceph_cap_string(dirty),
918 seq, issue_seq, mseq, follows, size, max_size,
919 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
921 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc), 0, 0, NULL);
925 msg->hdr.tid = cpu_to_le64(flush_tid);
927 fc = msg->front.iov_base;
928 memset(fc, 0, sizeof(*fc));
930 fc->cap_id = cpu_to_le64(cid);
931 fc->op = cpu_to_le32(op);
932 fc->seq = cpu_to_le32(seq);
933 fc->issue_seq = cpu_to_le32(issue_seq);
934 fc->migrate_seq = cpu_to_le32(mseq);
935 fc->caps = cpu_to_le32(caps);
936 fc->wanted = cpu_to_le32(wanted);
937 fc->dirty = cpu_to_le32(dirty);
938 fc->ino = cpu_to_le64(ino);
939 fc->snap_follows = cpu_to_le64(follows);
941 fc->size = cpu_to_le64(size);
942 fc->max_size = cpu_to_le64(max_size);
944 ceph_encode_timespec(&fc->mtime, mtime);
946 ceph_encode_timespec(&fc->atime, atime);
947 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
949 fc->uid = cpu_to_le32(uid);
950 fc->gid = cpu_to_le32(gid);
951 fc->mode = cpu_to_le32(mode);
953 fc->xattr_version = cpu_to_le64(xattr_version);
955 msg->middle = ceph_buffer_get(xattrs_buf);
956 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
957 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
960 ceph_con_send(&session->s_con, msg);
965 * Queue cap releases when an inode is dropped from our
968 void ceph_queue_caps_release(struct inode *inode)
970 struct ceph_inode_info *ci = ceph_inode(inode);
973 spin_lock(&inode->i_lock);
974 p = rb_first(&ci->i_caps);
976 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
977 struct ceph_mds_session *session = cap->session;
978 struct ceph_msg *msg;
979 struct ceph_mds_cap_release *head;
980 struct ceph_mds_cap_item *item;
982 spin_lock(&session->s_cap_lock);
983 BUG_ON(!session->s_num_cap_releases);
984 msg = list_first_entry(&session->s_cap_releases,
985 struct ceph_msg, list_head);
987 dout(" adding %p release to mds%d msg %p (%d left)\n",
988 inode, session->s_mds, msg, session->s_num_cap_releases);
990 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE);
991 head = msg->front.iov_base;
992 head->num = cpu_to_le32(le32_to_cpu(head->num) + 1);
993 item = msg->front.iov_base + msg->front.iov_len;
994 item->ino = cpu_to_le64(ceph_ino(inode));
995 item->cap_id = cpu_to_le64(cap->cap_id);
996 item->migrate_seq = cpu_to_le32(cap->mseq);
997 item->seq = cpu_to_le32(cap->issue_seq);
999 session->s_num_cap_releases--;
1001 msg->front.iov_len += sizeof(*item);
1002 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1003 dout(" release msg %p full\n", msg);
1004 list_move_tail(&msg->list_head,
1005 &session->s_cap_releases_done);
1007 dout(" release msg %p at %d/%d (%d)\n", msg,
1008 (int)le32_to_cpu(head->num),
1009 (int)CEPH_CAPS_PER_RELEASE,
1010 (int)msg->front.iov_len);
1012 spin_unlock(&session->s_cap_lock);
1014 __ceph_remove_cap(cap, NULL);
1017 spin_unlock(&inode->i_lock);
1021 * Send a cap msg on the given inode. Update our caps state, then
1022 * drop i_lock and send the message.
1024 * Make note of max_size reported/requested from mds, revoked caps
1025 * that have now been implemented.
1027 * Make half-hearted attempt ot to invalidate page cache if we are
1028 * dropping RDCACHE. Note that this will leave behind locked pages
1029 * that we'll then need to deal with elsewhere.
1031 * Return non-zero if delayed release, or we experienced an error
1032 * such that the caller should requeue + retry later.
1034 * called with i_lock, then drops it.
1035 * caller should hold snap_rwsem (read), s_mutex.
1037 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1038 int op, int used, int want, int retain, int flushing,
1039 unsigned *pflush_tid)
1040 __releases(cap->ci->vfs_inode->i_lock)
1042 struct ceph_inode_info *ci = cap->ci;
1043 struct inode *inode = &ci->vfs_inode;
1044 u64 cap_id = cap->cap_id;
1045 int held = cap->issued | cap->implemented;
1046 int revoking = cap->implemented & ~cap->issued;
1047 int dropping = cap->issued & ~retain;
1049 u64 seq, issue_seq, mseq, time_warp_seq, follows;
1051 struct timespec mtime, atime;
1056 struct ceph_mds_session *session;
1057 u64 xattr_version = 0;
1063 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1064 inode, cap, cap->session,
1065 ceph_cap_string(held), ceph_cap_string(held & retain),
1066 ceph_cap_string(revoking));
1067 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1069 session = cap->session;
1071 /* don't release wanted unless we've waited a bit. */
1072 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1073 time_before(jiffies, ci->i_hold_caps_min)) {
1074 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1075 ceph_cap_string(cap->issued),
1076 ceph_cap_string(cap->issued & retain),
1077 ceph_cap_string(cap->mds_wanted),
1078 ceph_cap_string(want));
1079 want |= cap->mds_wanted;
1080 retain |= cap->issued;
1083 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1085 cap->issued &= retain; /* drop bits we don't want */
1086 if (cap->implemented & ~cap->issued) {
1088 * Wake up any waiters on wanted -> needed transition.
1089 * This is due to the weird transition from buffered
1090 * to sync IO... we need to flush dirty pages _before_
1091 * allowing sync writes to avoid reordering.
1095 cap->implemented &= cap->issued | used;
1096 cap->mds_wanted = want;
1100 * assign a tid for flush operations so we can avoid
1101 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark
1102 * clean type races. track latest tid for every bit
1103 * so we can handle flush AxFw, flush Fw, and have the
1104 * first ack clean Ax.
1106 flush_tid = ++ci->i_cap_flush_last_tid;
1108 *pflush_tid = flush_tid;
1109 dout(" cap_flush_tid %d\n", (int)flush_tid);
1110 for (i = 0; i < CEPH_CAP_BITS; i++)
1111 if (flushing & (1 << i))
1112 ci->i_cap_flush_tid[i] = flush_tid;
1115 keep = cap->implemented;
1117 issue_seq = cap->issue_seq;
1119 size = inode->i_size;
1120 ci->i_reported_size = size;
1121 max_size = ci->i_wanted_max_size;
1122 ci->i_requested_max_size = max_size;
1123 mtime = inode->i_mtime;
1124 atime = inode->i_atime;
1125 time_warp_seq = ci->i_time_warp_seq;
1126 follows = ci->i_snap_realm->cached_context->seq;
1129 mode = inode->i_mode;
1131 if (dropping & CEPH_CAP_XATTR_EXCL) {
1132 __ceph_build_xattrs_blob(ci);
1133 xattr_version = ci->i_xattrs.version + 1;
1136 spin_unlock(&inode->i_lock);
1138 if (dropping & CEPH_CAP_FILE_CACHE) {
1139 /* invalidate what we can */
1140 dout("invalidating pages on %p\n", inode);
1141 invalidate_mapping_pages(&inode->i_data, 0, -1);
1144 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1145 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq,
1146 size, max_size, &mtime, &atime, time_warp_seq,
1149 (flushing & CEPH_CAP_XATTR_EXCL) ? ci->i_xattrs.blob : NULL,
1152 dout("error sending cap msg, must requeue %p\n", inode);
1157 wake_up(&ci->i_cap_wq);
1163 * When a snapshot is taken, clients accumulate dirty metadata on
1164 * inodes with capabilities in ceph_cap_snaps to describe the file
1165 * state at the time the snapshot was taken. This must be flushed
1166 * asynchronously back to the MDS once sync writes complete and dirty
1167 * data is written out.
1169 * Called under i_lock. Takes s_mutex as needed.
1171 void __ceph_flush_snaps(struct ceph_inode_info *ci,
1172 struct ceph_mds_session **psession)
1174 struct inode *inode = &ci->vfs_inode;
1176 struct ceph_cap_snap *capsnap;
1178 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
1179 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1181 u64 next_follows = 0; /* keep track of how far we've gotten through the
1182 i_cap_snaps list, and skip these entries next time
1183 around to avoid an infinite loop */
1186 session = *psession;
1188 dout("__flush_snaps %p\n", inode);
1190 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1191 /* avoid an infiniute loop after retry */
1192 if (capsnap->follows < next_follows)
1195 * we need to wait for sync writes to complete and for dirty
1196 * pages to be written out.
1198 if (capsnap->dirty_pages || capsnap->writing)
1201 /* pick mds, take s_mutex */
1202 mds = __ceph_get_cap_mds(ci, &mseq);
1203 if (session && session->s_mds != mds) {
1204 dout("oops, wrong session %p mutex\n", session);
1205 mutex_unlock(&session->s_mutex);
1206 ceph_put_mds_session(session);
1210 spin_unlock(&inode->i_lock);
1211 mutex_lock(&mdsc->mutex);
1212 session = __ceph_lookup_mds_session(mdsc, mds);
1213 mutex_unlock(&mdsc->mutex);
1215 dout("inverting session/ino locks on %p\n",
1217 mutex_lock(&session->s_mutex);
1220 * if session == NULL, we raced against a cap
1221 * deletion. retry, and we'll get a better
1222 * @mds value next time.
1224 spin_lock(&inode->i_lock);
1228 capsnap->flush_tid = ++ci->i_cap_flush_last_tid;
1229 atomic_inc(&capsnap->nref);
1230 if (!list_empty(&capsnap->flushing_item))
1231 list_del_init(&capsnap->flushing_item);
1232 list_add_tail(&capsnap->flushing_item,
1233 &session->s_cap_snaps_flushing);
1234 spin_unlock(&inode->i_lock);
1236 dout("flush_snaps %p cap_snap %p follows %lld size %llu\n",
1237 inode, capsnap, next_follows, capsnap->size);
1238 send_cap_msg(session, ceph_vino(inode).ino, 0,
1239 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1240 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq,
1242 &capsnap->mtime, &capsnap->atime,
1243 capsnap->time_warp_seq,
1244 capsnap->uid, capsnap->gid, capsnap->mode,
1248 next_follows = capsnap->follows + 1;
1249 ceph_put_cap_snap(capsnap);
1251 spin_lock(&inode->i_lock);
1255 /* we flushed them all; remove this inode from the queue */
1256 spin_lock(&mdsc->snap_flush_lock);
1257 list_del_init(&ci->i_snap_flush_item);
1258 spin_unlock(&mdsc->snap_flush_lock);
1261 *psession = session;
1263 mutex_unlock(&session->s_mutex);
1264 ceph_put_mds_session(session);
1268 static void ceph_flush_snaps(struct ceph_inode_info *ci)
1270 struct inode *inode = &ci->vfs_inode;
1272 spin_lock(&inode->i_lock);
1273 __ceph_flush_snaps(ci, NULL);
1274 spin_unlock(&inode->i_lock);
1278 * Mark caps dirty. If inode is newly dirty, add to the global dirty
1281 void __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1283 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc;
1284 struct inode *inode = &ci->vfs_inode;
1285 int was = ci->i_dirty_caps;
1288 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1289 ceph_cap_string(mask), ceph_cap_string(was),
1290 ceph_cap_string(was | mask));
1291 ci->i_dirty_caps |= mask;
1293 dout(" inode %p now dirty\n", &ci->vfs_inode);
1294 BUG_ON(!list_empty(&ci->i_dirty_item));
1295 spin_lock(&mdsc->cap_dirty_lock);
1296 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1297 spin_unlock(&mdsc->cap_dirty_lock);
1298 if (ci->i_flushing_caps == 0) {
1300 dirty |= I_DIRTY_SYNC;
1303 BUG_ON(list_empty(&ci->i_dirty_item));
1304 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1305 (mask & CEPH_CAP_FILE_BUFFER))
1306 dirty |= I_DIRTY_DATASYNC;
1308 __mark_inode_dirty(inode, dirty);
1309 __cap_delay_requeue(mdsc, ci);
1313 * Add dirty inode to the flushing list. Assigned a seq number so we
1314 * can wait for caps to flush without starving.
1316 * Called under i_lock.
1318 static int __mark_caps_flushing(struct inode *inode,
1319 struct ceph_mds_session *session)
1321 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1322 struct ceph_inode_info *ci = ceph_inode(inode);
1325 BUG_ON(ci->i_dirty_caps == 0);
1326 BUG_ON(list_empty(&ci->i_dirty_item));
1328 flushing = ci->i_dirty_caps;
1329 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1330 ceph_cap_string(flushing),
1331 ceph_cap_string(ci->i_flushing_caps),
1332 ceph_cap_string(ci->i_flushing_caps | flushing));
1333 ci->i_flushing_caps |= flushing;
1334 ci->i_dirty_caps = 0;
1335 dout(" inode %p now !dirty\n", inode);
1337 spin_lock(&mdsc->cap_dirty_lock);
1338 list_del_init(&ci->i_dirty_item);
1340 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq;
1341 if (list_empty(&ci->i_flushing_item)) {
1342 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1343 mdsc->num_cap_flushing++;
1344 dout(" inode %p now flushing seq %lld\n", inode,
1345 ci->i_cap_flush_seq);
1347 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1348 dout(" inode %p now flushing (more) seq %lld\n", inode,
1349 ci->i_cap_flush_seq);
1351 spin_unlock(&mdsc->cap_dirty_lock);
1357 * Swiss army knife function to examine currently used and wanted
1358 * versus held caps. Release, flush, ack revoked caps to mds as
1361 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1362 * cap release further.
1363 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1364 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1367 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1368 struct ceph_mds_session *session)
1370 struct ceph_client *client = ceph_inode_to_client(&ci->vfs_inode);
1371 struct ceph_mds_client *mdsc = &client->mdsc;
1372 struct inode *inode = &ci->vfs_inode;
1373 struct ceph_cap *cap;
1374 int file_wanted, used;
1375 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1376 int drop_session_lock = session ? 0 : 1;
1377 int want, retain, revoking, flushing = 0;
1378 int mds = -1; /* keep track of how far we've gone through i_caps list
1379 to avoid an infinite loop on retry */
1381 int tried_invalidate = 0;
1382 int delayed = 0, sent = 0, force_requeue = 0, num;
1383 int is_delayed = flags & CHECK_CAPS_NODELAY;
1385 /* if we are unmounting, flush any unused caps immediately. */
1389 spin_lock(&inode->i_lock);
1391 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1392 flags |= CHECK_CAPS_FLUSH;
1394 /* flush snaps first time around only */
1395 if (!list_empty(&ci->i_cap_snaps))
1396 __ceph_flush_snaps(ci, &session);
1399 spin_lock(&inode->i_lock);
1401 file_wanted = __ceph_caps_file_wanted(ci);
1402 used = __ceph_caps_used(ci);
1403 want = file_wanted | used;
1405 retain = want | CEPH_CAP_PIN;
1406 if (!mdsc->stopping && inode->i_nlink > 0) {
1408 retain |= CEPH_CAP_ANY; /* be greedy */
1410 retain |= CEPH_CAP_ANY_SHARED;
1412 * keep RD only if we didn't have the file open RW,
1413 * because then the mds would revoke it anyway to
1414 * journal max_size=0.
1416 if (ci->i_max_size == 0)
1417 retain |= CEPH_CAP_ANY_RD;
1421 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1422 " issued %s retain %s %s%s%s\n", inode,
1423 ceph_cap_string(file_wanted),
1424 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1425 ceph_cap_string(ci->i_flushing_caps),
1426 ceph_cap_string(__ceph_caps_issued(ci, NULL)),
1427 ceph_cap_string(retain),
1428 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1429 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1430 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1433 * If we no longer need to hold onto old our caps, and we may
1434 * have cached pages, but don't want them, then try to invalidate.
1435 * If we fail, it's because pages are locked.... try again later.
1437 if ((!is_delayed || mdsc->stopping) &&
1438 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */
1439 ci->i_rdcache_gen && /* may have cached pages */
1440 file_wanted == 0 && /* no open files */
1441 !ci->i_truncate_pending &&
1442 !tried_invalidate) {
1443 u32 invalidating_gen = ci->i_rdcache_gen;
1446 dout("check_caps trying to invalidate on %p\n", inode);
1447 spin_unlock(&inode->i_lock);
1448 ret = invalidate_mapping_pages(&inode->i_data, 0, -1);
1449 spin_lock(&inode->i_lock);
1450 if (ret == 0 && invalidating_gen == ci->i_rdcache_gen) {
1452 ci->i_rdcache_gen = 0;
1453 ci->i_rdcache_revoking = 0;
1455 dout("check_caps failed to invalidate pages\n");
1456 /* we failed to invalidate pages. check these
1457 caps again later. */
1459 __cap_set_timeouts(mdsc, ci);
1461 tried_invalidate = 1;
1466 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1467 cap = rb_entry(p, struct ceph_cap, ci_node);
1470 /* avoid looping forever */
1471 if (mds >= cap->mds ||
1472 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1475 /* NOTE: no side-effects allowed, until we take s_mutex */
1477 revoking = cap->implemented & ~cap->issued;
1479 dout("mds%d revoking %s\n", cap->mds,
1480 ceph_cap_string(revoking));
1482 if (cap == ci->i_auth_cap &&
1483 (cap->issued & CEPH_CAP_FILE_WR)) {
1484 /* request larger max_size from MDS? */
1485 if (ci->i_wanted_max_size > ci->i_max_size &&
1486 ci->i_wanted_max_size > ci->i_requested_max_size) {
1487 dout("requesting new max_size\n");
1491 /* approaching file_max? */
1492 if ((inode->i_size << 1) >= ci->i_max_size &&
1493 (ci->i_reported_size << 1) < ci->i_max_size) {
1494 dout("i_size approaching max_size\n");
1498 /* flush anything dirty? */
1499 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1501 dout("flushing dirty caps\n");
1505 /* completed revocation? going down and there are no caps? */
1506 if (revoking && (revoking & used) == 0) {
1507 dout("completed revocation of %s\n",
1508 ceph_cap_string(cap->implemented & ~cap->issued));
1512 /* want more caps from mds? */
1513 if (want & ~(cap->mds_wanted | cap->issued))
1516 /* things we might delay */
1517 if ((cap->issued & ~retain) == 0 &&
1518 cap->mds_wanted == want)
1519 continue; /* nope, all good */
1525 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1526 time_before(jiffies, ci->i_hold_caps_max)) {
1527 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1528 ceph_cap_string(cap->issued),
1529 ceph_cap_string(cap->issued & retain),
1530 ceph_cap_string(cap->mds_wanted),
1531 ceph_cap_string(want));
1537 if (session && session != cap->session) {
1538 dout("oops, wrong session %p mutex\n", session);
1539 mutex_unlock(&session->s_mutex);
1543 session = cap->session;
1544 if (mutex_trylock(&session->s_mutex) == 0) {
1545 dout("inverting session/ino locks on %p\n",
1547 spin_unlock(&inode->i_lock);
1548 if (took_snap_rwsem) {
1549 up_read(&mdsc->snap_rwsem);
1550 took_snap_rwsem = 0;
1552 mutex_lock(&session->s_mutex);
1556 /* take snap_rwsem after session mutex */
1557 if (!took_snap_rwsem) {
1558 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1559 dout("inverting snap/in locks on %p\n",
1561 spin_unlock(&inode->i_lock);
1562 down_read(&mdsc->snap_rwsem);
1563 took_snap_rwsem = 1;
1566 took_snap_rwsem = 1;
1569 if (cap == ci->i_auth_cap && ci->i_dirty_caps)
1570 flushing = __mark_caps_flushing(inode, session);
1572 mds = cap->mds; /* remember mds, so we don't repeat */
1575 /* __send_cap drops i_lock */
1576 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, used, want,
1577 retain, flushing, NULL);
1578 goto retry; /* retake i_lock and restart our cap scan. */
1582 * Reschedule delayed caps release if we delayed anything,
1585 if (delayed && is_delayed)
1586 force_requeue = 1; /* __send_cap delayed release; requeue */
1587 if (!delayed && !is_delayed)
1588 __cap_delay_cancel(mdsc, ci);
1589 else if (!is_delayed || force_requeue)
1590 __cap_delay_requeue(mdsc, ci);
1592 spin_unlock(&inode->i_lock);
1594 if (session && drop_session_lock)
1595 mutex_unlock(&session->s_mutex);
1596 if (took_snap_rwsem)
1597 up_read(&mdsc->snap_rwsem);
1601 * Try to flush dirty caps back to the auth mds.
1603 static int try_flush_caps(struct inode *inode, struct ceph_mds_session *session,
1604 unsigned *flush_tid)
1606 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1607 struct ceph_inode_info *ci = ceph_inode(inode);
1608 int unlock_session = session ? 0 : 1;
1612 spin_lock(&inode->i_lock);
1613 if (ci->i_dirty_caps && ci->i_auth_cap) {
1614 struct ceph_cap *cap = ci->i_auth_cap;
1615 int used = __ceph_caps_used(ci);
1616 int want = __ceph_caps_wanted(ci);
1620 spin_unlock(&inode->i_lock);
1621 session = cap->session;
1622 mutex_lock(&session->s_mutex);
1625 BUG_ON(session != cap->session);
1626 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1629 flushing = __mark_caps_flushing(inode, session);
1631 /* __send_cap drops i_lock */
1632 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1633 cap->issued | cap->implemented, flushing,
1638 spin_lock(&inode->i_lock);
1639 __cap_delay_requeue(mdsc, ci);
1642 spin_unlock(&inode->i_lock);
1644 if (session && unlock_session)
1645 mutex_unlock(&session->s_mutex);
1650 * Return true if we've flushed caps through the given flush_tid.
1652 static int caps_are_flushed(struct inode *inode, unsigned tid)
1654 struct ceph_inode_info *ci = ceph_inode(inode);
1655 int dirty, i, ret = 1;
1657 spin_lock(&inode->i_lock);
1658 dirty = __ceph_caps_dirty(ci);
1659 for (i = 0; i < CEPH_CAP_BITS; i++)
1660 if ((ci->i_flushing_caps & (1 << i)) &&
1661 ci->i_cap_flush_tid[i] <= tid) {
1662 /* still flushing this bit */
1666 spin_unlock(&inode->i_lock);
1671 * Wait on any unsafe replies for the given inode. First wait on the
1672 * newest request, and make that the upper bound. Then, if there are
1673 * more requests, keep waiting on the oldest as long as it is still older
1674 * than the original request.
1676 static void sync_write_wait(struct inode *inode)
1678 struct ceph_inode_info *ci = ceph_inode(inode);
1679 struct list_head *head = &ci->i_unsafe_writes;
1680 struct ceph_osd_request *req;
1683 spin_lock(&ci->i_unsafe_lock);
1684 if (list_empty(head))
1687 /* set upper bound as _last_ entry in chain */
1688 req = list_entry(head->prev, struct ceph_osd_request,
1690 last_tid = req->r_tid;
1693 ceph_osdc_get_request(req);
1694 spin_unlock(&ci->i_unsafe_lock);
1695 dout("sync_write_wait on tid %llu (until %llu)\n",
1696 req->r_tid, last_tid);
1697 wait_for_completion(&req->r_safe_completion);
1698 spin_lock(&ci->i_unsafe_lock);
1699 ceph_osdc_put_request(req);
1702 * from here on look at first entry in chain, since we
1703 * only want to wait for anything older than last_tid
1705 if (list_empty(head))
1707 req = list_entry(head->next, struct ceph_osd_request,
1709 } while (req->r_tid < last_tid);
1711 spin_unlock(&ci->i_unsafe_lock);
1714 int ceph_fsync(struct file *file, struct dentry *dentry, int datasync)
1716 struct inode *inode = dentry->d_inode;
1717 struct ceph_inode_info *ci = ceph_inode(inode);
1722 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
1723 sync_write_wait(inode);
1725 ret = filemap_write_and_wait(inode->i_mapping);
1729 dirty = try_flush_caps(inode, NULL, &flush_tid);
1730 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1733 * only wait on non-file metadata writeback (the mds
1734 * can recover size and mtime, so we don't need to
1737 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1738 dout("fsync waiting for flush_tid %u\n", flush_tid);
1739 ret = wait_event_interruptible(ci->i_cap_wq,
1740 caps_are_flushed(inode, flush_tid));
1743 dout("fsync %p%s done\n", inode, datasync ? " datasync" : "");
1748 * Flush any dirty caps back to the mds. If we aren't asked to wait,
1749 * queue inode for flush but don't do so immediately, because we can
1750 * get by with fewer MDS messages if we wait for data writeback to
1753 int ceph_write_inode(struct inode *inode, int wait)
1755 struct ceph_inode_info *ci = ceph_inode(inode);
1760 dout("write_inode %p wait=%d\n", inode, wait);
1762 dirty = try_flush_caps(inode, NULL, &flush_tid);
1764 err = wait_event_interruptible(ci->i_cap_wq,
1765 caps_are_flushed(inode, flush_tid));
1767 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1769 spin_lock(&inode->i_lock);
1770 if (__ceph_caps_dirty(ci))
1771 __cap_delay_requeue_front(mdsc, ci);
1772 spin_unlock(&inode->i_lock);
1778 * After a recovering MDS goes active, we need to resend any caps
1781 * Caller holds session->s_mutex.
1783 static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
1784 struct ceph_mds_session *session)
1786 struct ceph_cap_snap *capsnap;
1788 dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
1789 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
1791 struct ceph_inode_info *ci = capsnap->ci;
1792 struct inode *inode = &ci->vfs_inode;
1793 struct ceph_cap *cap;
1795 spin_lock(&inode->i_lock);
1796 cap = ci->i_auth_cap;
1797 if (cap && cap->session == session) {
1798 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
1800 __ceph_flush_snaps(ci, &session);
1802 pr_err("%p auth cap %p not mds%d ???\n", inode,
1803 cap, session->s_mds);
1804 spin_unlock(&inode->i_lock);
1809 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
1810 struct ceph_mds_session *session)
1812 struct ceph_inode_info *ci;
1814 kick_flushing_capsnaps(mdsc, session);
1816 dout("kick_flushing_caps mds%d\n", session->s_mds);
1817 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
1818 struct inode *inode = &ci->vfs_inode;
1819 struct ceph_cap *cap;
1822 spin_lock(&inode->i_lock);
1823 cap = ci->i_auth_cap;
1824 if (cap && cap->session == session) {
1825 dout("kick_flushing_caps %p cap %p %s\n", inode,
1826 cap, ceph_cap_string(ci->i_flushing_caps));
1827 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1828 __ceph_caps_used(ci),
1829 __ceph_caps_wanted(ci),
1830 cap->issued | cap->implemented,
1831 ci->i_flushing_caps, NULL);
1833 spin_lock(&inode->i_lock);
1834 __cap_delay_requeue(mdsc, ci);
1835 spin_unlock(&inode->i_lock);
1838 pr_err("%p auth cap %p not mds%d ???\n", inode,
1839 cap, session->s_mds);
1840 spin_unlock(&inode->i_lock);
1847 * Take references to capabilities we hold, so that we don't release
1848 * them to the MDS prematurely.
1850 * Protected by i_lock.
1852 static void __take_cap_refs(struct ceph_inode_info *ci, int got)
1854 if (got & CEPH_CAP_PIN)
1856 if (got & CEPH_CAP_FILE_RD)
1858 if (got & CEPH_CAP_FILE_CACHE)
1859 ci->i_rdcache_ref++;
1860 if (got & CEPH_CAP_FILE_WR)
1862 if (got & CEPH_CAP_FILE_BUFFER) {
1863 if (ci->i_wrbuffer_ref == 0)
1864 igrab(&ci->vfs_inode);
1865 ci->i_wrbuffer_ref++;
1866 dout("__take_cap_refs %p wrbuffer %d -> %d (?)\n",
1867 &ci->vfs_inode, ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref);
1872 * Try to grab cap references. Specify those refs we @want, and the
1873 * minimal set we @need. Also include the larger offset we are writing
1874 * to (when applicable), and check against max_size here as well.
1875 * Note that caller is responsible for ensuring max_size increases are
1876 * requested from the MDS.
1878 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
1879 int *got, loff_t endoff, int *check_max, int *err)
1881 struct inode *inode = &ci->vfs_inode;
1883 int have, implemented;
1885 dout("get_cap_refs %p need %s want %s\n", inode,
1886 ceph_cap_string(need), ceph_cap_string(want));
1887 spin_lock(&inode->i_lock);
1889 /* make sure we _have_ some caps! */
1890 if (!__ceph_is_any_caps(ci)) {
1891 dout("get_cap_refs %p no real caps\n", inode);
1897 if (need & CEPH_CAP_FILE_WR) {
1898 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
1899 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
1900 inode, endoff, ci->i_max_size);
1901 if (endoff > ci->i_wanted_max_size) {
1908 * If a sync write is in progress, we must wait, so that we
1909 * can get a final snapshot value for size+mtime.
1911 if (__ceph_have_pending_cap_snap(ci)) {
1912 dout("get_cap_refs %p cap_snap_pending\n", inode);
1916 have = __ceph_caps_issued(ci, &implemented);
1919 * disallow writes while a truncate is pending
1921 if (ci->i_truncate_pending)
1922 have &= ~CEPH_CAP_FILE_WR;
1924 if ((have & need) == need) {
1926 * Look at (implemented & ~have & not) so that we keep waiting
1927 * on transition from wanted -> needed caps. This is needed
1928 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
1929 * going before a prior buffered writeback happens.
1931 int not = want & ~(have & need);
1932 int revoking = implemented & ~have;
1933 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
1934 inode, ceph_cap_string(have), ceph_cap_string(not),
1935 ceph_cap_string(revoking));
1936 if ((revoking & not) == 0) {
1937 *got = need | (have & want);
1938 __take_cap_refs(ci, *got);
1942 dout("get_cap_refs %p have %s needed %s\n", inode,
1943 ceph_cap_string(have), ceph_cap_string(need));
1946 spin_unlock(&inode->i_lock);
1947 dout("get_cap_refs %p ret %d got %s\n", inode,
1948 ret, ceph_cap_string(*got));
1953 * Check the offset we are writing up to against our current
1954 * max_size. If necessary, tell the MDS we want to write to
1957 static void check_max_size(struct inode *inode, loff_t endoff)
1959 struct ceph_inode_info *ci = ceph_inode(inode);
1962 /* do we need to explicitly request a larger max_size? */
1963 spin_lock(&inode->i_lock);
1964 if ((endoff >= ci->i_max_size ||
1965 endoff > (inode->i_size << 1)) &&
1966 endoff > ci->i_wanted_max_size) {
1967 dout("write %p at large endoff %llu, req max_size\n",
1969 ci->i_wanted_max_size = endoff;
1972 spin_unlock(&inode->i_lock);
1974 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
1978 * Wait for caps, and take cap references. If we can't get a WR cap
1979 * due to a small max_size, make sure we check_max_size (and possibly
1980 * ask the mds) so we don't get hung up indefinitely.
1982 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, int *got,
1985 int check_max, ret, err;
1989 check_max_size(&ci->vfs_inode, endoff);
1992 ret = wait_event_interruptible(ci->i_cap_wq,
1993 try_get_cap_refs(ci, need, want,
2004 * Take cap refs. Caller must already know we hold at least one ref
2005 * on the caps in question or we don't know this is safe.
2007 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2009 spin_lock(&ci->vfs_inode.i_lock);
2010 __take_cap_refs(ci, caps);
2011 spin_unlock(&ci->vfs_inode.i_lock);
2017 * If we released the last ref on any given cap, call ceph_check_caps
2018 * to release (or schedule a release).
2020 * If we are releasing a WR cap (from a sync write), finalize any affected
2021 * cap_snap, and wake up any waiters.
2023 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2025 struct inode *inode = &ci->vfs_inode;
2026 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2027 struct ceph_cap_snap *capsnap;
2029 spin_lock(&inode->i_lock);
2030 if (had & CEPH_CAP_PIN)
2032 if (had & CEPH_CAP_FILE_RD)
2033 if (--ci->i_rd_ref == 0)
2035 if (had & CEPH_CAP_FILE_CACHE)
2036 if (--ci->i_rdcache_ref == 0)
2038 if (had & CEPH_CAP_FILE_BUFFER) {
2039 if (--ci->i_wrbuffer_ref == 0) {
2043 dout("put_cap_refs %p wrbuffer %d -> %d (?)\n",
2044 inode, ci->i_wrbuffer_ref+1, ci->i_wrbuffer_ref);
2046 if (had & CEPH_CAP_FILE_WR)
2047 if (--ci->i_wr_ref == 0) {
2049 if (!list_empty(&ci->i_cap_snaps)) {
2050 capsnap = list_first_entry(&ci->i_cap_snaps,
2051 struct ceph_cap_snap,
2053 if (capsnap->writing) {
2054 capsnap->writing = 0;
2056 __ceph_finish_cap_snap(ci,
2062 spin_unlock(&inode->i_lock);
2064 dout("put_cap_refs %p had %s %s\n", inode, ceph_cap_string(had),
2065 last ? "last" : "");
2067 if (last && !flushsnaps)
2068 ceph_check_caps(ci, 0, NULL);
2069 else if (flushsnaps)
2070 ceph_flush_snaps(ci);
2072 wake_up(&ci->i_cap_wq);
2078 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2079 * context. Adjust per-snap dirty page accounting as appropriate.
2080 * Once all dirty data for a cap_snap is flushed, flush snapped file
2081 * metadata back to the MDS. If we dropped the last ref, call
2084 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2085 struct ceph_snap_context *snapc)
2087 struct inode *inode = &ci->vfs_inode;
2091 struct ceph_cap_snap *capsnap = NULL;
2093 spin_lock(&inode->i_lock);
2094 ci->i_wrbuffer_ref -= nr;
2095 last = !ci->i_wrbuffer_ref;
2097 if (ci->i_head_snapc == snapc) {
2098 ci->i_wrbuffer_ref_head -= nr;
2099 if (!ci->i_wrbuffer_ref_head) {
2100 ceph_put_snap_context(ci->i_head_snapc);
2101 ci->i_head_snapc = NULL;
2103 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2105 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2106 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2107 last ? " LAST" : "");
2109 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2110 if (capsnap->context == snapc) {
2112 capsnap->dirty_pages -= nr;
2113 last_snap = !capsnap->dirty_pages;
2118 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2119 " snap %lld %d/%d -> %d/%d %s%s\n",
2120 inode, capsnap, capsnap->context->seq,
2121 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2122 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2123 last ? " (wrbuffer last)" : "",
2124 last_snap ? " (capsnap last)" : "");
2127 spin_unlock(&inode->i_lock);
2130 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2132 } else if (last_snap) {
2133 ceph_flush_snaps(ci);
2134 wake_up(&ci->i_cap_wq);
2139 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2140 * actually be a revocation if it specifies a smaller cap set.)
2142 * caller holds s_mutex.
2145 * 1 - check_caps on auth cap only (writeback)
2146 * 2 - check_caps (ack revoke)
2148 static int handle_cap_grant(struct inode *inode, struct ceph_mds_caps *grant,
2149 struct ceph_mds_session *session,
2150 struct ceph_cap *cap,
2151 struct ceph_buffer *xattr_buf)
2152 __releases(inode->i_lock)
2155 struct ceph_inode_info *ci = ceph_inode(inode);
2156 int mds = session->s_mds;
2157 int seq = le32_to_cpu(grant->seq);
2158 int newcaps = le32_to_cpu(grant->caps);
2159 int issued, implemented, used, wanted, dirty;
2160 u64 size = le64_to_cpu(grant->size);
2161 u64 max_size = le64_to_cpu(grant->max_size);
2162 struct timespec mtime, atime, ctime;
2166 int revoked_rdcache = 0;
2167 int invalidate_async = 0;
2168 int tried_invalidate = 0;
2171 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2172 inode, cap, mds, seq, ceph_cap_string(newcaps));
2173 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2177 * If CACHE is being revoked, and we have no dirty buffers,
2178 * try to invalidate (once). (If there are dirty buffers, we
2179 * will invalidate _after_ writeback.)
2182 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2183 !ci->i_wrbuffer_ref && !tried_invalidate) {
2184 dout("CACHE invalidation\n");
2185 spin_unlock(&inode->i_lock);
2186 tried_invalidate = 1;
2188 ret = invalidate_mapping_pages(&inode->i_data, 0, -1);
2189 spin_lock(&inode->i_lock);
2191 /* there were locked pages.. invalidate later
2192 in a separate thread. */
2193 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2194 invalidate_async = 1;
2195 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2198 /* we successfully invalidated those pages */
2199 revoked_rdcache = 1;
2200 ci->i_rdcache_gen = 0;
2201 ci->i_rdcache_revoking = 0;
2206 /* side effects now are allowed */
2208 issued = __ceph_caps_issued(ci, &implemented);
2209 issued |= implemented | __ceph_caps_dirty(ci);
2211 cap->cap_gen = session->s_cap_gen;
2213 __check_cap_issue(ci, cap, newcaps);
2215 if ((issued & CEPH_CAP_AUTH_EXCL) == 0) {
2216 inode->i_mode = le32_to_cpu(grant->mode);
2217 inode->i_uid = le32_to_cpu(grant->uid);
2218 inode->i_gid = le32_to_cpu(grant->gid);
2219 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2220 inode->i_uid, inode->i_gid);
2223 if ((issued & CEPH_CAP_LINK_EXCL) == 0)
2224 inode->i_nlink = le32_to_cpu(grant->nlink);
2226 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2227 int len = le32_to_cpu(grant->xattr_len);
2228 u64 version = le64_to_cpu(grant->xattr_version);
2230 if (version > ci->i_xattrs.version) {
2231 dout(" got new xattrs v%llu on %p len %d\n",
2232 version, inode, len);
2233 if (ci->i_xattrs.blob)
2234 ceph_buffer_put(ci->i_xattrs.blob);
2235 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2236 ci->i_xattrs.version = version;
2240 /* size/ctime/mtime/atime? */
2241 ceph_fill_file_size(inode, issued,
2242 le32_to_cpu(grant->truncate_seq),
2243 le64_to_cpu(grant->truncate_size), size);
2244 ceph_decode_timespec(&mtime, &grant->mtime);
2245 ceph_decode_timespec(&atime, &grant->atime);
2246 ceph_decode_timespec(&ctime, &grant->ctime);
2247 ceph_fill_file_time(inode, issued,
2248 le32_to_cpu(grant->time_warp_seq), &ctime, &mtime,
2251 /* max size increase? */
2252 if (max_size != ci->i_max_size) {
2253 dout("max_size %lld -> %llu\n", ci->i_max_size, max_size);
2254 ci->i_max_size = max_size;
2255 if (max_size >= ci->i_wanted_max_size) {
2256 ci->i_wanted_max_size = 0; /* reset */
2257 ci->i_requested_max_size = 0;
2262 /* check cap bits */
2263 wanted = __ceph_caps_wanted(ci);
2264 used = __ceph_caps_used(ci);
2265 dirty = __ceph_caps_dirty(ci);
2266 dout(" my wanted = %s, used = %s, dirty %s\n",
2267 ceph_cap_string(wanted),
2268 ceph_cap_string(used),
2269 ceph_cap_string(dirty));
2270 if (wanted != le32_to_cpu(grant->wanted)) {
2271 dout("mds wanted %s -> %s\n",
2272 ceph_cap_string(le32_to_cpu(grant->wanted)),
2273 ceph_cap_string(wanted));
2274 grant->wanted = cpu_to_le32(wanted);
2279 /* file layout may have changed */
2280 ci->i_layout = grant->layout;
2282 /* revocation, grant, or no-op? */
2283 if (cap->issued & ~newcaps) {
2284 dout("revocation: %s -> %s\n", ceph_cap_string(cap->issued),
2285 ceph_cap_string(newcaps));
2286 if ((used & ~newcaps) & CEPH_CAP_FILE_BUFFER)
2287 writeback = 1; /* will delay ack */
2288 else if (dirty & ~newcaps)
2289 reply = 1; /* initiate writeback in check_caps */
2290 else if (((used & ~newcaps) & CEPH_CAP_FILE_CACHE) == 0 ||
2292 reply = 2; /* send revoke ack in check_caps */
2293 cap->issued = newcaps;
2294 } else if (cap->issued == newcaps) {
2295 dout("caps unchanged: %s -> %s\n",
2296 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2298 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2299 ceph_cap_string(newcaps));
2300 cap->issued = newcaps;
2301 cap->implemented |= newcaps; /* add bits only, to
2302 * avoid stepping on a
2303 * pending revocation */
2307 spin_unlock(&inode->i_lock);
2310 * queue inode for writeback: we can't actually call
2311 * filemap_write_and_wait, etc. from message handler
2314 dout("queueing %p for writeback\n", inode);
2315 if (ceph_queue_writeback(inode))
2318 if (invalidate_async) {
2319 dout("queueing %p for page invalidation\n", inode);
2320 if (ceph_queue_page_invalidation(inode))
2324 wake_up(&ci->i_cap_wq);
2329 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2330 * MDS has been safely committed.
2332 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2333 struct ceph_mds_caps *m,
2334 struct ceph_mds_session *session,
2335 struct ceph_cap *cap)
2336 __releases(inode->i_lock)
2338 struct ceph_inode_info *ci = ceph_inode(inode);
2339 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
2340 unsigned seq = le32_to_cpu(m->seq);
2341 int dirty = le32_to_cpu(m->dirty);
2346 for (i = 0; i < CEPH_CAP_BITS; i++)
2347 if ((dirty & (1 << i)) &&
2348 flush_tid == ci->i_cap_flush_tid[i])
2351 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2352 " flushing %s -> %s\n",
2353 inode, session->s_mds, seq, ceph_cap_string(dirty),
2354 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2355 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2357 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned))
2360 ci->i_flushing_caps &= ~cleaned;
2362 spin_lock(&mdsc->cap_dirty_lock);
2363 if (ci->i_flushing_caps == 0) {
2364 list_del_init(&ci->i_flushing_item);
2365 if (!list_empty(&session->s_cap_flushing))
2366 dout(" mds%d still flushing cap on %p\n",
2368 &list_entry(session->s_cap_flushing.next,
2369 struct ceph_inode_info,
2370 i_flushing_item)->vfs_inode);
2371 mdsc->num_cap_flushing--;
2372 wake_up(&mdsc->cap_flushing_wq);
2373 dout(" inode %p now !flushing\n", inode);
2375 if (ci->i_dirty_caps == 0) {
2376 dout(" inode %p now clean\n", inode);
2377 BUG_ON(!list_empty(&ci->i_dirty_item));
2380 BUG_ON(list_empty(&ci->i_dirty_item));
2383 spin_unlock(&mdsc->cap_dirty_lock);
2384 wake_up(&ci->i_cap_wq);
2387 spin_unlock(&inode->i_lock);
2393 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
2394 * throw away our cap_snap.
2396 * Caller hold s_mutex.
2398 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
2399 struct ceph_mds_caps *m,
2400 struct ceph_mds_session *session)
2402 struct ceph_inode_info *ci = ceph_inode(inode);
2403 u64 follows = le64_to_cpu(m->snap_follows);
2404 struct ceph_cap_snap *capsnap;
2407 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
2408 inode, ci, session->s_mds, follows);
2410 spin_lock(&inode->i_lock);
2411 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2412 if (capsnap->follows == follows) {
2413 if (capsnap->flush_tid != flush_tid) {
2414 dout(" cap_snap %p follows %lld tid %lld !="
2415 " %lld\n", capsnap, follows,
2416 flush_tid, capsnap->flush_tid);
2419 WARN_ON(capsnap->dirty_pages || capsnap->writing);
2420 dout(" removing cap_snap %p follows %lld\n",
2422 ceph_put_snap_context(capsnap->context);
2423 list_del(&capsnap->ci_item);
2424 list_del(&capsnap->flushing_item);
2425 ceph_put_cap_snap(capsnap);
2429 dout(" skipping cap_snap %p follows %lld\n",
2430 capsnap, capsnap->follows);
2433 spin_unlock(&inode->i_lock);
2439 * Handle TRUNC from MDS, indicating file truncation.
2441 * caller hold s_mutex.
2443 static void handle_cap_trunc(struct inode *inode,
2444 struct ceph_mds_caps *trunc,
2445 struct ceph_mds_session *session)
2446 __releases(inode->i_lock)
2448 struct ceph_inode_info *ci = ceph_inode(inode);
2449 int mds = session->s_mds;
2450 int seq = le32_to_cpu(trunc->seq);
2451 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
2452 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
2453 u64 size = le64_to_cpu(trunc->size);
2454 int implemented = 0;
2455 int dirty = __ceph_caps_dirty(ci);
2456 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
2457 int queue_trunc = 0;
2459 issued |= implemented | dirty;
2461 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
2462 inode, mds, seq, truncate_size, truncate_seq);
2463 queue_trunc = ceph_fill_file_size(inode, issued,
2464 truncate_seq, truncate_size, size);
2465 spin_unlock(&inode->i_lock);
2468 if (queue_work(ceph_client(inode->i_sb)->trunc_wq,
2469 &ci->i_vmtruncate_work))
2474 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
2475 * different one. If we are the most recent migration we've seen (as
2476 * indicated by mseq), make note of the migrating cap bits for the
2477 * duration (until we see the corresponding IMPORT).
2479 * caller holds s_mutex
2481 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
2482 struct ceph_mds_session *session)
2484 struct ceph_inode_info *ci = ceph_inode(inode);
2485 int mds = session->s_mds;
2486 unsigned mseq = le32_to_cpu(ex->migrate_seq);
2487 struct ceph_cap *cap = NULL, *t;
2491 dout("handle_cap_export inode %p ci %p mds%d mseq %d\n",
2492 inode, ci, mds, mseq);
2494 spin_lock(&inode->i_lock);
2496 /* make sure we haven't seen a higher mseq */
2497 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2498 t = rb_entry(p, struct ceph_cap, ci_node);
2499 if (ceph_seq_cmp(t->mseq, mseq) > 0) {
2500 dout(" higher mseq on cap from mds%d\n",
2504 if (t->session->s_mds == mds)
2511 ci->i_cap_exporting_mds = mds;
2512 ci->i_cap_exporting_mseq = mseq;
2513 ci->i_cap_exporting_issued = cap->issued;
2515 __ceph_remove_cap(cap, NULL);
2520 spin_unlock(&inode->i_lock);
2524 * Handle cap IMPORT. If there are temp bits from an older EXPORT,
2527 * caller holds s_mutex.
2529 static void handle_cap_import(struct ceph_mds_client *mdsc,
2530 struct inode *inode, struct ceph_mds_caps *im,
2531 struct ceph_mds_session *session,
2532 void *snaptrace, int snaptrace_len)
2534 struct ceph_inode_info *ci = ceph_inode(inode);
2535 int mds = session->s_mds;
2536 unsigned issued = le32_to_cpu(im->caps);
2537 unsigned wanted = le32_to_cpu(im->wanted);
2538 unsigned seq = le32_to_cpu(im->seq);
2539 unsigned mseq = le32_to_cpu(im->migrate_seq);
2540 u64 realmino = le64_to_cpu(im->realm);
2541 u64 cap_id = le64_to_cpu(im->cap_id);
2543 if (ci->i_cap_exporting_mds >= 0 &&
2544 ceph_seq_cmp(ci->i_cap_exporting_mseq, mseq) < 0) {
2545 dout("handle_cap_import inode %p ci %p mds%d mseq %d"
2546 " - cleared exporting from mds%d\n",
2547 inode, ci, mds, mseq,
2548 ci->i_cap_exporting_mds);
2549 ci->i_cap_exporting_issued = 0;
2550 ci->i_cap_exporting_mseq = 0;
2551 ci->i_cap_exporting_mds = -1;
2553 dout("handle_cap_import inode %p ci %p mds%d mseq %d\n",
2554 inode, ci, mds, mseq);
2557 down_write(&mdsc->snap_rwsem);
2558 ceph_update_snap_trace(mdsc, snaptrace, snaptrace+snaptrace_len,
2560 downgrade_write(&mdsc->snap_rwsem);
2561 ceph_add_cap(inode, session, cap_id, -1,
2562 issued, wanted, seq, mseq, realmino, CEPH_CAP_FLAG_AUTH,
2563 NULL /* no caps context */);
2564 try_flush_caps(inode, session, NULL);
2565 up_read(&mdsc->snap_rwsem);
2569 * Handle a caps message from the MDS.
2571 * Identify the appropriate session, inode, and call the right handler
2572 * based on the cap op.
2574 void ceph_handle_caps(struct ceph_mds_session *session,
2575 struct ceph_msg *msg)
2577 struct ceph_mds_client *mdsc = session->s_mdsc;
2578 struct super_block *sb = mdsc->client->sb;
2579 struct inode *inode;
2580 struct ceph_cap *cap;
2581 struct ceph_mds_caps *h;
2582 int mds = le64_to_cpu(msg->hdr.src.name.num);
2585 struct ceph_vino vino;
2592 dout("handle_caps from mds%d\n", mds);
2595 tid = le64_to_cpu(msg->hdr.tid);
2596 if (msg->front.iov_len < sizeof(*h))
2598 h = msg->front.iov_base;
2599 op = le32_to_cpu(h->op);
2600 vino.ino = le64_to_cpu(h->ino);
2601 vino.snap = CEPH_NOSNAP;
2602 cap_id = le64_to_cpu(h->cap_id);
2603 seq = le32_to_cpu(h->seq);
2604 size = le64_to_cpu(h->size);
2605 max_size = le64_to_cpu(h->max_size);
2607 mutex_lock(&session->s_mutex);
2609 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
2613 inode = ceph_find_inode(sb, vino);
2614 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
2617 dout(" i don't have ino %llx\n", vino.ino);
2621 /* these will work even if we don't have a cap yet */
2623 case CEPH_CAP_OP_FLUSHSNAP_ACK:
2624 handle_cap_flushsnap_ack(inode, tid, h, session);
2627 case CEPH_CAP_OP_EXPORT:
2628 handle_cap_export(inode, h, session);
2631 case CEPH_CAP_OP_IMPORT:
2632 handle_cap_import(mdsc, inode, h, session,
2634 le32_to_cpu(h->snap_trace_len));
2635 check_caps = 1; /* we may have sent a RELEASE to the old auth */
2639 /* the rest require a cap */
2640 spin_lock(&inode->i_lock);
2641 cap = __get_cap_for_mds(ceph_inode(inode), mds);
2643 dout("no cap on %p ino %llx.%llx from mds%d, releasing\n",
2644 inode, ceph_ino(inode), ceph_snap(inode), mds);
2645 spin_unlock(&inode->i_lock);
2649 /* note that each of these drops i_lock for us */
2651 case CEPH_CAP_OP_REVOKE:
2652 case CEPH_CAP_OP_GRANT:
2653 r = handle_cap_grant(inode, h, session, cap, msg->middle);
2655 ceph_check_caps(ceph_inode(inode),
2656 CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2659 ceph_check_caps(ceph_inode(inode),
2664 case CEPH_CAP_OP_FLUSH_ACK:
2665 handle_cap_flush_ack(inode, tid, h, session, cap);
2668 case CEPH_CAP_OP_TRUNC:
2669 handle_cap_trunc(inode, h, session);
2673 spin_unlock(&inode->i_lock);
2674 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
2675 ceph_cap_op_name(op));
2679 mutex_unlock(&session->s_mutex);
2682 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_NODELAY, NULL);
2688 pr_err("ceph_handle_caps: corrupt message\n");
2694 * Delayed work handler to process end of delayed cap release LRU list.
2696 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
2698 struct ceph_inode_info *ci;
2699 int flags = CHECK_CAPS_NODELAY;
2701 dout("check_delayed_caps\n");
2703 spin_lock(&mdsc->cap_delay_lock);
2704 if (list_empty(&mdsc->cap_delay_list))
2706 ci = list_first_entry(&mdsc->cap_delay_list,
2707 struct ceph_inode_info,
2709 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
2710 time_before(jiffies, ci->i_hold_caps_max))
2712 list_del_init(&ci->i_cap_delay_list);
2713 spin_unlock(&mdsc->cap_delay_lock);
2714 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
2715 ceph_check_caps(ci, flags, NULL);
2717 spin_unlock(&mdsc->cap_delay_lock);
2721 * Flush all dirty caps to the mds
2723 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
2725 struct ceph_inode_info *ci;
2726 struct inode *inode;
2728 dout("flush_dirty_caps\n");
2729 spin_lock(&mdsc->cap_dirty_lock);
2730 while (!list_empty(&mdsc->cap_dirty)) {
2731 ci = list_first_entry(&mdsc->cap_dirty,
2732 struct ceph_inode_info,
2734 inode = igrab(&ci->vfs_inode);
2735 spin_unlock(&mdsc->cap_dirty_lock);
2737 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH,
2741 spin_lock(&mdsc->cap_dirty_lock);
2743 spin_unlock(&mdsc->cap_dirty_lock);
2747 * Drop open file reference. If we were the last open file,
2748 * we may need to release capabilities to the MDS (or schedule
2749 * their delayed release).
2751 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
2753 struct inode *inode = &ci->vfs_inode;
2756 spin_lock(&inode->i_lock);
2757 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
2758 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
2759 BUG_ON(ci->i_nr_by_mode[fmode] == 0);
2760 if (--ci->i_nr_by_mode[fmode] == 0)
2762 spin_unlock(&inode->i_lock);
2764 if (last && ci->i_vino.snap == CEPH_NOSNAP)
2765 ceph_check_caps(ci, 0, NULL);
2769 * Helpers for embedding cap and dentry lease releases into mds
2772 * @force is used by dentry_release (below) to force inclusion of a
2773 * record for the directory inode, even when there aren't any caps to
2776 int ceph_encode_inode_release(void **p, struct inode *inode,
2777 int mds, int drop, int unless, int force)
2779 struct ceph_inode_info *ci = ceph_inode(inode);
2780 struct ceph_cap *cap;
2781 struct ceph_mds_request_release *rel = *p;
2784 dout("encode_inode_release %p mds%d drop %s unless %s\n", inode,
2785 mds, ceph_cap_string(drop), ceph_cap_string(unless));
2787 spin_lock(&inode->i_lock);
2788 cap = __get_cap_for_mds(ci, mds);
2789 if (cap && __cap_is_valid(cap)) {
2791 ((cap->issued & drop) &&
2792 (cap->issued & unless) == 0)) {
2793 if ((cap->issued & drop) &&
2794 (cap->issued & unless) == 0) {
2795 dout("encode_inode_release %p cap %p %s -> "
2797 ceph_cap_string(cap->issued),
2798 ceph_cap_string(cap->issued & ~drop));
2799 cap->issued &= ~drop;
2800 cap->implemented &= ~drop;
2801 if (ci->i_ceph_flags & CEPH_I_NODELAY) {
2802 int wanted = __ceph_caps_wanted(ci);
2803 dout(" wanted %s -> %s (act %s)\n",
2804 ceph_cap_string(cap->mds_wanted),
2805 ceph_cap_string(cap->mds_wanted &
2807 ceph_cap_string(wanted));
2808 cap->mds_wanted &= wanted;
2811 dout("encode_inode_release %p cap %p %s"
2812 " (force)\n", inode, cap,
2813 ceph_cap_string(cap->issued));
2816 rel->ino = cpu_to_le64(ceph_ino(inode));
2817 rel->cap_id = cpu_to_le64(cap->cap_id);
2818 rel->seq = cpu_to_le32(cap->seq);
2819 rel->issue_seq = cpu_to_le32(cap->issue_seq),
2820 rel->mseq = cpu_to_le32(cap->mseq);
2821 rel->caps = cpu_to_le32(cap->issued);
2822 rel->wanted = cpu_to_le32(cap->mds_wanted);
2828 dout("encode_inode_release %p cap %p %s\n",
2829 inode, cap, ceph_cap_string(cap->issued));
2832 spin_unlock(&inode->i_lock);
2836 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
2837 int mds, int drop, int unless)
2839 struct inode *dir = dentry->d_parent->d_inode;
2840 struct ceph_mds_request_release *rel = *p;
2841 struct ceph_dentry_info *di = ceph_dentry(dentry);
2846 * force an record for the directory caps if we have a dentry lease.
2847 * this is racy (can't take i_lock and d_lock together), but it
2848 * doesn't have to be perfect; the mds will revoke anything we don't
2851 spin_lock(&dentry->d_lock);
2852 if (di->lease_session && di->lease_session->s_mds == mds)
2854 spin_unlock(&dentry->d_lock);
2856 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
2858 spin_lock(&dentry->d_lock);
2859 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
2860 dout("encode_dentry_release %p mds%d seq %d\n",
2861 dentry, mds, (int)di->lease_seq);
2862 rel->dname_len = cpu_to_le32(dentry->d_name.len);
2863 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
2864 *p += dentry->d_name.len;
2865 rel->dname_seq = cpu_to_le32(di->lease_seq);
2867 spin_unlock(&dentry->d_lock);