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 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
702 spin_lock(&s->s_cap_lock);
703 list_move_tail(&cap->session_caps, &s->s_caps);
704 spin_unlock(&s->s_cap_lock);
708 * Check if we hold the given mask. If so, move the cap(s) to the
709 * front of their respective LRUs. (This is the preferred way for
710 * callers to check for caps they want.)
712 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
714 struct ceph_cap *cap;
716 int have = ci->i_snap_caps;
718 if ((have & mask) == mask) {
719 dout("__ceph_caps_issued_mask %p snap issued %s"
720 " (mask %s)\n", &ci->vfs_inode,
721 ceph_cap_string(have),
722 ceph_cap_string(mask));
726 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
727 cap = rb_entry(p, struct ceph_cap, ci_node);
728 if (!__cap_is_valid(cap))
730 if ((cap->issued & mask) == mask) {
731 dout("__ceph_caps_issued_mask %p cap %p issued %s"
732 " (mask %s)\n", &ci->vfs_inode, cap,
733 ceph_cap_string(cap->issued),
734 ceph_cap_string(mask));
740 /* does a combination of caps satisfy mask? */
742 if ((have & mask) == mask) {
743 dout("__ceph_caps_issued_mask %p combo issued %s"
744 " (mask %s)\n", &ci->vfs_inode,
745 ceph_cap_string(cap->issued),
746 ceph_cap_string(mask));
750 /* touch this + preceeding caps */
752 for (q = rb_first(&ci->i_caps); q != p;
754 cap = rb_entry(q, struct ceph_cap,
756 if (!__cap_is_valid(cap))
769 * Return true if mask caps are currently being revoked by an MDS.
771 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
773 struct inode *inode = &ci->vfs_inode;
774 struct ceph_cap *cap;
778 spin_lock(&inode->i_lock);
779 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
780 cap = rb_entry(p, struct ceph_cap, ci_node);
781 if (__cap_is_valid(cap) &&
782 (cap->implemented & ~cap->issued & mask)) {
787 spin_unlock(&inode->i_lock);
788 dout("ceph_caps_revoking %p %s = %d\n", inode,
789 ceph_cap_string(mask), ret);
793 int __ceph_caps_used(struct ceph_inode_info *ci)
797 used |= CEPH_CAP_PIN;
799 used |= CEPH_CAP_FILE_RD;
800 if (ci->i_rdcache_ref || ci->i_rdcache_gen)
801 used |= CEPH_CAP_FILE_CACHE;
803 used |= CEPH_CAP_FILE_WR;
804 if (ci->i_wrbuffer_ref)
805 used |= CEPH_CAP_FILE_BUFFER;
810 * wanted, by virtue of open file modes
812 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
816 for (mode = 0; mode < 4; mode++)
817 if (ci->i_nr_by_mode[mode])
818 want |= ceph_caps_for_mode(mode);
823 * Return caps we have registered with the MDS(s) as 'wanted'.
825 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
827 struct ceph_cap *cap;
831 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
832 cap = rb_entry(p, struct ceph_cap, ci_node);
833 if (!__cap_is_valid(cap))
835 mds_wanted |= cap->mds_wanted;
841 * called under i_lock
843 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
845 return !RB_EMPTY_ROOT(&ci->i_caps) || ci->i_cap_exporting_mds >= 0;
849 * caller should hold i_lock, and session s_mutex.
850 * returns true if this is the last cap. if so, caller should iput.
852 void __ceph_remove_cap(struct ceph_cap *cap,
853 struct ceph_cap_reservation *ctx)
855 struct ceph_mds_session *session = cap->session;
856 struct ceph_inode_info *ci = cap->ci;
857 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc;
859 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
861 /* remove from session list */
862 spin_lock(&session->s_cap_lock);
863 list_del_init(&cap->session_caps);
864 session->s_nr_caps--;
865 spin_unlock(&session->s_cap_lock);
867 /* remove from inode list */
868 rb_erase(&cap->ci_node, &ci->i_caps);
870 if (ci->i_auth_cap == cap)
871 ci->i_auth_cap = NULL;
875 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) {
876 struct ceph_snap_realm *realm = ci->i_snap_realm;
877 spin_lock(&realm->inodes_with_caps_lock);
878 list_del_init(&ci->i_snap_realm_item);
879 ci->i_snap_realm_counter++;
880 ci->i_snap_realm = NULL;
881 spin_unlock(&realm->inodes_with_caps_lock);
882 ceph_put_snap_realm(mdsc, realm);
884 if (!__ceph_is_any_real_caps(ci))
885 __cap_delay_cancel(mdsc, ci);
889 * Build and send a cap message to the given MDS.
891 * Caller should be holding s_mutex.
893 static int send_cap_msg(struct ceph_mds_session *session,
894 u64 ino, u64 cid, int op,
895 int caps, int wanted, int dirty,
896 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq,
897 u64 size, u64 max_size,
898 struct timespec *mtime, struct timespec *atime,
900 uid_t uid, gid_t gid, mode_t mode,
902 struct ceph_buffer *xattrs_buf,
905 struct ceph_mds_caps *fc;
906 struct ceph_msg *msg;
908 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
909 " seq %u/%u mseq %u follows %lld size %llu/%llu"
910 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
911 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
912 ceph_cap_string(dirty),
913 seq, issue_seq, mseq, follows, size, max_size,
914 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
916 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc), 0, 0, NULL);
920 fc = msg->front.iov_base;
922 memset(fc, 0, sizeof(*fc));
924 fc->cap_id = cpu_to_le64(cid);
925 fc->op = cpu_to_le32(op);
926 fc->seq = cpu_to_le32(seq);
927 fc->client_tid = cpu_to_le64(flush_tid);
928 fc->issue_seq = cpu_to_le32(issue_seq);
929 fc->migrate_seq = cpu_to_le32(mseq);
930 fc->caps = cpu_to_le32(caps);
931 fc->wanted = cpu_to_le32(wanted);
932 fc->dirty = cpu_to_le32(dirty);
933 fc->ino = cpu_to_le64(ino);
934 fc->snap_follows = cpu_to_le64(follows);
936 fc->size = cpu_to_le64(size);
937 fc->max_size = cpu_to_le64(max_size);
939 ceph_encode_timespec(&fc->mtime, mtime);
941 ceph_encode_timespec(&fc->atime, atime);
942 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
944 fc->uid = cpu_to_le32(uid);
945 fc->gid = cpu_to_le32(gid);
946 fc->mode = cpu_to_le32(mode);
948 fc->xattr_version = cpu_to_le64(xattr_version);
950 msg->middle = ceph_buffer_get(xattrs_buf);
951 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
952 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
955 ceph_con_send(&session->s_con, msg);
960 * Queue cap releases when an inode is dropped from our
963 void ceph_queue_caps_release(struct inode *inode)
965 struct ceph_inode_info *ci = ceph_inode(inode);
968 spin_lock(&inode->i_lock);
969 p = rb_first(&ci->i_caps);
971 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
972 struct ceph_mds_session *session = cap->session;
973 struct ceph_msg *msg;
974 struct ceph_mds_cap_release *head;
975 struct ceph_mds_cap_item *item;
977 spin_lock(&session->s_cap_lock);
978 BUG_ON(!session->s_num_cap_releases);
979 msg = list_first_entry(&session->s_cap_releases,
980 struct ceph_msg, list_head);
982 dout(" adding %p release to mds%d msg %p (%d left)\n",
983 inode, session->s_mds, msg, session->s_num_cap_releases);
985 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE);
986 head = msg->front.iov_base;
987 head->num = cpu_to_le32(le32_to_cpu(head->num) + 1);
988 item = msg->front.iov_base + msg->front.iov_len;
989 item->ino = cpu_to_le64(ceph_ino(inode));
990 item->cap_id = cpu_to_le64(cap->cap_id);
991 item->migrate_seq = cpu_to_le32(cap->mseq);
992 item->seq = cpu_to_le32(cap->issue_seq);
994 session->s_num_cap_releases--;
996 msg->front.iov_len += sizeof(*item);
997 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
998 dout(" release msg %p full\n", msg);
999 list_move_tail(&msg->list_head,
1000 &session->s_cap_releases_done);
1002 dout(" release msg %p at %d/%d (%d)\n", msg,
1003 (int)le32_to_cpu(head->num),
1004 (int)CEPH_CAPS_PER_RELEASE,
1005 (int)msg->front.iov_len);
1007 spin_unlock(&session->s_cap_lock);
1009 __ceph_remove_cap(cap, NULL);
1012 spin_unlock(&inode->i_lock);
1016 * Send a cap msg on the given inode. Update our caps state, then
1017 * drop i_lock and send the message.
1019 * Make note of max_size reported/requested from mds, revoked caps
1020 * that have now been implemented.
1022 * Make half-hearted attempt ot to invalidate page cache if we are
1023 * dropping RDCACHE. Note that this will leave behind locked pages
1024 * that we'll then need to deal with elsewhere.
1026 * Return non-zero if delayed release, or we experienced an error
1027 * such that the caller should requeue + retry later.
1029 * called with i_lock, then drops it.
1030 * caller should hold snap_rwsem (read), s_mutex.
1032 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1033 int op, int used, int want, int retain, int flushing,
1034 unsigned *pflush_tid)
1035 __releases(cap->ci->vfs_inode->i_lock)
1037 struct ceph_inode_info *ci = cap->ci;
1038 struct inode *inode = &ci->vfs_inode;
1039 u64 cap_id = cap->cap_id;
1040 int held = cap->issued | cap->implemented;
1041 int revoking = cap->implemented & ~cap->issued;
1042 int dropping = cap->issued & ~retain;
1044 u64 seq, issue_seq, mseq, time_warp_seq, follows;
1046 struct timespec mtime, atime;
1051 struct ceph_mds_session *session;
1052 u64 xattr_version = 0;
1058 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1059 inode, cap, cap->session,
1060 ceph_cap_string(held), ceph_cap_string(held & retain),
1061 ceph_cap_string(revoking));
1062 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1064 session = cap->session;
1066 /* don't release wanted unless we've waited a bit. */
1067 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1068 time_before(jiffies, ci->i_hold_caps_min)) {
1069 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1070 ceph_cap_string(cap->issued),
1071 ceph_cap_string(cap->issued & retain),
1072 ceph_cap_string(cap->mds_wanted),
1073 ceph_cap_string(want));
1074 want |= cap->mds_wanted;
1075 retain |= cap->issued;
1078 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1080 cap->issued &= retain; /* drop bits we don't want */
1081 if (cap->implemented & ~cap->issued) {
1083 * Wake up any waiters on wanted -> needed transition.
1084 * This is due to the weird transition from buffered
1085 * to sync IO... we need to flush dirty pages _before_
1086 * allowing sync writes to avoid reordering.
1090 cap->implemented &= cap->issued | used;
1091 cap->mds_wanted = want;
1095 * assign a tid for flush operations so we can avoid
1096 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark
1097 * clean type races. track latest tid for every bit
1098 * so we can handle flush AxFw, flush Fw, and have the
1099 * first ack clean Ax.
1101 flush_tid = ++ci->i_cap_flush_last_tid;
1103 *pflush_tid = flush_tid;
1104 dout(" cap_flush_tid %d\n", (int)flush_tid);
1105 for (i = 0; i < CEPH_CAP_BITS; i++)
1106 if (flushing & (1 << i))
1107 ci->i_cap_flush_tid[i] = flush_tid;
1110 keep = cap->implemented;
1112 issue_seq = cap->issue_seq;
1114 size = inode->i_size;
1115 ci->i_reported_size = size;
1116 max_size = ci->i_wanted_max_size;
1117 ci->i_requested_max_size = max_size;
1118 mtime = inode->i_mtime;
1119 atime = inode->i_atime;
1120 time_warp_seq = ci->i_time_warp_seq;
1121 follows = ci->i_snap_realm->cached_context->seq;
1124 mode = inode->i_mode;
1126 if (dropping & CEPH_CAP_XATTR_EXCL) {
1127 __ceph_build_xattrs_blob(ci);
1128 xattr_version = ci->i_xattrs.version + 1;
1131 spin_unlock(&inode->i_lock);
1133 if (dropping & CEPH_CAP_FILE_CACHE) {
1134 /* invalidate what we can */
1135 dout("invalidating pages on %p\n", inode);
1136 invalidate_mapping_pages(&inode->i_data, 0, -1);
1139 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1140 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq,
1141 size, max_size, &mtime, &atime, time_warp_seq,
1144 (flushing & CEPH_CAP_XATTR_EXCL) ? ci->i_xattrs.blob : NULL,
1147 dout("error sending cap msg, must requeue %p\n", inode);
1152 wake_up(&ci->i_cap_wq);
1158 * When a snapshot is taken, clients accumulate dirty metadata on
1159 * inodes with capabilities in ceph_cap_snaps to describe the file
1160 * state at the time the snapshot was taken. This must be flushed
1161 * asynchronously back to the MDS once sync writes complete and dirty
1162 * data is written out.
1164 * Called under i_lock. Takes s_mutex as needed.
1166 void __ceph_flush_snaps(struct ceph_inode_info *ci,
1167 struct ceph_mds_session **psession)
1169 struct inode *inode = &ci->vfs_inode;
1171 struct ceph_cap_snap *capsnap;
1173 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
1174 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1176 u64 next_follows = 0; /* keep track of how far we've gotten through the
1177 i_cap_snaps list, and skip these entries next time
1178 around to avoid an infinite loop */
1181 session = *psession;
1183 dout("__flush_snaps %p\n", inode);
1185 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1186 /* avoid an infiniute loop after retry */
1187 if (capsnap->follows < next_follows)
1190 * we need to wait for sync writes to complete and for dirty
1191 * pages to be written out.
1193 if (capsnap->dirty_pages || capsnap->writing)
1196 /* pick mds, take s_mutex */
1197 mds = __ceph_get_cap_mds(ci, &mseq);
1198 if (session && session->s_mds != mds) {
1199 dout("oops, wrong session %p mutex\n", session);
1200 mutex_unlock(&session->s_mutex);
1201 ceph_put_mds_session(session);
1205 spin_unlock(&inode->i_lock);
1206 mutex_lock(&mdsc->mutex);
1207 session = __ceph_lookup_mds_session(mdsc, mds);
1208 mutex_unlock(&mdsc->mutex);
1210 dout("inverting session/ino locks on %p\n",
1212 mutex_lock(&session->s_mutex);
1215 * if session == NULL, we raced against a cap
1216 * deletion. retry, and we'll get a better
1217 * @mds value next time.
1219 spin_lock(&inode->i_lock);
1223 capsnap->flush_tid = ++ci->i_cap_flush_last_tid;
1224 atomic_inc(&capsnap->nref);
1225 if (!list_empty(&capsnap->flushing_item))
1226 list_del_init(&capsnap->flushing_item);
1227 list_add_tail(&capsnap->flushing_item,
1228 &session->s_cap_snaps_flushing);
1229 spin_unlock(&inode->i_lock);
1231 dout("flush_snaps %p cap_snap %p follows %lld size %llu\n",
1232 inode, capsnap, next_follows, capsnap->size);
1233 send_cap_msg(session, ceph_vino(inode).ino, 0,
1234 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1235 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq,
1237 &capsnap->mtime, &capsnap->atime,
1238 capsnap->time_warp_seq,
1239 capsnap->uid, capsnap->gid, capsnap->mode,
1243 next_follows = capsnap->follows + 1;
1244 ceph_put_cap_snap(capsnap);
1246 spin_lock(&inode->i_lock);
1250 /* we flushed them all; remove this inode from the queue */
1251 spin_lock(&mdsc->snap_flush_lock);
1252 list_del_init(&ci->i_snap_flush_item);
1253 spin_unlock(&mdsc->snap_flush_lock);
1256 *psession = session;
1258 mutex_unlock(&session->s_mutex);
1259 ceph_put_mds_session(session);
1263 static void ceph_flush_snaps(struct ceph_inode_info *ci)
1265 struct inode *inode = &ci->vfs_inode;
1267 spin_lock(&inode->i_lock);
1268 __ceph_flush_snaps(ci, NULL);
1269 spin_unlock(&inode->i_lock);
1273 * Mark caps dirty. If inode is newly dirty, add to the global dirty
1276 void __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1278 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc;
1279 struct inode *inode = &ci->vfs_inode;
1280 int was = ci->i_dirty_caps;
1283 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1284 ceph_cap_string(mask), ceph_cap_string(was),
1285 ceph_cap_string(was | mask));
1286 ci->i_dirty_caps |= mask;
1288 dout(" inode %p now dirty\n", &ci->vfs_inode);
1289 BUG_ON(!list_empty(&ci->i_dirty_item));
1290 spin_lock(&mdsc->cap_dirty_lock);
1291 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1292 spin_unlock(&mdsc->cap_dirty_lock);
1293 if (ci->i_flushing_caps == 0) {
1295 dirty |= I_DIRTY_SYNC;
1298 BUG_ON(list_empty(&ci->i_dirty_item));
1299 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1300 (mask & CEPH_CAP_FILE_BUFFER))
1301 dirty |= I_DIRTY_DATASYNC;
1303 __mark_inode_dirty(inode, dirty);
1304 __cap_delay_requeue(mdsc, ci);
1308 * Add dirty inode to the flushing list. Assigned a seq number so we
1309 * can wait for caps to flush without starving.
1311 * Called under i_lock.
1313 static int __mark_caps_flushing(struct inode *inode,
1314 struct ceph_mds_session *session)
1316 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1317 struct ceph_inode_info *ci = ceph_inode(inode);
1320 BUG_ON(ci->i_dirty_caps == 0);
1321 BUG_ON(list_empty(&ci->i_dirty_item));
1323 flushing = ci->i_dirty_caps;
1324 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1325 ceph_cap_string(flushing),
1326 ceph_cap_string(ci->i_flushing_caps),
1327 ceph_cap_string(ci->i_flushing_caps | flushing));
1328 ci->i_flushing_caps |= flushing;
1329 ci->i_dirty_caps = 0;
1330 dout(" inode %p now !dirty\n", inode);
1332 spin_lock(&mdsc->cap_dirty_lock);
1333 list_del_init(&ci->i_dirty_item);
1335 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq;
1336 if (list_empty(&ci->i_flushing_item)) {
1337 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1338 mdsc->num_cap_flushing++;
1339 dout(" inode %p now flushing seq %lld\n", inode,
1340 ci->i_cap_flush_seq);
1342 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1343 dout(" inode %p now flushing (more) seq %lld\n", inode,
1344 ci->i_cap_flush_seq);
1346 spin_unlock(&mdsc->cap_dirty_lock);
1352 * Swiss army knife function to examine currently used and wanted
1353 * versus held caps. Release, flush, ack revoked caps to mds as
1356 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1357 * cap release further.
1358 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1359 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1362 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1363 struct ceph_mds_session *session)
1365 struct ceph_client *client = ceph_inode_to_client(&ci->vfs_inode);
1366 struct ceph_mds_client *mdsc = &client->mdsc;
1367 struct inode *inode = &ci->vfs_inode;
1368 struct ceph_cap *cap;
1369 int file_wanted, used;
1370 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1371 int drop_session_lock = session ? 0 : 1;
1372 int want, retain, revoking, flushing = 0;
1373 int mds = -1; /* keep track of how far we've gone through i_caps list
1374 to avoid an infinite loop on retry */
1376 int tried_invalidate = 0;
1377 int delayed = 0, sent = 0, force_requeue = 0, num;
1378 int is_delayed = flags & CHECK_CAPS_NODELAY;
1380 /* if we are unmounting, flush any unused caps immediately. */
1384 spin_lock(&inode->i_lock);
1386 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1387 flags |= CHECK_CAPS_FLUSH;
1389 /* flush snaps first time around only */
1390 if (!list_empty(&ci->i_cap_snaps))
1391 __ceph_flush_snaps(ci, &session);
1394 spin_lock(&inode->i_lock);
1396 file_wanted = __ceph_caps_file_wanted(ci);
1397 used = __ceph_caps_used(ci);
1398 want = file_wanted | used;
1400 retain = want | CEPH_CAP_PIN;
1401 if (!mdsc->stopping && inode->i_nlink > 0) {
1403 retain |= CEPH_CAP_ANY; /* be greedy */
1405 retain |= CEPH_CAP_ANY_SHARED;
1407 * keep RD only if we didn't have the file open RW,
1408 * because then the mds would revoke it anyway to
1409 * journal max_size=0.
1411 if (ci->i_max_size == 0)
1412 retain |= CEPH_CAP_ANY_RD;
1416 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1417 " issued %s retain %s %s%s%s\n", inode,
1418 ceph_cap_string(file_wanted),
1419 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1420 ceph_cap_string(ci->i_flushing_caps),
1421 ceph_cap_string(__ceph_caps_issued(ci, NULL)),
1422 ceph_cap_string(retain),
1423 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1424 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1425 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1428 * If we no longer need to hold onto old our caps, and we may
1429 * have cached pages, but don't want them, then try to invalidate.
1430 * If we fail, it's because pages are locked.... try again later.
1432 if ((!is_delayed || mdsc->stopping) &&
1433 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */
1434 ci->i_rdcache_gen && /* may have cached pages */
1435 file_wanted == 0 && /* no open files */
1436 !ci->i_truncate_pending &&
1437 !tried_invalidate) {
1438 u32 invalidating_gen = ci->i_rdcache_gen;
1441 dout("check_caps trying to invalidate on %p\n", inode);
1442 spin_unlock(&inode->i_lock);
1443 ret = invalidate_inode_pages2(&inode->i_data);
1444 spin_lock(&inode->i_lock);
1445 if (ret == 0 && invalidating_gen == ci->i_rdcache_gen) {
1447 ci->i_rdcache_gen = 0;
1448 ci->i_rdcache_revoking = 0;
1450 dout("check_caps failed to invalidate pages\n");
1451 /* we failed to invalidate pages. check these
1452 caps again later. */
1454 __cap_set_timeouts(mdsc, ci);
1456 tried_invalidate = 1;
1461 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1462 cap = rb_entry(p, struct ceph_cap, ci_node);
1465 /* avoid looping forever */
1466 if (mds >= cap->mds ||
1467 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1470 /* NOTE: no side-effects allowed, until we take s_mutex */
1472 revoking = cap->implemented & ~cap->issued;
1474 dout("mds%d revoking %s\n", cap->mds,
1475 ceph_cap_string(revoking));
1477 if (cap == ci->i_auth_cap &&
1478 (cap->issued & CEPH_CAP_FILE_WR)) {
1479 /* request larger max_size from MDS? */
1480 if (ci->i_wanted_max_size > ci->i_max_size &&
1481 ci->i_wanted_max_size > ci->i_requested_max_size) {
1482 dout("requesting new max_size\n");
1486 /* approaching file_max? */
1487 if ((inode->i_size << 1) >= ci->i_max_size &&
1488 (ci->i_reported_size << 1) < ci->i_max_size) {
1489 dout("i_size approaching max_size\n");
1493 /* flush anything dirty? */
1494 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1496 dout("flushing dirty caps\n");
1500 /* completed revocation? going down and there are no caps? */
1501 if (revoking && (revoking & used) == 0) {
1502 dout("completed revocation of %s\n",
1503 ceph_cap_string(cap->implemented & ~cap->issued));
1507 /* want more caps from mds? */
1508 if (want & ~(cap->mds_wanted | cap->issued))
1511 /* things we might delay */
1512 if ((cap->issued & ~retain) == 0 &&
1513 cap->mds_wanted == want)
1514 continue; /* nope, all good */
1520 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1521 time_before(jiffies, ci->i_hold_caps_max)) {
1522 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1523 ceph_cap_string(cap->issued),
1524 ceph_cap_string(cap->issued & retain),
1525 ceph_cap_string(cap->mds_wanted),
1526 ceph_cap_string(want));
1532 if (session && session != cap->session) {
1533 dout("oops, wrong session %p mutex\n", session);
1534 mutex_unlock(&session->s_mutex);
1538 session = cap->session;
1539 if (mutex_trylock(&session->s_mutex) == 0) {
1540 dout("inverting session/ino locks on %p\n",
1542 spin_unlock(&inode->i_lock);
1543 if (took_snap_rwsem) {
1544 up_read(&mdsc->snap_rwsem);
1545 took_snap_rwsem = 0;
1547 mutex_lock(&session->s_mutex);
1551 /* take snap_rwsem after session mutex */
1552 if (!took_snap_rwsem) {
1553 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1554 dout("inverting snap/in locks on %p\n",
1556 spin_unlock(&inode->i_lock);
1557 down_read(&mdsc->snap_rwsem);
1558 took_snap_rwsem = 1;
1561 took_snap_rwsem = 1;
1564 if (cap == ci->i_auth_cap && ci->i_dirty_caps)
1565 flushing = __mark_caps_flushing(inode, session);
1567 mds = cap->mds; /* remember mds, so we don't repeat */
1570 /* __send_cap drops i_lock */
1571 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, used, want,
1572 retain, flushing, NULL);
1573 goto retry; /* retake i_lock and restart our cap scan. */
1577 * Reschedule delayed caps release if we delayed anything,
1580 if (delayed && is_delayed)
1581 force_requeue = 1; /* __send_cap delayed release; requeue */
1582 if (!delayed && !is_delayed)
1583 __cap_delay_cancel(mdsc, ci);
1584 else if (!is_delayed || force_requeue)
1585 __cap_delay_requeue(mdsc, ci);
1587 spin_unlock(&inode->i_lock);
1589 if (session && drop_session_lock)
1590 mutex_unlock(&session->s_mutex);
1591 if (took_snap_rwsem)
1592 up_read(&mdsc->snap_rwsem);
1596 * Try to flush dirty caps back to the auth mds.
1598 static int try_flush_caps(struct inode *inode, struct ceph_mds_session *session,
1599 unsigned *flush_tid)
1601 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1602 struct ceph_inode_info *ci = ceph_inode(inode);
1603 int unlock_session = session ? 0 : 1;
1607 spin_lock(&inode->i_lock);
1608 if (ci->i_dirty_caps && ci->i_auth_cap) {
1609 struct ceph_cap *cap = ci->i_auth_cap;
1610 int used = __ceph_caps_used(ci);
1611 int want = __ceph_caps_wanted(ci);
1615 spin_unlock(&inode->i_lock);
1616 session = cap->session;
1617 mutex_lock(&session->s_mutex);
1620 BUG_ON(session != cap->session);
1621 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1624 flushing = __mark_caps_flushing(inode, session);
1626 /* __send_cap drops i_lock */
1627 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1628 cap->issued | cap->implemented, flushing,
1633 spin_lock(&inode->i_lock);
1634 __cap_delay_requeue(mdsc, ci);
1637 spin_unlock(&inode->i_lock);
1639 if (session && unlock_session)
1640 mutex_unlock(&session->s_mutex);
1645 * Return true if we've flushed caps through the given flush_tid.
1647 static int caps_are_flushed(struct inode *inode, unsigned tid)
1649 struct ceph_inode_info *ci = ceph_inode(inode);
1650 int dirty, i, ret = 1;
1652 spin_lock(&inode->i_lock);
1653 dirty = __ceph_caps_dirty(ci);
1654 for (i = 0; i < CEPH_CAP_BITS; i++)
1655 if ((ci->i_flushing_caps & (1 << i)) &&
1656 ci->i_cap_flush_tid[i] <= tid) {
1657 /* still flushing this bit */
1661 spin_unlock(&inode->i_lock);
1666 * Wait on any unsafe replies for the given inode. First wait on the
1667 * newest request, and make that the upper bound. Then, if there are
1668 * more requests, keep waiting on the oldest as long as it is still older
1669 * than the original request.
1671 static void sync_write_wait(struct inode *inode)
1673 struct ceph_inode_info *ci = ceph_inode(inode);
1674 struct list_head *head = &ci->i_unsafe_writes;
1675 struct ceph_osd_request *req;
1678 spin_lock(&ci->i_unsafe_lock);
1679 if (list_empty(head))
1682 /* set upper bound as _last_ entry in chain */
1683 req = list_entry(head->prev, struct ceph_osd_request,
1685 last_tid = req->r_tid;
1688 ceph_osdc_get_request(req);
1689 spin_unlock(&ci->i_unsafe_lock);
1690 dout("sync_write_wait on tid %llu (until %llu)\n",
1691 req->r_tid, last_tid);
1692 wait_for_completion(&req->r_safe_completion);
1693 spin_lock(&ci->i_unsafe_lock);
1694 ceph_osdc_put_request(req);
1697 * from here on look at first entry in chain, since we
1698 * only want to wait for anything older than last_tid
1700 if (list_empty(head))
1702 req = list_entry(head->next, struct ceph_osd_request,
1704 } while (req->r_tid < last_tid);
1706 spin_unlock(&ci->i_unsafe_lock);
1709 int ceph_fsync(struct file *file, struct dentry *dentry, int datasync)
1711 struct inode *inode = dentry->d_inode;
1712 struct ceph_inode_info *ci = ceph_inode(inode);
1717 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
1718 sync_write_wait(inode);
1720 ret = filemap_write_and_wait(inode->i_mapping);
1724 dirty = try_flush_caps(inode, NULL, &flush_tid);
1725 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1728 * only wait on non-file metadata writeback (the mds
1729 * can recover size and mtime, so we don't need to
1732 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1733 dout("fsync waiting for flush_tid %u\n", flush_tid);
1734 ret = wait_event_interruptible(ci->i_cap_wq,
1735 caps_are_flushed(inode, flush_tid));
1738 dout("fsync %p%s done\n", inode, datasync ? " datasync" : "");
1743 * Flush any dirty caps back to the mds. If we aren't asked to wait,
1744 * queue inode for flush but don't do so immediately, because we can
1745 * get by with fewer MDS messages if we wait for data writeback to
1748 int ceph_write_inode(struct inode *inode, int wait)
1750 struct ceph_inode_info *ci = ceph_inode(inode);
1755 dout("write_inode %p wait=%d\n", inode, wait);
1757 dirty = try_flush_caps(inode, NULL, &flush_tid);
1759 err = wait_event_interruptible(ci->i_cap_wq,
1760 caps_are_flushed(inode, flush_tid));
1762 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1764 spin_lock(&inode->i_lock);
1765 if (__ceph_caps_dirty(ci))
1766 __cap_delay_requeue_front(mdsc, ci);
1767 spin_unlock(&inode->i_lock);
1773 * After a recovering MDS goes active, we need to resend any caps
1776 * Caller holds session->s_mutex.
1778 static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
1779 struct ceph_mds_session *session)
1781 struct ceph_cap_snap *capsnap;
1783 dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
1784 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
1786 struct ceph_inode_info *ci = capsnap->ci;
1787 struct inode *inode = &ci->vfs_inode;
1788 struct ceph_cap *cap;
1790 spin_lock(&inode->i_lock);
1791 cap = ci->i_auth_cap;
1792 if (cap && cap->session == session) {
1793 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
1795 __ceph_flush_snaps(ci, &session);
1797 pr_err("%p auth cap %p not mds%d ???\n", inode,
1798 cap, session->s_mds);
1799 spin_unlock(&inode->i_lock);
1804 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
1805 struct ceph_mds_session *session)
1807 struct ceph_inode_info *ci;
1809 kick_flushing_capsnaps(mdsc, session);
1811 dout("kick_flushing_caps mds%d\n", session->s_mds);
1812 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
1813 struct inode *inode = &ci->vfs_inode;
1814 struct ceph_cap *cap;
1817 spin_lock(&inode->i_lock);
1818 cap = ci->i_auth_cap;
1819 if (cap && cap->session == session) {
1820 dout("kick_flushing_caps %p cap %p %s\n", inode,
1821 cap, ceph_cap_string(ci->i_flushing_caps));
1822 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1823 __ceph_caps_used(ci),
1824 __ceph_caps_wanted(ci),
1825 cap->issued | cap->implemented,
1826 ci->i_flushing_caps, NULL);
1828 spin_lock(&inode->i_lock);
1829 __cap_delay_requeue(mdsc, ci);
1830 spin_unlock(&inode->i_lock);
1833 pr_err("%p auth cap %p not mds%d ???\n", inode,
1834 cap, session->s_mds);
1835 spin_unlock(&inode->i_lock);
1842 * Take references to capabilities we hold, so that we don't release
1843 * them to the MDS prematurely.
1845 * Protected by i_lock.
1847 static void __take_cap_refs(struct ceph_inode_info *ci, int got)
1849 if (got & CEPH_CAP_PIN)
1851 if (got & CEPH_CAP_FILE_RD)
1853 if (got & CEPH_CAP_FILE_CACHE)
1854 ci->i_rdcache_ref++;
1855 if (got & CEPH_CAP_FILE_WR)
1857 if (got & CEPH_CAP_FILE_BUFFER) {
1858 if (ci->i_wrbuffer_ref == 0)
1859 igrab(&ci->vfs_inode);
1860 ci->i_wrbuffer_ref++;
1861 dout("__take_cap_refs %p wrbuffer %d -> %d (?)\n",
1862 &ci->vfs_inode, ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref);
1867 * Try to grab cap references. Specify those refs we @want, and the
1868 * minimal set we @need. Also include the larger offset we are writing
1869 * to (when applicable), and check against max_size here as well.
1870 * Note that caller is responsible for ensuring max_size increases are
1871 * requested from the MDS.
1873 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
1874 int *got, loff_t endoff, int *check_max, int *err)
1876 struct inode *inode = &ci->vfs_inode;
1878 int have, implemented;
1880 dout("get_cap_refs %p need %s want %s\n", inode,
1881 ceph_cap_string(need), ceph_cap_string(want));
1882 spin_lock(&inode->i_lock);
1884 /* make sure we _have_ some caps! */
1885 if (!__ceph_is_any_caps(ci)) {
1886 dout("get_cap_refs %p no real caps\n", inode);
1892 if (need & CEPH_CAP_FILE_WR) {
1893 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
1894 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
1895 inode, endoff, ci->i_max_size);
1896 if (endoff > ci->i_wanted_max_size) {
1903 * If a sync write is in progress, we must wait, so that we
1904 * can get a final snapshot value for size+mtime.
1906 if (__ceph_have_pending_cap_snap(ci)) {
1907 dout("get_cap_refs %p cap_snap_pending\n", inode);
1911 have = __ceph_caps_issued(ci, &implemented);
1914 * disallow writes while a truncate is pending
1916 if (ci->i_truncate_pending)
1917 have &= ~CEPH_CAP_FILE_WR;
1919 if ((have & need) == need) {
1921 * Look at (implemented & ~have & not) so that we keep waiting
1922 * on transition from wanted -> needed caps. This is needed
1923 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
1924 * going before a prior buffered writeback happens.
1926 int not = want & ~(have & need);
1927 int revoking = implemented & ~have;
1928 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
1929 inode, ceph_cap_string(have), ceph_cap_string(not),
1930 ceph_cap_string(revoking));
1931 if ((revoking & not) == 0) {
1932 *got = need | (have & want);
1933 __take_cap_refs(ci, *got);
1937 dout("get_cap_refs %p have %s needed %s\n", inode,
1938 ceph_cap_string(have), ceph_cap_string(need));
1941 spin_unlock(&inode->i_lock);
1942 dout("get_cap_refs %p ret %d got %s\n", inode,
1943 ret, ceph_cap_string(*got));
1948 * Check the offset we are writing up to against our current
1949 * max_size. If necessary, tell the MDS we want to write to
1952 static void check_max_size(struct inode *inode, loff_t endoff)
1954 struct ceph_inode_info *ci = ceph_inode(inode);
1957 /* do we need to explicitly request a larger max_size? */
1958 spin_lock(&inode->i_lock);
1959 if ((endoff >= ci->i_max_size ||
1960 endoff > (inode->i_size << 1)) &&
1961 endoff > ci->i_wanted_max_size) {
1962 dout("write %p at large endoff %llu, req max_size\n",
1964 ci->i_wanted_max_size = endoff;
1967 spin_unlock(&inode->i_lock);
1969 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
1973 * Wait for caps, and take cap references. If we can't get a WR cap
1974 * due to a small max_size, make sure we check_max_size (and possibly
1975 * ask the mds) so we don't get hung up indefinitely.
1977 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, int *got,
1980 int check_max, ret, err;
1984 check_max_size(&ci->vfs_inode, endoff);
1987 ret = wait_event_interruptible(ci->i_cap_wq,
1988 try_get_cap_refs(ci, need, want,
1999 * Take cap refs. Caller must already know we hold at least one ref
2000 * on the caps in question or we don't know this is safe.
2002 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2004 spin_lock(&ci->vfs_inode.i_lock);
2005 __take_cap_refs(ci, caps);
2006 spin_unlock(&ci->vfs_inode.i_lock);
2012 * If we released the last ref on any given cap, call ceph_check_caps
2013 * to release (or schedule a release).
2015 * If we are releasing a WR cap (from a sync write), finalize any affected
2016 * cap_snap, and wake up any waiters.
2018 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2020 struct inode *inode = &ci->vfs_inode;
2021 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2022 struct ceph_cap_snap *capsnap;
2024 spin_lock(&inode->i_lock);
2025 if (had & CEPH_CAP_PIN)
2027 if (had & CEPH_CAP_FILE_RD)
2028 if (--ci->i_rd_ref == 0)
2030 if (had & CEPH_CAP_FILE_CACHE)
2031 if (--ci->i_rdcache_ref == 0)
2033 if (had & CEPH_CAP_FILE_BUFFER) {
2034 if (--ci->i_wrbuffer_ref == 0) {
2038 dout("put_cap_refs %p wrbuffer %d -> %d (?)\n",
2039 inode, ci->i_wrbuffer_ref+1, ci->i_wrbuffer_ref);
2041 if (had & CEPH_CAP_FILE_WR)
2042 if (--ci->i_wr_ref == 0) {
2044 if (!list_empty(&ci->i_cap_snaps)) {
2045 capsnap = list_first_entry(&ci->i_cap_snaps,
2046 struct ceph_cap_snap,
2048 if (capsnap->writing) {
2049 capsnap->writing = 0;
2051 __ceph_finish_cap_snap(ci,
2057 spin_unlock(&inode->i_lock);
2059 dout("put_cap_refs %p had %s %s\n", inode, ceph_cap_string(had),
2060 last ? "last" : "");
2062 if (last && !flushsnaps)
2063 ceph_check_caps(ci, 0, NULL);
2064 else if (flushsnaps)
2065 ceph_flush_snaps(ci);
2067 wake_up(&ci->i_cap_wq);
2073 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2074 * context. Adjust per-snap dirty page accounting as appropriate.
2075 * Once all dirty data for a cap_snap is flushed, flush snapped file
2076 * metadata back to the MDS. If we dropped the last ref, call
2079 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2080 struct ceph_snap_context *snapc)
2082 struct inode *inode = &ci->vfs_inode;
2086 struct ceph_cap_snap *capsnap = NULL;
2088 spin_lock(&inode->i_lock);
2089 ci->i_wrbuffer_ref -= nr;
2090 last = !ci->i_wrbuffer_ref;
2092 if (ci->i_head_snapc == snapc) {
2093 ci->i_wrbuffer_ref_head -= nr;
2094 if (!ci->i_wrbuffer_ref_head) {
2095 ceph_put_snap_context(ci->i_head_snapc);
2096 ci->i_head_snapc = NULL;
2098 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2100 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2101 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2102 last ? " LAST" : "");
2104 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2105 if (capsnap->context == snapc) {
2107 capsnap->dirty_pages -= nr;
2108 last_snap = !capsnap->dirty_pages;
2113 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2114 " snap %lld %d/%d -> %d/%d %s%s\n",
2115 inode, capsnap, capsnap->context->seq,
2116 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2117 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2118 last ? " (wrbuffer last)" : "",
2119 last_snap ? " (capsnap last)" : "");
2122 spin_unlock(&inode->i_lock);
2125 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2127 } else if (last_snap) {
2128 ceph_flush_snaps(ci);
2129 wake_up(&ci->i_cap_wq);
2134 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2135 * actually be a revocation if it specifies a smaller cap set.)
2137 * caller holds s_mutex.
2140 * 1 - check_caps on auth cap only (writeback)
2141 * 2 - check_caps (ack revoke)
2143 static int handle_cap_grant(struct inode *inode, struct ceph_mds_caps *grant,
2144 struct ceph_mds_session *session,
2145 struct ceph_cap *cap,
2146 struct ceph_buffer *xattr_buf)
2147 __releases(inode->i_lock)
2150 struct ceph_inode_info *ci = ceph_inode(inode);
2151 int mds = session->s_mds;
2152 int seq = le32_to_cpu(grant->seq);
2153 int newcaps = le32_to_cpu(grant->caps);
2154 int issued, implemented, used, wanted, dirty;
2155 u64 size = le64_to_cpu(grant->size);
2156 u64 max_size = le64_to_cpu(grant->max_size);
2157 struct timespec mtime, atime, ctime;
2161 int revoked_rdcache = 0;
2162 int invalidate_async = 0;
2163 int tried_invalidate = 0;
2166 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2167 inode, cap, mds, seq, ceph_cap_string(newcaps));
2168 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2172 * If CACHE is being revoked, and we have no dirty buffers,
2173 * try to invalidate (once). (If there are dirty buffers, we
2174 * will invalidate _after_ writeback.)
2177 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2178 !ci->i_wrbuffer_ref && !tried_invalidate) {
2179 dout("CACHE invalidation\n");
2180 spin_unlock(&inode->i_lock);
2181 tried_invalidate = 1;
2183 ret = invalidate_inode_pages2(&inode->i_data);
2184 spin_lock(&inode->i_lock);
2186 /* there were locked pages.. invalidate later
2187 in a separate thread. */
2188 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2189 invalidate_async = 1;
2190 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2193 /* we successfully invalidated those pages */
2194 revoked_rdcache = 1;
2195 ci->i_rdcache_gen = 0;
2196 ci->i_rdcache_revoking = 0;
2201 /* side effects now are allowed */
2203 issued = __ceph_caps_issued(ci, &implemented);
2204 issued |= implemented | __ceph_caps_dirty(ci);
2206 cap->cap_gen = session->s_cap_gen;
2208 __check_cap_issue(ci, cap, newcaps);
2210 if ((issued & CEPH_CAP_AUTH_EXCL) == 0) {
2211 inode->i_mode = le32_to_cpu(grant->mode);
2212 inode->i_uid = le32_to_cpu(grant->uid);
2213 inode->i_gid = le32_to_cpu(grant->gid);
2214 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2215 inode->i_uid, inode->i_gid);
2218 if ((issued & CEPH_CAP_LINK_EXCL) == 0)
2219 inode->i_nlink = le32_to_cpu(grant->nlink);
2221 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2222 int len = le32_to_cpu(grant->xattr_len);
2223 u64 version = le64_to_cpu(grant->xattr_version);
2225 if (version > ci->i_xattrs.version) {
2226 dout(" got new xattrs v%llu on %p len %d\n",
2227 version, inode, len);
2228 if (ci->i_xattrs.blob)
2229 ceph_buffer_put(ci->i_xattrs.blob);
2230 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2231 ci->i_xattrs.version = version;
2235 /* size/ctime/mtime/atime? */
2236 ceph_fill_file_size(inode, issued,
2237 le32_to_cpu(grant->truncate_seq),
2238 le64_to_cpu(grant->truncate_size), size);
2239 ceph_decode_timespec(&mtime, &grant->mtime);
2240 ceph_decode_timespec(&atime, &grant->atime);
2241 ceph_decode_timespec(&ctime, &grant->ctime);
2242 ceph_fill_file_time(inode, issued,
2243 le32_to_cpu(grant->time_warp_seq), &ctime, &mtime,
2246 /* max size increase? */
2247 if (max_size != ci->i_max_size) {
2248 dout("max_size %lld -> %llu\n", ci->i_max_size, max_size);
2249 ci->i_max_size = max_size;
2250 if (max_size >= ci->i_wanted_max_size) {
2251 ci->i_wanted_max_size = 0; /* reset */
2252 ci->i_requested_max_size = 0;
2257 /* check cap bits */
2258 wanted = __ceph_caps_wanted(ci);
2259 used = __ceph_caps_used(ci);
2260 dirty = __ceph_caps_dirty(ci);
2261 dout(" my wanted = %s, used = %s, dirty %s\n",
2262 ceph_cap_string(wanted),
2263 ceph_cap_string(used),
2264 ceph_cap_string(dirty));
2265 if (wanted != le32_to_cpu(grant->wanted)) {
2266 dout("mds wanted %s -> %s\n",
2267 ceph_cap_string(le32_to_cpu(grant->wanted)),
2268 ceph_cap_string(wanted));
2269 grant->wanted = cpu_to_le32(wanted);
2274 /* file layout may have changed */
2275 ci->i_layout = grant->layout;
2277 /* revocation, grant, or no-op? */
2278 if (cap->issued & ~newcaps) {
2279 dout("revocation: %s -> %s\n", ceph_cap_string(cap->issued),
2280 ceph_cap_string(newcaps));
2281 if ((used & ~newcaps) & CEPH_CAP_FILE_BUFFER)
2282 writeback = 1; /* will delay ack */
2283 else if (dirty & ~newcaps)
2284 reply = 1; /* initiate writeback in check_caps */
2285 else if (((used & ~newcaps) & CEPH_CAP_FILE_CACHE) == 0 ||
2287 reply = 2; /* send revoke ack in check_caps */
2288 cap->issued = newcaps;
2289 } else if (cap->issued == newcaps) {
2290 dout("caps unchanged: %s -> %s\n",
2291 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2293 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2294 ceph_cap_string(newcaps));
2295 cap->issued = newcaps;
2296 cap->implemented |= newcaps; /* add bits only, to
2297 * avoid stepping on a
2298 * pending revocation */
2302 spin_unlock(&inode->i_lock);
2305 * queue inode for writeback: we can't actually call
2306 * filemap_write_and_wait, etc. from message handler
2309 dout("queueing %p for writeback\n", inode);
2310 if (ceph_queue_writeback(inode))
2313 if (invalidate_async) {
2314 dout("queueing %p for page invalidation\n", inode);
2315 if (ceph_queue_page_invalidation(inode))
2319 wake_up(&ci->i_cap_wq);
2324 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2325 * MDS has been safely committed.
2327 static void handle_cap_flush_ack(struct inode *inode,
2328 struct ceph_mds_caps *m,
2329 struct ceph_mds_session *session,
2330 struct ceph_cap *cap)
2331 __releases(inode->i_lock)
2333 struct ceph_inode_info *ci = ceph_inode(inode);
2334 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
2335 unsigned seq = le32_to_cpu(m->seq);
2336 int dirty = le32_to_cpu(m->dirty);
2338 u64 flush_tid = le64_to_cpu(m->client_tid);
2342 for (i = 0; i < CEPH_CAP_BITS; i++)
2343 if ((dirty & (1 << i)) &&
2344 flush_tid == ci->i_cap_flush_tid[i])
2347 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2348 " flushing %s -> %s\n",
2349 inode, session->s_mds, seq, ceph_cap_string(dirty),
2350 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2351 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2353 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned))
2356 ci->i_flushing_caps &= ~cleaned;
2358 spin_lock(&mdsc->cap_dirty_lock);
2359 if (ci->i_flushing_caps == 0) {
2360 list_del_init(&ci->i_flushing_item);
2361 if (!list_empty(&session->s_cap_flushing))
2362 dout(" mds%d still flushing cap on %p\n",
2364 &list_entry(session->s_cap_flushing.next,
2365 struct ceph_inode_info,
2366 i_flushing_item)->vfs_inode);
2367 mdsc->num_cap_flushing--;
2368 wake_up(&mdsc->cap_flushing_wq);
2369 dout(" inode %p now !flushing\n", inode);
2371 if (ci->i_dirty_caps == 0) {
2372 dout(" inode %p now clean\n", inode);
2373 BUG_ON(!list_empty(&ci->i_dirty_item));
2376 BUG_ON(list_empty(&ci->i_dirty_item));
2379 spin_unlock(&mdsc->cap_dirty_lock);
2380 wake_up(&ci->i_cap_wq);
2383 spin_unlock(&inode->i_lock);
2389 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
2390 * throw away our cap_snap.
2392 * Caller hold s_mutex.
2394 static void handle_cap_flushsnap_ack(struct inode *inode,
2395 struct ceph_mds_caps *m,
2396 struct ceph_mds_session *session)
2398 struct ceph_inode_info *ci = ceph_inode(inode);
2399 u64 follows = le64_to_cpu(m->snap_follows);
2400 u64 flush_tid = le64_to_cpu(m->client_tid);
2401 struct ceph_cap_snap *capsnap;
2404 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
2405 inode, ci, session->s_mds, follows);
2407 spin_lock(&inode->i_lock);
2408 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2409 if (capsnap->follows == follows) {
2410 if (capsnap->flush_tid != flush_tid) {
2411 dout(" cap_snap %p follows %lld tid %lld !="
2412 " %lld\n", capsnap, follows,
2413 flush_tid, capsnap->flush_tid);
2416 WARN_ON(capsnap->dirty_pages || capsnap->writing);
2417 dout(" removing cap_snap %p follows %lld\n",
2419 ceph_put_snap_context(capsnap->context);
2420 list_del(&capsnap->ci_item);
2421 list_del(&capsnap->flushing_item);
2422 ceph_put_cap_snap(capsnap);
2426 dout(" skipping cap_snap %p follows %lld\n",
2427 capsnap, capsnap->follows);
2430 spin_unlock(&inode->i_lock);
2436 * Handle TRUNC from MDS, indicating file truncation.
2438 * caller hold s_mutex.
2440 static void handle_cap_trunc(struct inode *inode,
2441 struct ceph_mds_caps *trunc,
2442 struct ceph_mds_session *session)
2443 __releases(inode->i_lock)
2445 struct ceph_inode_info *ci = ceph_inode(inode);
2446 int mds = session->s_mds;
2447 int seq = le32_to_cpu(trunc->seq);
2448 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
2449 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
2450 u64 size = le64_to_cpu(trunc->size);
2451 int implemented = 0;
2452 int dirty = __ceph_caps_dirty(ci);
2453 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
2454 int queue_trunc = 0;
2456 issued |= implemented | dirty;
2458 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
2459 inode, mds, seq, truncate_size, truncate_seq);
2460 queue_trunc = ceph_fill_file_size(inode, issued,
2461 truncate_seq, truncate_size, size);
2462 spin_unlock(&inode->i_lock);
2465 if (queue_work(ceph_client(inode->i_sb)->trunc_wq,
2466 &ci->i_vmtruncate_work))
2471 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
2472 * different one. If we are the most recent migration we've seen (as
2473 * indicated by mseq), make note of the migrating cap bits for the
2474 * duration (until we see the corresponding IMPORT).
2476 * caller holds s_mutex
2478 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
2479 struct ceph_mds_session *session)
2481 struct ceph_inode_info *ci = ceph_inode(inode);
2482 int mds = session->s_mds;
2483 unsigned mseq = le32_to_cpu(ex->migrate_seq);
2484 struct ceph_cap *cap = NULL, *t;
2488 dout("handle_cap_export inode %p ci %p mds%d mseq %d\n",
2489 inode, ci, mds, mseq);
2491 spin_lock(&inode->i_lock);
2493 /* make sure we haven't seen a higher mseq */
2494 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2495 t = rb_entry(p, struct ceph_cap, ci_node);
2496 if (ceph_seq_cmp(t->mseq, mseq) > 0) {
2497 dout(" higher mseq on cap from mds%d\n",
2501 if (t->session->s_mds == mds)
2508 ci->i_cap_exporting_mds = mds;
2509 ci->i_cap_exporting_mseq = mseq;
2510 ci->i_cap_exporting_issued = cap->issued;
2512 __ceph_remove_cap(cap, NULL);
2517 spin_unlock(&inode->i_lock);
2521 * Handle cap IMPORT. If there are temp bits from an older EXPORT,
2524 * caller holds s_mutex.
2526 static void handle_cap_import(struct ceph_mds_client *mdsc,
2527 struct inode *inode, struct ceph_mds_caps *im,
2528 struct ceph_mds_session *session,
2529 void *snaptrace, int snaptrace_len)
2531 struct ceph_inode_info *ci = ceph_inode(inode);
2532 int mds = session->s_mds;
2533 unsigned issued = le32_to_cpu(im->caps);
2534 unsigned wanted = le32_to_cpu(im->wanted);
2535 unsigned seq = le32_to_cpu(im->seq);
2536 unsigned mseq = le32_to_cpu(im->migrate_seq);
2537 u64 realmino = le64_to_cpu(im->realm);
2538 u64 cap_id = le64_to_cpu(im->cap_id);
2540 if (ci->i_cap_exporting_mds >= 0 &&
2541 ceph_seq_cmp(ci->i_cap_exporting_mseq, mseq) < 0) {
2542 dout("handle_cap_import inode %p ci %p mds%d mseq %d"
2543 " - cleared exporting from mds%d\n",
2544 inode, ci, mds, mseq,
2545 ci->i_cap_exporting_mds);
2546 ci->i_cap_exporting_issued = 0;
2547 ci->i_cap_exporting_mseq = 0;
2548 ci->i_cap_exporting_mds = -1;
2550 dout("handle_cap_import inode %p ci %p mds%d mseq %d\n",
2551 inode, ci, mds, mseq);
2554 down_write(&mdsc->snap_rwsem);
2555 ceph_update_snap_trace(mdsc, snaptrace, snaptrace+snaptrace_len,
2557 downgrade_write(&mdsc->snap_rwsem);
2558 ceph_add_cap(inode, session, cap_id, -1,
2559 issued, wanted, seq, mseq, realmino, CEPH_CAP_FLAG_AUTH,
2560 NULL /* no caps context */);
2561 try_flush_caps(inode, session, NULL);
2562 up_read(&mdsc->snap_rwsem);
2566 * Handle a caps message from the MDS.
2568 * Identify the appropriate session, inode, and call the right handler
2569 * based on the cap op.
2571 void ceph_handle_caps(struct ceph_mds_session *session,
2572 struct ceph_msg *msg)
2574 struct ceph_mds_client *mdsc = session->s_mdsc;
2575 struct super_block *sb = mdsc->client->sb;
2576 struct inode *inode;
2577 struct ceph_cap *cap;
2578 struct ceph_mds_caps *h;
2579 int mds = le64_to_cpu(msg->hdr.src.name.num);
2582 struct ceph_vino vino;
2588 dout("handle_caps from mds%d\n", mds);
2591 if (msg->front.iov_len < sizeof(*h))
2593 h = msg->front.iov_base;
2594 op = le32_to_cpu(h->op);
2595 vino.ino = le64_to_cpu(h->ino);
2596 vino.snap = CEPH_NOSNAP;
2597 cap_id = le64_to_cpu(h->cap_id);
2598 seq = le32_to_cpu(h->seq);
2599 size = le64_to_cpu(h->size);
2600 max_size = le64_to_cpu(h->max_size);
2602 mutex_lock(&session->s_mutex);
2604 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
2608 inode = ceph_find_inode(sb, vino);
2609 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
2612 dout(" i don't have ino %llx\n", vino.ino);
2616 /* these will work even if we don't have a cap yet */
2618 case CEPH_CAP_OP_FLUSHSNAP_ACK:
2619 handle_cap_flushsnap_ack(inode, h, session);
2622 case CEPH_CAP_OP_EXPORT:
2623 handle_cap_export(inode, h, session);
2626 case CEPH_CAP_OP_IMPORT:
2627 handle_cap_import(mdsc, inode, h, session,
2629 le32_to_cpu(h->snap_trace_len));
2630 check_caps = 1; /* we may have sent a RELEASE to the old auth */
2634 /* the rest require a cap */
2635 spin_lock(&inode->i_lock);
2636 cap = __get_cap_for_mds(ceph_inode(inode), mds);
2638 dout("no cap on %p ino %llx.%llx from mds%d, releasing\n",
2639 inode, ceph_ino(inode), ceph_snap(inode), mds);
2640 spin_unlock(&inode->i_lock);
2644 /* note that each of these drops i_lock for us */
2646 case CEPH_CAP_OP_REVOKE:
2647 case CEPH_CAP_OP_GRANT:
2648 r = handle_cap_grant(inode, h, session, cap, msg->middle);
2650 ceph_check_caps(ceph_inode(inode),
2651 CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2654 ceph_check_caps(ceph_inode(inode),
2659 case CEPH_CAP_OP_FLUSH_ACK:
2660 handle_cap_flush_ack(inode, h, session, cap);
2663 case CEPH_CAP_OP_TRUNC:
2664 handle_cap_trunc(inode, h, session);
2668 spin_unlock(&inode->i_lock);
2669 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
2670 ceph_cap_op_name(op));
2674 mutex_unlock(&session->s_mutex);
2677 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_NODELAY, NULL);
2683 pr_err("ceph_handle_caps: corrupt message\n");
2688 * Delayed work handler to process end of delayed cap release LRU list.
2690 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
2692 struct ceph_inode_info *ci;
2693 int flags = CHECK_CAPS_NODELAY;
2695 dout("check_delayed_caps\n");
2697 spin_lock(&mdsc->cap_delay_lock);
2698 if (list_empty(&mdsc->cap_delay_list))
2700 ci = list_first_entry(&mdsc->cap_delay_list,
2701 struct ceph_inode_info,
2703 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
2704 time_before(jiffies, ci->i_hold_caps_max))
2706 list_del_init(&ci->i_cap_delay_list);
2707 spin_unlock(&mdsc->cap_delay_lock);
2708 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
2709 ceph_check_caps(ci, flags, NULL);
2711 spin_unlock(&mdsc->cap_delay_lock);
2715 * Flush all dirty caps to the mds
2717 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
2719 struct ceph_inode_info *ci;
2720 struct inode *inode;
2722 dout("flush_dirty_caps\n");
2723 spin_lock(&mdsc->cap_dirty_lock);
2724 while (!list_empty(&mdsc->cap_dirty)) {
2725 ci = list_first_entry(&mdsc->cap_dirty,
2726 struct ceph_inode_info,
2728 inode = igrab(&ci->vfs_inode);
2729 spin_unlock(&mdsc->cap_dirty_lock);
2731 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH,
2735 spin_lock(&mdsc->cap_dirty_lock);
2737 spin_unlock(&mdsc->cap_dirty_lock);
2741 * Drop open file reference. If we were the last open file,
2742 * we may need to release capabilities to the MDS (or schedule
2743 * their delayed release).
2745 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
2747 struct inode *inode = &ci->vfs_inode;
2750 spin_lock(&inode->i_lock);
2751 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
2752 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
2753 BUG_ON(ci->i_nr_by_mode[fmode] == 0);
2754 if (--ci->i_nr_by_mode[fmode] == 0)
2756 spin_unlock(&inode->i_lock);
2758 if (last && ci->i_vino.snap == CEPH_NOSNAP)
2759 ceph_check_caps(ci, 0, NULL);
2763 * Helpers for embedding cap and dentry lease releases into mds
2766 * @force is used by dentry_release (below) to force inclusion of a
2767 * record for the directory inode, even when there aren't any caps to
2770 int ceph_encode_inode_release(void **p, struct inode *inode,
2771 int mds, int drop, int unless, int force)
2773 struct ceph_inode_info *ci = ceph_inode(inode);
2774 struct ceph_cap *cap;
2775 struct ceph_mds_request_release *rel = *p;
2778 dout("encode_inode_release %p mds%d drop %s unless %s\n", inode,
2779 mds, ceph_cap_string(drop), ceph_cap_string(unless));
2781 spin_lock(&inode->i_lock);
2782 cap = __get_cap_for_mds(ci, mds);
2783 if (cap && __cap_is_valid(cap)) {
2785 ((cap->issued & drop) &&
2786 (cap->issued & unless) == 0)) {
2787 if ((cap->issued & drop) &&
2788 (cap->issued & unless) == 0) {
2789 dout("encode_inode_release %p cap %p %s -> "
2791 ceph_cap_string(cap->issued),
2792 ceph_cap_string(cap->issued & ~drop));
2793 cap->issued &= ~drop;
2794 cap->implemented &= ~drop;
2795 if (ci->i_ceph_flags & CEPH_I_NODELAY) {
2796 int wanted = __ceph_caps_wanted(ci);
2797 dout(" wanted %s -> %s (act %s)\n",
2798 ceph_cap_string(cap->mds_wanted),
2799 ceph_cap_string(cap->mds_wanted &
2801 ceph_cap_string(wanted));
2802 cap->mds_wanted &= wanted;
2805 dout("encode_inode_release %p cap %p %s"
2806 " (force)\n", inode, cap,
2807 ceph_cap_string(cap->issued));
2810 rel->ino = cpu_to_le64(ceph_ino(inode));
2811 rel->cap_id = cpu_to_le64(cap->cap_id);
2812 rel->seq = cpu_to_le32(cap->seq);
2813 rel->issue_seq = cpu_to_le32(cap->issue_seq),
2814 rel->mseq = cpu_to_le32(cap->mseq);
2815 rel->caps = cpu_to_le32(cap->issued);
2816 rel->wanted = cpu_to_le32(cap->mds_wanted);
2822 dout("encode_inode_release %p cap %p %s\n",
2823 inode, cap, ceph_cap_string(cap->issued));
2826 spin_unlock(&inode->i_lock);
2830 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
2831 int mds, int drop, int unless)
2833 struct inode *dir = dentry->d_parent->d_inode;
2834 struct ceph_mds_request_release *rel = *p;
2835 struct ceph_dentry_info *di = ceph_dentry(dentry);
2840 * force an record for the directory caps if we have a dentry lease.
2841 * this is racy (can't take i_lock and d_lock together), but it
2842 * doesn't have to be perfect; the mds will revoke anything we don't
2845 spin_lock(&dentry->d_lock);
2846 if (di->lease_session && di->lease_session->s_mds == mds)
2848 spin_unlock(&dentry->d_lock);
2850 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
2852 spin_lock(&dentry->d_lock);
2853 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
2854 dout("encode_dentry_release %p mds%d seq %d\n",
2855 dentry, mds, (int)di->lease_seq);
2856 rel->dname_len = cpu_to_le32(dentry->d_name.len);
2857 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
2858 *p += dentry->d_name.len;
2859 rel->dname_seq = cpu_to_le32(di->lease_seq);
2861 spin_unlock(&dentry->d_lock);