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 */
131 static int caps_min_count; /* keep at least this many (unreserved) */
133 void __init ceph_caps_init(void)
135 INIT_LIST_HEAD(&caps_list);
136 spin_lock_init(&caps_list_lock);
139 void ceph_caps_finalize(void)
141 struct ceph_cap *cap;
143 spin_lock(&caps_list_lock);
144 while (!list_empty(&caps_list)) {
145 cap = list_first_entry(&caps_list, struct ceph_cap, caps_item);
146 list_del(&cap->caps_item);
147 kmem_cache_free(ceph_cap_cachep, cap);
149 caps_total_count = 0;
150 caps_avail_count = 0;
152 caps_reserve_count = 0;
154 spin_unlock(&caps_list_lock);
157 void ceph_adjust_min_caps(int delta)
159 spin_lock(&caps_list_lock);
160 caps_min_count += delta;
161 BUG_ON(caps_min_count < 0);
162 spin_unlock(&caps_list_lock);
165 int ceph_reserve_caps(struct ceph_cap_reservation *ctx, int need)
168 struct ceph_cap *cap;
174 dout("reserve caps ctx=%p need=%d\n", ctx, need);
176 /* first reserve any caps that are already allocated */
177 spin_lock(&caps_list_lock);
178 if (caps_avail_count >= need)
181 have = caps_avail_count;
182 caps_avail_count -= have;
183 caps_reserve_count += have;
184 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
186 spin_unlock(&caps_list_lock);
188 for (i = have; i < need; i++) {
189 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
192 goto out_alloc_count;
194 list_add(&cap->caps_item, &newcaps);
197 BUG_ON(have + alloc != need);
199 spin_lock(&caps_list_lock);
200 caps_total_count += alloc;
201 caps_reserve_count += alloc;
202 list_splice(&newcaps, &caps_list);
204 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
206 spin_unlock(&caps_list_lock);
209 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
210 ctx, caps_total_count, caps_use_count, caps_reserve_count,
215 /* we didn't manage to reserve as much as we needed */
216 pr_warning("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
221 int ceph_unreserve_caps(struct ceph_cap_reservation *ctx)
223 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
225 spin_lock(&caps_list_lock);
226 BUG_ON(caps_reserve_count < ctx->count);
227 caps_reserve_count -= ctx->count;
228 caps_avail_count += ctx->count;
230 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
231 caps_total_count, caps_use_count, caps_reserve_count,
233 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
235 spin_unlock(&caps_list_lock);
240 static struct ceph_cap *get_cap(struct ceph_cap_reservation *ctx)
242 struct ceph_cap *cap = NULL;
244 /* temporary, until we do something about cap import/export */
246 return kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
248 spin_lock(&caps_list_lock);
249 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
250 ctx, ctx->count, caps_total_count, caps_use_count,
251 caps_reserve_count, caps_avail_count);
253 BUG_ON(ctx->count > caps_reserve_count);
254 BUG_ON(list_empty(&caps_list));
257 caps_reserve_count--;
260 cap = list_first_entry(&caps_list, struct ceph_cap, caps_item);
261 list_del(&cap->caps_item);
263 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
265 spin_unlock(&caps_list_lock);
269 void ceph_put_cap(struct ceph_cap *cap)
271 spin_lock(&caps_list_lock);
272 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
273 cap, caps_total_count, caps_use_count,
274 caps_reserve_count, caps_avail_count);
277 * Keep some preallocated caps around (ceph_min_count), to
278 * avoid lots of free/alloc churn.
280 if (caps_avail_count >= caps_reserve_count + caps_min_count) {
282 kmem_cache_free(ceph_cap_cachep, cap);
285 list_add(&cap->caps_item, &caps_list);
288 BUG_ON(caps_total_count != caps_use_count + caps_reserve_count +
290 spin_unlock(&caps_list_lock);
293 void ceph_reservation_status(struct ceph_client *client,
294 int *total, int *avail, int *used, int *reserved,
298 *total = caps_total_count;
300 *avail = caps_avail_count;
302 *used = caps_use_count;
304 *reserved = caps_reserve_count;
306 *min = caps_min_count;
310 * Find ceph_cap for given mds, if any.
312 * Called with i_lock held.
314 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
316 struct ceph_cap *cap;
317 struct rb_node *n = ci->i_caps.rb_node;
320 cap = rb_entry(n, struct ceph_cap, ci_node);
323 else if (mds > cap->mds)
332 * Return id of any MDS with a cap, preferably FILE_WR|WRBUFFER|EXCL, else
335 static int __ceph_get_cap_mds(struct ceph_inode_info *ci, u32 *mseq)
337 struct ceph_cap *cap;
341 /* prefer mds with WR|WRBUFFER|EXCL caps */
342 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
343 cap = rb_entry(p, struct ceph_cap, ci_node);
347 if (cap->issued & (CEPH_CAP_FILE_WR |
348 CEPH_CAP_FILE_BUFFER |
355 int ceph_get_cap_mds(struct inode *inode)
358 spin_lock(&inode->i_lock);
359 mds = __ceph_get_cap_mds(ceph_inode(inode), NULL);
360 spin_unlock(&inode->i_lock);
365 * Called under i_lock.
367 static void __insert_cap_node(struct ceph_inode_info *ci,
368 struct ceph_cap *new)
370 struct rb_node **p = &ci->i_caps.rb_node;
371 struct rb_node *parent = NULL;
372 struct ceph_cap *cap = NULL;
376 cap = rb_entry(parent, struct ceph_cap, ci_node);
377 if (new->mds < cap->mds)
379 else if (new->mds > cap->mds)
385 rb_link_node(&new->ci_node, parent, p);
386 rb_insert_color(&new->ci_node, &ci->i_caps);
390 * (re)set cap hold timeouts, which control the delayed release
391 * of unused caps back to the MDS. Should be called on cap use.
393 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
394 struct ceph_inode_info *ci)
396 struct ceph_mount_args *ma = mdsc->client->mount_args;
398 ci->i_hold_caps_min = round_jiffies(jiffies +
399 ma->caps_wanted_delay_min * HZ);
400 ci->i_hold_caps_max = round_jiffies(jiffies +
401 ma->caps_wanted_delay_max * HZ);
402 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
403 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
407 * (Re)queue cap at the end of the delayed cap release list.
409 * If I_FLUSH is set, leave the inode at the front of the list.
411 * Caller holds i_lock
412 * -> we take mdsc->cap_delay_lock
414 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
415 struct ceph_inode_info *ci)
417 __cap_set_timeouts(mdsc, ci);
418 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
419 ci->i_ceph_flags, ci->i_hold_caps_max);
420 if (!mdsc->stopping) {
421 spin_lock(&mdsc->cap_delay_lock);
422 if (!list_empty(&ci->i_cap_delay_list)) {
423 if (ci->i_ceph_flags & CEPH_I_FLUSH)
425 list_del_init(&ci->i_cap_delay_list);
427 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
429 spin_unlock(&mdsc->cap_delay_lock);
434 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
435 * indicating we should send a cap message to flush dirty metadata
436 * asap, and move to the front of the delayed cap list.
438 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
439 struct ceph_inode_info *ci)
441 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
442 spin_lock(&mdsc->cap_delay_lock);
443 ci->i_ceph_flags |= CEPH_I_FLUSH;
444 if (!list_empty(&ci->i_cap_delay_list))
445 list_del_init(&ci->i_cap_delay_list);
446 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
447 spin_unlock(&mdsc->cap_delay_lock);
451 * Cancel delayed work on cap.
453 * Caller must hold i_lock.
455 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
456 struct ceph_inode_info *ci)
458 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
459 if (list_empty(&ci->i_cap_delay_list))
461 spin_lock(&mdsc->cap_delay_lock);
462 list_del_init(&ci->i_cap_delay_list);
463 spin_unlock(&mdsc->cap_delay_lock);
467 * Common issue checks for add_cap, handle_cap_grant.
469 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
472 unsigned had = __ceph_caps_issued(ci, NULL);
475 * Each time we receive FILE_CACHE anew, we increment
478 if ((issued & CEPH_CAP_FILE_CACHE) &&
479 (had & CEPH_CAP_FILE_CACHE) == 0)
483 * if we are newly issued FILE_SHARED, clear I_COMPLETE; we
484 * don't know what happened to this directory while we didn't
487 if ((issued & CEPH_CAP_FILE_SHARED) &&
488 (had & CEPH_CAP_FILE_SHARED) == 0) {
490 if (S_ISDIR(ci->vfs_inode.i_mode)) {
491 dout(" marking %p NOT complete\n", &ci->vfs_inode);
492 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
498 * Add a capability under the given MDS session.
500 * Caller should hold session snap_rwsem (read) and s_mutex.
502 * @fmode is the open file mode, if we are opening a file, otherwise
503 * it is < 0. (This is so we can atomically add the cap and add an
504 * open file reference to it.)
506 int ceph_add_cap(struct inode *inode,
507 struct ceph_mds_session *session, u64 cap_id,
508 int fmode, unsigned issued, unsigned wanted,
509 unsigned seq, unsigned mseq, u64 realmino, int flags,
510 struct ceph_cap_reservation *caps_reservation)
512 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
513 struct ceph_inode_info *ci = ceph_inode(inode);
514 struct ceph_cap *new_cap = NULL;
515 struct ceph_cap *cap;
516 int mds = session->s_mds;
519 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
520 session->s_mds, cap_id, ceph_cap_string(issued), seq);
523 * If we are opening the file, include file mode wanted bits
527 wanted |= ceph_caps_for_mode(fmode);
530 spin_lock(&inode->i_lock);
531 cap = __get_cap_for_mds(ci, mds);
537 spin_unlock(&inode->i_lock);
538 new_cap = get_cap(caps_reservation);
545 cap->implemented = 0;
550 __insert_cap_node(ci, cap);
552 /* clear out old exporting info? (i.e. on cap import) */
553 if (ci->i_cap_exporting_mds == mds) {
554 ci->i_cap_exporting_issued = 0;
555 ci->i_cap_exporting_mseq = 0;
556 ci->i_cap_exporting_mds = -1;
559 /* add to session cap list */
560 cap->session = session;
561 spin_lock(&session->s_cap_lock);
562 list_add_tail(&cap->session_caps, &session->s_caps);
563 session->s_nr_caps++;
564 spin_unlock(&session->s_cap_lock);
567 if (!ci->i_snap_realm) {
569 * add this inode to the appropriate snap realm
571 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
574 ceph_get_snap_realm(mdsc, realm);
575 spin_lock(&realm->inodes_with_caps_lock);
576 ci->i_snap_realm = realm;
577 list_add(&ci->i_snap_realm_item,
578 &realm->inodes_with_caps);
579 spin_unlock(&realm->inodes_with_caps_lock);
581 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
586 __check_cap_issue(ci, cap, issued);
589 * If we are issued caps we don't want, or the mds' wanted
590 * value appears to be off, queue a check so we'll release
591 * later and/or update the mds wanted value.
593 actual_wanted = __ceph_caps_wanted(ci);
594 if ((wanted & ~actual_wanted) ||
595 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
596 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
597 ceph_cap_string(issued), ceph_cap_string(wanted),
598 ceph_cap_string(actual_wanted));
599 __cap_delay_requeue(mdsc, ci);
602 if (flags & CEPH_CAP_FLAG_AUTH)
603 ci->i_auth_cap = cap;
604 else if (ci->i_auth_cap == cap)
605 ci->i_auth_cap = NULL;
607 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
608 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
609 ceph_cap_string(issued|cap->issued), seq, mds);
610 cap->cap_id = cap_id;
611 cap->issued = issued;
612 cap->implemented |= issued;
613 cap->mds_wanted |= wanted;
615 cap->issue_seq = seq;
617 cap->cap_gen = session->s_cap_gen;
620 __ceph_get_fmode(ci, fmode);
621 spin_unlock(&inode->i_lock);
622 wake_up(&ci->i_cap_wq);
627 * Return true if cap has not timed out and belongs to the current
628 * generation of the MDS session (i.e. has not gone 'stale' due to
629 * us losing touch with the mds).
631 static int __cap_is_valid(struct ceph_cap *cap)
636 spin_lock(&cap->session->s_cap_lock);
637 gen = cap->session->s_cap_gen;
638 ttl = cap->session->s_cap_ttl;
639 spin_unlock(&cap->session->s_cap_lock);
641 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
642 dout("__cap_is_valid %p cap %p issued %s "
643 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
644 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
652 * Return set of valid cap bits issued to us. Note that caps time
653 * out, and may be invalidated in bulk if the client session times out
654 * and session->s_cap_gen is bumped.
656 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
658 int have = ci->i_snap_caps;
659 struct ceph_cap *cap;
664 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
665 cap = rb_entry(p, struct ceph_cap, ci_node);
666 if (!__cap_is_valid(cap))
668 dout("__ceph_caps_issued %p cap %p issued %s\n",
669 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
672 *implemented |= cap->implemented;
678 * Get cap bits issued by caps other than @ocap
680 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
682 int have = ci->i_snap_caps;
683 struct ceph_cap *cap;
686 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
687 cap = rb_entry(p, struct ceph_cap, ci_node);
690 if (!__cap_is_valid(cap))
698 * Move a cap to the end of the LRU (oldest caps at list head, newest
701 static void __touch_cap(struct ceph_cap *cap)
703 struct ceph_mds_session *s = cap->session;
705 spin_lock(&s->s_cap_lock);
706 if (s->s_cap_iterator == NULL) {
707 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
709 list_move_tail(&cap->session_caps, &s->s_caps);
711 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
712 &cap->ci->vfs_inode, cap, s->s_mds);
714 spin_unlock(&s->s_cap_lock);
718 * Check if we hold the given mask. If so, move the cap(s) to the
719 * front of their respective LRUs. (This is the preferred way for
720 * callers to check for caps they want.)
722 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
724 struct ceph_cap *cap;
726 int have = ci->i_snap_caps;
728 if ((have & mask) == mask) {
729 dout("__ceph_caps_issued_mask %p snap issued %s"
730 " (mask %s)\n", &ci->vfs_inode,
731 ceph_cap_string(have),
732 ceph_cap_string(mask));
736 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
737 cap = rb_entry(p, struct ceph_cap, ci_node);
738 if (!__cap_is_valid(cap))
740 if ((cap->issued & mask) == mask) {
741 dout("__ceph_caps_issued_mask %p cap %p issued %s"
742 " (mask %s)\n", &ci->vfs_inode, cap,
743 ceph_cap_string(cap->issued),
744 ceph_cap_string(mask));
750 /* does a combination of caps satisfy mask? */
752 if ((have & mask) == mask) {
753 dout("__ceph_caps_issued_mask %p combo issued %s"
754 " (mask %s)\n", &ci->vfs_inode,
755 ceph_cap_string(cap->issued),
756 ceph_cap_string(mask));
760 /* touch this + preceeding caps */
762 for (q = rb_first(&ci->i_caps); q != p;
764 cap = rb_entry(q, struct ceph_cap,
766 if (!__cap_is_valid(cap))
779 * Return true if mask caps are currently being revoked by an MDS.
781 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
783 struct inode *inode = &ci->vfs_inode;
784 struct ceph_cap *cap;
788 spin_lock(&inode->i_lock);
789 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
790 cap = rb_entry(p, struct ceph_cap, ci_node);
791 if (__cap_is_valid(cap) &&
792 (cap->implemented & ~cap->issued & mask)) {
797 spin_unlock(&inode->i_lock);
798 dout("ceph_caps_revoking %p %s = %d\n", inode,
799 ceph_cap_string(mask), ret);
803 int __ceph_caps_used(struct ceph_inode_info *ci)
807 used |= CEPH_CAP_PIN;
809 used |= CEPH_CAP_FILE_RD;
810 if (ci->i_rdcache_ref || ci->i_rdcache_gen)
811 used |= CEPH_CAP_FILE_CACHE;
813 used |= CEPH_CAP_FILE_WR;
814 if (ci->i_wrbuffer_ref)
815 used |= CEPH_CAP_FILE_BUFFER;
820 * wanted, by virtue of open file modes
822 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
826 for (mode = 0; mode < 4; mode++)
827 if (ci->i_nr_by_mode[mode])
828 want |= ceph_caps_for_mode(mode);
833 * Return caps we have registered with the MDS(s) as 'wanted'.
835 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
837 struct ceph_cap *cap;
841 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
842 cap = rb_entry(p, struct ceph_cap, ci_node);
843 if (!__cap_is_valid(cap))
845 mds_wanted |= cap->mds_wanted;
851 * called under i_lock
853 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
855 return !RB_EMPTY_ROOT(&ci->i_caps) || ci->i_cap_exporting_mds >= 0;
859 * caller should hold i_lock, and session s_mutex.
860 * returns true if this is the last cap. if so, caller should iput.
862 void __ceph_remove_cap(struct ceph_cap *cap)
864 struct ceph_mds_session *session = cap->session;
865 struct ceph_inode_info *ci = cap->ci;
866 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc;
868 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
870 /* remove from inode list */
871 rb_erase(&cap->ci_node, &ci->i_caps);
873 if (ci->i_auth_cap == cap)
874 ci->i_auth_cap = NULL;
876 /* remove from session list */
877 spin_lock(&session->s_cap_lock);
878 if (session->s_cap_iterator == cap) {
879 /* not yet, we are iterating over this very cap */
880 dout("__ceph_remove_cap delaying %p removal from session %p\n",
883 list_del_init(&cap->session_caps);
884 session->s_nr_caps--;
887 spin_unlock(&session->s_cap_lock);
889 if (cap->session == NULL)
892 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) {
893 struct ceph_snap_realm *realm = ci->i_snap_realm;
894 spin_lock(&realm->inodes_with_caps_lock);
895 list_del_init(&ci->i_snap_realm_item);
896 ci->i_snap_realm_counter++;
897 ci->i_snap_realm = NULL;
898 spin_unlock(&realm->inodes_with_caps_lock);
899 ceph_put_snap_realm(mdsc, realm);
901 if (!__ceph_is_any_real_caps(ci))
902 __cap_delay_cancel(mdsc, ci);
906 * Build and send a cap message to the given MDS.
908 * Caller should be holding s_mutex.
910 static int send_cap_msg(struct ceph_mds_session *session,
911 u64 ino, u64 cid, int op,
912 int caps, int wanted, int dirty,
913 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq,
914 u64 size, u64 max_size,
915 struct timespec *mtime, struct timespec *atime,
917 uid_t uid, gid_t gid, mode_t mode,
919 struct ceph_buffer *xattrs_buf,
922 struct ceph_mds_caps *fc;
923 struct ceph_msg *msg;
925 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
926 " seq %u/%u mseq %u follows %lld size %llu/%llu"
927 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
928 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
929 ceph_cap_string(dirty),
930 seq, issue_seq, mseq, follows, size, max_size,
931 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
933 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc), 0, 0, NULL);
937 msg->hdr.tid = cpu_to_le64(flush_tid);
939 fc = msg->front.iov_base;
940 memset(fc, 0, sizeof(*fc));
942 fc->cap_id = cpu_to_le64(cid);
943 fc->op = cpu_to_le32(op);
944 fc->seq = cpu_to_le32(seq);
945 fc->issue_seq = cpu_to_le32(issue_seq);
946 fc->migrate_seq = cpu_to_le32(mseq);
947 fc->caps = cpu_to_le32(caps);
948 fc->wanted = cpu_to_le32(wanted);
949 fc->dirty = cpu_to_le32(dirty);
950 fc->ino = cpu_to_le64(ino);
951 fc->snap_follows = cpu_to_le64(follows);
953 fc->size = cpu_to_le64(size);
954 fc->max_size = cpu_to_le64(max_size);
956 ceph_encode_timespec(&fc->mtime, mtime);
958 ceph_encode_timespec(&fc->atime, atime);
959 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
961 fc->uid = cpu_to_le32(uid);
962 fc->gid = cpu_to_le32(gid);
963 fc->mode = cpu_to_le32(mode);
965 fc->xattr_version = cpu_to_le64(xattr_version);
967 msg->middle = ceph_buffer_get(xattrs_buf);
968 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
969 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
972 ceph_con_send(&session->s_con, msg);
977 * Queue cap releases when an inode is dropped from our
980 void ceph_queue_caps_release(struct inode *inode)
982 struct ceph_inode_info *ci = ceph_inode(inode);
985 spin_lock(&inode->i_lock);
986 p = rb_first(&ci->i_caps);
988 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
989 struct ceph_mds_session *session = cap->session;
990 struct ceph_msg *msg;
991 struct ceph_mds_cap_release *head;
992 struct ceph_mds_cap_item *item;
994 spin_lock(&session->s_cap_lock);
995 BUG_ON(!session->s_num_cap_releases);
996 msg = list_first_entry(&session->s_cap_releases,
997 struct ceph_msg, list_head);
999 dout(" adding %p release to mds%d msg %p (%d left)\n",
1000 inode, session->s_mds, msg, session->s_num_cap_releases);
1002 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE);
1003 head = msg->front.iov_base;
1004 head->num = cpu_to_le32(le32_to_cpu(head->num) + 1);
1005 item = msg->front.iov_base + msg->front.iov_len;
1006 item->ino = cpu_to_le64(ceph_ino(inode));
1007 item->cap_id = cpu_to_le64(cap->cap_id);
1008 item->migrate_seq = cpu_to_le32(cap->mseq);
1009 item->seq = cpu_to_le32(cap->issue_seq);
1011 session->s_num_cap_releases--;
1013 msg->front.iov_len += sizeof(*item);
1014 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1015 dout(" release msg %p full\n", msg);
1016 list_move_tail(&msg->list_head,
1017 &session->s_cap_releases_done);
1019 dout(" release msg %p at %d/%d (%d)\n", msg,
1020 (int)le32_to_cpu(head->num),
1021 (int)CEPH_CAPS_PER_RELEASE,
1022 (int)msg->front.iov_len);
1024 spin_unlock(&session->s_cap_lock);
1026 __ceph_remove_cap(cap);
1029 spin_unlock(&inode->i_lock);
1033 * Send a cap msg on the given inode. Update our caps state, then
1034 * drop i_lock and send the message.
1036 * Make note of max_size reported/requested from mds, revoked caps
1037 * that have now been implemented.
1039 * Make half-hearted attempt ot to invalidate page cache if we are
1040 * dropping RDCACHE. Note that this will leave behind locked pages
1041 * that we'll then need to deal with elsewhere.
1043 * Return non-zero if delayed release, or we experienced an error
1044 * such that the caller should requeue + retry later.
1046 * called with i_lock, then drops it.
1047 * caller should hold snap_rwsem (read), s_mutex.
1049 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1050 int op, int used, int want, int retain, int flushing,
1051 unsigned *pflush_tid)
1052 __releases(cap->ci->vfs_inode->i_lock)
1054 struct ceph_inode_info *ci = cap->ci;
1055 struct inode *inode = &ci->vfs_inode;
1056 u64 cap_id = cap->cap_id;
1057 int held, revoking, dropping, keep;
1058 u64 seq, issue_seq, mseq, time_warp_seq, follows;
1060 struct timespec mtime, atime;
1065 struct ceph_mds_session *session;
1066 u64 xattr_version = 0;
1072 held = cap->issued | cap->implemented;
1073 revoking = cap->implemented & ~cap->issued;
1074 retain &= ~revoking;
1075 dropping = cap->issued & ~retain;
1077 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1078 inode, cap, cap->session,
1079 ceph_cap_string(held), ceph_cap_string(held & retain),
1080 ceph_cap_string(revoking));
1081 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1083 session = cap->session;
1085 /* don't release wanted unless we've waited a bit. */
1086 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1087 time_before(jiffies, ci->i_hold_caps_min)) {
1088 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1089 ceph_cap_string(cap->issued),
1090 ceph_cap_string(cap->issued & retain),
1091 ceph_cap_string(cap->mds_wanted),
1092 ceph_cap_string(want));
1093 want |= cap->mds_wanted;
1094 retain |= cap->issued;
1097 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1099 cap->issued &= retain; /* drop bits we don't want */
1100 if (cap->implemented & ~cap->issued) {
1102 * Wake up any waiters on wanted -> needed transition.
1103 * This is due to the weird transition from buffered
1104 * to sync IO... we need to flush dirty pages _before_
1105 * allowing sync writes to avoid reordering.
1109 cap->implemented &= cap->issued | used;
1110 cap->mds_wanted = want;
1114 * assign a tid for flush operations so we can avoid
1115 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark
1116 * clean type races. track latest tid for every bit
1117 * so we can handle flush AxFw, flush Fw, and have the
1118 * first ack clean Ax.
1120 flush_tid = ++ci->i_cap_flush_last_tid;
1122 *pflush_tid = flush_tid;
1123 dout(" cap_flush_tid %d\n", (int)flush_tid);
1124 for (i = 0; i < CEPH_CAP_BITS; i++)
1125 if (flushing & (1 << i))
1126 ci->i_cap_flush_tid[i] = flush_tid;
1129 keep = cap->implemented;
1131 issue_seq = cap->issue_seq;
1133 size = inode->i_size;
1134 ci->i_reported_size = size;
1135 max_size = ci->i_wanted_max_size;
1136 ci->i_requested_max_size = max_size;
1137 mtime = inode->i_mtime;
1138 atime = inode->i_atime;
1139 time_warp_seq = ci->i_time_warp_seq;
1140 follows = ci->i_snap_realm->cached_context->seq;
1143 mode = inode->i_mode;
1145 if (dropping & CEPH_CAP_XATTR_EXCL) {
1146 __ceph_build_xattrs_blob(ci);
1147 xattr_version = ci->i_xattrs.version + 1;
1150 spin_unlock(&inode->i_lock);
1152 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1153 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq,
1154 size, max_size, &mtime, &atime, time_warp_seq,
1157 (flushing & CEPH_CAP_XATTR_EXCL) ? ci->i_xattrs.blob : NULL,
1160 dout("error sending cap msg, must requeue %p\n", inode);
1165 wake_up(&ci->i_cap_wq);
1171 * When a snapshot is taken, clients accumulate dirty metadata on
1172 * inodes with capabilities in ceph_cap_snaps to describe the file
1173 * state at the time the snapshot was taken. This must be flushed
1174 * asynchronously back to the MDS once sync writes complete and dirty
1175 * data is written out.
1177 * Called under i_lock. Takes s_mutex as needed.
1179 void __ceph_flush_snaps(struct ceph_inode_info *ci,
1180 struct ceph_mds_session **psession)
1182 struct inode *inode = &ci->vfs_inode;
1184 struct ceph_cap_snap *capsnap;
1186 struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
1187 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1189 u64 next_follows = 0; /* keep track of how far we've gotten through the
1190 i_cap_snaps list, and skip these entries next time
1191 around to avoid an infinite loop */
1194 session = *psession;
1196 dout("__flush_snaps %p\n", inode);
1198 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1199 /* avoid an infiniute loop after retry */
1200 if (capsnap->follows < next_follows)
1203 * we need to wait for sync writes to complete and for dirty
1204 * pages to be written out.
1206 if (capsnap->dirty_pages || capsnap->writing)
1209 /* pick mds, take s_mutex */
1210 mds = __ceph_get_cap_mds(ci, &mseq);
1211 if (session && session->s_mds != mds) {
1212 dout("oops, wrong session %p mutex\n", session);
1213 mutex_unlock(&session->s_mutex);
1214 ceph_put_mds_session(session);
1218 spin_unlock(&inode->i_lock);
1219 mutex_lock(&mdsc->mutex);
1220 session = __ceph_lookup_mds_session(mdsc, mds);
1221 mutex_unlock(&mdsc->mutex);
1223 dout("inverting session/ino locks on %p\n",
1225 mutex_lock(&session->s_mutex);
1228 * if session == NULL, we raced against a cap
1229 * deletion. retry, and we'll get a better
1230 * @mds value next time.
1232 spin_lock(&inode->i_lock);
1236 capsnap->flush_tid = ++ci->i_cap_flush_last_tid;
1237 atomic_inc(&capsnap->nref);
1238 if (!list_empty(&capsnap->flushing_item))
1239 list_del_init(&capsnap->flushing_item);
1240 list_add_tail(&capsnap->flushing_item,
1241 &session->s_cap_snaps_flushing);
1242 spin_unlock(&inode->i_lock);
1244 dout("flush_snaps %p cap_snap %p follows %lld size %llu\n",
1245 inode, capsnap, next_follows, capsnap->size);
1246 send_cap_msg(session, ceph_vino(inode).ino, 0,
1247 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1248 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq,
1250 &capsnap->mtime, &capsnap->atime,
1251 capsnap->time_warp_seq,
1252 capsnap->uid, capsnap->gid, capsnap->mode,
1256 next_follows = capsnap->follows + 1;
1257 ceph_put_cap_snap(capsnap);
1259 spin_lock(&inode->i_lock);
1263 /* we flushed them all; remove this inode from the queue */
1264 spin_lock(&mdsc->snap_flush_lock);
1265 list_del_init(&ci->i_snap_flush_item);
1266 spin_unlock(&mdsc->snap_flush_lock);
1269 *psession = session;
1271 mutex_unlock(&session->s_mutex);
1272 ceph_put_mds_session(session);
1276 static void ceph_flush_snaps(struct ceph_inode_info *ci)
1278 struct inode *inode = &ci->vfs_inode;
1280 spin_lock(&inode->i_lock);
1281 __ceph_flush_snaps(ci, NULL);
1282 spin_unlock(&inode->i_lock);
1286 * Mark caps dirty. If inode is newly dirty, add to the global dirty
1289 void __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1291 struct ceph_mds_client *mdsc = &ceph_client(ci->vfs_inode.i_sb)->mdsc;
1292 struct inode *inode = &ci->vfs_inode;
1293 int was = ci->i_dirty_caps;
1296 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1297 ceph_cap_string(mask), ceph_cap_string(was),
1298 ceph_cap_string(was | mask));
1299 ci->i_dirty_caps |= mask;
1301 dout(" inode %p now dirty\n", &ci->vfs_inode);
1302 BUG_ON(!list_empty(&ci->i_dirty_item));
1303 spin_lock(&mdsc->cap_dirty_lock);
1304 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1305 spin_unlock(&mdsc->cap_dirty_lock);
1306 if (ci->i_flushing_caps == 0) {
1308 dirty |= I_DIRTY_SYNC;
1311 BUG_ON(list_empty(&ci->i_dirty_item));
1312 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1313 (mask & CEPH_CAP_FILE_BUFFER))
1314 dirty |= I_DIRTY_DATASYNC;
1316 __mark_inode_dirty(inode, dirty);
1317 __cap_delay_requeue(mdsc, ci);
1321 * Add dirty inode to the flushing list. Assigned a seq number so we
1322 * can wait for caps to flush without starving.
1324 * Called under i_lock.
1326 static int __mark_caps_flushing(struct inode *inode,
1327 struct ceph_mds_session *session)
1329 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1330 struct ceph_inode_info *ci = ceph_inode(inode);
1333 BUG_ON(ci->i_dirty_caps == 0);
1334 BUG_ON(list_empty(&ci->i_dirty_item));
1336 flushing = ci->i_dirty_caps;
1337 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1338 ceph_cap_string(flushing),
1339 ceph_cap_string(ci->i_flushing_caps),
1340 ceph_cap_string(ci->i_flushing_caps | flushing));
1341 ci->i_flushing_caps |= flushing;
1342 ci->i_dirty_caps = 0;
1343 dout(" inode %p now !dirty\n", inode);
1345 spin_lock(&mdsc->cap_dirty_lock);
1346 list_del_init(&ci->i_dirty_item);
1348 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq;
1349 if (list_empty(&ci->i_flushing_item)) {
1350 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1351 mdsc->num_cap_flushing++;
1352 dout(" inode %p now flushing seq %lld\n", inode,
1353 ci->i_cap_flush_seq);
1355 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1356 dout(" inode %p now flushing (more) seq %lld\n", inode,
1357 ci->i_cap_flush_seq);
1359 spin_unlock(&mdsc->cap_dirty_lock);
1365 * try to invalidate mapping pages without blocking.
1367 static int mapping_is_empty(struct address_space *mapping)
1369 struct page *page = find_get_page(mapping, 0);
1378 static int try_nonblocking_invalidate(struct inode *inode)
1380 struct ceph_inode_info *ci = ceph_inode(inode);
1381 u32 invalidating_gen = ci->i_rdcache_gen;
1383 spin_unlock(&inode->i_lock);
1384 invalidate_mapping_pages(&inode->i_data, 0, -1);
1385 spin_lock(&inode->i_lock);
1387 if (mapping_is_empty(&inode->i_data) &&
1388 invalidating_gen == ci->i_rdcache_gen) {
1390 dout("try_nonblocking_invalidate %p success\n", inode);
1391 ci->i_rdcache_gen = 0;
1392 ci->i_rdcache_revoking = 0;
1395 dout("try_nonblocking_invalidate %p failed\n", inode);
1400 * Swiss army knife function to examine currently used and wanted
1401 * versus held caps. Release, flush, ack revoked caps to mds as
1404 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1405 * cap release further.
1406 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1407 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1410 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1411 struct ceph_mds_session *session)
1413 struct ceph_client *client = ceph_inode_to_client(&ci->vfs_inode);
1414 struct ceph_mds_client *mdsc = &client->mdsc;
1415 struct inode *inode = &ci->vfs_inode;
1416 struct ceph_cap *cap;
1417 int file_wanted, used;
1418 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1419 int drop_session_lock = session ? 0 : 1;
1420 int issued, implemented, want, retain, revoking, flushing = 0;
1421 int mds = -1; /* keep track of how far we've gone through i_caps list
1422 to avoid an infinite loop on retry */
1424 int tried_invalidate = 0;
1425 int delayed = 0, sent = 0, force_requeue = 0, num;
1426 int queue_invalidate = 0;
1427 int is_delayed = flags & CHECK_CAPS_NODELAY;
1429 /* if we are unmounting, flush any unused caps immediately. */
1433 spin_lock(&inode->i_lock);
1435 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1436 flags |= CHECK_CAPS_FLUSH;
1438 /* flush snaps first time around only */
1439 if (!list_empty(&ci->i_cap_snaps))
1440 __ceph_flush_snaps(ci, &session);
1443 spin_lock(&inode->i_lock);
1445 file_wanted = __ceph_caps_file_wanted(ci);
1446 used = __ceph_caps_used(ci);
1447 want = file_wanted | used;
1448 issued = __ceph_caps_issued(ci, &implemented);
1449 revoking = implemented & ~issued;
1451 retain = want | CEPH_CAP_PIN;
1452 if (!mdsc->stopping && inode->i_nlink > 0) {
1454 retain |= CEPH_CAP_ANY; /* be greedy */
1456 retain |= CEPH_CAP_ANY_SHARED;
1458 * keep RD only if we didn't have the file open RW,
1459 * because then the mds would revoke it anyway to
1460 * journal max_size=0.
1462 if (ci->i_max_size == 0)
1463 retain |= CEPH_CAP_ANY_RD;
1467 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1468 " issued %s revoking %s retain %s %s%s%s\n", inode,
1469 ceph_cap_string(file_wanted),
1470 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1471 ceph_cap_string(ci->i_flushing_caps),
1472 ceph_cap_string(issued), ceph_cap_string(revoking),
1473 ceph_cap_string(retain),
1474 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1475 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1476 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1479 * If we no longer need to hold onto old our caps, and we may
1480 * have cached pages, but don't want them, then try to invalidate.
1481 * If we fail, it's because pages are locked.... try again later.
1483 if ((!is_delayed || mdsc->stopping) &&
1484 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */
1485 ci->i_rdcache_gen && /* may have cached pages */
1486 (file_wanted == 0 || /* no open files */
1487 (revoking & CEPH_CAP_FILE_CACHE)) && /* or revoking cache */
1488 !tried_invalidate) {
1489 dout("check_caps trying to invalidate on %p\n", inode);
1490 if (try_nonblocking_invalidate(inode) < 0) {
1491 if (revoking & CEPH_CAP_FILE_CACHE) {
1492 dout("check_caps queuing invalidate\n");
1493 queue_invalidate = 1;
1494 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1496 dout("check_caps failed to invalidate pages\n");
1497 /* we failed to invalidate pages. check these
1498 caps again later. */
1500 __cap_set_timeouts(mdsc, ci);
1503 tried_invalidate = 1;
1508 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1509 cap = rb_entry(p, struct ceph_cap, ci_node);
1512 /* avoid looping forever */
1513 if (mds >= cap->mds ||
1514 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1517 /* NOTE: no side-effects allowed, until we take s_mutex */
1519 revoking = cap->implemented & ~cap->issued;
1521 dout(" mds%d revoking %s\n", cap->mds,
1522 ceph_cap_string(revoking));
1524 if (cap == ci->i_auth_cap &&
1525 (cap->issued & CEPH_CAP_FILE_WR)) {
1526 /* request larger max_size from MDS? */
1527 if (ci->i_wanted_max_size > ci->i_max_size &&
1528 ci->i_wanted_max_size > ci->i_requested_max_size) {
1529 dout("requesting new max_size\n");
1533 /* approaching file_max? */
1534 if ((inode->i_size << 1) >= ci->i_max_size &&
1535 (ci->i_reported_size << 1) < ci->i_max_size) {
1536 dout("i_size approaching max_size\n");
1540 /* flush anything dirty? */
1541 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1543 dout("flushing dirty caps\n");
1547 /* completed revocation? going down and there are no caps? */
1548 if (revoking && (revoking & used) == 0) {
1549 dout("completed revocation of %s\n",
1550 ceph_cap_string(cap->implemented & ~cap->issued));
1554 /* want more caps from mds? */
1555 if (want & ~(cap->mds_wanted | cap->issued))
1558 /* things we might delay */
1559 if ((cap->issued & ~retain) == 0 &&
1560 cap->mds_wanted == want)
1561 continue; /* nope, all good */
1567 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1568 time_before(jiffies, ci->i_hold_caps_max)) {
1569 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1570 ceph_cap_string(cap->issued),
1571 ceph_cap_string(cap->issued & retain),
1572 ceph_cap_string(cap->mds_wanted),
1573 ceph_cap_string(want));
1579 if (session && session != cap->session) {
1580 dout("oops, wrong session %p mutex\n", session);
1581 mutex_unlock(&session->s_mutex);
1585 session = cap->session;
1586 if (mutex_trylock(&session->s_mutex) == 0) {
1587 dout("inverting session/ino locks on %p\n",
1589 spin_unlock(&inode->i_lock);
1590 if (took_snap_rwsem) {
1591 up_read(&mdsc->snap_rwsem);
1592 took_snap_rwsem = 0;
1594 mutex_lock(&session->s_mutex);
1598 /* take snap_rwsem after session mutex */
1599 if (!took_snap_rwsem) {
1600 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1601 dout("inverting snap/in locks on %p\n",
1603 spin_unlock(&inode->i_lock);
1604 down_read(&mdsc->snap_rwsem);
1605 took_snap_rwsem = 1;
1608 took_snap_rwsem = 1;
1611 if (cap == ci->i_auth_cap && ci->i_dirty_caps)
1612 flushing = __mark_caps_flushing(inode, session);
1614 mds = cap->mds; /* remember mds, so we don't repeat */
1617 /* __send_cap drops i_lock */
1618 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, used, want,
1619 retain, flushing, NULL);
1620 goto retry; /* retake i_lock and restart our cap scan. */
1624 * Reschedule delayed caps release if we delayed anything,
1627 if (delayed && is_delayed)
1628 force_requeue = 1; /* __send_cap delayed release; requeue */
1629 if (!delayed && !is_delayed)
1630 __cap_delay_cancel(mdsc, ci);
1631 else if (!is_delayed || force_requeue)
1632 __cap_delay_requeue(mdsc, ci);
1634 spin_unlock(&inode->i_lock);
1636 if (queue_invalidate)
1637 ceph_queue_invalidate(inode);
1639 if (session && drop_session_lock)
1640 mutex_unlock(&session->s_mutex);
1641 if (took_snap_rwsem)
1642 up_read(&mdsc->snap_rwsem);
1646 * Try to flush dirty caps back to the auth mds.
1648 static int try_flush_caps(struct inode *inode, struct ceph_mds_session *session,
1649 unsigned *flush_tid)
1651 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1652 struct ceph_inode_info *ci = ceph_inode(inode);
1653 int unlock_session = session ? 0 : 1;
1657 spin_lock(&inode->i_lock);
1658 if (ci->i_dirty_caps && ci->i_auth_cap) {
1659 struct ceph_cap *cap = ci->i_auth_cap;
1660 int used = __ceph_caps_used(ci);
1661 int want = __ceph_caps_wanted(ci);
1665 spin_unlock(&inode->i_lock);
1666 session = cap->session;
1667 mutex_lock(&session->s_mutex);
1670 BUG_ON(session != cap->session);
1671 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1674 flushing = __mark_caps_flushing(inode, session);
1676 /* __send_cap drops i_lock */
1677 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1678 cap->issued | cap->implemented, flushing,
1683 spin_lock(&inode->i_lock);
1684 __cap_delay_requeue(mdsc, ci);
1687 spin_unlock(&inode->i_lock);
1689 if (session && unlock_session)
1690 mutex_unlock(&session->s_mutex);
1695 * Return true if we've flushed caps through the given flush_tid.
1697 static int caps_are_flushed(struct inode *inode, unsigned tid)
1699 struct ceph_inode_info *ci = ceph_inode(inode);
1700 int dirty, i, ret = 1;
1702 spin_lock(&inode->i_lock);
1703 dirty = __ceph_caps_dirty(ci);
1704 for (i = 0; i < CEPH_CAP_BITS; i++)
1705 if ((ci->i_flushing_caps & (1 << i)) &&
1706 ci->i_cap_flush_tid[i] <= tid) {
1707 /* still flushing this bit */
1711 spin_unlock(&inode->i_lock);
1716 * Wait on any unsafe replies for the given inode. First wait on the
1717 * newest request, and make that the upper bound. Then, if there are
1718 * more requests, keep waiting on the oldest as long as it is still older
1719 * than the original request.
1721 static void sync_write_wait(struct inode *inode)
1723 struct ceph_inode_info *ci = ceph_inode(inode);
1724 struct list_head *head = &ci->i_unsafe_writes;
1725 struct ceph_osd_request *req;
1728 spin_lock(&ci->i_unsafe_lock);
1729 if (list_empty(head))
1732 /* set upper bound as _last_ entry in chain */
1733 req = list_entry(head->prev, struct ceph_osd_request,
1735 last_tid = req->r_tid;
1738 ceph_osdc_get_request(req);
1739 spin_unlock(&ci->i_unsafe_lock);
1740 dout("sync_write_wait on tid %llu (until %llu)\n",
1741 req->r_tid, last_tid);
1742 wait_for_completion(&req->r_safe_completion);
1743 spin_lock(&ci->i_unsafe_lock);
1744 ceph_osdc_put_request(req);
1747 * from here on look at first entry in chain, since we
1748 * only want to wait for anything older than last_tid
1750 if (list_empty(head))
1752 req = list_entry(head->next, struct ceph_osd_request,
1754 } while (req->r_tid < last_tid);
1756 spin_unlock(&ci->i_unsafe_lock);
1759 int ceph_fsync(struct file *file, struct dentry *dentry, int datasync)
1761 struct inode *inode = dentry->d_inode;
1762 struct ceph_inode_info *ci = ceph_inode(inode);
1767 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
1768 sync_write_wait(inode);
1770 ret = filemap_write_and_wait(inode->i_mapping);
1774 dirty = try_flush_caps(inode, NULL, &flush_tid);
1775 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1778 * only wait on non-file metadata writeback (the mds
1779 * can recover size and mtime, so we don't need to
1782 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1783 dout("fsync waiting for flush_tid %u\n", flush_tid);
1784 ret = wait_event_interruptible(ci->i_cap_wq,
1785 caps_are_flushed(inode, flush_tid));
1788 dout("fsync %p%s done\n", inode, datasync ? " datasync" : "");
1793 * Flush any dirty caps back to the mds. If we aren't asked to wait,
1794 * queue inode for flush but don't do so immediately, because we can
1795 * get by with fewer MDS messages if we wait for data writeback to
1798 int ceph_write_inode(struct inode *inode, int wait)
1800 struct ceph_inode_info *ci = ceph_inode(inode);
1805 dout("write_inode %p wait=%d\n", inode, wait);
1807 dirty = try_flush_caps(inode, NULL, &flush_tid);
1809 err = wait_event_interruptible(ci->i_cap_wq,
1810 caps_are_flushed(inode, flush_tid));
1812 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
1814 spin_lock(&inode->i_lock);
1815 if (__ceph_caps_dirty(ci))
1816 __cap_delay_requeue_front(mdsc, ci);
1817 spin_unlock(&inode->i_lock);
1823 * After a recovering MDS goes active, we need to resend any caps
1826 * Caller holds session->s_mutex.
1828 static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
1829 struct ceph_mds_session *session)
1831 struct ceph_cap_snap *capsnap;
1833 dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
1834 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
1836 struct ceph_inode_info *ci = capsnap->ci;
1837 struct inode *inode = &ci->vfs_inode;
1838 struct ceph_cap *cap;
1840 spin_lock(&inode->i_lock);
1841 cap = ci->i_auth_cap;
1842 if (cap && cap->session == session) {
1843 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
1845 __ceph_flush_snaps(ci, &session);
1847 pr_err("%p auth cap %p not mds%d ???\n", inode,
1848 cap, session->s_mds);
1849 spin_unlock(&inode->i_lock);
1854 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
1855 struct ceph_mds_session *session)
1857 struct ceph_inode_info *ci;
1859 kick_flushing_capsnaps(mdsc, session);
1861 dout("kick_flushing_caps mds%d\n", session->s_mds);
1862 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
1863 struct inode *inode = &ci->vfs_inode;
1864 struct ceph_cap *cap;
1867 spin_lock(&inode->i_lock);
1868 cap = ci->i_auth_cap;
1869 if (cap && cap->session == session) {
1870 dout("kick_flushing_caps %p cap %p %s\n", inode,
1871 cap, ceph_cap_string(ci->i_flushing_caps));
1872 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
1873 __ceph_caps_used(ci),
1874 __ceph_caps_wanted(ci),
1875 cap->issued | cap->implemented,
1876 ci->i_flushing_caps, NULL);
1878 spin_lock(&inode->i_lock);
1879 __cap_delay_requeue(mdsc, ci);
1880 spin_unlock(&inode->i_lock);
1883 pr_err("%p auth cap %p not mds%d ???\n", inode,
1884 cap, session->s_mds);
1885 spin_unlock(&inode->i_lock);
1892 * Take references to capabilities we hold, so that we don't release
1893 * them to the MDS prematurely.
1895 * Protected by i_lock.
1897 static void __take_cap_refs(struct ceph_inode_info *ci, int got)
1899 if (got & CEPH_CAP_PIN)
1901 if (got & CEPH_CAP_FILE_RD)
1903 if (got & CEPH_CAP_FILE_CACHE)
1904 ci->i_rdcache_ref++;
1905 if (got & CEPH_CAP_FILE_WR)
1907 if (got & CEPH_CAP_FILE_BUFFER) {
1908 if (ci->i_wrbuffer_ref == 0)
1909 igrab(&ci->vfs_inode);
1910 ci->i_wrbuffer_ref++;
1911 dout("__take_cap_refs %p wrbuffer %d -> %d (?)\n",
1912 &ci->vfs_inode, ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref);
1917 * Try to grab cap references. Specify those refs we @want, and the
1918 * minimal set we @need. Also include the larger offset we are writing
1919 * to (when applicable), and check against max_size here as well.
1920 * Note that caller is responsible for ensuring max_size increases are
1921 * requested from the MDS.
1923 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
1924 int *got, loff_t endoff, int *check_max, int *err)
1926 struct inode *inode = &ci->vfs_inode;
1928 int have, implemented;
1930 dout("get_cap_refs %p need %s want %s\n", inode,
1931 ceph_cap_string(need), ceph_cap_string(want));
1932 spin_lock(&inode->i_lock);
1934 /* make sure we _have_ some caps! */
1935 if (!__ceph_is_any_caps(ci)) {
1936 dout("get_cap_refs %p no real caps\n", inode);
1942 if (need & CEPH_CAP_FILE_WR) {
1943 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
1944 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
1945 inode, endoff, ci->i_max_size);
1946 if (endoff > ci->i_wanted_max_size) {
1953 * If a sync write is in progress, we must wait, so that we
1954 * can get a final snapshot value for size+mtime.
1956 if (__ceph_have_pending_cap_snap(ci)) {
1957 dout("get_cap_refs %p cap_snap_pending\n", inode);
1961 have = __ceph_caps_issued(ci, &implemented);
1964 * disallow writes while a truncate is pending
1966 if (ci->i_truncate_pending)
1967 have &= ~CEPH_CAP_FILE_WR;
1969 if ((have & need) == need) {
1971 * Look at (implemented & ~have & not) so that we keep waiting
1972 * on transition from wanted -> needed caps. This is needed
1973 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
1974 * going before a prior buffered writeback happens.
1976 int not = want & ~(have & need);
1977 int revoking = implemented & ~have;
1978 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
1979 inode, ceph_cap_string(have), ceph_cap_string(not),
1980 ceph_cap_string(revoking));
1981 if ((revoking & not) == 0) {
1982 *got = need | (have & want);
1983 __take_cap_refs(ci, *got);
1987 dout("get_cap_refs %p have %s needed %s\n", inode,
1988 ceph_cap_string(have), ceph_cap_string(need));
1991 spin_unlock(&inode->i_lock);
1992 dout("get_cap_refs %p ret %d got %s\n", inode,
1993 ret, ceph_cap_string(*got));
1998 * Check the offset we are writing up to against our current
1999 * max_size. If necessary, tell the MDS we want to write to
2002 static void check_max_size(struct inode *inode, loff_t endoff)
2004 struct ceph_inode_info *ci = ceph_inode(inode);
2007 /* do we need to explicitly request a larger max_size? */
2008 spin_lock(&inode->i_lock);
2009 if ((endoff >= ci->i_max_size ||
2010 endoff > (inode->i_size << 1)) &&
2011 endoff > ci->i_wanted_max_size) {
2012 dout("write %p at large endoff %llu, req max_size\n",
2014 ci->i_wanted_max_size = endoff;
2017 spin_unlock(&inode->i_lock);
2019 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2023 * Wait for caps, and take cap references. If we can't get a WR cap
2024 * due to a small max_size, make sure we check_max_size (and possibly
2025 * ask the mds) so we don't get hung up indefinitely.
2027 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, int *got,
2030 int check_max, ret, err;
2034 check_max_size(&ci->vfs_inode, endoff);
2037 ret = wait_event_interruptible(ci->i_cap_wq,
2038 try_get_cap_refs(ci, need, want,
2049 * Take cap refs. Caller must already know we hold at least one ref
2050 * on the caps in question or we don't know this is safe.
2052 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2054 spin_lock(&ci->vfs_inode.i_lock);
2055 __take_cap_refs(ci, caps);
2056 spin_unlock(&ci->vfs_inode.i_lock);
2062 * If we released the last ref on any given cap, call ceph_check_caps
2063 * to release (or schedule a release).
2065 * If we are releasing a WR cap (from a sync write), finalize any affected
2066 * cap_snap, and wake up any waiters.
2068 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2070 struct inode *inode = &ci->vfs_inode;
2071 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2072 struct ceph_cap_snap *capsnap;
2074 spin_lock(&inode->i_lock);
2075 if (had & CEPH_CAP_PIN)
2077 if (had & CEPH_CAP_FILE_RD)
2078 if (--ci->i_rd_ref == 0)
2080 if (had & CEPH_CAP_FILE_CACHE)
2081 if (--ci->i_rdcache_ref == 0)
2083 if (had & CEPH_CAP_FILE_BUFFER) {
2084 if (--ci->i_wrbuffer_ref == 0) {
2088 dout("put_cap_refs %p wrbuffer %d -> %d (?)\n",
2089 inode, ci->i_wrbuffer_ref+1, ci->i_wrbuffer_ref);
2091 if (had & CEPH_CAP_FILE_WR)
2092 if (--ci->i_wr_ref == 0) {
2094 if (!list_empty(&ci->i_cap_snaps)) {
2095 capsnap = list_first_entry(&ci->i_cap_snaps,
2096 struct ceph_cap_snap,
2098 if (capsnap->writing) {
2099 capsnap->writing = 0;
2101 __ceph_finish_cap_snap(ci,
2107 spin_unlock(&inode->i_lock);
2109 dout("put_cap_refs %p had %s %s\n", inode, ceph_cap_string(had),
2110 last ? "last" : "");
2112 if (last && !flushsnaps)
2113 ceph_check_caps(ci, 0, NULL);
2114 else if (flushsnaps)
2115 ceph_flush_snaps(ci);
2117 wake_up(&ci->i_cap_wq);
2123 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2124 * context. Adjust per-snap dirty page accounting as appropriate.
2125 * Once all dirty data for a cap_snap is flushed, flush snapped file
2126 * metadata back to the MDS. If we dropped the last ref, call
2129 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2130 struct ceph_snap_context *snapc)
2132 struct inode *inode = &ci->vfs_inode;
2136 struct ceph_cap_snap *capsnap = NULL;
2138 spin_lock(&inode->i_lock);
2139 ci->i_wrbuffer_ref -= nr;
2140 last = !ci->i_wrbuffer_ref;
2142 if (ci->i_head_snapc == snapc) {
2143 ci->i_wrbuffer_ref_head -= nr;
2144 if (!ci->i_wrbuffer_ref_head) {
2145 ceph_put_snap_context(ci->i_head_snapc);
2146 ci->i_head_snapc = NULL;
2148 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2150 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2151 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2152 last ? " LAST" : "");
2154 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2155 if (capsnap->context == snapc) {
2157 capsnap->dirty_pages -= nr;
2158 last_snap = !capsnap->dirty_pages;
2163 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2164 " snap %lld %d/%d -> %d/%d %s%s\n",
2165 inode, capsnap, capsnap->context->seq,
2166 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2167 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2168 last ? " (wrbuffer last)" : "",
2169 last_snap ? " (capsnap last)" : "");
2172 spin_unlock(&inode->i_lock);
2175 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2177 } else if (last_snap) {
2178 ceph_flush_snaps(ci);
2179 wake_up(&ci->i_cap_wq);
2184 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2185 * actually be a revocation if it specifies a smaller cap set.)
2187 * caller holds s_mutex.
2190 * 1 - check_caps on auth cap only (writeback)
2191 * 2 - check_caps (ack revoke)
2193 static int handle_cap_grant(struct inode *inode, struct ceph_mds_caps *grant,
2194 struct ceph_mds_session *session,
2195 struct ceph_cap *cap,
2196 struct ceph_buffer *xattr_buf)
2197 __releases(inode->i_lock)
2200 struct ceph_inode_info *ci = ceph_inode(inode);
2201 int mds = session->s_mds;
2202 int seq = le32_to_cpu(grant->seq);
2203 int newcaps = le32_to_cpu(grant->caps);
2204 int issued, implemented, used, wanted, dirty;
2205 u64 size = le64_to_cpu(grant->size);
2206 u64 max_size = le64_to_cpu(grant->max_size);
2207 struct timespec mtime, atime, ctime;
2211 int revoked_rdcache = 0;
2212 int queue_invalidate = 0;
2214 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2215 inode, cap, mds, seq, ceph_cap_string(newcaps));
2216 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2220 * If CACHE is being revoked, and we have no dirty buffers,
2221 * try to invalidate (once). (If there are dirty buffers, we
2222 * will invalidate _after_ writeback.)
2224 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2225 !ci->i_wrbuffer_ref) {
2226 if (try_nonblocking_invalidate(inode) == 0) {
2227 revoked_rdcache = 1;
2229 /* there were locked pages.. invalidate later
2230 in a separate thread. */
2231 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2232 queue_invalidate = 1;
2233 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2238 /* side effects now are allowed */
2240 issued = __ceph_caps_issued(ci, &implemented);
2241 issued |= implemented | __ceph_caps_dirty(ci);
2243 cap->cap_gen = session->s_cap_gen;
2245 __check_cap_issue(ci, cap, newcaps);
2247 if ((issued & CEPH_CAP_AUTH_EXCL) == 0) {
2248 inode->i_mode = le32_to_cpu(grant->mode);
2249 inode->i_uid = le32_to_cpu(grant->uid);
2250 inode->i_gid = le32_to_cpu(grant->gid);
2251 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2252 inode->i_uid, inode->i_gid);
2255 if ((issued & CEPH_CAP_LINK_EXCL) == 0)
2256 inode->i_nlink = le32_to_cpu(grant->nlink);
2258 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2259 int len = le32_to_cpu(grant->xattr_len);
2260 u64 version = le64_to_cpu(grant->xattr_version);
2262 if (version > ci->i_xattrs.version) {
2263 dout(" got new xattrs v%llu on %p len %d\n",
2264 version, inode, len);
2265 if (ci->i_xattrs.blob)
2266 ceph_buffer_put(ci->i_xattrs.blob);
2267 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2268 ci->i_xattrs.version = version;
2272 /* size/ctime/mtime/atime? */
2273 ceph_fill_file_size(inode, issued,
2274 le32_to_cpu(grant->truncate_seq),
2275 le64_to_cpu(grant->truncate_size), size);
2276 ceph_decode_timespec(&mtime, &grant->mtime);
2277 ceph_decode_timespec(&atime, &grant->atime);
2278 ceph_decode_timespec(&ctime, &grant->ctime);
2279 ceph_fill_file_time(inode, issued,
2280 le32_to_cpu(grant->time_warp_seq), &ctime, &mtime,
2283 /* max size increase? */
2284 if (max_size != ci->i_max_size) {
2285 dout("max_size %lld -> %llu\n", ci->i_max_size, max_size);
2286 ci->i_max_size = max_size;
2287 if (max_size >= ci->i_wanted_max_size) {
2288 ci->i_wanted_max_size = 0; /* reset */
2289 ci->i_requested_max_size = 0;
2294 /* check cap bits */
2295 wanted = __ceph_caps_wanted(ci);
2296 used = __ceph_caps_used(ci);
2297 dirty = __ceph_caps_dirty(ci);
2298 dout(" my wanted = %s, used = %s, dirty %s\n",
2299 ceph_cap_string(wanted),
2300 ceph_cap_string(used),
2301 ceph_cap_string(dirty));
2302 if (wanted != le32_to_cpu(grant->wanted)) {
2303 dout("mds wanted %s -> %s\n",
2304 ceph_cap_string(le32_to_cpu(grant->wanted)),
2305 ceph_cap_string(wanted));
2306 grant->wanted = cpu_to_le32(wanted);
2311 /* file layout may have changed */
2312 ci->i_layout = grant->layout;
2314 /* revocation, grant, or no-op? */
2315 if (cap->issued & ~newcaps) {
2316 dout("revocation: %s -> %s\n", ceph_cap_string(cap->issued),
2317 ceph_cap_string(newcaps));
2318 if ((used & ~newcaps) & CEPH_CAP_FILE_BUFFER)
2319 writeback = 1; /* will delay ack */
2320 else if (dirty & ~newcaps)
2321 reply = 1; /* initiate writeback in check_caps */
2322 else if (((used & ~newcaps) & CEPH_CAP_FILE_CACHE) == 0 ||
2324 reply = 2; /* send revoke ack in check_caps */
2325 cap->issued = newcaps;
2326 } else if (cap->issued == newcaps) {
2327 dout("caps unchanged: %s -> %s\n",
2328 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2330 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2331 ceph_cap_string(newcaps));
2332 cap->issued = newcaps;
2333 cap->implemented |= newcaps; /* add bits only, to
2334 * avoid stepping on a
2335 * pending revocation */
2339 spin_unlock(&inode->i_lock);
2342 * queue inode for writeback: we can't actually call
2343 * filemap_write_and_wait, etc. from message handler
2346 ceph_queue_writeback(inode);
2347 if (queue_invalidate)
2348 ceph_queue_invalidate(inode);
2350 wake_up(&ci->i_cap_wq);
2355 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2356 * MDS has been safely committed.
2358 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2359 struct ceph_mds_caps *m,
2360 struct ceph_mds_session *session,
2361 struct ceph_cap *cap)
2362 __releases(inode->i_lock)
2364 struct ceph_inode_info *ci = ceph_inode(inode);
2365 struct ceph_mds_client *mdsc = &ceph_client(inode->i_sb)->mdsc;
2366 unsigned seq = le32_to_cpu(m->seq);
2367 int dirty = le32_to_cpu(m->dirty);
2372 for (i = 0; i < CEPH_CAP_BITS; i++)
2373 if ((dirty & (1 << i)) &&
2374 flush_tid == ci->i_cap_flush_tid[i])
2377 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2378 " flushing %s -> %s\n",
2379 inode, session->s_mds, seq, ceph_cap_string(dirty),
2380 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2381 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2383 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned))
2386 ci->i_flushing_caps &= ~cleaned;
2388 spin_lock(&mdsc->cap_dirty_lock);
2389 if (ci->i_flushing_caps == 0) {
2390 list_del_init(&ci->i_flushing_item);
2391 if (!list_empty(&session->s_cap_flushing))
2392 dout(" mds%d still flushing cap on %p\n",
2394 &list_entry(session->s_cap_flushing.next,
2395 struct ceph_inode_info,
2396 i_flushing_item)->vfs_inode);
2397 mdsc->num_cap_flushing--;
2398 wake_up(&mdsc->cap_flushing_wq);
2399 dout(" inode %p now !flushing\n", inode);
2401 if (ci->i_dirty_caps == 0) {
2402 dout(" inode %p now clean\n", inode);
2403 BUG_ON(!list_empty(&ci->i_dirty_item));
2406 BUG_ON(list_empty(&ci->i_dirty_item));
2409 spin_unlock(&mdsc->cap_dirty_lock);
2410 wake_up(&ci->i_cap_wq);
2413 spin_unlock(&inode->i_lock);
2419 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
2420 * throw away our cap_snap.
2422 * Caller hold s_mutex.
2424 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
2425 struct ceph_mds_caps *m,
2426 struct ceph_mds_session *session)
2428 struct ceph_inode_info *ci = ceph_inode(inode);
2429 u64 follows = le64_to_cpu(m->snap_follows);
2430 struct ceph_cap_snap *capsnap;
2433 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
2434 inode, ci, session->s_mds, follows);
2436 spin_lock(&inode->i_lock);
2437 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2438 if (capsnap->follows == follows) {
2439 if (capsnap->flush_tid != flush_tid) {
2440 dout(" cap_snap %p follows %lld tid %lld !="
2441 " %lld\n", capsnap, follows,
2442 flush_tid, capsnap->flush_tid);
2445 WARN_ON(capsnap->dirty_pages || capsnap->writing);
2446 dout(" removing cap_snap %p follows %lld\n",
2448 ceph_put_snap_context(capsnap->context);
2449 list_del(&capsnap->ci_item);
2450 list_del(&capsnap->flushing_item);
2451 ceph_put_cap_snap(capsnap);
2455 dout(" skipping cap_snap %p follows %lld\n",
2456 capsnap, capsnap->follows);
2459 spin_unlock(&inode->i_lock);
2465 * Handle TRUNC from MDS, indicating file truncation.
2467 * caller hold s_mutex.
2469 static void handle_cap_trunc(struct inode *inode,
2470 struct ceph_mds_caps *trunc,
2471 struct ceph_mds_session *session)
2472 __releases(inode->i_lock)
2474 struct ceph_inode_info *ci = ceph_inode(inode);
2475 int mds = session->s_mds;
2476 int seq = le32_to_cpu(trunc->seq);
2477 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
2478 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
2479 u64 size = le64_to_cpu(trunc->size);
2480 int implemented = 0;
2481 int dirty = __ceph_caps_dirty(ci);
2482 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
2483 int queue_trunc = 0;
2485 issued |= implemented | dirty;
2487 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
2488 inode, mds, seq, truncate_size, truncate_seq);
2489 queue_trunc = ceph_fill_file_size(inode, issued,
2490 truncate_seq, truncate_size, size);
2491 spin_unlock(&inode->i_lock);
2494 ceph_queue_vmtruncate(inode);
2498 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
2499 * different one. If we are the most recent migration we've seen (as
2500 * indicated by mseq), make note of the migrating cap bits for the
2501 * duration (until we see the corresponding IMPORT).
2503 * caller holds s_mutex
2505 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
2506 struct ceph_mds_session *session)
2508 struct ceph_inode_info *ci = ceph_inode(inode);
2509 int mds = session->s_mds;
2510 unsigned mseq = le32_to_cpu(ex->migrate_seq);
2511 struct ceph_cap *cap = NULL, *t;
2515 dout("handle_cap_export inode %p ci %p mds%d mseq %d\n",
2516 inode, ci, mds, mseq);
2518 spin_lock(&inode->i_lock);
2520 /* make sure we haven't seen a higher mseq */
2521 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2522 t = rb_entry(p, struct ceph_cap, ci_node);
2523 if (ceph_seq_cmp(t->mseq, mseq) > 0) {
2524 dout(" higher mseq on cap from mds%d\n",
2528 if (t->session->s_mds == mds)
2535 ci->i_cap_exporting_mds = mds;
2536 ci->i_cap_exporting_mseq = mseq;
2537 ci->i_cap_exporting_issued = cap->issued;
2539 __ceph_remove_cap(cap);
2544 spin_unlock(&inode->i_lock);
2548 * Handle cap IMPORT. If there are temp bits from an older EXPORT,
2551 * caller holds s_mutex.
2553 static void handle_cap_import(struct ceph_mds_client *mdsc,
2554 struct inode *inode, struct ceph_mds_caps *im,
2555 struct ceph_mds_session *session,
2556 void *snaptrace, int snaptrace_len)
2558 struct ceph_inode_info *ci = ceph_inode(inode);
2559 int mds = session->s_mds;
2560 unsigned issued = le32_to_cpu(im->caps);
2561 unsigned wanted = le32_to_cpu(im->wanted);
2562 unsigned seq = le32_to_cpu(im->seq);
2563 unsigned mseq = le32_to_cpu(im->migrate_seq);
2564 u64 realmino = le64_to_cpu(im->realm);
2565 u64 cap_id = le64_to_cpu(im->cap_id);
2567 if (ci->i_cap_exporting_mds >= 0 &&
2568 ceph_seq_cmp(ci->i_cap_exporting_mseq, mseq) < 0) {
2569 dout("handle_cap_import inode %p ci %p mds%d mseq %d"
2570 " - cleared exporting from mds%d\n",
2571 inode, ci, mds, mseq,
2572 ci->i_cap_exporting_mds);
2573 ci->i_cap_exporting_issued = 0;
2574 ci->i_cap_exporting_mseq = 0;
2575 ci->i_cap_exporting_mds = -1;
2577 dout("handle_cap_import inode %p ci %p mds%d mseq %d\n",
2578 inode, ci, mds, mseq);
2581 down_write(&mdsc->snap_rwsem);
2582 ceph_update_snap_trace(mdsc, snaptrace, snaptrace+snaptrace_len,
2584 downgrade_write(&mdsc->snap_rwsem);
2585 ceph_add_cap(inode, session, cap_id, -1,
2586 issued, wanted, seq, mseq, realmino, CEPH_CAP_FLAG_AUTH,
2587 NULL /* no caps context */);
2588 try_flush_caps(inode, session, NULL);
2589 up_read(&mdsc->snap_rwsem);
2593 * Handle a caps message from the MDS.
2595 * Identify the appropriate session, inode, and call the right handler
2596 * based on the cap op.
2598 void ceph_handle_caps(struct ceph_mds_session *session,
2599 struct ceph_msg *msg)
2601 struct ceph_mds_client *mdsc = session->s_mdsc;
2602 struct super_block *sb = mdsc->client->sb;
2603 struct inode *inode;
2604 struct ceph_cap *cap;
2605 struct ceph_mds_caps *h;
2606 int mds = le64_to_cpu(msg->hdr.src.name.num);
2609 struct ceph_vino vino;
2616 dout("handle_caps from mds%d\n", mds);
2619 tid = le64_to_cpu(msg->hdr.tid);
2620 if (msg->front.iov_len < sizeof(*h))
2622 h = msg->front.iov_base;
2623 op = le32_to_cpu(h->op);
2624 vino.ino = le64_to_cpu(h->ino);
2625 vino.snap = CEPH_NOSNAP;
2626 cap_id = le64_to_cpu(h->cap_id);
2627 seq = le32_to_cpu(h->seq);
2628 size = le64_to_cpu(h->size);
2629 max_size = le64_to_cpu(h->max_size);
2631 mutex_lock(&session->s_mutex);
2633 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
2637 inode = ceph_find_inode(sb, vino);
2638 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
2641 dout(" i don't have ino %llx\n", vino.ino);
2645 /* these will work even if we don't have a cap yet */
2647 case CEPH_CAP_OP_FLUSHSNAP_ACK:
2648 handle_cap_flushsnap_ack(inode, tid, h, session);
2651 case CEPH_CAP_OP_EXPORT:
2652 handle_cap_export(inode, h, session);
2655 case CEPH_CAP_OP_IMPORT:
2656 handle_cap_import(mdsc, inode, h, session,
2658 le32_to_cpu(h->snap_trace_len));
2659 check_caps = 1; /* we may have sent a RELEASE to the old auth */
2663 /* the rest require a cap */
2664 spin_lock(&inode->i_lock);
2665 cap = __get_cap_for_mds(ceph_inode(inode), mds);
2667 dout("no cap on %p ino %llx.%llx from mds%d, releasing\n",
2668 inode, ceph_ino(inode), ceph_snap(inode), mds);
2669 spin_unlock(&inode->i_lock);
2673 /* note that each of these drops i_lock for us */
2675 case CEPH_CAP_OP_REVOKE:
2676 case CEPH_CAP_OP_GRANT:
2677 r = handle_cap_grant(inode, h, session, cap, msg->middle);
2679 ceph_check_caps(ceph_inode(inode),
2680 CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2683 ceph_check_caps(ceph_inode(inode),
2688 case CEPH_CAP_OP_FLUSH_ACK:
2689 handle_cap_flush_ack(inode, tid, h, session, cap);
2692 case CEPH_CAP_OP_TRUNC:
2693 handle_cap_trunc(inode, h, session);
2697 spin_unlock(&inode->i_lock);
2698 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
2699 ceph_cap_op_name(op));
2703 mutex_unlock(&session->s_mutex);
2706 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_NODELAY, NULL);
2712 pr_err("ceph_handle_caps: corrupt message\n");
2718 * Delayed work handler to process end of delayed cap release LRU list.
2720 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
2722 struct ceph_inode_info *ci;
2723 int flags = CHECK_CAPS_NODELAY;
2725 dout("check_delayed_caps\n");
2727 spin_lock(&mdsc->cap_delay_lock);
2728 if (list_empty(&mdsc->cap_delay_list))
2730 ci = list_first_entry(&mdsc->cap_delay_list,
2731 struct ceph_inode_info,
2733 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
2734 time_before(jiffies, ci->i_hold_caps_max))
2736 list_del_init(&ci->i_cap_delay_list);
2737 spin_unlock(&mdsc->cap_delay_lock);
2738 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
2739 ceph_check_caps(ci, flags, NULL);
2741 spin_unlock(&mdsc->cap_delay_lock);
2745 * Flush all dirty caps to the mds
2747 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
2749 struct ceph_inode_info *ci;
2750 struct inode *inode;
2752 dout("flush_dirty_caps\n");
2753 spin_lock(&mdsc->cap_dirty_lock);
2754 while (!list_empty(&mdsc->cap_dirty)) {
2755 ci = list_first_entry(&mdsc->cap_dirty,
2756 struct ceph_inode_info,
2758 inode = igrab(&ci->vfs_inode);
2759 spin_unlock(&mdsc->cap_dirty_lock);
2761 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH,
2765 spin_lock(&mdsc->cap_dirty_lock);
2767 spin_unlock(&mdsc->cap_dirty_lock);
2771 * Drop open file reference. If we were the last open file,
2772 * we may need to release capabilities to the MDS (or schedule
2773 * their delayed release).
2775 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
2777 struct inode *inode = &ci->vfs_inode;
2780 spin_lock(&inode->i_lock);
2781 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
2782 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
2783 BUG_ON(ci->i_nr_by_mode[fmode] == 0);
2784 if (--ci->i_nr_by_mode[fmode] == 0)
2786 spin_unlock(&inode->i_lock);
2788 if (last && ci->i_vino.snap == CEPH_NOSNAP)
2789 ceph_check_caps(ci, 0, NULL);
2793 * Helpers for embedding cap and dentry lease releases into mds
2796 * @force is used by dentry_release (below) to force inclusion of a
2797 * record for the directory inode, even when there aren't any caps to
2800 int ceph_encode_inode_release(void **p, struct inode *inode,
2801 int mds, int drop, int unless, int force)
2803 struct ceph_inode_info *ci = ceph_inode(inode);
2804 struct ceph_cap *cap;
2805 struct ceph_mds_request_release *rel = *p;
2808 dout("encode_inode_release %p mds%d drop %s unless %s\n", inode,
2809 mds, ceph_cap_string(drop), ceph_cap_string(unless));
2811 spin_lock(&inode->i_lock);
2812 cap = __get_cap_for_mds(ci, mds);
2813 if (cap && __cap_is_valid(cap)) {
2815 ((cap->issued & drop) &&
2816 (cap->issued & unless) == 0)) {
2817 if ((cap->issued & drop) &&
2818 (cap->issued & unless) == 0) {
2819 dout("encode_inode_release %p cap %p %s -> "
2821 ceph_cap_string(cap->issued),
2822 ceph_cap_string(cap->issued & ~drop));
2823 cap->issued &= ~drop;
2824 cap->implemented &= ~drop;
2825 if (ci->i_ceph_flags & CEPH_I_NODELAY) {
2826 int wanted = __ceph_caps_wanted(ci);
2827 dout(" wanted %s -> %s (act %s)\n",
2828 ceph_cap_string(cap->mds_wanted),
2829 ceph_cap_string(cap->mds_wanted &
2831 ceph_cap_string(wanted));
2832 cap->mds_wanted &= wanted;
2835 dout("encode_inode_release %p cap %p %s"
2836 " (force)\n", inode, cap,
2837 ceph_cap_string(cap->issued));
2840 rel->ino = cpu_to_le64(ceph_ino(inode));
2841 rel->cap_id = cpu_to_le64(cap->cap_id);
2842 rel->seq = cpu_to_le32(cap->seq);
2843 rel->issue_seq = cpu_to_le32(cap->issue_seq),
2844 rel->mseq = cpu_to_le32(cap->mseq);
2845 rel->caps = cpu_to_le32(cap->issued);
2846 rel->wanted = cpu_to_le32(cap->mds_wanted);
2852 dout("encode_inode_release %p cap %p %s\n",
2853 inode, cap, ceph_cap_string(cap->issued));
2856 spin_unlock(&inode->i_lock);
2860 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
2861 int mds, int drop, int unless)
2863 struct inode *dir = dentry->d_parent->d_inode;
2864 struct ceph_mds_request_release *rel = *p;
2865 struct ceph_dentry_info *di = ceph_dentry(dentry);
2870 * force an record for the directory caps if we have a dentry lease.
2871 * this is racy (can't take i_lock and d_lock together), but it
2872 * doesn't have to be perfect; the mds will revoke anything we don't
2875 spin_lock(&dentry->d_lock);
2876 if (di->lease_session && di->lease_session->s_mds == mds)
2878 spin_unlock(&dentry->d_lock);
2880 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
2882 spin_lock(&dentry->d_lock);
2883 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
2884 dout("encode_dentry_release %p mds%d seq %d\n",
2885 dentry, mds, (int)di->lease_seq);
2886 rel->dname_len = cpu_to_le32(dentry->d_name.len);
2887 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
2888 *p += dentry->d_name.len;
2889 rel->dname_seq = cpu_to_le32(di->lease_seq);
2891 spin_unlock(&dentry->d_lock);