1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/kernel.h>
5 #include <linux/sched.h>
6 #include <linux/slab.h>
7 #include <linux/vmalloc.h>
8 #include <linux/wait.h>
9 #include <linux/writeback.h>
12 #include "mds_client.h"
14 #include <linux/ceph/decode.h>
15 #include <linux/ceph/messenger.h>
18 * Capability management
20 * The Ceph metadata servers control client access to inode metadata
21 * and file data by issuing capabilities, granting clients permission
22 * to read and/or write both inode field and file data to OSDs
23 * (storage nodes). Each capability consists of a set of bits
24 * indicating which operations are allowed.
26 * If the client holds a *_SHARED cap, the client has a coherent value
27 * that can be safely read from the cached inode.
29 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
30 * client is allowed to change inode attributes (e.g., file size,
31 * mtime), note its dirty state in the ceph_cap, and asynchronously
32 * flush that metadata change to the MDS.
34 * In the event of a conflicting operation (perhaps by another
35 * client), the MDS will revoke the conflicting client capabilities.
37 * In order for a client to cache an inode, it must hold a capability
38 * with at least one MDS server. When inodes are released, release
39 * notifications are batched and periodically sent en masse to the MDS
40 * cluster to release server state.
43 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc);
44 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
45 struct ceph_mds_session *session,
46 struct ceph_inode_info *ci,
47 u64 oldest_flush_tid);
50 * Generate readable cap strings for debugging output.
52 #define MAX_CAP_STR 20
53 static char cap_str[MAX_CAP_STR][40];
54 static DEFINE_SPINLOCK(cap_str_lock);
55 static int last_cap_str;
57 static char *gcap_string(char *s, int c)
59 if (c & CEPH_CAP_GSHARED)
61 if (c & CEPH_CAP_GEXCL)
63 if (c & CEPH_CAP_GCACHE)
69 if (c & CEPH_CAP_GBUFFER)
71 if (c & CEPH_CAP_GLAZYIO)
76 const char *ceph_cap_string(int caps)
82 spin_lock(&cap_str_lock);
84 if (last_cap_str == MAX_CAP_STR)
86 spin_unlock(&cap_str_lock);
90 if (caps & CEPH_CAP_PIN)
93 c = (caps >> CEPH_CAP_SAUTH) & 3;
96 s = gcap_string(s, c);
99 c = (caps >> CEPH_CAP_SLINK) & 3;
102 s = gcap_string(s, c);
105 c = (caps >> CEPH_CAP_SXATTR) & 3;
108 s = gcap_string(s, c);
111 c = caps >> CEPH_CAP_SFILE;
114 s = gcap_string(s, c);
123 void ceph_caps_init(struct ceph_mds_client *mdsc)
125 INIT_LIST_HEAD(&mdsc->caps_list);
126 spin_lock_init(&mdsc->caps_list_lock);
129 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
131 struct ceph_cap *cap;
133 spin_lock(&mdsc->caps_list_lock);
134 while (!list_empty(&mdsc->caps_list)) {
135 cap = list_first_entry(&mdsc->caps_list,
136 struct ceph_cap, caps_item);
137 list_del(&cap->caps_item);
138 kmem_cache_free(ceph_cap_cachep, cap);
140 mdsc->caps_total_count = 0;
141 mdsc->caps_avail_count = 0;
142 mdsc->caps_use_count = 0;
143 mdsc->caps_reserve_count = 0;
144 mdsc->caps_min_count = 0;
145 spin_unlock(&mdsc->caps_list_lock);
148 void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta)
150 spin_lock(&mdsc->caps_list_lock);
151 mdsc->caps_min_count += delta;
152 BUG_ON(mdsc->caps_min_count < 0);
153 spin_unlock(&mdsc->caps_list_lock);
156 void ceph_reserve_caps(struct ceph_mds_client *mdsc,
157 struct ceph_cap_reservation *ctx, int need)
160 struct ceph_cap *cap;
165 dout("reserve caps ctx=%p need=%d\n", ctx, need);
167 /* first reserve any caps that are already allocated */
168 spin_lock(&mdsc->caps_list_lock);
169 if (mdsc->caps_avail_count >= need)
172 have = mdsc->caps_avail_count;
173 mdsc->caps_avail_count -= have;
174 mdsc->caps_reserve_count += have;
175 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
176 mdsc->caps_reserve_count +
177 mdsc->caps_avail_count);
178 spin_unlock(&mdsc->caps_list_lock);
180 for (i = have; i < need; i++) {
181 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
184 list_add(&cap->caps_item, &newcaps);
187 /* we didn't manage to reserve as much as we needed */
188 if (have + alloc != need)
189 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
190 ctx, need, have + alloc);
192 spin_lock(&mdsc->caps_list_lock);
193 mdsc->caps_total_count += alloc;
194 mdsc->caps_reserve_count += alloc;
195 list_splice(&newcaps, &mdsc->caps_list);
197 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
198 mdsc->caps_reserve_count +
199 mdsc->caps_avail_count);
200 spin_unlock(&mdsc->caps_list_lock);
203 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
204 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
205 mdsc->caps_reserve_count, mdsc->caps_avail_count);
208 int ceph_unreserve_caps(struct ceph_mds_client *mdsc,
209 struct ceph_cap_reservation *ctx)
211 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
213 spin_lock(&mdsc->caps_list_lock);
214 BUG_ON(mdsc->caps_reserve_count < ctx->count);
215 mdsc->caps_reserve_count -= ctx->count;
216 mdsc->caps_avail_count += ctx->count;
218 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
219 mdsc->caps_total_count, mdsc->caps_use_count,
220 mdsc->caps_reserve_count, mdsc->caps_avail_count);
221 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
222 mdsc->caps_reserve_count +
223 mdsc->caps_avail_count);
224 spin_unlock(&mdsc->caps_list_lock);
229 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
230 struct ceph_cap_reservation *ctx)
232 struct ceph_cap *cap = NULL;
234 /* temporary, until we do something about cap import/export */
236 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
238 spin_lock(&mdsc->caps_list_lock);
239 mdsc->caps_use_count++;
240 mdsc->caps_total_count++;
241 spin_unlock(&mdsc->caps_list_lock);
246 spin_lock(&mdsc->caps_list_lock);
247 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
248 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
249 mdsc->caps_reserve_count, mdsc->caps_avail_count);
251 BUG_ON(ctx->count > mdsc->caps_reserve_count);
252 BUG_ON(list_empty(&mdsc->caps_list));
255 mdsc->caps_reserve_count--;
256 mdsc->caps_use_count++;
258 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
259 list_del(&cap->caps_item);
261 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
262 mdsc->caps_reserve_count + mdsc->caps_avail_count);
263 spin_unlock(&mdsc->caps_list_lock);
267 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
269 spin_lock(&mdsc->caps_list_lock);
270 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
271 cap, mdsc->caps_total_count, mdsc->caps_use_count,
272 mdsc->caps_reserve_count, mdsc->caps_avail_count);
273 mdsc->caps_use_count--;
275 * Keep some preallocated caps around (ceph_min_count), to
276 * avoid lots of free/alloc churn.
278 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
279 mdsc->caps_min_count) {
280 mdsc->caps_total_count--;
281 kmem_cache_free(ceph_cap_cachep, cap);
283 mdsc->caps_avail_count++;
284 list_add(&cap->caps_item, &mdsc->caps_list);
287 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
288 mdsc->caps_reserve_count + mdsc->caps_avail_count);
289 spin_unlock(&mdsc->caps_list_lock);
292 void ceph_reservation_status(struct ceph_fs_client *fsc,
293 int *total, int *avail, int *used, int *reserved,
296 struct ceph_mds_client *mdsc = fsc->mdsc;
299 *total = mdsc->caps_total_count;
301 *avail = mdsc->caps_avail_count;
303 *used = mdsc->caps_use_count;
305 *reserved = mdsc->caps_reserve_count;
307 *min = mdsc->caps_min_count;
311 * Find ceph_cap for given mds, if any.
313 * Called with i_ceph_lock held.
315 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
317 struct ceph_cap *cap;
318 struct rb_node *n = ci->i_caps.rb_node;
321 cap = rb_entry(n, struct ceph_cap, ci_node);
324 else if (mds > cap->mds)
332 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
334 struct ceph_cap *cap;
336 spin_lock(&ci->i_ceph_lock);
337 cap = __get_cap_for_mds(ci, mds);
338 spin_unlock(&ci->i_ceph_lock);
343 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
345 static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
347 struct ceph_cap *cap;
351 /* prefer mds with WR|BUFFER|EXCL caps */
352 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
353 cap = rb_entry(p, struct ceph_cap, ci_node);
355 if (cap->issued & (CEPH_CAP_FILE_WR |
356 CEPH_CAP_FILE_BUFFER |
363 int ceph_get_cap_mds(struct inode *inode)
365 struct ceph_inode_info *ci = ceph_inode(inode);
367 spin_lock(&ci->i_ceph_lock);
368 mds = __ceph_get_cap_mds(ceph_inode(inode));
369 spin_unlock(&ci->i_ceph_lock);
374 * Called under i_ceph_lock.
376 static void __insert_cap_node(struct ceph_inode_info *ci,
377 struct ceph_cap *new)
379 struct rb_node **p = &ci->i_caps.rb_node;
380 struct rb_node *parent = NULL;
381 struct ceph_cap *cap = NULL;
385 cap = rb_entry(parent, struct ceph_cap, ci_node);
386 if (new->mds < cap->mds)
388 else if (new->mds > cap->mds)
394 rb_link_node(&new->ci_node, parent, p);
395 rb_insert_color(&new->ci_node, &ci->i_caps);
399 * (re)set cap hold timeouts, which control the delayed release
400 * of unused caps back to the MDS. Should be called on cap use.
402 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
403 struct ceph_inode_info *ci)
405 struct ceph_mount_options *ma = mdsc->fsc->mount_options;
407 ci->i_hold_caps_min = round_jiffies(jiffies +
408 ma->caps_wanted_delay_min * HZ);
409 ci->i_hold_caps_max = round_jiffies(jiffies +
410 ma->caps_wanted_delay_max * HZ);
411 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
412 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
416 * (Re)queue cap at the end of the delayed cap release list.
418 * If I_FLUSH is set, leave the inode at the front of the list.
420 * Caller holds i_ceph_lock
421 * -> we take mdsc->cap_delay_lock
423 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
424 struct ceph_inode_info *ci)
426 __cap_set_timeouts(mdsc, ci);
427 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
428 ci->i_ceph_flags, ci->i_hold_caps_max);
429 if (!mdsc->stopping) {
430 spin_lock(&mdsc->cap_delay_lock);
431 if (!list_empty(&ci->i_cap_delay_list)) {
432 if (ci->i_ceph_flags & CEPH_I_FLUSH)
434 list_del_init(&ci->i_cap_delay_list);
436 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
438 spin_unlock(&mdsc->cap_delay_lock);
443 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
444 * indicating we should send a cap message to flush dirty metadata
445 * asap, and move to the front of the delayed cap list.
447 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
448 struct ceph_inode_info *ci)
450 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
451 spin_lock(&mdsc->cap_delay_lock);
452 ci->i_ceph_flags |= CEPH_I_FLUSH;
453 if (!list_empty(&ci->i_cap_delay_list))
454 list_del_init(&ci->i_cap_delay_list);
455 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
456 spin_unlock(&mdsc->cap_delay_lock);
460 * Cancel delayed work on cap.
462 * Caller must hold i_ceph_lock.
464 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
465 struct ceph_inode_info *ci)
467 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
468 if (list_empty(&ci->i_cap_delay_list))
470 spin_lock(&mdsc->cap_delay_lock);
471 list_del_init(&ci->i_cap_delay_list);
472 spin_unlock(&mdsc->cap_delay_lock);
476 * Common issue checks for add_cap, handle_cap_grant.
478 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
481 unsigned had = __ceph_caps_issued(ci, NULL);
484 * Each time we receive FILE_CACHE anew, we increment
487 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
488 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
493 * if we are newly issued FILE_SHARED, mark dir not complete; we
494 * don't know what happened to this directory while we didn't
497 if ((issued & CEPH_CAP_FILE_SHARED) &&
498 (had & CEPH_CAP_FILE_SHARED) == 0) {
500 if (S_ISDIR(ci->vfs_inode.i_mode)) {
501 dout(" marking %p NOT complete\n", &ci->vfs_inode);
502 __ceph_dir_clear_complete(ci);
508 * Add a capability under the given MDS session.
510 * Caller should hold session snap_rwsem (read) and s_mutex.
512 * @fmode is the open file mode, if we are opening a file, otherwise
513 * it is < 0. (This is so we can atomically add the cap and add an
514 * open file reference to it.)
516 void ceph_add_cap(struct inode *inode,
517 struct ceph_mds_session *session, u64 cap_id,
518 int fmode, unsigned issued, unsigned wanted,
519 unsigned seq, unsigned mseq, u64 realmino, int flags,
520 struct ceph_cap **new_cap)
522 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
523 struct ceph_inode_info *ci = ceph_inode(inode);
524 struct ceph_cap *cap;
525 int mds = session->s_mds;
528 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
529 session->s_mds, cap_id, ceph_cap_string(issued), seq);
532 * If we are opening the file, include file mode wanted bits
536 wanted |= ceph_caps_for_mode(fmode);
538 cap = __get_cap_for_mds(ci, mds);
544 cap->implemented = 0;
550 __insert_cap_node(ci, cap);
552 /* add to session cap list */
553 cap->session = session;
554 spin_lock(&session->s_cap_lock);
555 list_add_tail(&cap->session_caps, &session->s_caps);
556 session->s_nr_caps++;
557 spin_unlock(&session->s_cap_lock);
560 * auth mds of the inode changed. we received the cap export
561 * message, but still haven't received the cap import message.
562 * handle_cap_export() updated the new auth MDS' cap.
564 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
565 * a message that was send before the cap import message. So
568 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
569 WARN_ON(cap != ci->i_auth_cap);
570 WARN_ON(cap->cap_id != cap_id);
573 issued |= cap->issued;
574 flags |= CEPH_CAP_FLAG_AUTH;
578 if (!ci->i_snap_realm) {
580 * add this inode to the appropriate snap realm
582 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
585 spin_lock(&realm->inodes_with_caps_lock);
586 ci->i_snap_realm = realm;
587 list_add(&ci->i_snap_realm_item,
588 &realm->inodes_with_caps);
589 spin_unlock(&realm->inodes_with_caps_lock);
591 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
597 __check_cap_issue(ci, cap, issued);
600 * If we are issued caps we don't want, or the mds' wanted
601 * value appears to be off, queue a check so we'll release
602 * later and/or update the mds wanted value.
604 actual_wanted = __ceph_caps_wanted(ci);
605 if ((wanted & ~actual_wanted) ||
606 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
607 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
608 ceph_cap_string(issued), ceph_cap_string(wanted),
609 ceph_cap_string(actual_wanted));
610 __cap_delay_requeue(mdsc, ci);
613 if (flags & CEPH_CAP_FLAG_AUTH) {
614 if (ci->i_auth_cap == NULL ||
615 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
616 ci->i_auth_cap = cap;
617 cap->mds_wanted = wanted;
620 WARN_ON(ci->i_auth_cap == cap);
623 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
624 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
625 ceph_cap_string(issued|cap->issued), seq, mds);
626 cap->cap_id = cap_id;
627 cap->issued = issued;
628 cap->implemented |= issued;
629 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
630 cap->mds_wanted = wanted;
632 cap->mds_wanted |= wanted;
634 cap->issue_seq = seq;
636 cap->cap_gen = session->s_cap_gen;
639 __ceph_get_fmode(ci, fmode);
643 * Return true if cap has not timed out and belongs to the current
644 * generation of the MDS session (i.e. has not gone 'stale' due to
645 * us losing touch with the mds).
647 static int __cap_is_valid(struct ceph_cap *cap)
652 spin_lock(&cap->session->s_gen_ttl_lock);
653 gen = cap->session->s_cap_gen;
654 ttl = cap->session->s_cap_ttl;
655 spin_unlock(&cap->session->s_gen_ttl_lock);
657 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
658 dout("__cap_is_valid %p cap %p issued %s "
659 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
660 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
668 * Return set of valid cap bits issued to us. Note that caps time
669 * out, and may be invalidated in bulk if the client session times out
670 * and session->s_cap_gen is bumped.
672 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
674 int have = ci->i_snap_caps;
675 struct ceph_cap *cap;
680 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
681 cap = rb_entry(p, struct ceph_cap, ci_node);
682 if (!__cap_is_valid(cap))
684 dout("__ceph_caps_issued %p cap %p issued %s\n",
685 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
688 *implemented |= cap->implemented;
691 * exclude caps issued by non-auth MDS, but are been revoking
692 * by the auth MDS. The non-auth MDS should be revoking/exporting
693 * these caps, but the message is delayed.
695 if (ci->i_auth_cap) {
696 cap = ci->i_auth_cap;
697 have &= ~cap->implemented | cap->issued;
703 * Get cap bits issued by caps other than @ocap
705 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
707 int have = ci->i_snap_caps;
708 struct ceph_cap *cap;
711 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
712 cap = rb_entry(p, struct ceph_cap, ci_node);
715 if (!__cap_is_valid(cap))
723 * Move a cap to the end of the LRU (oldest caps at list head, newest
726 static void __touch_cap(struct ceph_cap *cap)
728 struct ceph_mds_session *s = cap->session;
730 spin_lock(&s->s_cap_lock);
731 if (s->s_cap_iterator == NULL) {
732 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
734 list_move_tail(&cap->session_caps, &s->s_caps);
736 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
737 &cap->ci->vfs_inode, cap, s->s_mds);
739 spin_unlock(&s->s_cap_lock);
743 * Check if we hold the given mask. If so, move the cap(s) to the
744 * front of their respective LRUs. (This is the preferred way for
745 * callers to check for caps they want.)
747 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
749 struct ceph_cap *cap;
751 int have = ci->i_snap_caps;
753 if ((have & mask) == mask) {
754 dout("__ceph_caps_issued_mask %p snap issued %s"
755 " (mask %s)\n", &ci->vfs_inode,
756 ceph_cap_string(have),
757 ceph_cap_string(mask));
761 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
762 cap = rb_entry(p, struct ceph_cap, ci_node);
763 if (!__cap_is_valid(cap))
765 if ((cap->issued & mask) == mask) {
766 dout("__ceph_caps_issued_mask %p cap %p issued %s"
767 " (mask %s)\n", &ci->vfs_inode, cap,
768 ceph_cap_string(cap->issued),
769 ceph_cap_string(mask));
775 /* does a combination of caps satisfy mask? */
777 if ((have & mask) == mask) {
778 dout("__ceph_caps_issued_mask %p combo issued %s"
779 " (mask %s)\n", &ci->vfs_inode,
780 ceph_cap_string(cap->issued),
781 ceph_cap_string(mask));
785 /* touch this + preceding caps */
787 for (q = rb_first(&ci->i_caps); q != p;
789 cap = rb_entry(q, struct ceph_cap,
791 if (!__cap_is_valid(cap))
804 * Return true if mask caps are currently being revoked by an MDS.
806 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
807 struct ceph_cap *ocap, int mask)
809 struct ceph_cap *cap;
812 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
813 cap = rb_entry(p, struct ceph_cap, ci_node);
815 (cap->implemented & ~cap->issued & mask))
821 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
823 struct inode *inode = &ci->vfs_inode;
826 spin_lock(&ci->i_ceph_lock);
827 ret = __ceph_caps_revoking_other(ci, NULL, mask);
828 spin_unlock(&ci->i_ceph_lock);
829 dout("ceph_caps_revoking %p %s = %d\n", inode,
830 ceph_cap_string(mask), ret);
834 int __ceph_caps_used(struct ceph_inode_info *ci)
838 used |= CEPH_CAP_PIN;
840 used |= CEPH_CAP_FILE_RD;
841 if (ci->i_rdcache_ref ||
842 (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */
843 ci->vfs_inode.i_data.nrpages))
844 used |= CEPH_CAP_FILE_CACHE;
846 used |= CEPH_CAP_FILE_WR;
847 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
848 used |= CEPH_CAP_FILE_BUFFER;
853 * wanted, by virtue of open file modes
855 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
858 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
859 if (ci->i_nr_by_mode[i])
864 return ceph_caps_for_mode(bits >> 1);
868 * Return caps we have registered with the MDS(s) as 'wanted'.
870 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
872 struct ceph_cap *cap;
876 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
877 cap = rb_entry(p, struct ceph_cap, ci_node);
878 if (!__cap_is_valid(cap))
880 if (cap == ci->i_auth_cap)
881 mds_wanted |= cap->mds_wanted;
883 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
889 * called under i_ceph_lock
891 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
893 return !RB_EMPTY_ROOT(&ci->i_caps);
896 int ceph_is_any_caps(struct inode *inode)
898 struct ceph_inode_info *ci = ceph_inode(inode);
901 spin_lock(&ci->i_ceph_lock);
902 ret = __ceph_is_any_caps(ci);
903 spin_unlock(&ci->i_ceph_lock);
908 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
910 struct ceph_snap_realm *realm = ci->i_snap_realm;
911 spin_lock(&realm->inodes_with_caps_lock);
912 list_del_init(&ci->i_snap_realm_item);
913 ci->i_snap_realm_counter++;
914 ci->i_snap_realm = NULL;
915 spin_unlock(&realm->inodes_with_caps_lock);
916 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
921 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
923 * caller should hold i_ceph_lock.
924 * caller will not hold session s_mutex if called from destroy_inode.
926 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
928 struct ceph_mds_session *session = cap->session;
929 struct ceph_inode_info *ci = cap->ci;
930 struct ceph_mds_client *mdsc =
931 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
934 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
936 /* remove from session list */
937 spin_lock(&session->s_cap_lock);
938 if (session->s_cap_iterator == cap) {
939 /* not yet, we are iterating over this very cap */
940 dout("__ceph_remove_cap delaying %p removal from session %p\n",
943 list_del_init(&cap->session_caps);
944 session->s_nr_caps--;
948 /* protect backpointer with s_cap_lock: see iterate_session_caps */
952 * s_cap_reconnect is protected by s_cap_lock. no one changes
953 * s_cap_gen while session is in the reconnect state.
956 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
957 cap->queue_release = 1;
959 list_add_tail(&cap->session_caps,
960 &session->s_cap_releases);
961 session->s_num_cap_releases++;
965 cap->queue_release = 0;
967 cap->cap_ino = ci->i_vino.ino;
969 spin_unlock(&session->s_cap_lock);
971 /* remove from inode list */
972 rb_erase(&cap->ci_node, &ci->i_caps);
973 if (ci->i_auth_cap == cap)
974 ci->i_auth_cap = NULL;
977 ceph_put_cap(mdsc, cap);
979 /* when reconnect denied, we remove session caps forcibly,
980 * i_wr_ref can be non-zero. If there are ongoing write,
983 if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm)
984 drop_inode_snap_realm(ci);
986 if (!__ceph_is_any_real_caps(ci))
987 __cap_delay_cancel(mdsc, ci);
991 * Build and send a cap message to the given MDS.
993 * Caller should be holding s_mutex.
995 static int send_cap_msg(struct ceph_mds_session *session,
996 u64 ino, u64 cid, int op,
997 int caps, int wanted, int dirty,
998 u32 seq, u64 flush_tid, u64 oldest_flush_tid,
999 u32 issue_seq, u32 mseq, u64 size, u64 max_size,
1000 struct timespec *mtime, struct timespec *atime,
1001 struct timespec *ctime, u32 time_warp_seq,
1002 kuid_t uid, kgid_t gid, umode_t mode,
1004 struct ceph_buffer *xattrs_buf,
1005 u64 follows, bool inline_data)
1007 struct ceph_mds_caps *fc;
1008 struct ceph_msg *msg;
1012 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1013 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1014 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
1015 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
1016 ceph_cap_string(dirty),
1017 seq, issue_seq, flush_tid, oldest_flush_tid,
1018 mseq, follows, size, max_size,
1019 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
1021 /* flock buffer size + inline version + inline data size +
1022 * osd_epoch_barrier + oldest_flush_tid */
1023 extra_len = 4 + 8 + 4 + 4 + 8;
1024 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1029 msg->hdr.version = cpu_to_le16(6);
1030 msg->hdr.tid = cpu_to_le64(flush_tid);
1032 fc = msg->front.iov_base;
1033 memset(fc, 0, sizeof(*fc));
1035 fc->cap_id = cpu_to_le64(cid);
1036 fc->op = cpu_to_le32(op);
1037 fc->seq = cpu_to_le32(seq);
1038 fc->issue_seq = cpu_to_le32(issue_seq);
1039 fc->migrate_seq = cpu_to_le32(mseq);
1040 fc->caps = cpu_to_le32(caps);
1041 fc->wanted = cpu_to_le32(wanted);
1042 fc->dirty = cpu_to_le32(dirty);
1043 fc->ino = cpu_to_le64(ino);
1044 fc->snap_follows = cpu_to_le64(follows);
1046 fc->size = cpu_to_le64(size);
1047 fc->max_size = cpu_to_le64(max_size);
1049 ceph_encode_timespec(&fc->mtime, mtime);
1051 ceph_encode_timespec(&fc->atime, atime);
1053 ceph_encode_timespec(&fc->ctime, ctime);
1054 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
1056 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, uid));
1057 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, gid));
1058 fc->mode = cpu_to_le32(mode);
1061 /* flock buffer size */
1062 ceph_encode_32(&p, 0);
1063 /* inline version */
1064 ceph_encode_64(&p, inline_data ? 0 : CEPH_INLINE_NONE);
1065 /* inline data size */
1066 ceph_encode_32(&p, 0);
1067 /* osd_epoch_barrier */
1068 ceph_encode_32(&p, 0);
1069 /* oldest_flush_tid */
1070 ceph_encode_64(&p, oldest_flush_tid);
1072 fc->xattr_version = cpu_to_le64(xattr_version);
1074 msg->middle = ceph_buffer_get(xattrs_buf);
1075 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1076 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1079 ceph_con_send(&session->s_con, msg);
1084 * Queue cap releases when an inode is dropped from our cache. Since
1085 * inode is about to be destroyed, there is no need for i_ceph_lock.
1087 void ceph_queue_caps_release(struct inode *inode)
1089 struct ceph_inode_info *ci = ceph_inode(inode);
1092 p = rb_first(&ci->i_caps);
1094 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1096 __ceph_remove_cap(cap, true);
1101 * Send a cap msg on the given inode. Update our caps state, then
1102 * drop i_ceph_lock and send the message.
1104 * Make note of max_size reported/requested from mds, revoked caps
1105 * that have now been implemented.
1107 * Make half-hearted attempt ot to invalidate page cache if we are
1108 * dropping RDCACHE. Note that this will leave behind locked pages
1109 * that we'll then need to deal with elsewhere.
1111 * Return non-zero if delayed release, or we experienced an error
1112 * such that the caller should requeue + retry later.
1114 * called with i_ceph_lock, then drops it.
1115 * caller should hold snap_rwsem (read), s_mutex.
1117 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1118 int op, int used, int want, int retain, int flushing,
1119 u64 flush_tid, u64 oldest_flush_tid)
1120 __releases(cap->ci->i_ceph_lock)
1122 struct ceph_inode_info *ci = cap->ci;
1123 struct inode *inode = &ci->vfs_inode;
1124 u64 cap_id = cap->cap_id;
1125 int held, revoking, dropping, keep;
1126 u64 follows, size, max_size;
1127 u32 seq, issue_seq, mseq, time_warp_seq;
1128 struct timespec mtime, atime, ctime;
1133 struct ceph_mds_session *session;
1134 u64 xattr_version = 0;
1135 struct ceph_buffer *xattr_blob = NULL;
1140 held = cap->issued | cap->implemented;
1141 revoking = cap->implemented & ~cap->issued;
1142 retain &= ~revoking;
1143 dropping = cap->issued & ~retain;
1145 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1146 inode, cap, cap->session,
1147 ceph_cap_string(held), ceph_cap_string(held & retain),
1148 ceph_cap_string(revoking));
1149 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1151 session = cap->session;
1153 /* don't release wanted unless we've waited a bit. */
1154 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1155 time_before(jiffies, ci->i_hold_caps_min)) {
1156 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1157 ceph_cap_string(cap->issued),
1158 ceph_cap_string(cap->issued & retain),
1159 ceph_cap_string(cap->mds_wanted),
1160 ceph_cap_string(want));
1161 want |= cap->mds_wanted;
1162 retain |= cap->issued;
1165 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1167 cap->issued &= retain; /* drop bits we don't want */
1168 if (cap->implemented & ~cap->issued) {
1170 * Wake up any waiters on wanted -> needed transition.
1171 * This is due to the weird transition from buffered
1172 * to sync IO... we need to flush dirty pages _before_
1173 * allowing sync writes to avoid reordering.
1177 cap->implemented &= cap->issued | used;
1178 cap->mds_wanted = want;
1180 follows = flushing ? ci->i_head_snapc->seq : 0;
1182 keep = cap->implemented;
1184 issue_seq = cap->issue_seq;
1186 size = inode->i_size;
1187 ci->i_reported_size = size;
1188 max_size = ci->i_wanted_max_size;
1189 ci->i_requested_max_size = max_size;
1190 mtime = inode->i_mtime;
1191 atime = inode->i_atime;
1192 ctime = inode->i_ctime;
1193 time_warp_seq = ci->i_time_warp_seq;
1196 mode = inode->i_mode;
1198 if (flushing & CEPH_CAP_XATTR_EXCL) {
1199 __ceph_build_xattrs_blob(ci);
1200 xattr_blob = ci->i_xattrs.blob;
1201 xattr_version = ci->i_xattrs.version;
1204 inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1206 spin_unlock(&ci->i_ceph_lock);
1208 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1209 op, keep, want, flushing, seq,
1210 flush_tid, oldest_flush_tid, issue_seq, mseq,
1211 size, max_size, &mtime, &atime, &ctime, time_warp_seq,
1212 uid, gid, mode, xattr_version, xattr_blob,
1213 follows, inline_data);
1215 dout("error sending cap msg, must requeue %p\n", inode);
1220 wake_up_all(&ci->i_cap_wq);
1225 static inline int __send_flush_snap(struct inode *inode,
1226 struct ceph_mds_session *session,
1227 struct ceph_cap_snap *capsnap,
1228 u32 mseq, u64 oldest_flush_tid)
1230 return send_cap_msg(session, ceph_vino(inode).ino, 0,
1231 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1232 capsnap->dirty, 0, capsnap->cap_flush.tid,
1233 oldest_flush_tid, 0, mseq, capsnap->size, 0,
1234 &capsnap->mtime, &capsnap->atime,
1235 &capsnap->ctime, capsnap->time_warp_seq,
1236 capsnap->uid, capsnap->gid, capsnap->mode,
1237 capsnap->xattr_version, capsnap->xattr_blob,
1238 capsnap->follows, capsnap->inline_data);
1242 * When a snapshot is taken, clients accumulate dirty metadata on
1243 * inodes with capabilities in ceph_cap_snaps to describe the file
1244 * state at the time the snapshot was taken. This must be flushed
1245 * asynchronously back to the MDS once sync writes complete and dirty
1246 * data is written out.
1248 * Called under i_ceph_lock. Takes s_mutex as needed.
1250 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1251 struct ceph_mds_session *session)
1252 __releases(ci->i_ceph_lock)
1253 __acquires(ci->i_ceph_lock)
1255 struct inode *inode = &ci->vfs_inode;
1256 struct ceph_mds_client *mdsc = session->s_mdsc;
1257 struct ceph_cap_snap *capsnap;
1258 u64 oldest_flush_tid = 0;
1259 u64 first_tid = 1, last_tid = 0;
1261 dout("__flush_snaps %p session %p\n", inode, session);
1263 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1265 * we need to wait for sync writes to complete and for dirty
1266 * pages to be written out.
1268 if (capsnap->dirty_pages || capsnap->writing)
1271 /* should be removed by ceph_try_drop_cap_snap() */
1272 BUG_ON(!capsnap->need_flush);
1274 /* only flush each capsnap once */
1275 if (capsnap->cap_flush.tid > 0) {
1276 dout(" already flushed %p, skipping\n", capsnap);
1280 spin_lock(&mdsc->cap_dirty_lock);
1281 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1282 list_add_tail(&capsnap->cap_flush.g_list,
1283 &mdsc->cap_flush_list);
1284 if (oldest_flush_tid == 0)
1285 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1286 if (list_empty(&ci->i_flushing_item)) {
1287 list_add_tail(&ci->i_flushing_item,
1288 &session->s_cap_flushing);
1290 spin_unlock(&mdsc->cap_dirty_lock);
1292 list_add_tail(&capsnap->cap_flush.i_list,
1293 &ci->i_cap_flush_list);
1296 first_tid = capsnap->cap_flush.tid;
1297 last_tid = capsnap->cap_flush.tid;
1300 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1302 while (first_tid <= last_tid) {
1303 struct ceph_cap *cap = ci->i_auth_cap;
1304 struct ceph_cap_flush *cf;
1307 if (!(cap && cap->session == session)) {
1308 dout("__flush_snaps %p auth cap %p not mds%d, "
1309 "stop\n", inode, cap, session->s_mds);
1314 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1315 if (cf->tid >= first_tid) {
1323 first_tid = cf->tid + 1;
1325 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1326 atomic_inc(&capsnap->nref);
1327 spin_unlock(&ci->i_ceph_lock);
1329 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1330 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1332 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1335 pr_err("__flush_snaps: error sending cap flushsnap, "
1336 "ino (%llx.%llx) tid %llu follows %llu\n",
1337 ceph_vinop(inode), cf->tid, capsnap->follows);
1340 ceph_put_cap_snap(capsnap);
1341 spin_lock(&ci->i_ceph_lock);
1345 void ceph_flush_snaps(struct ceph_inode_info *ci,
1346 struct ceph_mds_session **psession)
1348 struct inode *inode = &ci->vfs_inode;
1349 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1350 struct ceph_mds_session *session = *psession;
1352 dout("ceph_flush_snaps %p\n", inode);
1354 spin_lock(&ci->i_ceph_lock);
1355 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1356 dout(" no capsnap needs flush, doing nothing\n");
1359 if (!ci->i_auth_cap) {
1360 dout(" no auth cap (migrating?), doing nothing\n");
1364 mds = ci->i_auth_cap->session->s_mds;
1365 if (session && session->s_mds != mds) {
1366 dout(" oops, wrong session %p mutex\n", session);
1367 mutex_unlock(&session->s_mutex);
1368 ceph_put_mds_session(session);
1372 spin_unlock(&ci->i_ceph_lock);
1373 mutex_lock(&mdsc->mutex);
1374 session = __ceph_lookup_mds_session(mdsc, mds);
1375 mutex_unlock(&mdsc->mutex);
1377 dout(" inverting session/ino locks on %p\n", session);
1378 mutex_lock(&session->s_mutex);
1383 __ceph_flush_snaps(ci, session);
1385 spin_unlock(&ci->i_ceph_lock);
1388 *psession = session;
1390 mutex_unlock(&session->s_mutex);
1391 ceph_put_mds_session(session);
1393 /* we flushed them all; remove this inode from the queue */
1394 spin_lock(&mdsc->snap_flush_lock);
1395 list_del_init(&ci->i_snap_flush_item);
1396 spin_unlock(&mdsc->snap_flush_lock);
1400 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1401 * Caller is then responsible for calling __mark_inode_dirty with the
1402 * returned flags value.
1404 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1405 struct ceph_cap_flush **pcf)
1407 struct ceph_mds_client *mdsc =
1408 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1409 struct inode *inode = &ci->vfs_inode;
1410 int was = ci->i_dirty_caps;
1413 if (!ci->i_auth_cap) {
1414 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1415 "but no auth cap (session was closed?)\n",
1416 inode, ceph_ino(inode), ceph_cap_string(mask));
1420 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1421 ceph_cap_string(mask), ceph_cap_string(was),
1422 ceph_cap_string(was | mask));
1423 ci->i_dirty_caps |= mask;
1425 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1426 swap(ci->i_prealloc_cap_flush, *pcf);
1428 if (!ci->i_head_snapc) {
1429 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1430 ci->i_head_snapc = ceph_get_snap_context(
1431 ci->i_snap_realm->cached_context);
1433 dout(" inode %p now dirty snapc %p auth cap %p\n",
1434 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1435 BUG_ON(!list_empty(&ci->i_dirty_item));
1436 spin_lock(&mdsc->cap_dirty_lock);
1437 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1438 spin_unlock(&mdsc->cap_dirty_lock);
1439 if (ci->i_flushing_caps == 0) {
1441 dirty |= I_DIRTY_SYNC;
1444 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1446 BUG_ON(list_empty(&ci->i_dirty_item));
1447 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1448 (mask & CEPH_CAP_FILE_BUFFER))
1449 dirty |= I_DIRTY_DATASYNC;
1450 __cap_delay_requeue(mdsc, ci);
1454 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1456 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1459 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1462 kmem_cache_free(ceph_cap_flush_cachep, cf);
1465 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1467 if (!list_empty(&mdsc->cap_flush_list)) {
1468 struct ceph_cap_flush *cf =
1469 list_first_entry(&mdsc->cap_flush_list,
1470 struct ceph_cap_flush, g_list);
1477 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1478 * Return true if caller needs to wake up flush waiters.
1480 static bool __finish_cap_flush(struct ceph_mds_client *mdsc,
1481 struct ceph_inode_info *ci,
1482 struct ceph_cap_flush *cf)
1484 struct ceph_cap_flush *prev;
1485 bool wake = cf->wake;
1487 /* are there older pending cap flushes? */
1488 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1489 prev = list_prev_entry(cf, g_list);
1493 list_del(&cf->g_list);
1495 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1496 prev = list_prev_entry(cf, i_list);
1500 list_del(&cf->i_list);
1508 * Add dirty inode to the flushing list. Assigned a seq number so we
1509 * can wait for caps to flush without starving.
1511 * Called under i_ceph_lock.
1513 static int __mark_caps_flushing(struct inode *inode,
1514 struct ceph_mds_session *session, bool wake,
1515 u64 *flush_tid, u64 *oldest_flush_tid)
1517 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1518 struct ceph_inode_info *ci = ceph_inode(inode);
1519 struct ceph_cap_flush *cf = NULL;
1522 BUG_ON(ci->i_dirty_caps == 0);
1523 BUG_ON(list_empty(&ci->i_dirty_item));
1524 BUG_ON(!ci->i_prealloc_cap_flush);
1526 flushing = ci->i_dirty_caps;
1527 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1528 ceph_cap_string(flushing),
1529 ceph_cap_string(ci->i_flushing_caps),
1530 ceph_cap_string(ci->i_flushing_caps | flushing));
1531 ci->i_flushing_caps |= flushing;
1532 ci->i_dirty_caps = 0;
1533 dout(" inode %p now !dirty\n", inode);
1535 swap(cf, ci->i_prealloc_cap_flush);
1536 cf->caps = flushing;
1539 spin_lock(&mdsc->cap_dirty_lock);
1540 list_del_init(&ci->i_dirty_item);
1542 cf->tid = ++mdsc->last_cap_flush_tid;
1543 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1544 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1546 if (list_empty(&ci->i_flushing_item)) {
1547 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1548 mdsc->num_cap_flushing++;
1550 spin_unlock(&mdsc->cap_dirty_lock);
1552 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1554 *flush_tid = cf->tid;
1559 * try to invalidate mapping pages without blocking.
1561 static int try_nonblocking_invalidate(struct inode *inode)
1563 struct ceph_inode_info *ci = ceph_inode(inode);
1564 u32 invalidating_gen = ci->i_rdcache_gen;
1566 spin_unlock(&ci->i_ceph_lock);
1567 invalidate_mapping_pages(&inode->i_data, 0, -1);
1568 spin_lock(&ci->i_ceph_lock);
1570 if (inode->i_data.nrpages == 0 &&
1571 invalidating_gen == ci->i_rdcache_gen) {
1573 dout("try_nonblocking_invalidate %p success\n", inode);
1574 /* save any racing async invalidate some trouble */
1575 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1578 dout("try_nonblocking_invalidate %p failed\n", inode);
1583 * Swiss army knife function to examine currently used and wanted
1584 * versus held caps. Release, flush, ack revoked caps to mds as
1587 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1588 * cap release further.
1589 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1590 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1593 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1594 struct ceph_mds_session *session)
1596 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1597 struct ceph_mds_client *mdsc = fsc->mdsc;
1598 struct inode *inode = &ci->vfs_inode;
1599 struct ceph_cap *cap;
1600 u64 flush_tid, oldest_flush_tid;
1601 int file_wanted, used, cap_used;
1602 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1603 int issued, implemented, want, retain, revoking, flushing = 0;
1604 int mds = -1; /* keep track of how far we've gone through i_caps list
1605 to avoid an infinite loop on retry */
1607 int delayed = 0, sent = 0, num;
1608 bool is_delayed = flags & CHECK_CAPS_NODELAY;
1609 bool queue_invalidate = false;
1610 bool force_requeue = false;
1611 bool tried_invalidate = false;
1613 /* if we are unmounting, flush any unused caps immediately. */
1617 spin_lock(&ci->i_ceph_lock);
1619 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1620 flags |= CHECK_CAPS_FLUSH;
1624 spin_lock(&ci->i_ceph_lock);
1626 file_wanted = __ceph_caps_file_wanted(ci);
1627 used = __ceph_caps_used(ci);
1628 issued = __ceph_caps_issued(ci, &implemented);
1629 revoking = implemented & ~issued;
1632 retain = file_wanted | used | CEPH_CAP_PIN;
1633 if (!mdsc->stopping && inode->i_nlink > 0) {
1635 retain |= CEPH_CAP_ANY; /* be greedy */
1636 } else if (S_ISDIR(inode->i_mode) &&
1637 (issued & CEPH_CAP_FILE_SHARED) &&
1638 __ceph_dir_is_complete(ci)) {
1640 * If a directory is complete, we want to keep
1641 * the exclusive cap. So that MDS does not end up
1642 * revoking the shared cap on every create/unlink
1645 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1649 retain |= CEPH_CAP_ANY_SHARED;
1651 * keep RD only if we didn't have the file open RW,
1652 * because then the mds would revoke it anyway to
1653 * journal max_size=0.
1655 if (ci->i_max_size == 0)
1656 retain |= CEPH_CAP_ANY_RD;
1660 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1661 " issued %s revoking %s retain %s %s%s%s\n", inode,
1662 ceph_cap_string(file_wanted),
1663 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1664 ceph_cap_string(ci->i_flushing_caps),
1665 ceph_cap_string(issued), ceph_cap_string(revoking),
1666 ceph_cap_string(retain),
1667 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1668 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1669 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1672 * If we no longer need to hold onto old our caps, and we may
1673 * have cached pages, but don't want them, then try to invalidate.
1674 * If we fail, it's because pages are locked.... try again later.
1676 if ((!is_delayed || mdsc->stopping) &&
1677 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */
1678 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1679 inode->i_data.nrpages && /* have cached pages */
1680 (revoking & (CEPH_CAP_FILE_CACHE|
1681 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1682 !tried_invalidate) {
1683 dout("check_caps trying to invalidate on %p\n", inode);
1684 if (try_nonblocking_invalidate(inode) < 0) {
1685 if (revoking & (CEPH_CAP_FILE_CACHE|
1686 CEPH_CAP_FILE_LAZYIO)) {
1687 dout("check_caps queuing invalidate\n");
1688 queue_invalidate = true;
1689 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1691 dout("check_caps failed to invalidate pages\n");
1692 /* we failed to invalidate pages. check these
1693 caps again later. */
1694 force_requeue = true;
1695 __cap_set_timeouts(mdsc, ci);
1698 tried_invalidate = true;
1703 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1704 cap = rb_entry(p, struct ceph_cap, ci_node);
1707 /* avoid looping forever */
1708 if (mds >= cap->mds ||
1709 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1712 /* NOTE: no side-effects allowed, until we take s_mutex */
1715 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1716 cap_used &= ~ci->i_auth_cap->issued;
1718 revoking = cap->implemented & ~cap->issued;
1719 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1720 cap->mds, cap, ceph_cap_string(cap_used),
1721 ceph_cap_string(cap->issued),
1722 ceph_cap_string(cap->implemented),
1723 ceph_cap_string(revoking));
1725 if (cap == ci->i_auth_cap &&
1726 (cap->issued & CEPH_CAP_FILE_WR)) {
1727 /* request larger max_size from MDS? */
1728 if (ci->i_wanted_max_size > ci->i_max_size &&
1729 ci->i_wanted_max_size > ci->i_requested_max_size) {
1730 dout("requesting new max_size\n");
1734 /* approaching file_max? */
1735 if ((inode->i_size << 1) >= ci->i_max_size &&
1736 (ci->i_reported_size << 1) < ci->i_max_size) {
1737 dout("i_size approaching max_size\n");
1741 /* flush anything dirty? */
1742 if (cap == ci->i_auth_cap) {
1743 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
1744 dout("flushing dirty caps\n");
1747 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
1748 dout("flushing snap caps\n");
1753 /* completed revocation? going down and there are no caps? */
1754 if (revoking && (revoking & cap_used) == 0) {
1755 dout("completed revocation of %s\n",
1756 ceph_cap_string(cap->implemented & ~cap->issued));
1760 /* want more caps from mds? */
1761 if (want & ~(cap->mds_wanted | cap->issued))
1764 /* things we might delay */
1765 if ((cap->issued & ~retain) == 0 &&
1766 cap->mds_wanted == want)
1767 continue; /* nope, all good */
1773 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1774 time_before(jiffies, ci->i_hold_caps_max)) {
1775 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1776 ceph_cap_string(cap->issued),
1777 ceph_cap_string(cap->issued & retain),
1778 ceph_cap_string(cap->mds_wanted),
1779 ceph_cap_string(want));
1785 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1786 dout(" skipping %p I_NOFLUSH set\n", inode);
1790 if (session && session != cap->session) {
1791 dout("oops, wrong session %p mutex\n", session);
1792 mutex_unlock(&session->s_mutex);
1796 session = cap->session;
1797 if (mutex_trylock(&session->s_mutex) == 0) {
1798 dout("inverting session/ino locks on %p\n",
1800 spin_unlock(&ci->i_ceph_lock);
1801 if (took_snap_rwsem) {
1802 up_read(&mdsc->snap_rwsem);
1803 took_snap_rwsem = 0;
1805 mutex_lock(&session->s_mutex);
1810 /* kick flushing and flush snaps before sending normal
1812 if (cap == ci->i_auth_cap &&
1814 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
1815 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
1816 spin_lock(&mdsc->cap_dirty_lock);
1817 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1818 spin_unlock(&mdsc->cap_dirty_lock);
1819 __kick_flushing_caps(mdsc, session, ci,
1821 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
1823 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
1824 __ceph_flush_snaps(ci, session);
1829 /* take snap_rwsem after session mutex */
1830 if (!took_snap_rwsem) {
1831 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1832 dout("inverting snap/in locks on %p\n",
1834 spin_unlock(&ci->i_ceph_lock);
1835 down_read(&mdsc->snap_rwsem);
1836 took_snap_rwsem = 1;
1839 took_snap_rwsem = 1;
1842 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
1843 flushing = __mark_caps_flushing(inode, session, false,
1849 spin_lock(&mdsc->cap_dirty_lock);
1850 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1851 spin_unlock(&mdsc->cap_dirty_lock);
1854 mds = cap->mds; /* remember mds, so we don't repeat */
1857 /* __send_cap drops i_ceph_lock */
1858 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used,
1859 want, retain, flushing,
1860 flush_tid, oldest_flush_tid);
1861 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1865 * Reschedule delayed caps release if we delayed anything,
1868 if (delayed && is_delayed)
1869 force_requeue = true; /* __send_cap delayed release; requeue */
1870 if (!delayed && !is_delayed)
1871 __cap_delay_cancel(mdsc, ci);
1872 else if (!is_delayed || force_requeue)
1873 __cap_delay_requeue(mdsc, ci);
1875 spin_unlock(&ci->i_ceph_lock);
1877 if (queue_invalidate)
1878 ceph_queue_invalidate(inode);
1881 mutex_unlock(&session->s_mutex);
1882 if (took_snap_rwsem)
1883 up_read(&mdsc->snap_rwsem);
1887 * Try to flush dirty caps back to the auth mds.
1889 static int try_flush_caps(struct inode *inode, u64 *ptid)
1891 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1892 struct ceph_inode_info *ci = ceph_inode(inode);
1893 struct ceph_mds_session *session = NULL;
1895 u64 flush_tid = 0, oldest_flush_tid = 0;
1898 spin_lock(&ci->i_ceph_lock);
1899 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1900 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1903 if (ci->i_dirty_caps && ci->i_auth_cap) {
1904 struct ceph_cap *cap = ci->i_auth_cap;
1905 int used = __ceph_caps_used(ci);
1906 int want = __ceph_caps_wanted(ci);
1909 if (!session || session != cap->session) {
1910 spin_unlock(&ci->i_ceph_lock);
1912 mutex_unlock(&session->s_mutex);
1913 session = cap->session;
1914 mutex_lock(&session->s_mutex);
1917 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1920 flushing = __mark_caps_flushing(inode, session, true,
1921 &flush_tid, &oldest_flush_tid);
1923 /* __send_cap drops i_ceph_lock */
1924 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1925 (cap->issued | cap->implemented),
1926 flushing, flush_tid, oldest_flush_tid);
1929 spin_lock(&ci->i_ceph_lock);
1930 __cap_delay_requeue(mdsc, ci);
1931 spin_unlock(&ci->i_ceph_lock);
1934 if (!list_empty(&ci->i_cap_flush_list)) {
1935 struct ceph_cap_flush *cf =
1936 list_last_entry(&ci->i_cap_flush_list,
1937 struct ceph_cap_flush, i_list);
1939 flush_tid = cf->tid;
1941 flushing = ci->i_flushing_caps;
1942 spin_unlock(&ci->i_ceph_lock);
1946 mutex_unlock(&session->s_mutex);
1953 * Return true if we've flushed caps through the given flush_tid.
1955 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
1957 struct ceph_inode_info *ci = ceph_inode(inode);
1960 spin_lock(&ci->i_ceph_lock);
1961 if (!list_empty(&ci->i_cap_flush_list)) {
1962 struct ceph_cap_flush * cf =
1963 list_first_entry(&ci->i_cap_flush_list,
1964 struct ceph_cap_flush, i_list);
1965 if (cf->tid <= flush_tid)
1968 spin_unlock(&ci->i_ceph_lock);
1973 * wait for any unsafe requests to complete.
1975 static int unsafe_request_wait(struct inode *inode)
1977 struct ceph_inode_info *ci = ceph_inode(inode);
1978 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
1981 spin_lock(&ci->i_unsafe_lock);
1982 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
1983 req1 = list_last_entry(&ci->i_unsafe_dirops,
1984 struct ceph_mds_request,
1986 ceph_mdsc_get_request(req1);
1988 if (!list_empty(&ci->i_unsafe_iops)) {
1989 req2 = list_last_entry(&ci->i_unsafe_iops,
1990 struct ceph_mds_request,
1991 r_unsafe_target_item);
1992 ceph_mdsc_get_request(req2);
1994 spin_unlock(&ci->i_unsafe_lock);
1996 dout("unsafe_requeset_wait %p wait on tid %llu %llu\n",
1997 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
1999 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2000 ceph_timeout_jiffies(req1->r_timeout));
2003 ceph_mdsc_put_request(req1);
2006 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2007 ceph_timeout_jiffies(req2->r_timeout));
2010 ceph_mdsc_put_request(req2);
2015 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2017 struct inode *inode = file->f_mapping->host;
2018 struct ceph_inode_info *ci = ceph_inode(inode);
2023 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2025 ceph_sync_write_wait(inode);
2027 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
2036 dirty = try_flush_caps(inode, &flush_tid);
2037 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2039 ret = unsafe_request_wait(inode);
2042 * only wait on non-file metadata writeback (the mds
2043 * can recover size and mtime, so we don't need to
2046 if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2047 ret = wait_event_interruptible(ci->i_cap_wq,
2048 caps_are_flushed(inode, flush_tid));
2050 inode_unlock(inode);
2052 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2057 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2058 * queue inode for flush but don't do so immediately, because we can
2059 * get by with fewer MDS messages if we wait for data writeback to
2062 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2064 struct ceph_inode_info *ci = ceph_inode(inode);
2068 int wait = wbc->sync_mode == WB_SYNC_ALL;
2070 dout("write_inode %p wait=%d\n", inode, wait);
2072 dirty = try_flush_caps(inode, &flush_tid);
2074 err = wait_event_interruptible(ci->i_cap_wq,
2075 caps_are_flushed(inode, flush_tid));
2077 struct ceph_mds_client *mdsc =
2078 ceph_sb_to_client(inode->i_sb)->mdsc;
2080 spin_lock(&ci->i_ceph_lock);
2081 if (__ceph_caps_dirty(ci))
2082 __cap_delay_requeue_front(mdsc, ci);
2083 spin_unlock(&ci->i_ceph_lock);
2088 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2089 struct ceph_mds_session *session,
2090 struct ceph_inode_info *ci,
2091 u64 oldest_flush_tid)
2092 __releases(ci->i_ceph_lock)
2093 __acquires(ci->i_ceph_lock)
2095 struct inode *inode = &ci->vfs_inode;
2096 struct ceph_cap *cap;
2097 struct ceph_cap_flush *cf;
2101 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2102 if (cf->tid < first_tid)
2105 cap = ci->i_auth_cap;
2106 if (!(cap && cap->session == session)) {
2107 pr_err("%p auth cap %p not mds%d ???\n",
2108 inode, cap, session->s_mds);
2112 first_tid = cf->tid + 1;
2115 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2116 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2117 ci->i_ceph_flags |= CEPH_I_NODELAY;
2118 ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2119 __ceph_caps_used(ci),
2120 __ceph_caps_wanted(ci),
2121 cap->issued | cap->implemented,
2122 cf->caps, cf->tid, oldest_flush_tid);
2124 pr_err("kick_flushing_caps: error sending "
2125 "cap flush, ino (%llx.%llx) "
2126 "tid %llu flushing %s\n",
2127 ceph_vinop(inode), cf->tid,
2128 ceph_cap_string(cf->caps));
2131 struct ceph_cap_snap *capsnap =
2132 container_of(cf, struct ceph_cap_snap,
2134 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2135 inode, capsnap, cf->tid,
2136 ceph_cap_string(capsnap->dirty));
2138 atomic_inc(&capsnap->nref);
2139 spin_unlock(&ci->i_ceph_lock);
2141 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2144 pr_err("kick_flushing_caps: error sending "
2145 "cap flushsnap, ino (%llx.%llx) "
2146 "tid %llu follows %llu\n",
2147 ceph_vinop(inode), cf->tid,
2151 ceph_put_cap_snap(capsnap);
2154 spin_lock(&ci->i_ceph_lock);
2158 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2159 struct ceph_mds_session *session)
2161 struct ceph_inode_info *ci;
2162 struct ceph_cap *cap;
2163 u64 oldest_flush_tid;
2165 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2167 spin_lock(&mdsc->cap_dirty_lock);
2168 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2169 spin_unlock(&mdsc->cap_dirty_lock);
2171 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2172 spin_lock(&ci->i_ceph_lock);
2173 cap = ci->i_auth_cap;
2174 if (!(cap && cap->session == session)) {
2175 pr_err("%p auth cap %p not mds%d ???\n",
2176 &ci->vfs_inode, cap, session->s_mds);
2177 spin_unlock(&ci->i_ceph_lock);
2183 * if flushing caps were revoked, we re-send the cap flush
2184 * in client reconnect stage. This guarantees MDS * processes
2185 * the cap flush message before issuing the flushing caps to
2188 if ((cap->issued & ci->i_flushing_caps) !=
2189 ci->i_flushing_caps) {
2190 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2191 __kick_flushing_caps(mdsc, session, ci,
2194 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2197 spin_unlock(&ci->i_ceph_lock);
2201 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2202 struct ceph_mds_session *session)
2204 struct ceph_inode_info *ci;
2205 struct ceph_cap *cap;
2206 u64 oldest_flush_tid;
2208 dout("kick_flushing_caps mds%d\n", session->s_mds);
2210 spin_lock(&mdsc->cap_dirty_lock);
2211 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2212 spin_unlock(&mdsc->cap_dirty_lock);
2214 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2215 spin_lock(&ci->i_ceph_lock);
2216 cap = ci->i_auth_cap;
2217 if (!(cap && cap->session == session)) {
2218 pr_err("%p auth cap %p not mds%d ???\n",
2219 &ci->vfs_inode, cap, session->s_mds);
2220 spin_unlock(&ci->i_ceph_lock);
2223 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2224 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2225 __kick_flushing_caps(mdsc, session, ci,
2228 spin_unlock(&ci->i_ceph_lock);
2232 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2233 struct ceph_mds_session *session,
2234 struct inode *inode)
2235 __releases(ci->i_ceph_lock)
2237 struct ceph_inode_info *ci = ceph_inode(inode);
2238 struct ceph_cap *cap;
2240 cap = ci->i_auth_cap;
2241 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2242 ceph_cap_string(ci->i_flushing_caps));
2244 if (!list_empty(&ci->i_cap_flush_list)) {
2245 u64 oldest_flush_tid;
2246 spin_lock(&mdsc->cap_dirty_lock);
2247 list_move_tail(&ci->i_flushing_item,
2248 &cap->session->s_cap_flushing);
2249 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2250 spin_unlock(&mdsc->cap_dirty_lock);
2252 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2253 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2254 spin_unlock(&ci->i_ceph_lock);
2256 spin_unlock(&ci->i_ceph_lock);
2262 * Take references to capabilities we hold, so that we don't release
2263 * them to the MDS prematurely.
2265 * Protected by i_ceph_lock.
2267 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2268 bool snap_rwsem_locked)
2270 if (got & CEPH_CAP_PIN)
2272 if (got & CEPH_CAP_FILE_RD)
2274 if (got & CEPH_CAP_FILE_CACHE)
2275 ci->i_rdcache_ref++;
2276 if (got & CEPH_CAP_FILE_WR) {
2277 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2278 BUG_ON(!snap_rwsem_locked);
2279 ci->i_head_snapc = ceph_get_snap_context(
2280 ci->i_snap_realm->cached_context);
2284 if (got & CEPH_CAP_FILE_BUFFER) {
2285 if (ci->i_wb_ref == 0)
2286 ihold(&ci->vfs_inode);
2288 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2289 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2294 * Try to grab cap references. Specify those refs we @want, and the
2295 * minimal set we @need. Also include the larger offset we are writing
2296 * to (when applicable), and check against max_size here as well.
2297 * Note that caller is responsible for ensuring max_size increases are
2298 * requested from the MDS.
2300 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2301 loff_t endoff, bool nonblock, int *got, int *err)
2303 struct inode *inode = &ci->vfs_inode;
2304 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2306 int have, implemented;
2308 bool snap_rwsem_locked = false;
2310 dout("get_cap_refs %p need %s want %s\n", inode,
2311 ceph_cap_string(need), ceph_cap_string(want));
2314 spin_lock(&ci->i_ceph_lock);
2316 /* make sure file is actually open */
2317 file_wanted = __ceph_caps_file_wanted(ci);
2318 if ((file_wanted & need) != need) {
2319 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2320 ceph_cap_string(need), ceph_cap_string(file_wanted));
2326 /* finish pending truncate */
2327 while (ci->i_truncate_pending) {
2328 spin_unlock(&ci->i_ceph_lock);
2329 if (snap_rwsem_locked) {
2330 up_read(&mdsc->snap_rwsem);
2331 snap_rwsem_locked = false;
2333 __ceph_do_pending_vmtruncate(inode);
2334 spin_lock(&ci->i_ceph_lock);
2337 have = __ceph_caps_issued(ci, &implemented);
2339 if (have & need & CEPH_CAP_FILE_WR) {
2340 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2341 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2342 inode, endoff, ci->i_max_size);
2343 if (endoff > ci->i_requested_max_size) {
2350 * If a sync write is in progress, we must wait, so that we
2351 * can get a final snapshot value for size+mtime.
2353 if (__ceph_have_pending_cap_snap(ci)) {
2354 dout("get_cap_refs %p cap_snap_pending\n", inode);
2359 if ((have & need) == need) {
2361 * Look at (implemented & ~have & not) so that we keep waiting
2362 * on transition from wanted -> needed caps. This is needed
2363 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2364 * going before a prior buffered writeback happens.
2366 int not = want & ~(have & need);
2367 int revoking = implemented & ~have;
2368 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2369 inode, ceph_cap_string(have), ceph_cap_string(not),
2370 ceph_cap_string(revoking));
2371 if ((revoking & not) == 0) {
2372 if (!snap_rwsem_locked &&
2373 !ci->i_head_snapc &&
2374 (need & CEPH_CAP_FILE_WR)) {
2375 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2377 * we can not call down_read() when
2378 * task isn't in TASK_RUNNING state
2386 spin_unlock(&ci->i_ceph_lock);
2387 down_read(&mdsc->snap_rwsem);
2388 snap_rwsem_locked = true;
2391 snap_rwsem_locked = true;
2393 *got = need | (have & want);
2394 if ((need & CEPH_CAP_FILE_RD) &&
2395 !(*got & CEPH_CAP_FILE_CACHE))
2396 ceph_disable_fscache_readpage(ci);
2397 __take_cap_refs(ci, *got, true);
2401 int session_readonly = false;
2402 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2403 struct ceph_mds_session *s = ci->i_auth_cap->session;
2404 spin_lock(&s->s_cap_lock);
2405 session_readonly = s->s_readonly;
2406 spin_unlock(&s->s_cap_lock);
2408 if (session_readonly) {
2409 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2410 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2416 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) {
2418 if (ACCESS_ONCE(mdsc->fsc->mount_state) ==
2419 CEPH_MOUNT_SHUTDOWN) {
2420 dout("get_cap_refs %p forced umount\n", inode);
2425 mds_wanted = __ceph_caps_mds_wanted(ci);
2426 if ((mds_wanted & need) != need) {
2427 dout("get_cap_refs %p caps were dropped"
2428 " (session killed?)\n", inode);
2433 if ((mds_wanted & file_wanted) ==
2434 (file_wanted & (CEPH_CAP_FILE_RD|CEPH_CAP_FILE_WR)))
2435 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED;
2438 dout("get_cap_refs %p have %s needed %s\n", inode,
2439 ceph_cap_string(have), ceph_cap_string(need));
2442 spin_unlock(&ci->i_ceph_lock);
2443 if (snap_rwsem_locked)
2444 up_read(&mdsc->snap_rwsem);
2446 dout("get_cap_refs %p ret %d got %s\n", inode,
2447 ret, ceph_cap_string(*got));
2452 * Check the offset we are writing up to against our current
2453 * max_size. If necessary, tell the MDS we want to write to
2456 static void check_max_size(struct inode *inode, loff_t endoff)
2458 struct ceph_inode_info *ci = ceph_inode(inode);
2461 /* do we need to explicitly request a larger max_size? */
2462 spin_lock(&ci->i_ceph_lock);
2463 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2464 dout("write %p at large endoff %llu, req max_size\n",
2466 ci->i_wanted_max_size = endoff;
2468 /* duplicate ceph_check_caps()'s logic */
2469 if (ci->i_auth_cap &&
2470 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2471 ci->i_wanted_max_size > ci->i_max_size &&
2472 ci->i_wanted_max_size > ci->i_requested_max_size)
2474 spin_unlock(&ci->i_ceph_lock);
2476 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2480 * Wait for caps, and take cap references. If we can't get a WR cap
2481 * due to a small max_size, make sure we check_max_size (and possibly
2482 * ask the mds) so we don't get hung up indefinitely.
2484 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2485 loff_t endoff, int *got, struct page **pinned_page)
2487 int _got, ret, err = 0;
2489 ret = ceph_pool_perm_check(ci, need);
2495 check_max_size(&ci->vfs_inode, endoff);
2499 ret = try_get_cap_refs(ci, need, want, endoff,
2500 false, &_got, &err);
2507 ret = wait_event_interruptible(ci->i_cap_wq,
2508 try_get_cap_refs(ci, need, want, endoff,
2509 true, &_got, &err));
2516 if (err == -ESTALE) {
2517 /* session was killed, try renew caps */
2518 ret = ceph_renew_caps(&ci->vfs_inode);
2525 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2526 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2527 i_size_read(&ci->vfs_inode) > 0) {
2529 find_get_page(ci->vfs_inode.i_mapping, 0);
2531 if (PageUptodate(page)) {
2532 *pinned_page = page;
2538 * drop cap refs first because getattr while
2539 * holding * caps refs can cause deadlock.
2541 ceph_put_cap_refs(ci, _got);
2545 * getattr request will bring inline data into
2548 ret = __ceph_do_getattr(&ci->vfs_inode, NULL,
2549 CEPH_STAT_CAP_INLINE_DATA,
2558 if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
2559 ceph_fscache_revalidate_cookie(ci);
2566 * Take cap refs. Caller must already know we hold at least one ref
2567 * on the caps in question or we don't know this is safe.
2569 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2571 spin_lock(&ci->i_ceph_lock);
2572 __take_cap_refs(ci, caps, false);
2573 spin_unlock(&ci->i_ceph_lock);
2578 * drop cap_snap that is not associated with any snapshot.
2579 * we don't need to send FLUSHSNAP message for it.
2581 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2582 struct ceph_cap_snap *capsnap)
2584 if (!capsnap->need_flush &&
2585 !capsnap->writing && !capsnap->dirty_pages) {
2586 dout("dropping cap_snap %p follows %llu\n",
2587 capsnap, capsnap->follows);
2588 BUG_ON(capsnap->cap_flush.tid > 0);
2589 ceph_put_snap_context(capsnap->context);
2590 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2591 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2593 list_del(&capsnap->ci_item);
2594 ceph_put_cap_snap(capsnap);
2603 * If we released the last ref on any given cap, call ceph_check_caps
2604 * to release (or schedule a release).
2606 * If we are releasing a WR cap (from a sync write), finalize any affected
2607 * cap_snap, and wake up any waiters.
2609 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2611 struct inode *inode = &ci->vfs_inode;
2612 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2614 spin_lock(&ci->i_ceph_lock);
2615 if (had & CEPH_CAP_PIN)
2617 if (had & CEPH_CAP_FILE_RD)
2618 if (--ci->i_rd_ref == 0)
2620 if (had & CEPH_CAP_FILE_CACHE)
2621 if (--ci->i_rdcache_ref == 0)
2623 if (had & CEPH_CAP_FILE_BUFFER) {
2624 if (--ci->i_wb_ref == 0) {
2628 dout("put_cap_refs %p wb %d -> %d (?)\n",
2629 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2631 if (had & CEPH_CAP_FILE_WR)
2632 if (--ci->i_wr_ref == 0) {
2634 if (__ceph_have_pending_cap_snap(ci)) {
2635 struct ceph_cap_snap *capsnap =
2636 list_last_entry(&ci->i_cap_snaps,
2637 struct ceph_cap_snap,
2639 capsnap->writing = 0;
2640 if (ceph_try_drop_cap_snap(ci, capsnap))
2642 else if (__ceph_finish_cap_snap(ci, capsnap))
2646 if (ci->i_wrbuffer_ref_head == 0 &&
2647 ci->i_dirty_caps == 0 &&
2648 ci->i_flushing_caps == 0) {
2649 BUG_ON(!ci->i_head_snapc);
2650 ceph_put_snap_context(ci->i_head_snapc);
2651 ci->i_head_snapc = NULL;
2653 /* see comment in __ceph_remove_cap() */
2654 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2655 drop_inode_snap_realm(ci);
2657 spin_unlock(&ci->i_ceph_lock);
2659 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2660 last ? " last" : "", put ? " put" : "");
2662 if (last && !flushsnaps)
2663 ceph_check_caps(ci, 0, NULL);
2664 else if (flushsnaps)
2665 ceph_flush_snaps(ci, NULL);
2667 wake_up_all(&ci->i_cap_wq);
2673 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2674 * context. Adjust per-snap dirty page accounting as appropriate.
2675 * Once all dirty data for a cap_snap is flushed, flush snapped file
2676 * metadata back to the MDS. If we dropped the last ref, call
2679 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2680 struct ceph_snap_context *snapc)
2682 struct inode *inode = &ci->vfs_inode;
2683 struct ceph_cap_snap *capsnap = NULL;
2687 bool flush_snaps = false;
2688 bool complete_capsnap = false;
2690 spin_lock(&ci->i_ceph_lock);
2691 ci->i_wrbuffer_ref -= nr;
2692 if (ci->i_wrbuffer_ref == 0) {
2697 if (ci->i_head_snapc == snapc) {
2698 ci->i_wrbuffer_ref_head -= nr;
2699 if (ci->i_wrbuffer_ref_head == 0 &&
2700 ci->i_wr_ref == 0 &&
2701 ci->i_dirty_caps == 0 &&
2702 ci->i_flushing_caps == 0) {
2703 BUG_ON(!ci->i_head_snapc);
2704 ceph_put_snap_context(ci->i_head_snapc);
2705 ci->i_head_snapc = NULL;
2707 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2709 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2710 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2711 last ? " LAST" : "");
2713 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2714 if (capsnap->context == snapc) {
2720 capsnap->dirty_pages -= nr;
2721 if (capsnap->dirty_pages == 0) {
2722 complete_capsnap = true;
2723 if (!capsnap->writing) {
2724 if (ceph_try_drop_cap_snap(ci, capsnap)) {
2727 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2732 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2733 " snap %lld %d/%d -> %d/%d %s%s\n",
2734 inode, capsnap, capsnap->context->seq,
2735 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2736 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2737 last ? " (wrbuffer last)" : "",
2738 complete_capsnap ? " (complete capsnap)" : "");
2741 spin_unlock(&ci->i_ceph_lock);
2744 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2745 } else if (flush_snaps) {
2746 ceph_flush_snaps(ci, NULL);
2748 if (complete_capsnap)
2749 wake_up_all(&ci->i_cap_wq);
2755 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2757 static void invalidate_aliases(struct inode *inode)
2759 struct dentry *dn, *prev = NULL;
2761 dout("invalidate_aliases inode %p\n", inode);
2762 d_prune_aliases(inode);
2764 * For non-directory inode, d_find_alias() only returns
2765 * hashed dentry. After calling d_invalidate(), the
2766 * dentry becomes unhashed.
2768 * For directory inode, d_find_alias() can return
2769 * unhashed dentry. But directory inode should have
2770 * one alias at most.
2772 while ((dn = d_find_alias(inode))) {
2787 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2788 * actually be a revocation if it specifies a smaller cap set.)
2790 * caller holds s_mutex and i_ceph_lock, we drop both.
2792 static void handle_cap_grant(struct ceph_mds_client *mdsc,
2793 struct inode *inode, struct ceph_mds_caps *grant,
2794 struct ceph_string **pns, u64 inline_version,
2795 void *inline_data, u32 inline_len,
2796 struct ceph_buffer *xattr_buf,
2797 struct ceph_mds_session *session,
2798 struct ceph_cap *cap, int issued)
2799 __releases(ci->i_ceph_lock)
2800 __releases(mdsc->snap_rwsem)
2802 struct ceph_inode_info *ci = ceph_inode(inode);
2803 int mds = session->s_mds;
2804 int seq = le32_to_cpu(grant->seq);
2805 int newcaps = le32_to_cpu(grant->caps);
2806 int used, wanted, dirty;
2807 u64 size = le64_to_cpu(grant->size);
2808 u64 max_size = le64_to_cpu(grant->max_size);
2809 struct timespec mtime, atime, ctime;
2812 bool writeback = false;
2813 bool queue_trunc = false;
2814 bool queue_invalidate = false;
2815 bool deleted_inode = false;
2816 bool fill_inline = false;
2818 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2819 inode, cap, mds, seq, ceph_cap_string(newcaps));
2820 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2825 * auth mds of the inode changed. we received the cap export message,
2826 * but still haven't received the cap import message. handle_cap_export
2827 * updated the new auth MDS' cap.
2829 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2830 * that was sent before the cap import message. So don't remove caps.
2832 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
2833 WARN_ON(cap != ci->i_auth_cap);
2834 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
2836 newcaps |= cap->issued;
2840 * If CACHE is being revoked, and we have no dirty buffers,
2841 * try to invalidate (once). (If there are dirty buffers, we
2842 * will invalidate _after_ writeback.)
2844 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
2845 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2846 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2847 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
2848 if (try_nonblocking_invalidate(inode)) {
2849 /* there were locked pages.. invalidate later
2850 in a separate thread. */
2851 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2852 queue_invalidate = true;
2853 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2858 /* side effects now are allowed */
2859 cap->cap_gen = session->s_cap_gen;
2862 __check_cap_issue(ci, cap, newcaps);
2864 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2865 (issued & CEPH_CAP_AUTH_EXCL) == 0) {
2866 inode->i_mode = le32_to_cpu(grant->mode);
2867 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
2868 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
2869 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2870 from_kuid(&init_user_ns, inode->i_uid),
2871 from_kgid(&init_user_ns, inode->i_gid));
2874 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2875 (issued & CEPH_CAP_LINK_EXCL) == 0) {
2876 set_nlink(inode, le32_to_cpu(grant->nlink));
2877 if (inode->i_nlink == 0 &&
2878 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
2879 deleted_inode = true;
2882 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2883 int len = le32_to_cpu(grant->xattr_len);
2884 u64 version = le64_to_cpu(grant->xattr_version);
2886 if (version > ci->i_xattrs.version) {
2887 dout(" got new xattrs v%llu on %p len %d\n",
2888 version, inode, len);
2889 if (ci->i_xattrs.blob)
2890 ceph_buffer_put(ci->i_xattrs.blob);
2891 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2892 ci->i_xattrs.version = version;
2893 ceph_forget_all_cached_acls(inode);
2897 if (newcaps & CEPH_CAP_ANY_RD) {
2898 /* ctime/mtime/atime? */
2899 ceph_decode_timespec(&mtime, &grant->mtime);
2900 ceph_decode_timespec(&atime, &grant->atime);
2901 ceph_decode_timespec(&ctime, &grant->ctime);
2902 ceph_fill_file_time(inode, issued,
2903 le32_to_cpu(grant->time_warp_seq),
2904 &ctime, &mtime, &atime);
2907 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
2908 /* file layout may have changed */
2909 s64 old_pool = ci->i_layout.pool_id;
2910 struct ceph_string *old_ns;
2912 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
2913 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
2914 lockdep_is_held(&ci->i_ceph_lock));
2915 rcu_assign_pointer(ci->i_layout.pool_ns, *pns);
2917 if (ci->i_layout.pool_id != old_pool || *pns != old_ns)
2918 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
2922 /* size/truncate_seq? */
2923 queue_trunc = ceph_fill_file_size(inode, issued,
2924 le32_to_cpu(grant->truncate_seq),
2925 le64_to_cpu(grant->truncate_size),
2927 /* max size increase? */
2928 if (ci->i_auth_cap == cap && max_size != ci->i_max_size) {
2929 dout("max_size %lld -> %llu\n",
2930 ci->i_max_size, max_size);
2931 ci->i_max_size = max_size;
2932 if (max_size >= ci->i_wanted_max_size) {
2933 ci->i_wanted_max_size = 0; /* reset */
2934 ci->i_requested_max_size = 0;
2940 /* check cap bits */
2941 wanted = __ceph_caps_wanted(ci);
2942 used = __ceph_caps_used(ci);
2943 dirty = __ceph_caps_dirty(ci);
2944 dout(" my wanted = %s, used = %s, dirty %s\n",
2945 ceph_cap_string(wanted),
2946 ceph_cap_string(used),
2947 ceph_cap_string(dirty));
2948 if (wanted != le32_to_cpu(grant->wanted)) {
2949 dout("mds wanted %s -> %s\n",
2950 ceph_cap_string(le32_to_cpu(grant->wanted)),
2951 ceph_cap_string(wanted));
2952 /* imported cap may not have correct mds_wanted */
2953 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
2957 /* revocation, grant, or no-op? */
2958 if (cap->issued & ~newcaps) {
2959 int revoking = cap->issued & ~newcaps;
2961 dout("revocation: %s -> %s (revoking %s)\n",
2962 ceph_cap_string(cap->issued),
2963 ceph_cap_string(newcaps),
2964 ceph_cap_string(revoking));
2965 if (revoking & used & CEPH_CAP_FILE_BUFFER)
2966 writeback = true; /* initiate writeback; will delay ack */
2967 else if (revoking == CEPH_CAP_FILE_CACHE &&
2968 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2970 ; /* do nothing yet, invalidation will be queued */
2971 else if (cap == ci->i_auth_cap)
2972 check_caps = 1; /* check auth cap only */
2974 check_caps = 2; /* check all caps */
2975 cap->issued = newcaps;
2976 cap->implemented |= newcaps;
2977 } else if (cap->issued == newcaps) {
2978 dout("caps unchanged: %s -> %s\n",
2979 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2981 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2982 ceph_cap_string(newcaps));
2983 /* non-auth MDS is revoking the newly grant caps ? */
2984 if (cap == ci->i_auth_cap &&
2985 __ceph_caps_revoking_other(ci, cap, newcaps))
2988 cap->issued = newcaps;
2989 cap->implemented |= newcaps; /* add bits only, to
2990 * avoid stepping on a
2991 * pending revocation */
2994 BUG_ON(cap->issued & ~cap->implemented);
2996 if (inline_version > 0 && inline_version >= ci->i_inline_version) {
2997 ci->i_inline_version = inline_version;
2998 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2999 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3003 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3004 if (newcaps & ~issued)
3006 kick_flushing_inode_caps(mdsc, session, inode);
3007 up_read(&mdsc->snap_rwsem);
3009 spin_unlock(&ci->i_ceph_lock);
3013 ceph_fill_inline_data(inode, NULL, inline_data, inline_len);
3016 ceph_queue_vmtruncate(inode);
3020 * queue inode for writeback: we can't actually call
3021 * filemap_write_and_wait, etc. from message handler
3024 ceph_queue_writeback(inode);
3025 if (queue_invalidate)
3026 ceph_queue_invalidate(inode);
3028 invalidate_aliases(inode);
3030 wake_up_all(&ci->i_cap_wq);
3032 if (check_caps == 1)
3033 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
3035 else if (check_caps == 2)
3036 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
3038 mutex_unlock(&session->s_mutex);
3042 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3043 * MDS has been safely committed.
3045 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3046 struct ceph_mds_caps *m,
3047 struct ceph_mds_session *session,
3048 struct ceph_cap *cap)
3049 __releases(ci->i_ceph_lock)
3051 struct ceph_inode_info *ci = ceph_inode(inode);
3052 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3053 struct ceph_cap_flush *cf, *tmp_cf;
3054 LIST_HEAD(to_remove);
3055 unsigned seq = le32_to_cpu(m->seq);
3056 int dirty = le32_to_cpu(m->dirty);
3062 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3063 if (cf->tid == flush_tid)
3065 if (cf->caps == 0) /* capsnap */
3067 if (cf->tid <= flush_tid) {
3068 if (__finish_cap_flush(NULL, ci, cf))
3070 list_add_tail(&cf->i_list, &to_remove);
3072 cleaned &= ~cf->caps;
3078 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3079 " flushing %s -> %s\n",
3080 inode, session->s_mds, seq, ceph_cap_string(dirty),
3081 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3082 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3084 if (list_empty(&to_remove) && !cleaned)
3087 ci->i_flushing_caps &= ~cleaned;
3089 spin_lock(&mdsc->cap_dirty_lock);
3091 list_for_each_entry(cf, &to_remove, i_list) {
3092 if (__finish_cap_flush(mdsc, NULL, cf))
3096 if (ci->i_flushing_caps == 0) {
3097 if (list_empty(&ci->i_cap_flush_list)) {
3098 list_del_init(&ci->i_flushing_item);
3099 if (!list_empty(&session->s_cap_flushing)) {
3100 dout(" mds%d still flushing cap on %p\n",
3102 &list_first_entry(&session->s_cap_flushing,
3103 struct ceph_inode_info,
3104 i_flushing_item)->vfs_inode);
3107 mdsc->num_cap_flushing--;
3108 dout(" inode %p now !flushing\n", inode);
3110 if (ci->i_dirty_caps == 0) {
3111 dout(" inode %p now clean\n", inode);
3112 BUG_ON(!list_empty(&ci->i_dirty_item));
3114 if (ci->i_wr_ref == 0 &&
3115 ci->i_wrbuffer_ref_head == 0) {
3116 BUG_ON(!ci->i_head_snapc);
3117 ceph_put_snap_context(ci->i_head_snapc);
3118 ci->i_head_snapc = NULL;
3121 BUG_ON(list_empty(&ci->i_dirty_item));
3124 spin_unlock(&mdsc->cap_dirty_lock);
3127 spin_unlock(&ci->i_ceph_lock);
3129 while (!list_empty(&to_remove)) {
3130 cf = list_first_entry(&to_remove,
3131 struct ceph_cap_flush, i_list);
3132 list_del(&cf->i_list);
3133 ceph_free_cap_flush(cf);
3137 wake_up_all(&ci->i_cap_wq);
3139 wake_up_all(&mdsc->cap_flushing_wq);
3145 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3146 * throw away our cap_snap.
3148 * Caller hold s_mutex.
3150 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3151 struct ceph_mds_caps *m,
3152 struct ceph_mds_session *session)
3154 struct ceph_inode_info *ci = ceph_inode(inode);
3155 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3156 u64 follows = le64_to_cpu(m->snap_follows);
3157 struct ceph_cap_snap *capsnap;
3158 bool flushed = false;
3159 bool wake_ci = false;
3160 bool wake_mdsc = false;
3162 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3163 inode, ci, session->s_mds, follows);
3165 spin_lock(&ci->i_ceph_lock);
3166 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3167 if (capsnap->follows == follows) {
3168 if (capsnap->cap_flush.tid != flush_tid) {
3169 dout(" cap_snap %p follows %lld tid %lld !="
3170 " %lld\n", capsnap, follows,
3171 flush_tid, capsnap->cap_flush.tid);
3177 dout(" skipping cap_snap %p follows %lld\n",
3178 capsnap, capsnap->follows);
3182 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3183 dout(" removing %p cap_snap %p follows %lld\n",
3184 inode, capsnap, follows);
3185 list_del(&capsnap->ci_item);
3186 if (__finish_cap_flush(NULL, ci, &capsnap->cap_flush))
3189 spin_lock(&mdsc->cap_dirty_lock);
3191 if (list_empty(&ci->i_cap_flush_list))
3192 list_del_init(&ci->i_flushing_item);
3194 if (__finish_cap_flush(mdsc, NULL, &capsnap->cap_flush))
3197 spin_unlock(&mdsc->cap_dirty_lock);
3199 spin_unlock(&ci->i_ceph_lock);
3201 ceph_put_snap_context(capsnap->context);
3202 ceph_put_cap_snap(capsnap);
3204 wake_up_all(&ci->i_cap_wq);
3206 wake_up_all(&mdsc->cap_flushing_wq);
3212 * Handle TRUNC from MDS, indicating file truncation.
3214 * caller hold s_mutex.
3216 static void handle_cap_trunc(struct inode *inode,
3217 struct ceph_mds_caps *trunc,
3218 struct ceph_mds_session *session)
3219 __releases(ci->i_ceph_lock)
3221 struct ceph_inode_info *ci = ceph_inode(inode);
3222 int mds = session->s_mds;
3223 int seq = le32_to_cpu(trunc->seq);
3224 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3225 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3226 u64 size = le64_to_cpu(trunc->size);
3227 int implemented = 0;
3228 int dirty = __ceph_caps_dirty(ci);
3229 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3230 int queue_trunc = 0;
3232 issued |= implemented | dirty;
3234 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3235 inode, mds, seq, truncate_size, truncate_seq);
3236 queue_trunc = ceph_fill_file_size(inode, issued,
3237 truncate_seq, truncate_size, size);
3238 spin_unlock(&ci->i_ceph_lock);
3241 ceph_queue_vmtruncate(inode);
3245 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3246 * different one. If we are the most recent migration we've seen (as
3247 * indicated by mseq), make note of the migrating cap bits for the
3248 * duration (until we see the corresponding IMPORT).
3250 * caller holds s_mutex
3252 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3253 struct ceph_mds_cap_peer *ph,
3254 struct ceph_mds_session *session)
3256 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3257 struct ceph_mds_session *tsession = NULL;
3258 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3259 struct ceph_inode_info *ci = ceph_inode(inode);
3261 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3262 unsigned t_seq, t_mseq;
3264 int mds = session->s_mds;
3267 t_cap_id = le64_to_cpu(ph->cap_id);
3268 t_seq = le32_to_cpu(ph->seq);
3269 t_mseq = le32_to_cpu(ph->mseq);
3270 target = le32_to_cpu(ph->mds);
3272 t_cap_id = t_seq = t_mseq = 0;
3276 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3277 inode, ci, mds, mseq, target);
3279 spin_lock(&ci->i_ceph_lock);
3280 cap = __get_cap_for_mds(ci, mds);
3281 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3285 __ceph_remove_cap(cap, false);
3286 if (!ci->i_auth_cap)
3287 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
3292 * now we know we haven't received the cap import message yet
3293 * because the exported cap still exist.
3296 issued = cap->issued;
3297 WARN_ON(issued != cap->implemented);
3299 tcap = __get_cap_for_mds(ci, target);
3301 /* already have caps from the target */
3302 if (tcap->cap_id != t_cap_id ||
3303 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3304 dout(" updating import cap %p mds%d\n", tcap, target);
3305 tcap->cap_id = t_cap_id;
3306 tcap->seq = t_seq - 1;
3307 tcap->issue_seq = t_seq - 1;
3308 tcap->mseq = t_mseq;
3309 tcap->issued |= issued;
3310 tcap->implemented |= issued;
3311 if (cap == ci->i_auth_cap)
3312 ci->i_auth_cap = tcap;
3313 if (!list_empty(&ci->i_cap_flush_list) &&
3314 ci->i_auth_cap == tcap) {
3315 spin_lock(&mdsc->cap_dirty_lock);
3316 list_move_tail(&ci->i_flushing_item,
3317 &tcap->session->s_cap_flushing);
3318 spin_unlock(&mdsc->cap_dirty_lock);
3321 __ceph_remove_cap(cap, false);
3323 } else if (tsession) {
3324 /* add placeholder for the export tagert */
3325 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3326 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3327 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3329 __ceph_remove_cap(cap, false);
3333 spin_unlock(&ci->i_ceph_lock);
3334 mutex_unlock(&session->s_mutex);
3336 /* open target session */
3337 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3338 if (!IS_ERR(tsession)) {
3340 mutex_lock(&session->s_mutex);
3341 mutex_lock_nested(&tsession->s_mutex,
3342 SINGLE_DEPTH_NESTING);
3344 mutex_lock(&tsession->s_mutex);
3345 mutex_lock_nested(&session->s_mutex,
3346 SINGLE_DEPTH_NESTING);
3348 new_cap = ceph_get_cap(mdsc, NULL);
3357 spin_unlock(&ci->i_ceph_lock);
3358 mutex_unlock(&session->s_mutex);
3360 mutex_unlock(&tsession->s_mutex);
3361 ceph_put_mds_session(tsession);
3364 ceph_put_cap(mdsc, new_cap);
3368 * Handle cap IMPORT.
3370 * caller holds s_mutex. acquires i_ceph_lock
3372 static void handle_cap_import(struct ceph_mds_client *mdsc,
3373 struct inode *inode, struct ceph_mds_caps *im,
3374 struct ceph_mds_cap_peer *ph,
3375 struct ceph_mds_session *session,
3376 struct ceph_cap **target_cap, int *old_issued)
3377 __acquires(ci->i_ceph_lock)
3379 struct ceph_inode_info *ci = ceph_inode(inode);
3380 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3381 int mds = session->s_mds;
3383 unsigned caps = le32_to_cpu(im->caps);
3384 unsigned wanted = le32_to_cpu(im->wanted);
3385 unsigned seq = le32_to_cpu(im->seq);
3386 unsigned mseq = le32_to_cpu(im->migrate_seq);
3387 u64 realmino = le64_to_cpu(im->realm);
3388 u64 cap_id = le64_to_cpu(im->cap_id);
3393 p_cap_id = le64_to_cpu(ph->cap_id);
3394 peer = le32_to_cpu(ph->mds);
3400 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3401 inode, ci, mds, mseq, peer);
3404 spin_lock(&ci->i_ceph_lock);
3405 cap = __get_cap_for_mds(ci, mds);
3408 spin_unlock(&ci->i_ceph_lock);
3409 new_cap = ceph_get_cap(mdsc, NULL);
3415 ceph_put_cap(mdsc, new_cap);
3420 __ceph_caps_issued(ci, &issued);
3421 issued |= __ceph_caps_dirty(ci);
3423 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3424 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3426 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3427 if (ocap && ocap->cap_id == p_cap_id) {
3428 dout(" remove export cap %p mds%d flags %d\n",
3429 ocap, peer, ph->flags);
3430 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3431 (ocap->seq != le32_to_cpu(ph->seq) ||
3432 ocap->mseq != le32_to_cpu(ph->mseq))) {
3433 pr_err("handle_cap_import: mismatched seq/mseq: "
3434 "ino (%llx.%llx) mds%d seq %d mseq %d "
3435 "importer mds%d has peer seq %d mseq %d\n",
3436 ceph_vinop(inode), peer, ocap->seq,
3437 ocap->mseq, mds, le32_to_cpu(ph->seq),
3438 le32_to_cpu(ph->mseq));
3440 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3443 /* make sure we re-request max_size, if necessary */
3444 ci->i_wanted_max_size = 0;
3445 ci->i_requested_max_size = 0;
3447 *old_issued = issued;
3452 * Handle a caps message from the MDS.
3454 * Identify the appropriate session, inode, and call the right handler
3455 * based on the cap op.
3457 void ceph_handle_caps(struct ceph_mds_session *session,
3458 struct ceph_msg *msg)
3460 struct ceph_mds_client *mdsc = session->s_mdsc;
3461 struct super_block *sb = mdsc->fsc->sb;
3462 struct inode *inode;
3463 struct ceph_inode_info *ci;
3464 struct ceph_cap *cap;
3465 struct ceph_mds_caps *h;
3466 struct ceph_mds_cap_peer *peer = NULL;
3467 struct ceph_snap_realm *realm = NULL;
3468 struct ceph_string *pool_ns = NULL;
3469 int mds = session->s_mds;
3472 struct ceph_vino vino;
3474 u64 inline_version = 0;
3475 void *inline_data = NULL;
3478 size_t snaptrace_len;
3481 dout("handle_caps from mds%d\n", mds);
3484 end = msg->front.iov_base + msg->front.iov_len;
3485 tid = le64_to_cpu(msg->hdr.tid);
3486 if (msg->front.iov_len < sizeof(*h))
3488 h = msg->front.iov_base;
3489 op = le32_to_cpu(h->op);
3490 vino.ino = le64_to_cpu(h->ino);
3491 vino.snap = CEPH_NOSNAP;
3492 seq = le32_to_cpu(h->seq);
3493 mseq = le32_to_cpu(h->migrate_seq);
3496 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3497 p = snaptrace + snaptrace_len;
3499 if (le16_to_cpu(msg->hdr.version) >= 2) {
3501 ceph_decode_32_safe(&p, end, flock_len, bad);
3502 if (p + flock_len > end)
3507 if (le16_to_cpu(msg->hdr.version) >= 3) {
3508 if (op == CEPH_CAP_OP_IMPORT) {
3509 if (p + sizeof(*peer) > end)
3513 } else if (op == CEPH_CAP_OP_EXPORT) {
3514 /* recorded in unused fields */
3515 peer = (void *)&h->size;
3519 if (le16_to_cpu(msg->hdr.version) >= 4) {
3520 ceph_decode_64_safe(&p, end, inline_version, bad);
3521 ceph_decode_32_safe(&p, end, inline_len, bad);
3522 if (p + inline_len > end)
3528 if (le16_to_cpu(msg->hdr.version) >= 8) {
3530 u32 caller_uid, caller_gid;
3531 u32 osd_epoch_barrier;
3534 ceph_decode_32_safe(&p, end, osd_epoch_barrier, bad);
3536 ceph_decode_64_safe(&p, end, flush_tid, bad);
3538 ceph_decode_32_safe(&p, end, caller_uid, bad);
3539 ceph_decode_32_safe(&p, end, caller_gid, bad);
3541 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
3542 if (pool_ns_len > 0) {
3543 ceph_decode_need(&p, end, pool_ns_len, bad);
3544 pool_ns = ceph_find_or_create_string(p, pool_ns_len);
3550 inode = ceph_find_inode(sb, vino);
3551 ci = ceph_inode(inode);
3552 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3555 mutex_lock(&session->s_mutex);
3557 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3561 dout(" i don't have ino %llx\n", vino.ino);
3563 if (op == CEPH_CAP_OP_IMPORT) {
3564 cap = ceph_get_cap(mdsc, NULL);
3565 cap->cap_ino = vino.ino;
3566 cap->queue_release = 1;
3567 cap->cap_id = le64_to_cpu(h->cap_id);
3570 spin_lock(&session->s_cap_lock);
3571 list_add_tail(&cap->session_caps,
3572 &session->s_cap_releases);
3573 session->s_num_cap_releases++;
3574 spin_unlock(&session->s_cap_lock);
3576 goto flush_cap_releases;
3579 /* these will work even if we don't have a cap yet */
3581 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3582 handle_cap_flushsnap_ack(inode, tid, h, session);
3585 case CEPH_CAP_OP_EXPORT:
3586 handle_cap_export(inode, h, peer, session);
3589 case CEPH_CAP_OP_IMPORT:
3591 if (snaptrace_len) {
3592 down_write(&mdsc->snap_rwsem);
3593 ceph_update_snap_trace(mdsc, snaptrace,
3594 snaptrace + snaptrace_len,
3596 downgrade_write(&mdsc->snap_rwsem);
3598 down_read(&mdsc->snap_rwsem);
3600 handle_cap_import(mdsc, inode, h, peer, session,
3602 handle_cap_grant(mdsc, inode, h, &pool_ns,
3603 inline_version, inline_data, inline_len,
3604 msg->middle, session, cap, issued);
3606 ceph_put_snap_realm(mdsc, realm);
3610 /* the rest require a cap */
3611 spin_lock(&ci->i_ceph_lock);
3612 cap = __get_cap_for_mds(ceph_inode(inode), mds);
3614 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3615 inode, ceph_ino(inode), ceph_snap(inode), mds);
3616 spin_unlock(&ci->i_ceph_lock);
3617 goto flush_cap_releases;
3620 /* note that each of these drops i_ceph_lock for us */
3622 case CEPH_CAP_OP_REVOKE:
3623 case CEPH_CAP_OP_GRANT:
3624 __ceph_caps_issued(ci, &issued);
3625 issued |= __ceph_caps_dirty(ci);
3626 handle_cap_grant(mdsc, inode, h, &pool_ns,
3627 inline_version, inline_data, inline_len,
3628 msg->middle, session, cap, issued);
3631 case CEPH_CAP_OP_FLUSH_ACK:
3632 handle_cap_flush_ack(inode, tid, h, session, cap);
3635 case CEPH_CAP_OP_TRUNC:
3636 handle_cap_trunc(inode, h, session);
3640 spin_unlock(&ci->i_ceph_lock);
3641 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3642 ceph_cap_op_name(op));
3649 * send any cap release message to try to move things
3650 * along for the mds (who clearly thinks we still have this
3653 ceph_send_cap_releases(mdsc, session);
3656 mutex_unlock(&session->s_mutex);
3659 ceph_put_string(pool_ns);
3663 pr_err("ceph_handle_caps: corrupt message\n");
3669 * Delayed work handler to process end of delayed cap release LRU list.
3671 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3673 struct ceph_inode_info *ci;
3674 int flags = CHECK_CAPS_NODELAY;
3676 dout("check_delayed_caps\n");
3678 spin_lock(&mdsc->cap_delay_lock);
3679 if (list_empty(&mdsc->cap_delay_list))
3681 ci = list_first_entry(&mdsc->cap_delay_list,
3682 struct ceph_inode_info,
3684 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3685 time_before(jiffies, ci->i_hold_caps_max))
3687 list_del_init(&ci->i_cap_delay_list);
3688 spin_unlock(&mdsc->cap_delay_lock);
3689 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
3690 ceph_check_caps(ci, flags, NULL);
3692 spin_unlock(&mdsc->cap_delay_lock);
3696 * Flush all dirty caps to the mds
3698 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
3700 struct ceph_inode_info *ci;
3701 struct inode *inode;
3703 dout("flush_dirty_caps\n");
3704 spin_lock(&mdsc->cap_dirty_lock);
3705 while (!list_empty(&mdsc->cap_dirty)) {
3706 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
3708 inode = &ci->vfs_inode;
3710 dout("flush_dirty_caps %p\n", inode);
3711 spin_unlock(&mdsc->cap_dirty_lock);
3712 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
3714 spin_lock(&mdsc->cap_dirty_lock);
3716 spin_unlock(&mdsc->cap_dirty_lock);
3717 dout("flush_dirty_caps done\n");
3720 void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode)
3723 int bits = (fmode << 1) | 1;
3724 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
3725 if (bits & (1 << i))
3726 ci->i_nr_by_mode[i]++;
3731 * Drop open file reference. If we were the last open file,
3732 * we may need to release capabilities to the MDS (or schedule
3733 * their delayed release).
3735 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
3738 int bits = (fmode << 1) | 1;
3739 spin_lock(&ci->i_ceph_lock);
3740 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
3741 if (bits & (1 << i)) {
3742 BUG_ON(ci->i_nr_by_mode[i] == 0);
3743 if (--ci->i_nr_by_mode[i] == 0)
3747 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
3748 &ci->vfs_inode, fmode,
3749 ci->i_nr_by_mode[0], ci->i_nr_by_mode[1],
3750 ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]);
3751 spin_unlock(&ci->i_ceph_lock);
3753 if (last && ci->i_vino.snap == CEPH_NOSNAP)
3754 ceph_check_caps(ci, 0, NULL);
3758 * Helpers for embedding cap and dentry lease releases into mds
3761 * @force is used by dentry_release (below) to force inclusion of a
3762 * record for the directory inode, even when there aren't any caps to
3765 int ceph_encode_inode_release(void **p, struct inode *inode,
3766 int mds, int drop, int unless, int force)
3768 struct ceph_inode_info *ci = ceph_inode(inode);
3769 struct ceph_cap *cap;
3770 struct ceph_mds_request_release *rel = *p;
3774 spin_lock(&ci->i_ceph_lock);
3775 used = __ceph_caps_used(ci);
3776 dirty = __ceph_caps_dirty(ci);
3778 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3779 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3780 ceph_cap_string(unless));
3782 /* only drop unused, clean caps */
3783 drop &= ~(used | dirty);
3785 cap = __get_cap_for_mds(ci, mds);
3786 if (cap && __cap_is_valid(cap)) {
3788 ((cap->issued & drop) &&
3789 (cap->issued & unless) == 0)) {
3790 if ((cap->issued & drop) &&
3791 (cap->issued & unless) == 0) {
3792 int wanted = __ceph_caps_wanted(ci);
3793 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
3794 wanted |= cap->mds_wanted;
3795 dout("encode_inode_release %p cap %p "
3796 "%s -> %s, wanted %s -> %s\n", inode, cap,
3797 ceph_cap_string(cap->issued),
3798 ceph_cap_string(cap->issued & ~drop),
3799 ceph_cap_string(cap->mds_wanted),
3800 ceph_cap_string(wanted));
3802 cap->issued &= ~drop;
3803 cap->implemented &= ~drop;
3804 cap->mds_wanted = wanted;
3806 dout("encode_inode_release %p cap %p %s"
3807 " (force)\n", inode, cap,
3808 ceph_cap_string(cap->issued));
3811 rel->ino = cpu_to_le64(ceph_ino(inode));
3812 rel->cap_id = cpu_to_le64(cap->cap_id);
3813 rel->seq = cpu_to_le32(cap->seq);
3814 rel->issue_seq = cpu_to_le32(cap->issue_seq);
3815 rel->mseq = cpu_to_le32(cap->mseq);
3816 rel->caps = cpu_to_le32(cap->implemented);
3817 rel->wanted = cpu_to_le32(cap->mds_wanted);
3823 dout("encode_inode_release %p cap %p %s\n",
3824 inode, cap, ceph_cap_string(cap->issued));
3827 spin_unlock(&ci->i_ceph_lock);
3831 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3832 int mds, int drop, int unless)
3834 struct inode *dir = d_inode(dentry->d_parent);
3835 struct ceph_mds_request_release *rel = *p;
3836 struct ceph_dentry_info *di = ceph_dentry(dentry);
3841 * force an record for the directory caps if we have a dentry lease.
3842 * this is racy (can't take i_ceph_lock and d_lock together), but it
3843 * doesn't have to be perfect; the mds will revoke anything we don't
3846 spin_lock(&dentry->d_lock);
3847 if (di->lease_session && di->lease_session->s_mds == mds)
3849 spin_unlock(&dentry->d_lock);
3851 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3853 spin_lock(&dentry->d_lock);
3854 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3855 dout("encode_dentry_release %p mds%d seq %d\n",
3856 dentry, mds, (int)di->lease_seq);
3857 rel->dname_len = cpu_to_le32(dentry->d_name.len);
3858 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3859 *p += dentry->d_name.len;
3860 rel->dname_seq = cpu_to_le32(di->lease_seq);
3861 __ceph_mdsc_drop_dentry_lease(dentry);
3863 spin_unlock(&dentry->d_lock);