1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
5 #include <linux/kernel.h>
6 #include <linux/sched/signal.h>
7 #include <linux/slab.h>
8 #include <linux/vmalloc.h>
9 #include <linux/wait.h>
10 #include <linux/writeback.h>
11 #include <linux/iversion.h>
14 #include "mds_client.h"
16 #include <linux/ceph/decode.h>
17 #include <linux/ceph/messenger.h>
20 * Capability management
22 * The Ceph metadata servers control client access to inode metadata
23 * and file data by issuing capabilities, granting clients permission
24 * to read and/or write both inode field and file data to OSDs
25 * (storage nodes). Each capability consists of a set of bits
26 * indicating which operations are allowed.
28 * If the client holds a *_SHARED cap, the client has a coherent value
29 * that can be safely read from the cached inode.
31 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
32 * client is allowed to change inode attributes (e.g., file size,
33 * mtime), note its dirty state in the ceph_cap, and asynchronously
34 * flush that metadata change to the MDS.
36 * In the event of a conflicting operation (perhaps by another
37 * client), the MDS will revoke the conflicting client capabilities.
39 * In order for a client to cache an inode, it must hold a capability
40 * with at least one MDS server. When inodes are released, release
41 * notifications are batched and periodically sent en masse to the MDS
42 * cluster to release server state.
45 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc);
46 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
47 struct ceph_mds_session *session,
48 struct ceph_inode_info *ci,
49 u64 oldest_flush_tid);
52 * Generate readable cap strings for debugging output.
54 #define MAX_CAP_STR 20
55 static char cap_str[MAX_CAP_STR][40];
56 static DEFINE_SPINLOCK(cap_str_lock);
57 static int last_cap_str;
59 static char *gcap_string(char *s, int c)
61 if (c & CEPH_CAP_GSHARED)
63 if (c & CEPH_CAP_GEXCL)
65 if (c & CEPH_CAP_GCACHE)
71 if (c & CEPH_CAP_GBUFFER)
73 if (c & CEPH_CAP_GWREXTEND)
75 if (c & CEPH_CAP_GLAZYIO)
80 const char *ceph_cap_string(int caps)
86 spin_lock(&cap_str_lock);
88 if (last_cap_str == MAX_CAP_STR)
90 spin_unlock(&cap_str_lock);
94 if (caps & CEPH_CAP_PIN)
97 c = (caps >> CEPH_CAP_SAUTH) & 3;
100 s = gcap_string(s, c);
103 c = (caps >> CEPH_CAP_SLINK) & 3;
106 s = gcap_string(s, c);
109 c = (caps >> CEPH_CAP_SXATTR) & 3;
112 s = gcap_string(s, c);
115 c = caps >> CEPH_CAP_SFILE;
118 s = gcap_string(s, c);
127 void ceph_caps_init(struct ceph_mds_client *mdsc)
129 INIT_LIST_HEAD(&mdsc->caps_list);
130 spin_lock_init(&mdsc->caps_list_lock);
133 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
135 struct ceph_cap *cap;
137 spin_lock(&mdsc->caps_list_lock);
138 while (!list_empty(&mdsc->caps_list)) {
139 cap = list_first_entry(&mdsc->caps_list,
140 struct ceph_cap, caps_item);
141 list_del(&cap->caps_item);
142 kmem_cache_free(ceph_cap_cachep, cap);
144 mdsc->caps_total_count = 0;
145 mdsc->caps_avail_count = 0;
146 mdsc->caps_use_count = 0;
147 mdsc->caps_reserve_count = 0;
148 mdsc->caps_min_count = 0;
149 spin_unlock(&mdsc->caps_list_lock);
152 void ceph_adjust_caps_max_min(struct ceph_mds_client *mdsc,
153 struct ceph_mount_options *fsopt)
155 spin_lock(&mdsc->caps_list_lock);
156 mdsc->caps_min_count = fsopt->max_readdir;
157 if (mdsc->caps_min_count < 1024)
158 mdsc->caps_min_count = 1024;
159 mdsc->caps_use_max = fsopt->caps_max;
160 if (mdsc->caps_use_max > 0 &&
161 mdsc->caps_use_max < mdsc->caps_min_count)
162 mdsc->caps_use_max = mdsc->caps_min_count;
163 spin_unlock(&mdsc->caps_list_lock);
166 static void __ceph_unreserve_caps(struct ceph_mds_client *mdsc, int nr_caps)
168 struct ceph_cap *cap;
172 BUG_ON(mdsc->caps_reserve_count < nr_caps);
173 mdsc->caps_reserve_count -= nr_caps;
174 if (mdsc->caps_avail_count >=
175 mdsc->caps_reserve_count + mdsc->caps_min_count) {
176 mdsc->caps_total_count -= nr_caps;
177 for (i = 0; i < nr_caps; i++) {
178 cap = list_first_entry(&mdsc->caps_list,
179 struct ceph_cap, caps_item);
180 list_del(&cap->caps_item);
181 kmem_cache_free(ceph_cap_cachep, cap);
184 mdsc->caps_avail_count += nr_caps;
187 dout("%s: caps %d = %d used + %d resv + %d avail\n",
189 mdsc->caps_total_count, mdsc->caps_use_count,
190 mdsc->caps_reserve_count, mdsc->caps_avail_count);
191 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
192 mdsc->caps_reserve_count +
193 mdsc->caps_avail_count);
198 * Called under mdsc->mutex.
200 int ceph_reserve_caps(struct ceph_mds_client *mdsc,
201 struct ceph_cap_reservation *ctx, int need)
204 struct ceph_cap *cap;
209 bool trimmed = false;
210 struct ceph_mds_session *s;
213 dout("reserve caps ctx=%p need=%d\n", ctx, need);
215 /* first reserve any caps that are already allocated */
216 spin_lock(&mdsc->caps_list_lock);
217 if (mdsc->caps_avail_count >= need)
220 have = mdsc->caps_avail_count;
221 mdsc->caps_avail_count -= have;
222 mdsc->caps_reserve_count += have;
223 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
224 mdsc->caps_reserve_count +
225 mdsc->caps_avail_count);
226 spin_unlock(&mdsc->caps_list_lock);
228 for (i = have; i < need; ) {
229 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
231 list_add(&cap->caps_item, &newcaps);
238 for (j = 0; j < mdsc->max_sessions; j++) {
239 s = __ceph_lookup_mds_session(mdsc, j);
242 mutex_unlock(&mdsc->mutex);
244 mutex_lock(&s->s_mutex);
245 max_caps = s->s_nr_caps - (need - i);
246 ceph_trim_caps(mdsc, s, max_caps);
247 mutex_unlock(&s->s_mutex);
249 ceph_put_mds_session(s);
250 mutex_lock(&mdsc->mutex);
254 spin_lock(&mdsc->caps_list_lock);
255 if (mdsc->caps_avail_count) {
257 if (mdsc->caps_avail_count >= need - i)
258 more_have = need - i;
260 more_have = mdsc->caps_avail_count;
264 mdsc->caps_avail_count -= more_have;
265 mdsc->caps_reserve_count += more_have;
268 spin_unlock(&mdsc->caps_list_lock);
273 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
274 ctx, need, have + alloc);
280 BUG_ON(have + alloc != need);
285 spin_lock(&mdsc->caps_list_lock);
286 mdsc->caps_total_count += alloc;
287 mdsc->caps_reserve_count += alloc;
288 list_splice(&newcaps, &mdsc->caps_list);
290 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
291 mdsc->caps_reserve_count +
292 mdsc->caps_avail_count);
295 __ceph_unreserve_caps(mdsc, have + alloc);
297 spin_unlock(&mdsc->caps_list_lock);
299 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
300 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
301 mdsc->caps_reserve_count, mdsc->caps_avail_count);
305 void ceph_unreserve_caps(struct ceph_mds_client *mdsc,
306 struct ceph_cap_reservation *ctx)
308 bool reclaim = false;
312 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
313 spin_lock(&mdsc->caps_list_lock);
314 __ceph_unreserve_caps(mdsc, ctx->count);
317 if (mdsc->caps_use_max > 0 &&
318 mdsc->caps_use_count > mdsc->caps_use_max)
320 spin_unlock(&mdsc->caps_list_lock);
323 ceph_reclaim_caps_nr(mdsc, ctx->used);
326 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
327 struct ceph_cap_reservation *ctx)
329 struct ceph_cap *cap = NULL;
331 /* temporary, until we do something about cap import/export */
333 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
335 spin_lock(&mdsc->caps_list_lock);
336 mdsc->caps_use_count++;
337 mdsc->caps_total_count++;
338 spin_unlock(&mdsc->caps_list_lock);
340 spin_lock(&mdsc->caps_list_lock);
341 if (mdsc->caps_avail_count) {
342 BUG_ON(list_empty(&mdsc->caps_list));
344 mdsc->caps_avail_count--;
345 mdsc->caps_use_count++;
346 cap = list_first_entry(&mdsc->caps_list,
347 struct ceph_cap, caps_item);
348 list_del(&cap->caps_item);
350 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
351 mdsc->caps_reserve_count + mdsc->caps_avail_count);
353 spin_unlock(&mdsc->caps_list_lock);
359 spin_lock(&mdsc->caps_list_lock);
360 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
361 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
362 mdsc->caps_reserve_count, mdsc->caps_avail_count);
364 BUG_ON(ctx->count > mdsc->caps_reserve_count);
365 BUG_ON(list_empty(&mdsc->caps_list));
369 mdsc->caps_reserve_count--;
370 mdsc->caps_use_count++;
372 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
373 list_del(&cap->caps_item);
375 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
376 mdsc->caps_reserve_count + mdsc->caps_avail_count);
377 spin_unlock(&mdsc->caps_list_lock);
381 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
383 spin_lock(&mdsc->caps_list_lock);
384 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
385 cap, mdsc->caps_total_count, mdsc->caps_use_count,
386 mdsc->caps_reserve_count, mdsc->caps_avail_count);
387 mdsc->caps_use_count--;
389 * Keep some preallocated caps around (ceph_min_count), to
390 * avoid lots of free/alloc churn.
392 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
393 mdsc->caps_min_count) {
394 mdsc->caps_total_count--;
395 kmem_cache_free(ceph_cap_cachep, cap);
397 mdsc->caps_avail_count++;
398 list_add(&cap->caps_item, &mdsc->caps_list);
401 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
402 mdsc->caps_reserve_count + mdsc->caps_avail_count);
403 spin_unlock(&mdsc->caps_list_lock);
406 void ceph_reservation_status(struct ceph_fs_client *fsc,
407 int *total, int *avail, int *used, int *reserved,
410 struct ceph_mds_client *mdsc = fsc->mdsc;
412 spin_lock(&mdsc->caps_list_lock);
415 *total = mdsc->caps_total_count;
417 *avail = mdsc->caps_avail_count;
419 *used = mdsc->caps_use_count;
421 *reserved = mdsc->caps_reserve_count;
423 *min = mdsc->caps_min_count;
425 spin_unlock(&mdsc->caps_list_lock);
429 * Find ceph_cap for given mds, if any.
431 * Called with i_ceph_lock held.
433 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
435 struct ceph_cap *cap;
436 struct rb_node *n = ci->i_caps.rb_node;
439 cap = rb_entry(n, struct ceph_cap, ci_node);
442 else if (mds > cap->mds)
450 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
452 struct ceph_cap *cap;
454 spin_lock(&ci->i_ceph_lock);
455 cap = __get_cap_for_mds(ci, mds);
456 spin_unlock(&ci->i_ceph_lock);
461 * Called under i_ceph_lock.
463 static void __insert_cap_node(struct ceph_inode_info *ci,
464 struct ceph_cap *new)
466 struct rb_node **p = &ci->i_caps.rb_node;
467 struct rb_node *parent = NULL;
468 struct ceph_cap *cap = NULL;
472 cap = rb_entry(parent, struct ceph_cap, ci_node);
473 if (new->mds < cap->mds)
475 else if (new->mds > cap->mds)
481 rb_link_node(&new->ci_node, parent, p);
482 rb_insert_color(&new->ci_node, &ci->i_caps);
486 * (re)set cap hold timeouts, which control the delayed release
487 * of unused caps back to the MDS. Should be called on cap use.
489 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
490 struct ceph_inode_info *ci)
492 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
493 ci->i_hold_caps_max = round_jiffies(jiffies +
494 opt->caps_wanted_delay_max * HZ);
495 dout("__cap_set_timeouts %p %lu\n", &ci->vfs_inode,
496 ci->i_hold_caps_max - jiffies);
500 * (Re)queue cap at the end of the delayed cap release list.
502 * If I_FLUSH is set, leave the inode at the front of the list.
504 * Caller holds i_ceph_lock
505 * -> we take mdsc->cap_delay_lock
507 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
508 struct ceph_inode_info *ci)
510 dout("__cap_delay_requeue %p flags 0x%lx at %lu\n", &ci->vfs_inode,
511 ci->i_ceph_flags, ci->i_hold_caps_max);
512 if (!mdsc->stopping) {
513 spin_lock(&mdsc->cap_delay_lock);
514 if (!list_empty(&ci->i_cap_delay_list)) {
515 if (ci->i_ceph_flags & CEPH_I_FLUSH)
517 list_del_init(&ci->i_cap_delay_list);
519 __cap_set_timeouts(mdsc, ci);
520 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
522 spin_unlock(&mdsc->cap_delay_lock);
527 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
528 * indicating we should send a cap message to flush dirty metadata
529 * asap, and move to the front of the delayed cap list.
531 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
532 struct ceph_inode_info *ci)
534 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
535 spin_lock(&mdsc->cap_delay_lock);
536 ci->i_ceph_flags |= CEPH_I_FLUSH;
537 if (!list_empty(&ci->i_cap_delay_list))
538 list_del_init(&ci->i_cap_delay_list);
539 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
540 spin_unlock(&mdsc->cap_delay_lock);
544 * Cancel delayed work on cap.
546 * Caller must hold i_ceph_lock.
548 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
549 struct ceph_inode_info *ci)
551 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
552 if (list_empty(&ci->i_cap_delay_list))
554 spin_lock(&mdsc->cap_delay_lock);
555 list_del_init(&ci->i_cap_delay_list);
556 spin_unlock(&mdsc->cap_delay_lock);
559 /* Common issue checks for add_cap, handle_cap_grant. */
560 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
563 unsigned had = __ceph_caps_issued(ci, NULL);
565 lockdep_assert_held(&ci->i_ceph_lock);
568 * Each time we receive FILE_CACHE anew, we increment
571 if (S_ISREG(ci->vfs_inode.i_mode) &&
572 (issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
573 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
578 * If FILE_SHARED is newly issued, mark dir not complete. We don't
579 * know what happened to this directory while we didn't have the cap.
580 * If FILE_SHARED is being revoked, also mark dir not complete. It
581 * stops on-going cached readdir.
583 if ((issued & CEPH_CAP_FILE_SHARED) != (had & CEPH_CAP_FILE_SHARED)) {
584 if (issued & CEPH_CAP_FILE_SHARED)
585 atomic_inc(&ci->i_shared_gen);
586 if (S_ISDIR(ci->vfs_inode.i_mode)) {
587 dout(" marking %p NOT complete\n", &ci->vfs_inode);
588 __ceph_dir_clear_complete(ci);
592 /* Wipe saved layout if we're losing DIR_CREATE caps */
593 if (S_ISDIR(ci->vfs_inode.i_mode) && (had & CEPH_CAP_DIR_CREATE) &&
594 !(issued & CEPH_CAP_DIR_CREATE)) {
595 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
596 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
601 * change_auth_cap_ses - move inode to appropriate lists when auth caps change
602 * @ci: inode to be moved
603 * @session: new auth caps session
605 static void change_auth_cap_ses(struct ceph_inode_info *ci,
606 struct ceph_mds_session *session)
608 lockdep_assert_held(&ci->i_ceph_lock);
610 if (list_empty(&ci->i_dirty_item) && list_empty(&ci->i_flushing_item))
613 spin_lock(&session->s_mdsc->cap_dirty_lock);
614 if (!list_empty(&ci->i_dirty_item))
615 list_move(&ci->i_dirty_item, &session->s_cap_dirty);
616 if (!list_empty(&ci->i_flushing_item))
617 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
618 spin_unlock(&session->s_mdsc->cap_dirty_lock);
622 * Add a capability under the given MDS session.
624 * Caller should hold session snap_rwsem (read) and ci->i_ceph_lock
626 * @fmode is the open file mode, if we are opening a file, otherwise
627 * it is < 0. (This is so we can atomically add the cap and add an
628 * open file reference to it.)
630 void ceph_add_cap(struct inode *inode,
631 struct ceph_mds_session *session, u64 cap_id,
632 unsigned issued, unsigned wanted,
633 unsigned seq, unsigned mseq, u64 realmino, int flags,
634 struct ceph_cap **new_cap)
636 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
637 struct ceph_inode_info *ci = ceph_inode(inode);
638 struct ceph_cap *cap;
639 int mds = session->s_mds;
643 lockdep_assert_held(&ci->i_ceph_lock);
645 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
646 session->s_mds, cap_id, ceph_cap_string(issued), seq);
648 spin_lock(&session->s_gen_ttl_lock);
649 gen = session->s_cap_gen;
650 spin_unlock(&session->s_gen_ttl_lock);
652 cap = __get_cap_for_mds(ci, mds);
658 cap->implemented = 0;
664 __insert_cap_node(ci, cap);
666 /* add to session cap list */
667 cap->session = session;
668 spin_lock(&session->s_cap_lock);
669 list_add_tail(&cap->session_caps, &session->s_caps);
670 session->s_nr_caps++;
671 atomic64_inc(&mdsc->metric.total_caps);
672 spin_unlock(&session->s_cap_lock);
674 spin_lock(&session->s_cap_lock);
675 list_move_tail(&cap->session_caps, &session->s_caps);
676 spin_unlock(&session->s_cap_lock);
678 if (cap->cap_gen < gen)
679 cap->issued = cap->implemented = CEPH_CAP_PIN;
682 * auth mds of the inode changed. we received the cap export
683 * message, but still haven't received the cap import message.
684 * handle_cap_export() updated the new auth MDS' cap.
686 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
687 * a message that was send before the cap import message. So
690 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
691 WARN_ON(cap != ci->i_auth_cap);
692 WARN_ON(cap->cap_id != cap_id);
695 issued |= cap->issued;
696 flags |= CEPH_CAP_FLAG_AUTH;
700 if (!ci->i_snap_realm ||
701 ((flags & CEPH_CAP_FLAG_AUTH) &&
702 realmino != (u64)-1 && ci->i_snap_realm->ino != realmino)) {
704 * add this inode to the appropriate snap realm
706 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
709 struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
711 spin_lock(&oldrealm->inodes_with_caps_lock);
712 list_del_init(&ci->i_snap_realm_item);
713 spin_unlock(&oldrealm->inodes_with_caps_lock);
716 spin_lock(&realm->inodes_with_caps_lock);
717 list_add(&ci->i_snap_realm_item,
718 &realm->inodes_with_caps);
719 ci->i_snap_realm = realm;
720 if (realm->ino == ci->i_vino.ino)
721 realm->inode = inode;
722 spin_unlock(&realm->inodes_with_caps_lock);
725 ceph_put_snap_realm(mdsc, oldrealm);
727 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
733 __check_cap_issue(ci, cap, issued);
736 * If we are issued caps we don't want, or the mds' wanted
737 * value appears to be off, queue a check so we'll release
738 * later and/or update the mds wanted value.
740 actual_wanted = __ceph_caps_wanted(ci);
741 if ((wanted & ~actual_wanted) ||
742 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
743 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
744 ceph_cap_string(issued), ceph_cap_string(wanted),
745 ceph_cap_string(actual_wanted));
746 __cap_delay_requeue(mdsc, ci);
749 if (flags & CEPH_CAP_FLAG_AUTH) {
750 if (!ci->i_auth_cap ||
751 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
752 if (ci->i_auth_cap &&
753 ci->i_auth_cap->session != cap->session)
754 change_auth_cap_ses(ci, cap->session);
755 ci->i_auth_cap = cap;
756 cap->mds_wanted = wanted;
759 WARN_ON(ci->i_auth_cap == cap);
762 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
763 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
764 ceph_cap_string(issued|cap->issued), seq, mds);
765 cap->cap_id = cap_id;
766 cap->issued = issued;
767 cap->implemented |= issued;
768 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
769 cap->mds_wanted = wanted;
771 cap->mds_wanted |= wanted;
773 cap->issue_seq = seq;
779 * Return true if cap has not timed out and belongs to the current
780 * generation of the MDS session (i.e. has not gone 'stale' due to
781 * us losing touch with the mds).
783 static int __cap_is_valid(struct ceph_cap *cap)
788 spin_lock(&cap->session->s_gen_ttl_lock);
789 gen = cap->session->s_cap_gen;
790 ttl = cap->session->s_cap_ttl;
791 spin_unlock(&cap->session->s_gen_ttl_lock);
793 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
794 dout("__cap_is_valid %p cap %p issued %s "
795 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
796 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
804 * Return set of valid cap bits issued to us. Note that caps time
805 * out, and may be invalidated in bulk if the client session times out
806 * and session->s_cap_gen is bumped.
808 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
810 int have = ci->i_snap_caps;
811 struct ceph_cap *cap;
816 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
817 cap = rb_entry(p, struct ceph_cap, ci_node);
818 if (!__cap_is_valid(cap))
820 dout("__ceph_caps_issued %p cap %p issued %s\n",
821 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
824 *implemented |= cap->implemented;
827 * exclude caps issued by non-auth MDS, but are been revoking
828 * by the auth MDS. The non-auth MDS should be revoking/exporting
829 * these caps, but the message is delayed.
831 if (ci->i_auth_cap) {
832 cap = ci->i_auth_cap;
833 have &= ~cap->implemented | cap->issued;
839 * Get cap bits issued by caps other than @ocap
841 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
843 int have = ci->i_snap_caps;
844 struct ceph_cap *cap;
847 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
848 cap = rb_entry(p, struct ceph_cap, ci_node);
851 if (!__cap_is_valid(cap))
859 * Move a cap to the end of the LRU (oldest caps at list head, newest
862 static void __touch_cap(struct ceph_cap *cap)
864 struct ceph_mds_session *s = cap->session;
866 spin_lock(&s->s_cap_lock);
867 if (!s->s_cap_iterator) {
868 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
870 list_move_tail(&cap->session_caps, &s->s_caps);
872 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
873 &cap->ci->vfs_inode, cap, s->s_mds);
875 spin_unlock(&s->s_cap_lock);
879 * Check if we hold the given mask. If so, move the cap(s) to the
880 * front of their respective LRUs. (This is the preferred way for
881 * callers to check for caps they want.)
883 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
885 struct ceph_cap *cap;
887 int have = ci->i_snap_caps;
889 if ((have & mask) == mask) {
890 dout("__ceph_caps_issued_mask ino 0x%llx snap issued %s"
891 " (mask %s)\n", ceph_ino(&ci->vfs_inode),
892 ceph_cap_string(have),
893 ceph_cap_string(mask));
897 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
898 cap = rb_entry(p, struct ceph_cap, ci_node);
899 if (!__cap_is_valid(cap))
901 if ((cap->issued & mask) == mask) {
902 dout("__ceph_caps_issued_mask ino 0x%llx cap %p issued %s"
903 " (mask %s)\n", ceph_ino(&ci->vfs_inode), cap,
904 ceph_cap_string(cap->issued),
905 ceph_cap_string(mask));
911 /* does a combination of caps satisfy mask? */
913 if ((have & mask) == mask) {
914 dout("__ceph_caps_issued_mask ino 0x%llx combo issued %s"
915 " (mask %s)\n", ceph_ino(&ci->vfs_inode),
916 ceph_cap_string(cap->issued),
917 ceph_cap_string(mask));
921 /* touch this + preceding caps */
923 for (q = rb_first(&ci->i_caps); q != p;
925 cap = rb_entry(q, struct ceph_cap,
927 if (!__cap_is_valid(cap))
929 if (cap->issued & mask)
940 int __ceph_caps_issued_mask_metric(struct ceph_inode_info *ci, int mask,
943 struct ceph_fs_client *fsc = ceph_sb_to_client(ci->vfs_inode.i_sb);
946 r = __ceph_caps_issued_mask(ci, mask, touch);
948 ceph_update_cap_hit(&fsc->mdsc->metric);
950 ceph_update_cap_mis(&fsc->mdsc->metric);
955 * Return true if mask caps are currently being revoked by an MDS.
957 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
958 struct ceph_cap *ocap, int mask)
960 struct ceph_cap *cap;
963 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
964 cap = rb_entry(p, struct ceph_cap, ci_node);
966 (cap->implemented & ~cap->issued & mask))
972 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
974 struct inode *inode = &ci->vfs_inode;
977 spin_lock(&ci->i_ceph_lock);
978 ret = __ceph_caps_revoking_other(ci, NULL, mask);
979 spin_unlock(&ci->i_ceph_lock);
980 dout("ceph_caps_revoking %p %s = %d\n", inode,
981 ceph_cap_string(mask), ret);
985 int __ceph_caps_used(struct ceph_inode_info *ci)
989 used |= CEPH_CAP_PIN;
991 used |= CEPH_CAP_FILE_RD;
992 if (ci->i_rdcache_ref ||
993 (S_ISREG(ci->vfs_inode.i_mode) &&
994 ci->vfs_inode.i_data.nrpages))
995 used |= CEPH_CAP_FILE_CACHE;
997 used |= CEPH_CAP_FILE_WR;
998 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
999 used |= CEPH_CAP_FILE_BUFFER;
1001 used |= CEPH_CAP_FILE_EXCL;
1005 #define FMODE_WAIT_BIAS 1000
1008 * wanted, by virtue of open file modes
1010 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
1012 const int PIN_SHIFT = ffs(CEPH_FILE_MODE_PIN);
1013 const int RD_SHIFT = ffs(CEPH_FILE_MODE_RD);
1014 const int WR_SHIFT = ffs(CEPH_FILE_MODE_WR);
1015 const int LAZY_SHIFT = ffs(CEPH_FILE_MODE_LAZY);
1016 struct ceph_mount_options *opt =
1017 ceph_inode_to_client(&ci->vfs_inode)->mount_options;
1018 unsigned long used_cutoff = jiffies - opt->caps_wanted_delay_max * HZ;
1019 unsigned long idle_cutoff = jiffies - opt->caps_wanted_delay_min * HZ;
1021 if (S_ISDIR(ci->vfs_inode.i_mode)) {
1024 /* use used_cutoff here, to keep dir's wanted caps longer */
1025 if (ci->i_nr_by_mode[RD_SHIFT] > 0 ||
1026 time_after(ci->i_last_rd, used_cutoff))
1027 want |= CEPH_CAP_ANY_SHARED;
1029 if (ci->i_nr_by_mode[WR_SHIFT] > 0 ||
1030 time_after(ci->i_last_wr, used_cutoff)) {
1031 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1032 if (opt->flags & CEPH_MOUNT_OPT_ASYNC_DIROPS)
1033 want |= CEPH_CAP_ANY_DIR_OPS;
1036 if (want || ci->i_nr_by_mode[PIN_SHIFT] > 0)
1037 want |= CEPH_CAP_PIN;
1043 if (ci->i_nr_by_mode[RD_SHIFT] > 0) {
1044 if (ci->i_nr_by_mode[RD_SHIFT] >= FMODE_WAIT_BIAS ||
1045 time_after(ci->i_last_rd, used_cutoff))
1046 bits |= 1 << RD_SHIFT;
1047 } else if (time_after(ci->i_last_rd, idle_cutoff)) {
1048 bits |= 1 << RD_SHIFT;
1051 if (ci->i_nr_by_mode[WR_SHIFT] > 0) {
1052 if (ci->i_nr_by_mode[WR_SHIFT] >= FMODE_WAIT_BIAS ||
1053 time_after(ci->i_last_wr, used_cutoff))
1054 bits |= 1 << WR_SHIFT;
1055 } else if (time_after(ci->i_last_wr, idle_cutoff)) {
1056 bits |= 1 << WR_SHIFT;
1059 /* check lazyio only when read/write is wanted */
1060 if ((bits & (CEPH_FILE_MODE_RDWR << 1)) &&
1061 ci->i_nr_by_mode[LAZY_SHIFT] > 0)
1062 bits |= 1 << LAZY_SHIFT;
1064 return bits ? ceph_caps_for_mode(bits >> 1) : 0;
1069 * wanted, by virtue of open file modes AND cap refs (buffered/cached data)
1071 int __ceph_caps_wanted(struct ceph_inode_info *ci)
1073 int w = __ceph_caps_file_wanted(ci) | __ceph_caps_used(ci);
1074 if (S_ISDIR(ci->vfs_inode.i_mode)) {
1075 /* we want EXCL if holding caps of dir ops */
1076 if (w & CEPH_CAP_ANY_DIR_OPS)
1077 w |= CEPH_CAP_FILE_EXCL;
1079 /* we want EXCL if dirty data */
1080 if (w & CEPH_CAP_FILE_BUFFER)
1081 w |= CEPH_CAP_FILE_EXCL;
1087 * Return caps we have registered with the MDS(s) as 'wanted'.
1089 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
1091 struct ceph_cap *cap;
1095 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1096 cap = rb_entry(p, struct ceph_cap, ci_node);
1097 if (check && !__cap_is_valid(cap))
1099 if (cap == ci->i_auth_cap)
1100 mds_wanted |= cap->mds_wanted;
1102 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
1107 int ceph_is_any_caps(struct inode *inode)
1109 struct ceph_inode_info *ci = ceph_inode(inode);
1112 spin_lock(&ci->i_ceph_lock);
1113 ret = __ceph_is_any_real_caps(ci);
1114 spin_unlock(&ci->i_ceph_lock);
1119 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
1121 struct ceph_snap_realm *realm = ci->i_snap_realm;
1122 spin_lock(&realm->inodes_with_caps_lock);
1123 list_del_init(&ci->i_snap_realm_item);
1124 ci->i_snap_realm_counter++;
1125 ci->i_snap_realm = NULL;
1126 if (realm->ino == ci->i_vino.ino)
1127 realm->inode = NULL;
1128 spin_unlock(&realm->inodes_with_caps_lock);
1129 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
1134 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
1136 * caller should hold i_ceph_lock.
1137 * caller will not hold session s_mutex if called from destroy_inode.
1139 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1141 struct ceph_mds_session *session = cap->session;
1142 struct ceph_inode_info *ci = cap->ci;
1143 struct ceph_mds_client *mdsc;
1146 /* 'ci' being NULL means the remove have already occurred */
1148 dout("%s: cap inode is NULL\n", __func__);
1152 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1154 mdsc = ceph_inode_to_client(&ci->vfs_inode)->mdsc;
1156 /* remove from inode's cap rbtree, and clear auth cap */
1157 rb_erase(&cap->ci_node, &ci->i_caps);
1158 if (ci->i_auth_cap == cap) {
1159 WARN_ON_ONCE(!list_empty(&ci->i_dirty_item));
1160 ci->i_auth_cap = NULL;
1163 /* remove from session list */
1164 spin_lock(&session->s_cap_lock);
1165 if (session->s_cap_iterator == cap) {
1166 /* not yet, we are iterating over this very cap */
1167 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1170 list_del_init(&cap->session_caps);
1171 session->s_nr_caps--;
1172 atomic64_dec(&mdsc->metric.total_caps);
1173 cap->session = NULL;
1176 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1180 * s_cap_reconnect is protected by s_cap_lock. no one changes
1181 * s_cap_gen while session is in the reconnect state.
1183 if (queue_release &&
1184 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
1185 cap->queue_release = 1;
1187 __ceph_queue_cap_release(session, cap);
1191 cap->queue_release = 0;
1193 cap->cap_ino = ci->i_vino.ino;
1195 spin_unlock(&session->s_cap_lock);
1198 ceph_put_cap(mdsc, cap);
1200 if (!__ceph_is_any_real_caps(ci)) {
1201 /* when reconnect denied, we remove session caps forcibly,
1202 * i_wr_ref can be non-zero. If there are ongoing write,
1203 * keep i_snap_realm.
1205 if (ci->i_wr_ref == 0 && ci->i_snap_realm)
1206 drop_inode_snap_realm(ci);
1208 __cap_delay_cancel(mdsc, ci);
1212 struct cap_msg_args {
1213 struct ceph_mds_session *session;
1214 u64 ino, cid, follows;
1215 u64 flush_tid, oldest_flush_tid, size, max_size;
1218 struct ceph_buffer *xattr_buf;
1219 struct ceph_buffer *old_xattr_buf;
1220 struct timespec64 atime, mtime, ctime, btime;
1221 int op, caps, wanted, dirty;
1222 u32 seq, issue_seq, mseq, time_warp_seq;
1232 * cap struct size + flock buffer size + inline version + inline data size +
1233 * osd_epoch_barrier + oldest_flush_tid
1235 #define CAP_MSG_SIZE (sizeof(struct ceph_mds_caps) + \
1236 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4)
1238 /* Marshal up the cap msg to the MDS */
1239 static void encode_cap_msg(struct ceph_msg *msg, struct cap_msg_args *arg)
1241 struct ceph_mds_caps *fc;
1243 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1245 dout("%s %s %llx %llx caps %s wanted %s dirty %s seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu xattr_ver %llu xattr_len %d\n",
1246 __func__, ceph_cap_op_name(arg->op), arg->cid, arg->ino,
1247 ceph_cap_string(arg->caps), ceph_cap_string(arg->wanted),
1248 ceph_cap_string(arg->dirty), arg->seq, arg->issue_seq,
1249 arg->flush_tid, arg->oldest_flush_tid, arg->mseq, arg->follows,
1250 arg->size, arg->max_size, arg->xattr_version,
1251 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1253 msg->hdr.version = cpu_to_le16(10);
1254 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1256 fc = msg->front.iov_base;
1257 memset(fc, 0, sizeof(*fc));
1259 fc->cap_id = cpu_to_le64(arg->cid);
1260 fc->op = cpu_to_le32(arg->op);
1261 fc->seq = cpu_to_le32(arg->seq);
1262 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1263 fc->migrate_seq = cpu_to_le32(arg->mseq);
1264 fc->caps = cpu_to_le32(arg->caps);
1265 fc->wanted = cpu_to_le32(arg->wanted);
1266 fc->dirty = cpu_to_le32(arg->dirty);
1267 fc->ino = cpu_to_le64(arg->ino);
1268 fc->snap_follows = cpu_to_le64(arg->follows);
1270 fc->size = cpu_to_le64(arg->size);
1271 fc->max_size = cpu_to_le64(arg->max_size);
1272 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1273 ceph_encode_timespec64(&fc->atime, &arg->atime);
1274 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1275 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1277 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1278 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1279 fc->mode = cpu_to_le32(arg->mode);
1281 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1282 if (arg->xattr_buf) {
1283 msg->middle = ceph_buffer_get(arg->xattr_buf);
1284 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1285 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1289 /* flock buffer size (version 2) */
1290 ceph_encode_32(&p, 0);
1291 /* inline version (version 4) */
1292 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1293 /* inline data size */
1294 ceph_encode_32(&p, 0);
1296 * osd_epoch_barrier (version 5)
1297 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1298 * case it was recently changed
1300 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1301 /* oldest_flush_tid (version 6) */
1302 ceph_encode_64(&p, arg->oldest_flush_tid);
1305 * caller_uid/caller_gid (version 7)
1307 * Currently, we don't properly track which caller dirtied the caps
1308 * last, and force a flush of them when there is a conflict. For now,
1309 * just set this to 0:0, to emulate how the MDS has worked up to now.
1311 ceph_encode_32(&p, 0);
1312 ceph_encode_32(&p, 0);
1314 /* pool namespace (version 8) (mds always ignores this) */
1315 ceph_encode_32(&p, 0);
1317 /* btime and change_attr (version 9) */
1318 ceph_encode_timespec64(p, &arg->btime);
1319 p += sizeof(struct ceph_timespec);
1320 ceph_encode_64(&p, arg->change_attr);
1322 /* Advisory flags (version 10) */
1323 ceph_encode_32(&p, arg->flags);
1327 * Queue cap releases when an inode is dropped from our cache.
1329 void __ceph_remove_caps(struct ceph_inode_info *ci)
1333 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1334 * may call __ceph_caps_issued_mask() on a freeing inode. */
1335 spin_lock(&ci->i_ceph_lock);
1336 p = rb_first(&ci->i_caps);
1338 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1340 __ceph_remove_cap(cap, true);
1342 spin_unlock(&ci->i_ceph_lock);
1346 * Prepare to send a cap message to an MDS. Update the cap state, and populate
1347 * the arg struct with the parameters that will need to be sent. This should
1348 * be done under the i_ceph_lock to guard against changes to cap state.
1350 * Make note of max_size reported/requested from mds, revoked caps
1351 * that have now been implemented.
1353 static void __prep_cap(struct cap_msg_args *arg, struct ceph_cap *cap,
1354 int op, int flags, int used, int want, int retain,
1355 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1357 struct ceph_inode_info *ci = cap->ci;
1358 struct inode *inode = &ci->vfs_inode;
1361 lockdep_assert_held(&ci->i_ceph_lock);
1363 held = cap->issued | cap->implemented;
1364 revoking = cap->implemented & ~cap->issued;
1365 retain &= ~revoking;
1367 dout("%s %p cap %p session %p %s -> %s (revoking %s)\n",
1368 __func__, inode, cap, cap->session,
1369 ceph_cap_string(held), ceph_cap_string(held & retain),
1370 ceph_cap_string(revoking));
1371 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1373 ci->i_ceph_flags &= ~CEPH_I_FLUSH;
1375 cap->issued &= retain; /* drop bits we don't want */
1377 * Wake up any waiters on wanted -> needed transition. This is due to
1378 * the weird transition from buffered to sync IO... we need to flush
1379 * dirty pages _before_ allowing sync writes to avoid reordering.
1381 arg->wake = cap->implemented & ~cap->issued;
1382 cap->implemented &= cap->issued | used;
1383 cap->mds_wanted = want;
1385 arg->session = cap->session;
1386 arg->ino = ceph_vino(inode).ino;
1387 arg->cid = cap->cap_id;
1388 arg->follows = flushing ? ci->i_head_snapc->seq : 0;
1389 arg->flush_tid = flush_tid;
1390 arg->oldest_flush_tid = oldest_flush_tid;
1392 arg->size = inode->i_size;
1393 ci->i_reported_size = arg->size;
1394 arg->max_size = ci->i_wanted_max_size;
1395 if (cap == ci->i_auth_cap) {
1396 if (want & CEPH_CAP_ANY_FILE_WR)
1397 ci->i_requested_max_size = arg->max_size;
1399 ci->i_requested_max_size = 0;
1402 if (flushing & CEPH_CAP_XATTR_EXCL) {
1403 arg->old_xattr_buf = __ceph_build_xattrs_blob(ci);
1404 arg->xattr_version = ci->i_xattrs.version;
1405 arg->xattr_buf = ci->i_xattrs.blob;
1407 arg->xattr_buf = NULL;
1408 arg->old_xattr_buf = NULL;
1411 arg->mtime = inode->i_mtime;
1412 arg->atime = inode->i_atime;
1413 arg->ctime = inode->i_ctime;
1414 arg->btime = ci->i_btime;
1415 arg->change_attr = inode_peek_iversion_raw(inode);
1418 arg->caps = cap->implemented;
1420 arg->dirty = flushing;
1422 arg->seq = cap->seq;
1423 arg->issue_seq = cap->issue_seq;
1424 arg->mseq = cap->mseq;
1425 arg->time_warp_seq = ci->i_time_warp_seq;
1427 arg->uid = inode->i_uid;
1428 arg->gid = inode->i_gid;
1429 arg->mode = inode->i_mode;
1431 arg->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1432 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1433 !list_empty(&ci->i_cap_snaps)) {
1434 struct ceph_cap_snap *capsnap;
1435 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1436 if (capsnap->cap_flush.tid)
1438 if (capsnap->need_flush) {
1439 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1448 * Send a cap msg on the given inode.
1450 * Caller should hold snap_rwsem (read), s_mutex.
1452 static void __send_cap(struct cap_msg_args *arg, struct ceph_inode_info *ci)
1454 struct ceph_msg *msg;
1455 struct inode *inode = &ci->vfs_inode;
1457 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1459 pr_err("error allocating cap msg: ino (%llx.%llx) flushing %s tid %llu, requeuing cap.\n",
1460 ceph_vinop(inode), ceph_cap_string(arg->dirty),
1462 spin_lock(&ci->i_ceph_lock);
1463 __cap_delay_requeue(arg->session->s_mdsc, ci);
1464 spin_unlock(&ci->i_ceph_lock);
1468 encode_cap_msg(msg, arg);
1469 ceph_con_send(&arg->session->s_con, msg);
1470 ceph_buffer_put(arg->old_xattr_buf);
1472 wake_up_all(&ci->i_cap_wq);
1475 static inline int __send_flush_snap(struct inode *inode,
1476 struct ceph_mds_session *session,
1477 struct ceph_cap_snap *capsnap,
1478 u32 mseq, u64 oldest_flush_tid)
1480 struct cap_msg_args arg;
1481 struct ceph_msg *msg;
1483 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1487 arg.session = session;
1488 arg.ino = ceph_vino(inode).ino;
1490 arg.follows = capsnap->follows;
1491 arg.flush_tid = capsnap->cap_flush.tid;
1492 arg.oldest_flush_tid = oldest_flush_tid;
1494 arg.size = capsnap->size;
1496 arg.xattr_version = capsnap->xattr_version;
1497 arg.xattr_buf = capsnap->xattr_blob;
1498 arg.old_xattr_buf = NULL;
1500 arg.atime = capsnap->atime;
1501 arg.mtime = capsnap->mtime;
1502 arg.ctime = capsnap->ctime;
1503 arg.btime = capsnap->btime;
1504 arg.change_attr = capsnap->change_attr;
1506 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1507 arg.caps = capsnap->issued;
1509 arg.dirty = capsnap->dirty;
1514 arg.time_warp_seq = capsnap->time_warp_seq;
1516 arg.uid = capsnap->uid;
1517 arg.gid = capsnap->gid;
1518 arg.mode = capsnap->mode;
1520 arg.inline_data = capsnap->inline_data;
1524 encode_cap_msg(msg, &arg);
1525 ceph_con_send(&arg.session->s_con, msg);
1530 * When a snapshot is taken, clients accumulate dirty metadata on
1531 * inodes with capabilities in ceph_cap_snaps to describe the file
1532 * state at the time the snapshot was taken. This must be flushed
1533 * asynchronously back to the MDS once sync writes complete and dirty
1534 * data is written out.
1536 * Called under i_ceph_lock. Takes s_mutex as needed.
1538 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1539 struct ceph_mds_session *session)
1540 __releases(ci->i_ceph_lock)
1541 __acquires(ci->i_ceph_lock)
1543 struct inode *inode = &ci->vfs_inode;
1544 struct ceph_mds_client *mdsc = session->s_mdsc;
1545 struct ceph_cap_snap *capsnap;
1546 u64 oldest_flush_tid = 0;
1547 u64 first_tid = 1, last_tid = 0;
1549 dout("__flush_snaps %p session %p\n", inode, session);
1551 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1553 * we need to wait for sync writes to complete and for dirty
1554 * pages to be written out.
1556 if (capsnap->dirty_pages || capsnap->writing)
1559 /* should be removed by ceph_try_drop_cap_snap() */
1560 BUG_ON(!capsnap->need_flush);
1562 /* only flush each capsnap once */
1563 if (capsnap->cap_flush.tid > 0) {
1564 dout(" already flushed %p, skipping\n", capsnap);
1568 spin_lock(&mdsc->cap_dirty_lock);
1569 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1570 list_add_tail(&capsnap->cap_flush.g_list,
1571 &mdsc->cap_flush_list);
1572 if (oldest_flush_tid == 0)
1573 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1574 if (list_empty(&ci->i_flushing_item)) {
1575 list_add_tail(&ci->i_flushing_item,
1576 &session->s_cap_flushing);
1578 spin_unlock(&mdsc->cap_dirty_lock);
1580 list_add_tail(&capsnap->cap_flush.i_list,
1581 &ci->i_cap_flush_list);
1584 first_tid = capsnap->cap_flush.tid;
1585 last_tid = capsnap->cap_flush.tid;
1588 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1590 while (first_tid <= last_tid) {
1591 struct ceph_cap *cap = ci->i_auth_cap;
1592 struct ceph_cap_flush *cf;
1595 if (!(cap && cap->session == session)) {
1596 dout("__flush_snaps %p auth cap %p not mds%d, "
1597 "stop\n", inode, cap, session->s_mds);
1602 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1603 if (cf->tid >= first_tid) {
1611 first_tid = cf->tid + 1;
1613 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1614 refcount_inc(&capsnap->nref);
1615 spin_unlock(&ci->i_ceph_lock);
1617 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1618 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1620 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1623 pr_err("__flush_snaps: error sending cap flushsnap, "
1624 "ino (%llx.%llx) tid %llu follows %llu\n",
1625 ceph_vinop(inode), cf->tid, capsnap->follows);
1628 ceph_put_cap_snap(capsnap);
1629 spin_lock(&ci->i_ceph_lock);
1633 void ceph_flush_snaps(struct ceph_inode_info *ci,
1634 struct ceph_mds_session **psession)
1636 struct inode *inode = &ci->vfs_inode;
1637 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1638 struct ceph_mds_session *session = NULL;
1641 dout("ceph_flush_snaps %p\n", inode);
1643 session = *psession;
1645 spin_lock(&ci->i_ceph_lock);
1646 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1647 dout(" no capsnap needs flush, doing nothing\n");
1650 if (!ci->i_auth_cap) {
1651 dout(" no auth cap (migrating?), doing nothing\n");
1655 mds = ci->i_auth_cap->session->s_mds;
1656 if (session && session->s_mds != mds) {
1657 dout(" oops, wrong session %p mutex\n", session);
1658 mutex_unlock(&session->s_mutex);
1659 ceph_put_mds_session(session);
1663 spin_unlock(&ci->i_ceph_lock);
1664 mutex_lock(&mdsc->mutex);
1665 session = __ceph_lookup_mds_session(mdsc, mds);
1666 mutex_unlock(&mdsc->mutex);
1668 dout(" inverting session/ino locks on %p\n", session);
1669 mutex_lock(&session->s_mutex);
1674 // make sure flushsnap messages are sent in proper order.
1675 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1676 __kick_flushing_caps(mdsc, session, ci, 0);
1678 __ceph_flush_snaps(ci, session);
1680 spin_unlock(&ci->i_ceph_lock);
1683 *psession = session;
1684 } else if (session) {
1685 mutex_unlock(&session->s_mutex);
1686 ceph_put_mds_session(session);
1688 /* we flushed them all; remove this inode from the queue */
1689 spin_lock(&mdsc->snap_flush_lock);
1690 list_del_init(&ci->i_snap_flush_item);
1691 spin_unlock(&mdsc->snap_flush_lock);
1695 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1696 * Caller is then responsible for calling __mark_inode_dirty with the
1697 * returned flags value.
1699 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1700 struct ceph_cap_flush **pcf)
1702 struct ceph_mds_client *mdsc =
1703 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1704 struct inode *inode = &ci->vfs_inode;
1705 int was = ci->i_dirty_caps;
1708 lockdep_assert_held(&ci->i_ceph_lock);
1710 if (!ci->i_auth_cap) {
1711 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1712 "but no auth cap (session was closed?)\n",
1713 inode, ceph_ino(inode), ceph_cap_string(mask));
1717 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1718 ceph_cap_string(mask), ceph_cap_string(was),
1719 ceph_cap_string(was | mask));
1720 ci->i_dirty_caps |= mask;
1722 struct ceph_mds_session *session = ci->i_auth_cap->session;
1724 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1725 swap(ci->i_prealloc_cap_flush, *pcf);
1727 if (!ci->i_head_snapc) {
1728 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1729 ci->i_head_snapc = ceph_get_snap_context(
1730 ci->i_snap_realm->cached_context);
1732 dout(" inode %p now dirty snapc %p auth cap %p\n",
1733 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1734 BUG_ON(!list_empty(&ci->i_dirty_item));
1735 spin_lock(&mdsc->cap_dirty_lock);
1736 list_add(&ci->i_dirty_item, &session->s_cap_dirty);
1737 spin_unlock(&mdsc->cap_dirty_lock);
1738 if (ci->i_flushing_caps == 0) {
1740 dirty |= I_DIRTY_SYNC;
1743 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1745 BUG_ON(list_empty(&ci->i_dirty_item));
1746 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1747 (mask & CEPH_CAP_FILE_BUFFER))
1748 dirty |= I_DIRTY_DATASYNC;
1749 __cap_delay_requeue(mdsc, ci);
1753 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1755 struct ceph_cap_flush *cf;
1757 cf = kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1761 cf->is_capsnap = false;
1765 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1768 kmem_cache_free(ceph_cap_flush_cachep, cf);
1771 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1773 if (!list_empty(&mdsc->cap_flush_list)) {
1774 struct ceph_cap_flush *cf =
1775 list_first_entry(&mdsc->cap_flush_list,
1776 struct ceph_cap_flush, g_list);
1783 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1784 * Return true if caller needs to wake up flush waiters.
1786 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1787 struct ceph_cap_flush *cf)
1789 struct ceph_cap_flush *prev;
1790 bool wake = cf->wake;
1792 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1793 prev = list_prev_entry(cf, g_list);
1797 list_del_init(&cf->g_list);
1801 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1802 struct ceph_cap_flush *cf)
1804 struct ceph_cap_flush *prev;
1805 bool wake = cf->wake;
1807 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1808 prev = list_prev_entry(cf, i_list);
1812 list_del_init(&cf->i_list);
1817 * Add dirty inode to the flushing list. Assigned a seq number so we
1818 * can wait for caps to flush without starving.
1820 * Called under i_ceph_lock. Returns the flush tid.
1822 static u64 __mark_caps_flushing(struct inode *inode,
1823 struct ceph_mds_session *session, bool wake,
1824 u64 *oldest_flush_tid)
1826 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1827 struct ceph_inode_info *ci = ceph_inode(inode);
1828 struct ceph_cap_flush *cf = NULL;
1831 lockdep_assert_held(&ci->i_ceph_lock);
1832 BUG_ON(ci->i_dirty_caps == 0);
1833 BUG_ON(list_empty(&ci->i_dirty_item));
1834 BUG_ON(!ci->i_prealloc_cap_flush);
1836 flushing = ci->i_dirty_caps;
1837 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1838 ceph_cap_string(flushing),
1839 ceph_cap_string(ci->i_flushing_caps),
1840 ceph_cap_string(ci->i_flushing_caps | flushing));
1841 ci->i_flushing_caps |= flushing;
1842 ci->i_dirty_caps = 0;
1843 dout(" inode %p now !dirty\n", inode);
1845 swap(cf, ci->i_prealloc_cap_flush);
1846 cf->caps = flushing;
1849 spin_lock(&mdsc->cap_dirty_lock);
1850 list_del_init(&ci->i_dirty_item);
1852 cf->tid = ++mdsc->last_cap_flush_tid;
1853 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1854 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1856 if (list_empty(&ci->i_flushing_item)) {
1857 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1858 mdsc->num_cap_flushing++;
1860 spin_unlock(&mdsc->cap_dirty_lock);
1862 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1868 * try to invalidate mapping pages without blocking.
1870 static int try_nonblocking_invalidate(struct inode *inode)
1871 __releases(ci->i_ceph_lock)
1872 __acquires(ci->i_ceph_lock)
1874 struct ceph_inode_info *ci = ceph_inode(inode);
1875 u32 invalidating_gen = ci->i_rdcache_gen;
1877 spin_unlock(&ci->i_ceph_lock);
1878 ceph_fscache_invalidate(inode);
1879 invalidate_mapping_pages(&inode->i_data, 0, -1);
1880 spin_lock(&ci->i_ceph_lock);
1882 if (inode->i_data.nrpages == 0 &&
1883 invalidating_gen == ci->i_rdcache_gen) {
1885 dout("try_nonblocking_invalidate %p success\n", inode);
1886 /* save any racing async invalidate some trouble */
1887 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1890 dout("try_nonblocking_invalidate %p failed\n", inode);
1894 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1896 loff_t size = ci->vfs_inode.i_size;
1897 /* mds will adjust max size according to the reported size */
1898 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1900 if (size >= ci->i_max_size)
1902 /* half of previous max_size increment has been used */
1903 if (ci->i_max_size > ci->i_reported_size &&
1904 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1910 * Swiss army knife function to examine currently used and wanted
1911 * versus held caps. Release, flush, ack revoked caps to mds as
1914 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1915 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1918 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1919 struct ceph_mds_session *session)
1921 struct inode *inode = &ci->vfs_inode;
1922 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1923 struct ceph_cap *cap;
1924 u64 flush_tid, oldest_flush_tid;
1925 int file_wanted, used, cap_used;
1926 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1927 int issued, implemented, want, retain, revoking, flushing = 0;
1928 int mds = -1; /* keep track of how far we've gone through i_caps list
1929 to avoid an infinite loop on retry */
1931 bool queue_invalidate = false;
1932 bool tried_invalidate = false;
1934 spin_lock(&ci->i_ceph_lock);
1935 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1936 flags |= CHECK_CAPS_FLUSH;
1940 spin_lock(&ci->i_ceph_lock);
1942 /* Caps wanted by virtue of active open files. */
1943 file_wanted = __ceph_caps_file_wanted(ci);
1945 /* Caps which have active references against them */
1946 used = __ceph_caps_used(ci);
1949 * "issued" represents the current caps that the MDS wants us to have.
1950 * "implemented" is the set that we have been granted, and includes the
1951 * ones that have not yet been returned to the MDS (the "revoking" set,
1952 * usually because they have outstanding references).
1954 issued = __ceph_caps_issued(ci, &implemented);
1955 revoking = implemented & ~issued;
1959 /* The ones we currently want to retain (may be adjusted below) */
1960 retain = file_wanted | used | CEPH_CAP_PIN;
1961 if (!mdsc->stopping && inode->i_nlink > 0) {
1963 retain |= CEPH_CAP_ANY; /* be greedy */
1964 } else if (S_ISDIR(inode->i_mode) &&
1965 (issued & CEPH_CAP_FILE_SHARED) &&
1966 __ceph_dir_is_complete(ci)) {
1968 * If a directory is complete, we want to keep
1969 * the exclusive cap. So that MDS does not end up
1970 * revoking the shared cap on every create/unlink
1973 if (IS_RDONLY(inode)) {
1974 want = CEPH_CAP_ANY_SHARED;
1976 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1981 retain |= CEPH_CAP_ANY_SHARED;
1983 * keep RD only if we didn't have the file open RW,
1984 * because then the mds would revoke it anyway to
1985 * journal max_size=0.
1987 if (ci->i_max_size == 0)
1988 retain |= CEPH_CAP_ANY_RD;
1992 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1993 " issued %s revoking %s retain %s %s%s\n", inode,
1994 ceph_cap_string(file_wanted),
1995 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1996 ceph_cap_string(ci->i_flushing_caps),
1997 ceph_cap_string(issued), ceph_cap_string(revoking),
1998 ceph_cap_string(retain),
1999 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
2000 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
2003 * If we no longer need to hold onto old our caps, and we may
2004 * have cached pages, but don't want them, then try to invalidate.
2005 * If we fail, it's because pages are locked.... try again later.
2007 if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
2008 S_ISREG(inode->i_mode) &&
2009 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
2010 inode->i_data.nrpages && /* have cached pages */
2011 (revoking & (CEPH_CAP_FILE_CACHE|
2012 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
2013 !tried_invalidate) {
2014 dout("check_caps trying to invalidate on %p\n", inode);
2015 if (try_nonblocking_invalidate(inode) < 0) {
2016 dout("check_caps queuing invalidate\n");
2017 queue_invalidate = true;
2018 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2020 tried_invalidate = true;
2024 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2026 struct cap_msg_args arg;
2028 cap = rb_entry(p, struct ceph_cap, ci_node);
2030 /* avoid looping forever */
2031 if (mds >= cap->mds ||
2032 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
2035 /* NOTE: no side-effects allowed, until we take s_mutex */
2038 * If we have an auth cap, we don't need to consider any
2039 * overlapping caps as used.
2042 if (ci->i_auth_cap && cap != ci->i_auth_cap)
2043 cap_used &= ~ci->i_auth_cap->issued;
2045 revoking = cap->implemented & ~cap->issued;
2046 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
2047 cap->mds, cap, ceph_cap_string(cap_used),
2048 ceph_cap_string(cap->issued),
2049 ceph_cap_string(cap->implemented),
2050 ceph_cap_string(revoking));
2052 if (cap == ci->i_auth_cap &&
2053 (cap->issued & CEPH_CAP_FILE_WR)) {
2054 /* request larger max_size from MDS? */
2055 if (ci->i_wanted_max_size > ci->i_max_size &&
2056 ci->i_wanted_max_size > ci->i_requested_max_size) {
2057 dout("requesting new max_size\n");
2061 /* approaching file_max? */
2062 if (__ceph_should_report_size(ci)) {
2063 dout("i_size approaching max_size\n");
2067 /* flush anything dirty? */
2068 if (cap == ci->i_auth_cap) {
2069 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
2070 dout("flushing dirty caps\n");
2073 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
2074 dout("flushing snap caps\n");
2079 /* completed revocation? going down and there are no caps? */
2080 if (revoking && (revoking & cap_used) == 0) {
2081 dout("completed revocation of %s\n",
2082 ceph_cap_string(cap->implemented & ~cap->issued));
2086 /* want more caps from mds? */
2087 if (want & ~cap->mds_wanted) {
2088 if (want & ~(cap->mds_wanted | cap->issued))
2090 if (!__cap_is_valid(cap))
2094 /* things we might delay */
2095 if ((cap->issued & ~retain) == 0)
2096 continue; /* nope, all good */
2099 if (session && session != cap->session) {
2100 dout("oops, wrong session %p mutex\n", session);
2101 mutex_unlock(&session->s_mutex);
2105 session = cap->session;
2106 if (mutex_trylock(&session->s_mutex) == 0) {
2107 dout("inverting session/ino locks on %p\n",
2109 session = ceph_get_mds_session(session);
2110 spin_unlock(&ci->i_ceph_lock);
2111 if (took_snap_rwsem) {
2112 up_read(&mdsc->snap_rwsem);
2113 took_snap_rwsem = 0;
2116 mutex_lock(&session->s_mutex);
2117 ceph_put_mds_session(session);
2120 * Because we take the reference while
2121 * holding the i_ceph_lock, it should
2122 * never be NULL. Throw a warning if it
2131 /* kick flushing and flush snaps before sending normal
2133 if (cap == ci->i_auth_cap &&
2135 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2136 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2137 __kick_flushing_caps(mdsc, session, ci, 0);
2138 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2139 __ceph_flush_snaps(ci, session);
2144 /* take snap_rwsem after session mutex */
2145 if (!took_snap_rwsem) {
2146 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
2147 dout("inverting snap/in locks on %p\n",
2149 spin_unlock(&ci->i_ceph_lock);
2150 down_read(&mdsc->snap_rwsem);
2151 took_snap_rwsem = 1;
2154 took_snap_rwsem = 1;
2157 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2158 flushing = ci->i_dirty_caps;
2159 flush_tid = __mark_caps_flushing(inode, session, false,
2161 if (flags & CHECK_CAPS_FLUSH &&
2162 list_empty(&session->s_cap_dirty))
2163 mflags |= CEPH_CLIENT_CAPS_SYNC;
2167 spin_lock(&mdsc->cap_dirty_lock);
2168 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2169 spin_unlock(&mdsc->cap_dirty_lock);
2172 mds = cap->mds; /* remember mds, so we don't repeat */
2174 __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2175 want, retain, flushing, flush_tid, oldest_flush_tid);
2176 spin_unlock(&ci->i_ceph_lock);
2178 __send_cap(&arg, ci);
2180 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2183 /* periodically re-calculate caps wanted by open files */
2184 if (__ceph_is_any_real_caps(ci) &&
2185 list_empty(&ci->i_cap_delay_list) &&
2186 (file_wanted & ~CEPH_CAP_PIN) &&
2187 !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2188 __cap_delay_requeue(mdsc, ci);
2191 spin_unlock(&ci->i_ceph_lock);
2193 if (queue_invalidate)
2194 ceph_queue_invalidate(inode);
2197 mutex_unlock(&session->s_mutex);
2198 if (took_snap_rwsem)
2199 up_read(&mdsc->snap_rwsem);
2203 * Try to flush dirty caps back to the auth mds.
2205 static int try_flush_caps(struct inode *inode, u64 *ptid)
2207 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2208 struct ceph_inode_info *ci = ceph_inode(inode);
2209 struct ceph_mds_session *session = NULL;
2211 u64 flush_tid = 0, oldest_flush_tid = 0;
2214 spin_lock(&ci->i_ceph_lock);
2216 if (ci->i_dirty_caps && ci->i_auth_cap) {
2217 struct ceph_cap *cap = ci->i_auth_cap;
2218 struct cap_msg_args arg;
2220 if (session != cap->session) {
2221 spin_unlock(&ci->i_ceph_lock);
2223 mutex_unlock(&session->s_mutex);
2224 session = cap->session;
2225 mutex_lock(&session->s_mutex);
2228 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) {
2229 spin_unlock(&ci->i_ceph_lock);
2233 if (ci->i_ceph_flags &
2234 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2235 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2236 __kick_flushing_caps(mdsc, session, ci, 0);
2237 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2238 __ceph_flush_snaps(ci, session);
2242 flushing = ci->i_dirty_caps;
2243 flush_tid = __mark_caps_flushing(inode, session, true,
2246 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2247 __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2248 (cap->issued | cap->implemented),
2249 flushing, flush_tid, oldest_flush_tid);
2250 spin_unlock(&ci->i_ceph_lock);
2252 __send_cap(&arg, ci);
2254 if (!list_empty(&ci->i_cap_flush_list)) {
2255 struct ceph_cap_flush *cf =
2256 list_last_entry(&ci->i_cap_flush_list,
2257 struct ceph_cap_flush, i_list);
2259 flush_tid = cf->tid;
2261 flushing = ci->i_flushing_caps;
2262 spin_unlock(&ci->i_ceph_lock);
2266 mutex_unlock(&session->s_mutex);
2273 * Return true if we've flushed caps through the given flush_tid.
2275 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2277 struct ceph_inode_info *ci = ceph_inode(inode);
2280 spin_lock(&ci->i_ceph_lock);
2281 if (!list_empty(&ci->i_cap_flush_list)) {
2282 struct ceph_cap_flush * cf =
2283 list_first_entry(&ci->i_cap_flush_list,
2284 struct ceph_cap_flush, i_list);
2285 if (cf->tid <= flush_tid)
2288 spin_unlock(&ci->i_ceph_lock);
2293 * wait for any unsafe requests to complete.
2295 static int unsafe_request_wait(struct inode *inode)
2297 struct ceph_inode_info *ci = ceph_inode(inode);
2298 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2301 spin_lock(&ci->i_unsafe_lock);
2302 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2303 req1 = list_last_entry(&ci->i_unsafe_dirops,
2304 struct ceph_mds_request,
2306 ceph_mdsc_get_request(req1);
2308 if (!list_empty(&ci->i_unsafe_iops)) {
2309 req2 = list_last_entry(&ci->i_unsafe_iops,
2310 struct ceph_mds_request,
2311 r_unsafe_target_item);
2312 ceph_mdsc_get_request(req2);
2314 spin_unlock(&ci->i_unsafe_lock);
2316 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2317 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2319 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2320 ceph_timeout_jiffies(req1->r_timeout));
2323 ceph_mdsc_put_request(req1);
2326 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2327 ceph_timeout_jiffies(req2->r_timeout));
2330 ceph_mdsc_put_request(req2);
2335 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2337 struct inode *inode = file->f_mapping->host;
2338 struct ceph_inode_info *ci = ceph_inode(inode);
2343 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2345 ret = file_write_and_wait_range(file, start, end);
2349 ret = ceph_wait_on_async_create(inode);
2353 dirty = try_flush_caps(inode, &flush_tid);
2354 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2356 err = unsafe_request_wait(inode);
2359 * only wait on non-file metadata writeback (the mds
2360 * can recover size and mtime, so we don't need to
2363 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2364 err = wait_event_interruptible(ci->i_cap_wq,
2365 caps_are_flushed(inode, flush_tid));
2371 err = file_check_and_advance_wb_err(file);
2375 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2380 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2381 * queue inode for flush but don't do so immediately, because we can
2382 * get by with fewer MDS messages if we wait for data writeback to
2385 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2387 struct ceph_inode_info *ci = ceph_inode(inode);
2391 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2393 dout("write_inode %p wait=%d\n", inode, wait);
2395 dirty = try_flush_caps(inode, &flush_tid);
2397 err = wait_event_interruptible(ci->i_cap_wq,
2398 caps_are_flushed(inode, flush_tid));
2400 struct ceph_mds_client *mdsc =
2401 ceph_sb_to_client(inode->i_sb)->mdsc;
2403 spin_lock(&ci->i_ceph_lock);
2404 if (__ceph_caps_dirty(ci))
2405 __cap_delay_requeue_front(mdsc, ci);
2406 spin_unlock(&ci->i_ceph_lock);
2411 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2412 struct ceph_mds_session *session,
2413 struct ceph_inode_info *ci,
2414 u64 oldest_flush_tid)
2415 __releases(ci->i_ceph_lock)
2416 __acquires(ci->i_ceph_lock)
2418 struct inode *inode = &ci->vfs_inode;
2419 struct ceph_cap *cap;
2420 struct ceph_cap_flush *cf;
2423 u64 last_snap_flush = 0;
2425 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2427 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2428 if (cf->is_capsnap) {
2429 last_snap_flush = cf->tid;
2434 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2435 if (cf->tid < first_tid)
2438 cap = ci->i_auth_cap;
2439 if (!(cap && cap->session == session)) {
2440 pr_err("%p auth cap %p not mds%d ???\n",
2441 inode, cap, session->s_mds);
2445 first_tid = cf->tid + 1;
2447 if (!cf->is_capsnap) {
2448 struct cap_msg_args arg;
2450 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2451 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2452 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2453 (cf->tid < last_snap_flush ?
2454 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2455 __ceph_caps_used(ci),
2456 __ceph_caps_wanted(ci),
2457 (cap->issued | cap->implemented),
2458 cf->caps, cf->tid, oldest_flush_tid);
2459 spin_unlock(&ci->i_ceph_lock);
2460 __send_cap(&arg, ci);
2462 struct ceph_cap_snap *capsnap =
2463 container_of(cf, struct ceph_cap_snap,
2465 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2466 inode, capsnap, cf->tid,
2467 ceph_cap_string(capsnap->dirty));
2469 refcount_inc(&capsnap->nref);
2470 spin_unlock(&ci->i_ceph_lock);
2472 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2475 pr_err("kick_flushing_caps: error sending "
2476 "cap flushsnap, ino (%llx.%llx) "
2477 "tid %llu follows %llu\n",
2478 ceph_vinop(inode), cf->tid,
2482 ceph_put_cap_snap(capsnap);
2485 spin_lock(&ci->i_ceph_lock);
2489 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2490 struct ceph_mds_session *session)
2492 struct ceph_inode_info *ci;
2493 struct ceph_cap *cap;
2494 u64 oldest_flush_tid;
2496 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2498 spin_lock(&mdsc->cap_dirty_lock);
2499 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2500 spin_unlock(&mdsc->cap_dirty_lock);
2502 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2503 spin_lock(&ci->i_ceph_lock);
2504 cap = ci->i_auth_cap;
2505 if (!(cap && cap->session == session)) {
2506 pr_err("%p auth cap %p not mds%d ???\n",
2507 &ci->vfs_inode, cap, session->s_mds);
2508 spin_unlock(&ci->i_ceph_lock);
2514 * if flushing caps were revoked, we re-send the cap flush
2515 * in client reconnect stage. This guarantees MDS * processes
2516 * the cap flush message before issuing the flushing caps to
2519 if ((cap->issued & ci->i_flushing_caps) !=
2520 ci->i_flushing_caps) {
2521 /* encode_caps_cb() also will reset these sequence
2522 * numbers. make sure sequence numbers in cap flush
2523 * message match later reconnect message */
2527 __kick_flushing_caps(mdsc, session, ci,
2530 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2533 spin_unlock(&ci->i_ceph_lock);
2537 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2538 struct ceph_mds_session *session)
2540 struct ceph_inode_info *ci;
2541 struct ceph_cap *cap;
2542 u64 oldest_flush_tid;
2544 lockdep_assert_held(&session->s_mutex);
2546 dout("kick_flushing_caps mds%d\n", session->s_mds);
2548 spin_lock(&mdsc->cap_dirty_lock);
2549 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2550 spin_unlock(&mdsc->cap_dirty_lock);
2552 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2553 spin_lock(&ci->i_ceph_lock);
2554 cap = ci->i_auth_cap;
2555 if (!(cap && cap->session == session)) {
2556 pr_err("%p auth cap %p not mds%d ???\n",
2557 &ci->vfs_inode, cap, session->s_mds);
2558 spin_unlock(&ci->i_ceph_lock);
2561 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2562 __kick_flushing_caps(mdsc, session, ci,
2565 spin_unlock(&ci->i_ceph_lock);
2569 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2570 struct ceph_inode_info *ci)
2572 struct ceph_mds_client *mdsc = session->s_mdsc;
2573 struct ceph_cap *cap = ci->i_auth_cap;
2575 lockdep_assert_held(&ci->i_ceph_lock);
2577 dout("%s %p flushing %s\n", __func__, &ci->vfs_inode,
2578 ceph_cap_string(ci->i_flushing_caps));
2580 if (!list_empty(&ci->i_cap_flush_list)) {
2581 u64 oldest_flush_tid;
2582 spin_lock(&mdsc->cap_dirty_lock);
2583 list_move_tail(&ci->i_flushing_item,
2584 &cap->session->s_cap_flushing);
2585 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2586 spin_unlock(&mdsc->cap_dirty_lock);
2588 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2594 * Take references to capabilities we hold, so that we don't release
2595 * them to the MDS prematurely.
2597 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2598 bool snap_rwsem_locked)
2600 lockdep_assert_held(&ci->i_ceph_lock);
2602 if (got & CEPH_CAP_PIN)
2604 if (got & CEPH_CAP_FILE_RD)
2606 if (got & CEPH_CAP_FILE_CACHE)
2607 ci->i_rdcache_ref++;
2608 if (got & CEPH_CAP_FILE_EXCL)
2610 if (got & CEPH_CAP_FILE_WR) {
2611 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2612 BUG_ON(!snap_rwsem_locked);
2613 ci->i_head_snapc = ceph_get_snap_context(
2614 ci->i_snap_realm->cached_context);
2618 if (got & CEPH_CAP_FILE_BUFFER) {
2619 if (ci->i_wb_ref == 0)
2620 ihold(&ci->vfs_inode);
2622 dout("%s %p wb %d -> %d (?)\n", __func__,
2623 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2628 * Try to grab cap references. Specify those refs we @want, and the
2629 * minimal set we @need. Also include the larger offset we are writing
2630 * to (when applicable), and check against max_size here as well.
2631 * Note that caller is responsible for ensuring max_size increases are
2632 * requested from the MDS.
2634 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2635 * or a negative error code. There are 3 speical error codes:
2636 * -EAGAIN: need to sleep but non-blocking is specified
2637 * -EFBIG: ask caller to call check_max_size() and try again.
2638 * -ESTALE: ask caller to call ceph_renew_caps() and try again.
2641 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2642 NON_BLOCKING = (1 << 8),
2643 CHECK_FILELOCK = (1 << 9),
2646 static int try_get_cap_refs(struct inode *inode, int need, int want,
2647 loff_t endoff, int flags, int *got)
2649 struct ceph_inode_info *ci = ceph_inode(inode);
2650 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2652 int have, implemented;
2653 bool snap_rwsem_locked = false;
2655 dout("get_cap_refs %p need %s want %s\n", inode,
2656 ceph_cap_string(need), ceph_cap_string(want));
2659 spin_lock(&ci->i_ceph_lock);
2661 if ((flags & CHECK_FILELOCK) &&
2662 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2663 dout("try_get_cap_refs %p error filelock\n", inode);
2668 /* finish pending truncate */
2669 while (ci->i_truncate_pending) {
2670 spin_unlock(&ci->i_ceph_lock);
2671 if (snap_rwsem_locked) {
2672 up_read(&mdsc->snap_rwsem);
2673 snap_rwsem_locked = false;
2675 __ceph_do_pending_vmtruncate(inode);
2676 spin_lock(&ci->i_ceph_lock);
2679 have = __ceph_caps_issued(ci, &implemented);
2681 if (have & need & CEPH_CAP_FILE_WR) {
2682 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2683 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2684 inode, endoff, ci->i_max_size);
2685 if (endoff > ci->i_requested_max_size)
2686 ret = ci->i_auth_cap ? -EFBIG : -ESTALE;
2690 * If a sync write is in progress, we must wait, so that we
2691 * can get a final snapshot value for size+mtime.
2693 if (__ceph_have_pending_cap_snap(ci)) {
2694 dout("get_cap_refs %p cap_snap_pending\n", inode);
2699 if ((have & need) == need) {
2701 * Look at (implemented & ~have & not) so that we keep waiting
2702 * on transition from wanted -> needed caps. This is needed
2703 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2704 * going before a prior buffered writeback happens.
2706 int not = want & ~(have & need);
2707 int revoking = implemented & ~have;
2708 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2709 inode, ceph_cap_string(have), ceph_cap_string(not),
2710 ceph_cap_string(revoking));
2711 if ((revoking & not) == 0) {
2712 if (!snap_rwsem_locked &&
2713 !ci->i_head_snapc &&
2714 (need & CEPH_CAP_FILE_WR)) {
2715 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2717 * we can not call down_read() when
2718 * task isn't in TASK_RUNNING state
2720 if (flags & NON_BLOCKING) {
2725 spin_unlock(&ci->i_ceph_lock);
2726 down_read(&mdsc->snap_rwsem);
2727 snap_rwsem_locked = true;
2730 snap_rwsem_locked = true;
2732 if ((have & want) == want)
2736 if (S_ISREG(inode->i_mode) &&
2737 (need & CEPH_CAP_FILE_RD) &&
2738 !(*got & CEPH_CAP_FILE_CACHE))
2739 ceph_disable_fscache_readpage(ci);
2740 ceph_take_cap_refs(ci, *got, true);
2744 int session_readonly = false;
2746 if (ci->i_auth_cap &&
2747 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2748 struct ceph_mds_session *s = ci->i_auth_cap->session;
2749 spin_lock(&s->s_cap_lock);
2750 session_readonly = s->s_readonly;
2751 spin_unlock(&s->s_cap_lock);
2753 if (session_readonly) {
2754 dout("get_cap_refs %p need %s but mds%d readonly\n",
2755 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2760 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2761 dout("get_cap_refs %p forced umount\n", inode);
2765 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2766 if (need & ~mds_wanted) {
2767 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2768 inode, ceph_cap_string(need),
2769 ceph_cap_string(mds_wanted));
2774 dout("get_cap_refs %p have %s need %s\n", inode,
2775 ceph_cap_string(have), ceph_cap_string(need));
2779 __ceph_touch_fmode(ci, mdsc, flags);
2781 spin_unlock(&ci->i_ceph_lock);
2782 if (snap_rwsem_locked)
2783 up_read(&mdsc->snap_rwsem);
2786 ceph_update_cap_mis(&mdsc->metric);
2788 ceph_update_cap_hit(&mdsc->metric);
2790 dout("get_cap_refs %p ret %d got %s\n", inode,
2791 ret, ceph_cap_string(*got));
2796 * Check the offset we are writing up to against our current
2797 * max_size. If necessary, tell the MDS we want to write to
2800 static void check_max_size(struct inode *inode, loff_t endoff)
2802 struct ceph_inode_info *ci = ceph_inode(inode);
2805 /* do we need to explicitly request a larger max_size? */
2806 spin_lock(&ci->i_ceph_lock);
2807 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2808 dout("write %p at large endoff %llu, req max_size\n",
2810 ci->i_wanted_max_size = endoff;
2812 /* duplicate ceph_check_caps()'s logic */
2813 if (ci->i_auth_cap &&
2814 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2815 ci->i_wanted_max_size > ci->i_max_size &&
2816 ci->i_wanted_max_size > ci->i_requested_max_size)
2818 spin_unlock(&ci->i_ceph_lock);
2820 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2823 static inline int get_used_fmode(int caps)
2826 if (caps & CEPH_CAP_FILE_RD)
2827 fmode |= CEPH_FILE_MODE_RD;
2828 if (caps & CEPH_CAP_FILE_WR)
2829 fmode |= CEPH_FILE_MODE_WR;
2833 int ceph_try_get_caps(struct inode *inode, int need, int want,
2834 bool nonblock, int *got)
2838 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2839 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2840 CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2841 CEPH_CAP_ANY_DIR_OPS));
2843 ret = ceph_pool_perm_check(inode, need);
2848 flags = get_used_fmode(need | want);
2850 flags |= NON_BLOCKING;
2852 ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2853 /* three special error codes */
2854 if (ret == -EAGAIN || ret == -EFBIG || ret == -ESTALE)
2860 * Wait for caps, and take cap references. If we can't get a WR cap
2861 * due to a small max_size, make sure we check_max_size (and possibly
2862 * ask the mds) so we don't get hung up indefinitely.
2864 int ceph_get_caps(struct file *filp, int need, int want,
2865 loff_t endoff, int *got, struct page **pinned_page)
2867 struct ceph_file_info *fi = filp->private_data;
2868 struct inode *inode = file_inode(filp);
2869 struct ceph_inode_info *ci = ceph_inode(inode);
2870 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2871 int ret, _got, flags;
2873 ret = ceph_pool_perm_check(inode, need);
2877 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2878 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2881 flags = get_used_fmode(need | want);
2884 flags &= CEPH_FILE_MODE_MASK;
2885 if (atomic_read(&fi->num_locks))
2886 flags |= CHECK_FILELOCK;
2888 ret = try_get_cap_refs(inode, need, want, endoff,
2890 WARN_ON_ONCE(ret == -EAGAIN);
2892 struct ceph_mds_client *mdsc = fsc->mdsc;
2894 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2896 cw.ino = ceph_ino(inode);
2897 cw.tgid = current->tgid;
2901 spin_lock(&mdsc->caps_list_lock);
2902 list_add(&cw.list, &mdsc->cap_wait_list);
2903 spin_unlock(&mdsc->caps_list_lock);
2905 /* make sure used fmode not timeout */
2906 ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2907 add_wait_queue(&ci->i_cap_wq, &wait);
2909 flags |= NON_BLOCKING;
2910 while (!(ret = try_get_cap_refs(inode, need, want,
2911 endoff, flags, &_got))) {
2912 if (signal_pending(current)) {
2916 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2919 remove_wait_queue(&ci->i_cap_wq, &wait);
2920 ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
2922 spin_lock(&mdsc->caps_list_lock);
2924 spin_unlock(&mdsc->caps_list_lock);
2930 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2931 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2932 if (ret >= 0 && _got)
2933 ceph_put_cap_refs(ci, _got);
2938 if (ret == -EFBIG || ret == -ESTALE) {
2939 int ret2 = ceph_wait_on_async_create(inode);
2943 if (ret == -EFBIG) {
2944 check_max_size(inode, endoff);
2947 if (ret == -ESTALE) {
2948 /* session was killed, try renew caps */
2949 ret = ceph_renew_caps(inode, flags);
2956 if (S_ISREG(ci->vfs_inode.i_mode) &&
2957 ci->i_inline_version != CEPH_INLINE_NONE &&
2958 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2959 i_size_read(inode) > 0) {
2961 find_get_page(inode->i_mapping, 0);
2963 if (PageUptodate(page)) {
2964 *pinned_page = page;
2970 * drop cap refs first because getattr while
2971 * holding * caps refs can cause deadlock.
2973 ceph_put_cap_refs(ci, _got);
2977 * getattr request will bring inline data into
2980 ret = __ceph_do_getattr(inode, NULL,
2981 CEPH_STAT_CAP_INLINE_DATA,
2990 if (S_ISREG(ci->vfs_inode.i_mode) &&
2991 (_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
2992 ceph_fscache_revalidate_cookie(ci);
2999 * Take cap refs. Caller must already know we hold at least one ref
3000 * on the caps in question or we don't know this is safe.
3002 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
3004 spin_lock(&ci->i_ceph_lock);
3005 ceph_take_cap_refs(ci, caps, false);
3006 spin_unlock(&ci->i_ceph_lock);
3011 * drop cap_snap that is not associated with any snapshot.
3012 * we don't need to send FLUSHSNAP message for it.
3014 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
3015 struct ceph_cap_snap *capsnap)
3017 if (!capsnap->need_flush &&
3018 !capsnap->writing && !capsnap->dirty_pages) {
3019 dout("dropping cap_snap %p follows %llu\n",
3020 capsnap, capsnap->follows);
3021 BUG_ON(capsnap->cap_flush.tid > 0);
3022 ceph_put_snap_context(capsnap->context);
3023 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
3024 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3026 list_del(&capsnap->ci_item);
3027 ceph_put_cap_snap(capsnap);
3036 * If we released the last ref on any given cap, call ceph_check_caps
3037 * to release (or schedule a release).
3039 * If we are releasing a WR cap (from a sync write), finalize any affected
3040 * cap_snap, and wake up any waiters.
3042 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
3043 bool skip_checking_caps)
3045 struct inode *inode = &ci->vfs_inode;
3046 int last = 0, put = 0, flushsnaps = 0, wake = 0;
3048 spin_lock(&ci->i_ceph_lock);
3049 if (had & CEPH_CAP_PIN)
3051 if (had & CEPH_CAP_FILE_RD)
3052 if (--ci->i_rd_ref == 0)
3054 if (had & CEPH_CAP_FILE_CACHE)
3055 if (--ci->i_rdcache_ref == 0)
3057 if (had & CEPH_CAP_FILE_EXCL)
3058 if (--ci->i_fx_ref == 0)
3060 if (had & CEPH_CAP_FILE_BUFFER) {
3061 if (--ci->i_wb_ref == 0) {
3065 dout("put_cap_refs %p wb %d -> %d (?)\n",
3066 inode, ci->i_wb_ref+1, ci->i_wb_ref);
3068 if (had & CEPH_CAP_FILE_WR)
3069 if (--ci->i_wr_ref == 0) {
3071 if (__ceph_have_pending_cap_snap(ci)) {
3072 struct ceph_cap_snap *capsnap =
3073 list_last_entry(&ci->i_cap_snaps,
3074 struct ceph_cap_snap,
3076 capsnap->writing = 0;
3077 if (ceph_try_drop_cap_snap(ci, capsnap))
3079 else if (__ceph_finish_cap_snap(ci, capsnap))
3083 if (ci->i_wrbuffer_ref_head == 0 &&
3084 ci->i_dirty_caps == 0 &&
3085 ci->i_flushing_caps == 0) {
3086 BUG_ON(!ci->i_head_snapc);
3087 ceph_put_snap_context(ci->i_head_snapc);
3088 ci->i_head_snapc = NULL;
3090 /* see comment in __ceph_remove_cap() */
3091 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3092 drop_inode_snap_realm(ci);
3094 spin_unlock(&ci->i_ceph_lock);
3096 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3097 last ? " last" : "", put ? " put" : "");
3099 if (!skip_checking_caps) {
3101 ceph_check_caps(ci, 0, NULL);
3102 else if (flushsnaps)
3103 ceph_flush_snaps(ci, NULL);
3106 wake_up_all(&ci->i_cap_wq);
3111 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3113 __ceph_put_cap_refs(ci, had, false);
3116 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3118 __ceph_put_cap_refs(ci, had, true);
3122 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3123 * context. Adjust per-snap dirty page accounting as appropriate.
3124 * Once all dirty data for a cap_snap is flushed, flush snapped file
3125 * metadata back to the MDS. If we dropped the last ref, call
3128 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3129 struct ceph_snap_context *snapc)
3131 struct inode *inode = &ci->vfs_inode;
3132 struct ceph_cap_snap *capsnap = NULL;
3136 bool flush_snaps = false;
3137 bool complete_capsnap = false;
3139 spin_lock(&ci->i_ceph_lock);
3140 ci->i_wrbuffer_ref -= nr;
3141 if (ci->i_wrbuffer_ref == 0) {
3146 if (ci->i_head_snapc == snapc) {
3147 ci->i_wrbuffer_ref_head -= nr;
3148 if (ci->i_wrbuffer_ref_head == 0 &&
3149 ci->i_wr_ref == 0 &&
3150 ci->i_dirty_caps == 0 &&
3151 ci->i_flushing_caps == 0) {
3152 BUG_ON(!ci->i_head_snapc);
3153 ceph_put_snap_context(ci->i_head_snapc);
3154 ci->i_head_snapc = NULL;
3156 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3158 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3159 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3160 last ? " LAST" : "");
3162 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3163 if (capsnap->context == snapc) {
3171 * The capsnap should already be removed when removing
3172 * auth cap in the case of a forced unmount.
3174 WARN_ON_ONCE(ci->i_auth_cap);
3178 capsnap->dirty_pages -= nr;
3179 if (capsnap->dirty_pages == 0) {
3180 complete_capsnap = true;
3181 if (!capsnap->writing) {
3182 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3185 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3190 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3191 " snap %lld %d/%d -> %d/%d %s%s\n",
3192 inode, capsnap, capsnap->context->seq,
3193 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3194 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3195 last ? " (wrbuffer last)" : "",
3196 complete_capsnap ? " (complete capsnap)" : "");
3200 spin_unlock(&ci->i_ceph_lock);
3203 ceph_check_caps(ci, 0, NULL);
3204 } else if (flush_snaps) {
3205 ceph_flush_snaps(ci, NULL);
3207 if (complete_capsnap)
3208 wake_up_all(&ci->i_cap_wq);
3210 /* avoid calling iput_final() in osd dispatch threads */
3211 ceph_async_iput(inode);
3216 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3218 static void invalidate_aliases(struct inode *inode)
3220 struct dentry *dn, *prev = NULL;
3222 dout("invalidate_aliases inode %p\n", inode);
3223 d_prune_aliases(inode);
3225 * For non-directory inode, d_find_alias() only returns
3226 * hashed dentry. After calling d_invalidate(), the
3227 * dentry becomes unhashed.
3229 * For directory inode, d_find_alias() can return
3230 * unhashed dentry. But directory inode should have
3231 * one alias at most.
3233 while ((dn = d_find_alias(inode))) {
3247 struct cap_extra_info {
3248 struct ceph_string *pool_ns;
3258 /* currently issued */
3260 struct timespec64 btime;
3264 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3265 * actually be a revocation if it specifies a smaller cap set.)
3267 * caller holds s_mutex and i_ceph_lock, we drop both.
3269 static void handle_cap_grant(struct inode *inode,
3270 struct ceph_mds_session *session,
3271 struct ceph_cap *cap,
3272 struct ceph_mds_caps *grant,
3273 struct ceph_buffer *xattr_buf,
3274 struct cap_extra_info *extra_info)
3275 __releases(ci->i_ceph_lock)
3276 __releases(session->s_mdsc->snap_rwsem)
3278 struct ceph_inode_info *ci = ceph_inode(inode);
3279 int seq = le32_to_cpu(grant->seq);
3280 int newcaps = le32_to_cpu(grant->caps);
3281 int used, wanted, dirty;
3282 u64 size = le64_to_cpu(grant->size);
3283 u64 max_size = le64_to_cpu(grant->max_size);
3284 unsigned char check_caps = 0;
3285 bool was_stale = cap->cap_gen < session->s_cap_gen;
3287 bool writeback = false;
3288 bool queue_trunc = false;
3289 bool queue_invalidate = false;
3290 bool deleted_inode = false;
3291 bool fill_inline = false;
3293 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3294 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3295 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3300 * If CACHE is being revoked, and we have no dirty buffers,
3301 * try to invalidate (once). (If there are dirty buffers, we
3302 * will invalidate _after_ writeback.)
3304 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3305 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3306 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3307 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3308 if (try_nonblocking_invalidate(inode)) {
3309 /* there were locked pages.. invalidate later
3310 in a separate thread. */
3311 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3312 queue_invalidate = true;
3313 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3319 cap->issued = cap->implemented = CEPH_CAP_PIN;
3322 * auth mds of the inode changed. we received the cap export message,
3323 * but still haven't received the cap import message. handle_cap_export
3324 * updated the new auth MDS' cap.
3326 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3327 * that was sent before the cap import message. So don't remove caps.
3329 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3330 WARN_ON(cap != ci->i_auth_cap);
3331 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3333 newcaps |= cap->issued;
3336 /* side effects now are allowed */
3337 cap->cap_gen = session->s_cap_gen;
3340 __check_cap_issue(ci, cap, newcaps);
3342 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3344 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3345 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3346 inode->i_mode = le32_to_cpu(grant->mode);
3347 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3348 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3349 ci->i_btime = extra_info->btime;
3350 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3351 from_kuid(&init_user_ns, inode->i_uid),
3352 from_kgid(&init_user_ns, inode->i_gid));
3355 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3356 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3357 set_nlink(inode, le32_to_cpu(grant->nlink));
3358 if (inode->i_nlink == 0 &&
3359 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3360 deleted_inode = true;
3363 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3365 int len = le32_to_cpu(grant->xattr_len);
3366 u64 version = le64_to_cpu(grant->xattr_version);
3368 if (version > ci->i_xattrs.version) {
3369 dout(" got new xattrs v%llu on %p len %d\n",
3370 version, inode, len);
3371 if (ci->i_xattrs.blob)
3372 ceph_buffer_put(ci->i_xattrs.blob);
3373 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3374 ci->i_xattrs.version = version;
3375 ceph_forget_all_cached_acls(inode);
3376 ceph_security_invalidate_secctx(inode);
3380 if (newcaps & CEPH_CAP_ANY_RD) {
3381 struct timespec64 mtime, atime, ctime;
3382 /* ctime/mtime/atime? */
3383 ceph_decode_timespec64(&mtime, &grant->mtime);
3384 ceph_decode_timespec64(&atime, &grant->atime);
3385 ceph_decode_timespec64(&ctime, &grant->ctime);
3386 ceph_fill_file_time(inode, extra_info->issued,
3387 le32_to_cpu(grant->time_warp_seq),
3388 &ctime, &mtime, &atime);
3391 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3392 ci->i_files = extra_info->nfiles;
3393 ci->i_subdirs = extra_info->nsubdirs;
3396 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3397 /* file layout may have changed */
3398 s64 old_pool = ci->i_layout.pool_id;
3399 struct ceph_string *old_ns;
3401 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3402 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3403 lockdep_is_held(&ci->i_ceph_lock));
3404 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3406 if (ci->i_layout.pool_id != old_pool ||
3407 extra_info->pool_ns != old_ns)
3408 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3410 extra_info->pool_ns = old_ns;
3412 /* size/truncate_seq? */
3413 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3414 le32_to_cpu(grant->truncate_seq),
3415 le64_to_cpu(grant->truncate_size),
3419 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3420 if (max_size != ci->i_max_size) {
3421 dout("max_size %lld -> %llu\n",
3422 ci->i_max_size, max_size);
3423 ci->i_max_size = max_size;
3424 if (max_size >= ci->i_wanted_max_size) {
3425 ci->i_wanted_max_size = 0; /* reset */
3426 ci->i_requested_max_size = 0;
3432 /* check cap bits */
3433 wanted = __ceph_caps_wanted(ci);
3434 used = __ceph_caps_used(ci);
3435 dirty = __ceph_caps_dirty(ci);
3436 dout(" my wanted = %s, used = %s, dirty %s\n",
3437 ceph_cap_string(wanted),
3438 ceph_cap_string(used),
3439 ceph_cap_string(dirty));
3441 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3442 (wanted & ~(cap->mds_wanted | newcaps))) {
3444 * If mds is importing cap, prior cap messages that update
3445 * 'wanted' may get dropped by mds (migrate seq mismatch).
3447 * We don't send cap message to update 'wanted' if what we
3448 * want are already issued. If mds revokes caps, cap message
3449 * that releases caps also tells mds what we want. But if
3450 * caps got revoked by mds forcedly (session stale). We may
3451 * haven't told mds what we want.
3456 /* revocation, grant, or no-op? */
3457 if (cap->issued & ~newcaps) {
3458 int revoking = cap->issued & ~newcaps;
3460 dout("revocation: %s -> %s (revoking %s)\n",
3461 ceph_cap_string(cap->issued),
3462 ceph_cap_string(newcaps),
3463 ceph_cap_string(revoking));
3464 if (S_ISREG(inode->i_mode) &&
3465 (revoking & used & CEPH_CAP_FILE_BUFFER))
3466 writeback = true; /* initiate writeback; will delay ack */
3467 else if (queue_invalidate &&
3468 revoking == CEPH_CAP_FILE_CACHE &&
3469 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3470 ; /* do nothing yet, invalidation will be queued */
3471 else if (cap == ci->i_auth_cap)
3472 check_caps = 1; /* check auth cap only */
3474 check_caps = 2; /* check all caps */
3475 cap->issued = newcaps;
3476 cap->implemented |= newcaps;
3477 } else if (cap->issued == newcaps) {
3478 dout("caps unchanged: %s -> %s\n",
3479 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3481 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3482 ceph_cap_string(newcaps));
3483 /* non-auth MDS is revoking the newly grant caps ? */
3484 if (cap == ci->i_auth_cap &&
3485 __ceph_caps_revoking_other(ci, cap, newcaps))
3488 cap->issued = newcaps;
3489 cap->implemented |= newcaps; /* add bits only, to
3490 * avoid stepping on a
3491 * pending revocation */
3494 BUG_ON(cap->issued & ~cap->implemented);
3496 if (extra_info->inline_version > 0 &&
3497 extra_info->inline_version >= ci->i_inline_version) {
3498 ci->i_inline_version = extra_info->inline_version;
3499 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3500 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3504 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3505 if (ci->i_auth_cap == cap) {
3506 if (newcaps & ~extra_info->issued)
3509 if (ci->i_requested_max_size > max_size ||
3510 !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3511 /* re-request max_size if necessary */
3512 ci->i_requested_max_size = 0;
3516 ceph_kick_flushing_inode_caps(session, ci);
3518 up_read(&session->s_mdsc->snap_rwsem);
3520 spin_unlock(&ci->i_ceph_lock);
3523 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3524 extra_info->inline_len);
3527 ceph_queue_vmtruncate(inode);
3531 * queue inode for writeback: we can't actually call
3532 * filemap_write_and_wait, etc. from message handler
3535 ceph_queue_writeback(inode);
3536 if (queue_invalidate)
3537 ceph_queue_invalidate(inode);
3539 invalidate_aliases(inode);
3541 wake_up_all(&ci->i_cap_wq);
3543 if (check_caps == 1)
3544 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3546 else if (check_caps == 2)
3547 ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3549 mutex_unlock(&session->s_mutex);
3553 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3554 * MDS has been safely committed.
3556 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3557 struct ceph_mds_caps *m,
3558 struct ceph_mds_session *session,
3559 struct ceph_cap *cap)
3560 __releases(ci->i_ceph_lock)
3562 struct ceph_inode_info *ci = ceph_inode(inode);
3563 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3564 struct ceph_cap_flush *cf, *tmp_cf;
3565 LIST_HEAD(to_remove);
3566 unsigned seq = le32_to_cpu(m->seq);
3567 int dirty = le32_to_cpu(m->dirty);
3570 bool wake_ci = false;
3571 bool wake_mdsc = false;
3573 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3574 /* Is this the one that was flushed? */
3575 if (cf->tid == flush_tid)
3578 /* Is this a capsnap? */
3582 if (cf->tid <= flush_tid) {
3584 * An earlier or current tid. The FLUSH_ACK should
3585 * represent a superset of this flush's caps.
3587 wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3588 list_add_tail(&cf->i_list, &to_remove);
3591 * This is a later one. Any caps in it are still dirty
3592 * so don't count them as cleaned.
3594 cleaned &= ~cf->caps;
3600 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3601 " flushing %s -> %s\n",
3602 inode, session->s_mds, seq, ceph_cap_string(dirty),
3603 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3604 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3606 if (list_empty(&to_remove) && !cleaned)
3609 ci->i_flushing_caps &= ~cleaned;
3611 spin_lock(&mdsc->cap_dirty_lock);
3613 list_for_each_entry(cf, &to_remove, i_list)
3614 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3616 if (ci->i_flushing_caps == 0) {
3617 if (list_empty(&ci->i_cap_flush_list)) {
3618 list_del_init(&ci->i_flushing_item);
3619 if (!list_empty(&session->s_cap_flushing)) {
3620 dout(" mds%d still flushing cap on %p\n",
3622 &list_first_entry(&session->s_cap_flushing,
3623 struct ceph_inode_info,
3624 i_flushing_item)->vfs_inode);
3627 mdsc->num_cap_flushing--;
3628 dout(" inode %p now !flushing\n", inode);
3630 if (ci->i_dirty_caps == 0) {
3631 dout(" inode %p now clean\n", inode);
3632 BUG_ON(!list_empty(&ci->i_dirty_item));
3634 if (ci->i_wr_ref == 0 &&
3635 ci->i_wrbuffer_ref_head == 0) {
3636 BUG_ON(!ci->i_head_snapc);
3637 ceph_put_snap_context(ci->i_head_snapc);
3638 ci->i_head_snapc = NULL;
3641 BUG_ON(list_empty(&ci->i_dirty_item));
3644 spin_unlock(&mdsc->cap_dirty_lock);
3647 spin_unlock(&ci->i_ceph_lock);
3649 while (!list_empty(&to_remove)) {
3650 cf = list_first_entry(&to_remove,
3651 struct ceph_cap_flush, i_list);
3652 list_del_init(&cf->i_list);
3653 if (!cf->is_capsnap)
3654 ceph_free_cap_flush(cf);
3658 wake_up_all(&ci->i_cap_wq);
3660 wake_up_all(&mdsc->cap_flushing_wq);
3665 void __ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3666 bool *wake_ci, bool *wake_mdsc)
3668 struct ceph_inode_info *ci = ceph_inode(inode);
3669 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3672 lockdep_assert_held(&ci->i_ceph_lock);
3674 dout("removing capsnap %p, inode %p ci %p\n", capsnap, inode, ci);
3676 list_del_init(&capsnap->ci_item);
3677 ret = __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3681 spin_lock(&mdsc->cap_dirty_lock);
3682 if (list_empty(&ci->i_cap_flush_list))
3683 list_del_init(&ci->i_flushing_item);
3685 ret = __detach_cap_flush_from_mdsc(mdsc, &capsnap->cap_flush);
3688 spin_unlock(&mdsc->cap_dirty_lock);
3691 void ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3692 bool *wake_ci, bool *wake_mdsc)
3694 struct ceph_inode_info *ci = ceph_inode(inode);
3696 lockdep_assert_held(&ci->i_ceph_lock);
3698 WARN_ON_ONCE(capsnap->dirty_pages || capsnap->writing);
3699 __ceph_remove_capsnap(inode, capsnap, wake_ci, wake_mdsc);
3703 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3704 * throw away our cap_snap.
3706 * Caller hold s_mutex.
3708 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3709 struct ceph_mds_caps *m,
3710 struct ceph_mds_session *session)
3712 struct ceph_inode_info *ci = ceph_inode(inode);
3713 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3714 u64 follows = le64_to_cpu(m->snap_follows);
3715 struct ceph_cap_snap *capsnap;
3716 bool flushed = false;
3717 bool wake_ci = false;
3718 bool wake_mdsc = false;
3720 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3721 inode, ci, session->s_mds, follows);
3723 spin_lock(&ci->i_ceph_lock);
3724 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3725 if (capsnap->follows == follows) {
3726 if (capsnap->cap_flush.tid != flush_tid) {
3727 dout(" cap_snap %p follows %lld tid %lld !="
3728 " %lld\n", capsnap, follows,
3729 flush_tid, capsnap->cap_flush.tid);
3735 dout(" skipping cap_snap %p follows %lld\n",
3736 capsnap, capsnap->follows);
3740 ceph_remove_capsnap(inode, capsnap, &wake_ci, &wake_mdsc);
3741 spin_unlock(&ci->i_ceph_lock);
3744 ceph_put_snap_context(capsnap->context);
3745 ceph_put_cap_snap(capsnap);
3747 wake_up_all(&ci->i_cap_wq);
3749 wake_up_all(&mdsc->cap_flushing_wq);
3755 * Handle TRUNC from MDS, indicating file truncation.
3757 * caller hold s_mutex.
3759 static bool handle_cap_trunc(struct inode *inode,
3760 struct ceph_mds_caps *trunc,
3761 struct ceph_mds_session *session)
3763 struct ceph_inode_info *ci = ceph_inode(inode);
3764 int mds = session->s_mds;
3765 int seq = le32_to_cpu(trunc->seq);
3766 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3767 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3768 u64 size = le64_to_cpu(trunc->size);
3769 int implemented = 0;
3770 int dirty = __ceph_caps_dirty(ci);
3771 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3772 bool queue_trunc = false;
3774 lockdep_assert_held(&ci->i_ceph_lock);
3776 issued |= implemented | dirty;
3778 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3779 inode, mds, seq, truncate_size, truncate_seq);
3780 queue_trunc = ceph_fill_file_size(inode, issued,
3781 truncate_seq, truncate_size, size);
3786 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3787 * different one. If we are the most recent migration we've seen (as
3788 * indicated by mseq), make note of the migrating cap bits for the
3789 * duration (until we see the corresponding IMPORT).
3791 * caller holds s_mutex
3793 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3794 struct ceph_mds_cap_peer *ph,
3795 struct ceph_mds_session *session)
3797 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3798 struct ceph_mds_session *tsession = NULL;
3799 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3800 struct ceph_inode_info *ci = ceph_inode(inode);
3802 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3803 unsigned t_seq, t_mseq;
3805 int mds = session->s_mds;
3808 t_cap_id = le64_to_cpu(ph->cap_id);
3809 t_seq = le32_to_cpu(ph->seq);
3810 t_mseq = le32_to_cpu(ph->mseq);
3811 target = le32_to_cpu(ph->mds);
3813 t_cap_id = t_seq = t_mseq = 0;
3817 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3818 inode, ci, mds, mseq, target);
3820 spin_lock(&ci->i_ceph_lock);
3821 cap = __get_cap_for_mds(ci, mds);
3822 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3826 __ceph_remove_cap(cap, false);
3831 * now we know we haven't received the cap import message yet
3832 * because the exported cap still exist.
3835 issued = cap->issued;
3836 if (issued != cap->implemented)
3837 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3838 "ino (%llx.%llx) mds%d seq %d mseq %d "
3839 "issued %s implemented %s\n",
3840 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3841 ceph_cap_string(issued),
3842 ceph_cap_string(cap->implemented));
3845 tcap = __get_cap_for_mds(ci, target);
3847 /* already have caps from the target */
3848 if (tcap->cap_id == t_cap_id &&
3849 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3850 dout(" updating import cap %p mds%d\n", tcap, target);
3851 tcap->cap_id = t_cap_id;
3852 tcap->seq = t_seq - 1;
3853 tcap->issue_seq = t_seq - 1;
3854 tcap->issued |= issued;
3855 tcap->implemented |= issued;
3856 if (cap == ci->i_auth_cap) {
3857 ci->i_auth_cap = tcap;
3858 change_auth_cap_ses(ci, tcap->session);
3861 __ceph_remove_cap(cap, false);
3863 } else if (tsession) {
3864 /* add placeholder for the export tagert */
3865 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3867 ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3868 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3870 if (!list_empty(&ci->i_cap_flush_list) &&
3871 ci->i_auth_cap == tcap) {
3872 spin_lock(&mdsc->cap_dirty_lock);
3873 list_move_tail(&ci->i_flushing_item,
3874 &tcap->session->s_cap_flushing);
3875 spin_unlock(&mdsc->cap_dirty_lock);
3878 __ceph_remove_cap(cap, false);
3882 spin_unlock(&ci->i_ceph_lock);
3883 mutex_unlock(&session->s_mutex);
3885 /* open target session */
3886 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3887 if (!IS_ERR(tsession)) {
3889 mutex_lock(&session->s_mutex);
3890 mutex_lock_nested(&tsession->s_mutex,
3891 SINGLE_DEPTH_NESTING);
3893 mutex_lock(&tsession->s_mutex);
3894 mutex_lock_nested(&session->s_mutex,
3895 SINGLE_DEPTH_NESTING);
3897 new_cap = ceph_get_cap(mdsc, NULL);
3902 mutex_lock(&session->s_mutex);
3907 spin_unlock(&ci->i_ceph_lock);
3908 mutex_unlock(&session->s_mutex);
3910 mutex_unlock(&tsession->s_mutex);
3911 ceph_put_mds_session(tsession);
3914 ceph_put_cap(mdsc, new_cap);
3918 * Handle cap IMPORT.
3920 * caller holds s_mutex. acquires i_ceph_lock
3922 static void handle_cap_import(struct ceph_mds_client *mdsc,
3923 struct inode *inode, struct ceph_mds_caps *im,
3924 struct ceph_mds_cap_peer *ph,
3925 struct ceph_mds_session *session,
3926 struct ceph_cap **target_cap, int *old_issued)
3928 struct ceph_inode_info *ci = ceph_inode(inode);
3929 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3930 int mds = session->s_mds;
3932 unsigned caps = le32_to_cpu(im->caps);
3933 unsigned wanted = le32_to_cpu(im->wanted);
3934 unsigned seq = le32_to_cpu(im->seq);
3935 unsigned mseq = le32_to_cpu(im->migrate_seq);
3936 u64 realmino = le64_to_cpu(im->realm);
3937 u64 cap_id = le64_to_cpu(im->cap_id);
3942 p_cap_id = le64_to_cpu(ph->cap_id);
3943 peer = le32_to_cpu(ph->mds);
3949 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3950 inode, ci, mds, mseq, peer);
3952 cap = __get_cap_for_mds(ci, mds);
3955 spin_unlock(&ci->i_ceph_lock);
3956 new_cap = ceph_get_cap(mdsc, NULL);
3957 spin_lock(&ci->i_ceph_lock);
3963 ceph_put_cap(mdsc, new_cap);
3968 __ceph_caps_issued(ci, &issued);
3969 issued |= __ceph_caps_dirty(ci);
3971 ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
3972 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3974 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3975 if (ocap && ocap->cap_id == p_cap_id) {
3976 dout(" remove export cap %p mds%d flags %d\n",
3977 ocap, peer, ph->flags);
3978 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3979 (ocap->seq != le32_to_cpu(ph->seq) ||
3980 ocap->mseq != le32_to_cpu(ph->mseq))) {
3981 pr_err_ratelimited("handle_cap_import: "
3982 "mismatched seq/mseq: ino (%llx.%llx) "
3983 "mds%d seq %d mseq %d importer mds%d "
3984 "has peer seq %d mseq %d\n",
3985 ceph_vinop(inode), peer, ocap->seq,
3986 ocap->mseq, mds, le32_to_cpu(ph->seq),
3987 le32_to_cpu(ph->mseq));
3989 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3992 *old_issued = issued;
3997 * Handle a caps message from the MDS.
3999 * Identify the appropriate session, inode, and call the right handler
4000 * based on the cap op.
4002 void ceph_handle_caps(struct ceph_mds_session *session,
4003 struct ceph_msg *msg)
4005 struct ceph_mds_client *mdsc = session->s_mdsc;
4006 struct inode *inode;
4007 struct ceph_inode_info *ci;
4008 struct ceph_cap *cap;
4009 struct ceph_mds_caps *h;
4010 struct ceph_mds_cap_peer *peer = NULL;
4011 struct ceph_snap_realm *realm = NULL;
4013 int msg_version = le16_to_cpu(msg->hdr.version);
4015 struct ceph_vino vino;
4017 size_t snaptrace_len;
4019 struct cap_extra_info extra_info = {};
4022 dout("handle_caps from mds%d\n", session->s_mds);
4025 end = msg->front.iov_base + msg->front.iov_len;
4026 if (msg->front.iov_len < sizeof(*h))
4028 h = msg->front.iov_base;
4029 op = le32_to_cpu(h->op);
4030 vino.ino = le64_to_cpu(h->ino);
4031 vino.snap = CEPH_NOSNAP;
4032 seq = le32_to_cpu(h->seq);
4033 mseq = le32_to_cpu(h->migrate_seq);
4036 snaptrace_len = le32_to_cpu(h->snap_trace_len);
4037 p = snaptrace + snaptrace_len;
4039 if (msg_version >= 2) {
4041 ceph_decode_32_safe(&p, end, flock_len, bad);
4042 if (p + flock_len > end)
4047 if (msg_version >= 3) {
4048 if (op == CEPH_CAP_OP_IMPORT) {
4049 if (p + sizeof(*peer) > end)
4053 } else if (op == CEPH_CAP_OP_EXPORT) {
4054 /* recorded in unused fields */
4055 peer = (void *)&h->size;
4059 if (msg_version >= 4) {
4060 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
4061 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
4062 if (p + extra_info.inline_len > end)
4064 extra_info.inline_data = p;
4065 p += extra_info.inline_len;
4068 if (msg_version >= 5) {
4069 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
4072 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
4073 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
4076 if (msg_version >= 8) {
4078 u32 caller_uid, caller_gid;
4082 ceph_decode_64_safe(&p, end, flush_tid, bad);
4084 ceph_decode_32_safe(&p, end, caller_uid, bad);
4085 ceph_decode_32_safe(&p, end, caller_gid, bad);
4087 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4088 if (pool_ns_len > 0) {
4089 ceph_decode_need(&p, end, pool_ns_len, bad);
4090 extra_info.pool_ns =
4091 ceph_find_or_create_string(p, pool_ns_len);
4096 if (msg_version >= 9) {
4097 struct ceph_timespec *btime;
4099 if (p + sizeof(*btime) > end)
4102 ceph_decode_timespec64(&extra_info.btime, btime);
4103 p += sizeof(*btime);
4104 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4107 if (msg_version >= 11) {
4110 ceph_decode_32_safe(&p, end, flags, bad);
4112 extra_info.dirstat_valid = true;
4113 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4114 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4118 inode = ceph_find_inode(mdsc->fsc->sb, vino);
4119 ci = ceph_inode(inode);
4120 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4123 mutex_lock(&session->s_mutex);
4124 inc_session_sequence(session);
4125 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4129 dout(" i don't have ino %llx\n", vino.ino);
4131 if (op == CEPH_CAP_OP_IMPORT) {
4132 cap = ceph_get_cap(mdsc, NULL);
4133 cap->cap_ino = vino.ino;
4134 cap->queue_release = 1;
4135 cap->cap_id = le64_to_cpu(h->cap_id);
4138 cap->issue_seq = seq;
4139 spin_lock(&session->s_cap_lock);
4140 __ceph_queue_cap_release(session, cap);
4141 spin_unlock(&session->s_cap_lock);
4143 goto flush_cap_releases;
4146 /* these will work even if we don't have a cap yet */
4148 case CEPH_CAP_OP_FLUSHSNAP_ACK:
4149 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4153 case CEPH_CAP_OP_EXPORT:
4154 handle_cap_export(inode, h, peer, session);
4157 case CEPH_CAP_OP_IMPORT:
4159 if (snaptrace_len) {
4160 down_write(&mdsc->snap_rwsem);
4161 ceph_update_snap_trace(mdsc, snaptrace,
4162 snaptrace + snaptrace_len,
4164 downgrade_write(&mdsc->snap_rwsem);
4166 down_read(&mdsc->snap_rwsem);
4168 spin_lock(&ci->i_ceph_lock);
4169 handle_cap_import(mdsc, inode, h, peer, session,
4170 &cap, &extra_info.issued);
4171 handle_cap_grant(inode, session, cap,
4172 h, msg->middle, &extra_info);
4174 ceph_put_snap_realm(mdsc, realm);
4178 /* the rest require a cap */
4179 spin_lock(&ci->i_ceph_lock);
4180 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4182 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4183 inode, ceph_ino(inode), ceph_snap(inode),
4185 spin_unlock(&ci->i_ceph_lock);
4186 goto flush_cap_releases;
4189 /* note that each of these drops i_ceph_lock for us */
4191 case CEPH_CAP_OP_REVOKE:
4192 case CEPH_CAP_OP_GRANT:
4193 __ceph_caps_issued(ci, &extra_info.issued);
4194 extra_info.issued |= __ceph_caps_dirty(ci);
4195 handle_cap_grant(inode, session, cap,
4196 h, msg->middle, &extra_info);
4199 case CEPH_CAP_OP_FLUSH_ACK:
4200 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4204 case CEPH_CAP_OP_TRUNC:
4205 queue_trunc = handle_cap_trunc(inode, h, session);
4206 spin_unlock(&ci->i_ceph_lock);
4208 ceph_queue_vmtruncate(inode);
4212 spin_unlock(&ci->i_ceph_lock);
4213 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4214 ceph_cap_op_name(op));
4218 mutex_unlock(&session->s_mutex);
4220 ceph_put_string(extra_info.pool_ns);
4221 /* avoid calling iput_final() in mds dispatch threads */
4222 ceph_async_iput(inode);
4227 * send any cap release message to try to move things
4228 * along for the mds (who clearly thinks we still have this
4231 ceph_flush_cap_releases(mdsc, session);
4235 pr_err("ceph_handle_caps: corrupt message\n");
4241 * Delayed work handler to process end of delayed cap release LRU list.
4243 * If new caps are added to the list while processing it, these won't get
4244 * processed in this run. In this case, the ci->i_hold_caps_max will be
4245 * returned so that the work can be scheduled accordingly.
4247 unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4249 struct inode *inode;
4250 struct ceph_inode_info *ci;
4251 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
4252 unsigned long delay_max = opt->caps_wanted_delay_max * HZ;
4253 unsigned long loop_start = jiffies;
4254 unsigned long delay = 0;
4256 dout("check_delayed_caps\n");
4257 spin_lock(&mdsc->cap_delay_lock);
4258 while (!list_empty(&mdsc->cap_delay_list)) {
4259 ci = list_first_entry(&mdsc->cap_delay_list,
4260 struct ceph_inode_info,
4262 if (time_before(loop_start, ci->i_hold_caps_max - delay_max)) {
4263 dout("%s caps added recently. Exiting loop", __func__);
4264 delay = ci->i_hold_caps_max;
4267 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4268 time_before(jiffies, ci->i_hold_caps_max))
4270 list_del_init(&ci->i_cap_delay_list);
4272 inode = igrab(&ci->vfs_inode);
4274 spin_unlock(&mdsc->cap_delay_lock);
4275 dout("check_delayed_caps on %p\n", inode);
4276 ceph_check_caps(ci, 0, NULL);
4277 /* avoid calling iput_final() in tick thread */
4278 ceph_async_iput(inode);
4279 spin_lock(&mdsc->cap_delay_lock);
4282 spin_unlock(&mdsc->cap_delay_lock);
4288 * Flush all dirty caps to the mds
4290 static void flush_dirty_session_caps(struct ceph_mds_session *s)
4292 struct ceph_mds_client *mdsc = s->s_mdsc;
4293 struct ceph_inode_info *ci;
4294 struct inode *inode;
4296 dout("flush_dirty_caps\n");
4297 spin_lock(&mdsc->cap_dirty_lock);
4298 while (!list_empty(&s->s_cap_dirty)) {
4299 ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4301 inode = &ci->vfs_inode;
4303 dout("flush_dirty_caps %p\n", inode);
4304 spin_unlock(&mdsc->cap_dirty_lock);
4305 ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4307 spin_lock(&mdsc->cap_dirty_lock);
4309 spin_unlock(&mdsc->cap_dirty_lock);
4310 dout("flush_dirty_caps done\n");
4313 static void iterate_sessions(struct ceph_mds_client *mdsc,
4314 void (*cb)(struct ceph_mds_session *))
4318 mutex_lock(&mdsc->mutex);
4319 for (mds = 0; mds < mdsc->max_sessions; ++mds) {
4320 struct ceph_mds_session *s;
4322 if (!mdsc->sessions[mds])
4325 s = ceph_get_mds_session(mdsc->sessions[mds]);
4329 mutex_unlock(&mdsc->mutex);
4331 ceph_put_mds_session(s);
4332 mutex_lock(&mdsc->mutex);
4334 mutex_unlock(&mdsc->mutex);
4337 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4339 iterate_sessions(mdsc, flush_dirty_session_caps);
4342 void __ceph_touch_fmode(struct ceph_inode_info *ci,
4343 struct ceph_mds_client *mdsc, int fmode)
4345 unsigned long now = jiffies;
4346 if (fmode & CEPH_FILE_MODE_RD)
4347 ci->i_last_rd = now;
4348 if (fmode & CEPH_FILE_MODE_WR)
4349 ci->i_last_wr = now;
4350 /* queue periodic check */
4352 __ceph_is_any_real_caps(ci) &&
4353 list_empty(&ci->i_cap_delay_list))
4354 __cap_delay_requeue(mdsc, ci);
4357 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4359 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4360 int bits = (fmode << 1) | 1;
4361 bool already_opened = false;
4365 atomic64_inc(&mdsc->metric.opened_files);
4367 spin_lock(&ci->i_ceph_lock);
4368 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4370 * If any of the mode ref is larger than 0,
4371 * that means it has been already opened by
4372 * others. Just skip checking the PIN ref.
4374 if (i && ci->i_nr_by_mode[i])
4375 already_opened = true;
4377 if (bits & (1 << i))
4378 ci->i_nr_by_mode[i] += count;
4381 if (!already_opened)
4382 percpu_counter_inc(&mdsc->metric.opened_inodes);
4383 spin_unlock(&ci->i_ceph_lock);
4387 * Drop open file reference. If we were the last open file,
4388 * we may need to release capabilities to the MDS (or schedule
4389 * their delayed release).
4391 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4393 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4394 int bits = (fmode << 1) | 1;
4395 bool is_closed = true;
4399 atomic64_dec(&mdsc->metric.opened_files);
4401 spin_lock(&ci->i_ceph_lock);
4402 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4403 if (bits & (1 << i)) {
4404 BUG_ON(ci->i_nr_by_mode[i] < count);
4405 ci->i_nr_by_mode[i] -= count;
4409 * If any of the mode ref is not 0 after
4410 * decreased, that means it is still opened
4411 * by others. Just skip checking the PIN ref.
4413 if (i && ci->i_nr_by_mode[i])
4418 percpu_counter_dec(&mdsc->metric.opened_inodes);
4419 spin_unlock(&ci->i_ceph_lock);
4423 * For a soon-to-be unlinked file, drop the LINK caps. If it
4424 * looks like the link count will hit 0, drop any other caps (other
4425 * than PIN) we don't specifically want (due to the file still being
4428 int ceph_drop_caps_for_unlink(struct inode *inode)
4430 struct ceph_inode_info *ci = ceph_inode(inode);
4431 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4433 spin_lock(&ci->i_ceph_lock);
4434 if (inode->i_nlink == 1) {
4435 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4437 if (__ceph_caps_dirty(ci)) {
4438 struct ceph_mds_client *mdsc =
4439 ceph_inode_to_client(inode)->mdsc;
4440 __cap_delay_requeue_front(mdsc, ci);
4443 spin_unlock(&ci->i_ceph_lock);
4448 * Helpers for embedding cap and dentry lease releases into mds
4451 * @force is used by dentry_release (below) to force inclusion of a
4452 * record for the directory inode, even when there aren't any caps to
4455 int ceph_encode_inode_release(void **p, struct inode *inode,
4456 int mds, int drop, int unless, int force)
4458 struct ceph_inode_info *ci = ceph_inode(inode);
4459 struct ceph_cap *cap;
4460 struct ceph_mds_request_release *rel = *p;
4464 spin_lock(&ci->i_ceph_lock);
4465 used = __ceph_caps_used(ci);
4466 dirty = __ceph_caps_dirty(ci);
4468 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4469 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4470 ceph_cap_string(unless));
4472 /* only drop unused, clean caps */
4473 drop &= ~(used | dirty);
4475 cap = __get_cap_for_mds(ci, mds);
4476 if (cap && __cap_is_valid(cap)) {
4477 unless &= cap->issued;
4479 if (unless & CEPH_CAP_AUTH_EXCL)
4480 drop &= ~CEPH_CAP_AUTH_SHARED;
4481 if (unless & CEPH_CAP_LINK_EXCL)
4482 drop &= ~CEPH_CAP_LINK_SHARED;
4483 if (unless & CEPH_CAP_XATTR_EXCL)
4484 drop &= ~CEPH_CAP_XATTR_SHARED;
4485 if (unless & CEPH_CAP_FILE_EXCL)
4486 drop &= ~CEPH_CAP_FILE_SHARED;
4489 if (force || (cap->issued & drop)) {
4490 if (cap->issued & drop) {
4491 int wanted = __ceph_caps_wanted(ci);
4492 dout("encode_inode_release %p cap %p "
4493 "%s -> %s, wanted %s -> %s\n", inode, cap,
4494 ceph_cap_string(cap->issued),
4495 ceph_cap_string(cap->issued & ~drop),
4496 ceph_cap_string(cap->mds_wanted),
4497 ceph_cap_string(wanted));
4499 cap->issued &= ~drop;
4500 cap->implemented &= ~drop;
4501 cap->mds_wanted = wanted;
4502 if (cap == ci->i_auth_cap &&
4503 !(wanted & CEPH_CAP_ANY_FILE_WR))
4504 ci->i_requested_max_size = 0;
4506 dout("encode_inode_release %p cap %p %s"
4507 " (force)\n", inode, cap,
4508 ceph_cap_string(cap->issued));
4511 rel->ino = cpu_to_le64(ceph_ino(inode));
4512 rel->cap_id = cpu_to_le64(cap->cap_id);
4513 rel->seq = cpu_to_le32(cap->seq);
4514 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4515 rel->mseq = cpu_to_le32(cap->mseq);
4516 rel->caps = cpu_to_le32(cap->implemented);
4517 rel->wanted = cpu_to_le32(cap->mds_wanted);
4523 dout("encode_inode_release %p cap %p %s (noop)\n",
4524 inode, cap, ceph_cap_string(cap->issued));
4527 spin_unlock(&ci->i_ceph_lock);
4531 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4533 int mds, int drop, int unless)
4535 struct dentry *parent = NULL;
4536 struct ceph_mds_request_release *rel = *p;
4537 struct ceph_dentry_info *di = ceph_dentry(dentry);
4542 * force an record for the directory caps if we have a dentry lease.
4543 * this is racy (can't take i_ceph_lock and d_lock together), but it
4544 * doesn't have to be perfect; the mds will revoke anything we don't
4547 spin_lock(&dentry->d_lock);
4548 if (di->lease_session && di->lease_session->s_mds == mds)
4551 parent = dget(dentry->d_parent);
4552 dir = d_inode(parent);
4554 spin_unlock(&dentry->d_lock);
4556 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4559 spin_lock(&dentry->d_lock);
4560 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4561 dout("encode_dentry_release %p mds%d seq %d\n",
4562 dentry, mds, (int)di->lease_seq);
4563 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4564 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4565 *p += dentry->d_name.len;
4566 rel->dname_seq = cpu_to_le32(di->lease_seq);
4567 __ceph_mdsc_drop_dentry_lease(dentry);
4569 spin_unlock(&dentry->d_lock);