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 !mdsc->fsc->blocklisted);
1161 ci->i_auth_cap = NULL;
1164 /* remove from session list */
1165 spin_lock(&session->s_cap_lock);
1166 if (session->s_cap_iterator == cap) {
1167 /* not yet, we are iterating over this very cap */
1168 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1171 list_del_init(&cap->session_caps);
1172 session->s_nr_caps--;
1173 atomic64_dec(&mdsc->metric.total_caps);
1174 cap->session = NULL;
1177 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1181 * s_cap_reconnect is protected by s_cap_lock. no one changes
1182 * s_cap_gen while session is in the reconnect state.
1184 if (queue_release &&
1185 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
1186 cap->queue_release = 1;
1188 __ceph_queue_cap_release(session, cap);
1192 cap->queue_release = 0;
1194 cap->cap_ino = ci->i_vino.ino;
1196 spin_unlock(&session->s_cap_lock);
1199 ceph_put_cap(mdsc, cap);
1201 if (!__ceph_is_any_real_caps(ci)) {
1202 /* when reconnect denied, we remove session caps forcibly,
1203 * i_wr_ref can be non-zero. If there are ongoing write,
1204 * keep i_snap_realm.
1206 if (ci->i_wr_ref == 0 && ci->i_snap_realm)
1207 drop_inode_snap_realm(ci);
1209 __cap_delay_cancel(mdsc, ci);
1213 struct cap_msg_args {
1214 struct ceph_mds_session *session;
1215 u64 ino, cid, follows;
1216 u64 flush_tid, oldest_flush_tid, size, max_size;
1219 struct ceph_buffer *xattr_buf;
1220 struct ceph_buffer *old_xattr_buf;
1221 struct timespec64 atime, mtime, ctime, btime;
1222 int op, caps, wanted, dirty;
1223 u32 seq, issue_seq, mseq, time_warp_seq;
1233 * cap struct size + flock buffer size + inline version + inline data size +
1234 * osd_epoch_barrier + oldest_flush_tid
1236 #define CAP_MSG_SIZE (sizeof(struct ceph_mds_caps) + \
1237 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4)
1239 /* Marshal up the cap msg to the MDS */
1240 static void encode_cap_msg(struct ceph_msg *msg, struct cap_msg_args *arg)
1242 struct ceph_mds_caps *fc;
1244 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1246 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",
1247 __func__, ceph_cap_op_name(arg->op), arg->cid, arg->ino,
1248 ceph_cap_string(arg->caps), ceph_cap_string(arg->wanted),
1249 ceph_cap_string(arg->dirty), arg->seq, arg->issue_seq,
1250 arg->flush_tid, arg->oldest_flush_tid, arg->mseq, arg->follows,
1251 arg->size, arg->max_size, arg->xattr_version,
1252 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1254 msg->hdr.version = cpu_to_le16(10);
1255 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1257 fc = msg->front.iov_base;
1258 memset(fc, 0, sizeof(*fc));
1260 fc->cap_id = cpu_to_le64(arg->cid);
1261 fc->op = cpu_to_le32(arg->op);
1262 fc->seq = cpu_to_le32(arg->seq);
1263 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1264 fc->migrate_seq = cpu_to_le32(arg->mseq);
1265 fc->caps = cpu_to_le32(arg->caps);
1266 fc->wanted = cpu_to_le32(arg->wanted);
1267 fc->dirty = cpu_to_le32(arg->dirty);
1268 fc->ino = cpu_to_le64(arg->ino);
1269 fc->snap_follows = cpu_to_le64(arg->follows);
1271 fc->size = cpu_to_le64(arg->size);
1272 fc->max_size = cpu_to_le64(arg->max_size);
1273 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1274 ceph_encode_timespec64(&fc->atime, &arg->atime);
1275 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1276 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1278 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1279 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1280 fc->mode = cpu_to_le32(arg->mode);
1282 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1283 if (arg->xattr_buf) {
1284 msg->middle = ceph_buffer_get(arg->xattr_buf);
1285 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1286 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1290 /* flock buffer size (version 2) */
1291 ceph_encode_32(&p, 0);
1292 /* inline version (version 4) */
1293 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1294 /* inline data size */
1295 ceph_encode_32(&p, 0);
1297 * osd_epoch_barrier (version 5)
1298 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1299 * case it was recently changed
1301 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1302 /* oldest_flush_tid (version 6) */
1303 ceph_encode_64(&p, arg->oldest_flush_tid);
1306 * caller_uid/caller_gid (version 7)
1308 * Currently, we don't properly track which caller dirtied the caps
1309 * last, and force a flush of them when there is a conflict. For now,
1310 * just set this to 0:0, to emulate how the MDS has worked up to now.
1312 ceph_encode_32(&p, 0);
1313 ceph_encode_32(&p, 0);
1315 /* pool namespace (version 8) (mds always ignores this) */
1316 ceph_encode_32(&p, 0);
1318 /* btime and change_attr (version 9) */
1319 ceph_encode_timespec64(p, &arg->btime);
1320 p += sizeof(struct ceph_timespec);
1321 ceph_encode_64(&p, arg->change_attr);
1323 /* Advisory flags (version 10) */
1324 ceph_encode_32(&p, arg->flags);
1328 * Queue cap releases when an inode is dropped from our cache.
1330 void __ceph_remove_caps(struct ceph_inode_info *ci)
1334 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1335 * may call __ceph_caps_issued_mask() on a freeing inode. */
1336 spin_lock(&ci->i_ceph_lock);
1337 p = rb_first(&ci->i_caps);
1339 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1341 __ceph_remove_cap(cap, true);
1343 spin_unlock(&ci->i_ceph_lock);
1347 * Prepare to send a cap message to an MDS. Update the cap state, and populate
1348 * the arg struct with the parameters that will need to be sent. This should
1349 * be done under the i_ceph_lock to guard against changes to cap state.
1351 * Make note of max_size reported/requested from mds, revoked caps
1352 * that have now been implemented.
1354 static void __prep_cap(struct cap_msg_args *arg, struct ceph_cap *cap,
1355 int op, int flags, int used, int want, int retain,
1356 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1358 struct ceph_inode_info *ci = cap->ci;
1359 struct inode *inode = &ci->vfs_inode;
1362 lockdep_assert_held(&ci->i_ceph_lock);
1364 held = cap->issued | cap->implemented;
1365 revoking = cap->implemented & ~cap->issued;
1366 retain &= ~revoking;
1368 dout("%s %p cap %p session %p %s -> %s (revoking %s)\n",
1369 __func__, inode, cap, cap->session,
1370 ceph_cap_string(held), ceph_cap_string(held & retain),
1371 ceph_cap_string(revoking));
1372 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1374 ci->i_ceph_flags &= ~CEPH_I_FLUSH;
1376 cap->issued &= retain; /* drop bits we don't want */
1378 * Wake up any waiters on wanted -> needed transition. This is due to
1379 * the weird transition from buffered to sync IO... we need to flush
1380 * dirty pages _before_ allowing sync writes to avoid reordering.
1382 arg->wake = cap->implemented & ~cap->issued;
1383 cap->implemented &= cap->issued | used;
1384 cap->mds_wanted = want;
1386 arg->session = cap->session;
1387 arg->ino = ceph_vino(inode).ino;
1388 arg->cid = cap->cap_id;
1389 arg->follows = flushing ? ci->i_head_snapc->seq : 0;
1390 arg->flush_tid = flush_tid;
1391 arg->oldest_flush_tid = oldest_flush_tid;
1393 arg->size = i_size_read(inode);
1394 ci->i_reported_size = arg->size;
1395 arg->max_size = ci->i_wanted_max_size;
1396 if (cap == ci->i_auth_cap) {
1397 if (want & CEPH_CAP_ANY_FILE_WR)
1398 ci->i_requested_max_size = arg->max_size;
1400 ci->i_requested_max_size = 0;
1403 if (flushing & CEPH_CAP_XATTR_EXCL) {
1404 arg->old_xattr_buf = __ceph_build_xattrs_blob(ci);
1405 arg->xattr_version = ci->i_xattrs.version;
1406 arg->xattr_buf = ci->i_xattrs.blob;
1408 arg->xattr_buf = NULL;
1409 arg->old_xattr_buf = NULL;
1412 arg->mtime = inode->i_mtime;
1413 arg->atime = inode->i_atime;
1414 arg->ctime = inode->i_ctime;
1415 arg->btime = ci->i_btime;
1416 arg->change_attr = inode_peek_iversion_raw(inode);
1419 arg->caps = cap->implemented;
1421 arg->dirty = flushing;
1423 arg->seq = cap->seq;
1424 arg->issue_seq = cap->issue_seq;
1425 arg->mseq = cap->mseq;
1426 arg->time_warp_seq = ci->i_time_warp_seq;
1428 arg->uid = inode->i_uid;
1429 arg->gid = inode->i_gid;
1430 arg->mode = inode->i_mode;
1432 arg->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1433 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1434 !list_empty(&ci->i_cap_snaps)) {
1435 struct ceph_cap_snap *capsnap;
1436 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1437 if (capsnap->cap_flush.tid)
1439 if (capsnap->need_flush) {
1440 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1449 * Send a cap msg on the given inode.
1451 * Caller should hold snap_rwsem (read), s_mutex.
1453 static void __send_cap(struct cap_msg_args *arg, struct ceph_inode_info *ci)
1455 struct ceph_msg *msg;
1456 struct inode *inode = &ci->vfs_inode;
1458 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1460 pr_err("error allocating cap msg: ino (%llx.%llx) flushing %s tid %llu, requeuing cap.\n",
1461 ceph_vinop(inode), ceph_cap_string(arg->dirty),
1463 spin_lock(&ci->i_ceph_lock);
1464 __cap_delay_requeue(arg->session->s_mdsc, ci);
1465 spin_unlock(&ci->i_ceph_lock);
1469 encode_cap_msg(msg, arg);
1470 ceph_con_send(&arg->session->s_con, msg);
1471 ceph_buffer_put(arg->old_xattr_buf);
1473 wake_up_all(&ci->i_cap_wq);
1476 static inline int __send_flush_snap(struct inode *inode,
1477 struct ceph_mds_session *session,
1478 struct ceph_cap_snap *capsnap,
1479 u32 mseq, u64 oldest_flush_tid)
1481 struct cap_msg_args arg;
1482 struct ceph_msg *msg;
1484 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1488 arg.session = session;
1489 arg.ino = ceph_vino(inode).ino;
1491 arg.follows = capsnap->follows;
1492 arg.flush_tid = capsnap->cap_flush.tid;
1493 arg.oldest_flush_tid = oldest_flush_tid;
1495 arg.size = capsnap->size;
1497 arg.xattr_version = capsnap->xattr_version;
1498 arg.xattr_buf = capsnap->xattr_blob;
1499 arg.old_xattr_buf = NULL;
1501 arg.atime = capsnap->atime;
1502 arg.mtime = capsnap->mtime;
1503 arg.ctime = capsnap->ctime;
1504 arg.btime = capsnap->btime;
1505 arg.change_attr = capsnap->change_attr;
1507 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1508 arg.caps = capsnap->issued;
1510 arg.dirty = capsnap->dirty;
1515 arg.time_warp_seq = capsnap->time_warp_seq;
1517 arg.uid = capsnap->uid;
1518 arg.gid = capsnap->gid;
1519 arg.mode = capsnap->mode;
1521 arg.inline_data = capsnap->inline_data;
1525 encode_cap_msg(msg, &arg);
1526 ceph_con_send(&arg.session->s_con, msg);
1531 * When a snapshot is taken, clients accumulate dirty metadata on
1532 * inodes with capabilities in ceph_cap_snaps to describe the file
1533 * state at the time the snapshot was taken. This must be flushed
1534 * asynchronously back to the MDS once sync writes complete and dirty
1535 * data is written out.
1537 * Called under i_ceph_lock. Takes s_mutex as needed.
1539 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1540 struct ceph_mds_session *session)
1541 __releases(ci->i_ceph_lock)
1542 __acquires(ci->i_ceph_lock)
1544 struct inode *inode = &ci->vfs_inode;
1545 struct ceph_mds_client *mdsc = session->s_mdsc;
1546 struct ceph_cap_snap *capsnap;
1547 u64 oldest_flush_tid = 0;
1548 u64 first_tid = 1, last_tid = 0;
1550 dout("__flush_snaps %p session %p\n", inode, session);
1552 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1554 * we need to wait for sync writes to complete and for dirty
1555 * pages to be written out.
1557 if (capsnap->dirty_pages || capsnap->writing)
1560 /* should be removed by ceph_try_drop_cap_snap() */
1561 BUG_ON(!capsnap->need_flush);
1563 /* only flush each capsnap once */
1564 if (capsnap->cap_flush.tid > 0) {
1565 dout(" already flushed %p, skipping\n", capsnap);
1569 spin_lock(&mdsc->cap_dirty_lock);
1570 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1571 list_add_tail(&capsnap->cap_flush.g_list,
1572 &mdsc->cap_flush_list);
1573 if (oldest_flush_tid == 0)
1574 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1575 if (list_empty(&ci->i_flushing_item)) {
1576 list_add_tail(&ci->i_flushing_item,
1577 &session->s_cap_flushing);
1579 spin_unlock(&mdsc->cap_dirty_lock);
1581 list_add_tail(&capsnap->cap_flush.i_list,
1582 &ci->i_cap_flush_list);
1585 first_tid = capsnap->cap_flush.tid;
1586 last_tid = capsnap->cap_flush.tid;
1589 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1591 while (first_tid <= last_tid) {
1592 struct ceph_cap *cap = ci->i_auth_cap;
1593 struct ceph_cap_flush *cf;
1596 if (!(cap && cap->session == session)) {
1597 dout("__flush_snaps %p auth cap %p not mds%d, "
1598 "stop\n", inode, cap, session->s_mds);
1603 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1604 if (cf->tid >= first_tid) {
1612 first_tid = cf->tid + 1;
1614 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1615 refcount_inc(&capsnap->nref);
1616 spin_unlock(&ci->i_ceph_lock);
1618 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1619 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1621 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1624 pr_err("__flush_snaps: error sending cap flushsnap, "
1625 "ino (%llx.%llx) tid %llu follows %llu\n",
1626 ceph_vinop(inode), cf->tid, capsnap->follows);
1629 ceph_put_cap_snap(capsnap);
1630 spin_lock(&ci->i_ceph_lock);
1634 void ceph_flush_snaps(struct ceph_inode_info *ci,
1635 struct ceph_mds_session **psession)
1637 struct inode *inode = &ci->vfs_inode;
1638 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1639 struct ceph_mds_session *session = NULL;
1642 dout("ceph_flush_snaps %p\n", inode);
1644 session = *psession;
1646 spin_lock(&ci->i_ceph_lock);
1647 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1648 dout(" no capsnap needs flush, doing nothing\n");
1651 if (!ci->i_auth_cap) {
1652 dout(" no auth cap (migrating?), doing nothing\n");
1656 mds = ci->i_auth_cap->session->s_mds;
1657 if (session && session->s_mds != mds) {
1658 dout(" oops, wrong session %p mutex\n", session);
1659 mutex_unlock(&session->s_mutex);
1660 ceph_put_mds_session(session);
1664 spin_unlock(&ci->i_ceph_lock);
1665 mutex_lock(&mdsc->mutex);
1666 session = __ceph_lookup_mds_session(mdsc, mds);
1667 mutex_unlock(&mdsc->mutex);
1669 dout(" inverting session/ino locks on %p\n", session);
1670 mutex_lock(&session->s_mutex);
1675 // make sure flushsnap messages are sent in proper order.
1676 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1677 __kick_flushing_caps(mdsc, session, ci, 0);
1679 __ceph_flush_snaps(ci, session);
1681 spin_unlock(&ci->i_ceph_lock);
1684 *psession = session;
1685 } else if (session) {
1686 mutex_unlock(&session->s_mutex);
1687 ceph_put_mds_session(session);
1689 /* we flushed them all; remove this inode from the queue */
1690 spin_lock(&mdsc->snap_flush_lock);
1691 list_del_init(&ci->i_snap_flush_item);
1692 spin_unlock(&mdsc->snap_flush_lock);
1696 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1697 * Caller is then responsible for calling __mark_inode_dirty with the
1698 * returned flags value.
1700 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1701 struct ceph_cap_flush **pcf)
1703 struct ceph_mds_client *mdsc =
1704 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1705 struct inode *inode = &ci->vfs_inode;
1706 int was = ci->i_dirty_caps;
1709 lockdep_assert_held(&ci->i_ceph_lock);
1711 if (!ci->i_auth_cap) {
1712 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1713 "but no auth cap (session was closed?)\n",
1714 inode, ceph_ino(inode), ceph_cap_string(mask));
1718 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1719 ceph_cap_string(mask), ceph_cap_string(was),
1720 ceph_cap_string(was | mask));
1721 ci->i_dirty_caps |= mask;
1723 struct ceph_mds_session *session = ci->i_auth_cap->session;
1725 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1726 swap(ci->i_prealloc_cap_flush, *pcf);
1728 if (!ci->i_head_snapc) {
1729 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1730 ci->i_head_snapc = ceph_get_snap_context(
1731 ci->i_snap_realm->cached_context);
1733 dout(" inode %p now dirty snapc %p auth cap %p\n",
1734 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1735 BUG_ON(!list_empty(&ci->i_dirty_item));
1736 spin_lock(&mdsc->cap_dirty_lock);
1737 list_add(&ci->i_dirty_item, &session->s_cap_dirty);
1738 spin_unlock(&mdsc->cap_dirty_lock);
1739 if (ci->i_flushing_caps == 0) {
1741 dirty |= I_DIRTY_SYNC;
1744 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1746 BUG_ON(list_empty(&ci->i_dirty_item));
1747 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1748 (mask & CEPH_CAP_FILE_BUFFER))
1749 dirty |= I_DIRTY_DATASYNC;
1750 __cap_delay_requeue(mdsc, ci);
1754 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1756 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1759 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1762 kmem_cache_free(ceph_cap_flush_cachep, cf);
1765 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1767 if (!list_empty(&mdsc->cap_flush_list)) {
1768 struct ceph_cap_flush *cf =
1769 list_first_entry(&mdsc->cap_flush_list,
1770 struct ceph_cap_flush, g_list);
1777 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1778 * Return true if caller needs to wake up flush waiters.
1780 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1781 struct ceph_cap_flush *cf)
1783 struct ceph_cap_flush *prev;
1784 bool wake = cf->wake;
1786 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1787 prev = list_prev_entry(cf, g_list);
1791 list_del(&cf->g_list);
1795 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1796 struct ceph_cap_flush *cf)
1798 struct ceph_cap_flush *prev;
1799 bool wake = cf->wake;
1801 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1802 prev = list_prev_entry(cf, i_list);
1806 list_del(&cf->i_list);
1811 * Add dirty inode to the flushing list. Assigned a seq number so we
1812 * can wait for caps to flush without starving.
1814 * Called under i_ceph_lock. Returns the flush tid.
1816 static u64 __mark_caps_flushing(struct inode *inode,
1817 struct ceph_mds_session *session, bool wake,
1818 u64 *oldest_flush_tid)
1820 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1821 struct ceph_inode_info *ci = ceph_inode(inode);
1822 struct ceph_cap_flush *cf = NULL;
1825 lockdep_assert_held(&ci->i_ceph_lock);
1826 BUG_ON(ci->i_dirty_caps == 0);
1827 BUG_ON(list_empty(&ci->i_dirty_item));
1828 BUG_ON(!ci->i_prealloc_cap_flush);
1830 flushing = ci->i_dirty_caps;
1831 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1832 ceph_cap_string(flushing),
1833 ceph_cap_string(ci->i_flushing_caps),
1834 ceph_cap_string(ci->i_flushing_caps | flushing));
1835 ci->i_flushing_caps |= flushing;
1836 ci->i_dirty_caps = 0;
1837 dout(" inode %p now !dirty\n", inode);
1839 swap(cf, ci->i_prealloc_cap_flush);
1840 cf->caps = flushing;
1843 spin_lock(&mdsc->cap_dirty_lock);
1844 list_del_init(&ci->i_dirty_item);
1846 cf->tid = ++mdsc->last_cap_flush_tid;
1847 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1848 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1850 if (list_empty(&ci->i_flushing_item)) {
1851 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1852 mdsc->num_cap_flushing++;
1854 spin_unlock(&mdsc->cap_dirty_lock);
1856 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1862 * try to invalidate mapping pages without blocking.
1864 static int try_nonblocking_invalidate(struct inode *inode)
1866 struct ceph_inode_info *ci = ceph_inode(inode);
1867 u32 invalidating_gen = ci->i_rdcache_gen;
1869 spin_unlock(&ci->i_ceph_lock);
1870 ceph_fscache_invalidate(inode);
1871 invalidate_mapping_pages(&inode->i_data, 0, -1);
1872 spin_lock(&ci->i_ceph_lock);
1874 if (inode->i_data.nrpages == 0 &&
1875 invalidating_gen == ci->i_rdcache_gen) {
1877 dout("try_nonblocking_invalidate %p success\n", inode);
1878 /* save any racing async invalidate some trouble */
1879 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1882 dout("try_nonblocking_invalidate %p failed\n", inode);
1886 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1888 loff_t size = i_size_read(&ci->vfs_inode);
1889 /* mds will adjust max size according to the reported size */
1890 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1892 if (size >= ci->i_max_size)
1894 /* half of previous max_size increment has been used */
1895 if (ci->i_max_size > ci->i_reported_size &&
1896 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1902 * Swiss army knife function to examine currently used and wanted
1903 * versus held caps. Release, flush, ack revoked caps to mds as
1906 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1907 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1910 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1911 struct ceph_mds_session *session)
1913 struct inode *inode = &ci->vfs_inode;
1914 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1915 struct ceph_cap *cap;
1916 u64 flush_tid, oldest_flush_tid;
1917 int file_wanted, used, cap_used;
1918 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1919 int issued, implemented, want, retain, revoking, flushing = 0;
1920 int mds = -1; /* keep track of how far we've gone through i_caps list
1921 to avoid an infinite loop on retry */
1923 bool queue_invalidate = false;
1924 bool tried_invalidate = false;
1926 spin_lock(&ci->i_ceph_lock);
1927 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1928 flags |= CHECK_CAPS_FLUSH;
1932 spin_lock(&ci->i_ceph_lock);
1934 /* Caps wanted by virtue of active open files. */
1935 file_wanted = __ceph_caps_file_wanted(ci);
1937 /* Caps which have active references against them */
1938 used = __ceph_caps_used(ci);
1941 * "issued" represents the current caps that the MDS wants us to have.
1942 * "implemented" is the set that we have been granted, and includes the
1943 * ones that have not yet been returned to the MDS (the "revoking" set,
1944 * usually because they have outstanding references).
1946 issued = __ceph_caps_issued(ci, &implemented);
1947 revoking = implemented & ~issued;
1951 /* The ones we currently want to retain (may be adjusted below) */
1952 retain = file_wanted | used | CEPH_CAP_PIN;
1953 if (!mdsc->stopping && inode->i_nlink > 0) {
1955 retain |= CEPH_CAP_ANY; /* be greedy */
1956 } else if (S_ISDIR(inode->i_mode) &&
1957 (issued & CEPH_CAP_FILE_SHARED) &&
1958 __ceph_dir_is_complete(ci)) {
1960 * If a directory is complete, we want to keep
1961 * the exclusive cap. So that MDS does not end up
1962 * revoking the shared cap on every create/unlink
1965 if (IS_RDONLY(inode)) {
1966 want = CEPH_CAP_ANY_SHARED;
1968 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1973 retain |= CEPH_CAP_ANY_SHARED;
1975 * keep RD only if we didn't have the file open RW,
1976 * because then the mds would revoke it anyway to
1977 * journal max_size=0.
1979 if (ci->i_max_size == 0)
1980 retain |= CEPH_CAP_ANY_RD;
1984 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1985 " issued %s revoking %s retain %s %s%s\n", inode,
1986 ceph_cap_string(file_wanted),
1987 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1988 ceph_cap_string(ci->i_flushing_caps),
1989 ceph_cap_string(issued), ceph_cap_string(revoking),
1990 ceph_cap_string(retain),
1991 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1992 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1995 * If we no longer need to hold onto old our caps, and we may
1996 * have cached pages, but don't want them, then try to invalidate.
1997 * If we fail, it's because pages are locked.... try again later.
1999 if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
2000 S_ISREG(inode->i_mode) &&
2001 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
2002 inode->i_data.nrpages && /* have cached pages */
2003 (revoking & (CEPH_CAP_FILE_CACHE|
2004 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
2005 !tried_invalidate) {
2006 dout("check_caps trying to invalidate on %p\n", inode);
2007 if (try_nonblocking_invalidate(inode) < 0) {
2008 dout("check_caps queuing invalidate\n");
2009 queue_invalidate = true;
2010 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2012 tried_invalidate = true;
2016 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2018 struct cap_msg_args arg;
2020 cap = rb_entry(p, struct ceph_cap, ci_node);
2022 /* avoid looping forever */
2023 if (mds >= cap->mds ||
2024 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
2027 /* NOTE: no side-effects allowed, until we take s_mutex */
2030 * If we have an auth cap, we don't need to consider any
2031 * overlapping caps as used.
2034 if (ci->i_auth_cap && cap != ci->i_auth_cap)
2035 cap_used &= ~ci->i_auth_cap->issued;
2037 revoking = cap->implemented & ~cap->issued;
2038 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
2039 cap->mds, cap, ceph_cap_string(cap_used),
2040 ceph_cap_string(cap->issued),
2041 ceph_cap_string(cap->implemented),
2042 ceph_cap_string(revoking));
2044 if (cap == ci->i_auth_cap &&
2045 (cap->issued & CEPH_CAP_FILE_WR)) {
2046 /* request larger max_size from MDS? */
2047 if (ci->i_wanted_max_size > ci->i_max_size &&
2048 ci->i_wanted_max_size > ci->i_requested_max_size) {
2049 dout("requesting new max_size\n");
2053 /* approaching file_max? */
2054 if (__ceph_should_report_size(ci)) {
2055 dout("i_size approaching max_size\n");
2059 /* flush anything dirty? */
2060 if (cap == ci->i_auth_cap) {
2061 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
2062 dout("flushing dirty caps\n");
2065 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
2066 dout("flushing snap caps\n");
2071 /* completed revocation? going down and there are no caps? */
2072 if (revoking && (revoking & cap_used) == 0) {
2073 dout("completed revocation of %s\n",
2074 ceph_cap_string(cap->implemented & ~cap->issued));
2078 /* want more caps from mds? */
2079 if (want & ~cap->mds_wanted) {
2080 if (want & ~(cap->mds_wanted | cap->issued))
2082 if (!__cap_is_valid(cap))
2086 /* things we might delay */
2087 if ((cap->issued & ~retain) == 0)
2088 continue; /* nope, all good */
2091 if (session && session != cap->session) {
2092 dout("oops, wrong session %p mutex\n", session);
2093 mutex_unlock(&session->s_mutex);
2097 session = cap->session;
2098 if (mutex_trylock(&session->s_mutex) == 0) {
2099 dout("inverting session/ino locks on %p\n",
2101 session = ceph_get_mds_session(session);
2102 spin_unlock(&ci->i_ceph_lock);
2103 if (took_snap_rwsem) {
2104 up_read(&mdsc->snap_rwsem);
2105 took_snap_rwsem = 0;
2108 mutex_lock(&session->s_mutex);
2109 ceph_put_mds_session(session);
2112 * Because we take the reference while
2113 * holding the i_ceph_lock, it should
2114 * never be NULL. Throw a warning if it
2123 /* kick flushing and flush snaps before sending normal
2125 if (cap == ci->i_auth_cap &&
2127 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2128 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2129 __kick_flushing_caps(mdsc, session, ci, 0);
2130 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2131 __ceph_flush_snaps(ci, session);
2136 /* take snap_rwsem after session mutex */
2137 if (!took_snap_rwsem) {
2138 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
2139 dout("inverting snap/in locks on %p\n",
2141 spin_unlock(&ci->i_ceph_lock);
2142 down_read(&mdsc->snap_rwsem);
2143 took_snap_rwsem = 1;
2146 took_snap_rwsem = 1;
2149 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2150 flushing = ci->i_dirty_caps;
2151 flush_tid = __mark_caps_flushing(inode, session, false,
2153 if (flags & CHECK_CAPS_FLUSH &&
2154 list_empty(&session->s_cap_dirty))
2155 mflags |= CEPH_CLIENT_CAPS_SYNC;
2159 spin_lock(&mdsc->cap_dirty_lock);
2160 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2161 spin_unlock(&mdsc->cap_dirty_lock);
2164 mds = cap->mds; /* remember mds, so we don't repeat */
2166 __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2167 want, retain, flushing, flush_tid, oldest_flush_tid);
2168 spin_unlock(&ci->i_ceph_lock);
2170 __send_cap(&arg, ci);
2172 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2175 /* periodically re-calculate caps wanted by open files */
2176 if (__ceph_is_any_real_caps(ci) &&
2177 list_empty(&ci->i_cap_delay_list) &&
2178 (file_wanted & ~CEPH_CAP_PIN) &&
2179 !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2180 __cap_delay_requeue(mdsc, ci);
2183 spin_unlock(&ci->i_ceph_lock);
2185 if (queue_invalidate)
2186 ceph_queue_invalidate(inode);
2189 mutex_unlock(&session->s_mutex);
2190 if (took_snap_rwsem)
2191 up_read(&mdsc->snap_rwsem);
2195 * Try to flush dirty caps back to the auth mds.
2197 static int try_flush_caps(struct inode *inode, u64 *ptid)
2199 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2200 struct ceph_inode_info *ci = ceph_inode(inode);
2201 struct ceph_mds_session *session = NULL;
2203 u64 flush_tid = 0, oldest_flush_tid = 0;
2206 spin_lock(&ci->i_ceph_lock);
2208 if (ci->i_dirty_caps && ci->i_auth_cap) {
2209 struct ceph_cap *cap = ci->i_auth_cap;
2210 struct cap_msg_args arg;
2212 if (session != cap->session) {
2213 spin_unlock(&ci->i_ceph_lock);
2215 mutex_unlock(&session->s_mutex);
2216 session = cap->session;
2217 mutex_lock(&session->s_mutex);
2220 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) {
2221 spin_unlock(&ci->i_ceph_lock);
2225 if (ci->i_ceph_flags &
2226 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2227 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2228 __kick_flushing_caps(mdsc, session, ci, 0);
2229 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2230 __ceph_flush_snaps(ci, session);
2234 flushing = ci->i_dirty_caps;
2235 flush_tid = __mark_caps_flushing(inode, session, true,
2238 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2239 __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2240 (cap->issued | cap->implemented),
2241 flushing, flush_tid, oldest_flush_tid);
2242 spin_unlock(&ci->i_ceph_lock);
2244 __send_cap(&arg, ci);
2246 if (!list_empty(&ci->i_cap_flush_list)) {
2247 struct ceph_cap_flush *cf =
2248 list_last_entry(&ci->i_cap_flush_list,
2249 struct ceph_cap_flush, i_list);
2251 flush_tid = cf->tid;
2253 flushing = ci->i_flushing_caps;
2254 spin_unlock(&ci->i_ceph_lock);
2258 mutex_unlock(&session->s_mutex);
2265 * Return true if we've flushed caps through the given flush_tid.
2267 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2269 struct ceph_inode_info *ci = ceph_inode(inode);
2272 spin_lock(&ci->i_ceph_lock);
2273 if (!list_empty(&ci->i_cap_flush_list)) {
2274 struct ceph_cap_flush * cf =
2275 list_first_entry(&ci->i_cap_flush_list,
2276 struct ceph_cap_flush, i_list);
2277 if (cf->tid <= flush_tid)
2280 spin_unlock(&ci->i_ceph_lock);
2285 * wait for any unsafe requests to complete.
2287 static int unsafe_request_wait(struct inode *inode)
2289 struct ceph_inode_info *ci = ceph_inode(inode);
2290 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2293 spin_lock(&ci->i_unsafe_lock);
2294 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2295 req1 = list_last_entry(&ci->i_unsafe_dirops,
2296 struct ceph_mds_request,
2298 ceph_mdsc_get_request(req1);
2300 if (!list_empty(&ci->i_unsafe_iops)) {
2301 req2 = list_last_entry(&ci->i_unsafe_iops,
2302 struct ceph_mds_request,
2303 r_unsafe_target_item);
2304 ceph_mdsc_get_request(req2);
2306 spin_unlock(&ci->i_unsafe_lock);
2308 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2309 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2311 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2312 ceph_timeout_jiffies(req1->r_timeout));
2315 ceph_mdsc_put_request(req1);
2318 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2319 ceph_timeout_jiffies(req2->r_timeout));
2322 ceph_mdsc_put_request(req2);
2327 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2329 struct ceph_file_info *fi = file->private_data;
2330 struct inode *inode = file->f_mapping->host;
2331 struct ceph_inode_info *ci = ceph_inode(inode);
2336 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2338 ret = file_write_and_wait_range(file, start, end);
2342 ret = ceph_wait_on_async_create(inode);
2346 dirty = try_flush_caps(inode, &flush_tid);
2347 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2349 err = unsafe_request_wait(inode);
2352 * only wait on non-file metadata writeback (the mds
2353 * can recover size and mtime, so we don't need to
2356 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2357 err = wait_event_interruptible(ci->i_cap_wq,
2358 caps_are_flushed(inode, flush_tid));
2364 if (errseq_check(&ci->i_meta_err, READ_ONCE(fi->meta_err))) {
2365 spin_lock(&file->f_lock);
2366 err = errseq_check_and_advance(&ci->i_meta_err,
2368 spin_unlock(&file->f_lock);
2373 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2378 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2379 * queue inode for flush but don't do so immediately, because we can
2380 * get by with fewer MDS messages if we wait for data writeback to
2383 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2385 struct ceph_inode_info *ci = ceph_inode(inode);
2389 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2391 dout("write_inode %p wait=%d\n", inode, wait);
2393 dirty = try_flush_caps(inode, &flush_tid);
2395 err = wait_event_interruptible(ci->i_cap_wq,
2396 caps_are_flushed(inode, flush_tid));
2398 struct ceph_mds_client *mdsc =
2399 ceph_sb_to_client(inode->i_sb)->mdsc;
2401 spin_lock(&ci->i_ceph_lock);
2402 if (__ceph_caps_dirty(ci))
2403 __cap_delay_requeue_front(mdsc, ci);
2404 spin_unlock(&ci->i_ceph_lock);
2409 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2410 struct ceph_mds_session *session,
2411 struct ceph_inode_info *ci,
2412 u64 oldest_flush_tid)
2413 __releases(ci->i_ceph_lock)
2414 __acquires(ci->i_ceph_lock)
2416 struct inode *inode = &ci->vfs_inode;
2417 struct ceph_cap *cap;
2418 struct ceph_cap_flush *cf;
2421 u64 last_snap_flush = 0;
2423 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2425 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2427 last_snap_flush = cf->tid;
2432 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2433 if (cf->tid < first_tid)
2436 cap = ci->i_auth_cap;
2437 if (!(cap && cap->session == session)) {
2438 pr_err("%p auth cap %p not mds%d ???\n",
2439 inode, cap, session->s_mds);
2443 first_tid = cf->tid + 1;
2446 struct cap_msg_args arg;
2448 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2449 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2450 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2451 (cf->tid < last_snap_flush ?
2452 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2453 __ceph_caps_used(ci),
2454 __ceph_caps_wanted(ci),
2455 (cap->issued | cap->implemented),
2456 cf->caps, cf->tid, oldest_flush_tid);
2457 spin_unlock(&ci->i_ceph_lock);
2458 __send_cap(&arg, ci);
2460 struct ceph_cap_snap *capsnap =
2461 container_of(cf, struct ceph_cap_snap,
2463 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2464 inode, capsnap, cf->tid,
2465 ceph_cap_string(capsnap->dirty));
2467 refcount_inc(&capsnap->nref);
2468 spin_unlock(&ci->i_ceph_lock);
2470 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2473 pr_err("kick_flushing_caps: error sending "
2474 "cap flushsnap, ino (%llx.%llx) "
2475 "tid %llu follows %llu\n",
2476 ceph_vinop(inode), cf->tid,
2480 ceph_put_cap_snap(capsnap);
2483 spin_lock(&ci->i_ceph_lock);
2487 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2488 struct ceph_mds_session *session)
2490 struct ceph_inode_info *ci;
2491 struct ceph_cap *cap;
2492 u64 oldest_flush_tid;
2494 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2496 spin_lock(&mdsc->cap_dirty_lock);
2497 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2498 spin_unlock(&mdsc->cap_dirty_lock);
2500 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2501 spin_lock(&ci->i_ceph_lock);
2502 cap = ci->i_auth_cap;
2503 if (!(cap && cap->session == session)) {
2504 pr_err("%p auth cap %p not mds%d ???\n",
2505 &ci->vfs_inode, cap, session->s_mds);
2506 spin_unlock(&ci->i_ceph_lock);
2512 * if flushing caps were revoked, we re-send the cap flush
2513 * in client reconnect stage. This guarantees MDS * processes
2514 * the cap flush message before issuing the flushing caps to
2517 if ((cap->issued & ci->i_flushing_caps) !=
2518 ci->i_flushing_caps) {
2519 /* encode_caps_cb() also will reset these sequence
2520 * numbers. make sure sequence numbers in cap flush
2521 * message match later reconnect message */
2525 __kick_flushing_caps(mdsc, session, ci,
2528 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2531 spin_unlock(&ci->i_ceph_lock);
2535 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2536 struct ceph_mds_session *session)
2538 struct ceph_inode_info *ci;
2539 struct ceph_cap *cap;
2540 u64 oldest_flush_tid;
2542 lockdep_assert_held(&session->s_mutex);
2544 dout("kick_flushing_caps mds%d\n", session->s_mds);
2546 spin_lock(&mdsc->cap_dirty_lock);
2547 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2548 spin_unlock(&mdsc->cap_dirty_lock);
2550 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2551 spin_lock(&ci->i_ceph_lock);
2552 cap = ci->i_auth_cap;
2553 if (!(cap && cap->session == session)) {
2554 pr_err("%p auth cap %p not mds%d ???\n",
2555 &ci->vfs_inode, cap, session->s_mds);
2556 spin_unlock(&ci->i_ceph_lock);
2559 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2560 __kick_flushing_caps(mdsc, session, ci,
2563 spin_unlock(&ci->i_ceph_lock);
2567 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2568 struct ceph_inode_info *ci)
2570 struct ceph_mds_client *mdsc = session->s_mdsc;
2571 struct ceph_cap *cap = ci->i_auth_cap;
2573 lockdep_assert_held(&ci->i_ceph_lock);
2575 dout("%s %p flushing %s\n", __func__, &ci->vfs_inode,
2576 ceph_cap_string(ci->i_flushing_caps));
2578 if (!list_empty(&ci->i_cap_flush_list)) {
2579 u64 oldest_flush_tid;
2580 spin_lock(&mdsc->cap_dirty_lock);
2581 list_move_tail(&ci->i_flushing_item,
2582 &cap->session->s_cap_flushing);
2583 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2584 spin_unlock(&mdsc->cap_dirty_lock);
2586 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2592 * Take references to capabilities we hold, so that we don't release
2593 * them to the MDS prematurely.
2595 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2596 bool snap_rwsem_locked)
2598 lockdep_assert_held(&ci->i_ceph_lock);
2600 if (got & CEPH_CAP_PIN)
2602 if (got & CEPH_CAP_FILE_RD)
2604 if (got & CEPH_CAP_FILE_CACHE)
2605 ci->i_rdcache_ref++;
2606 if (got & CEPH_CAP_FILE_EXCL)
2608 if (got & CEPH_CAP_FILE_WR) {
2609 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2610 BUG_ON(!snap_rwsem_locked);
2611 ci->i_head_snapc = ceph_get_snap_context(
2612 ci->i_snap_realm->cached_context);
2616 if (got & CEPH_CAP_FILE_BUFFER) {
2617 if (ci->i_wb_ref == 0)
2618 ihold(&ci->vfs_inode);
2620 dout("%s %p wb %d -> %d (?)\n", __func__,
2621 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2626 * Try to grab cap references. Specify those refs we @want, and the
2627 * minimal set we @need. Also include the larger offset we are writing
2628 * to (when applicable), and check against max_size here as well.
2629 * Note that caller is responsible for ensuring max_size increases are
2630 * requested from the MDS.
2632 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2633 * or a negative error code. There are 3 speical error codes:
2634 * -EAGAIN: need to sleep but non-blocking is specified
2635 * -EFBIG: ask caller to call check_max_size() and try again.
2636 * -ESTALE: ask caller to call ceph_renew_caps() and try again.
2639 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2640 NON_BLOCKING = (1 << 8),
2641 CHECK_FILELOCK = (1 << 9),
2644 static int try_get_cap_refs(struct inode *inode, int need, int want,
2645 loff_t endoff, int flags, int *got)
2647 struct ceph_inode_info *ci = ceph_inode(inode);
2648 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2650 int have, implemented;
2651 bool snap_rwsem_locked = false;
2653 dout("get_cap_refs %p need %s want %s\n", inode,
2654 ceph_cap_string(need), ceph_cap_string(want));
2657 spin_lock(&ci->i_ceph_lock);
2659 if ((flags & CHECK_FILELOCK) &&
2660 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2661 dout("try_get_cap_refs %p error filelock\n", inode);
2666 /* finish pending truncate */
2667 while (ci->i_truncate_pending) {
2668 spin_unlock(&ci->i_ceph_lock);
2669 if (snap_rwsem_locked) {
2670 up_read(&mdsc->snap_rwsem);
2671 snap_rwsem_locked = false;
2673 __ceph_do_pending_vmtruncate(inode);
2674 spin_lock(&ci->i_ceph_lock);
2677 have = __ceph_caps_issued(ci, &implemented);
2679 if (have & need & CEPH_CAP_FILE_WR) {
2680 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2681 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2682 inode, endoff, ci->i_max_size);
2683 if (endoff > ci->i_requested_max_size)
2684 ret = ci->i_auth_cap ? -EFBIG : -ESTALE;
2688 * If a sync write is in progress, we must wait, so that we
2689 * can get a final snapshot value for size+mtime.
2691 if (__ceph_have_pending_cap_snap(ci)) {
2692 dout("get_cap_refs %p cap_snap_pending\n", inode);
2697 if ((have & need) == need) {
2699 * Look at (implemented & ~have & not) so that we keep waiting
2700 * on transition from wanted -> needed caps. This is needed
2701 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2702 * going before a prior buffered writeback happens.
2704 int not = want & ~(have & need);
2705 int revoking = implemented & ~have;
2706 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2707 inode, ceph_cap_string(have), ceph_cap_string(not),
2708 ceph_cap_string(revoking));
2709 if ((revoking & not) == 0) {
2710 if (!snap_rwsem_locked &&
2711 !ci->i_head_snapc &&
2712 (need & CEPH_CAP_FILE_WR)) {
2713 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2715 * we can not call down_read() when
2716 * task isn't in TASK_RUNNING state
2718 if (flags & NON_BLOCKING) {
2723 spin_unlock(&ci->i_ceph_lock);
2724 down_read(&mdsc->snap_rwsem);
2725 snap_rwsem_locked = true;
2728 snap_rwsem_locked = true;
2730 if ((have & want) == want)
2734 ceph_take_cap_refs(ci, *got, true);
2738 int session_readonly = false;
2740 if (ci->i_auth_cap &&
2741 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2742 struct ceph_mds_session *s = ci->i_auth_cap->session;
2743 spin_lock(&s->s_cap_lock);
2744 session_readonly = s->s_readonly;
2745 spin_unlock(&s->s_cap_lock);
2747 if (session_readonly) {
2748 dout("get_cap_refs %p need %s but mds%d readonly\n",
2749 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2754 if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
2755 dout("get_cap_refs %p forced umount\n", inode);
2759 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2760 if (need & ~mds_wanted) {
2761 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2762 inode, ceph_cap_string(need),
2763 ceph_cap_string(mds_wanted));
2768 dout("get_cap_refs %p have %s need %s\n", inode,
2769 ceph_cap_string(have), ceph_cap_string(need));
2773 __ceph_touch_fmode(ci, mdsc, flags);
2775 spin_unlock(&ci->i_ceph_lock);
2776 if (snap_rwsem_locked)
2777 up_read(&mdsc->snap_rwsem);
2780 ceph_update_cap_mis(&mdsc->metric);
2782 ceph_update_cap_hit(&mdsc->metric);
2784 dout("get_cap_refs %p ret %d got %s\n", inode,
2785 ret, ceph_cap_string(*got));
2790 * Check the offset we are writing up to against our current
2791 * max_size. If necessary, tell the MDS we want to write to
2794 static void check_max_size(struct inode *inode, loff_t endoff)
2796 struct ceph_inode_info *ci = ceph_inode(inode);
2799 /* do we need to explicitly request a larger max_size? */
2800 spin_lock(&ci->i_ceph_lock);
2801 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2802 dout("write %p at large endoff %llu, req max_size\n",
2804 ci->i_wanted_max_size = endoff;
2806 /* duplicate ceph_check_caps()'s logic */
2807 if (ci->i_auth_cap &&
2808 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2809 ci->i_wanted_max_size > ci->i_max_size &&
2810 ci->i_wanted_max_size > ci->i_requested_max_size)
2812 spin_unlock(&ci->i_ceph_lock);
2814 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2817 static inline int get_used_fmode(int caps)
2820 if (caps & CEPH_CAP_FILE_RD)
2821 fmode |= CEPH_FILE_MODE_RD;
2822 if (caps & CEPH_CAP_FILE_WR)
2823 fmode |= CEPH_FILE_MODE_WR;
2827 int ceph_try_get_caps(struct inode *inode, int need, int want,
2828 bool nonblock, int *got)
2832 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2833 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2834 CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2835 CEPH_CAP_ANY_DIR_OPS));
2837 ret = ceph_pool_perm_check(inode, need);
2842 flags = get_used_fmode(need | want);
2844 flags |= NON_BLOCKING;
2846 ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2847 /* three special error codes */
2848 if (ret == -EAGAIN || ret == -EFBIG || ret == -ESTALE)
2854 * Wait for caps, and take cap references. If we can't get a WR cap
2855 * due to a small max_size, make sure we check_max_size (and possibly
2856 * ask the mds) so we don't get hung up indefinitely.
2858 int ceph_get_caps(struct file *filp, int need, int want, loff_t endoff, int *got)
2860 struct ceph_file_info *fi = filp->private_data;
2861 struct inode *inode = file_inode(filp);
2862 struct ceph_inode_info *ci = ceph_inode(inode);
2863 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2864 int ret, _got, flags;
2866 ret = ceph_pool_perm_check(inode, need);
2870 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2871 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2874 flags = get_used_fmode(need | want);
2877 flags &= CEPH_FILE_MODE_MASK;
2878 if (atomic_read(&fi->num_locks))
2879 flags |= CHECK_FILELOCK;
2881 ret = try_get_cap_refs(inode, need, want, endoff,
2883 WARN_ON_ONCE(ret == -EAGAIN);
2885 struct ceph_mds_client *mdsc = fsc->mdsc;
2887 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2889 cw.ino = ceph_ino(inode);
2890 cw.tgid = current->tgid;
2894 spin_lock(&mdsc->caps_list_lock);
2895 list_add(&cw.list, &mdsc->cap_wait_list);
2896 spin_unlock(&mdsc->caps_list_lock);
2898 /* make sure used fmode not timeout */
2899 ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2900 add_wait_queue(&ci->i_cap_wq, &wait);
2902 flags |= NON_BLOCKING;
2903 while (!(ret = try_get_cap_refs(inode, need, want,
2904 endoff, flags, &_got))) {
2905 if (signal_pending(current)) {
2909 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2912 remove_wait_queue(&ci->i_cap_wq, &wait);
2913 ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
2915 spin_lock(&mdsc->caps_list_lock);
2917 spin_unlock(&mdsc->caps_list_lock);
2923 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2924 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2925 if (ret >= 0 && _got)
2926 ceph_put_cap_refs(ci, _got);
2931 if (ret == -EFBIG || ret == -ESTALE) {
2932 int ret2 = ceph_wait_on_async_create(inode);
2936 if (ret == -EFBIG) {
2937 check_max_size(inode, endoff);
2940 if (ret == -ESTALE) {
2941 /* session was killed, try renew caps */
2942 ret = ceph_renew_caps(inode, flags);
2949 if (S_ISREG(ci->vfs_inode.i_mode) &&
2950 ci->i_inline_version != CEPH_INLINE_NONE &&
2951 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2952 i_size_read(inode) > 0) {
2954 find_get_page(inode->i_mapping, 0);
2956 bool uptodate = PageUptodate(page);
2963 * drop cap refs first because getattr while
2964 * holding * caps refs can cause deadlock.
2966 ceph_put_cap_refs(ci, _got);
2970 * getattr request will bring inline data into
2973 ret = __ceph_do_getattr(inode, NULL,
2974 CEPH_STAT_CAP_INLINE_DATA,
2987 * Take cap refs. Caller must already know we hold at least one ref
2988 * on the caps in question or we don't know this is safe.
2990 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2992 spin_lock(&ci->i_ceph_lock);
2993 ceph_take_cap_refs(ci, caps, false);
2994 spin_unlock(&ci->i_ceph_lock);
2999 * drop cap_snap that is not associated with any snapshot.
3000 * we don't need to send FLUSHSNAP message for it.
3002 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
3003 struct ceph_cap_snap *capsnap)
3005 if (!capsnap->need_flush &&
3006 !capsnap->writing && !capsnap->dirty_pages) {
3007 dout("dropping cap_snap %p follows %llu\n",
3008 capsnap, capsnap->follows);
3009 BUG_ON(capsnap->cap_flush.tid > 0);
3010 ceph_put_snap_context(capsnap->context);
3011 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
3012 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3014 list_del(&capsnap->ci_item);
3015 ceph_put_cap_snap(capsnap);
3021 enum put_cap_refs_mode {
3022 PUT_CAP_REFS_SYNC = 0,
3023 PUT_CAP_REFS_NO_CHECK,
3030 * If we released the last ref on any given cap, call ceph_check_caps
3031 * to release (or schedule a release).
3033 * If we are releasing a WR cap (from a sync write), finalize any affected
3034 * cap_snap, and wake up any waiters.
3036 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
3037 enum put_cap_refs_mode mode)
3039 struct inode *inode = &ci->vfs_inode;
3040 int last = 0, put = 0, flushsnaps = 0, wake = 0;
3041 bool check_flushsnaps = false;
3043 spin_lock(&ci->i_ceph_lock);
3044 if (had & CEPH_CAP_PIN)
3046 if (had & CEPH_CAP_FILE_RD)
3047 if (--ci->i_rd_ref == 0)
3049 if (had & CEPH_CAP_FILE_CACHE)
3050 if (--ci->i_rdcache_ref == 0)
3052 if (had & CEPH_CAP_FILE_EXCL)
3053 if (--ci->i_fx_ref == 0)
3055 if (had & CEPH_CAP_FILE_BUFFER) {
3056 if (--ci->i_wb_ref == 0) {
3058 /* put the ref held by ceph_take_cap_refs() */
3060 check_flushsnaps = true;
3062 dout("put_cap_refs %p wb %d -> %d (?)\n",
3063 inode, ci->i_wb_ref+1, ci->i_wb_ref);
3065 if (had & CEPH_CAP_FILE_WR) {
3066 if (--ci->i_wr_ref == 0) {
3068 check_flushsnaps = true;
3069 if (ci->i_wrbuffer_ref_head == 0 &&
3070 ci->i_dirty_caps == 0 &&
3071 ci->i_flushing_caps == 0) {
3072 BUG_ON(!ci->i_head_snapc);
3073 ceph_put_snap_context(ci->i_head_snapc);
3074 ci->i_head_snapc = NULL;
3076 /* see comment in __ceph_remove_cap() */
3077 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3078 drop_inode_snap_realm(ci);
3081 if (check_flushsnaps && __ceph_have_pending_cap_snap(ci)) {
3082 struct ceph_cap_snap *capsnap =
3083 list_last_entry(&ci->i_cap_snaps,
3084 struct ceph_cap_snap,
3087 capsnap->writing = 0;
3088 if (ceph_try_drop_cap_snap(ci, capsnap))
3089 /* put the ref held by ceph_queue_cap_snap() */
3091 else if (__ceph_finish_cap_snap(ci, capsnap))
3095 spin_unlock(&ci->i_ceph_lock);
3097 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3098 last ? " last" : "", put ? " put" : "");
3101 case PUT_CAP_REFS_SYNC:
3103 ceph_check_caps(ci, 0, NULL);
3104 else if (flushsnaps)
3105 ceph_flush_snaps(ci, NULL);
3107 case PUT_CAP_REFS_ASYNC:
3109 ceph_queue_check_caps(inode);
3110 else if (flushsnaps)
3111 ceph_queue_flush_snaps(inode);
3117 wake_up_all(&ci->i_cap_wq);
3122 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3124 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_SYNC);
3127 void ceph_put_cap_refs_async(struct ceph_inode_info *ci, int had)
3129 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_ASYNC);
3132 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3134 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_NO_CHECK);
3138 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3139 * context. Adjust per-snap dirty page accounting as appropriate.
3140 * Once all dirty data for a cap_snap is flushed, flush snapped file
3141 * metadata back to the MDS. If we dropped the last ref, call
3144 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3145 struct ceph_snap_context *snapc)
3147 struct inode *inode = &ci->vfs_inode;
3148 struct ceph_cap_snap *capsnap = NULL;
3152 bool flush_snaps = false;
3153 bool complete_capsnap = false;
3155 spin_lock(&ci->i_ceph_lock);
3156 ci->i_wrbuffer_ref -= nr;
3157 if (ci->i_wrbuffer_ref == 0) {
3162 if (ci->i_head_snapc == snapc) {
3163 ci->i_wrbuffer_ref_head -= nr;
3164 if (ci->i_wrbuffer_ref_head == 0 &&
3165 ci->i_wr_ref == 0 &&
3166 ci->i_dirty_caps == 0 &&
3167 ci->i_flushing_caps == 0) {
3168 BUG_ON(!ci->i_head_snapc);
3169 ceph_put_snap_context(ci->i_head_snapc);
3170 ci->i_head_snapc = NULL;
3172 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3174 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3175 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3176 last ? " LAST" : "");
3178 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3179 if (capsnap->context == snapc) {
3185 capsnap->dirty_pages -= nr;
3186 if (capsnap->dirty_pages == 0) {
3187 complete_capsnap = true;
3188 if (!capsnap->writing) {
3189 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3192 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3197 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3198 " snap %lld %d/%d -> %d/%d %s%s\n",
3199 inode, capsnap, capsnap->context->seq,
3200 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3201 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3202 last ? " (wrbuffer last)" : "",
3203 complete_capsnap ? " (complete capsnap)" : "");
3206 spin_unlock(&ci->i_ceph_lock);
3209 ceph_check_caps(ci, 0, NULL);
3210 } else if (flush_snaps) {
3211 ceph_flush_snaps(ci, NULL);
3213 if (complete_capsnap)
3214 wake_up_all(&ci->i_cap_wq);
3216 /* avoid calling iput_final() in osd dispatch threads */
3217 ceph_async_iput(inode);
3222 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3224 static void invalidate_aliases(struct inode *inode)
3226 struct dentry *dn, *prev = NULL;
3228 dout("invalidate_aliases inode %p\n", inode);
3229 d_prune_aliases(inode);
3231 * For non-directory inode, d_find_alias() only returns
3232 * hashed dentry. After calling d_invalidate(), the
3233 * dentry becomes unhashed.
3235 * For directory inode, d_find_alias() can return
3236 * unhashed dentry. But directory inode should have
3237 * one alias at most.
3239 while ((dn = d_find_alias(inode))) {
3253 struct cap_extra_info {
3254 struct ceph_string *pool_ns;
3264 /* currently issued */
3266 struct timespec64 btime;
3270 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3271 * actually be a revocation if it specifies a smaller cap set.)
3273 * caller holds s_mutex and i_ceph_lock, we drop both.
3275 static void handle_cap_grant(struct inode *inode,
3276 struct ceph_mds_session *session,
3277 struct ceph_cap *cap,
3278 struct ceph_mds_caps *grant,
3279 struct ceph_buffer *xattr_buf,
3280 struct cap_extra_info *extra_info)
3281 __releases(ci->i_ceph_lock)
3282 __releases(session->s_mdsc->snap_rwsem)
3284 struct ceph_inode_info *ci = ceph_inode(inode);
3285 int seq = le32_to_cpu(grant->seq);
3286 int newcaps = le32_to_cpu(grant->caps);
3287 int used, wanted, dirty;
3288 u64 size = le64_to_cpu(grant->size);
3289 u64 max_size = le64_to_cpu(grant->max_size);
3290 unsigned char check_caps = 0;
3291 bool was_stale = cap->cap_gen < session->s_cap_gen;
3293 bool writeback = false;
3294 bool queue_trunc = false;
3295 bool queue_invalidate = false;
3296 bool deleted_inode = false;
3297 bool fill_inline = false;
3299 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3300 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3301 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3302 i_size_read(inode));
3306 * If CACHE is being revoked, and we have no dirty buffers,
3307 * try to invalidate (once). (If there are dirty buffers, we
3308 * will invalidate _after_ writeback.)
3310 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3311 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3312 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3313 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3314 if (try_nonblocking_invalidate(inode)) {
3315 /* there were locked pages.. invalidate later
3316 in a separate thread. */
3317 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3318 queue_invalidate = true;
3319 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3325 cap->issued = cap->implemented = CEPH_CAP_PIN;
3328 * auth mds of the inode changed. we received the cap export message,
3329 * but still haven't received the cap import message. handle_cap_export
3330 * updated the new auth MDS' cap.
3332 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3333 * that was sent before the cap import message. So don't remove caps.
3335 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3336 WARN_ON(cap != ci->i_auth_cap);
3337 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3339 newcaps |= cap->issued;
3342 /* side effects now are allowed */
3343 cap->cap_gen = session->s_cap_gen;
3346 __check_cap_issue(ci, cap, newcaps);
3348 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3350 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3351 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3352 umode_t mode = le32_to_cpu(grant->mode);
3354 if (inode_wrong_type(inode, mode))
3355 pr_warn_once("inode type changed! (ino %llx.%llx is 0%o, mds says 0%o)\n",
3356 ceph_vinop(inode), inode->i_mode, mode);
3358 inode->i_mode = mode;
3359 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3360 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3361 ci->i_btime = extra_info->btime;
3362 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3363 from_kuid(&init_user_ns, inode->i_uid),
3364 from_kgid(&init_user_ns, inode->i_gid));
3367 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3368 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3369 set_nlink(inode, le32_to_cpu(grant->nlink));
3370 if (inode->i_nlink == 0 &&
3371 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3372 deleted_inode = true;
3375 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3377 int len = le32_to_cpu(grant->xattr_len);
3378 u64 version = le64_to_cpu(grant->xattr_version);
3380 if (version > ci->i_xattrs.version) {
3381 dout(" got new xattrs v%llu on %p len %d\n",
3382 version, inode, len);
3383 if (ci->i_xattrs.blob)
3384 ceph_buffer_put(ci->i_xattrs.blob);
3385 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3386 ci->i_xattrs.version = version;
3387 ceph_forget_all_cached_acls(inode);
3388 ceph_security_invalidate_secctx(inode);
3392 if (newcaps & CEPH_CAP_ANY_RD) {
3393 struct timespec64 mtime, atime, ctime;
3394 /* ctime/mtime/atime? */
3395 ceph_decode_timespec64(&mtime, &grant->mtime);
3396 ceph_decode_timespec64(&atime, &grant->atime);
3397 ceph_decode_timespec64(&ctime, &grant->ctime);
3398 ceph_fill_file_time(inode, extra_info->issued,
3399 le32_to_cpu(grant->time_warp_seq),
3400 &ctime, &mtime, &atime);
3403 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3404 ci->i_files = extra_info->nfiles;
3405 ci->i_subdirs = extra_info->nsubdirs;
3408 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3409 /* file layout may have changed */
3410 s64 old_pool = ci->i_layout.pool_id;
3411 struct ceph_string *old_ns;
3413 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3414 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3415 lockdep_is_held(&ci->i_ceph_lock));
3416 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3418 if (ci->i_layout.pool_id != old_pool ||
3419 extra_info->pool_ns != old_ns)
3420 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3422 extra_info->pool_ns = old_ns;
3424 /* size/truncate_seq? */
3425 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3426 le32_to_cpu(grant->truncate_seq),
3427 le64_to_cpu(grant->truncate_size),
3431 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3432 if (max_size != ci->i_max_size) {
3433 dout("max_size %lld -> %llu\n",
3434 ci->i_max_size, max_size);
3435 ci->i_max_size = max_size;
3436 if (max_size >= ci->i_wanted_max_size) {
3437 ci->i_wanted_max_size = 0; /* reset */
3438 ci->i_requested_max_size = 0;
3444 /* check cap bits */
3445 wanted = __ceph_caps_wanted(ci);
3446 used = __ceph_caps_used(ci);
3447 dirty = __ceph_caps_dirty(ci);
3448 dout(" my wanted = %s, used = %s, dirty %s\n",
3449 ceph_cap_string(wanted),
3450 ceph_cap_string(used),
3451 ceph_cap_string(dirty));
3453 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3454 (wanted & ~(cap->mds_wanted | newcaps))) {
3456 * If mds is importing cap, prior cap messages that update
3457 * 'wanted' may get dropped by mds (migrate seq mismatch).
3459 * We don't send cap message to update 'wanted' if what we
3460 * want are already issued. If mds revokes caps, cap message
3461 * that releases caps also tells mds what we want. But if
3462 * caps got revoked by mds forcedly (session stale). We may
3463 * haven't told mds what we want.
3468 /* revocation, grant, or no-op? */
3469 if (cap->issued & ~newcaps) {
3470 int revoking = cap->issued & ~newcaps;
3472 dout("revocation: %s -> %s (revoking %s)\n",
3473 ceph_cap_string(cap->issued),
3474 ceph_cap_string(newcaps),
3475 ceph_cap_string(revoking));
3476 if (S_ISREG(inode->i_mode) &&
3477 (revoking & used & CEPH_CAP_FILE_BUFFER))
3478 writeback = true; /* initiate writeback; will delay ack */
3479 else if (queue_invalidate &&
3480 revoking == CEPH_CAP_FILE_CACHE &&
3481 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3482 ; /* do nothing yet, invalidation will be queued */
3483 else if (cap == ci->i_auth_cap)
3484 check_caps = 1; /* check auth cap only */
3486 check_caps = 2; /* check all caps */
3487 cap->issued = newcaps;
3488 cap->implemented |= newcaps;
3489 } else if (cap->issued == newcaps) {
3490 dout("caps unchanged: %s -> %s\n",
3491 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3493 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3494 ceph_cap_string(newcaps));
3495 /* non-auth MDS is revoking the newly grant caps ? */
3496 if (cap == ci->i_auth_cap &&
3497 __ceph_caps_revoking_other(ci, cap, newcaps))
3500 cap->issued = newcaps;
3501 cap->implemented |= newcaps; /* add bits only, to
3502 * avoid stepping on a
3503 * pending revocation */
3506 BUG_ON(cap->issued & ~cap->implemented);
3508 if (extra_info->inline_version > 0 &&
3509 extra_info->inline_version >= ci->i_inline_version) {
3510 ci->i_inline_version = extra_info->inline_version;
3511 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3512 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3516 if (ci->i_auth_cap == cap &&
3517 le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3518 if (newcaps & ~extra_info->issued)
3521 if (ci->i_requested_max_size > max_size ||
3522 !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3523 /* re-request max_size if necessary */
3524 ci->i_requested_max_size = 0;
3528 ceph_kick_flushing_inode_caps(session, ci);
3529 spin_unlock(&ci->i_ceph_lock);
3530 up_read(&session->s_mdsc->snap_rwsem);
3532 spin_unlock(&ci->i_ceph_lock);
3536 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3537 extra_info->inline_len);
3540 ceph_queue_vmtruncate(inode);
3544 * queue inode for writeback: we can't actually call
3545 * filemap_write_and_wait, etc. from message handler
3548 ceph_queue_writeback(inode);
3549 if (queue_invalidate)
3550 ceph_queue_invalidate(inode);
3552 invalidate_aliases(inode);
3554 wake_up_all(&ci->i_cap_wq);
3556 if (check_caps == 1)
3557 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3559 else if (check_caps == 2)
3560 ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3562 mutex_unlock(&session->s_mutex);
3566 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3567 * MDS has been safely committed.
3569 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3570 struct ceph_mds_caps *m,
3571 struct ceph_mds_session *session,
3572 struct ceph_cap *cap)
3573 __releases(ci->i_ceph_lock)
3575 struct ceph_inode_info *ci = ceph_inode(inode);
3576 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3577 struct ceph_cap_flush *cf, *tmp_cf;
3578 LIST_HEAD(to_remove);
3579 unsigned seq = le32_to_cpu(m->seq);
3580 int dirty = le32_to_cpu(m->dirty);
3583 bool wake_ci = false;
3584 bool wake_mdsc = false;
3586 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3587 /* Is this the one that was flushed? */
3588 if (cf->tid == flush_tid)
3591 /* Is this a capsnap? */
3595 if (cf->tid <= flush_tid) {
3597 * An earlier or current tid. The FLUSH_ACK should
3598 * represent a superset of this flush's caps.
3600 wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3601 list_add_tail(&cf->i_list, &to_remove);
3604 * This is a later one. Any caps in it are still dirty
3605 * so don't count them as cleaned.
3607 cleaned &= ~cf->caps;
3613 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3614 " flushing %s -> %s\n",
3615 inode, session->s_mds, seq, ceph_cap_string(dirty),
3616 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3617 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3619 if (list_empty(&to_remove) && !cleaned)
3622 ci->i_flushing_caps &= ~cleaned;
3624 spin_lock(&mdsc->cap_dirty_lock);
3626 list_for_each_entry(cf, &to_remove, i_list)
3627 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3629 if (ci->i_flushing_caps == 0) {
3630 if (list_empty(&ci->i_cap_flush_list)) {
3631 list_del_init(&ci->i_flushing_item);
3632 if (!list_empty(&session->s_cap_flushing)) {
3633 dout(" mds%d still flushing cap on %p\n",
3635 &list_first_entry(&session->s_cap_flushing,
3636 struct ceph_inode_info,
3637 i_flushing_item)->vfs_inode);
3640 mdsc->num_cap_flushing--;
3641 dout(" inode %p now !flushing\n", inode);
3643 if (ci->i_dirty_caps == 0) {
3644 dout(" inode %p now clean\n", inode);
3645 BUG_ON(!list_empty(&ci->i_dirty_item));
3647 if (ci->i_wr_ref == 0 &&
3648 ci->i_wrbuffer_ref_head == 0) {
3649 BUG_ON(!ci->i_head_snapc);
3650 ceph_put_snap_context(ci->i_head_snapc);
3651 ci->i_head_snapc = NULL;
3654 BUG_ON(list_empty(&ci->i_dirty_item));
3657 spin_unlock(&mdsc->cap_dirty_lock);
3660 spin_unlock(&ci->i_ceph_lock);
3662 while (!list_empty(&to_remove)) {
3663 cf = list_first_entry(&to_remove,
3664 struct ceph_cap_flush, i_list);
3665 list_del(&cf->i_list);
3666 ceph_free_cap_flush(cf);
3670 wake_up_all(&ci->i_cap_wq);
3672 wake_up_all(&mdsc->cap_flushing_wq);
3678 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3679 * throw away our cap_snap.
3681 * Caller hold s_mutex.
3683 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3684 struct ceph_mds_caps *m,
3685 struct ceph_mds_session *session)
3687 struct ceph_inode_info *ci = ceph_inode(inode);
3688 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3689 u64 follows = le64_to_cpu(m->snap_follows);
3690 struct ceph_cap_snap *capsnap;
3691 bool flushed = false;
3692 bool wake_ci = false;
3693 bool wake_mdsc = false;
3695 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3696 inode, ci, session->s_mds, follows);
3698 spin_lock(&ci->i_ceph_lock);
3699 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3700 if (capsnap->follows == follows) {
3701 if (capsnap->cap_flush.tid != flush_tid) {
3702 dout(" cap_snap %p follows %lld tid %lld !="
3703 " %lld\n", capsnap, follows,
3704 flush_tid, capsnap->cap_flush.tid);
3710 dout(" skipping cap_snap %p follows %lld\n",
3711 capsnap, capsnap->follows);
3715 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3716 dout(" removing %p cap_snap %p follows %lld\n",
3717 inode, capsnap, follows);
3718 list_del(&capsnap->ci_item);
3719 wake_ci |= __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3721 spin_lock(&mdsc->cap_dirty_lock);
3723 if (list_empty(&ci->i_cap_flush_list))
3724 list_del_init(&ci->i_flushing_item);
3726 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc,
3727 &capsnap->cap_flush);
3728 spin_unlock(&mdsc->cap_dirty_lock);
3730 spin_unlock(&ci->i_ceph_lock);
3732 ceph_put_snap_context(capsnap->context);
3733 ceph_put_cap_snap(capsnap);
3735 wake_up_all(&ci->i_cap_wq);
3737 wake_up_all(&mdsc->cap_flushing_wq);
3743 * Handle TRUNC from MDS, indicating file truncation.
3745 * caller hold s_mutex.
3747 static bool handle_cap_trunc(struct inode *inode,
3748 struct ceph_mds_caps *trunc,
3749 struct ceph_mds_session *session)
3751 struct ceph_inode_info *ci = ceph_inode(inode);
3752 int mds = session->s_mds;
3753 int seq = le32_to_cpu(trunc->seq);
3754 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3755 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3756 u64 size = le64_to_cpu(trunc->size);
3757 int implemented = 0;
3758 int dirty = __ceph_caps_dirty(ci);
3759 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3760 bool queue_trunc = false;
3762 lockdep_assert_held(&ci->i_ceph_lock);
3764 issued |= implemented | dirty;
3766 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3767 inode, mds, seq, truncate_size, truncate_seq);
3768 queue_trunc = ceph_fill_file_size(inode, issued,
3769 truncate_seq, truncate_size, size);
3774 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3775 * different one. If we are the most recent migration we've seen (as
3776 * indicated by mseq), make note of the migrating cap bits for the
3777 * duration (until we see the corresponding IMPORT).
3779 * caller holds s_mutex
3781 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3782 struct ceph_mds_cap_peer *ph,
3783 struct ceph_mds_session *session)
3785 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3786 struct ceph_mds_session *tsession = NULL;
3787 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3788 struct ceph_inode_info *ci = ceph_inode(inode);
3790 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3791 unsigned t_seq, t_mseq;
3793 int mds = session->s_mds;
3796 t_cap_id = le64_to_cpu(ph->cap_id);
3797 t_seq = le32_to_cpu(ph->seq);
3798 t_mseq = le32_to_cpu(ph->mseq);
3799 target = le32_to_cpu(ph->mds);
3801 t_cap_id = t_seq = t_mseq = 0;
3805 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3806 inode, ci, mds, mseq, target);
3808 spin_lock(&ci->i_ceph_lock);
3809 cap = __get_cap_for_mds(ci, mds);
3810 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3814 __ceph_remove_cap(cap, false);
3819 * now we know we haven't received the cap import message yet
3820 * because the exported cap still exist.
3823 issued = cap->issued;
3824 if (issued != cap->implemented)
3825 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3826 "ino (%llx.%llx) mds%d seq %d mseq %d "
3827 "issued %s implemented %s\n",
3828 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3829 ceph_cap_string(issued),
3830 ceph_cap_string(cap->implemented));
3833 tcap = __get_cap_for_mds(ci, target);
3835 /* already have caps from the target */
3836 if (tcap->cap_id == t_cap_id &&
3837 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3838 dout(" updating import cap %p mds%d\n", tcap, target);
3839 tcap->cap_id = t_cap_id;
3840 tcap->seq = t_seq - 1;
3841 tcap->issue_seq = t_seq - 1;
3842 tcap->issued |= issued;
3843 tcap->implemented |= issued;
3844 if (cap == ci->i_auth_cap) {
3845 ci->i_auth_cap = tcap;
3846 change_auth_cap_ses(ci, tcap->session);
3849 __ceph_remove_cap(cap, false);
3851 } else if (tsession) {
3852 /* add placeholder for the export tagert */
3853 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3855 ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3856 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3858 if (!list_empty(&ci->i_cap_flush_list) &&
3859 ci->i_auth_cap == tcap) {
3860 spin_lock(&mdsc->cap_dirty_lock);
3861 list_move_tail(&ci->i_flushing_item,
3862 &tcap->session->s_cap_flushing);
3863 spin_unlock(&mdsc->cap_dirty_lock);
3866 __ceph_remove_cap(cap, false);
3870 spin_unlock(&ci->i_ceph_lock);
3871 mutex_unlock(&session->s_mutex);
3873 /* open target session */
3874 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3875 if (!IS_ERR(tsession)) {
3877 mutex_lock(&session->s_mutex);
3878 mutex_lock_nested(&tsession->s_mutex,
3879 SINGLE_DEPTH_NESTING);
3881 mutex_lock(&tsession->s_mutex);
3882 mutex_lock_nested(&session->s_mutex,
3883 SINGLE_DEPTH_NESTING);
3885 new_cap = ceph_get_cap(mdsc, NULL);
3890 mutex_lock(&session->s_mutex);
3895 spin_unlock(&ci->i_ceph_lock);
3896 mutex_unlock(&session->s_mutex);
3898 mutex_unlock(&tsession->s_mutex);
3899 ceph_put_mds_session(tsession);
3902 ceph_put_cap(mdsc, new_cap);
3906 * Handle cap IMPORT.
3908 * caller holds s_mutex. acquires i_ceph_lock
3910 static void handle_cap_import(struct ceph_mds_client *mdsc,
3911 struct inode *inode, struct ceph_mds_caps *im,
3912 struct ceph_mds_cap_peer *ph,
3913 struct ceph_mds_session *session,
3914 struct ceph_cap **target_cap, int *old_issued)
3916 struct ceph_inode_info *ci = ceph_inode(inode);
3917 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3918 int mds = session->s_mds;
3920 unsigned caps = le32_to_cpu(im->caps);
3921 unsigned wanted = le32_to_cpu(im->wanted);
3922 unsigned seq = le32_to_cpu(im->seq);
3923 unsigned mseq = le32_to_cpu(im->migrate_seq);
3924 u64 realmino = le64_to_cpu(im->realm);
3925 u64 cap_id = le64_to_cpu(im->cap_id);
3930 p_cap_id = le64_to_cpu(ph->cap_id);
3931 peer = le32_to_cpu(ph->mds);
3937 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3938 inode, ci, mds, mseq, peer);
3940 cap = __get_cap_for_mds(ci, mds);
3943 spin_unlock(&ci->i_ceph_lock);
3944 new_cap = ceph_get_cap(mdsc, NULL);
3945 spin_lock(&ci->i_ceph_lock);
3951 ceph_put_cap(mdsc, new_cap);
3956 __ceph_caps_issued(ci, &issued);
3957 issued |= __ceph_caps_dirty(ci);
3959 ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
3960 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3962 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3963 if (ocap && ocap->cap_id == p_cap_id) {
3964 dout(" remove export cap %p mds%d flags %d\n",
3965 ocap, peer, ph->flags);
3966 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3967 (ocap->seq != le32_to_cpu(ph->seq) ||
3968 ocap->mseq != le32_to_cpu(ph->mseq))) {
3969 pr_err_ratelimited("handle_cap_import: "
3970 "mismatched seq/mseq: ino (%llx.%llx) "
3971 "mds%d seq %d mseq %d importer mds%d "
3972 "has peer seq %d mseq %d\n",
3973 ceph_vinop(inode), peer, ocap->seq,
3974 ocap->mseq, mds, le32_to_cpu(ph->seq),
3975 le32_to_cpu(ph->mseq));
3977 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3980 *old_issued = issued;
3985 * Handle a caps message from the MDS.
3987 * Identify the appropriate session, inode, and call the right handler
3988 * based on the cap op.
3990 void ceph_handle_caps(struct ceph_mds_session *session,
3991 struct ceph_msg *msg)
3993 struct ceph_mds_client *mdsc = session->s_mdsc;
3994 struct inode *inode;
3995 struct ceph_inode_info *ci;
3996 struct ceph_cap *cap;
3997 struct ceph_mds_caps *h;
3998 struct ceph_mds_cap_peer *peer = NULL;
3999 struct ceph_snap_realm *realm = NULL;
4001 int msg_version = le16_to_cpu(msg->hdr.version);
4003 struct ceph_vino vino;
4005 size_t snaptrace_len;
4007 struct cap_extra_info extra_info = {};
4010 dout("handle_caps from mds%d\n", session->s_mds);
4013 end = msg->front.iov_base + msg->front.iov_len;
4014 if (msg->front.iov_len < sizeof(*h))
4016 h = msg->front.iov_base;
4017 op = le32_to_cpu(h->op);
4018 vino.ino = le64_to_cpu(h->ino);
4019 vino.snap = CEPH_NOSNAP;
4020 seq = le32_to_cpu(h->seq);
4021 mseq = le32_to_cpu(h->migrate_seq);
4024 snaptrace_len = le32_to_cpu(h->snap_trace_len);
4025 p = snaptrace + snaptrace_len;
4027 if (msg_version >= 2) {
4029 ceph_decode_32_safe(&p, end, flock_len, bad);
4030 if (p + flock_len > end)
4035 if (msg_version >= 3) {
4036 if (op == CEPH_CAP_OP_IMPORT) {
4037 if (p + sizeof(*peer) > end)
4041 } else if (op == CEPH_CAP_OP_EXPORT) {
4042 /* recorded in unused fields */
4043 peer = (void *)&h->size;
4047 if (msg_version >= 4) {
4048 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
4049 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
4050 if (p + extra_info.inline_len > end)
4052 extra_info.inline_data = p;
4053 p += extra_info.inline_len;
4056 if (msg_version >= 5) {
4057 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
4060 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
4061 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
4064 if (msg_version >= 8) {
4068 ceph_decode_skip_64(&p, end, bad); // flush_tid
4070 ceph_decode_skip_32(&p, end, bad); // caller_uid
4071 ceph_decode_skip_32(&p, end, bad); // caller_gid
4073 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4074 if (pool_ns_len > 0) {
4075 ceph_decode_need(&p, end, pool_ns_len, bad);
4076 extra_info.pool_ns =
4077 ceph_find_or_create_string(p, pool_ns_len);
4082 if (msg_version >= 9) {
4083 struct ceph_timespec *btime;
4085 if (p + sizeof(*btime) > end)
4088 ceph_decode_timespec64(&extra_info.btime, btime);
4089 p += sizeof(*btime);
4090 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4093 if (msg_version >= 11) {
4095 ceph_decode_skip_32(&p, end, bad); // flags
4097 extra_info.dirstat_valid = true;
4098 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4099 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4103 inode = ceph_find_inode(mdsc->fsc->sb, vino);
4104 ci = ceph_inode(inode);
4105 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4108 mutex_lock(&session->s_mutex);
4109 inc_session_sequence(session);
4110 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4114 dout(" i don't have ino %llx\n", vino.ino);
4116 if (op == CEPH_CAP_OP_IMPORT) {
4117 cap = ceph_get_cap(mdsc, NULL);
4118 cap->cap_ino = vino.ino;
4119 cap->queue_release = 1;
4120 cap->cap_id = le64_to_cpu(h->cap_id);
4123 cap->issue_seq = seq;
4124 spin_lock(&session->s_cap_lock);
4125 __ceph_queue_cap_release(session, cap);
4126 spin_unlock(&session->s_cap_lock);
4128 goto flush_cap_releases;
4131 /* these will work even if we don't have a cap yet */
4133 case CEPH_CAP_OP_FLUSHSNAP_ACK:
4134 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4138 case CEPH_CAP_OP_EXPORT:
4139 handle_cap_export(inode, h, peer, session);
4142 case CEPH_CAP_OP_IMPORT:
4144 if (snaptrace_len) {
4145 down_write(&mdsc->snap_rwsem);
4146 ceph_update_snap_trace(mdsc, snaptrace,
4147 snaptrace + snaptrace_len,
4149 downgrade_write(&mdsc->snap_rwsem);
4151 down_read(&mdsc->snap_rwsem);
4153 spin_lock(&ci->i_ceph_lock);
4154 handle_cap_import(mdsc, inode, h, peer, session,
4155 &cap, &extra_info.issued);
4156 handle_cap_grant(inode, session, cap,
4157 h, msg->middle, &extra_info);
4159 ceph_put_snap_realm(mdsc, realm);
4163 /* the rest require a cap */
4164 spin_lock(&ci->i_ceph_lock);
4165 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4167 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4168 inode, ceph_ino(inode), ceph_snap(inode),
4170 spin_unlock(&ci->i_ceph_lock);
4171 goto flush_cap_releases;
4174 /* note that each of these drops i_ceph_lock for us */
4176 case CEPH_CAP_OP_REVOKE:
4177 case CEPH_CAP_OP_GRANT:
4178 __ceph_caps_issued(ci, &extra_info.issued);
4179 extra_info.issued |= __ceph_caps_dirty(ci);
4180 handle_cap_grant(inode, session, cap,
4181 h, msg->middle, &extra_info);
4184 case CEPH_CAP_OP_FLUSH_ACK:
4185 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4189 case CEPH_CAP_OP_TRUNC:
4190 queue_trunc = handle_cap_trunc(inode, h, session);
4191 spin_unlock(&ci->i_ceph_lock);
4193 ceph_queue_vmtruncate(inode);
4197 spin_unlock(&ci->i_ceph_lock);
4198 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4199 ceph_cap_op_name(op));
4203 mutex_unlock(&session->s_mutex);
4205 ceph_put_string(extra_info.pool_ns);
4206 /* avoid calling iput_final() in mds dispatch threads */
4207 ceph_async_iput(inode);
4212 * send any cap release message to try to move things
4213 * along for the mds (who clearly thinks we still have this
4216 ceph_flush_cap_releases(mdsc, session);
4220 pr_err("ceph_handle_caps: corrupt message\n");
4226 * Delayed work handler to process end of delayed cap release LRU list.
4228 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4230 struct inode *inode;
4231 struct ceph_inode_info *ci;
4233 dout("check_delayed_caps\n");
4234 spin_lock(&mdsc->cap_delay_lock);
4235 while (!list_empty(&mdsc->cap_delay_list)) {
4236 ci = list_first_entry(&mdsc->cap_delay_list,
4237 struct ceph_inode_info,
4239 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4240 time_before(jiffies, ci->i_hold_caps_max))
4242 list_del_init(&ci->i_cap_delay_list);
4244 inode = igrab(&ci->vfs_inode);
4246 spin_unlock(&mdsc->cap_delay_lock);
4247 dout("check_delayed_caps on %p\n", inode);
4248 ceph_check_caps(ci, 0, NULL);
4249 /* avoid calling iput_final() in tick thread */
4250 ceph_async_iput(inode);
4251 spin_lock(&mdsc->cap_delay_lock);
4254 spin_unlock(&mdsc->cap_delay_lock);
4258 * Flush all dirty caps to the mds
4260 static void flush_dirty_session_caps(struct ceph_mds_session *s)
4262 struct ceph_mds_client *mdsc = s->s_mdsc;
4263 struct ceph_inode_info *ci;
4264 struct inode *inode;
4266 dout("flush_dirty_caps\n");
4267 spin_lock(&mdsc->cap_dirty_lock);
4268 while (!list_empty(&s->s_cap_dirty)) {
4269 ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4271 inode = &ci->vfs_inode;
4273 dout("flush_dirty_caps %p\n", inode);
4274 spin_unlock(&mdsc->cap_dirty_lock);
4275 ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4277 spin_lock(&mdsc->cap_dirty_lock);
4279 spin_unlock(&mdsc->cap_dirty_lock);
4280 dout("flush_dirty_caps done\n");
4283 static void iterate_sessions(struct ceph_mds_client *mdsc,
4284 void (*cb)(struct ceph_mds_session *))
4288 mutex_lock(&mdsc->mutex);
4289 for (mds = 0; mds < mdsc->max_sessions; ++mds) {
4290 struct ceph_mds_session *s;
4292 if (!mdsc->sessions[mds])
4295 s = ceph_get_mds_session(mdsc->sessions[mds]);
4299 mutex_unlock(&mdsc->mutex);
4301 ceph_put_mds_session(s);
4302 mutex_lock(&mdsc->mutex);
4304 mutex_unlock(&mdsc->mutex);
4307 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4309 iterate_sessions(mdsc, flush_dirty_session_caps);
4312 void __ceph_touch_fmode(struct ceph_inode_info *ci,
4313 struct ceph_mds_client *mdsc, int fmode)
4315 unsigned long now = jiffies;
4316 if (fmode & CEPH_FILE_MODE_RD)
4317 ci->i_last_rd = now;
4318 if (fmode & CEPH_FILE_MODE_WR)
4319 ci->i_last_wr = now;
4320 /* queue periodic check */
4322 __ceph_is_any_real_caps(ci) &&
4323 list_empty(&ci->i_cap_delay_list))
4324 __cap_delay_requeue(mdsc, ci);
4327 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4329 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4330 int bits = (fmode << 1) | 1;
4331 bool is_opened = false;
4335 atomic64_inc(&mdsc->metric.opened_files);
4337 spin_lock(&ci->i_ceph_lock);
4338 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4339 if (bits & (1 << i))
4340 ci->i_nr_by_mode[i] += count;
4343 * If any of the mode ref is larger than 1,
4344 * that means it has been already opened by
4345 * others. Just skip checking the PIN ref.
4347 if (i && ci->i_nr_by_mode[i] > 1)
4352 percpu_counter_inc(&mdsc->metric.opened_inodes);
4353 spin_unlock(&ci->i_ceph_lock);
4357 * Drop open file reference. If we were the last open file,
4358 * we may need to release capabilities to the MDS (or schedule
4359 * their delayed release).
4361 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4363 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4364 int bits = (fmode << 1) | 1;
4365 bool is_closed = true;
4369 atomic64_dec(&mdsc->metric.opened_files);
4371 spin_lock(&ci->i_ceph_lock);
4372 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4373 if (bits & (1 << i)) {
4374 BUG_ON(ci->i_nr_by_mode[i] < count);
4375 ci->i_nr_by_mode[i] -= count;
4379 * If any of the mode ref is not 0 after
4380 * decreased, that means it is still opened
4381 * by others. Just skip checking the PIN ref.
4383 if (i && ci->i_nr_by_mode[i])
4388 percpu_counter_dec(&mdsc->metric.opened_inodes);
4389 spin_unlock(&ci->i_ceph_lock);
4393 * For a soon-to-be unlinked file, drop the LINK caps. If it
4394 * looks like the link count will hit 0, drop any other caps (other
4395 * than PIN) we don't specifically want (due to the file still being
4398 int ceph_drop_caps_for_unlink(struct inode *inode)
4400 struct ceph_inode_info *ci = ceph_inode(inode);
4401 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4403 spin_lock(&ci->i_ceph_lock);
4404 if (inode->i_nlink == 1) {
4405 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4407 if (__ceph_caps_dirty(ci)) {
4408 struct ceph_mds_client *mdsc =
4409 ceph_inode_to_client(inode)->mdsc;
4410 __cap_delay_requeue_front(mdsc, ci);
4413 spin_unlock(&ci->i_ceph_lock);
4418 * Helpers for embedding cap and dentry lease releases into mds
4421 * @force is used by dentry_release (below) to force inclusion of a
4422 * record for the directory inode, even when there aren't any caps to
4425 int ceph_encode_inode_release(void **p, struct inode *inode,
4426 int mds, int drop, int unless, int force)
4428 struct ceph_inode_info *ci = ceph_inode(inode);
4429 struct ceph_cap *cap;
4430 struct ceph_mds_request_release *rel = *p;
4434 spin_lock(&ci->i_ceph_lock);
4435 used = __ceph_caps_used(ci);
4436 dirty = __ceph_caps_dirty(ci);
4438 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4439 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4440 ceph_cap_string(unless));
4442 /* only drop unused, clean caps */
4443 drop &= ~(used | dirty);
4445 cap = __get_cap_for_mds(ci, mds);
4446 if (cap && __cap_is_valid(cap)) {
4447 unless &= cap->issued;
4449 if (unless & CEPH_CAP_AUTH_EXCL)
4450 drop &= ~CEPH_CAP_AUTH_SHARED;
4451 if (unless & CEPH_CAP_LINK_EXCL)
4452 drop &= ~CEPH_CAP_LINK_SHARED;
4453 if (unless & CEPH_CAP_XATTR_EXCL)
4454 drop &= ~CEPH_CAP_XATTR_SHARED;
4455 if (unless & CEPH_CAP_FILE_EXCL)
4456 drop &= ~CEPH_CAP_FILE_SHARED;
4459 if (force || (cap->issued & drop)) {
4460 if (cap->issued & drop) {
4461 int wanted = __ceph_caps_wanted(ci);
4462 dout("encode_inode_release %p cap %p "
4463 "%s -> %s, wanted %s -> %s\n", inode, cap,
4464 ceph_cap_string(cap->issued),
4465 ceph_cap_string(cap->issued & ~drop),
4466 ceph_cap_string(cap->mds_wanted),
4467 ceph_cap_string(wanted));
4469 cap->issued &= ~drop;
4470 cap->implemented &= ~drop;
4471 cap->mds_wanted = wanted;
4472 if (cap == ci->i_auth_cap &&
4473 !(wanted & CEPH_CAP_ANY_FILE_WR))
4474 ci->i_requested_max_size = 0;
4476 dout("encode_inode_release %p cap %p %s"
4477 " (force)\n", inode, cap,
4478 ceph_cap_string(cap->issued));
4481 rel->ino = cpu_to_le64(ceph_ino(inode));
4482 rel->cap_id = cpu_to_le64(cap->cap_id);
4483 rel->seq = cpu_to_le32(cap->seq);
4484 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4485 rel->mseq = cpu_to_le32(cap->mseq);
4486 rel->caps = cpu_to_le32(cap->implemented);
4487 rel->wanted = cpu_to_le32(cap->mds_wanted);
4493 dout("encode_inode_release %p cap %p %s (noop)\n",
4494 inode, cap, ceph_cap_string(cap->issued));
4497 spin_unlock(&ci->i_ceph_lock);
4501 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4503 int mds, int drop, int unless)
4505 struct dentry *parent = NULL;
4506 struct ceph_mds_request_release *rel = *p;
4507 struct ceph_dentry_info *di = ceph_dentry(dentry);
4512 * force an record for the directory caps if we have a dentry lease.
4513 * this is racy (can't take i_ceph_lock and d_lock together), but it
4514 * doesn't have to be perfect; the mds will revoke anything we don't
4517 spin_lock(&dentry->d_lock);
4518 if (di->lease_session && di->lease_session->s_mds == mds)
4521 parent = dget(dentry->d_parent);
4522 dir = d_inode(parent);
4524 spin_unlock(&dentry->d_lock);
4526 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4529 spin_lock(&dentry->d_lock);
4530 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4531 dout("encode_dentry_release %p mds%d seq %d\n",
4532 dentry, mds, (int)di->lease_seq);
4533 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4534 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4535 *p += dentry->d_name.len;
4536 rel->dname_seq = cpu_to_le32(di->lease_seq);
4537 __ceph_mdsc_drop_dentry_lease(dentry);
4539 spin_unlock(&dentry->d_lock);