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 struct ceph_cap_flush *cf;
1758 cf = kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1759 cf->is_capsnap = false;
1763 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1766 kmem_cache_free(ceph_cap_flush_cachep, cf);
1769 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1771 if (!list_empty(&mdsc->cap_flush_list)) {
1772 struct ceph_cap_flush *cf =
1773 list_first_entry(&mdsc->cap_flush_list,
1774 struct ceph_cap_flush, g_list);
1781 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1782 * Return true if caller needs to wake up flush waiters.
1784 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1785 struct ceph_cap_flush *cf)
1787 struct ceph_cap_flush *prev;
1788 bool wake = cf->wake;
1790 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1791 prev = list_prev_entry(cf, g_list);
1795 list_del_init(&cf->g_list);
1799 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1800 struct ceph_cap_flush *cf)
1802 struct ceph_cap_flush *prev;
1803 bool wake = cf->wake;
1805 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1806 prev = list_prev_entry(cf, i_list);
1810 list_del_init(&cf->i_list);
1815 * Add dirty inode to the flushing list. Assigned a seq number so we
1816 * can wait for caps to flush without starving.
1818 * Called under i_ceph_lock. Returns the flush tid.
1820 static u64 __mark_caps_flushing(struct inode *inode,
1821 struct ceph_mds_session *session, bool wake,
1822 u64 *oldest_flush_tid)
1824 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1825 struct ceph_inode_info *ci = ceph_inode(inode);
1826 struct ceph_cap_flush *cf = NULL;
1829 lockdep_assert_held(&ci->i_ceph_lock);
1830 BUG_ON(ci->i_dirty_caps == 0);
1831 BUG_ON(list_empty(&ci->i_dirty_item));
1832 BUG_ON(!ci->i_prealloc_cap_flush);
1834 flushing = ci->i_dirty_caps;
1835 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1836 ceph_cap_string(flushing),
1837 ceph_cap_string(ci->i_flushing_caps),
1838 ceph_cap_string(ci->i_flushing_caps | flushing));
1839 ci->i_flushing_caps |= flushing;
1840 ci->i_dirty_caps = 0;
1841 dout(" inode %p now !dirty\n", inode);
1843 swap(cf, ci->i_prealloc_cap_flush);
1844 cf->caps = flushing;
1847 spin_lock(&mdsc->cap_dirty_lock);
1848 list_del_init(&ci->i_dirty_item);
1850 cf->tid = ++mdsc->last_cap_flush_tid;
1851 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1852 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1854 if (list_empty(&ci->i_flushing_item)) {
1855 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1856 mdsc->num_cap_flushing++;
1858 spin_unlock(&mdsc->cap_dirty_lock);
1860 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1866 * try to invalidate mapping pages without blocking.
1868 static int try_nonblocking_invalidate(struct inode *inode)
1870 struct ceph_inode_info *ci = ceph_inode(inode);
1871 u32 invalidating_gen = ci->i_rdcache_gen;
1873 spin_unlock(&ci->i_ceph_lock);
1874 ceph_fscache_invalidate(inode);
1875 invalidate_mapping_pages(&inode->i_data, 0, -1);
1876 spin_lock(&ci->i_ceph_lock);
1878 if (inode->i_data.nrpages == 0 &&
1879 invalidating_gen == ci->i_rdcache_gen) {
1881 dout("try_nonblocking_invalidate %p success\n", inode);
1882 /* save any racing async invalidate some trouble */
1883 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1886 dout("try_nonblocking_invalidate %p failed\n", inode);
1890 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1892 loff_t size = i_size_read(&ci->vfs_inode);
1893 /* mds will adjust max size according to the reported size */
1894 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1896 if (size >= ci->i_max_size)
1898 /* half of previous max_size increment has been used */
1899 if (ci->i_max_size > ci->i_reported_size &&
1900 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1906 * Swiss army knife function to examine currently used and wanted
1907 * versus held caps. Release, flush, ack revoked caps to mds as
1910 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1911 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1914 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1915 struct ceph_mds_session *session)
1917 struct inode *inode = &ci->vfs_inode;
1918 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1919 struct ceph_cap *cap;
1920 u64 flush_tid, oldest_flush_tid;
1921 int file_wanted, used, cap_used;
1922 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1923 int issued, implemented, want, retain, revoking, flushing = 0;
1924 int mds = -1; /* keep track of how far we've gone through i_caps list
1925 to avoid an infinite loop on retry */
1927 bool queue_invalidate = false;
1928 bool tried_invalidate = false;
1930 spin_lock(&ci->i_ceph_lock);
1931 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1932 flags |= CHECK_CAPS_FLUSH;
1936 spin_lock(&ci->i_ceph_lock);
1938 /* Caps wanted by virtue of active open files. */
1939 file_wanted = __ceph_caps_file_wanted(ci);
1941 /* Caps which have active references against them */
1942 used = __ceph_caps_used(ci);
1945 * "issued" represents the current caps that the MDS wants us to have.
1946 * "implemented" is the set that we have been granted, and includes the
1947 * ones that have not yet been returned to the MDS (the "revoking" set,
1948 * usually because they have outstanding references).
1950 issued = __ceph_caps_issued(ci, &implemented);
1951 revoking = implemented & ~issued;
1955 /* The ones we currently want to retain (may be adjusted below) */
1956 retain = file_wanted | used | CEPH_CAP_PIN;
1957 if (!mdsc->stopping && inode->i_nlink > 0) {
1959 retain |= CEPH_CAP_ANY; /* be greedy */
1960 } else if (S_ISDIR(inode->i_mode) &&
1961 (issued & CEPH_CAP_FILE_SHARED) &&
1962 __ceph_dir_is_complete(ci)) {
1964 * If a directory is complete, we want to keep
1965 * the exclusive cap. So that MDS does not end up
1966 * revoking the shared cap on every create/unlink
1969 if (IS_RDONLY(inode)) {
1970 want = CEPH_CAP_ANY_SHARED;
1972 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1977 retain |= CEPH_CAP_ANY_SHARED;
1979 * keep RD only if we didn't have the file open RW,
1980 * because then the mds would revoke it anyway to
1981 * journal max_size=0.
1983 if (ci->i_max_size == 0)
1984 retain |= CEPH_CAP_ANY_RD;
1988 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1989 " issued %s revoking %s retain %s %s%s\n", inode,
1990 ceph_cap_string(file_wanted),
1991 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1992 ceph_cap_string(ci->i_flushing_caps),
1993 ceph_cap_string(issued), ceph_cap_string(revoking),
1994 ceph_cap_string(retain),
1995 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1996 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1999 * If we no longer need to hold onto old our caps, and we may
2000 * have cached pages, but don't want them, then try to invalidate.
2001 * If we fail, it's because pages are locked.... try again later.
2003 if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
2004 S_ISREG(inode->i_mode) &&
2005 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
2006 inode->i_data.nrpages && /* have cached pages */
2007 (revoking & (CEPH_CAP_FILE_CACHE|
2008 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
2009 !tried_invalidate) {
2010 dout("check_caps trying to invalidate on %p\n", inode);
2011 if (try_nonblocking_invalidate(inode) < 0) {
2012 dout("check_caps queuing invalidate\n");
2013 queue_invalidate = true;
2014 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2016 tried_invalidate = true;
2020 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2022 struct cap_msg_args arg;
2024 cap = rb_entry(p, struct ceph_cap, ci_node);
2026 /* avoid looping forever */
2027 if (mds >= cap->mds ||
2028 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
2031 /* NOTE: no side-effects allowed, until we take s_mutex */
2034 * If we have an auth cap, we don't need to consider any
2035 * overlapping caps as used.
2038 if (ci->i_auth_cap && cap != ci->i_auth_cap)
2039 cap_used &= ~ci->i_auth_cap->issued;
2041 revoking = cap->implemented & ~cap->issued;
2042 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
2043 cap->mds, cap, ceph_cap_string(cap_used),
2044 ceph_cap_string(cap->issued),
2045 ceph_cap_string(cap->implemented),
2046 ceph_cap_string(revoking));
2048 if (cap == ci->i_auth_cap &&
2049 (cap->issued & CEPH_CAP_FILE_WR)) {
2050 /* request larger max_size from MDS? */
2051 if (ci->i_wanted_max_size > ci->i_max_size &&
2052 ci->i_wanted_max_size > ci->i_requested_max_size) {
2053 dout("requesting new max_size\n");
2057 /* approaching file_max? */
2058 if (__ceph_should_report_size(ci)) {
2059 dout("i_size approaching max_size\n");
2063 /* flush anything dirty? */
2064 if (cap == ci->i_auth_cap) {
2065 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
2066 dout("flushing dirty caps\n");
2069 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
2070 dout("flushing snap caps\n");
2075 /* completed revocation? going down and there are no caps? */
2076 if (revoking && (revoking & cap_used) == 0) {
2077 dout("completed revocation of %s\n",
2078 ceph_cap_string(cap->implemented & ~cap->issued));
2082 /* want more caps from mds? */
2083 if (want & ~cap->mds_wanted) {
2084 if (want & ~(cap->mds_wanted | cap->issued))
2086 if (!__cap_is_valid(cap))
2090 /* things we might delay */
2091 if ((cap->issued & ~retain) == 0)
2092 continue; /* nope, all good */
2095 if (session && session != cap->session) {
2096 dout("oops, wrong session %p mutex\n", session);
2097 mutex_unlock(&session->s_mutex);
2101 session = cap->session;
2102 if (mutex_trylock(&session->s_mutex) == 0) {
2103 dout("inverting session/ino locks on %p\n",
2105 session = ceph_get_mds_session(session);
2106 spin_unlock(&ci->i_ceph_lock);
2107 if (took_snap_rwsem) {
2108 up_read(&mdsc->snap_rwsem);
2109 took_snap_rwsem = 0;
2112 mutex_lock(&session->s_mutex);
2113 ceph_put_mds_session(session);
2116 * Because we take the reference while
2117 * holding the i_ceph_lock, it should
2118 * never be NULL. Throw a warning if it
2127 /* kick flushing and flush snaps before sending normal
2129 if (cap == ci->i_auth_cap &&
2131 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2132 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2133 __kick_flushing_caps(mdsc, session, ci, 0);
2134 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2135 __ceph_flush_snaps(ci, session);
2140 /* take snap_rwsem after session mutex */
2141 if (!took_snap_rwsem) {
2142 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
2143 dout("inverting snap/in locks on %p\n",
2145 spin_unlock(&ci->i_ceph_lock);
2146 down_read(&mdsc->snap_rwsem);
2147 took_snap_rwsem = 1;
2150 took_snap_rwsem = 1;
2153 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2154 flushing = ci->i_dirty_caps;
2155 flush_tid = __mark_caps_flushing(inode, session, false,
2157 if (flags & CHECK_CAPS_FLUSH &&
2158 list_empty(&session->s_cap_dirty))
2159 mflags |= CEPH_CLIENT_CAPS_SYNC;
2163 spin_lock(&mdsc->cap_dirty_lock);
2164 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2165 spin_unlock(&mdsc->cap_dirty_lock);
2168 mds = cap->mds; /* remember mds, so we don't repeat */
2170 __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2171 want, retain, flushing, flush_tid, oldest_flush_tid);
2172 spin_unlock(&ci->i_ceph_lock);
2174 __send_cap(&arg, ci);
2176 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2179 /* periodically re-calculate caps wanted by open files */
2180 if (__ceph_is_any_real_caps(ci) &&
2181 list_empty(&ci->i_cap_delay_list) &&
2182 (file_wanted & ~CEPH_CAP_PIN) &&
2183 !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2184 __cap_delay_requeue(mdsc, ci);
2187 spin_unlock(&ci->i_ceph_lock);
2189 if (queue_invalidate)
2190 ceph_queue_invalidate(inode);
2193 mutex_unlock(&session->s_mutex);
2194 if (took_snap_rwsem)
2195 up_read(&mdsc->snap_rwsem);
2199 * Try to flush dirty caps back to the auth mds.
2201 static int try_flush_caps(struct inode *inode, u64 *ptid)
2203 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2204 struct ceph_inode_info *ci = ceph_inode(inode);
2205 struct ceph_mds_session *session = NULL;
2207 u64 flush_tid = 0, oldest_flush_tid = 0;
2210 spin_lock(&ci->i_ceph_lock);
2212 if (ci->i_dirty_caps && ci->i_auth_cap) {
2213 struct ceph_cap *cap = ci->i_auth_cap;
2214 struct cap_msg_args arg;
2216 if (session != cap->session) {
2217 spin_unlock(&ci->i_ceph_lock);
2219 mutex_unlock(&session->s_mutex);
2220 session = cap->session;
2221 mutex_lock(&session->s_mutex);
2224 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) {
2225 spin_unlock(&ci->i_ceph_lock);
2229 if (ci->i_ceph_flags &
2230 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2231 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2232 __kick_flushing_caps(mdsc, session, ci, 0);
2233 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2234 __ceph_flush_snaps(ci, session);
2238 flushing = ci->i_dirty_caps;
2239 flush_tid = __mark_caps_flushing(inode, session, true,
2242 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2243 __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2244 (cap->issued | cap->implemented),
2245 flushing, flush_tid, oldest_flush_tid);
2246 spin_unlock(&ci->i_ceph_lock);
2248 __send_cap(&arg, ci);
2250 if (!list_empty(&ci->i_cap_flush_list)) {
2251 struct ceph_cap_flush *cf =
2252 list_last_entry(&ci->i_cap_flush_list,
2253 struct ceph_cap_flush, i_list);
2255 flush_tid = cf->tid;
2257 flushing = ci->i_flushing_caps;
2258 spin_unlock(&ci->i_ceph_lock);
2262 mutex_unlock(&session->s_mutex);
2269 * Return true if we've flushed caps through the given flush_tid.
2271 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2273 struct ceph_inode_info *ci = ceph_inode(inode);
2276 spin_lock(&ci->i_ceph_lock);
2277 if (!list_empty(&ci->i_cap_flush_list)) {
2278 struct ceph_cap_flush * cf =
2279 list_first_entry(&ci->i_cap_flush_list,
2280 struct ceph_cap_flush, i_list);
2281 if (cf->tid <= flush_tid)
2284 spin_unlock(&ci->i_ceph_lock);
2289 * wait for any unsafe requests to complete.
2291 static int unsafe_request_wait(struct inode *inode)
2293 struct ceph_inode_info *ci = ceph_inode(inode);
2294 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2297 spin_lock(&ci->i_unsafe_lock);
2298 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2299 req1 = list_last_entry(&ci->i_unsafe_dirops,
2300 struct ceph_mds_request,
2302 ceph_mdsc_get_request(req1);
2304 if (!list_empty(&ci->i_unsafe_iops)) {
2305 req2 = list_last_entry(&ci->i_unsafe_iops,
2306 struct ceph_mds_request,
2307 r_unsafe_target_item);
2308 ceph_mdsc_get_request(req2);
2310 spin_unlock(&ci->i_unsafe_lock);
2312 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2313 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2315 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2316 ceph_timeout_jiffies(req1->r_timeout));
2319 ceph_mdsc_put_request(req1);
2322 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2323 ceph_timeout_jiffies(req2->r_timeout));
2326 ceph_mdsc_put_request(req2);
2331 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2333 struct ceph_file_info *fi = file->private_data;
2334 struct inode *inode = file->f_mapping->host;
2335 struct ceph_inode_info *ci = ceph_inode(inode);
2340 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2342 ret = file_write_and_wait_range(file, start, end);
2346 ret = ceph_wait_on_async_create(inode);
2350 dirty = try_flush_caps(inode, &flush_tid);
2351 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2353 err = unsafe_request_wait(inode);
2356 * only wait on non-file metadata writeback (the mds
2357 * can recover size and mtime, so we don't need to
2360 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2361 err = wait_event_interruptible(ci->i_cap_wq,
2362 caps_are_flushed(inode, flush_tid));
2368 if (errseq_check(&ci->i_meta_err, READ_ONCE(fi->meta_err))) {
2369 spin_lock(&file->f_lock);
2370 err = errseq_check_and_advance(&ci->i_meta_err,
2372 spin_unlock(&file->f_lock);
2377 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2382 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2383 * queue inode for flush but don't do so immediately, because we can
2384 * get by with fewer MDS messages if we wait for data writeback to
2387 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2389 struct ceph_inode_info *ci = ceph_inode(inode);
2393 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2395 dout("write_inode %p wait=%d\n", inode, wait);
2397 dirty = try_flush_caps(inode, &flush_tid);
2399 err = wait_event_interruptible(ci->i_cap_wq,
2400 caps_are_flushed(inode, flush_tid));
2402 struct ceph_mds_client *mdsc =
2403 ceph_sb_to_client(inode->i_sb)->mdsc;
2405 spin_lock(&ci->i_ceph_lock);
2406 if (__ceph_caps_dirty(ci))
2407 __cap_delay_requeue_front(mdsc, ci);
2408 spin_unlock(&ci->i_ceph_lock);
2413 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2414 struct ceph_mds_session *session,
2415 struct ceph_inode_info *ci,
2416 u64 oldest_flush_tid)
2417 __releases(ci->i_ceph_lock)
2418 __acquires(ci->i_ceph_lock)
2420 struct inode *inode = &ci->vfs_inode;
2421 struct ceph_cap *cap;
2422 struct ceph_cap_flush *cf;
2425 u64 last_snap_flush = 0;
2427 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2429 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2430 if (cf->is_capsnap) {
2431 last_snap_flush = cf->tid;
2436 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2437 if (cf->tid < first_tid)
2440 cap = ci->i_auth_cap;
2441 if (!(cap && cap->session == session)) {
2442 pr_err("%p auth cap %p not mds%d ???\n",
2443 inode, cap, session->s_mds);
2447 first_tid = cf->tid + 1;
2449 if (!cf->is_capsnap) {
2450 struct cap_msg_args arg;
2452 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2453 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2454 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2455 (cf->tid < last_snap_flush ?
2456 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2457 __ceph_caps_used(ci),
2458 __ceph_caps_wanted(ci),
2459 (cap->issued | cap->implemented),
2460 cf->caps, cf->tid, oldest_flush_tid);
2461 spin_unlock(&ci->i_ceph_lock);
2462 __send_cap(&arg, ci);
2464 struct ceph_cap_snap *capsnap =
2465 container_of(cf, struct ceph_cap_snap,
2467 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2468 inode, capsnap, cf->tid,
2469 ceph_cap_string(capsnap->dirty));
2471 refcount_inc(&capsnap->nref);
2472 spin_unlock(&ci->i_ceph_lock);
2474 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2477 pr_err("kick_flushing_caps: error sending "
2478 "cap flushsnap, ino (%llx.%llx) "
2479 "tid %llu follows %llu\n",
2480 ceph_vinop(inode), cf->tid,
2484 ceph_put_cap_snap(capsnap);
2487 spin_lock(&ci->i_ceph_lock);
2491 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2492 struct ceph_mds_session *session)
2494 struct ceph_inode_info *ci;
2495 struct ceph_cap *cap;
2496 u64 oldest_flush_tid;
2498 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2500 spin_lock(&mdsc->cap_dirty_lock);
2501 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2502 spin_unlock(&mdsc->cap_dirty_lock);
2504 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2505 spin_lock(&ci->i_ceph_lock);
2506 cap = ci->i_auth_cap;
2507 if (!(cap && cap->session == session)) {
2508 pr_err("%p auth cap %p not mds%d ???\n",
2509 &ci->vfs_inode, cap, session->s_mds);
2510 spin_unlock(&ci->i_ceph_lock);
2516 * if flushing caps were revoked, we re-send the cap flush
2517 * in client reconnect stage. This guarantees MDS * processes
2518 * the cap flush message before issuing the flushing caps to
2521 if ((cap->issued & ci->i_flushing_caps) !=
2522 ci->i_flushing_caps) {
2523 /* encode_caps_cb() also will reset these sequence
2524 * numbers. make sure sequence numbers in cap flush
2525 * message match later reconnect message */
2529 __kick_flushing_caps(mdsc, session, ci,
2532 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2535 spin_unlock(&ci->i_ceph_lock);
2539 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2540 struct ceph_mds_session *session)
2542 struct ceph_inode_info *ci;
2543 struct ceph_cap *cap;
2544 u64 oldest_flush_tid;
2546 lockdep_assert_held(&session->s_mutex);
2548 dout("kick_flushing_caps mds%d\n", session->s_mds);
2550 spin_lock(&mdsc->cap_dirty_lock);
2551 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2552 spin_unlock(&mdsc->cap_dirty_lock);
2554 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2555 spin_lock(&ci->i_ceph_lock);
2556 cap = ci->i_auth_cap;
2557 if (!(cap && cap->session == session)) {
2558 pr_err("%p auth cap %p not mds%d ???\n",
2559 &ci->vfs_inode, cap, session->s_mds);
2560 spin_unlock(&ci->i_ceph_lock);
2563 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2564 __kick_flushing_caps(mdsc, session, ci,
2567 spin_unlock(&ci->i_ceph_lock);
2571 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2572 struct ceph_inode_info *ci)
2574 struct ceph_mds_client *mdsc = session->s_mdsc;
2575 struct ceph_cap *cap = ci->i_auth_cap;
2577 lockdep_assert_held(&ci->i_ceph_lock);
2579 dout("%s %p flushing %s\n", __func__, &ci->vfs_inode,
2580 ceph_cap_string(ci->i_flushing_caps));
2582 if (!list_empty(&ci->i_cap_flush_list)) {
2583 u64 oldest_flush_tid;
2584 spin_lock(&mdsc->cap_dirty_lock);
2585 list_move_tail(&ci->i_flushing_item,
2586 &cap->session->s_cap_flushing);
2587 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2588 spin_unlock(&mdsc->cap_dirty_lock);
2590 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2596 * Take references to capabilities we hold, so that we don't release
2597 * them to the MDS prematurely.
2599 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2600 bool snap_rwsem_locked)
2602 lockdep_assert_held(&ci->i_ceph_lock);
2604 if (got & CEPH_CAP_PIN)
2606 if (got & CEPH_CAP_FILE_RD)
2608 if (got & CEPH_CAP_FILE_CACHE)
2609 ci->i_rdcache_ref++;
2610 if (got & CEPH_CAP_FILE_EXCL)
2612 if (got & CEPH_CAP_FILE_WR) {
2613 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2614 BUG_ON(!snap_rwsem_locked);
2615 ci->i_head_snapc = ceph_get_snap_context(
2616 ci->i_snap_realm->cached_context);
2620 if (got & CEPH_CAP_FILE_BUFFER) {
2621 if (ci->i_wb_ref == 0)
2622 ihold(&ci->vfs_inode);
2624 dout("%s %p wb %d -> %d (?)\n", __func__,
2625 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2630 * Try to grab cap references. Specify those refs we @want, and the
2631 * minimal set we @need. Also include the larger offset we are writing
2632 * to (when applicable), and check against max_size here as well.
2633 * Note that caller is responsible for ensuring max_size increases are
2634 * requested from the MDS.
2636 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2637 * or a negative error code. There are 3 speical error codes:
2638 * -EAGAIN: need to sleep but non-blocking is specified
2639 * -EFBIG: ask caller to call check_max_size() and try again.
2640 * -ESTALE: ask caller to call ceph_renew_caps() and try again.
2643 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2644 NON_BLOCKING = (1 << 8),
2645 CHECK_FILELOCK = (1 << 9),
2648 static int try_get_cap_refs(struct inode *inode, int need, int want,
2649 loff_t endoff, int flags, int *got)
2651 struct ceph_inode_info *ci = ceph_inode(inode);
2652 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2654 int have, implemented;
2655 bool snap_rwsem_locked = false;
2657 dout("get_cap_refs %p need %s want %s\n", inode,
2658 ceph_cap_string(need), ceph_cap_string(want));
2661 spin_lock(&ci->i_ceph_lock);
2663 if ((flags & CHECK_FILELOCK) &&
2664 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2665 dout("try_get_cap_refs %p error filelock\n", inode);
2670 /* finish pending truncate */
2671 while (ci->i_truncate_pending) {
2672 spin_unlock(&ci->i_ceph_lock);
2673 if (snap_rwsem_locked) {
2674 up_read(&mdsc->snap_rwsem);
2675 snap_rwsem_locked = false;
2677 __ceph_do_pending_vmtruncate(inode);
2678 spin_lock(&ci->i_ceph_lock);
2681 have = __ceph_caps_issued(ci, &implemented);
2683 if (have & need & CEPH_CAP_FILE_WR) {
2684 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2685 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2686 inode, endoff, ci->i_max_size);
2687 if (endoff > ci->i_requested_max_size)
2688 ret = ci->i_auth_cap ? -EFBIG : -ESTALE;
2692 * If a sync write is in progress, we must wait, so that we
2693 * can get a final snapshot value for size+mtime.
2695 if (__ceph_have_pending_cap_snap(ci)) {
2696 dout("get_cap_refs %p cap_snap_pending\n", inode);
2701 if ((have & need) == need) {
2703 * Look at (implemented & ~have & not) so that we keep waiting
2704 * on transition from wanted -> needed caps. This is needed
2705 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2706 * going before a prior buffered writeback happens.
2708 int not = want & ~(have & need);
2709 int revoking = implemented & ~have;
2710 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2711 inode, ceph_cap_string(have), ceph_cap_string(not),
2712 ceph_cap_string(revoking));
2713 if ((revoking & not) == 0) {
2714 if (!snap_rwsem_locked &&
2715 !ci->i_head_snapc &&
2716 (need & CEPH_CAP_FILE_WR)) {
2717 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2719 * we can not call down_read() when
2720 * task isn't in TASK_RUNNING state
2722 if (flags & NON_BLOCKING) {
2727 spin_unlock(&ci->i_ceph_lock);
2728 down_read(&mdsc->snap_rwsem);
2729 snap_rwsem_locked = true;
2732 snap_rwsem_locked = true;
2734 if ((have & want) == want)
2738 ceph_take_cap_refs(ci, *got, true);
2742 int session_readonly = false;
2744 if (ci->i_auth_cap &&
2745 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2746 struct ceph_mds_session *s = ci->i_auth_cap->session;
2747 spin_lock(&s->s_cap_lock);
2748 session_readonly = s->s_readonly;
2749 spin_unlock(&s->s_cap_lock);
2751 if (session_readonly) {
2752 dout("get_cap_refs %p need %s but mds%d readonly\n",
2753 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2758 if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
2759 dout("get_cap_refs %p forced umount\n", inode);
2763 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2764 if (need & ~mds_wanted) {
2765 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2766 inode, ceph_cap_string(need),
2767 ceph_cap_string(mds_wanted));
2772 dout("get_cap_refs %p have %s need %s\n", inode,
2773 ceph_cap_string(have), ceph_cap_string(need));
2777 __ceph_touch_fmode(ci, mdsc, flags);
2779 spin_unlock(&ci->i_ceph_lock);
2780 if (snap_rwsem_locked)
2781 up_read(&mdsc->snap_rwsem);
2784 ceph_update_cap_mis(&mdsc->metric);
2786 ceph_update_cap_hit(&mdsc->metric);
2788 dout("get_cap_refs %p ret %d got %s\n", inode,
2789 ret, ceph_cap_string(*got));
2794 * Check the offset we are writing up to against our current
2795 * max_size. If necessary, tell the MDS we want to write to
2798 static void check_max_size(struct inode *inode, loff_t endoff)
2800 struct ceph_inode_info *ci = ceph_inode(inode);
2803 /* do we need to explicitly request a larger max_size? */
2804 spin_lock(&ci->i_ceph_lock);
2805 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2806 dout("write %p at large endoff %llu, req max_size\n",
2808 ci->i_wanted_max_size = endoff;
2810 /* duplicate ceph_check_caps()'s logic */
2811 if (ci->i_auth_cap &&
2812 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2813 ci->i_wanted_max_size > ci->i_max_size &&
2814 ci->i_wanted_max_size > ci->i_requested_max_size)
2816 spin_unlock(&ci->i_ceph_lock);
2818 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2821 static inline int get_used_fmode(int caps)
2824 if (caps & CEPH_CAP_FILE_RD)
2825 fmode |= CEPH_FILE_MODE_RD;
2826 if (caps & CEPH_CAP_FILE_WR)
2827 fmode |= CEPH_FILE_MODE_WR;
2831 int ceph_try_get_caps(struct inode *inode, int need, int want,
2832 bool nonblock, int *got)
2836 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2837 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2838 CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2839 CEPH_CAP_ANY_DIR_OPS));
2841 ret = ceph_pool_perm_check(inode, need);
2846 flags = get_used_fmode(need | want);
2848 flags |= NON_BLOCKING;
2850 ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2851 /* three special error codes */
2852 if (ret == -EAGAIN || ret == -EFBIG || ret == -ESTALE)
2858 * Wait for caps, and take cap references. If we can't get a WR cap
2859 * due to a small max_size, make sure we check_max_size (and possibly
2860 * ask the mds) so we don't get hung up indefinitely.
2862 int ceph_get_caps(struct file *filp, int need, int want, loff_t endoff, int *got)
2864 struct ceph_file_info *fi = filp->private_data;
2865 struct inode *inode = file_inode(filp);
2866 struct ceph_inode_info *ci = ceph_inode(inode);
2867 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2868 int ret, _got, flags;
2870 ret = ceph_pool_perm_check(inode, need);
2874 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2875 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2878 flags = get_used_fmode(need | want);
2881 flags &= CEPH_FILE_MODE_MASK;
2882 if (atomic_read(&fi->num_locks))
2883 flags |= CHECK_FILELOCK;
2885 ret = try_get_cap_refs(inode, need, want, endoff,
2887 WARN_ON_ONCE(ret == -EAGAIN);
2889 struct ceph_mds_client *mdsc = fsc->mdsc;
2891 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2893 cw.ino = ceph_ino(inode);
2894 cw.tgid = current->tgid;
2898 spin_lock(&mdsc->caps_list_lock);
2899 list_add(&cw.list, &mdsc->cap_wait_list);
2900 spin_unlock(&mdsc->caps_list_lock);
2902 /* make sure used fmode not timeout */
2903 ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2904 add_wait_queue(&ci->i_cap_wq, &wait);
2906 flags |= NON_BLOCKING;
2907 while (!(ret = try_get_cap_refs(inode, need, want,
2908 endoff, flags, &_got))) {
2909 if (signal_pending(current)) {
2913 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2916 remove_wait_queue(&ci->i_cap_wq, &wait);
2917 ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
2919 spin_lock(&mdsc->caps_list_lock);
2921 spin_unlock(&mdsc->caps_list_lock);
2927 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2928 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2929 if (ret >= 0 && _got)
2930 ceph_put_cap_refs(ci, _got);
2935 if (ret == -EFBIG || ret == -ESTALE) {
2936 int ret2 = ceph_wait_on_async_create(inode);
2940 if (ret == -EFBIG) {
2941 check_max_size(inode, endoff);
2944 if (ret == -ESTALE) {
2945 /* session was killed, try renew caps */
2946 ret = ceph_renew_caps(inode, flags);
2953 if (S_ISREG(ci->vfs_inode.i_mode) &&
2954 ci->i_inline_version != CEPH_INLINE_NONE &&
2955 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2956 i_size_read(inode) > 0) {
2958 find_get_page(inode->i_mapping, 0);
2960 bool uptodate = PageUptodate(page);
2967 * drop cap refs first because getattr while
2968 * holding * caps refs can cause deadlock.
2970 ceph_put_cap_refs(ci, _got);
2974 * getattr request will bring inline data into
2977 ret = __ceph_do_getattr(inode, NULL,
2978 CEPH_STAT_CAP_INLINE_DATA,
2991 * Take cap refs. Caller must already know we hold at least one ref
2992 * on the caps in question or we don't know this is safe.
2994 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2996 spin_lock(&ci->i_ceph_lock);
2997 ceph_take_cap_refs(ci, caps, false);
2998 spin_unlock(&ci->i_ceph_lock);
3003 * drop cap_snap that is not associated with any snapshot.
3004 * we don't need to send FLUSHSNAP message for it.
3006 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
3007 struct ceph_cap_snap *capsnap)
3009 if (!capsnap->need_flush &&
3010 !capsnap->writing && !capsnap->dirty_pages) {
3011 dout("dropping cap_snap %p follows %llu\n",
3012 capsnap, capsnap->follows);
3013 BUG_ON(capsnap->cap_flush.tid > 0);
3014 ceph_put_snap_context(capsnap->context);
3015 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
3016 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3018 list_del(&capsnap->ci_item);
3019 ceph_put_cap_snap(capsnap);
3025 enum put_cap_refs_mode {
3026 PUT_CAP_REFS_SYNC = 0,
3027 PUT_CAP_REFS_NO_CHECK,
3034 * If we released the last ref on any given cap, call ceph_check_caps
3035 * to release (or schedule a release).
3037 * If we are releasing a WR cap (from a sync write), finalize any affected
3038 * cap_snap, and wake up any waiters.
3040 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
3041 enum put_cap_refs_mode mode)
3043 struct inode *inode = &ci->vfs_inode;
3044 int last = 0, put = 0, flushsnaps = 0, wake = 0;
3045 bool check_flushsnaps = false;
3047 spin_lock(&ci->i_ceph_lock);
3048 if (had & CEPH_CAP_PIN)
3050 if (had & CEPH_CAP_FILE_RD)
3051 if (--ci->i_rd_ref == 0)
3053 if (had & CEPH_CAP_FILE_CACHE)
3054 if (--ci->i_rdcache_ref == 0)
3056 if (had & CEPH_CAP_FILE_EXCL)
3057 if (--ci->i_fx_ref == 0)
3059 if (had & CEPH_CAP_FILE_BUFFER) {
3060 if (--ci->i_wb_ref == 0) {
3062 /* put the ref held by ceph_take_cap_refs() */
3064 check_flushsnaps = true;
3066 dout("put_cap_refs %p wb %d -> %d (?)\n",
3067 inode, ci->i_wb_ref+1, ci->i_wb_ref);
3069 if (had & CEPH_CAP_FILE_WR) {
3070 if (--ci->i_wr_ref == 0) {
3072 check_flushsnaps = true;
3073 if (ci->i_wrbuffer_ref_head == 0 &&
3074 ci->i_dirty_caps == 0 &&
3075 ci->i_flushing_caps == 0) {
3076 BUG_ON(!ci->i_head_snapc);
3077 ceph_put_snap_context(ci->i_head_snapc);
3078 ci->i_head_snapc = NULL;
3080 /* see comment in __ceph_remove_cap() */
3081 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3082 drop_inode_snap_realm(ci);
3085 if (check_flushsnaps && __ceph_have_pending_cap_snap(ci)) {
3086 struct ceph_cap_snap *capsnap =
3087 list_last_entry(&ci->i_cap_snaps,
3088 struct ceph_cap_snap,
3091 capsnap->writing = 0;
3092 if (ceph_try_drop_cap_snap(ci, capsnap))
3093 /* put the ref held by ceph_queue_cap_snap() */
3095 else if (__ceph_finish_cap_snap(ci, capsnap))
3099 spin_unlock(&ci->i_ceph_lock);
3101 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3102 last ? " last" : "", put ? " put" : "");
3105 case PUT_CAP_REFS_SYNC:
3107 ceph_check_caps(ci, 0, NULL);
3108 else if (flushsnaps)
3109 ceph_flush_snaps(ci, NULL);
3111 case PUT_CAP_REFS_ASYNC:
3113 ceph_queue_check_caps(inode);
3114 else if (flushsnaps)
3115 ceph_queue_flush_snaps(inode);
3121 wake_up_all(&ci->i_cap_wq);
3126 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3128 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_SYNC);
3131 void ceph_put_cap_refs_async(struct ceph_inode_info *ci, int had)
3133 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_ASYNC);
3136 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3138 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_NO_CHECK);
3142 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3143 * context. Adjust per-snap dirty page accounting as appropriate.
3144 * Once all dirty data for a cap_snap is flushed, flush snapped file
3145 * metadata back to the MDS. If we dropped the last ref, call
3148 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3149 struct ceph_snap_context *snapc)
3151 struct inode *inode = &ci->vfs_inode;
3152 struct ceph_cap_snap *capsnap = NULL;
3156 bool flush_snaps = false;
3157 bool complete_capsnap = false;
3159 spin_lock(&ci->i_ceph_lock);
3160 ci->i_wrbuffer_ref -= nr;
3161 if (ci->i_wrbuffer_ref == 0) {
3166 if (ci->i_head_snapc == snapc) {
3167 ci->i_wrbuffer_ref_head -= nr;
3168 if (ci->i_wrbuffer_ref_head == 0 &&
3169 ci->i_wr_ref == 0 &&
3170 ci->i_dirty_caps == 0 &&
3171 ci->i_flushing_caps == 0) {
3172 BUG_ON(!ci->i_head_snapc);
3173 ceph_put_snap_context(ci->i_head_snapc);
3174 ci->i_head_snapc = NULL;
3176 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3178 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3179 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3180 last ? " LAST" : "");
3182 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3183 if (capsnap->context == snapc) {
3189 capsnap->dirty_pages -= nr;
3190 if (capsnap->dirty_pages == 0) {
3191 complete_capsnap = true;
3192 if (!capsnap->writing) {
3193 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3196 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3201 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3202 " snap %lld %d/%d -> %d/%d %s%s\n",
3203 inode, capsnap, capsnap->context->seq,
3204 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3205 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3206 last ? " (wrbuffer last)" : "",
3207 complete_capsnap ? " (complete capsnap)" : "");
3210 spin_unlock(&ci->i_ceph_lock);
3213 ceph_check_caps(ci, 0, NULL);
3214 } else if (flush_snaps) {
3215 ceph_flush_snaps(ci, NULL);
3217 if (complete_capsnap)
3218 wake_up_all(&ci->i_cap_wq);
3220 /* avoid calling iput_final() in osd dispatch threads */
3221 ceph_async_iput(inode);
3226 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3228 static void invalidate_aliases(struct inode *inode)
3230 struct dentry *dn, *prev = NULL;
3232 dout("invalidate_aliases inode %p\n", inode);
3233 d_prune_aliases(inode);
3235 * For non-directory inode, d_find_alias() only returns
3236 * hashed dentry. After calling d_invalidate(), the
3237 * dentry becomes unhashed.
3239 * For directory inode, d_find_alias() can return
3240 * unhashed dentry. But directory inode should have
3241 * one alias at most.
3243 while ((dn = d_find_alias(inode))) {
3257 struct cap_extra_info {
3258 struct ceph_string *pool_ns;
3268 /* currently issued */
3270 struct timespec64 btime;
3274 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3275 * actually be a revocation if it specifies a smaller cap set.)
3277 * caller holds s_mutex and i_ceph_lock, we drop both.
3279 static void handle_cap_grant(struct inode *inode,
3280 struct ceph_mds_session *session,
3281 struct ceph_cap *cap,
3282 struct ceph_mds_caps *grant,
3283 struct ceph_buffer *xattr_buf,
3284 struct cap_extra_info *extra_info)
3285 __releases(ci->i_ceph_lock)
3286 __releases(session->s_mdsc->snap_rwsem)
3288 struct ceph_inode_info *ci = ceph_inode(inode);
3289 int seq = le32_to_cpu(grant->seq);
3290 int newcaps = le32_to_cpu(grant->caps);
3291 int used, wanted, dirty;
3292 u64 size = le64_to_cpu(grant->size);
3293 u64 max_size = le64_to_cpu(grant->max_size);
3294 unsigned char check_caps = 0;
3295 bool was_stale = cap->cap_gen < session->s_cap_gen;
3297 bool writeback = false;
3298 bool queue_trunc = false;
3299 bool queue_invalidate = false;
3300 bool deleted_inode = false;
3301 bool fill_inline = false;
3303 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3304 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3305 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3306 i_size_read(inode));
3310 * If CACHE is being revoked, and we have no dirty buffers,
3311 * try to invalidate (once). (If there are dirty buffers, we
3312 * will invalidate _after_ writeback.)
3314 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3315 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3316 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3317 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3318 if (try_nonblocking_invalidate(inode)) {
3319 /* there were locked pages.. invalidate later
3320 in a separate thread. */
3321 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3322 queue_invalidate = true;
3323 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3329 cap->issued = cap->implemented = CEPH_CAP_PIN;
3332 * auth mds of the inode changed. we received the cap export message,
3333 * but still haven't received the cap import message. handle_cap_export
3334 * updated the new auth MDS' cap.
3336 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3337 * that was sent before the cap import message. So don't remove caps.
3339 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3340 WARN_ON(cap != ci->i_auth_cap);
3341 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3343 newcaps |= cap->issued;
3346 /* side effects now are allowed */
3347 cap->cap_gen = session->s_cap_gen;
3350 __check_cap_issue(ci, cap, newcaps);
3352 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3354 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3355 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3356 umode_t mode = le32_to_cpu(grant->mode);
3358 if (inode_wrong_type(inode, mode))
3359 pr_warn_once("inode type changed! (ino %llx.%llx is 0%o, mds says 0%o)\n",
3360 ceph_vinop(inode), inode->i_mode, mode);
3362 inode->i_mode = mode;
3363 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3364 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3365 ci->i_btime = extra_info->btime;
3366 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3367 from_kuid(&init_user_ns, inode->i_uid),
3368 from_kgid(&init_user_ns, inode->i_gid));
3371 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3372 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3373 set_nlink(inode, le32_to_cpu(grant->nlink));
3374 if (inode->i_nlink == 0 &&
3375 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3376 deleted_inode = true;
3379 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3381 int len = le32_to_cpu(grant->xattr_len);
3382 u64 version = le64_to_cpu(grant->xattr_version);
3384 if (version > ci->i_xattrs.version) {
3385 dout(" got new xattrs v%llu on %p len %d\n",
3386 version, inode, len);
3387 if (ci->i_xattrs.blob)
3388 ceph_buffer_put(ci->i_xattrs.blob);
3389 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3390 ci->i_xattrs.version = version;
3391 ceph_forget_all_cached_acls(inode);
3392 ceph_security_invalidate_secctx(inode);
3396 if (newcaps & CEPH_CAP_ANY_RD) {
3397 struct timespec64 mtime, atime, ctime;
3398 /* ctime/mtime/atime? */
3399 ceph_decode_timespec64(&mtime, &grant->mtime);
3400 ceph_decode_timespec64(&atime, &grant->atime);
3401 ceph_decode_timespec64(&ctime, &grant->ctime);
3402 ceph_fill_file_time(inode, extra_info->issued,
3403 le32_to_cpu(grant->time_warp_seq),
3404 &ctime, &mtime, &atime);
3407 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3408 ci->i_files = extra_info->nfiles;
3409 ci->i_subdirs = extra_info->nsubdirs;
3412 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3413 /* file layout may have changed */
3414 s64 old_pool = ci->i_layout.pool_id;
3415 struct ceph_string *old_ns;
3417 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3418 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3419 lockdep_is_held(&ci->i_ceph_lock));
3420 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3422 if (ci->i_layout.pool_id != old_pool ||
3423 extra_info->pool_ns != old_ns)
3424 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3426 extra_info->pool_ns = old_ns;
3428 /* size/truncate_seq? */
3429 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3430 le32_to_cpu(grant->truncate_seq),
3431 le64_to_cpu(grant->truncate_size),
3435 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3436 if (max_size != ci->i_max_size) {
3437 dout("max_size %lld -> %llu\n",
3438 ci->i_max_size, max_size);
3439 ci->i_max_size = max_size;
3440 if (max_size >= ci->i_wanted_max_size) {
3441 ci->i_wanted_max_size = 0; /* reset */
3442 ci->i_requested_max_size = 0;
3448 /* check cap bits */
3449 wanted = __ceph_caps_wanted(ci);
3450 used = __ceph_caps_used(ci);
3451 dirty = __ceph_caps_dirty(ci);
3452 dout(" my wanted = %s, used = %s, dirty %s\n",
3453 ceph_cap_string(wanted),
3454 ceph_cap_string(used),
3455 ceph_cap_string(dirty));
3457 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3458 (wanted & ~(cap->mds_wanted | newcaps))) {
3460 * If mds is importing cap, prior cap messages that update
3461 * 'wanted' may get dropped by mds (migrate seq mismatch).
3463 * We don't send cap message to update 'wanted' if what we
3464 * want are already issued. If mds revokes caps, cap message
3465 * that releases caps also tells mds what we want. But if
3466 * caps got revoked by mds forcedly (session stale). We may
3467 * haven't told mds what we want.
3472 /* revocation, grant, or no-op? */
3473 if (cap->issued & ~newcaps) {
3474 int revoking = cap->issued & ~newcaps;
3476 dout("revocation: %s -> %s (revoking %s)\n",
3477 ceph_cap_string(cap->issued),
3478 ceph_cap_string(newcaps),
3479 ceph_cap_string(revoking));
3480 if (S_ISREG(inode->i_mode) &&
3481 (revoking & used & CEPH_CAP_FILE_BUFFER))
3482 writeback = true; /* initiate writeback; will delay ack */
3483 else if (queue_invalidate &&
3484 revoking == CEPH_CAP_FILE_CACHE &&
3485 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3486 ; /* do nothing yet, invalidation will be queued */
3487 else if (cap == ci->i_auth_cap)
3488 check_caps = 1; /* check auth cap only */
3490 check_caps = 2; /* check all caps */
3491 cap->issued = newcaps;
3492 cap->implemented |= newcaps;
3493 } else if (cap->issued == newcaps) {
3494 dout("caps unchanged: %s -> %s\n",
3495 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3497 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3498 ceph_cap_string(newcaps));
3499 /* non-auth MDS is revoking the newly grant caps ? */
3500 if (cap == ci->i_auth_cap &&
3501 __ceph_caps_revoking_other(ci, cap, newcaps))
3504 cap->issued = newcaps;
3505 cap->implemented |= newcaps; /* add bits only, to
3506 * avoid stepping on a
3507 * pending revocation */
3510 BUG_ON(cap->issued & ~cap->implemented);
3512 if (extra_info->inline_version > 0 &&
3513 extra_info->inline_version >= ci->i_inline_version) {
3514 ci->i_inline_version = extra_info->inline_version;
3515 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3516 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3520 if (ci->i_auth_cap == cap &&
3521 le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3522 if (newcaps & ~extra_info->issued)
3525 if (ci->i_requested_max_size > max_size ||
3526 !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3527 /* re-request max_size if necessary */
3528 ci->i_requested_max_size = 0;
3532 ceph_kick_flushing_inode_caps(session, ci);
3533 spin_unlock(&ci->i_ceph_lock);
3534 up_read(&session->s_mdsc->snap_rwsem);
3536 spin_unlock(&ci->i_ceph_lock);
3540 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3541 extra_info->inline_len);
3544 ceph_queue_vmtruncate(inode);
3548 * queue inode for writeback: we can't actually call
3549 * filemap_write_and_wait, etc. from message handler
3552 ceph_queue_writeback(inode);
3553 if (queue_invalidate)
3554 ceph_queue_invalidate(inode);
3556 invalidate_aliases(inode);
3558 wake_up_all(&ci->i_cap_wq);
3560 if (check_caps == 1)
3561 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3563 else if (check_caps == 2)
3564 ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3566 mutex_unlock(&session->s_mutex);
3570 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3571 * MDS has been safely committed.
3573 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3574 struct ceph_mds_caps *m,
3575 struct ceph_mds_session *session,
3576 struct ceph_cap *cap)
3577 __releases(ci->i_ceph_lock)
3579 struct ceph_inode_info *ci = ceph_inode(inode);
3580 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3581 struct ceph_cap_flush *cf, *tmp_cf;
3582 LIST_HEAD(to_remove);
3583 unsigned seq = le32_to_cpu(m->seq);
3584 int dirty = le32_to_cpu(m->dirty);
3587 bool wake_ci = false;
3588 bool wake_mdsc = false;
3590 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3591 /* Is this the one that was flushed? */
3592 if (cf->tid == flush_tid)
3595 /* Is this a capsnap? */
3599 if (cf->tid <= flush_tid) {
3601 * An earlier or current tid. The FLUSH_ACK should
3602 * represent a superset of this flush's caps.
3604 wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3605 list_add_tail(&cf->i_list, &to_remove);
3608 * This is a later one. Any caps in it are still dirty
3609 * so don't count them as cleaned.
3611 cleaned &= ~cf->caps;
3617 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3618 " flushing %s -> %s\n",
3619 inode, session->s_mds, seq, ceph_cap_string(dirty),
3620 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3621 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3623 if (list_empty(&to_remove) && !cleaned)
3626 ci->i_flushing_caps &= ~cleaned;
3628 spin_lock(&mdsc->cap_dirty_lock);
3630 list_for_each_entry(cf, &to_remove, i_list)
3631 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3633 if (ci->i_flushing_caps == 0) {
3634 if (list_empty(&ci->i_cap_flush_list)) {
3635 list_del_init(&ci->i_flushing_item);
3636 if (!list_empty(&session->s_cap_flushing)) {
3637 dout(" mds%d still flushing cap on %p\n",
3639 &list_first_entry(&session->s_cap_flushing,
3640 struct ceph_inode_info,
3641 i_flushing_item)->vfs_inode);
3644 mdsc->num_cap_flushing--;
3645 dout(" inode %p now !flushing\n", inode);
3647 if (ci->i_dirty_caps == 0) {
3648 dout(" inode %p now clean\n", inode);
3649 BUG_ON(!list_empty(&ci->i_dirty_item));
3651 if (ci->i_wr_ref == 0 &&
3652 ci->i_wrbuffer_ref_head == 0) {
3653 BUG_ON(!ci->i_head_snapc);
3654 ceph_put_snap_context(ci->i_head_snapc);
3655 ci->i_head_snapc = NULL;
3658 BUG_ON(list_empty(&ci->i_dirty_item));
3661 spin_unlock(&mdsc->cap_dirty_lock);
3664 spin_unlock(&ci->i_ceph_lock);
3666 while (!list_empty(&to_remove)) {
3667 cf = list_first_entry(&to_remove,
3668 struct ceph_cap_flush, i_list);
3669 list_del_init(&cf->i_list);
3670 if (!cf->is_capsnap)
3671 ceph_free_cap_flush(cf);
3675 wake_up_all(&ci->i_cap_wq);
3677 wake_up_all(&mdsc->cap_flushing_wq);
3683 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3684 * throw away our cap_snap.
3686 * Caller hold s_mutex.
3688 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3689 struct ceph_mds_caps *m,
3690 struct ceph_mds_session *session)
3692 struct ceph_inode_info *ci = ceph_inode(inode);
3693 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3694 u64 follows = le64_to_cpu(m->snap_follows);
3695 struct ceph_cap_snap *capsnap;
3696 bool flushed = false;
3697 bool wake_ci = false;
3698 bool wake_mdsc = false;
3700 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3701 inode, ci, session->s_mds, follows);
3703 spin_lock(&ci->i_ceph_lock);
3704 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3705 if (capsnap->follows == follows) {
3706 if (capsnap->cap_flush.tid != flush_tid) {
3707 dout(" cap_snap %p follows %lld tid %lld !="
3708 " %lld\n", capsnap, follows,
3709 flush_tid, capsnap->cap_flush.tid);
3715 dout(" skipping cap_snap %p follows %lld\n",
3716 capsnap, capsnap->follows);
3720 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3721 dout(" removing %p cap_snap %p follows %lld\n",
3722 inode, capsnap, follows);
3723 list_del(&capsnap->ci_item);
3724 wake_ci |= __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3726 spin_lock(&mdsc->cap_dirty_lock);
3728 if (list_empty(&ci->i_cap_flush_list))
3729 list_del_init(&ci->i_flushing_item);
3731 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc,
3732 &capsnap->cap_flush);
3733 spin_unlock(&mdsc->cap_dirty_lock);
3735 spin_unlock(&ci->i_ceph_lock);
3737 ceph_put_snap_context(capsnap->context);
3738 ceph_put_cap_snap(capsnap);
3740 wake_up_all(&ci->i_cap_wq);
3742 wake_up_all(&mdsc->cap_flushing_wq);
3748 * Handle TRUNC from MDS, indicating file truncation.
3750 * caller hold s_mutex.
3752 static bool handle_cap_trunc(struct inode *inode,
3753 struct ceph_mds_caps *trunc,
3754 struct ceph_mds_session *session)
3756 struct ceph_inode_info *ci = ceph_inode(inode);
3757 int mds = session->s_mds;
3758 int seq = le32_to_cpu(trunc->seq);
3759 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3760 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3761 u64 size = le64_to_cpu(trunc->size);
3762 int implemented = 0;
3763 int dirty = __ceph_caps_dirty(ci);
3764 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3765 bool queue_trunc = false;
3767 lockdep_assert_held(&ci->i_ceph_lock);
3769 issued |= implemented | dirty;
3771 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3772 inode, mds, seq, truncate_size, truncate_seq);
3773 queue_trunc = ceph_fill_file_size(inode, issued,
3774 truncate_seq, truncate_size, size);
3779 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3780 * different one. If we are the most recent migration we've seen (as
3781 * indicated by mseq), make note of the migrating cap bits for the
3782 * duration (until we see the corresponding IMPORT).
3784 * caller holds s_mutex
3786 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3787 struct ceph_mds_cap_peer *ph,
3788 struct ceph_mds_session *session)
3790 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3791 struct ceph_mds_session *tsession = NULL;
3792 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3793 struct ceph_inode_info *ci = ceph_inode(inode);
3795 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3796 unsigned t_seq, t_mseq;
3798 int mds = session->s_mds;
3801 t_cap_id = le64_to_cpu(ph->cap_id);
3802 t_seq = le32_to_cpu(ph->seq);
3803 t_mseq = le32_to_cpu(ph->mseq);
3804 target = le32_to_cpu(ph->mds);
3806 t_cap_id = t_seq = t_mseq = 0;
3810 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3811 inode, ci, mds, mseq, target);
3813 spin_lock(&ci->i_ceph_lock);
3814 cap = __get_cap_for_mds(ci, mds);
3815 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3819 __ceph_remove_cap(cap, false);
3824 * now we know we haven't received the cap import message yet
3825 * because the exported cap still exist.
3828 issued = cap->issued;
3829 if (issued != cap->implemented)
3830 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3831 "ino (%llx.%llx) mds%d seq %d mseq %d "
3832 "issued %s implemented %s\n",
3833 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3834 ceph_cap_string(issued),
3835 ceph_cap_string(cap->implemented));
3838 tcap = __get_cap_for_mds(ci, target);
3840 /* already have caps from the target */
3841 if (tcap->cap_id == t_cap_id &&
3842 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3843 dout(" updating import cap %p mds%d\n", tcap, target);
3844 tcap->cap_id = t_cap_id;
3845 tcap->seq = t_seq - 1;
3846 tcap->issue_seq = t_seq - 1;
3847 tcap->issued |= issued;
3848 tcap->implemented |= issued;
3849 if (cap == ci->i_auth_cap) {
3850 ci->i_auth_cap = tcap;
3851 change_auth_cap_ses(ci, tcap->session);
3854 __ceph_remove_cap(cap, false);
3856 } else if (tsession) {
3857 /* add placeholder for the export tagert */
3858 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3860 ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3861 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3863 if (!list_empty(&ci->i_cap_flush_list) &&
3864 ci->i_auth_cap == tcap) {
3865 spin_lock(&mdsc->cap_dirty_lock);
3866 list_move_tail(&ci->i_flushing_item,
3867 &tcap->session->s_cap_flushing);
3868 spin_unlock(&mdsc->cap_dirty_lock);
3871 __ceph_remove_cap(cap, false);
3875 spin_unlock(&ci->i_ceph_lock);
3876 mutex_unlock(&session->s_mutex);
3878 /* open target session */
3879 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3880 if (!IS_ERR(tsession)) {
3882 mutex_lock(&session->s_mutex);
3883 mutex_lock_nested(&tsession->s_mutex,
3884 SINGLE_DEPTH_NESTING);
3886 mutex_lock(&tsession->s_mutex);
3887 mutex_lock_nested(&session->s_mutex,
3888 SINGLE_DEPTH_NESTING);
3890 new_cap = ceph_get_cap(mdsc, NULL);
3895 mutex_lock(&session->s_mutex);
3900 spin_unlock(&ci->i_ceph_lock);
3901 mutex_unlock(&session->s_mutex);
3903 mutex_unlock(&tsession->s_mutex);
3904 ceph_put_mds_session(tsession);
3907 ceph_put_cap(mdsc, new_cap);
3911 * Handle cap IMPORT.
3913 * caller holds s_mutex. acquires i_ceph_lock
3915 static void handle_cap_import(struct ceph_mds_client *mdsc,
3916 struct inode *inode, struct ceph_mds_caps *im,
3917 struct ceph_mds_cap_peer *ph,
3918 struct ceph_mds_session *session,
3919 struct ceph_cap **target_cap, int *old_issued)
3921 struct ceph_inode_info *ci = ceph_inode(inode);
3922 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3923 int mds = session->s_mds;
3925 unsigned caps = le32_to_cpu(im->caps);
3926 unsigned wanted = le32_to_cpu(im->wanted);
3927 unsigned seq = le32_to_cpu(im->seq);
3928 unsigned mseq = le32_to_cpu(im->migrate_seq);
3929 u64 realmino = le64_to_cpu(im->realm);
3930 u64 cap_id = le64_to_cpu(im->cap_id);
3935 p_cap_id = le64_to_cpu(ph->cap_id);
3936 peer = le32_to_cpu(ph->mds);
3942 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3943 inode, ci, mds, mseq, peer);
3945 cap = __get_cap_for_mds(ci, mds);
3948 spin_unlock(&ci->i_ceph_lock);
3949 new_cap = ceph_get_cap(mdsc, NULL);
3950 spin_lock(&ci->i_ceph_lock);
3956 ceph_put_cap(mdsc, new_cap);
3961 __ceph_caps_issued(ci, &issued);
3962 issued |= __ceph_caps_dirty(ci);
3964 ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
3965 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3967 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3968 if (ocap && ocap->cap_id == p_cap_id) {
3969 dout(" remove export cap %p mds%d flags %d\n",
3970 ocap, peer, ph->flags);
3971 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3972 (ocap->seq != le32_to_cpu(ph->seq) ||
3973 ocap->mseq != le32_to_cpu(ph->mseq))) {
3974 pr_err_ratelimited("handle_cap_import: "
3975 "mismatched seq/mseq: ino (%llx.%llx) "
3976 "mds%d seq %d mseq %d importer mds%d "
3977 "has peer seq %d mseq %d\n",
3978 ceph_vinop(inode), peer, ocap->seq,
3979 ocap->mseq, mds, le32_to_cpu(ph->seq),
3980 le32_to_cpu(ph->mseq));
3982 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3985 *old_issued = issued;
3990 * Handle a caps message from the MDS.
3992 * Identify the appropriate session, inode, and call the right handler
3993 * based on the cap op.
3995 void ceph_handle_caps(struct ceph_mds_session *session,
3996 struct ceph_msg *msg)
3998 struct ceph_mds_client *mdsc = session->s_mdsc;
3999 struct inode *inode;
4000 struct ceph_inode_info *ci;
4001 struct ceph_cap *cap;
4002 struct ceph_mds_caps *h;
4003 struct ceph_mds_cap_peer *peer = NULL;
4004 struct ceph_snap_realm *realm = NULL;
4006 int msg_version = le16_to_cpu(msg->hdr.version);
4008 struct ceph_vino vino;
4010 size_t snaptrace_len;
4012 struct cap_extra_info extra_info = {};
4015 dout("handle_caps from mds%d\n", session->s_mds);
4018 end = msg->front.iov_base + msg->front.iov_len;
4019 if (msg->front.iov_len < sizeof(*h))
4021 h = msg->front.iov_base;
4022 op = le32_to_cpu(h->op);
4023 vino.ino = le64_to_cpu(h->ino);
4024 vino.snap = CEPH_NOSNAP;
4025 seq = le32_to_cpu(h->seq);
4026 mseq = le32_to_cpu(h->migrate_seq);
4029 snaptrace_len = le32_to_cpu(h->snap_trace_len);
4030 p = snaptrace + snaptrace_len;
4032 if (msg_version >= 2) {
4034 ceph_decode_32_safe(&p, end, flock_len, bad);
4035 if (p + flock_len > end)
4040 if (msg_version >= 3) {
4041 if (op == CEPH_CAP_OP_IMPORT) {
4042 if (p + sizeof(*peer) > end)
4046 } else if (op == CEPH_CAP_OP_EXPORT) {
4047 /* recorded in unused fields */
4048 peer = (void *)&h->size;
4052 if (msg_version >= 4) {
4053 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
4054 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
4055 if (p + extra_info.inline_len > end)
4057 extra_info.inline_data = p;
4058 p += extra_info.inline_len;
4061 if (msg_version >= 5) {
4062 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
4065 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
4066 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
4069 if (msg_version >= 8) {
4073 ceph_decode_skip_64(&p, end, bad); // flush_tid
4075 ceph_decode_skip_32(&p, end, bad); // caller_uid
4076 ceph_decode_skip_32(&p, end, bad); // caller_gid
4078 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4079 if (pool_ns_len > 0) {
4080 ceph_decode_need(&p, end, pool_ns_len, bad);
4081 extra_info.pool_ns =
4082 ceph_find_or_create_string(p, pool_ns_len);
4087 if (msg_version >= 9) {
4088 struct ceph_timespec *btime;
4090 if (p + sizeof(*btime) > end)
4093 ceph_decode_timespec64(&extra_info.btime, btime);
4094 p += sizeof(*btime);
4095 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4098 if (msg_version >= 11) {
4100 ceph_decode_skip_32(&p, end, bad); // flags
4102 extra_info.dirstat_valid = true;
4103 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4104 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4108 inode = ceph_find_inode(mdsc->fsc->sb, vino);
4109 ci = ceph_inode(inode);
4110 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4113 mutex_lock(&session->s_mutex);
4114 inc_session_sequence(session);
4115 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4119 dout(" i don't have ino %llx\n", vino.ino);
4121 if (op == CEPH_CAP_OP_IMPORT) {
4122 cap = ceph_get_cap(mdsc, NULL);
4123 cap->cap_ino = vino.ino;
4124 cap->queue_release = 1;
4125 cap->cap_id = le64_to_cpu(h->cap_id);
4128 cap->issue_seq = seq;
4129 spin_lock(&session->s_cap_lock);
4130 __ceph_queue_cap_release(session, cap);
4131 spin_unlock(&session->s_cap_lock);
4133 goto flush_cap_releases;
4136 /* these will work even if we don't have a cap yet */
4138 case CEPH_CAP_OP_FLUSHSNAP_ACK:
4139 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4143 case CEPH_CAP_OP_EXPORT:
4144 handle_cap_export(inode, h, peer, session);
4147 case CEPH_CAP_OP_IMPORT:
4149 if (snaptrace_len) {
4150 down_write(&mdsc->snap_rwsem);
4151 ceph_update_snap_trace(mdsc, snaptrace,
4152 snaptrace + snaptrace_len,
4154 downgrade_write(&mdsc->snap_rwsem);
4156 down_read(&mdsc->snap_rwsem);
4158 spin_lock(&ci->i_ceph_lock);
4159 handle_cap_import(mdsc, inode, h, peer, session,
4160 &cap, &extra_info.issued);
4161 handle_cap_grant(inode, session, cap,
4162 h, msg->middle, &extra_info);
4164 ceph_put_snap_realm(mdsc, realm);
4168 /* the rest require a cap */
4169 spin_lock(&ci->i_ceph_lock);
4170 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4172 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4173 inode, ceph_ino(inode), ceph_snap(inode),
4175 spin_unlock(&ci->i_ceph_lock);
4176 goto flush_cap_releases;
4179 /* note that each of these drops i_ceph_lock for us */
4181 case CEPH_CAP_OP_REVOKE:
4182 case CEPH_CAP_OP_GRANT:
4183 __ceph_caps_issued(ci, &extra_info.issued);
4184 extra_info.issued |= __ceph_caps_dirty(ci);
4185 handle_cap_grant(inode, session, cap,
4186 h, msg->middle, &extra_info);
4189 case CEPH_CAP_OP_FLUSH_ACK:
4190 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4194 case CEPH_CAP_OP_TRUNC:
4195 queue_trunc = handle_cap_trunc(inode, h, session);
4196 spin_unlock(&ci->i_ceph_lock);
4198 ceph_queue_vmtruncate(inode);
4202 spin_unlock(&ci->i_ceph_lock);
4203 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4204 ceph_cap_op_name(op));
4208 mutex_unlock(&session->s_mutex);
4210 ceph_put_string(extra_info.pool_ns);
4211 /* avoid calling iput_final() in mds dispatch threads */
4212 ceph_async_iput(inode);
4217 * send any cap release message to try to move things
4218 * along for the mds (who clearly thinks we still have this
4221 ceph_flush_cap_releases(mdsc, session);
4225 pr_err("ceph_handle_caps: corrupt message\n");
4231 * Delayed work handler to process end of delayed cap release LRU list.
4233 * If new caps are added to the list while processing it, these won't get
4234 * processed in this run. In this case, the ci->i_hold_caps_max will be
4235 * returned so that the work can be scheduled accordingly.
4237 unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4239 struct inode *inode;
4240 struct ceph_inode_info *ci;
4241 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
4242 unsigned long delay_max = opt->caps_wanted_delay_max * HZ;
4243 unsigned long loop_start = jiffies;
4244 unsigned long delay = 0;
4246 dout("check_delayed_caps\n");
4247 spin_lock(&mdsc->cap_delay_lock);
4248 while (!list_empty(&mdsc->cap_delay_list)) {
4249 ci = list_first_entry(&mdsc->cap_delay_list,
4250 struct ceph_inode_info,
4252 if (time_before(loop_start, ci->i_hold_caps_max - delay_max)) {
4253 dout("%s caps added recently. Exiting loop", __func__);
4254 delay = ci->i_hold_caps_max;
4257 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4258 time_before(jiffies, ci->i_hold_caps_max))
4260 list_del_init(&ci->i_cap_delay_list);
4262 inode = igrab(&ci->vfs_inode);
4264 spin_unlock(&mdsc->cap_delay_lock);
4265 dout("check_delayed_caps on %p\n", inode);
4266 ceph_check_caps(ci, 0, NULL);
4267 /* avoid calling iput_final() in tick thread */
4268 ceph_async_iput(inode);
4269 spin_lock(&mdsc->cap_delay_lock);
4272 spin_unlock(&mdsc->cap_delay_lock);
4278 * Flush all dirty caps to the mds
4280 static void flush_dirty_session_caps(struct ceph_mds_session *s)
4282 struct ceph_mds_client *mdsc = s->s_mdsc;
4283 struct ceph_inode_info *ci;
4284 struct inode *inode;
4286 dout("flush_dirty_caps\n");
4287 spin_lock(&mdsc->cap_dirty_lock);
4288 while (!list_empty(&s->s_cap_dirty)) {
4289 ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4291 inode = &ci->vfs_inode;
4293 dout("flush_dirty_caps %p\n", inode);
4294 spin_unlock(&mdsc->cap_dirty_lock);
4295 ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4297 spin_lock(&mdsc->cap_dirty_lock);
4299 spin_unlock(&mdsc->cap_dirty_lock);
4300 dout("flush_dirty_caps done\n");
4303 static void iterate_sessions(struct ceph_mds_client *mdsc,
4304 void (*cb)(struct ceph_mds_session *))
4308 mutex_lock(&mdsc->mutex);
4309 for (mds = 0; mds < mdsc->max_sessions; ++mds) {
4310 struct ceph_mds_session *s;
4312 if (!mdsc->sessions[mds])
4315 s = ceph_get_mds_session(mdsc->sessions[mds]);
4319 mutex_unlock(&mdsc->mutex);
4321 ceph_put_mds_session(s);
4322 mutex_lock(&mdsc->mutex);
4324 mutex_unlock(&mdsc->mutex);
4327 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4329 iterate_sessions(mdsc, flush_dirty_session_caps);
4332 void __ceph_touch_fmode(struct ceph_inode_info *ci,
4333 struct ceph_mds_client *mdsc, int fmode)
4335 unsigned long now = jiffies;
4336 if (fmode & CEPH_FILE_MODE_RD)
4337 ci->i_last_rd = now;
4338 if (fmode & CEPH_FILE_MODE_WR)
4339 ci->i_last_wr = now;
4340 /* queue periodic check */
4342 __ceph_is_any_real_caps(ci) &&
4343 list_empty(&ci->i_cap_delay_list))
4344 __cap_delay_requeue(mdsc, ci);
4347 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4349 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4350 int bits = (fmode << 1) | 1;
4351 bool is_opened = false;
4355 atomic64_inc(&mdsc->metric.opened_files);
4357 spin_lock(&ci->i_ceph_lock);
4358 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4359 if (bits & (1 << i))
4360 ci->i_nr_by_mode[i] += count;
4363 * If any of the mode ref is larger than 1,
4364 * that means it has been already opened by
4365 * others. Just skip checking the PIN ref.
4367 if (i && ci->i_nr_by_mode[i] > 1)
4372 percpu_counter_inc(&mdsc->metric.opened_inodes);
4373 spin_unlock(&ci->i_ceph_lock);
4377 * Drop open file reference. If we were the last open file,
4378 * we may need to release capabilities to the MDS (or schedule
4379 * their delayed release).
4381 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4383 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4384 int bits = (fmode << 1) | 1;
4385 bool is_closed = true;
4389 atomic64_dec(&mdsc->metric.opened_files);
4391 spin_lock(&ci->i_ceph_lock);
4392 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4393 if (bits & (1 << i)) {
4394 BUG_ON(ci->i_nr_by_mode[i] < count);
4395 ci->i_nr_by_mode[i] -= count;
4399 * If any of the mode ref is not 0 after
4400 * decreased, that means it is still opened
4401 * by others. Just skip checking the PIN ref.
4403 if (i && ci->i_nr_by_mode[i])
4408 percpu_counter_dec(&mdsc->metric.opened_inodes);
4409 spin_unlock(&ci->i_ceph_lock);
4413 * For a soon-to-be unlinked file, drop the LINK caps. If it
4414 * looks like the link count will hit 0, drop any other caps (other
4415 * than PIN) we don't specifically want (due to the file still being
4418 int ceph_drop_caps_for_unlink(struct inode *inode)
4420 struct ceph_inode_info *ci = ceph_inode(inode);
4421 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4423 spin_lock(&ci->i_ceph_lock);
4424 if (inode->i_nlink == 1) {
4425 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4427 if (__ceph_caps_dirty(ci)) {
4428 struct ceph_mds_client *mdsc =
4429 ceph_inode_to_client(inode)->mdsc;
4430 __cap_delay_requeue_front(mdsc, ci);
4433 spin_unlock(&ci->i_ceph_lock);
4438 * Helpers for embedding cap and dentry lease releases into mds
4441 * @force is used by dentry_release (below) to force inclusion of a
4442 * record for the directory inode, even when there aren't any caps to
4445 int ceph_encode_inode_release(void **p, struct inode *inode,
4446 int mds, int drop, int unless, int force)
4448 struct ceph_inode_info *ci = ceph_inode(inode);
4449 struct ceph_cap *cap;
4450 struct ceph_mds_request_release *rel = *p;
4454 spin_lock(&ci->i_ceph_lock);
4455 used = __ceph_caps_used(ci);
4456 dirty = __ceph_caps_dirty(ci);
4458 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4459 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4460 ceph_cap_string(unless));
4462 /* only drop unused, clean caps */
4463 drop &= ~(used | dirty);
4465 cap = __get_cap_for_mds(ci, mds);
4466 if (cap && __cap_is_valid(cap)) {
4467 unless &= cap->issued;
4469 if (unless & CEPH_CAP_AUTH_EXCL)
4470 drop &= ~CEPH_CAP_AUTH_SHARED;
4471 if (unless & CEPH_CAP_LINK_EXCL)
4472 drop &= ~CEPH_CAP_LINK_SHARED;
4473 if (unless & CEPH_CAP_XATTR_EXCL)
4474 drop &= ~CEPH_CAP_XATTR_SHARED;
4475 if (unless & CEPH_CAP_FILE_EXCL)
4476 drop &= ~CEPH_CAP_FILE_SHARED;
4479 if (force || (cap->issued & drop)) {
4480 if (cap->issued & drop) {
4481 int wanted = __ceph_caps_wanted(ci);
4482 dout("encode_inode_release %p cap %p "
4483 "%s -> %s, wanted %s -> %s\n", inode, cap,
4484 ceph_cap_string(cap->issued),
4485 ceph_cap_string(cap->issued & ~drop),
4486 ceph_cap_string(cap->mds_wanted),
4487 ceph_cap_string(wanted));
4489 cap->issued &= ~drop;
4490 cap->implemented &= ~drop;
4491 cap->mds_wanted = wanted;
4492 if (cap == ci->i_auth_cap &&
4493 !(wanted & CEPH_CAP_ANY_FILE_WR))
4494 ci->i_requested_max_size = 0;
4496 dout("encode_inode_release %p cap %p %s"
4497 " (force)\n", inode, cap,
4498 ceph_cap_string(cap->issued));
4501 rel->ino = cpu_to_le64(ceph_ino(inode));
4502 rel->cap_id = cpu_to_le64(cap->cap_id);
4503 rel->seq = cpu_to_le32(cap->seq);
4504 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4505 rel->mseq = cpu_to_le32(cap->mseq);
4506 rel->caps = cpu_to_le32(cap->implemented);
4507 rel->wanted = cpu_to_le32(cap->mds_wanted);
4513 dout("encode_inode_release %p cap %p %s (noop)\n",
4514 inode, cap, ceph_cap_string(cap->issued));
4517 spin_unlock(&ci->i_ceph_lock);
4521 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4523 int mds, int drop, int unless)
4525 struct dentry *parent = NULL;
4526 struct ceph_mds_request_release *rel = *p;
4527 struct ceph_dentry_info *di = ceph_dentry(dentry);
4532 * force an record for the directory caps if we have a dentry lease.
4533 * this is racy (can't take i_ceph_lock and d_lock together), but it
4534 * doesn't have to be perfect; the mds will revoke anything we don't
4537 spin_lock(&dentry->d_lock);
4538 if (di->lease_session && di->lease_session->s_mds == mds)
4541 parent = dget(dentry->d_parent);
4542 dir = d_inode(parent);
4544 spin_unlock(&dentry->d_lock);
4546 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4549 spin_lock(&dentry->d_lock);
4550 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4551 dout("encode_dentry_release %p mds%d seq %d\n",
4552 dentry, mds, (int)di->lease_seq);
4553 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4554 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4555 *p += dentry->d_name.len;
4556 rel->dname_seq = cpu_to_le32(di->lease_seq);
4557 __ceph_mdsc_drop_dentry_lease(dentry);
4559 spin_unlock(&dentry->d_lock);