1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
5 #include <linux/kernel.h>
6 #include <linux/sched/signal.h>
7 #include <linux/slab.h>
8 #include <linux/vmalloc.h>
9 #include <linux/wait.h>
10 #include <linux/writeback.h>
11 #include <linux/iversion.h>
14 #include "mds_client.h"
16 #include <linux/ceph/decode.h>
17 #include <linux/ceph/messenger.h>
20 * Capability management
22 * The Ceph metadata servers control client access to inode metadata
23 * and file data by issuing capabilities, granting clients permission
24 * to read and/or write both inode field and file data to OSDs
25 * (storage nodes). Each capability consists of a set of bits
26 * indicating which operations are allowed.
28 * If the client holds a *_SHARED cap, the client has a coherent value
29 * that can be safely read from the cached inode.
31 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
32 * client is allowed to change inode attributes (e.g., file size,
33 * mtime), note its dirty state in the ceph_cap, and asynchronously
34 * flush that metadata change to the MDS.
36 * In the event of a conflicting operation (perhaps by another
37 * client), the MDS will revoke the conflicting client capabilities.
39 * In order for a client to cache an inode, it must hold a capability
40 * with at least one MDS server. When inodes are released, release
41 * notifications are batched and periodically sent en masse to the MDS
42 * cluster to release server state.
45 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc);
46 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
47 struct ceph_mds_session *session,
48 struct ceph_inode_info *ci,
49 u64 oldest_flush_tid);
52 * Generate readable cap strings for debugging output.
54 #define MAX_CAP_STR 20
55 static char cap_str[MAX_CAP_STR][40];
56 static DEFINE_SPINLOCK(cap_str_lock);
57 static int last_cap_str;
59 static char *gcap_string(char *s, int c)
61 if (c & CEPH_CAP_GSHARED)
63 if (c & CEPH_CAP_GEXCL)
65 if (c & CEPH_CAP_GCACHE)
71 if (c & CEPH_CAP_GBUFFER)
73 if (c & CEPH_CAP_GWREXTEND)
75 if (c & CEPH_CAP_GLAZYIO)
80 const char *ceph_cap_string(int caps)
86 spin_lock(&cap_str_lock);
88 if (last_cap_str == MAX_CAP_STR)
90 spin_unlock(&cap_str_lock);
94 if (caps & CEPH_CAP_PIN)
97 c = (caps >> CEPH_CAP_SAUTH) & 3;
100 s = gcap_string(s, c);
103 c = (caps >> CEPH_CAP_SLINK) & 3;
106 s = gcap_string(s, c);
109 c = (caps >> CEPH_CAP_SXATTR) & 3;
112 s = gcap_string(s, c);
115 c = caps >> CEPH_CAP_SFILE;
118 s = gcap_string(s, c);
127 void ceph_caps_init(struct ceph_mds_client *mdsc)
129 INIT_LIST_HEAD(&mdsc->caps_list);
130 spin_lock_init(&mdsc->caps_list_lock);
133 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
135 struct ceph_cap *cap;
137 spin_lock(&mdsc->caps_list_lock);
138 while (!list_empty(&mdsc->caps_list)) {
139 cap = list_first_entry(&mdsc->caps_list,
140 struct ceph_cap, caps_item);
141 list_del(&cap->caps_item);
142 kmem_cache_free(ceph_cap_cachep, cap);
144 mdsc->caps_total_count = 0;
145 mdsc->caps_avail_count = 0;
146 mdsc->caps_use_count = 0;
147 mdsc->caps_reserve_count = 0;
148 mdsc->caps_min_count = 0;
149 spin_unlock(&mdsc->caps_list_lock);
152 void ceph_adjust_caps_max_min(struct ceph_mds_client *mdsc,
153 struct ceph_mount_options *fsopt)
155 spin_lock(&mdsc->caps_list_lock);
156 mdsc->caps_min_count = fsopt->max_readdir;
157 if (mdsc->caps_min_count < 1024)
158 mdsc->caps_min_count = 1024;
159 mdsc->caps_use_max = fsopt->caps_max;
160 if (mdsc->caps_use_max > 0 &&
161 mdsc->caps_use_max < mdsc->caps_min_count)
162 mdsc->caps_use_max = mdsc->caps_min_count;
163 spin_unlock(&mdsc->caps_list_lock);
166 static void __ceph_unreserve_caps(struct ceph_mds_client *mdsc, int nr_caps)
168 struct ceph_cap *cap;
172 BUG_ON(mdsc->caps_reserve_count < nr_caps);
173 mdsc->caps_reserve_count -= nr_caps;
174 if (mdsc->caps_avail_count >=
175 mdsc->caps_reserve_count + mdsc->caps_min_count) {
176 mdsc->caps_total_count -= nr_caps;
177 for (i = 0; i < nr_caps; i++) {
178 cap = list_first_entry(&mdsc->caps_list,
179 struct ceph_cap, caps_item);
180 list_del(&cap->caps_item);
181 kmem_cache_free(ceph_cap_cachep, cap);
184 mdsc->caps_avail_count += nr_caps;
187 dout("%s: caps %d = %d used + %d resv + %d avail\n",
189 mdsc->caps_total_count, mdsc->caps_use_count,
190 mdsc->caps_reserve_count, mdsc->caps_avail_count);
191 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
192 mdsc->caps_reserve_count +
193 mdsc->caps_avail_count);
198 * Called under mdsc->mutex.
200 int ceph_reserve_caps(struct ceph_mds_client *mdsc,
201 struct ceph_cap_reservation *ctx, int need)
204 struct ceph_cap *cap;
209 bool trimmed = false;
210 struct ceph_mds_session *s;
213 dout("reserve caps ctx=%p need=%d\n", ctx, need);
215 /* first reserve any caps that are already allocated */
216 spin_lock(&mdsc->caps_list_lock);
217 if (mdsc->caps_avail_count >= need)
220 have = mdsc->caps_avail_count;
221 mdsc->caps_avail_count -= have;
222 mdsc->caps_reserve_count += have;
223 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
224 mdsc->caps_reserve_count +
225 mdsc->caps_avail_count);
226 spin_unlock(&mdsc->caps_list_lock);
228 for (i = have; i < need; ) {
229 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
231 list_add(&cap->caps_item, &newcaps);
238 for (j = 0; j < mdsc->max_sessions; j++) {
239 s = __ceph_lookup_mds_session(mdsc, j);
242 mutex_unlock(&mdsc->mutex);
244 mutex_lock(&s->s_mutex);
245 max_caps = s->s_nr_caps - (need - i);
246 ceph_trim_caps(mdsc, s, max_caps);
247 mutex_unlock(&s->s_mutex);
249 ceph_put_mds_session(s);
250 mutex_lock(&mdsc->mutex);
254 spin_lock(&mdsc->caps_list_lock);
255 if (mdsc->caps_avail_count) {
257 if (mdsc->caps_avail_count >= need - i)
258 more_have = need - i;
260 more_have = mdsc->caps_avail_count;
264 mdsc->caps_avail_count -= more_have;
265 mdsc->caps_reserve_count += more_have;
268 spin_unlock(&mdsc->caps_list_lock);
273 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
274 ctx, need, have + alloc);
280 BUG_ON(have + alloc != need);
285 spin_lock(&mdsc->caps_list_lock);
286 mdsc->caps_total_count += alloc;
287 mdsc->caps_reserve_count += alloc;
288 list_splice(&newcaps, &mdsc->caps_list);
290 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
291 mdsc->caps_reserve_count +
292 mdsc->caps_avail_count);
295 __ceph_unreserve_caps(mdsc, have + alloc);
297 spin_unlock(&mdsc->caps_list_lock);
299 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
300 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
301 mdsc->caps_reserve_count, mdsc->caps_avail_count);
305 void ceph_unreserve_caps(struct ceph_mds_client *mdsc,
306 struct ceph_cap_reservation *ctx)
308 bool reclaim = false;
312 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
313 spin_lock(&mdsc->caps_list_lock);
314 __ceph_unreserve_caps(mdsc, ctx->count);
317 if (mdsc->caps_use_max > 0 &&
318 mdsc->caps_use_count > mdsc->caps_use_max)
320 spin_unlock(&mdsc->caps_list_lock);
323 ceph_reclaim_caps_nr(mdsc, ctx->used);
326 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
327 struct ceph_cap_reservation *ctx)
329 struct ceph_cap *cap = NULL;
331 /* temporary, until we do something about cap import/export */
333 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
335 spin_lock(&mdsc->caps_list_lock);
336 mdsc->caps_use_count++;
337 mdsc->caps_total_count++;
338 spin_unlock(&mdsc->caps_list_lock);
340 spin_lock(&mdsc->caps_list_lock);
341 if (mdsc->caps_avail_count) {
342 BUG_ON(list_empty(&mdsc->caps_list));
344 mdsc->caps_avail_count--;
345 mdsc->caps_use_count++;
346 cap = list_first_entry(&mdsc->caps_list,
347 struct ceph_cap, caps_item);
348 list_del(&cap->caps_item);
350 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
351 mdsc->caps_reserve_count + mdsc->caps_avail_count);
353 spin_unlock(&mdsc->caps_list_lock);
359 spin_lock(&mdsc->caps_list_lock);
360 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
361 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
362 mdsc->caps_reserve_count, mdsc->caps_avail_count);
364 BUG_ON(ctx->count > mdsc->caps_reserve_count);
365 BUG_ON(list_empty(&mdsc->caps_list));
369 mdsc->caps_reserve_count--;
370 mdsc->caps_use_count++;
372 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
373 list_del(&cap->caps_item);
375 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
376 mdsc->caps_reserve_count + mdsc->caps_avail_count);
377 spin_unlock(&mdsc->caps_list_lock);
381 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
383 spin_lock(&mdsc->caps_list_lock);
384 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
385 cap, mdsc->caps_total_count, mdsc->caps_use_count,
386 mdsc->caps_reserve_count, mdsc->caps_avail_count);
387 mdsc->caps_use_count--;
389 * Keep some preallocated caps around (ceph_min_count), to
390 * avoid lots of free/alloc churn.
392 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
393 mdsc->caps_min_count) {
394 mdsc->caps_total_count--;
395 kmem_cache_free(ceph_cap_cachep, cap);
397 mdsc->caps_avail_count++;
398 list_add(&cap->caps_item, &mdsc->caps_list);
401 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
402 mdsc->caps_reserve_count + mdsc->caps_avail_count);
403 spin_unlock(&mdsc->caps_list_lock);
406 void ceph_reservation_status(struct ceph_fs_client *fsc,
407 int *total, int *avail, int *used, int *reserved,
410 struct ceph_mds_client *mdsc = fsc->mdsc;
412 spin_lock(&mdsc->caps_list_lock);
415 *total = mdsc->caps_total_count;
417 *avail = mdsc->caps_avail_count;
419 *used = mdsc->caps_use_count;
421 *reserved = mdsc->caps_reserve_count;
423 *min = mdsc->caps_min_count;
425 spin_unlock(&mdsc->caps_list_lock);
429 * Find ceph_cap for given mds, if any.
431 * Called with i_ceph_lock held.
433 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
435 struct ceph_cap *cap;
436 struct rb_node *n = ci->i_caps.rb_node;
439 cap = rb_entry(n, struct ceph_cap, ci_node);
442 else if (mds > cap->mds)
450 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
452 struct ceph_cap *cap;
454 spin_lock(&ci->i_ceph_lock);
455 cap = __get_cap_for_mds(ci, mds);
456 spin_unlock(&ci->i_ceph_lock);
461 * Called under i_ceph_lock.
463 static void __insert_cap_node(struct ceph_inode_info *ci,
464 struct ceph_cap *new)
466 struct rb_node **p = &ci->i_caps.rb_node;
467 struct rb_node *parent = NULL;
468 struct ceph_cap *cap = NULL;
472 cap = rb_entry(parent, struct ceph_cap, ci_node);
473 if (new->mds < cap->mds)
475 else if (new->mds > cap->mds)
481 rb_link_node(&new->ci_node, parent, p);
482 rb_insert_color(&new->ci_node, &ci->i_caps);
486 * (re)set cap hold timeouts, which control the delayed release
487 * of unused caps back to the MDS. Should be called on cap use.
489 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
490 struct ceph_inode_info *ci)
492 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
493 ci->i_hold_caps_max = round_jiffies(jiffies +
494 opt->caps_wanted_delay_max * HZ);
495 dout("__cap_set_timeouts %p %lu\n", &ci->vfs_inode,
496 ci->i_hold_caps_max - jiffies);
500 * (Re)queue cap at the end of the delayed cap release list.
502 * If I_FLUSH is set, leave the inode at the front of the list.
504 * Caller holds i_ceph_lock
505 * -> we take mdsc->cap_delay_lock
507 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
508 struct ceph_inode_info *ci)
510 dout("__cap_delay_requeue %p flags 0x%lx at %lu\n", &ci->vfs_inode,
511 ci->i_ceph_flags, ci->i_hold_caps_max);
512 if (!mdsc->stopping) {
513 spin_lock(&mdsc->cap_delay_lock);
514 if (!list_empty(&ci->i_cap_delay_list)) {
515 if (ci->i_ceph_flags & CEPH_I_FLUSH)
517 list_del_init(&ci->i_cap_delay_list);
519 __cap_set_timeouts(mdsc, ci);
520 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
522 spin_unlock(&mdsc->cap_delay_lock);
527 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
528 * indicating we should send a cap message to flush dirty metadata
529 * asap, and move to the front of the delayed cap list.
531 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
532 struct ceph_inode_info *ci)
534 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
535 spin_lock(&mdsc->cap_delay_lock);
536 ci->i_ceph_flags |= CEPH_I_FLUSH;
537 if (!list_empty(&ci->i_cap_delay_list))
538 list_del_init(&ci->i_cap_delay_list);
539 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
540 spin_unlock(&mdsc->cap_delay_lock);
544 * Cancel delayed work on cap.
546 * Caller must hold i_ceph_lock.
548 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
549 struct ceph_inode_info *ci)
551 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
552 if (list_empty(&ci->i_cap_delay_list))
554 spin_lock(&mdsc->cap_delay_lock);
555 list_del_init(&ci->i_cap_delay_list);
556 spin_unlock(&mdsc->cap_delay_lock);
559 /* Common issue checks for add_cap, handle_cap_grant. */
560 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
563 unsigned had = __ceph_caps_issued(ci, NULL);
565 lockdep_assert_held(&ci->i_ceph_lock);
568 * Each time we receive FILE_CACHE anew, we increment
571 if (S_ISREG(ci->vfs_inode.i_mode) &&
572 (issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
573 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
578 * If FILE_SHARED is newly issued, mark dir not complete. We don't
579 * know what happened to this directory while we didn't have the cap.
580 * If FILE_SHARED is being revoked, also mark dir not complete. It
581 * stops on-going cached readdir.
583 if ((issued & CEPH_CAP_FILE_SHARED) != (had & CEPH_CAP_FILE_SHARED)) {
584 if (issued & CEPH_CAP_FILE_SHARED)
585 atomic_inc(&ci->i_shared_gen);
586 if (S_ISDIR(ci->vfs_inode.i_mode)) {
587 dout(" marking %p NOT complete\n", &ci->vfs_inode);
588 __ceph_dir_clear_complete(ci);
592 /* Wipe saved layout if we're losing DIR_CREATE caps */
593 if (S_ISDIR(ci->vfs_inode.i_mode) && (had & CEPH_CAP_DIR_CREATE) &&
594 !(issued & CEPH_CAP_DIR_CREATE)) {
595 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
596 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
601 * change_auth_cap_ses - move inode to appropriate lists when auth caps change
602 * @ci: inode to be moved
603 * @session: new auth caps session
605 static void change_auth_cap_ses(struct ceph_inode_info *ci,
606 struct ceph_mds_session *session)
608 lockdep_assert_held(&ci->i_ceph_lock);
610 if (list_empty(&ci->i_dirty_item) && list_empty(&ci->i_flushing_item))
613 spin_lock(&session->s_mdsc->cap_dirty_lock);
614 if (!list_empty(&ci->i_dirty_item))
615 list_move(&ci->i_dirty_item, &session->s_cap_dirty);
616 if (!list_empty(&ci->i_flushing_item))
617 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
618 spin_unlock(&session->s_mdsc->cap_dirty_lock);
622 * Add a capability under the given MDS session.
624 * Caller should hold session snap_rwsem (read) and ci->i_ceph_lock
626 * @fmode is the open file mode, if we are opening a file, otherwise
627 * it is < 0. (This is so we can atomically add the cap and add an
628 * open file reference to it.)
630 void ceph_add_cap(struct inode *inode,
631 struct ceph_mds_session *session, u64 cap_id,
632 unsigned issued, unsigned wanted,
633 unsigned seq, unsigned mseq, u64 realmino, int flags,
634 struct ceph_cap **new_cap)
636 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
637 struct ceph_inode_info *ci = ceph_inode(inode);
638 struct ceph_cap *cap;
639 int mds = session->s_mds;
643 lockdep_assert_held(&ci->i_ceph_lock);
645 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
646 session->s_mds, cap_id, ceph_cap_string(issued), seq);
648 spin_lock(&session->s_gen_ttl_lock);
649 gen = session->s_cap_gen;
650 spin_unlock(&session->s_gen_ttl_lock);
652 cap = __get_cap_for_mds(ci, mds);
658 cap->implemented = 0;
664 __insert_cap_node(ci, cap);
666 /* add to session cap list */
667 cap->session = session;
668 spin_lock(&session->s_cap_lock);
669 list_add_tail(&cap->session_caps, &session->s_caps);
670 session->s_nr_caps++;
671 atomic64_inc(&mdsc->metric.total_caps);
672 spin_unlock(&session->s_cap_lock);
674 spin_lock(&session->s_cap_lock);
675 list_move_tail(&cap->session_caps, &session->s_caps);
676 spin_unlock(&session->s_cap_lock);
678 if (cap->cap_gen < gen)
679 cap->issued = cap->implemented = CEPH_CAP_PIN;
682 * auth mds of the inode changed. we received the cap export
683 * message, but still haven't received the cap import message.
684 * handle_cap_export() updated the new auth MDS' cap.
686 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
687 * a message that was send before the cap import message. So
690 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
691 WARN_ON(cap != ci->i_auth_cap);
692 WARN_ON(cap->cap_id != cap_id);
695 issued |= cap->issued;
696 flags |= CEPH_CAP_FLAG_AUTH;
700 if (!ci->i_snap_realm ||
701 ((flags & CEPH_CAP_FLAG_AUTH) &&
702 realmino != (u64)-1 && ci->i_snap_realm->ino != realmino)) {
704 * add this inode to the appropriate snap realm
706 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
709 struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
711 spin_lock(&oldrealm->inodes_with_caps_lock);
712 list_del_init(&ci->i_snap_realm_item);
713 spin_unlock(&oldrealm->inodes_with_caps_lock);
716 spin_lock(&realm->inodes_with_caps_lock);
717 list_add(&ci->i_snap_realm_item,
718 &realm->inodes_with_caps);
719 ci->i_snap_realm = realm;
720 if (realm->ino == ci->i_vino.ino)
721 realm->inode = inode;
722 spin_unlock(&realm->inodes_with_caps_lock);
725 ceph_put_snap_realm(mdsc, oldrealm);
727 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
733 __check_cap_issue(ci, cap, issued);
736 * If we are issued caps we don't want, or the mds' wanted
737 * value appears to be off, queue a check so we'll release
738 * later and/or update the mds wanted value.
740 actual_wanted = __ceph_caps_wanted(ci);
741 if ((wanted & ~actual_wanted) ||
742 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
743 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
744 ceph_cap_string(issued), ceph_cap_string(wanted),
745 ceph_cap_string(actual_wanted));
746 __cap_delay_requeue(mdsc, ci);
749 if (flags & CEPH_CAP_FLAG_AUTH) {
750 if (!ci->i_auth_cap ||
751 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
752 if (ci->i_auth_cap &&
753 ci->i_auth_cap->session != cap->session)
754 change_auth_cap_ses(ci, cap->session);
755 ci->i_auth_cap = cap;
756 cap->mds_wanted = wanted;
759 WARN_ON(ci->i_auth_cap == cap);
762 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
763 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
764 ceph_cap_string(issued|cap->issued), seq, mds);
765 cap->cap_id = cap_id;
766 cap->issued = issued;
767 cap->implemented |= issued;
768 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
769 cap->mds_wanted = wanted;
771 cap->mds_wanted |= wanted;
773 cap->issue_seq = seq;
779 * Return true if cap has not timed out and belongs to the current
780 * generation of the MDS session (i.e. has not gone 'stale' due to
781 * us losing touch with the mds).
783 static int __cap_is_valid(struct ceph_cap *cap)
788 spin_lock(&cap->session->s_gen_ttl_lock);
789 gen = cap->session->s_cap_gen;
790 ttl = cap->session->s_cap_ttl;
791 spin_unlock(&cap->session->s_gen_ttl_lock);
793 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
794 dout("__cap_is_valid %p cap %p issued %s "
795 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
796 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
804 * Return set of valid cap bits issued to us. Note that caps time
805 * out, and may be invalidated in bulk if the client session times out
806 * and session->s_cap_gen is bumped.
808 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
810 int have = ci->i_snap_caps;
811 struct ceph_cap *cap;
816 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
817 cap = rb_entry(p, struct ceph_cap, ci_node);
818 if (!__cap_is_valid(cap))
820 dout("__ceph_caps_issued %p cap %p issued %s\n",
821 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
824 *implemented |= cap->implemented;
827 * exclude caps issued by non-auth MDS, but are been revoking
828 * by the auth MDS. The non-auth MDS should be revoking/exporting
829 * these caps, but the message is delayed.
831 if (ci->i_auth_cap) {
832 cap = ci->i_auth_cap;
833 have &= ~cap->implemented | cap->issued;
839 * Get cap bits issued by caps other than @ocap
841 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
843 int have = ci->i_snap_caps;
844 struct ceph_cap *cap;
847 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
848 cap = rb_entry(p, struct ceph_cap, ci_node);
851 if (!__cap_is_valid(cap))
859 * Move a cap to the end of the LRU (oldest caps at list head, newest
862 static void __touch_cap(struct ceph_cap *cap)
864 struct ceph_mds_session *s = cap->session;
866 spin_lock(&s->s_cap_lock);
867 if (!s->s_cap_iterator) {
868 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
870 list_move_tail(&cap->session_caps, &s->s_caps);
872 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
873 &cap->ci->vfs_inode, cap, s->s_mds);
875 spin_unlock(&s->s_cap_lock);
879 * Check if we hold the given mask. If so, move the cap(s) to the
880 * front of their respective LRUs. (This is the preferred way for
881 * callers to check for caps they want.)
883 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
885 struct ceph_cap *cap;
887 int have = ci->i_snap_caps;
889 if ((have & mask) == mask) {
890 dout("__ceph_caps_issued_mask ino 0x%llx snap issued %s"
891 " (mask %s)\n", ceph_ino(&ci->vfs_inode),
892 ceph_cap_string(have),
893 ceph_cap_string(mask));
897 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
898 cap = rb_entry(p, struct ceph_cap, ci_node);
899 if (!__cap_is_valid(cap))
901 if ((cap->issued & mask) == mask) {
902 dout("__ceph_caps_issued_mask ino 0x%llx cap %p issued %s"
903 " (mask %s)\n", ceph_ino(&ci->vfs_inode), cap,
904 ceph_cap_string(cap->issued),
905 ceph_cap_string(mask));
911 /* does a combination of caps satisfy mask? */
913 if ((have & mask) == mask) {
914 dout("__ceph_caps_issued_mask ino 0x%llx combo issued %s"
915 " (mask %s)\n", ceph_ino(&ci->vfs_inode),
916 ceph_cap_string(cap->issued),
917 ceph_cap_string(mask));
921 /* touch this + preceding caps */
923 for (q = rb_first(&ci->i_caps); q != p;
925 cap = rb_entry(q, struct ceph_cap,
927 if (!__cap_is_valid(cap))
929 if (cap->issued & mask)
940 int __ceph_caps_issued_mask_metric(struct ceph_inode_info *ci, int mask,
943 struct ceph_fs_client *fsc = ceph_sb_to_client(ci->vfs_inode.i_sb);
946 r = __ceph_caps_issued_mask(ci, mask, touch);
948 ceph_update_cap_hit(&fsc->mdsc->metric);
950 ceph_update_cap_mis(&fsc->mdsc->metric);
955 * Return true if mask caps are currently being revoked by an MDS.
957 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
958 struct ceph_cap *ocap, int mask)
960 struct ceph_cap *cap;
963 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
964 cap = rb_entry(p, struct ceph_cap, ci_node);
966 (cap->implemented & ~cap->issued & mask))
972 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
974 struct inode *inode = &ci->vfs_inode;
977 spin_lock(&ci->i_ceph_lock);
978 ret = __ceph_caps_revoking_other(ci, NULL, mask);
979 spin_unlock(&ci->i_ceph_lock);
980 dout("ceph_caps_revoking %p %s = %d\n", inode,
981 ceph_cap_string(mask), ret);
985 int __ceph_caps_used(struct ceph_inode_info *ci)
989 used |= CEPH_CAP_PIN;
991 used |= CEPH_CAP_FILE_RD;
992 if (ci->i_rdcache_ref ||
993 (S_ISREG(ci->vfs_inode.i_mode) &&
994 ci->vfs_inode.i_data.nrpages))
995 used |= CEPH_CAP_FILE_CACHE;
997 used |= CEPH_CAP_FILE_WR;
998 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
999 used |= CEPH_CAP_FILE_BUFFER;
1001 used |= CEPH_CAP_FILE_EXCL;
1005 #define FMODE_WAIT_BIAS 1000
1008 * wanted, by virtue of open file modes
1010 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
1012 const int PIN_SHIFT = ffs(CEPH_FILE_MODE_PIN);
1013 const int RD_SHIFT = ffs(CEPH_FILE_MODE_RD);
1014 const int WR_SHIFT = ffs(CEPH_FILE_MODE_WR);
1015 const int LAZY_SHIFT = ffs(CEPH_FILE_MODE_LAZY);
1016 struct ceph_mount_options *opt =
1017 ceph_inode_to_client(&ci->vfs_inode)->mount_options;
1018 unsigned long used_cutoff = jiffies - opt->caps_wanted_delay_max * HZ;
1019 unsigned long idle_cutoff = jiffies - opt->caps_wanted_delay_min * HZ;
1021 if (S_ISDIR(ci->vfs_inode.i_mode)) {
1024 /* use used_cutoff here, to keep dir's wanted caps longer */
1025 if (ci->i_nr_by_mode[RD_SHIFT] > 0 ||
1026 time_after(ci->i_last_rd, used_cutoff))
1027 want |= CEPH_CAP_ANY_SHARED;
1029 if (ci->i_nr_by_mode[WR_SHIFT] > 0 ||
1030 time_after(ci->i_last_wr, used_cutoff)) {
1031 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1032 if (opt->flags & CEPH_MOUNT_OPT_ASYNC_DIROPS)
1033 want |= CEPH_CAP_ANY_DIR_OPS;
1036 if (want || ci->i_nr_by_mode[PIN_SHIFT] > 0)
1037 want |= CEPH_CAP_PIN;
1043 if (ci->i_nr_by_mode[RD_SHIFT] > 0) {
1044 if (ci->i_nr_by_mode[RD_SHIFT] >= FMODE_WAIT_BIAS ||
1045 time_after(ci->i_last_rd, used_cutoff))
1046 bits |= 1 << RD_SHIFT;
1047 } else if (time_after(ci->i_last_rd, idle_cutoff)) {
1048 bits |= 1 << RD_SHIFT;
1051 if (ci->i_nr_by_mode[WR_SHIFT] > 0) {
1052 if (ci->i_nr_by_mode[WR_SHIFT] >= FMODE_WAIT_BIAS ||
1053 time_after(ci->i_last_wr, used_cutoff))
1054 bits |= 1 << WR_SHIFT;
1055 } else if (time_after(ci->i_last_wr, idle_cutoff)) {
1056 bits |= 1 << WR_SHIFT;
1059 /* check lazyio only when read/write is wanted */
1060 if ((bits & (CEPH_FILE_MODE_RDWR << 1)) &&
1061 ci->i_nr_by_mode[LAZY_SHIFT] > 0)
1062 bits |= 1 << LAZY_SHIFT;
1064 return bits ? ceph_caps_for_mode(bits >> 1) : 0;
1069 * wanted, by virtue of open file modes AND cap refs (buffered/cached data)
1071 int __ceph_caps_wanted(struct ceph_inode_info *ci)
1073 int w = __ceph_caps_file_wanted(ci) | __ceph_caps_used(ci);
1074 if (S_ISDIR(ci->vfs_inode.i_mode)) {
1075 /* we want EXCL if holding caps of dir ops */
1076 if (w & CEPH_CAP_ANY_DIR_OPS)
1077 w |= CEPH_CAP_FILE_EXCL;
1079 /* we want EXCL if dirty data */
1080 if (w & CEPH_CAP_FILE_BUFFER)
1081 w |= CEPH_CAP_FILE_EXCL;
1087 * Return caps we have registered with the MDS(s) as 'wanted'.
1089 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
1091 struct ceph_cap *cap;
1095 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1096 cap = rb_entry(p, struct ceph_cap, ci_node);
1097 if (check && !__cap_is_valid(cap))
1099 if (cap == ci->i_auth_cap)
1100 mds_wanted |= cap->mds_wanted;
1102 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
1107 int ceph_is_any_caps(struct inode *inode)
1109 struct ceph_inode_info *ci = ceph_inode(inode);
1112 spin_lock(&ci->i_ceph_lock);
1113 ret = __ceph_is_any_real_caps(ci);
1114 spin_unlock(&ci->i_ceph_lock);
1119 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
1121 struct ceph_snap_realm *realm = ci->i_snap_realm;
1122 spin_lock(&realm->inodes_with_caps_lock);
1123 list_del_init(&ci->i_snap_realm_item);
1124 ci->i_snap_realm_counter++;
1125 ci->i_snap_realm = NULL;
1126 if (realm->ino == ci->i_vino.ino)
1127 realm->inode = NULL;
1128 spin_unlock(&realm->inodes_with_caps_lock);
1129 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
1134 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
1136 * caller should hold i_ceph_lock.
1137 * caller will not hold session s_mutex if called from destroy_inode.
1139 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1141 struct ceph_mds_session *session = cap->session;
1142 struct ceph_inode_info *ci = cap->ci;
1143 struct ceph_mds_client *mdsc;
1146 /* 'ci' being NULL means the remove have already occurred */
1148 dout("%s: cap inode is NULL\n", __func__);
1152 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1154 mdsc = ceph_inode_to_client(&ci->vfs_inode)->mdsc;
1156 /* remove from inode's cap rbtree, and clear auth cap */
1157 rb_erase(&cap->ci_node, &ci->i_caps);
1158 if (ci->i_auth_cap == cap) {
1159 WARN_ON_ONCE(!list_empty(&ci->i_dirty_item));
1160 ci->i_auth_cap = NULL;
1163 /* remove from session list */
1164 spin_lock(&session->s_cap_lock);
1165 if (session->s_cap_iterator == cap) {
1166 /* not yet, we are iterating over this very cap */
1167 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1170 list_del_init(&cap->session_caps);
1171 session->s_nr_caps--;
1172 atomic64_dec(&mdsc->metric.total_caps);
1173 cap->session = NULL;
1176 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1180 * s_cap_reconnect is protected by s_cap_lock. no one changes
1181 * s_cap_gen while session is in the reconnect state.
1183 if (queue_release &&
1184 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
1185 cap->queue_release = 1;
1187 __ceph_queue_cap_release(session, cap);
1191 cap->queue_release = 0;
1193 cap->cap_ino = ci->i_vino.ino;
1195 spin_unlock(&session->s_cap_lock);
1198 ceph_put_cap(mdsc, cap);
1200 if (!__ceph_is_any_real_caps(ci)) {
1201 /* when reconnect denied, we remove session caps forcibly,
1202 * i_wr_ref can be non-zero. If there are ongoing write,
1203 * keep i_snap_realm.
1205 if (ci->i_wr_ref == 0 && ci->i_snap_realm)
1206 drop_inode_snap_realm(ci);
1208 __cap_delay_cancel(mdsc, ci);
1212 struct cap_msg_args {
1213 struct ceph_mds_session *session;
1214 u64 ino, cid, follows;
1215 u64 flush_tid, oldest_flush_tid, size, max_size;
1218 struct ceph_buffer *xattr_buf;
1219 struct ceph_buffer *old_xattr_buf;
1220 struct timespec64 atime, mtime, ctime, btime;
1221 int op, caps, wanted, dirty;
1222 u32 seq, issue_seq, mseq, time_warp_seq;
1232 * cap struct size + flock buffer size + inline version + inline data size +
1233 * osd_epoch_barrier + oldest_flush_tid
1235 #define CAP_MSG_SIZE (sizeof(struct ceph_mds_caps) + \
1236 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4)
1238 /* Marshal up the cap msg to the MDS */
1239 static void encode_cap_msg(struct ceph_msg *msg, struct cap_msg_args *arg)
1241 struct ceph_mds_caps *fc;
1243 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1245 dout("%s %s %llx %llx caps %s wanted %s dirty %s seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu xattr_ver %llu xattr_len %d\n",
1246 __func__, ceph_cap_op_name(arg->op), arg->cid, arg->ino,
1247 ceph_cap_string(arg->caps), ceph_cap_string(arg->wanted),
1248 ceph_cap_string(arg->dirty), arg->seq, arg->issue_seq,
1249 arg->flush_tid, arg->oldest_flush_tid, arg->mseq, arg->follows,
1250 arg->size, arg->max_size, arg->xattr_version,
1251 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1253 msg->hdr.version = cpu_to_le16(10);
1254 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1256 fc = msg->front.iov_base;
1257 memset(fc, 0, sizeof(*fc));
1259 fc->cap_id = cpu_to_le64(arg->cid);
1260 fc->op = cpu_to_le32(arg->op);
1261 fc->seq = cpu_to_le32(arg->seq);
1262 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1263 fc->migrate_seq = cpu_to_le32(arg->mseq);
1264 fc->caps = cpu_to_le32(arg->caps);
1265 fc->wanted = cpu_to_le32(arg->wanted);
1266 fc->dirty = cpu_to_le32(arg->dirty);
1267 fc->ino = cpu_to_le64(arg->ino);
1268 fc->snap_follows = cpu_to_le64(arg->follows);
1270 fc->size = cpu_to_le64(arg->size);
1271 fc->max_size = cpu_to_le64(arg->max_size);
1272 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1273 ceph_encode_timespec64(&fc->atime, &arg->atime);
1274 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1275 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1277 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1278 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1279 fc->mode = cpu_to_le32(arg->mode);
1281 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1282 if (arg->xattr_buf) {
1283 msg->middle = ceph_buffer_get(arg->xattr_buf);
1284 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1285 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1289 /* flock buffer size (version 2) */
1290 ceph_encode_32(&p, 0);
1291 /* inline version (version 4) */
1292 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1293 /* inline data size */
1294 ceph_encode_32(&p, 0);
1296 * osd_epoch_barrier (version 5)
1297 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1298 * case it was recently changed
1300 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1301 /* oldest_flush_tid (version 6) */
1302 ceph_encode_64(&p, arg->oldest_flush_tid);
1305 * caller_uid/caller_gid (version 7)
1307 * Currently, we don't properly track which caller dirtied the caps
1308 * last, and force a flush of them when there is a conflict. For now,
1309 * just set this to 0:0, to emulate how the MDS has worked up to now.
1311 ceph_encode_32(&p, 0);
1312 ceph_encode_32(&p, 0);
1314 /* pool namespace (version 8) (mds always ignores this) */
1315 ceph_encode_32(&p, 0);
1317 /* btime and change_attr (version 9) */
1318 ceph_encode_timespec64(p, &arg->btime);
1319 p += sizeof(struct ceph_timespec);
1320 ceph_encode_64(&p, arg->change_attr);
1322 /* Advisory flags (version 10) */
1323 ceph_encode_32(&p, arg->flags);
1327 * Queue cap releases when an inode is dropped from our cache.
1329 void __ceph_remove_caps(struct ceph_inode_info *ci)
1333 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1334 * may call __ceph_caps_issued_mask() on a freeing inode. */
1335 spin_lock(&ci->i_ceph_lock);
1336 p = rb_first(&ci->i_caps);
1338 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1340 __ceph_remove_cap(cap, true);
1342 spin_unlock(&ci->i_ceph_lock);
1346 * Prepare to send a cap message to an MDS. Update the cap state, and populate
1347 * the arg struct with the parameters that will need to be sent. This should
1348 * be done under the i_ceph_lock to guard against changes to cap state.
1350 * Make note of max_size reported/requested from mds, revoked caps
1351 * that have now been implemented.
1353 static void __prep_cap(struct cap_msg_args *arg, struct ceph_cap *cap,
1354 int op, int flags, int used, int want, int retain,
1355 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1357 struct ceph_inode_info *ci = cap->ci;
1358 struct inode *inode = &ci->vfs_inode;
1361 lockdep_assert_held(&ci->i_ceph_lock);
1363 held = cap->issued | cap->implemented;
1364 revoking = cap->implemented & ~cap->issued;
1365 retain &= ~revoking;
1367 dout("%s %p cap %p session %p %s -> %s (revoking %s)\n",
1368 __func__, inode, cap, cap->session,
1369 ceph_cap_string(held), ceph_cap_string(held & retain),
1370 ceph_cap_string(revoking));
1371 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1373 ci->i_ceph_flags &= ~CEPH_I_FLUSH;
1375 cap->issued &= retain; /* drop bits we don't want */
1377 * Wake up any waiters on wanted -> needed transition. This is due to
1378 * the weird transition from buffered to sync IO... we need to flush
1379 * dirty pages _before_ allowing sync writes to avoid reordering.
1381 arg->wake = cap->implemented & ~cap->issued;
1382 cap->implemented &= cap->issued | used;
1383 cap->mds_wanted = want;
1385 arg->session = cap->session;
1386 arg->ino = ceph_vino(inode).ino;
1387 arg->cid = cap->cap_id;
1388 arg->follows = flushing ? ci->i_head_snapc->seq : 0;
1389 arg->flush_tid = flush_tid;
1390 arg->oldest_flush_tid = oldest_flush_tid;
1392 arg->size = inode->i_size;
1393 ci->i_reported_size = arg->size;
1394 arg->max_size = ci->i_wanted_max_size;
1395 if (cap == ci->i_auth_cap) {
1396 if (want & CEPH_CAP_ANY_FILE_WR)
1397 ci->i_requested_max_size = arg->max_size;
1399 ci->i_requested_max_size = 0;
1402 if (flushing & CEPH_CAP_XATTR_EXCL) {
1403 arg->old_xattr_buf = __ceph_build_xattrs_blob(ci);
1404 arg->xattr_version = ci->i_xattrs.version;
1405 arg->xattr_buf = ceph_buffer_get(ci->i_xattrs.blob);
1407 arg->xattr_buf = NULL;
1408 arg->old_xattr_buf = NULL;
1411 arg->mtime = inode->i_mtime;
1412 arg->atime = inode->i_atime;
1413 arg->ctime = inode->i_ctime;
1414 arg->btime = ci->i_btime;
1415 arg->change_attr = inode_peek_iversion_raw(inode);
1418 arg->caps = cap->implemented;
1420 arg->dirty = flushing;
1422 arg->seq = cap->seq;
1423 arg->issue_seq = cap->issue_seq;
1424 arg->mseq = cap->mseq;
1425 arg->time_warp_seq = ci->i_time_warp_seq;
1427 arg->uid = inode->i_uid;
1428 arg->gid = inode->i_gid;
1429 arg->mode = inode->i_mode;
1431 arg->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1432 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1433 !list_empty(&ci->i_cap_snaps)) {
1434 struct ceph_cap_snap *capsnap;
1435 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1436 if (capsnap->cap_flush.tid)
1438 if (capsnap->need_flush) {
1439 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1448 * Send a cap msg on the given inode.
1450 * Caller should hold snap_rwsem (read), s_mutex.
1452 static void __send_cap(struct cap_msg_args *arg, struct ceph_inode_info *ci)
1454 struct ceph_msg *msg;
1455 struct inode *inode = &ci->vfs_inode;
1457 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1459 pr_err("error allocating cap msg: ino (%llx.%llx) flushing %s tid %llu, requeuing cap.\n",
1460 ceph_vinop(inode), ceph_cap_string(arg->dirty),
1462 spin_lock(&ci->i_ceph_lock);
1463 __cap_delay_requeue(arg->session->s_mdsc, ci);
1464 spin_unlock(&ci->i_ceph_lock);
1468 encode_cap_msg(msg, arg);
1469 ceph_con_send(&arg->session->s_con, msg);
1470 ceph_buffer_put(arg->old_xattr_buf);
1471 ceph_buffer_put(arg->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;
1640 bool need_put = false;
1643 dout("ceph_flush_snaps %p\n", inode);
1645 session = *psession;
1647 spin_lock(&ci->i_ceph_lock);
1648 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1649 dout(" no capsnap needs flush, doing nothing\n");
1652 if (!ci->i_auth_cap) {
1653 dout(" no auth cap (migrating?), doing nothing\n");
1657 mds = ci->i_auth_cap->session->s_mds;
1658 if (session && session->s_mds != mds) {
1659 dout(" oops, wrong session %p mutex\n", session);
1660 mutex_unlock(&session->s_mutex);
1661 ceph_put_mds_session(session);
1665 spin_unlock(&ci->i_ceph_lock);
1666 mutex_lock(&mdsc->mutex);
1667 session = __ceph_lookup_mds_session(mdsc, mds);
1668 mutex_unlock(&mdsc->mutex);
1670 dout(" inverting session/ino locks on %p\n", session);
1671 mutex_lock(&session->s_mutex);
1676 // make sure flushsnap messages are sent in proper order.
1677 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1678 __kick_flushing_caps(mdsc, session, ci, 0);
1680 __ceph_flush_snaps(ci, session);
1682 spin_unlock(&ci->i_ceph_lock);
1685 *psession = session;
1686 } else if (session) {
1687 mutex_unlock(&session->s_mutex);
1688 ceph_put_mds_session(session);
1690 /* we flushed them all; remove this inode from the queue */
1691 spin_lock(&mdsc->snap_flush_lock);
1692 if (!list_empty(&ci->i_snap_flush_item))
1694 list_del_init(&ci->i_snap_flush_item);
1695 spin_unlock(&mdsc->snap_flush_lock);
1702 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1703 * Caller is then responsible for calling __mark_inode_dirty with the
1704 * returned flags value.
1706 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1707 struct ceph_cap_flush **pcf)
1709 struct ceph_mds_client *mdsc =
1710 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1711 struct inode *inode = &ci->vfs_inode;
1712 int was = ci->i_dirty_caps;
1715 lockdep_assert_held(&ci->i_ceph_lock);
1717 if (!ci->i_auth_cap) {
1718 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1719 "but no auth cap (session was closed?)\n",
1720 inode, ceph_ino(inode), ceph_cap_string(mask));
1724 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1725 ceph_cap_string(mask), ceph_cap_string(was),
1726 ceph_cap_string(was | mask));
1727 ci->i_dirty_caps |= mask;
1729 struct ceph_mds_session *session = ci->i_auth_cap->session;
1731 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1732 swap(ci->i_prealloc_cap_flush, *pcf);
1734 if (!ci->i_head_snapc) {
1735 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1736 ci->i_head_snapc = ceph_get_snap_context(
1737 ci->i_snap_realm->cached_context);
1739 dout(" inode %p now dirty snapc %p auth cap %p\n",
1740 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1741 BUG_ON(!list_empty(&ci->i_dirty_item));
1742 spin_lock(&mdsc->cap_dirty_lock);
1743 list_add(&ci->i_dirty_item, &session->s_cap_dirty);
1744 spin_unlock(&mdsc->cap_dirty_lock);
1745 if (ci->i_flushing_caps == 0) {
1747 dirty |= I_DIRTY_SYNC;
1750 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1752 BUG_ON(list_empty(&ci->i_dirty_item));
1753 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1754 (mask & CEPH_CAP_FILE_BUFFER))
1755 dirty |= I_DIRTY_DATASYNC;
1756 __cap_delay_requeue(mdsc, ci);
1760 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1762 struct ceph_cap_flush *cf;
1764 cf = kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1768 cf->is_capsnap = false;
1772 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1775 kmem_cache_free(ceph_cap_flush_cachep, cf);
1778 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1780 if (!list_empty(&mdsc->cap_flush_list)) {
1781 struct ceph_cap_flush *cf =
1782 list_first_entry(&mdsc->cap_flush_list,
1783 struct ceph_cap_flush, g_list);
1790 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1791 * Return true if caller needs to wake up flush waiters.
1793 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1794 struct ceph_cap_flush *cf)
1796 struct ceph_cap_flush *prev;
1797 bool wake = cf->wake;
1799 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1800 prev = list_prev_entry(cf, g_list);
1804 list_del_init(&cf->g_list);
1808 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1809 struct ceph_cap_flush *cf)
1811 struct ceph_cap_flush *prev;
1812 bool wake = cf->wake;
1814 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1815 prev = list_prev_entry(cf, i_list);
1819 list_del_init(&cf->i_list);
1824 * Add dirty inode to the flushing list. Assigned a seq number so we
1825 * can wait for caps to flush without starving.
1827 * Called under i_ceph_lock. Returns the flush tid.
1829 static u64 __mark_caps_flushing(struct inode *inode,
1830 struct ceph_mds_session *session, bool wake,
1831 u64 *oldest_flush_tid)
1833 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1834 struct ceph_inode_info *ci = ceph_inode(inode);
1835 struct ceph_cap_flush *cf = NULL;
1838 lockdep_assert_held(&ci->i_ceph_lock);
1839 BUG_ON(ci->i_dirty_caps == 0);
1840 BUG_ON(list_empty(&ci->i_dirty_item));
1841 BUG_ON(!ci->i_prealloc_cap_flush);
1843 flushing = ci->i_dirty_caps;
1844 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1845 ceph_cap_string(flushing),
1846 ceph_cap_string(ci->i_flushing_caps),
1847 ceph_cap_string(ci->i_flushing_caps | flushing));
1848 ci->i_flushing_caps |= flushing;
1849 ci->i_dirty_caps = 0;
1850 dout(" inode %p now !dirty\n", inode);
1852 swap(cf, ci->i_prealloc_cap_flush);
1853 cf->caps = flushing;
1856 spin_lock(&mdsc->cap_dirty_lock);
1857 list_del_init(&ci->i_dirty_item);
1859 cf->tid = ++mdsc->last_cap_flush_tid;
1860 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1861 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1863 if (list_empty(&ci->i_flushing_item)) {
1864 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1865 mdsc->num_cap_flushing++;
1867 spin_unlock(&mdsc->cap_dirty_lock);
1869 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1875 * try to invalidate mapping pages without blocking.
1877 static int try_nonblocking_invalidate(struct inode *inode)
1878 __releases(ci->i_ceph_lock)
1879 __acquires(ci->i_ceph_lock)
1881 struct ceph_inode_info *ci = ceph_inode(inode);
1882 u32 invalidating_gen = ci->i_rdcache_gen;
1884 spin_unlock(&ci->i_ceph_lock);
1885 ceph_fscache_invalidate(inode);
1886 invalidate_mapping_pages(&inode->i_data, 0, -1);
1887 spin_lock(&ci->i_ceph_lock);
1889 if (inode->i_data.nrpages == 0 &&
1890 invalidating_gen == ci->i_rdcache_gen) {
1892 dout("try_nonblocking_invalidate %p success\n", inode);
1893 /* save any racing async invalidate some trouble */
1894 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1897 dout("try_nonblocking_invalidate %p failed\n", inode);
1901 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1903 loff_t size = ci->vfs_inode.i_size;
1904 /* mds will adjust max size according to the reported size */
1905 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1907 if (size >= ci->i_max_size)
1909 /* half of previous max_size increment has been used */
1910 if (ci->i_max_size > ci->i_reported_size &&
1911 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1917 * Swiss army knife function to examine currently used and wanted
1918 * versus held caps. Release, flush, ack revoked caps to mds as
1921 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1922 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1925 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1926 struct ceph_mds_session *session)
1928 struct inode *inode = &ci->vfs_inode;
1929 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1930 struct ceph_cap *cap;
1931 u64 flush_tid, oldest_flush_tid;
1932 int file_wanted, used, cap_used;
1933 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1934 int issued, implemented, want, retain, revoking, flushing = 0;
1935 int mds = -1; /* keep track of how far we've gone through i_caps list
1936 to avoid an infinite loop on retry */
1938 bool queue_invalidate = false;
1939 bool tried_invalidate = false;
1941 spin_lock(&ci->i_ceph_lock);
1942 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1943 flags |= CHECK_CAPS_FLUSH;
1947 spin_lock(&ci->i_ceph_lock);
1949 /* Caps wanted by virtue of active open files. */
1950 file_wanted = __ceph_caps_file_wanted(ci);
1952 /* Caps which have active references against them */
1953 used = __ceph_caps_used(ci);
1956 * "issued" represents the current caps that the MDS wants us to have.
1957 * "implemented" is the set that we have been granted, and includes the
1958 * ones that have not yet been returned to the MDS (the "revoking" set,
1959 * usually because they have outstanding references).
1961 issued = __ceph_caps_issued(ci, &implemented);
1962 revoking = implemented & ~issued;
1966 /* The ones we currently want to retain (may be adjusted below) */
1967 retain = file_wanted | used | CEPH_CAP_PIN;
1968 if (!mdsc->stopping && inode->i_nlink > 0) {
1970 retain |= CEPH_CAP_ANY; /* be greedy */
1971 } else if (S_ISDIR(inode->i_mode) &&
1972 (issued & CEPH_CAP_FILE_SHARED) &&
1973 __ceph_dir_is_complete(ci)) {
1975 * If a directory is complete, we want to keep
1976 * the exclusive cap. So that MDS does not end up
1977 * revoking the shared cap on every create/unlink
1980 if (IS_RDONLY(inode)) {
1981 want = CEPH_CAP_ANY_SHARED;
1983 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1988 retain |= CEPH_CAP_ANY_SHARED;
1990 * keep RD only if we didn't have the file open RW,
1991 * because then the mds would revoke it anyway to
1992 * journal max_size=0.
1994 if (ci->i_max_size == 0)
1995 retain |= CEPH_CAP_ANY_RD;
1999 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
2000 " issued %s revoking %s retain %s %s%s\n", inode,
2001 ceph_cap_string(file_wanted),
2002 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
2003 ceph_cap_string(ci->i_flushing_caps),
2004 ceph_cap_string(issued), ceph_cap_string(revoking),
2005 ceph_cap_string(retain),
2006 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
2007 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
2010 * If we no longer need to hold onto old our caps, and we may
2011 * have cached pages, but don't want them, then try to invalidate.
2012 * If we fail, it's because pages are locked.... try again later.
2014 if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
2015 S_ISREG(inode->i_mode) &&
2016 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
2017 inode->i_data.nrpages && /* have cached pages */
2018 (revoking & (CEPH_CAP_FILE_CACHE|
2019 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
2020 !tried_invalidate) {
2021 dout("check_caps trying to invalidate on %p\n", inode);
2022 if (try_nonblocking_invalidate(inode) < 0) {
2023 dout("check_caps queuing invalidate\n");
2024 queue_invalidate = true;
2025 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2027 tried_invalidate = true;
2031 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2033 struct cap_msg_args arg;
2035 cap = rb_entry(p, struct ceph_cap, ci_node);
2037 /* avoid looping forever */
2038 if (mds >= cap->mds ||
2039 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
2042 /* NOTE: no side-effects allowed, until we take s_mutex */
2045 * If we have an auth cap, we don't need to consider any
2046 * overlapping caps as used.
2049 if (ci->i_auth_cap && cap != ci->i_auth_cap)
2050 cap_used &= ~ci->i_auth_cap->issued;
2052 revoking = cap->implemented & ~cap->issued;
2053 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
2054 cap->mds, cap, ceph_cap_string(cap_used),
2055 ceph_cap_string(cap->issued),
2056 ceph_cap_string(cap->implemented),
2057 ceph_cap_string(revoking));
2059 if (cap == ci->i_auth_cap &&
2060 (cap->issued & CEPH_CAP_FILE_WR)) {
2061 /* request larger max_size from MDS? */
2062 if (ci->i_wanted_max_size > ci->i_max_size &&
2063 ci->i_wanted_max_size > ci->i_requested_max_size) {
2064 dout("requesting new max_size\n");
2068 /* approaching file_max? */
2069 if (__ceph_should_report_size(ci)) {
2070 dout("i_size approaching max_size\n");
2074 /* flush anything dirty? */
2075 if (cap == ci->i_auth_cap) {
2076 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
2077 dout("flushing dirty caps\n");
2080 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
2081 dout("flushing snap caps\n");
2086 /* completed revocation? going down and there are no caps? */
2087 if (revoking && (revoking & cap_used) == 0) {
2088 dout("completed revocation of %s\n",
2089 ceph_cap_string(cap->implemented & ~cap->issued));
2093 /* want more caps from mds? */
2094 if (want & ~cap->mds_wanted) {
2095 if (want & ~(cap->mds_wanted | cap->issued))
2097 if (!__cap_is_valid(cap))
2101 /* things we might delay */
2102 if ((cap->issued & ~retain) == 0)
2103 continue; /* nope, all good */
2106 if (session && session != cap->session) {
2107 dout("oops, wrong session %p mutex\n", session);
2108 mutex_unlock(&session->s_mutex);
2112 session = cap->session;
2113 if (mutex_trylock(&session->s_mutex) == 0) {
2114 dout("inverting session/ino locks on %p\n",
2116 session = ceph_get_mds_session(session);
2117 spin_unlock(&ci->i_ceph_lock);
2118 if (took_snap_rwsem) {
2119 up_read(&mdsc->snap_rwsem);
2120 took_snap_rwsem = 0;
2123 mutex_lock(&session->s_mutex);
2124 ceph_put_mds_session(session);
2127 * Because we take the reference while
2128 * holding the i_ceph_lock, it should
2129 * never be NULL. Throw a warning if it
2138 /* kick flushing and flush snaps before sending normal
2140 if (cap == ci->i_auth_cap &&
2142 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2143 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2144 __kick_flushing_caps(mdsc, session, ci, 0);
2145 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2146 __ceph_flush_snaps(ci, session);
2151 /* take snap_rwsem after session mutex */
2152 if (!took_snap_rwsem) {
2153 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
2154 dout("inverting snap/in locks on %p\n",
2156 spin_unlock(&ci->i_ceph_lock);
2157 down_read(&mdsc->snap_rwsem);
2158 took_snap_rwsem = 1;
2161 took_snap_rwsem = 1;
2164 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2165 flushing = ci->i_dirty_caps;
2166 flush_tid = __mark_caps_flushing(inode, session, false,
2168 if (flags & CHECK_CAPS_FLUSH &&
2169 list_empty(&session->s_cap_dirty))
2170 mflags |= CEPH_CLIENT_CAPS_SYNC;
2174 spin_lock(&mdsc->cap_dirty_lock);
2175 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2176 spin_unlock(&mdsc->cap_dirty_lock);
2179 mds = cap->mds; /* remember mds, so we don't repeat */
2181 __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2182 want, retain, flushing, flush_tid, oldest_flush_tid);
2183 spin_unlock(&ci->i_ceph_lock);
2185 __send_cap(&arg, ci);
2187 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2190 /* periodically re-calculate caps wanted by open files */
2191 if (__ceph_is_any_real_caps(ci) &&
2192 list_empty(&ci->i_cap_delay_list) &&
2193 (file_wanted & ~CEPH_CAP_PIN) &&
2194 !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2195 __cap_delay_requeue(mdsc, ci);
2198 spin_unlock(&ci->i_ceph_lock);
2200 if (queue_invalidate)
2201 ceph_queue_invalidate(inode);
2204 mutex_unlock(&session->s_mutex);
2205 if (took_snap_rwsem)
2206 up_read(&mdsc->snap_rwsem);
2210 * Try to flush dirty caps back to the auth mds.
2212 static int try_flush_caps(struct inode *inode, u64 *ptid)
2214 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2215 struct ceph_inode_info *ci = ceph_inode(inode);
2216 struct ceph_mds_session *session = NULL;
2218 u64 flush_tid = 0, oldest_flush_tid = 0;
2221 spin_lock(&ci->i_ceph_lock);
2223 if (ci->i_dirty_caps && ci->i_auth_cap) {
2224 struct ceph_cap *cap = ci->i_auth_cap;
2225 struct cap_msg_args arg;
2227 if (session != cap->session) {
2228 spin_unlock(&ci->i_ceph_lock);
2230 mutex_unlock(&session->s_mutex);
2231 session = cap->session;
2232 mutex_lock(&session->s_mutex);
2235 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) {
2236 spin_unlock(&ci->i_ceph_lock);
2240 if (ci->i_ceph_flags &
2241 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2242 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2243 __kick_flushing_caps(mdsc, session, ci, 0);
2244 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2245 __ceph_flush_snaps(ci, session);
2249 flushing = ci->i_dirty_caps;
2250 flush_tid = __mark_caps_flushing(inode, session, true,
2253 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2254 __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2255 (cap->issued | cap->implemented),
2256 flushing, flush_tid, oldest_flush_tid);
2257 spin_unlock(&ci->i_ceph_lock);
2259 __send_cap(&arg, ci);
2261 if (!list_empty(&ci->i_cap_flush_list)) {
2262 struct ceph_cap_flush *cf =
2263 list_last_entry(&ci->i_cap_flush_list,
2264 struct ceph_cap_flush, i_list);
2266 flush_tid = cf->tid;
2268 flushing = ci->i_flushing_caps;
2269 spin_unlock(&ci->i_ceph_lock);
2273 mutex_unlock(&session->s_mutex);
2280 * Return true if we've flushed caps through the given flush_tid.
2282 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2284 struct ceph_inode_info *ci = ceph_inode(inode);
2287 spin_lock(&ci->i_ceph_lock);
2288 if (!list_empty(&ci->i_cap_flush_list)) {
2289 struct ceph_cap_flush * cf =
2290 list_first_entry(&ci->i_cap_flush_list,
2291 struct ceph_cap_flush, i_list);
2292 if (cf->tid <= flush_tid)
2295 spin_unlock(&ci->i_ceph_lock);
2300 * wait for any unsafe requests to complete.
2302 static int unsafe_request_wait(struct inode *inode)
2304 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2305 struct ceph_inode_info *ci = ceph_inode(inode);
2306 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2309 spin_lock(&ci->i_unsafe_lock);
2310 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2311 req1 = list_last_entry(&ci->i_unsafe_dirops,
2312 struct ceph_mds_request,
2314 ceph_mdsc_get_request(req1);
2316 if (!list_empty(&ci->i_unsafe_iops)) {
2317 req2 = list_last_entry(&ci->i_unsafe_iops,
2318 struct ceph_mds_request,
2319 r_unsafe_target_item);
2320 ceph_mdsc_get_request(req2);
2322 spin_unlock(&ci->i_unsafe_lock);
2325 * Trigger to flush the journal logs in all the relevant MDSes
2326 * manually, or in the worst case we must wait at most 5 seconds
2327 * to wait the journal logs to be flushed by the MDSes periodically.
2330 struct ceph_mds_request *req;
2331 struct ceph_mds_session **sessions;
2332 struct ceph_mds_session *s;
2333 unsigned int max_sessions;
2336 mutex_lock(&mdsc->mutex);
2337 max_sessions = mdsc->max_sessions;
2339 sessions = kcalloc(max_sessions, sizeof(s), GFP_KERNEL);
2341 mutex_unlock(&mdsc->mutex);
2346 spin_lock(&ci->i_unsafe_lock);
2348 list_for_each_entry(req, &ci->i_unsafe_dirops,
2349 r_unsafe_dir_item) {
2353 if (!sessions[s->s_mds]) {
2354 s = ceph_get_mds_session(s);
2355 sessions[s->s_mds] = s;
2360 list_for_each_entry(req, &ci->i_unsafe_iops,
2361 r_unsafe_target_item) {
2365 if (!sessions[s->s_mds]) {
2366 s = ceph_get_mds_session(s);
2367 sessions[s->s_mds] = s;
2371 spin_unlock(&ci->i_unsafe_lock);
2374 spin_lock(&ci->i_ceph_lock);
2375 if (ci->i_auth_cap) {
2376 s = ci->i_auth_cap->session;
2377 if (!sessions[s->s_mds])
2378 sessions[s->s_mds] = ceph_get_mds_session(s);
2380 spin_unlock(&ci->i_ceph_lock);
2381 mutex_unlock(&mdsc->mutex);
2383 /* send flush mdlog request to MDSes */
2384 for (i = 0; i < max_sessions; i++) {
2387 send_flush_mdlog(s);
2388 ceph_put_mds_session(s);
2394 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2395 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2397 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2398 ceph_timeout_jiffies(req1->r_timeout));
2403 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2404 ceph_timeout_jiffies(req2->r_timeout));
2411 ceph_mdsc_put_request(req1);
2413 ceph_mdsc_put_request(req2);
2417 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2419 struct inode *inode = file->f_mapping->host;
2420 struct ceph_inode_info *ci = ceph_inode(inode);
2425 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2427 ret = file_write_and_wait_range(file, start, end);
2431 ret = ceph_wait_on_async_create(inode);
2435 dirty = try_flush_caps(inode, &flush_tid);
2436 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2438 err = unsafe_request_wait(inode);
2441 * only wait on non-file metadata writeback (the mds
2442 * can recover size and mtime, so we don't need to
2445 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2446 err = wait_event_interruptible(ci->i_cap_wq,
2447 caps_are_flushed(inode, flush_tid));
2453 err = file_check_and_advance_wb_err(file);
2457 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2462 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2463 * queue inode for flush but don't do so immediately, because we can
2464 * get by with fewer MDS messages if we wait for data writeback to
2467 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2469 struct ceph_inode_info *ci = ceph_inode(inode);
2473 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2475 dout("write_inode %p wait=%d\n", inode, wait);
2477 dirty = try_flush_caps(inode, &flush_tid);
2479 err = wait_event_interruptible(ci->i_cap_wq,
2480 caps_are_flushed(inode, flush_tid));
2482 struct ceph_mds_client *mdsc =
2483 ceph_sb_to_client(inode->i_sb)->mdsc;
2485 spin_lock(&ci->i_ceph_lock);
2486 if (__ceph_caps_dirty(ci))
2487 __cap_delay_requeue_front(mdsc, ci);
2488 spin_unlock(&ci->i_ceph_lock);
2493 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2494 struct ceph_mds_session *session,
2495 struct ceph_inode_info *ci,
2496 u64 oldest_flush_tid)
2497 __releases(ci->i_ceph_lock)
2498 __acquires(ci->i_ceph_lock)
2500 struct inode *inode = &ci->vfs_inode;
2501 struct ceph_cap *cap;
2502 struct ceph_cap_flush *cf;
2505 u64 last_snap_flush = 0;
2507 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2509 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2510 if (cf->is_capsnap) {
2511 last_snap_flush = cf->tid;
2516 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2517 if (cf->tid < first_tid)
2520 cap = ci->i_auth_cap;
2521 if (!(cap && cap->session == session)) {
2522 pr_err("%p auth cap %p not mds%d ???\n",
2523 inode, cap, session->s_mds);
2527 first_tid = cf->tid + 1;
2529 if (!cf->is_capsnap) {
2530 struct cap_msg_args arg;
2532 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2533 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2534 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2535 (cf->tid < last_snap_flush ?
2536 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2537 __ceph_caps_used(ci),
2538 __ceph_caps_wanted(ci),
2539 (cap->issued | cap->implemented),
2540 cf->caps, cf->tid, oldest_flush_tid);
2541 spin_unlock(&ci->i_ceph_lock);
2542 __send_cap(&arg, ci);
2544 struct ceph_cap_snap *capsnap =
2545 container_of(cf, struct ceph_cap_snap,
2547 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2548 inode, capsnap, cf->tid,
2549 ceph_cap_string(capsnap->dirty));
2551 refcount_inc(&capsnap->nref);
2552 spin_unlock(&ci->i_ceph_lock);
2554 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2557 pr_err("kick_flushing_caps: error sending "
2558 "cap flushsnap, ino (%llx.%llx) "
2559 "tid %llu follows %llu\n",
2560 ceph_vinop(inode), cf->tid,
2564 ceph_put_cap_snap(capsnap);
2567 spin_lock(&ci->i_ceph_lock);
2571 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2572 struct ceph_mds_session *session)
2574 struct ceph_inode_info *ci;
2575 struct ceph_cap *cap;
2576 u64 oldest_flush_tid;
2578 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2580 spin_lock(&mdsc->cap_dirty_lock);
2581 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2582 spin_unlock(&mdsc->cap_dirty_lock);
2584 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2585 spin_lock(&ci->i_ceph_lock);
2586 cap = ci->i_auth_cap;
2587 if (!(cap && cap->session == session)) {
2588 pr_err("%p auth cap %p not mds%d ???\n",
2589 &ci->vfs_inode, cap, session->s_mds);
2590 spin_unlock(&ci->i_ceph_lock);
2596 * if flushing caps were revoked, we re-send the cap flush
2597 * in client reconnect stage. This guarantees MDS * processes
2598 * the cap flush message before issuing the flushing caps to
2601 if ((cap->issued & ci->i_flushing_caps) !=
2602 ci->i_flushing_caps) {
2603 /* encode_caps_cb() also will reset these sequence
2604 * numbers. make sure sequence numbers in cap flush
2605 * message match later reconnect message */
2609 __kick_flushing_caps(mdsc, session, ci,
2612 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2615 spin_unlock(&ci->i_ceph_lock);
2619 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2620 struct ceph_mds_session *session)
2622 struct ceph_inode_info *ci;
2623 struct ceph_cap *cap;
2624 u64 oldest_flush_tid;
2626 lockdep_assert_held(&session->s_mutex);
2628 dout("kick_flushing_caps mds%d\n", session->s_mds);
2630 spin_lock(&mdsc->cap_dirty_lock);
2631 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2632 spin_unlock(&mdsc->cap_dirty_lock);
2634 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2635 spin_lock(&ci->i_ceph_lock);
2636 cap = ci->i_auth_cap;
2637 if (!(cap && cap->session == session)) {
2638 pr_err("%p auth cap %p not mds%d ???\n",
2639 &ci->vfs_inode, cap, session->s_mds);
2640 spin_unlock(&ci->i_ceph_lock);
2643 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2644 __kick_flushing_caps(mdsc, session, ci,
2647 spin_unlock(&ci->i_ceph_lock);
2651 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2652 struct ceph_inode_info *ci)
2654 struct ceph_mds_client *mdsc = session->s_mdsc;
2655 struct ceph_cap *cap = ci->i_auth_cap;
2657 lockdep_assert_held(&ci->i_ceph_lock);
2659 dout("%s %p flushing %s\n", __func__, &ci->vfs_inode,
2660 ceph_cap_string(ci->i_flushing_caps));
2662 if (!list_empty(&ci->i_cap_flush_list)) {
2663 u64 oldest_flush_tid;
2664 spin_lock(&mdsc->cap_dirty_lock);
2665 list_move_tail(&ci->i_flushing_item,
2666 &cap->session->s_cap_flushing);
2667 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2668 spin_unlock(&mdsc->cap_dirty_lock);
2670 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2676 * Take references to capabilities we hold, so that we don't release
2677 * them to the MDS prematurely.
2679 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2680 bool snap_rwsem_locked)
2682 lockdep_assert_held(&ci->i_ceph_lock);
2684 if (got & CEPH_CAP_PIN)
2686 if (got & CEPH_CAP_FILE_RD)
2688 if (got & CEPH_CAP_FILE_CACHE)
2689 ci->i_rdcache_ref++;
2690 if (got & CEPH_CAP_FILE_EXCL)
2692 if (got & CEPH_CAP_FILE_WR) {
2693 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2694 BUG_ON(!snap_rwsem_locked);
2695 ci->i_head_snapc = ceph_get_snap_context(
2696 ci->i_snap_realm->cached_context);
2700 if (got & CEPH_CAP_FILE_BUFFER) {
2701 if (ci->i_wb_ref == 0)
2702 ihold(&ci->vfs_inode);
2704 dout("%s %p wb %d -> %d (?)\n", __func__,
2705 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2710 * Try to grab cap references. Specify those refs we @want, and the
2711 * minimal set we @need. Also include the larger offset we are writing
2712 * to (when applicable), and check against max_size here as well.
2713 * Note that caller is responsible for ensuring max_size increases are
2714 * requested from the MDS.
2716 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2717 * or a negative error code. There are 3 speical error codes:
2718 * -EAGAIN: need to sleep but non-blocking is specified
2719 * -EFBIG: ask caller to call check_max_size() and try again.
2720 * -ESTALE: ask caller to call ceph_renew_caps() and try again.
2723 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2724 NON_BLOCKING = (1 << 8),
2725 CHECK_FILELOCK = (1 << 9),
2728 static int try_get_cap_refs(struct inode *inode, int need, int want,
2729 loff_t endoff, int flags, int *got)
2731 struct ceph_inode_info *ci = ceph_inode(inode);
2732 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2734 int have, implemented;
2735 bool snap_rwsem_locked = false;
2737 dout("get_cap_refs %p need %s want %s\n", inode,
2738 ceph_cap_string(need), ceph_cap_string(want));
2741 spin_lock(&ci->i_ceph_lock);
2743 if ((flags & CHECK_FILELOCK) &&
2744 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2745 dout("try_get_cap_refs %p error filelock\n", inode);
2750 /* finish pending truncate */
2751 while (ci->i_truncate_pending) {
2752 spin_unlock(&ci->i_ceph_lock);
2753 if (snap_rwsem_locked) {
2754 up_read(&mdsc->snap_rwsem);
2755 snap_rwsem_locked = false;
2757 __ceph_do_pending_vmtruncate(inode);
2758 spin_lock(&ci->i_ceph_lock);
2761 have = __ceph_caps_issued(ci, &implemented);
2763 if (have & need & CEPH_CAP_FILE_WR) {
2764 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2765 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2766 inode, endoff, ci->i_max_size);
2767 if (endoff > ci->i_requested_max_size)
2768 ret = ci->i_auth_cap ? -EFBIG : -ESTALE;
2772 * If a sync write is in progress, we must wait, so that we
2773 * can get a final snapshot value for size+mtime.
2775 if (__ceph_have_pending_cap_snap(ci)) {
2776 dout("get_cap_refs %p cap_snap_pending\n", inode);
2781 if ((have & need) == need) {
2783 * Look at (implemented & ~have & not) so that we keep waiting
2784 * on transition from wanted -> needed caps. This is needed
2785 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2786 * going before a prior buffered writeback happens.
2788 int not = want & ~(have & need);
2789 int revoking = implemented & ~have;
2790 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2791 inode, ceph_cap_string(have), ceph_cap_string(not),
2792 ceph_cap_string(revoking));
2793 if ((revoking & not) == 0) {
2794 if (!snap_rwsem_locked &&
2795 !ci->i_head_snapc &&
2796 (need & CEPH_CAP_FILE_WR)) {
2797 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2799 * we can not call down_read() when
2800 * task isn't in TASK_RUNNING state
2802 if (flags & NON_BLOCKING) {
2807 spin_unlock(&ci->i_ceph_lock);
2808 down_read(&mdsc->snap_rwsem);
2809 snap_rwsem_locked = true;
2812 snap_rwsem_locked = true;
2814 if ((have & want) == want)
2818 if (S_ISREG(inode->i_mode) &&
2819 (need & CEPH_CAP_FILE_RD) &&
2820 !(*got & CEPH_CAP_FILE_CACHE))
2821 ceph_disable_fscache_readpage(ci);
2822 ceph_take_cap_refs(ci, *got, true);
2826 int session_readonly = false;
2828 if (ci->i_auth_cap &&
2829 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2830 struct ceph_mds_session *s = ci->i_auth_cap->session;
2831 spin_lock(&s->s_cap_lock);
2832 session_readonly = s->s_readonly;
2833 spin_unlock(&s->s_cap_lock);
2835 if (session_readonly) {
2836 dout("get_cap_refs %p need %s but mds%d readonly\n",
2837 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2842 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2843 dout("get_cap_refs %p forced umount\n", inode);
2847 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2848 if (need & ~mds_wanted) {
2849 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2850 inode, ceph_cap_string(need),
2851 ceph_cap_string(mds_wanted));
2856 dout("get_cap_refs %p have %s need %s\n", inode,
2857 ceph_cap_string(have), ceph_cap_string(need));
2861 __ceph_touch_fmode(ci, mdsc, flags);
2863 spin_unlock(&ci->i_ceph_lock);
2864 if (snap_rwsem_locked)
2865 up_read(&mdsc->snap_rwsem);
2868 ceph_update_cap_mis(&mdsc->metric);
2870 ceph_update_cap_hit(&mdsc->metric);
2872 dout("get_cap_refs %p ret %d got %s\n", inode,
2873 ret, ceph_cap_string(*got));
2878 * Check the offset we are writing up to against our current
2879 * max_size. If necessary, tell the MDS we want to write to
2882 static void check_max_size(struct inode *inode, loff_t endoff)
2884 struct ceph_inode_info *ci = ceph_inode(inode);
2887 /* do we need to explicitly request a larger max_size? */
2888 spin_lock(&ci->i_ceph_lock);
2889 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2890 dout("write %p at large endoff %llu, req max_size\n",
2892 ci->i_wanted_max_size = endoff;
2894 /* duplicate ceph_check_caps()'s logic */
2895 if (ci->i_auth_cap &&
2896 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2897 ci->i_wanted_max_size > ci->i_max_size &&
2898 ci->i_wanted_max_size > ci->i_requested_max_size)
2900 spin_unlock(&ci->i_ceph_lock);
2902 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2905 static inline int get_used_fmode(int caps)
2908 if (caps & CEPH_CAP_FILE_RD)
2909 fmode |= CEPH_FILE_MODE_RD;
2910 if (caps & CEPH_CAP_FILE_WR)
2911 fmode |= CEPH_FILE_MODE_WR;
2915 int ceph_try_get_caps(struct inode *inode, int need, int want,
2916 bool nonblock, int *got)
2920 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2921 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2922 CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2923 CEPH_CAP_ANY_DIR_OPS));
2925 ret = ceph_pool_perm_check(inode, need);
2930 flags = get_used_fmode(need | want);
2932 flags |= NON_BLOCKING;
2934 ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2935 /* three special error codes */
2936 if (ret == -EAGAIN || ret == -EFBIG || ret == -ESTALE)
2942 * Wait for caps, and take cap references. If we can't get a WR cap
2943 * due to a small max_size, make sure we check_max_size (and possibly
2944 * ask the mds) so we don't get hung up indefinitely.
2946 int ceph_get_caps(struct file *filp, int need, int want,
2947 loff_t endoff, int *got, struct page **pinned_page)
2949 struct ceph_file_info *fi = filp->private_data;
2950 struct inode *inode = file_inode(filp);
2951 struct ceph_inode_info *ci = ceph_inode(inode);
2952 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2953 int ret, _got, flags;
2955 ret = ceph_pool_perm_check(inode, need);
2959 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2960 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2963 flags = get_used_fmode(need | want);
2966 flags &= CEPH_FILE_MODE_MASK;
2967 if (vfs_inode_has_locks(inode))
2968 flags |= CHECK_FILELOCK;
2970 ret = try_get_cap_refs(inode, need, want, endoff,
2972 WARN_ON_ONCE(ret == -EAGAIN);
2974 struct ceph_mds_client *mdsc = fsc->mdsc;
2976 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2978 cw.ino = ceph_ino(inode);
2979 cw.tgid = current->tgid;
2983 spin_lock(&mdsc->caps_list_lock);
2984 list_add(&cw.list, &mdsc->cap_wait_list);
2985 spin_unlock(&mdsc->caps_list_lock);
2987 /* make sure used fmode not timeout */
2988 ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2989 add_wait_queue(&ci->i_cap_wq, &wait);
2991 flags |= NON_BLOCKING;
2992 while (!(ret = try_get_cap_refs(inode, need, want,
2993 endoff, flags, &_got))) {
2994 if (signal_pending(current)) {
2998 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3001 remove_wait_queue(&ci->i_cap_wq, &wait);
3002 ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
3004 spin_lock(&mdsc->caps_list_lock);
3006 spin_unlock(&mdsc->caps_list_lock);
3012 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
3013 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
3014 if (ret >= 0 && _got)
3015 ceph_put_cap_refs(ci, _got);
3020 if (ret == -EFBIG || ret == -ESTALE) {
3021 int ret2 = ceph_wait_on_async_create(inode);
3025 if (ret == -EFBIG) {
3026 check_max_size(inode, endoff);
3029 if (ret == -ESTALE) {
3030 /* session was killed, try renew caps */
3031 ret = ceph_renew_caps(inode, flags);
3038 if (S_ISREG(ci->vfs_inode.i_mode) &&
3039 ci->i_inline_version != CEPH_INLINE_NONE &&
3040 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
3041 i_size_read(inode) > 0) {
3043 find_get_page(inode->i_mapping, 0);
3045 if (PageUptodate(page)) {
3046 *pinned_page = page;
3052 * drop cap refs first because getattr while
3053 * holding * caps refs can cause deadlock.
3055 ceph_put_cap_refs(ci, _got);
3059 * getattr request will bring inline data into
3062 ret = __ceph_do_getattr(inode, NULL,
3063 CEPH_STAT_CAP_INLINE_DATA,
3072 if (S_ISREG(ci->vfs_inode.i_mode) &&
3073 (_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
3074 ceph_fscache_revalidate_cookie(ci);
3081 * Take cap refs. Caller must already know we hold at least one ref
3082 * on the caps in question or we don't know this is safe.
3084 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
3086 spin_lock(&ci->i_ceph_lock);
3087 ceph_take_cap_refs(ci, caps, false);
3088 spin_unlock(&ci->i_ceph_lock);
3093 * drop cap_snap that is not associated with any snapshot.
3094 * we don't need to send FLUSHSNAP message for it.
3096 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
3097 struct ceph_cap_snap *capsnap)
3099 if (!capsnap->need_flush &&
3100 !capsnap->writing && !capsnap->dirty_pages) {
3101 dout("dropping cap_snap %p follows %llu\n",
3102 capsnap, capsnap->follows);
3103 BUG_ON(capsnap->cap_flush.tid > 0);
3104 ceph_put_snap_context(capsnap->context);
3105 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
3106 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3108 list_del(&capsnap->ci_item);
3109 ceph_put_cap_snap(capsnap);
3118 * If we released the last ref on any given cap, call ceph_check_caps
3119 * to release (or schedule a release).
3121 * If we are releasing a WR cap (from a sync write), finalize any affected
3122 * cap_snap, and wake up any waiters.
3124 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
3125 bool skip_checking_caps)
3127 struct inode *inode = &ci->vfs_inode;
3128 int last = 0, put = 0, flushsnaps = 0, wake = 0;
3130 spin_lock(&ci->i_ceph_lock);
3131 if (had & CEPH_CAP_PIN)
3133 if (had & CEPH_CAP_FILE_RD)
3134 if (--ci->i_rd_ref == 0)
3136 if (had & CEPH_CAP_FILE_CACHE)
3137 if (--ci->i_rdcache_ref == 0)
3139 if (had & CEPH_CAP_FILE_EXCL)
3140 if (--ci->i_fx_ref == 0)
3142 if (had & CEPH_CAP_FILE_BUFFER) {
3143 if (--ci->i_wb_ref == 0) {
3147 dout("put_cap_refs %p wb %d -> %d (?)\n",
3148 inode, ci->i_wb_ref+1, ci->i_wb_ref);
3150 if (had & CEPH_CAP_FILE_WR)
3151 if (--ci->i_wr_ref == 0) {
3153 if (__ceph_have_pending_cap_snap(ci)) {
3154 struct ceph_cap_snap *capsnap =
3155 list_last_entry(&ci->i_cap_snaps,
3156 struct ceph_cap_snap,
3158 capsnap->writing = 0;
3159 if (ceph_try_drop_cap_snap(ci, capsnap))
3161 else if (__ceph_finish_cap_snap(ci, capsnap))
3165 if (ci->i_wrbuffer_ref_head == 0 &&
3166 ci->i_dirty_caps == 0 &&
3167 ci->i_flushing_caps == 0) {
3168 BUG_ON(!ci->i_head_snapc);
3169 ceph_put_snap_context(ci->i_head_snapc);
3170 ci->i_head_snapc = NULL;
3172 /* see comment in __ceph_remove_cap() */
3173 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3174 drop_inode_snap_realm(ci);
3176 spin_unlock(&ci->i_ceph_lock);
3178 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3179 last ? " last" : "", put ? " put" : "");
3181 if (!skip_checking_caps) {
3183 ceph_check_caps(ci, 0, NULL);
3184 else if (flushsnaps)
3185 ceph_flush_snaps(ci, NULL);
3188 wake_up_all(&ci->i_cap_wq);
3193 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3195 __ceph_put_cap_refs(ci, had, false);
3198 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3200 __ceph_put_cap_refs(ci, had, true);
3204 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3205 * context. Adjust per-snap dirty page accounting as appropriate.
3206 * Once all dirty data for a cap_snap is flushed, flush snapped file
3207 * metadata back to the MDS. If we dropped the last ref, call
3210 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3211 struct ceph_snap_context *snapc)
3213 struct inode *inode = &ci->vfs_inode;
3214 struct ceph_cap_snap *capsnap = NULL;
3218 bool flush_snaps = false;
3219 bool complete_capsnap = false;
3221 spin_lock(&ci->i_ceph_lock);
3222 ci->i_wrbuffer_ref -= nr;
3223 if (ci->i_wrbuffer_ref == 0) {
3228 if (ci->i_head_snapc == snapc) {
3229 ci->i_wrbuffer_ref_head -= nr;
3230 if (ci->i_wrbuffer_ref_head == 0 &&
3231 ci->i_wr_ref == 0 &&
3232 ci->i_dirty_caps == 0 &&
3233 ci->i_flushing_caps == 0) {
3234 BUG_ON(!ci->i_head_snapc);
3235 ceph_put_snap_context(ci->i_head_snapc);
3236 ci->i_head_snapc = NULL;
3238 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3240 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3241 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3242 last ? " LAST" : "");
3244 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3245 if (capsnap->context == snapc) {
3253 * The capsnap should already be removed when removing
3254 * auth cap in the case of a forced unmount.
3256 WARN_ON_ONCE(ci->i_auth_cap);
3260 capsnap->dirty_pages -= nr;
3261 if (capsnap->dirty_pages == 0) {
3262 complete_capsnap = true;
3263 if (!capsnap->writing) {
3264 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3267 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3272 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3273 " snap %lld %d/%d -> %d/%d %s%s\n",
3274 inode, capsnap, capsnap->context->seq,
3275 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3276 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3277 last ? " (wrbuffer last)" : "",
3278 complete_capsnap ? " (complete capsnap)" : "");
3282 spin_unlock(&ci->i_ceph_lock);
3285 ceph_check_caps(ci, 0, NULL);
3286 } else if (flush_snaps) {
3287 ceph_flush_snaps(ci, NULL);
3289 if (complete_capsnap)
3290 wake_up_all(&ci->i_cap_wq);
3292 /* avoid calling iput_final() in osd dispatch threads */
3293 ceph_async_iput(inode);
3298 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3300 static void invalidate_aliases(struct inode *inode)
3302 struct dentry *dn, *prev = NULL;
3304 dout("invalidate_aliases inode %p\n", inode);
3305 d_prune_aliases(inode);
3307 * For non-directory inode, d_find_alias() only returns
3308 * hashed dentry. After calling d_invalidate(), the
3309 * dentry becomes unhashed.
3311 * For directory inode, d_find_alias() can return
3312 * unhashed dentry. But directory inode should have
3313 * one alias at most.
3315 while ((dn = d_find_alias(inode))) {
3329 struct cap_extra_info {
3330 struct ceph_string *pool_ns;
3340 /* currently issued */
3342 struct timespec64 btime;
3346 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3347 * actually be a revocation if it specifies a smaller cap set.)
3349 * caller holds s_mutex and i_ceph_lock, we drop both.
3351 static void handle_cap_grant(struct inode *inode,
3352 struct ceph_mds_session *session,
3353 struct ceph_cap *cap,
3354 struct ceph_mds_caps *grant,
3355 struct ceph_buffer *xattr_buf,
3356 struct cap_extra_info *extra_info)
3357 __releases(ci->i_ceph_lock)
3358 __releases(session->s_mdsc->snap_rwsem)
3360 struct ceph_inode_info *ci = ceph_inode(inode);
3361 int seq = le32_to_cpu(grant->seq);
3362 int newcaps = le32_to_cpu(grant->caps);
3363 int used, wanted, dirty;
3364 u64 size = le64_to_cpu(grant->size);
3365 u64 max_size = le64_to_cpu(grant->max_size);
3366 unsigned char check_caps = 0;
3367 bool was_stale = cap->cap_gen < session->s_cap_gen;
3369 bool writeback = false;
3370 bool queue_trunc = false;
3371 bool queue_invalidate = false;
3372 bool deleted_inode = false;
3373 bool fill_inline = false;
3375 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3376 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3377 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3382 * If CACHE is being revoked, and we have no dirty buffers,
3383 * try to invalidate (once). (If there are dirty buffers, we
3384 * will invalidate _after_ writeback.)
3386 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3387 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3388 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3389 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3390 if (try_nonblocking_invalidate(inode)) {
3391 /* there were locked pages.. invalidate later
3392 in a separate thread. */
3393 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3394 queue_invalidate = true;
3395 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3401 cap->issued = cap->implemented = CEPH_CAP_PIN;
3404 * auth mds of the inode changed. we received the cap export message,
3405 * but still haven't received the cap import message. handle_cap_export
3406 * updated the new auth MDS' cap.
3408 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3409 * that was sent before the cap import message. So don't remove caps.
3411 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3412 WARN_ON(cap != ci->i_auth_cap);
3413 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3415 newcaps |= cap->issued;
3418 /* side effects now are allowed */
3419 cap->cap_gen = session->s_cap_gen;
3422 __check_cap_issue(ci, cap, newcaps);
3424 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3426 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3427 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3428 inode->i_mode = le32_to_cpu(grant->mode);
3429 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3430 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3431 ci->i_btime = extra_info->btime;
3432 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3433 from_kuid(&init_user_ns, inode->i_uid),
3434 from_kgid(&init_user_ns, inode->i_gid));
3437 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3438 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3439 set_nlink(inode, le32_to_cpu(grant->nlink));
3440 if (inode->i_nlink == 0 &&
3441 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3442 deleted_inode = true;
3445 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3447 int len = le32_to_cpu(grant->xattr_len);
3448 u64 version = le64_to_cpu(grant->xattr_version);
3450 if (version > ci->i_xattrs.version) {
3451 dout(" got new xattrs v%llu on %p len %d\n",
3452 version, inode, len);
3453 if (ci->i_xattrs.blob)
3454 ceph_buffer_put(ci->i_xattrs.blob);
3455 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3456 ci->i_xattrs.version = version;
3457 ceph_forget_all_cached_acls(inode);
3458 ceph_security_invalidate_secctx(inode);
3462 if (newcaps & CEPH_CAP_ANY_RD) {
3463 struct timespec64 mtime, atime, ctime;
3464 /* ctime/mtime/atime? */
3465 ceph_decode_timespec64(&mtime, &grant->mtime);
3466 ceph_decode_timespec64(&atime, &grant->atime);
3467 ceph_decode_timespec64(&ctime, &grant->ctime);
3468 ceph_fill_file_time(inode, extra_info->issued,
3469 le32_to_cpu(grant->time_warp_seq),
3470 &ctime, &mtime, &atime);
3473 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3474 ci->i_files = extra_info->nfiles;
3475 ci->i_subdirs = extra_info->nsubdirs;
3478 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3479 /* file layout may have changed */
3480 s64 old_pool = ci->i_layout.pool_id;
3481 struct ceph_string *old_ns;
3483 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3484 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3485 lockdep_is_held(&ci->i_ceph_lock));
3486 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3488 if (ci->i_layout.pool_id != old_pool ||
3489 extra_info->pool_ns != old_ns)
3490 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3492 extra_info->pool_ns = old_ns;
3494 /* size/truncate_seq? */
3495 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3496 le32_to_cpu(grant->truncate_seq),
3497 le64_to_cpu(grant->truncate_size),
3501 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3502 if (max_size != ci->i_max_size) {
3503 dout("max_size %lld -> %llu\n",
3504 ci->i_max_size, max_size);
3505 ci->i_max_size = max_size;
3506 if (max_size >= ci->i_wanted_max_size) {
3507 ci->i_wanted_max_size = 0; /* reset */
3508 ci->i_requested_max_size = 0;
3514 /* check cap bits */
3515 wanted = __ceph_caps_wanted(ci);
3516 used = __ceph_caps_used(ci);
3517 dirty = __ceph_caps_dirty(ci);
3518 dout(" my wanted = %s, used = %s, dirty %s\n",
3519 ceph_cap_string(wanted),
3520 ceph_cap_string(used),
3521 ceph_cap_string(dirty));
3523 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3524 (wanted & ~(cap->mds_wanted | newcaps))) {
3526 * If mds is importing cap, prior cap messages that update
3527 * 'wanted' may get dropped by mds (migrate seq mismatch).
3529 * We don't send cap message to update 'wanted' if what we
3530 * want are already issued. If mds revokes caps, cap message
3531 * that releases caps also tells mds what we want. But if
3532 * caps got revoked by mds forcedly (session stale). We may
3533 * haven't told mds what we want.
3538 /* revocation, grant, or no-op? */
3539 if (cap->issued & ~newcaps) {
3540 int revoking = cap->issued & ~newcaps;
3542 dout("revocation: %s -> %s (revoking %s)\n",
3543 ceph_cap_string(cap->issued),
3544 ceph_cap_string(newcaps),
3545 ceph_cap_string(revoking));
3546 if (S_ISREG(inode->i_mode) &&
3547 (revoking & used & CEPH_CAP_FILE_BUFFER))
3548 writeback = true; /* initiate writeback; will delay ack */
3549 else if (queue_invalidate &&
3550 revoking == CEPH_CAP_FILE_CACHE &&
3551 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3552 ; /* do nothing yet, invalidation will be queued */
3553 else if (cap == ci->i_auth_cap)
3554 check_caps = 1; /* check auth cap only */
3556 check_caps = 2; /* check all caps */
3557 cap->issued = newcaps;
3558 cap->implemented |= newcaps;
3559 } else if (cap->issued == newcaps) {
3560 dout("caps unchanged: %s -> %s\n",
3561 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3563 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3564 ceph_cap_string(newcaps));
3565 /* non-auth MDS is revoking the newly grant caps ? */
3566 if (cap == ci->i_auth_cap &&
3567 __ceph_caps_revoking_other(ci, cap, newcaps))
3570 cap->issued = newcaps;
3571 cap->implemented |= newcaps; /* add bits only, to
3572 * avoid stepping on a
3573 * pending revocation */
3576 BUG_ON(cap->issued & ~cap->implemented);
3578 /* don't let check_caps skip sending a response to MDS for revoke msgs */
3579 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_REVOKE) {
3580 cap->mds_wanted = 0;
3581 if (cap == ci->i_auth_cap)
3582 check_caps = 1; /* check auth cap only */
3584 check_caps = 2; /* check all caps */
3587 if (extra_info->inline_version > 0 &&
3588 extra_info->inline_version >= ci->i_inline_version) {
3589 ci->i_inline_version = extra_info->inline_version;
3590 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3591 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3595 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3596 if (ci->i_auth_cap == cap) {
3597 if (newcaps & ~extra_info->issued)
3600 if (ci->i_requested_max_size > max_size ||
3601 !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3602 /* re-request max_size if necessary */
3603 ci->i_requested_max_size = 0;
3607 ceph_kick_flushing_inode_caps(session, ci);
3609 up_read(&session->s_mdsc->snap_rwsem);
3611 spin_unlock(&ci->i_ceph_lock);
3614 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3615 extra_info->inline_len);
3618 ceph_queue_vmtruncate(inode);
3622 * queue inode for writeback: we can't actually call
3623 * filemap_write_and_wait, etc. from message handler
3626 ceph_queue_writeback(inode);
3627 if (queue_invalidate)
3628 ceph_queue_invalidate(inode);
3630 invalidate_aliases(inode);
3632 wake_up_all(&ci->i_cap_wq);
3634 if (check_caps == 1)
3635 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3637 else if (check_caps == 2)
3638 ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3640 mutex_unlock(&session->s_mutex);
3644 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3645 * MDS has been safely committed.
3647 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3648 struct ceph_mds_caps *m,
3649 struct ceph_mds_session *session,
3650 struct ceph_cap *cap)
3651 __releases(ci->i_ceph_lock)
3653 struct ceph_inode_info *ci = ceph_inode(inode);
3654 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3655 struct ceph_cap_flush *cf, *tmp_cf;
3656 LIST_HEAD(to_remove);
3657 unsigned seq = le32_to_cpu(m->seq);
3658 int dirty = le32_to_cpu(m->dirty);
3661 bool wake_ci = false;
3662 bool wake_mdsc = false;
3664 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3665 /* Is this the one that was flushed? */
3666 if (cf->tid == flush_tid)
3669 /* Is this a capsnap? */
3673 if (cf->tid <= flush_tid) {
3675 * An earlier or current tid. The FLUSH_ACK should
3676 * represent a superset of this flush's caps.
3678 wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3679 list_add_tail(&cf->i_list, &to_remove);
3682 * This is a later one. Any caps in it are still dirty
3683 * so don't count them as cleaned.
3685 cleaned &= ~cf->caps;
3691 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3692 " flushing %s -> %s\n",
3693 inode, session->s_mds, seq, ceph_cap_string(dirty),
3694 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3695 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3697 if (list_empty(&to_remove) && !cleaned)
3700 ci->i_flushing_caps &= ~cleaned;
3702 spin_lock(&mdsc->cap_dirty_lock);
3704 list_for_each_entry(cf, &to_remove, i_list)
3705 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3707 if (ci->i_flushing_caps == 0) {
3708 if (list_empty(&ci->i_cap_flush_list)) {
3709 list_del_init(&ci->i_flushing_item);
3710 if (!list_empty(&session->s_cap_flushing)) {
3711 dout(" mds%d still flushing cap on %p\n",
3713 &list_first_entry(&session->s_cap_flushing,
3714 struct ceph_inode_info,
3715 i_flushing_item)->vfs_inode);
3718 mdsc->num_cap_flushing--;
3719 dout(" inode %p now !flushing\n", inode);
3721 if (ci->i_dirty_caps == 0) {
3722 dout(" inode %p now clean\n", inode);
3723 BUG_ON(!list_empty(&ci->i_dirty_item));
3725 if (ci->i_wr_ref == 0 &&
3726 ci->i_wrbuffer_ref_head == 0) {
3727 BUG_ON(!ci->i_head_snapc);
3728 ceph_put_snap_context(ci->i_head_snapc);
3729 ci->i_head_snapc = NULL;
3732 BUG_ON(list_empty(&ci->i_dirty_item));
3735 spin_unlock(&mdsc->cap_dirty_lock);
3738 spin_unlock(&ci->i_ceph_lock);
3740 while (!list_empty(&to_remove)) {
3741 cf = list_first_entry(&to_remove,
3742 struct ceph_cap_flush, i_list);
3743 list_del_init(&cf->i_list);
3744 if (!cf->is_capsnap)
3745 ceph_free_cap_flush(cf);
3749 wake_up_all(&ci->i_cap_wq);
3751 wake_up_all(&mdsc->cap_flushing_wq);
3756 void __ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3757 bool *wake_ci, bool *wake_mdsc)
3759 struct ceph_inode_info *ci = ceph_inode(inode);
3760 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3763 lockdep_assert_held(&ci->i_ceph_lock);
3765 dout("removing capsnap %p, inode %p ci %p\n", capsnap, inode, ci);
3767 list_del_init(&capsnap->ci_item);
3768 ret = __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3772 spin_lock(&mdsc->cap_dirty_lock);
3773 if (list_empty(&ci->i_cap_flush_list))
3774 list_del_init(&ci->i_flushing_item);
3776 ret = __detach_cap_flush_from_mdsc(mdsc, &capsnap->cap_flush);
3779 spin_unlock(&mdsc->cap_dirty_lock);
3782 void ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3783 bool *wake_ci, bool *wake_mdsc)
3785 struct ceph_inode_info *ci = ceph_inode(inode);
3787 lockdep_assert_held(&ci->i_ceph_lock);
3789 WARN_ON_ONCE(capsnap->dirty_pages || capsnap->writing);
3790 __ceph_remove_capsnap(inode, capsnap, wake_ci, wake_mdsc);
3794 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3795 * throw away our cap_snap.
3797 * Caller hold s_mutex.
3799 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3800 struct ceph_mds_caps *m,
3801 struct ceph_mds_session *session)
3803 struct ceph_inode_info *ci = ceph_inode(inode);
3804 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3805 u64 follows = le64_to_cpu(m->snap_follows);
3806 struct ceph_cap_snap *capsnap;
3807 bool flushed = false;
3808 bool wake_ci = false;
3809 bool wake_mdsc = false;
3811 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3812 inode, ci, session->s_mds, follows);
3814 spin_lock(&ci->i_ceph_lock);
3815 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3816 if (capsnap->follows == follows) {
3817 if (capsnap->cap_flush.tid != flush_tid) {
3818 dout(" cap_snap %p follows %lld tid %lld !="
3819 " %lld\n", capsnap, follows,
3820 flush_tid, capsnap->cap_flush.tid);
3826 dout(" skipping cap_snap %p follows %lld\n",
3827 capsnap, capsnap->follows);
3831 ceph_remove_capsnap(inode, capsnap, &wake_ci, &wake_mdsc);
3832 spin_unlock(&ci->i_ceph_lock);
3835 ceph_put_snap_context(capsnap->context);
3836 ceph_put_cap_snap(capsnap);
3838 wake_up_all(&ci->i_cap_wq);
3840 wake_up_all(&mdsc->cap_flushing_wq);
3846 * Handle TRUNC from MDS, indicating file truncation.
3848 * caller hold s_mutex.
3850 static bool handle_cap_trunc(struct inode *inode,
3851 struct ceph_mds_caps *trunc,
3852 struct ceph_mds_session *session)
3854 struct ceph_inode_info *ci = ceph_inode(inode);
3855 int mds = session->s_mds;
3856 int seq = le32_to_cpu(trunc->seq);
3857 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3858 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3859 u64 size = le64_to_cpu(trunc->size);
3860 int implemented = 0;
3861 int dirty = __ceph_caps_dirty(ci);
3862 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3863 bool queue_trunc = false;
3865 lockdep_assert_held(&ci->i_ceph_lock);
3867 issued |= implemented | dirty;
3869 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3870 inode, mds, seq, truncate_size, truncate_seq);
3871 queue_trunc = ceph_fill_file_size(inode, issued,
3872 truncate_seq, truncate_size, size);
3877 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3878 * different one. If we are the most recent migration we've seen (as
3879 * indicated by mseq), make note of the migrating cap bits for the
3880 * duration (until we see the corresponding IMPORT).
3882 * caller holds s_mutex
3884 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3885 struct ceph_mds_cap_peer *ph,
3886 struct ceph_mds_session *session)
3888 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3889 struct ceph_mds_session *tsession = NULL;
3890 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3891 struct ceph_inode_info *ci = ceph_inode(inode);
3893 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3894 unsigned t_seq, t_mseq;
3896 int mds = session->s_mds;
3899 t_cap_id = le64_to_cpu(ph->cap_id);
3900 t_seq = le32_to_cpu(ph->seq);
3901 t_mseq = le32_to_cpu(ph->mseq);
3902 target = le32_to_cpu(ph->mds);
3904 t_cap_id = t_seq = t_mseq = 0;
3908 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3909 inode, ci, mds, mseq, target);
3911 spin_lock(&ci->i_ceph_lock);
3912 cap = __get_cap_for_mds(ci, mds);
3913 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3917 __ceph_remove_cap(cap, false);
3922 * now we know we haven't received the cap import message yet
3923 * because the exported cap still exist.
3926 issued = cap->issued;
3927 if (issued != cap->implemented)
3928 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3929 "ino (%llx.%llx) mds%d seq %d mseq %d "
3930 "issued %s implemented %s\n",
3931 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3932 ceph_cap_string(issued),
3933 ceph_cap_string(cap->implemented));
3936 tcap = __get_cap_for_mds(ci, target);
3938 /* already have caps from the target */
3939 if (tcap->cap_id == t_cap_id &&
3940 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3941 dout(" updating import cap %p mds%d\n", tcap, target);
3942 tcap->cap_id = t_cap_id;
3943 tcap->seq = t_seq - 1;
3944 tcap->issue_seq = t_seq - 1;
3945 tcap->issued |= issued;
3946 tcap->implemented |= issued;
3947 if (cap == ci->i_auth_cap) {
3948 ci->i_auth_cap = tcap;
3949 change_auth_cap_ses(ci, tcap->session);
3952 __ceph_remove_cap(cap, false);
3954 } else if (tsession) {
3955 /* add placeholder for the export tagert */
3956 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3958 ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3959 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3961 if (!list_empty(&ci->i_cap_flush_list) &&
3962 ci->i_auth_cap == tcap) {
3963 spin_lock(&mdsc->cap_dirty_lock);
3964 list_move_tail(&ci->i_flushing_item,
3965 &tcap->session->s_cap_flushing);
3966 spin_unlock(&mdsc->cap_dirty_lock);
3969 __ceph_remove_cap(cap, false);
3973 spin_unlock(&ci->i_ceph_lock);
3974 mutex_unlock(&session->s_mutex);
3976 /* open target session */
3977 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3978 if (!IS_ERR(tsession)) {
3980 mutex_lock(&session->s_mutex);
3981 mutex_lock_nested(&tsession->s_mutex,
3982 SINGLE_DEPTH_NESTING);
3984 mutex_lock(&tsession->s_mutex);
3985 mutex_lock_nested(&session->s_mutex,
3986 SINGLE_DEPTH_NESTING);
3988 new_cap = ceph_get_cap(mdsc, NULL);
3993 mutex_lock(&session->s_mutex);
3998 spin_unlock(&ci->i_ceph_lock);
3999 mutex_unlock(&session->s_mutex);
4001 mutex_unlock(&tsession->s_mutex);
4002 ceph_put_mds_session(tsession);
4005 ceph_put_cap(mdsc, new_cap);
4009 * Handle cap IMPORT.
4011 * caller holds s_mutex. acquires i_ceph_lock
4013 static void handle_cap_import(struct ceph_mds_client *mdsc,
4014 struct inode *inode, struct ceph_mds_caps *im,
4015 struct ceph_mds_cap_peer *ph,
4016 struct ceph_mds_session *session,
4017 struct ceph_cap **target_cap, int *old_issued)
4019 struct ceph_inode_info *ci = ceph_inode(inode);
4020 struct ceph_cap *cap, *ocap, *new_cap = NULL;
4021 int mds = session->s_mds;
4023 unsigned caps = le32_to_cpu(im->caps);
4024 unsigned wanted = le32_to_cpu(im->wanted);
4025 unsigned seq = le32_to_cpu(im->seq);
4026 unsigned mseq = le32_to_cpu(im->migrate_seq);
4027 u64 realmino = le64_to_cpu(im->realm);
4028 u64 cap_id = le64_to_cpu(im->cap_id);
4033 p_cap_id = le64_to_cpu(ph->cap_id);
4034 peer = le32_to_cpu(ph->mds);
4040 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
4041 inode, ci, mds, mseq, peer);
4043 cap = __get_cap_for_mds(ci, mds);
4046 spin_unlock(&ci->i_ceph_lock);
4047 new_cap = ceph_get_cap(mdsc, NULL);
4048 spin_lock(&ci->i_ceph_lock);
4054 ceph_put_cap(mdsc, new_cap);
4059 __ceph_caps_issued(ci, &issued);
4060 issued |= __ceph_caps_dirty(ci);
4062 ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
4063 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
4065 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
4066 if (ocap && ocap->cap_id == p_cap_id) {
4067 dout(" remove export cap %p mds%d flags %d\n",
4068 ocap, peer, ph->flags);
4069 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
4070 (ocap->seq != le32_to_cpu(ph->seq) ||
4071 ocap->mseq != le32_to_cpu(ph->mseq))) {
4072 pr_err_ratelimited("handle_cap_import: "
4073 "mismatched seq/mseq: ino (%llx.%llx) "
4074 "mds%d seq %d mseq %d importer mds%d "
4075 "has peer seq %d mseq %d\n",
4076 ceph_vinop(inode), peer, ocap->seq,
4077 ocap->mseq, mds, le32_to_cpu(ph->seq),
4078 le32_to_cpu(ph->mseq));
4080 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
4083 *old_issued = issued;
4088 * Handle a caps message from the MDS.
4090 * Identify the appropriate session, inode, and call the right handler
4091 * based on the cap op.
4093 void ceph_handle_caps(struct ceph_mds_session *session,
4094 struct ceph_msg *msg)
4096 struct ceph_mds_client *mdsc = session->s_mdsc;
4097 struct inode *inode;
4098 struct ceph_inode_info *ci;
4099 struct ceph_cap *cap;
4100 struct ceph_mds_caps *h;
4101 struct ceph_mds_cap_peer *peer = NULL;
4102 struct ceph_snap_realm *realm = NULL;
4104 int msg_version = le16_to_cpu(msg->hdr.version);
4106 struct ceph_vino vino;
4108 size_t snaptrace_len;
4110 struct cap_extra_info extra_info = {};
4113 dout("handle_caps from mds%d\n", session->s_mds);
4116 end = msg->front.iov_base + msg->front.iov_len;
4117 if (msg->front.iov_len < sizeof(*h))
4119 h = msg->front.iov_base;
4120 op = le32_to_cpu(h->op);
4121 vino.ino = le64_to_cpu(h->ino);
4122 vino.snap = CEPH_NOSNAP;
4123 seq = le32_to_cpu(h->seq);
4124 mseq = le32_to_cpu(h->migrate_seq);
4127 snaptrace_len = le32_to_cpu(h->snap_trace_len);
4128 p = snaptrace + snaptrace_len;
4130 if (msg_version >= 2) {
4132 ceph_decode_32_safe(&p, end, flock_len, bad);
4133 if (p + flock_len > end)
4138 if (msg_version >= 3) {
4139 if (op == CEPH_CAP_OP_IMPORT) {
4140 if (p + sizeof(*peer) > end)
4144 } else if (op == CEPH_CAP_OP_EXPORT) {
4145 /* recorded in unused fields */
4146 peer = (void *)&h->size;
4150 if (msg_version >= 4) {
4151 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
4152 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
4153 if (p + extra_info.inline_len > end)
4155 extra_info.inline_data = p;
4156 p += extra_info.inline_len;
4159 if (msg_version >= 5) {
4160 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
4163 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
4164 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
4167 if (msg_version >= 8) {
4169 u32 caller_uid, caller_gid;
4173 ceph_decode_64_safe(&p, end, flush_tid, bad);
4175 ceph_decode_32_safe(&p, end, caller_uid, bad);
4176 ceph_decode_32_safe(&p, end, caller_gid, bad);
4178 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4179 if (pool_ns_len > 0) {
4180 ceph_decode_need(&p, end, pool_ns_len, bad);
4181 extra_info.pool_ns =
4182 ceph_find_or_create_string(p, pool_ns_len);
4187 if (msg_version >= 9) {
4188 struct ceph_timespec *btime;
4190 if (p + sizeof(*btime) > end)
4193 ceph_decode_timespec64(&extra_info.btime, btime);
4194 p += sizeof(*btime);
4195 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4198 if (msg_version >= 11) {
4201 ceph_decode_32_safe(&p, end, flags, bad);
4203 extra_info.dirstat_valid = true;
4204 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4205 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4209 inode = ceph_find_inode(mdsc->fsc->sb, vino);
4210 ci = ceph_inode(inode);
4211 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4214 mutex_lock(&session->s_mutex);
4215 inc_session_sequence(session);
4216 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4220 dout(" i don't have ino %llx\n", vino.ino);
4222 if (op == CEPH_CAP_OP_IMPORT) {
4223 cap = ceph_get_cap(mdsc, NULL);
4224 cap->cap_ino = vino.ino;
4225 cap->queue_release = 1;
4226 cap->cap_id = le64_to_cpu(h->cap_id);
4229 cap->issue_seq = seq;
4230 spin_lock(&session->s_cap_lock);
4231 __ceph_queue_cap_release(session, cap);
4232 spin_unlock(&session->s_cap_lock);
4234 goto flush_cap_releases;
4237 /* these will work even if we don't have a cap yet */
4239 case CEPH_CAP_OP_FLUSHSNAP_ACK:
4240 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4244 case CEPH_CAP_OP_EXPORT:
4245 handle_cap_export(inode, h, peer, session);
4248 case CEPH_CAP_OP_IMPORT:
4250 if (snaptrace_len) {
4251 down_write(&mdsc->snap_rwsem);
4252 ceph_update_snap_trace(mdsc, snaptrace,
4253 snaptrace + snaptrace_len,
4255 downgrade_write(&mdsc->snap_rwsem);
4257 down_read(&mdsc->snap_rwsem);
4259 spin_lock(&ci->i_ceph_lock);
4260 handle_cap_import(mdsc, inode, h, peer, session,
4261 &cap, &extra_info.issued);
4262 handle_cap_grant(inode, session, cap,
4263 h, msg->middle, &extra_info);
4265 ceph_put_snap_realm(mdsc, realm);
4269 /* the rest require a cap */
4270 spin_lock(&ci->i_ceph_lock);
4271 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4273 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4274 inode, ceph_ino(inode), ceph_snap(inode),
4276 spin_unlock(&ci->i_ceph_lock);
4277 goto flush_cap_releases;
4280 /* note that each of these drops i_ceph_lock for us */
4282 case CEPH_CAP_OP_REVOKE:
4283 case CEPH_CAP_OP_GRANT:
4284 __ceph_caps_issued(ci, &extra_info.issued);
4285 extra_info.issued |= __ceph_caps_dirty(ci);
4286 handle_cap_grant(inode, session, cap,
4287 h, msg->middle, &extra_info);
4290 case CEPH_CAP_OP_FLUSH_ACK:
4291 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4295 case CEPH_CAP_OP_TRUNC:
4296 queue_trunc = handle_cap_trunc(inode, h, session);
4297 spin_unlock(&ci->i_ceph_lock);
4299 ceph_queue_vmtruncate(inode);
4303 spin_unlock(&ci->i_ceph_lock);
4304 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4305 ceph_cap_op_name(op));
4309 mutex_unlock(&session->s_mutex);
4311 ceph_put_string(extra_info.pool_ns);
4312 /* avoid calling iput_final() in mds dispatch threads */
4313 ceph_async_iput(inode);
4318 * send any cap release message to try to move things
4319 * along for the mds (who clearly thinks we still have this
4322 ceph_flush_cap_releases(mdsc, session);
4326 pr_err("ceph_handle_caps: corrupt message\n");
4332 * Delayed work handler to process end of delayed cap release LRU list.
4334 * If new caps are added to the list while processing it, these won't get
4335 * processed in this run. In this case, the ci->i_hold_caps_max will be
4336 * returned so that the work can be scheduled accordingly.
4338 unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4340 struct inode *inode;
4341 struct ceph_inode_info *ci;
4342 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
4343 unsigned long delay_max = opt->caps_wanted_delay_max * HZ;
4344 unsigned long loop_start = jiffies;
4345 unsigned long delay = 0;
4347 dout("check_delayed_caps\n");
4348 spin_lock(&mdsc->cap_delay_lock);
4349 while (!list_empty(&mdsc->cap_delay_list)) {
4350 ci = list_first_entry(&mdsc->cap_delay_list,
4351 struct ceph_inode_info,
4353 if (time_before(loop_start, ci->i_hold_caps_max - delay_max)) {
4354 dout("%s caps added recently. Exiting loop", __func__);
4355 delay = ci->i_hold_caps_max;
4358 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4359 time_before(jiffies, ci->i_hold_caps_max))
4361 list_del_init(&ci->i_cap_delay_list);
4363 inode = igrab(&ci->vfs_inode);
4365 spin_unlock(&mdsc->cap_delay_lock);
4366 dout("check_delayed_caps on %p\n", inode);
4367 ceph_check_caps(ci, 0, NULL);
4368 /* avoid calling iput_final() in tick thread */
4369 ceph_async_iput(inode);
4370 spin_lock(&mdsc->cap_delay_lock);
4373 spin_unlock(&mdsc->cap_delay_lock);
4379 * Flush all dirty caps to the mds
4381 static void flush_dirty_session_caps(struct ceph_mds_session *s)
4383 struct ceph_mds_client *mdsc = s->s_mdsc;
4384 struct ceph_inode_info *ci;
4385 struct inode *inode;
4387 dout("flush_dirty_caps\n");
4388 spin_lock(&mdsc->cap_dirty_lock);
4389 while (!list_empty(&s->s_cap_dirty)) {
4390 ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4392 inode = &ci->vfs_inode;
4394 dout("flush_dirty_caps %p\n", inode);
4395 spin_unlock(&mdsc->cap_dirty_lock);
4396 ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4398 spin_lock(&mdsc->cap_dirty_lock);
4400 spin_unlock(&mdsc->cap_dirty_lock);
4401 dout("flush_dirty_caps done\n");
4404 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4406 ceph_mdsc_iterate_sessions(mdsc, flush_dirty_session_caps, true);
4409 void __ceph_touch_fmode(struct ceph_inode_info *ci,
4410 struct ceph_mds_client *mdsc, int fmode)
4412 unsigned long now = jiffies;
4413 if (fmode & CEPH_FILE_MODE_RD)
4414 ci->i_last_rd = now;
4415 if (fmode & CEPH_FILE_MODE_WR)
4416 ci->i_last_wr = now;
4417 /* queue periodic check */
4419 __ceph_is_any_real_caps(ci) &&
4420 list_empty(&ci->i_cap_delay_list))
4421 __cap_delay_requeue(mdsc, ci);
4424 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4426 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4427 int bits = (fmode << 1) | 1;
4428 bool already_opened = false;
4432 atomic64_inc(&mdsc->metric.opened_files);
4434 spin_lock(&ci->i_ceph_lock);
4435 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4437 * If any of the mode ref is larger than 0,
4438 * that means it has been already opened by
4439 * others. Just skip checking the PIN ref.
4441 if (i && ci->i_nr_by_mode[i])
4442 already_opened = true;
4444 if (bits & (1 << i))
4445 ci->i_nr_by_mode[i] += count;
4448 if (!already_opened)
4449 percpu_counter_inc(&mdsc->metric.opened_inodes);
4450 spin_unlock(&ci->i_ceph_lock);
4454 * Drop open file reference. If we were the last open file,
4455 * we may need to release capabilities to the MDS (or schedule
4456 * their delayed release).
4458 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4460 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4461 int bits = (fmode << 1) | 1;
4462 bool is_closed = true;
4466 atomic64_dec(&mdsc->metric.opened_files);
4468 spin_lock(&ci->i_ceph_lock);
4469 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4470 if (bits & (1 << i)) {
4471 BUG_ON(ci->i_nr_by_mode[i] < count);
4472 ci->i_nr_by_mode[i] -= count;
4476 * If any of the mode ref is not 0 after
4477 * decreased, that means it is still opened
4478 * by others. Just skip checking the PIN ref.
4480 if (i && ci->i_nr_by_mode[i])
4485 percpu_counter_dec(&mdsc->metric.opened_inodes);
4486 spin_unlock(&ci->i_ceph_lock);
4490 * For a soon-to-be unlinked file, drop the LINK caps. If it
4491 * looks like the link count will hit 0, drop any other caps (other
4492 * than PIN) we don't specifically want (due to the file still being
4495 int ceph_drop_caps_for_unlink(struct inode *inode)
4497 struct ceph_inode_info *ci = ceph_inode(inode);
4498 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4500 spin_lock(&ci->i_ceph_lock);
4501 if (inode->i_nlink == 1) {
4502 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4504 if (__ceph_caps_dirty(ci)) {
4505 struct ceph_mds_client *mdsc =
4506 ceph_inode_to_client(inode)->mdsc;
4507 __cap_delay_requeue_front(mdsc, ci);
4510 spin_unlock(&ci->i_ceph_lock);
4515 * Helpers for embedding cap and dentry lease releases into mds
4518 * @force is used by dentry_release (below) to force inclusion of a
4519 * record for the directory inode, even when there aren't any caps to
4522 int ceph_encode_inode_release(void **p, struct inode *inode,
4523 int mds, int drop, int unless, int force)
4525 struct ceph_inode_info *ci = ceph_inode(inode);
4526 struct ceph_cap *cap;
4527 struct ceph_mds_request_release *rel = *p;
4531 spin_lock(&ci->i_ceph_lock);
4532 used = __ceph_caps_used(ci);
4533 dirty = __ceph_caps_dirty(ci);
4535 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4536 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4537 ceph_cap_string(unless));
4539 /* only drop unused, clean caps */
4540 drop &= ~(used | dirty);
4542 cap = __get_cap_for_mds(ci, mds);
4543 if (cap && __cap_is_valid(cap)) {
4544 unless &= cap->issued;
4546 if (unless & CEPH_CAP_AUTH_EXCL)
4547 drop &= ~CEPH_CAP_AUTH_SHARED;
4548 if (unless & CEPH_CAP_LINK_EXCL)
4549 drop &= ~CEPH_CAP_LINK_SHARED;
4550 if (unless & CEPH_CAP_XATTR_EXCL)
4551 drop &= ~CEPH_CAP_XATTR_SHARED;
4552 if (unless & CEPH_CAP_FILE_EXCL)
4553 drop &= ~CEPH_CAP_FILE_SHARED;
4556 if (force || (cap->issued & drop)) {
4557 if (cap->issued & drop) {
4558 int wanted = __ceph_caps_wanted(ci);
4559 dout("encode_inode_release %p cap %p "
4560 "%s -> %s, wanted %s -> %s\n", inode, cap,
4561 ceph_cap_string(cap->issued),
4562 ceph_cap_string(cap->issued & ~drop),
4563 ceph_cap_string(cap->mds_wanted),
4564 ceph_cap_string(wanted));
4566 cap->issued &= ~drop;
4567 cap->implemented &= ~drop;
4568 cap->mds_wanted = wanted;
4569 if (cap == ci->i_auth_cap &&
4570 !(wanted & CEPH_CAP_ANY_FILE_WR))
4571 ci->i_requested_max_size = 0;
4573 dout("encode_inode_release %p cap %p %s"
4574 " (force)\n", inode, cap,
4575 ceph_cap_string(cap->issued));
4578 rel->ino = cpu_to_le64(ceph_ino(inode));
4579 rel->cap_id = cpu_to_le64(cap->cap_id);
4580 rel->seq = cpu_to_le32(cap->seq);
4581 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4582 rel->mseq = cpu_to_le32(cap->mseq);
4583 rel->caps = cpu_to_le32(cap->implemented);
4584 rel->wanted = cpu_to_le32(cap->mds_wanted);
4590 dout("encode_inode_release %p cap %p %s (noop)\n",
4591 inode, cap, ceph_cap_string(cap->issued));
4594 spin_unlock(&ci->i_ceph_lock);
4598 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4600 int mds, int drop, int unless)
4602 struct ceph_mds_request_release *rel = *p;
4603 struct ceph_dentry_info *di = ceph_dentry(dentry);
4607 /* This shouldn't happen */
4611 * force an record for the directory caps if we have a dentry lease.
4612 * this is racy (can't take i_ceph_lock and d_lock together), but it
4613 * doesn't have to be perfect; the mds will revoke anything we don't
4616 spin_lock(&dentry->d_lock);
4617 if (di->lease_session && di->lease_session->s_mds == mds)
4619 spin_unlock(&dentry->d_lock);
4621 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4623 spin_lock(&dentry->d_lock);
4624 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4625 dout("encode_dentry_release %p mds%d seq %d\n",
4626 dentry, mds, (int)di->lease_seq);
4627 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4628 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4629 *p += dentry->d_name.len;
4630 rel->dname_seq = cpu_to_le32(di->lease_seq);
4631 __ceph_mdsc_drop_dentry_lease(dentry);
4633 spin_unlock(&dentry->d_lock);