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 gen = atomic_read(&session->s_cap_gen);
650 cap = __get_cap_for_mds(ci, mds);
656 cap->implemented = 0;
662 __insert_cap_node(ci, cap);
664 /* add to session cap list */
665 cap->session = session;
666 spin_lock(&session->s_cap_lock);
667 list_add_tail(&cap->session_caps, &session->s_caps);
668 session->s_nr_caps++;
669 atomic64_inc(&mdsc->metric.total_caps);
670 spin_unlock(&session->s_cap_lock);
672 spin_lock(&session->s_cap_lock);
673 list_move_tail(&cap->session_caps, &session->s_caps);
674 spin_unlock(&session->s_cap_lock);
676 if (cap->cap_gen < gen)
677 cap->issued = cap->implemented = CEPH_CAP_PIN;
680 * auth mds of the inode changed. we received the cap export
681 * message, but still haven't received the cap import message.
682 * handle_cap_export() updated the new auth MDS' cap.
684 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
685 * a message that was send before the cap import message. So
688 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
689 WARN_ON(cap != ci->i_auth_cap);
690 WARN_ON(cap->cap_id != cap_id);
693 issued |= cap->issued;
694 flags |= CEPH_CAP_FLAG_AUTH;
698 if (!ci->i_snap_realm ||
699 ((flags & CEPH_CAP_FLAG_AUTH) &&
700 realmino != (u64)-1 && ci->i_snap_realm->ino != realmino)) {
702 * add this inode to the appropriate snap realm
704 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
707 ceph_change_snap_realm(inode, realm);
709 WARN(1, "%s: couldn't find snap realm 0x%llx (ino 0x%llx oldrealm 0x%llx)\n",
710 __func__, realmino, ci->i_vino.ino,
711 ci->i_snap_realm ? ci->i_snap_realm->ino : 0);
714 __check_cap_issue(ci, cap, issued);
717 * If we are issued caps we don't want, or the mds' wanted
718 * value appears to be off, queue a check so we'll release
719 * later and/or update the mds wanted value.
721 actual_wanted = __ceph_caps_wanted(ci);
722 if ((wanted & ~actual_wanted) ||
723 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
724 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
725 ceph_cap_string(issued), ceph_cap_string(wanted),
726 ceph_cap_string(actual_wanted));
727 __cap_delay_requeue(mdsc, ci);
730 if (flags & CEPH_CAP_FLAG_AUTH) {
731 if (!ci->i_auth_cap ||
732 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
733 if (ci->i_auth_cap &&
734 ci->i_auth_cap->session != cap->session)
735 change_auth_cap_ses(ci, cap->session);
736 ci->i_auth_cap = cap;
737 cap->mds_wanted = wanted;
740 WARN_ON(ci->i_auth_cap == cap);
743 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
744 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
745 ceph_cap_string(issued|cap->issued), seq, mds);
746 cap->cap_id = cap_id;
747 cap->issued = issued;
748 cap->implemented |= issued;
749 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
750 cap->mds_wanted = wanted;
752 cap->mds_wanted |= wanted;
754 cap->issue_seq = seq;
760 * Return true if cap has not timed out and belongs to the current
761 * generation of the MDS session (i.e. has not gone 'stale' due to
762 * us losing touch with the mds).
764 static int __cap_is_valid(struct ceph_cap *cap)
769 gen = atomic_read(&cap->session->s_cap_gen);
770 ttl = cap->session->s_cap_ttl;
772 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
773 dout("__cap_is_valid %p cap %p issued %s "
774 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
775 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
783 * Return set of valid cap bits issued to us. Note that caps time
784 * out, and may be invalidated in bulk if the client session times out
785 * and session->s_cap_gen is bumped.
787 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
789 int have = ci->i_snap_caps;
790 struct ceph_cap *cap;
795 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
796 cap = rb_entry(p, struct ceph_cap, ci_node);
797 if (!__cap_is_valid(cap))
799 dout("__ceph_caps_issued %p cap %p issued %s\n",
800 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
803 *implemented |= cap->implemented;
806 * exclude caps issued by non-auth MDS, but are been revoking
807 * by the auth MDS. The non-auth MDS should be revoking/exporting
808 * these caps, but the message is delayed.
810 if (ci->i_auth_cap) {
811 cap = ci->i_auth_cap;
812 have &= ~cap->implemented | cap->issued;
818 * Get cap bits issued by caps other than @ocap
820 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
822 int have = ci->i_snap_caps;
823 struct ceph_cap *cap;
826 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
827 cap = rb_entry(p, struct ceph_cap, ci_node);
830 if (!__cap_is_valid(cap))
838 * Move a cap to the end of the LRU (oldest caps at list head, newest
841 static void __touch_cap(struct ceph_cap *cap)
843 struct ceph_mds_session *s = cap->session;
845 spin_lock(&s->s_cap_lock);
846 if (!s->s_cap_iterator) {
847 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
849 list_move_tail(&cap->session_caps, &s->s_caps);
851 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
852 &cap->ci->vfs_inode, cap, s->s_mds);
854 spin_unlock(&s->s_cap_lock);
858 * Check if we hold the given mask. If so, move the cap(s) to the
859 * front of their respective LRUs. (This is the preferred way for
860 * callers to check for caps they want.)
862 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
864 struct ceph_cap *cap;
866 int have = ci->i_snap_caps;
868 if ((have & mask) == mask) {
869 dout("__ceph_caps_issued_mask ino 0x%llx snap issued %s"
870 " (mask %s)\n", ceph_ino(&ci->vfs_inode),
871 ceph_cap_string(have),
872 ceph_cap_string(mask));
876 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
877 cap = rb_entry(p, struct ceph_cap, ci_node);
878 if (!__cap_is_valid(cap))
880 if ((cap->issued & mask) == mask) {
881 dout("__ceph_caps_issued_mask ino 0x%llx cap %p issued %s"
882 " (mask %s)\n", ceph_ino(&ci->vfs_inode), cap,
883 ceph_cap_string(cap->issued),
884 ceph_cap_string(mask));
890 /* does a combination of caps satisfy mask? */
892 if ((have & mask) == mask) {
893 dout("__ceph_caps_issued_mask ino 0x%llx combo issued %s"
894 " (mask %s)\n", ceph_ino(&ci->vfs_inode),
895 ceph_cap_string(cap->issued),
896 ceph_cap_string(mask));
900 /* touch this + preceding caps */
902 for (q = rb_first(&ci->i_caps); q != p;
904 cap = rb_entry(q, struct ceph_cap,
906 if (!__cap_is_valid(cap))
908 if (cap->issued & mask)
919 int __ceph_caps_issued_mask_metric(struct ceph_inode_info *ci, int mask,
922 struct ceph_fs_client *fsc = ceph_sb_to_client(ci->vfs_inode.i_sb);
925 r = __ceph_caps_issued_mask(ci, mask, touch);
927 ceph_update_cap_hit(&fsc->mdsc->metric);
929 ceph_update_cap_mis(&fsc->mdsc->metric);
934 * Return true if mask caps are currently being revoked by an MDS.
936 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
937 struct ceph_cap *ocap, int mask)
939 struct ceph_cap *cap;
942 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
943 cap = rb_entry(p, struct ceph_cap, ci_node);
945 (cap->implemented & ~cap->issued & mask))
951 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
953 struct inode *inode = &ci->vfs_inode;
956 spin_lock(&ci->i_ceph_lock);
957 ret = __ceph_caps_revoking_other(ci, NULL, mask);
958 spin_unlock(&ci->i_ceph_lock);
959 dout("ceph_caps_revoking %p %s = %d\n", inode,
960 ceph_cap_string(mask), ret);
964 int __ceph_caps_used(struct ceph_inode_info *ci)
968 used |= CEPH_CAP_PIN;
970 used |= CEPH_CAP_FILE_RD;
971 if (ci->i_rdcache_ref ||
972 (S_ISREG(ci->vfs_inode.i_mode) &&
973 ci->vfs_inode.i_data.nrpages))
974 used |= CEPH_CAP_FILE_CACHE;
976 used |= CEPH_CAP_FILE_WR;
977 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
978 used |= CEPH_CAP_FILE_BUFFER;
980 used |= CEPH_CAP_FILE_EXCL;
984 #define FMODE_WAIT_BIAS 1000
987 * wanted, by virtue of open file modes
989 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
991 const int PIN_SHIFT = ffs(CEPH_FILE_MODE_PIN);
992 const int RD_SHIFT = ffs(CEPH_FILE_MODE_RD);
993 const int WR_SHIFT = ffs(CEPH_FILE_MODE_WR);
994 const int LAZY_SHIFT = ffs(CEPH_FILE_MODE_LAZY);
995 struct ceph_mount_options *opt =
996 ceph_inode_to_client(&ci->vfs_inode)->mount_options;
997 unsigned long used_cutoff = jiffies - opt->caps_wanted_delay_max * HZ;
998 unsigned long idle_cutoff = jiffies - opt->caps_wanted_delay_min * HZ;
1000 if (S_ISDIR(ci->vfs_inode.i_mode)) {
1003 /* use used_cutoff here, to keep dir's wanted caps longer */
1004 if (ci->i_nr_by_mode[RD_SHIFT] > 0 ||
1005 time_after(ci->i_last_rd, used_cutoff))
1006 want |= CEPH_CAP_ANY_SHARED;
1008 if (ci->i_nr_by_mode[WR_SHIFT] > 0 ||
1009 time_after(ci->i_last_wr, used_cutoff)) {
1010 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1011 if (opt->flags & CEPH_MOUNT_OPT_ASYNC_DIROPS)
1012 want |= CEPH_CAP_ANY_DIR_OPS;
1015 if (want || ci->i_nr_by_mode[PIN_SHIFT] > 0)
1016 want |= CEPH_CAP_PIN;
1022 if (ci->i_nr_by_mode[RD_SHIFT] > 0) {
1023 if (ci->i_nr_by_mode[RD_SHIFT] >= FMODE_WAIT_BIAS ||
1024 time_after(ci->i_last_rd, used_cutoff))
1025 bits |= 1 << RD_SHIFT;
1026 } else if (time_after(ci->i_last_rd, idle_cutoff)) {
1027 bits |= 1 << RD_SHIFT;
1030 if (ci->i_nr_by_mode[WR_SHIFT] > 0) {
1031 if (ci->i_nr_by_mode[WR_SHIFT] >= FMODE_WAIT_BIAS ||
1032 time_after(ci->i_last_wr, used_cutoff))
1033 bits |= 1 << WR_SHIFT;
1034 } else if (time_after(ci->i_last_wr, idle_cutoff)) {
1035 bits |= 1 << WR_SHIFT;
1038 /* check lazyio only when read/write is wanted */
1039 if ((bits & (CEPH_FILE_MODE_RDWR << 1)) &&
1040 ci->i_nr_by_mode[LAZY_SHIFT] > 0)
1041 bits |= 1 << LAZY_SHIFT;
1043 return bits ? ceph_caps_for_mode(bits >> 1) : 0;
1048 * wanted, by virtue of open file modes AND cap refs (buffered/cached data)
1050 int __ceph_caps_wanted(struct ceph_inode_info *ci)
1052 int w = __ceph_caps_file_wanted(ci) | __ceph_caps_used(ci);
1053 if (S_ISDIR(ci->vfs_inode.i_mode)) {
1054 /* we want EXCL if holding caps of dir ops */
1055 if (w & CEPH_CAP_ANY_DIR_OPS)
1056 w |= CEPH_CAP_FILE_EXCL;
1058 /* we want EXCL if dirty data */
1059 if (w & CEPH_CAP_FILE_BUFFER)
1060 w |= CEPH_CAP_FILE_EXCL;
1066 * Return caps we have registered with the MDS(s) as 'wanted'.
1068 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
1070 struct ceph_cap *cap;
1074 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1075 cap = rb_entry(p, struct ceph_cap, ci_node);
1076 if (check && !__cap_is_valid(cap))
1078 if (cap == ci->i_auth_cap)
1079 mds_wanted |= cap->mds_wanted;
1081 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
1086 int ceph_is_any_caps(struct inode *inode)
1088 struct ceph_inode_info *ci = ceph_inode(inode);
1091 spin_lock(&ci->i_ceph_lock);
1092 ret = __ceph_is_any_real_caps(ci);
1093 spin_unlock(&ci->i_ceph_lock);
1099 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
1101 * caller should hold i_ceph_lock.
1102 * caller will not hold session s_mutex if called from destroy_inode.
1104 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1106 struct ceph_mds_session *session = cap->session;
1107 struct ceph_inode_info *ci = cap->ci;
1108 struct ceph_mds_client *mdsc;
1111 /* 'ci' being NULL means the remove have already occurred */
1113 dout("%s: cap inode is NULL\n", __func__);
1117 lockdep_assert_held(&ci->i_ceph_lock);
1119 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1121 mdsc = ceph_inode_to_client(&ci->vfs_inode)->mdsc;
1123 /* remove from inode's cap rbtree, and clear auth cap */
1124 rb_erase(&cap->ci_node, &ci->i_caps);
1125 if (ci->i_auth_cap == cap)
1126 ci->i_auth_cap = NULL;
1128 /* remove from session list */
1129 spin_lock(&session->s_cap_lock);
1130 if (session->s_cap_iterator == cap) {
1131 /* not yet, we are iterating over this very cap */
1132 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1135 list_del_init(&cap->session_caps);
1136 session->s_nr_caps--;
1137 atomic64_dec(&mdsc->metric.total_caps);
1138 cap->session = NULL;
1141 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1145 * s_cap_reconnect is protected by s_cap_lock. no one changes
1146 * s_cap_gen while session is in the reconnect state.
1148 if (queue_release &&
1149 (!session->s_cap_reconnect ||
1150 cap->cap_gen == atomic_read(&session->s_cap_gen))) {
1151 cap->queue_release = 1;
1153 __ceph_queue_cap_release(session, cap);
1157 cap->queue_release = 0;
1159 cap->cap_ino = ci->i_vino.ino;
1161 spin_unlock(&session->s_cap_lock);
1164 ceph_put_cap(mdsc, cap);
1166 if (!__ceph_is_any_real_caps(ci)) {
1167 /* when reconnect denied, we remove session caps forcibly,
1168 * i_wr_ref can be non-zero. If there are ongoing write,
1169 * keep i_snap_realm.
1171 if (ci->i_wr_ref == 0 && ci->i_snap_realm)
1172 ceph_change_snap_realm(&ci->vfs_inode, NULL);
1174 __cap_delay_cancel(mdsc, ci);
1178 void ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1180 struct ceph_inode_info *ci = cap->ci;
1181 struct ceph_fs_client *fsc;
1183 /* 'ci' being NULL means the remove have already occurred */
1185 dout("%s: cap inode is NULL\n", __func__);
1189 lockdep_assert_held(&ci->i_ceph_lock);
1191 fsc = ceph_inode_to_client(&ci->vfs_inode);
1192 WARN_ON_ONCE(ci->i_auth_cap == cap &&
1193 !list_empty(&ci->i_dirty_item) &&
1194 !fsc->blocklisted &&
1195 !ceph_inode_is_shutdown(&ci->vfs_inode));
1197 __ceph_remove_cap(cap, queue_release);
1200 struct cap_msg_args {
1201 struct ceph_mds_session *session;
1202 u64 ino, cid, follows;
1203 u64 flush_tid, oldest_flush_tid, size, max_size;
1206 struct ceph_buffer *xattr_buf;
1207 struct ceph_buffer *old_xattr_buf;
1208 struct timespec64 atime, mtime, ctime, btime;
1209 int op, caps, wanted, dirty;
1210 u32 seq, issue_seq, mseq, time_warp_seq;
1220 * cap struct size + flock buffer size + inline version + inline data size +
1221 * osd_epoch_barrier + oldest_flush_tid
1223 #define CAP_MSG_SIZE (sizeof(struct ceph_mds_caps) + \
1224 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4)
1226 /* Marshal up the cap msg to the MDS */
1227 static void encode_cap_msg(struct ceph_msg *msg, struct cap_msg_args *arg)
1229 struct ceph_mds_caps *fc;
1231 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1233 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",
1234 __func__, ceph_cap_op_name(arg->op), arg->cid, arg->ino,
1235 ceph_cap_string(arg->caps), ceph_cap_string(arg->wanted),
1236 ceph_cap_string(arg->dirty), arg->seq, arg->issue_seq,
1237 arg->flush_tid, arg->oldest_flush_tid, arg->mseq, arg->follows,
1238 arg->size, arg->max_size, arg->xattr_version,
1239 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1241 msg->hdr.version = cpu_to_le16(10);
1242 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1244 fc = msg->front.iov_base;
1245 memset(fc, 0, sizeof(*fc));
1247 fc->cap_id = cpu_to_le64(arg->cid);
1248 fc->op = cpu_to_le32(arg->op);
1249 fc->seq = cpu_to_le32(arg->seq);
1250 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1251 fc->migrate_seq = cpu_to_le32(arg->mseq);
1252 fc->caps = cpu_to_le32(arg->caps);
1253 fc->wanted = cpu_to_le32(arg->wanted);
1254 fc->dirty = cpu_to_le32(arg->dirty);
1255 fc->ino = cpu_to_le64(arg->ino);
1256 fc->snap_follows = cpu_to_le64(arg->follows);
1258 fc->size = cpu_to_le64(arg->size);
1259 fc->max_size = cpu_to_le64(arg->max_size);
1260 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1261 ceph_encode_timespec64(&fc->atime, &arg->atime);
1262 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1263 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1265 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1266 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1267 fc->mode = cpu_to_le32(arg->mode);
1269 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1270 if (arg->xattr_buf) {
1271 msg->middle = ceph_buffer_get(arg->xattr_buf);
1272 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1273 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1277 /* flock buffer size (version 2) */
1278 ceph_encode_32(&p, 0);
1279 /* inline version (version 4) */
1280 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1281 /* inline data size */
1282 ceph_encode_32(&p, 0);
1284 * osd_epoch_barrier (version 5)
1285 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1286 * case it was recently changed
1288 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1289 /* oldest_flush_tid (version 6) */
1290 ceph_encode_64(&p, arg->oldest_flush_tid);
1293 * caller_uid/caller_gid (version 7)
1295 * Currently, we don't properly track which caller dirtied the caps
1296 * last, and force a flush of them when there is a conflict. For now,
1297 * just set this to 0:0, to emulate how the MDS has worked up to now.
1299 ceph_encode_32(&p, 0);
1300 ceph_encode_32(&p, 0);
1302 /* pool namespace (version 8) (mds always ignores this) */
1303 ceph_encode_32(&p, 0);
1305 /* btime and change_attr (version 9) */
1306 ceph_encode_timespec64(p, &arg->btime);
1307 p += sizeof(struct ceph_timespec);
1308 ceph_encode_64(&p, arg->change_attr);
1310 /* Advisory flags (version 10) */
1311 ceph_encode_32(&p, arg->flags);
1315 * Queue cap releases when an inode is dropped from our cache.
1317 void __ceph_remove_caps(struct ceph_inode_info *ci)
1321 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1322 * may call __ceph_caps_issued_mask() on a freeing inode. */
1323 spin_lock(&ci->i_ceph_lock);
1324 p = rb_first(&ci->i_caps);
1326 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1328 ceph_remove_cap(cap, true);
1330 spin_unlock(&ci->i_ceph_lock);
1334 * Prepare to send a cap message to an MDS. Update the cap state, and populate
1335 * the arg struct with the parameters that will need to be sent. This should
1336 * be done under the i_ceph_lock to guard against changes to cap state.
1338 * Make note of max_size reported/requested from mds, revoked caps
1339 * that have now been implemented.
1341 static void __prep_cap(struct cap_msg_args *arg, struct ceph_cap *cap,
1342 int op, int flags, int used, int want, int retain,
1343 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1345 struct ceph_inode_info *ci = cap->ci;
1346 struct inode *inode = &ci->vfs_inode;
1349 lockdep_assert_held(&ci->i_ceph_lock);
1351 held = cap->issued | cap->implemented;
1352 revoking = cap->implemented & ~cap->issued;
1353 retain &= ~revoking;
1355 dout("%s %p cap %p session %p %s -> %s (revoking %s)\n",
1356 __func__, inode, cap, cap->session,
1357 ceph_cap_string(held), ceph_cap_string(held & retain),
1358 ceph_cap_string(revoking));
1359 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1361 ci->i_ceph_flags &= ~CEPH_I_FLUSH;
1363 cap->issued &= retain; /* drop bits we don't want */
1365 * Wake up any waiters on wanted -> needed transition. This is due to
1366 * the weird transition from buffered to sync IO... we need to flush
1367 * dirty pages _before_ allowing sync writes to avoid reordering.
1369 arg->wake = cap->implemented & ~cap->issued;
1370 cap->implemented &= cap->issued | used;
1371 cap->mds_wanted = want;
1373 arg->session = cap->session;
1374 arg->ino = ceph_vino(inode).ino;
1375 arg->cid = cap->cap_id;
1376 arg->follows = flushing ? ci->i_head_snapc->seq : 0;
1377 arg->flush_tid = flush_tid;
1378 arg->oldest_flush_tid = oldest_flush_tid;
1380 arg->size = i_size_read(inode);
1381 ci->i_reported_size = arg->size;
1382 arg->max_size = ci->i_wanted_max_size;
1383 if (cap == ci->i_auth_cap) {
1384 if (want & CEPH_CAP_ANY_FILE_WR)
1385 ci->i_requested_max_size = arg->max_size;
1387 ci->i_requested_max_size = 0;
1390 if (flushing & CEPH_CAP_XATTR_EXCL) {
1391 arg->old_xattr_buf = __ceph_build_xattrs_blob(ci);
1392 arg->xattr_version = ci->i_xattrs.version;
1393 arg->xattr_buf = ci->i_xattrs.blob;
1395 arg->xattr_buf = NULL;
1396 arg->old_xattr_buf = NULL;
1399 arg->mtime = inode->i_mtime;
1400 arg->atime = inode->i_atime;
1401 arg->ctime = inode->i_ctime;
1402 arg->btime = ci->i_btime;
1403 arg->change_attr = inode_peek_iversion_raw(inode);
1406 arg->caps = cap->implemented;
1408 arg->dirty = flushing;
1410 arg->seq = cap->seq;
1411 arg->issue_seq = cap->issue_seq;
1412 arg->mseq = cap->mseq;
1413 arg->time_warp_seq = ci->i_time_warp_seq;
1415 arg->uid = inode->i_uid;
1416 arg->gid = inode->i_gid;
1417 arg->mode = inode->i_mode;
1419 arg->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1420 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1421 !list_empty(&ci->i_cap_snaps)) {
1422 struct ceph_cap_snap *capsnap;
1423 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1424 if (capsnap->cap_flush.tid)
1426 if (capsnap->need_flush) {
1427 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1436 * Send a cap msg on the given inode.
1438 * Caller should hold snap_rwsem (read), s_mutex.
1440 static void __send_cap(struct cap_msg_args *arg, struct ceph_inode_info *ci)
1442 struct ceph_msg *msg;
1443 struct inode *inode = &ci->vfs_inode;
1445 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1447 pr_err("error allocating cap msg: ino (%llx.%llx) flushing %s tid %llu, requeuing cap.\n",
1448 ceph_vinop(inode), ceph_cap_string(arg->dirty),
1450 spin_lock(&ci->i_ceph_lock);
1451 __cap_delay_requeue(arg->session->s_mdsc, ci);
1452 spin_unlock(&ci->i_ceph_lock);
1456 encode_cap_msg(msg, arg);
1457 ceph_con_send(&arg->session->s_con, msg);
1458 ceph_buffer_put(arg->old_xattr_buf);
1460 wake_up_all(&ci->i_cap_wq);
1463 static inline int __send_flush_snap(struct inode *inode,
1464 struct ceph_mds_session *session,
1465 struct ceph_cap_snap *capsnap,
1466 u32 mseq, u64 oldest_flush_tid)
1468 struct cap_msg_args arg;
1469 struct ceph_msg *msg;
1471 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1475 arg.session = session;
1476 arg.ino = ceph_vino(inode).ino;
1478 arg.follows = capsnap->follows;
1479 arg.flush_tid = capsnap->cap_flush.tid;
1480 arg.oldest_flush_tid = oldest_flush_tid;
1482 arg.size = capsnap->size;
1484 arg.xattr_version = capsnap->xattr_version;
1485 arg.xattr_buf = capsnap->xattr_blob;
1486 arg.old_xattr_buf = NULL;
1488 arg.atime = capsnap->atime;
1489 arg.mtime = capsnap->mtime;
1490 arg.ctime = capsnap->ctime;
1491 arg.btime = capsnap->btime;
1492 arg.change_attr = capsnap->change_attr;
1494 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1495 arg.caps = capsnap->issued;
1497 arg.dirty = capsnap->dirty;
1502 arg.time_warp_seq = capsnap->time_warp_seq;
1504 arg.uid = capsnap->uid;
1505 arg.gid = capsnap->gid;
1506 arg.mode = capsnap->mode;
1508 arg.inline_data = capsnap->inline_data;
1512 encode_cap_msg(msg, &arg);
1513 ceph_con_send(&arg.session->s_con, msg);
1518 * When a snapshot is taken, clients accumulate dirty metadata on
1519 * inodes with capabilities in ceph_cap_snaps to describe the file
1520 * state at the time the snapshot was taken. This must be flushed
1521 * asynchronously back to the MDS once sync writes complete and dirty
1522 * data is written out.
1524 * Called under i_ceph_lock.
1526 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1527 struct ceph_mds_session *session)
1528 __releases(ci->i_ceph_lock)
1529 __acquires(ci->i_ceph_lock)
1531 struct inode *inode = &ci->vfs_inode;
1532 struct ceph_mds_client *mdsc = session->s_mdsc;
1533 struct ceph_cap_snap *capsnap;
1534 u64 oldest_flush_tid = 0;
1535 u64 first_tid = 1, last_tid = 0;
1537 dout("__flush_snaps %p session %p\n", inode, session);
1539 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1541 * we need to wait for sync writes to complete and for dirty
1542 * pages to be written out.
1544 if (capsnap->dirty_pages || capsnap->writing)
1547 /* should be removed by ceph_try_drop_cap_snap() */
1548 BUG_ON(!capsnap->need_flush);
1550 /* only flush each capsnap once */
1551 if (capsnap->cap_flush.tid > 0) {
1552 dout(" already flushed %p, skipping\n", capsnap);
1556 spin_lock(&mdsc->cap_dirty_lock);
1557 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1558 list_add_tail(&capsnap->cap_flush.g_list,
1559 &mdsc->cap_flush_list);
1560 if (oldest_flush_tid == 0)
1561 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1562 if (list_empty(&ci->i_flushing_item)) {
1563 list_add_tail(&ci->i_flushing_item,
1564 &session->s_cap_flushing);
1566 spin_unlock(&mdsc->cap_dirty_lock);
1568 list_add_tail(&capsnap->cap_flush.i_list,
1569 &ci->i_cap_flush_list);
1572 first_tid = capsnap->cap_flush.tid;
1573 last_tid = capsnap->cap_flush.tid;
1576 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1578 while (first_tid <= last_tid) {
1579 struct ceph_cap *cap = ci->i_auth_cap;
1580 struct ceph_cap_flush *cf;
1583 if (!(cap && cap->session == session)) {
1584 dout("__flush_snaps %p auth cap %p not mds%d, "
1585 "stop\n", inode, cap, session->s_mds);
1590 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1591 if (cf->tid >= first_tid) {
1599 first_tid = cf->tid + 1;
1601 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1602 refcount_inc(&capsnap->nref);
1603 spin_unlock(&ci->i_ceph_lock);
1605 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1606 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1608 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1611 pr_err("__flush_snaps: error sending cap flushsnap, "
1612 "ino (%llx.%llx) tid %llu follows %llu\n",
1613 ceph_vinop(inode), cf->tid, capsnap->follows);
1616 ceph_put_cap_snap(capsnap);
1617 spin_lock(&ci->i_ceph_lock);
1621 void ceph_flush_snaps(struct ceph_inode_info *ci,
1622 struct ceph_mds_session **psession)
1624 struct inode *inode = &ci->vfs_inode;
1625 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1626 struct ceph_mds_session *session = NULL;
1629 dout("ceph_flush_snaps %p\n", inode);
1631 session = *psession;
1633 spin_lock(&ci->i_ceph_lock);
1634 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1635 dout(" no capsnap needs flush, doing nothing\n");
1638 if (!ci->i_auth_cap) {
1639 dout(" no auth cap (migrating?), doing nothing\n");
1643 mds = ci->i_auth_cap->session->s_mds;
1644 if (session && session->s_mds != mds) {
1645 dout(" oops, wrong session %p mutex\n", session);
1646 ceph_put_mds_session(session);
1650 spin_unlock(&ci->i_ceph_lock);
1651 mutex_lock(&mdsc->mutex);
1652 session = __ceph_lookup_mds_session(mdsc, mds);
1653 mutex_unlock(&mdsc->mutex);
1657 // make sure flushsnap messages are sent in proper order.
1658 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1659 __kick_flushing_caps(mdsc, session, ci, 0);
1661 __ceph_flush_snaps(ci, session);
1663 spin_unlock(&ci->i_ceph_lock);
1666 *psession = session;
1668 ceph_put_mds_session(session);
1669 /* we flushed them all; remove this inode from the queue */
1670 spin_lock(&mdsc->snap_flush_lock);
1671 list_del_init(&ci->i_snap_flush_item);
1672 spin_unlock(&mdsc->snap_flush_lock);
1676 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1677 * Caller is then responsible for calling __mark_inode_dirty with the
1678 * returned flags value.
1680 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1681 struct ceph_cap_flush **pcf)
1683 struct ceph_mds_client *mdsc =
1684 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1685 struct inode *inode = &ci->vfs_inode;
1686 int was = ci->i_dirty_caps;
1689 lockdep_assert_held(&ci->i_ceph_lock);
1691 if (!ci->i_auth_cap) {
1692 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1693 "but no auth cap (session was closed?)\n",
1694 inode, ceph_ino(inode), ceph_cap_string(mask));
1698 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1699 ceph_cap_string(mask), ceph_cap_string(was),
1700 ceph_cap_string(was | mask));
1701 ci->i_dirty_caps |= mask;
1703 struct ceph_mds_session *session = ci->i_auth_cap->session;
1705 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1706 swap(ci->i_prealloc_cap_flush, *pcf);
1708 if (!ci->i_head_snapc) {
1709 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1710 ci->i_head_snapc = ceph_get_snap_context(
1711 ci->i_snap_realm->cached_context);
1713 dout(" inode %p now dirty snapc %p auth cap %p\n",
1714 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1715 BUG_ON(!list_empty(&ci->i_dirty_item));
1716 spin_lock(&mdsc->cap_dirty_lock);
1717 list_add(&ci->i_dirty_item, &session->s_cap_dirty);
1718 spin_unlock(&mdsc->cap_dirty_lock);
1719 if (ci->i_flushing_caps == 0) {
1721 dirty |= I_DIRTY_SYNC;
1724 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1726 BUG_ON(list_empty(&ci->i_dirty_item));
1727 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1728 (mask & CEPH_CAP_FILE_BUFFER))
1729 dirty |= I_DIRTY_DATASYNC;
1730 __cap_delay_requeue(mdsc, ci);
1734 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1736 struct ceph_cap_flush *cf;
1738 cf = kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1742 cf->is_capsnap = false;
1746 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1749 kmem_cache_free(ceph_cap_flush_cachep, cf);
1752 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1754 if (!list_empty(&mdsc->cap_flush_list)) {
1755 struct ceph_cap_flush *cf =
1756 list_first_entry(&mdsc->cap_flush_list,
1757 struct ceph_cap_flush, g_list);
1764 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1765 * Return true if caller needs to wake up flush waiters.
1767 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1768 struct ceph_cap_flush *cf)
1770 struct ceph_cap_flush *prev;
1771 bool wake = cf->wake;
1773 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1774 prev = list_prev_entry(cf, g_list);
1778 list_del_init(&cf->g_list);
1782 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1783 struct ceph_cap_flush *cf)
1785 struct ceph_cap_flush *prev;
1786 bool wake = cf->wake;
1788 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1789 prev = list_prev_entry(cf, i_list);
1793 list_del_init(&cf->i_list);
1798 * Add dirty inode to the flushing list. Assigned a seq number so we
1799 * can wait for caps to flush without starving.
1801 * Called under i_ceph_lock. Returns the flush tid.
1803 static u64 __mark_caps_flushing(struct inode *inode,
1804 struct ceph_mds_session *session, bool wake,
1805 u64 *oldest_flush_tid)
1807 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1808 struct ceph_inode_info *ci = ceph_inode(inode);
1809 struct ceph_cap_flush *cf = NULL;
1812 lockdep_assert_held(&ci->i_ceph_lock);
1813 BUG_ON(ci->i_dirty_caps == 0);
1814 BUG_ON(list_empty(&ci->i_dirty_item));
1815 BUG_ON(!ci->i_prealloc_cap_flush);
1817 flushing = ci->i_dirty_caps;
1818 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1819 ceph_cap_string(flushing),
1820 ceph_cap_string(ci->i_flushing_caps),
1821 ceph_cap_string(ci->i_flushing_caps | flushing));
1822 ci->i_flushing_caps |= flushing;
1823 ci->i_dirty_caps = 0;
1824 dout(" inode %p now !dirty\n", inode);
1826 swap(cf, ci->i_prealloc_cap_flush);
1827 cf->caps = flushing;
1830 spin_lock(&mdsc->cap_dirty_lock);
1831 list_del_init(&ci->i_dirty_item);
1833 cf->tid = ++mdsc->last_cap_flush_tid;
1834 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1835 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1837 if (list_empty(&ci->i_flushing_item)) {
1838 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1839 mdsc->num_cap_flushing++;
1841 spin_unlock(&mdsc->cap_dirty_lock);
1843 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1849 * try to invalidate mapping pages without blocking.
1851 static int try_nonblocking_invalidate(struct inode *inode)
1852 __releases(ci->i_ceph_lock)
1853 __acquires(ci->i_ceph_lock)
1855 struct ceph_inode_info *ci = ceph_inode(inode);
1856 u32 invalidating_gen = ci->i_rdcache_gen;
1858 spin_unlock(&ci->i_ceph_lock);
1859 ceph_fscache_invalidate(inode, false);
1860 invalidate_mapping_pages(&inode->i_data, 0, -1);
1861 spin_lock(&ci->i_ceph_lock);
1863 if (inode->i_data.nrpages == 0 &&
1864 invalidating_gen == ci->i_rdcache_gen) {
1866 dout("try_nonblocking_invalidate %p success\n", inode);
1867 /* save any racing async invalidate some trouble */
1868 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1871 dout("try_nonblocking_invalidate %p failed\n", inode);
1875 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1877 loff_t size = i_size_read(&ci->vfs_inode);
1878 /* mds will adjust max size according to the reported size */
1879 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1881 if (size >= ci->i_max_size)
1883 /* half of previous max_size increment has been used */
1884 if (ci->i_max_size > ci->i_reported_size &&
1885 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1891 * Swiss army knife function to examine currently used and wanted
1892 * versus held caps. Release, flush, ack revoked caps to mds as
1895 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1896 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1899 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1900 struct ceph_mds_session *session)
1902 struct inode *inode = &ci->vfs_inode;
1903 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1904 struct ceph_cap *cap;
1905 u64 flush_tid, oldest_flush_tid;
1906 int file_wanted, used, cap_used;
1907 int issued, implemented, want, retain, revoking, flushing = 0;
1908 int mds = -1; /* keep track of how far we've gone through i_caps list
1909 to avoid an infinite loop on retry */
1911 bool queue_invalidate = false;
1912 bool tried_invalidate = false;
1915 ceph_get_mds_session(session);
1917 spin_lock(&ci->i_ceph_lock);
1918 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
1919 /* Don't send messages until we get async create reply */
1920 spin_unlock(&ci->i_ceph_lock);
1921 ceph_put_mds_session(session);
1925 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1926 flags |= CHECK_CAPS_FLUSH;
1928 /* Caps wanted by virtue of active open files. */
1929 file_wanted = __ceph_caps_file_wanted(ci);
1931 /* Caps which have active references against them */
1932 used = __ceph_caps_used(ci);
1935 * "issued" represents the current caps that the MDS wants us to have.
1936 * "implemented" is the set that we have been granted, and includes the
1937 * ones that have not yet been returned to the MDS (the "revoking" set,
1938 * usually because they have outstanding references).
1940 issued = __ceph_caps_issued(ci, &implemented);
1941 revoking = implemented & ~issued;
1945 /* The ones we currently want to retain (may be adjusted below) */
1946 retain = file_wanted | used | CEPH_CAP_PIN;
1947 if (!mdsc->stopping && inode->i_nlink > 0) {
1949 retain |= CEPH_CAP_ANY; /* be greedy */
1950 } else if (S_ISDIR(inode->i_mode) &&
1951 (issued & CEPH_CAP_FILE_SHARED) &&
1952 __ceph_dir_is_complete(ci)) {
1954 * If a directory is complete, we want to keep
1955 * the exclusive cap. So that MDS does not end up
1956 * revoking the shared cap on every create/unlink
1959 if (IS_RDONLY(inode)) {
1960 want = CEPH_CAP_ANY_SHARED;
1962 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1967 retain |= CEPH_CAP_ANY_SHARED;
1969 * keep RD only if we didn't have the file open RW,
1970 * because then the mds would revoke it anyway to
1971 * journal max_size=0.
1973 if (ci->i_max_size == 0)
1974 retain |= CEPH_CAP_ANY_RD;
1978 dout("check_caps %llx.%llx file_want %s used %s dirty %s flushing %s"
1979 " issued %s revoking %s retain %s %s%s\n", ceph_vinop(inode),
1980 ceph_cap_string(file_wanted),
1981 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1982 ceph_cap_string(ci->i_flushing_caps),
1983 ceph_cap_string(issued), ceph_cap_string(revoking),
1984 ceph_cap_string(retain),
1985 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1986 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1989 * If we no longer need to hold onto old our caps, and we may
1990 * have cached pages, but don't want them, then try to invalidate.
1991 * If we fail, it's because pages are locked.... try again later.
1993 if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
1994 S_ISREG(inode->i_mode) &&
1995 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1996 inode->i_data.nrpages && /* have cached pages */
1997 (revoking & (CEPH_CAP_FILE_CACHE|
1998 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1999 !tried_invalidate) {
2000 dout("check_caps trying to invalidate on %llx.%llx\n",
2002 if (try_nonblocking_invalidate(inode) < 0) {
2003 dout("check_caps queuing invalidate\n");
2004 queue_invalidate = true;
2005 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2007 tried_invalidate = true;
2011 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2013 struct cap_msg_args arg;
2015 cap = rb_entry(p, struct ceph_cap, ci_node);
2017 /* avoid looping forever */
2018 if (mds >= cap->mds ||
2019 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
2023 * If we have an auth cap, we don't need to consider any
2024 * overlapping caps as used.
2027 if (ci->i_auth_cap && cap != ci->i_auth_cap)
2028 cap_used &= ~ci->i_auth_cap->issued;
2030 revoking = cap->implemented & ~cap->issued;
2031 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
2032 cap->mds, cap, ceph_cap_string(cap_used),
2033 ceph_cap_string(cap->issued),
2034 ceph_cap_string(cap->implemented),
2035 ceph_cap_string(revoking));
2037 if (cap == ci->i_auth_cap &&
2038 (cap->issued & CEPH_CAP_FILE_WR)) {
2039 /* request larger max_size from MDS? */
2040 if (ci->i_wanted_max_size > ci->i_max_size &&
2041 ci->i_wanted_max_size > ci->i_requested_max_size) {
2042 dout("requesting new max_size\n");
2046 /* approaching file_max? */
2047 if (__ceph_should_report_size(ci)) {
2048 dout("i_size approaching max_size\n");
2052 /* flush anything dirty? */
2053 if (cap == ci->i_auth_cap) {
2054 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
2055 dout("flushing dirty caps\n");
2058 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
2059 dout("flushing snap caps\n");
2064 /* completed revocation? going down and there are no caps? */
2065 if (revoking && (revoking & cap_used) == 0) {
2066 dout("completed revocation of %s\n",
2067 ceph_cap_string(cap->implemented & ~cap->issued));
2071 /* want more caps from mds? */
2072 if (want & ~cap->mds_wanted) {
2073 if (want & ~(cap->mds_wanted | cap->issued))
2075 if (!__cap_is_valid(cap))
2079 /* things we might delay */
2080 if ((cap->issued & ~retain) == 0)
2081 continue; /* nope, all good */
2084 ceph_put_mds_session(session);
2085 session = ceph_get_mds_session(cap->session);
2087 /* kick flushing and flush snaps before sending normal
2089 if (cap == ci->i_auth_cap &&
2091 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2092 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2093 __kick_flushing_caps(mdsc, session, ci, 0);
2094 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2095 __ceph_flush_snaps(ci, session);
2100 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2101 flushing = ci->i_dirty_caps;
2102 flush_tid = __mark_caps_flushing(inode, session, false,
2104 if (flags & CHECK_CAPS_FLUSH &&
2105 list_empty(&session->s_cap_dirty))
2106 mflags |= CEPH_CLIENT_CAPS_SYNC;
2110 spin_lock(&mdsc->cap_dirty_lock);
2111 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2112 spin_unlock(&mdsc->cap_dirty_lock);
2115 mds = cap->mds; /* remember mds, so we don't repeat */
2117 __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2118 want, retain, flushing, flush_tid, oldest_flush_tid);
2120 spin_unlock(&ci->i_ceph_lock);
2121 __send_cap(&arg, ci);
2122 spin_lock(&ci->i_ceph_lock);
2124 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2127 /* periodically re-calculate caps wanted by open files */
2128 if (__ceph_is_any_real_caps(ci) &&
2129 list_empty(&ci->i_cap_delay_list) &&
2130 (file_wanted & ~CEPH_CAP_PIN) &&
2131 !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2132 __cap_delay_requeue(mdsc, ci);
2135 spin_unlock(&ci->i_ceph_lock);
2137 ceph_put_mds_session(session);
2138 if (queue_invalidate)
2139 ceph_queue_invalidate(inode);
2143 * Try to flush dirty caps back to the auth mds.
2145 static int try_flush_caps(struct inode *inode, u64 *ptid)
2147 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2148 struct ceph_inode_info *ci = ceph_inode(inode);
2150 u64 flush_tid = 0, oldest_flush_tid = 0;
2152 spin_lock(&ci->i_ceph_lock);
2154 if (ci->i_dirty_caps && ci->i_auth_cap) {
2155 struct ceph_cap *cap = ci->i_auth_cap;
2156 struct cap_msg_args arg;
2157 struct ceph_mds_session *session = cap->session;
2159 if (session->s_state < CEPH_MDS_SESSION_OPEN) {
2160 spin_unlock(&ci->i_ceph_lock);
2164 if (ci->i_ceph_flags &
2165 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2166 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2167 __kick_flushing_caps(mdsc, session, ci, 0);
2168 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2169 __ceph_flush_snaps(ci, session);
2173 flushing = ci->i_dirty_caps;
2174 flush_tid = __mark_caps_flushing(inode, session, true,
2177 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2178 __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2179 (cap->issued | cap->implemented),
2180 flushing, flush_tid, oldest_flush_tid);
2181 spin_unlock(&ci->i_ceph_lock);
2183 __send_cap(&arg, ci);
2185 if (!list_empty(&ci->i_cap_flush_list)) {
2186 struct ceph_cap_flush *cf =
2187 list_last_entry(&ci->i_cap_flush_list,
2188 struct ceph_cap_flush, i_list);
2190 flush_tid = cf->tid;
2192 flushing = ci->i_flushing_caps;
2193 spin_unlock(&ci->i_ceph_lock);
2201 * Return true if we've flushed caps through the given flush_tid.
2203 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2205 struct ceph_inode_info *ci = ceph_inode(inode);
2208 spin_lock(&ci->i_ceph_lock);
2209 if (!list_empty(&ci->i_cap_flush_list)) {
2210 struct ceph_cap_flush * cf =
2211 list_first_entry(&ci->i_cap_flush_list,
2212 struct ceph_cap_flush, i_list);
2213 if (cf->tid <= flush_tid)
2216 spin_unlock(&ci->i_ceph_lock);
2221 * wait for any unsafe requests to complete.
2223 static int unsafe_request_wait(struct inode *inode)
2225 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2226 struct ceph_inode_info *ci = ceph_inode(inode);
2227 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2228 unsigned int max_sessions;
2231 spin_lock(&ci->i_unsafe_lock);
2232 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2233 req1 = list_last_entry(&ci->i_unsafe_dirops,
2234 struct ceph_mds_request,
2236 ceph_mdsc_get_request(req1);
2238 if (!list_empty(&ci->i_unsafe_iops)) {
2239 req2 = list_last_entry(&ci->i_unsafe_iops,
2240 struct ceph_mds_request,
2241 r_unsafe_target_item);
2242 ceph_mdsc_get_request(req2);
2244 spin_unlock(&ci->i_unsafe_lock);
2247 * The mdsc->max_sessions is unlikely to be changed
2248 * mostly, here we will retry it by reallocating the
2249 * sessions array memory to get rid of the mdsc->mutex
2253 max_sessions = mdsc->max_sessions;
2256 * Trigger to flush the journal logs in all the relevant MDSes
2257 * manually, or in the worst case we must wait at most 5 seconds
2258 * to wait the journal logs to be flushed by the MDSes periodically.
2260 if ((req1 || req2) && likely(max_sessions)) {
2261 struct ceph_mds_session **sessions = NULL;
2262 struct ceph_mds_session *s;
2263 struct ceph_mds_request *req;
2266 sessions = kzalloc(max_sessions * sizeof(s), GFP_KERNEL);
2272 spin_lock(&ci->i_unsafe_lock);
2274 list_for_each_entry(req, &ci->i_unsafe_dirops,
2275 r_unsafe_dir_item) {
2277 if (unlikely(s->s_mds >= max_sessions)) {
2278 spin_unlock(&ci->i_unsafe_lock);
2279 for (i = 0; i < max_sessions; i++) {
2282 ceph_put_mds_session(s);
2287 if (!sessions[s->s_mds]) {
2288 s = ceph_get_mds_session(s);
2289 sessions[s->s_mds] = s;
2294 list_for_each_entry(req, &ci->i_unsafe_iops,
2295 r_unsafe_target_item) {
2297 if (unlikely(s->s_mds >= max_sessions)) {
2298 spin_unlock(&ci->i_unsafe_lock);
2299 for (i = 0; i < max_sessions; i++) {
2302 ceph_put_mds_session(s);
2307 if (!sessions[s->s_mds]) {
2308 s = ceph_get_mds_session(s);
2309 sessions[s->s_mds] = s;
2313 spin_unlock(&ci->i_unsafe_lock);
2316 spin_lock(&ci->i_ceph_lock);
2317 if (ci->i_auth_cap) {
2318 s = ci->i_auth_cap->session;
2319 if (!sessions[s->s_mds])
2320 sessions[s->s_mds] = ceph_get_mds_session(s);
2322 spin_unlock(&ci->i_ceph_lock);
2324 /* send flush mdlog request to MDSes */
2325 for (i = 0; i < max_sessions; i++) {
2328 send_flush_mdlog(s);
2329 ceph_put_mds_session(s);
2335 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2336 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2338 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2339 ceph_timeout_jiffies(req1->r_timeout));
2344 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2345 ceph_timeout_jiffies(req2->r_timeout));
2352 ceph_mdsc_put_request(req1);
2354 ceph_mdsc_put_request(req2);
2358 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2360 struct inode *inode = file->f_mapping->host;
2361 struct ceph_inode_info *ci = ceph_inode(inode);
2366 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2368 ret = file_write_and_wait_range(file, start, end);
2372 ret = ceph_wait_on_async_create(inode);
2376 dirty = try_flush_caps(inode, &flush_tid);
2377 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2379 err = unsafe_request_wait(inode);
2382 * only wait on non-file metadata writeback (the mds
2383 * can recover size and mtime, so we don't need to
2386 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2387 err = wait_event_interruptible(ci->i_cap_wq,
2388 caps_are_flushed(inode, flush_tid));
2394 err = file_check_and_advance_wb_err(file);
2398 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2403 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2404 * queue inode for flush but don't do so immediately, because we can
2405 * get by with fewer MDS messages if we wait for data writeback to
2408 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2410 struct ceph_inode_info *ci = ceph_inode(inode);
2414 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2416 dout("write_inode %p wait=%d\n", inode, wait);
2417 ceph_fscache_unpin_writeback(inode, wbc);
2419 err = ceph_wait_on_async_create(inode);
2422 dirty = try_flush_caps(inode, &flush_tid);
2424 err = wait_event_interruptible(ci->i_cap_wq,
2425 caps_are_flushed(inode, flush_tid));
2427 struct ceph_mds_client *mdsc =
2428 ceph_sb_to_client(inode->i_sb)->mdsc;
2430 spin_lock(&ci->i_ceph_lock);
2431 if (__ceph_caps_dirty(ci))
2432 __cap_delay_requeue_front(mdsc, ci);
2433 spin_unlock(&ci->i_ceph_lock);
2438 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2439 struct ceph_mds_session *session,
2440 struct ceph_inode_info *ci,
2441 u64 oldest_flush_tid)
2442 __releases(ci->i_ceph_lock)
2443 __acquires(ci->i_ceph_lock)
2445 struct inode *inode = &ci->vfs_inode;
2446 struct ceph_cap *cap;
2447 struct ceph_cap_flush *cf;
2450 u64 last_snap_flush = 0;
2452 /* Don't do anything until create reply comes in */
2453 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE)
2456 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2458 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2459 if (cf->is_capsnap) {
2460 last_snap_flush = cf->tid;
2465 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2466 if (cf->tid < first_tid)
2469 cap = ci->i_auth_cap;
2470 if (!(cap && cap->session == session)) {
2471 pr_err("%p auth cap %p not mds%d ???\n",
2472 inode, cap, session->s_mds);
2476 first_tid = cf->tid + 1;
2478 if (!cf->is_capsnap) {
2479 struct cap_msg_args arg;
2481 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2482 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2483 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2484 (cf->tid < last_snap_flush ?
2485 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2486 __ceph_caps_used(ci),
2487 __ceph_caps_wanted(ci),
2488 (cap->issued | cap->implemented),
2489 cf->caps, cf->tid, oldest_flush_tid);
2490 spin_unlock(&ci->i_ceph_lock);
2491 __send_cap(&arg, ci);
2493 struct ceph_cap_snap *capsnap =
2494 container_of(cf, struct ceph_cap_snap,
2496 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2497 inode, capsnap, cf->tid,
2498 ceph_cap_string(capsnap->dirty));
2500 refcount_inc(&capsnap->nref);
2501 spin_unlock(&ci->i_ceph_lock);
2503 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2506 pr_err("kick_flushing_caps: error sending "
2507 "cap flushsnap, ino (%llx.%llx) "
2508 "tid %llu follows %llu\n",
2509 ceph_vinop(inode), cf->tid,
2513 ceph_put_cap_snap(capsnap);
2516 spin_lock(&ci->i_ceph_lock);
2520 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2521 struct ceph_mds_session *session)
2523 struct ceph_inode_info *ci;
2524 struct ceph_cap *cap;
2525 u64 oldest_flush_tid;
2527 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2529 spin_lock(&mdsc->cap_dirty_lock);
2530 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2531 spin_unlock(&mdsc->cap_dirty_lock);
2533 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2534 spin_lock(&ci->i_ceph_lock);
2535 cap = ci->i_auth_cap;
2536 if (!(cap && cap->session == session)) {
2537 pr_err("%p auth cap %p not mds%d ???\n",
2538 &ci->vfs_inode, cap, session->s_mds);
2539 spin_unlock(&ci->i_ceph_lock);
2545 * if flushing caps were revoked, we re-send the cap flush
2546 * in client reconnect stage. This guarantees MDS * processes
2547 * the cap flush message before issuing the flushing caps to
2550 if ((cap->issued & ci->i_flushing_caps) !=
2551 ci->i_flushing_caps) {
2552 /* encode_caps_cb() also will reset these sequence
2553 * numbers. make sure sequence numbers in cap flush
2554 * message match later reconnect message */
2558 __kick_flushing_caps(mdsc, session, ci,
2561 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2564 spin_unlock(&ci->i_ceph_lock);
2568 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2569 struct ceph_mds_session *session)
2571 struct ceph_inode_info *ci;
2572 struct ceph_cap *cap;
2573 u64 oldest_flush_tid;
2575 lockdep_assert_held(&session->s_mutex);
2577 dout("kick_flushing_caps mds%d\n", session->s_mds);
2579 spin_lock(&mdsc->cap_dirty_lock);
2580 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2581 spin_unlock(&mdsc->cap_dirty_lock);
2583 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2584 spin_lock(&ci->i_ceph_lock);
2585 cap = ci->i_auth_cap;
2586 if (!(cap && cap->session == session)) {
2587 pr_err("%p auth cap %p not mds%d ???\n",
2588 &ci->vfs_inode, cap, session->s_mds);
2589 spin_unlock(&ci->i_ceph_lock);
2592 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2593 __kick_flushing_caps(mdsc, session, ci,
2596 spin_unlock(&ci->i_ceph_lock);
2600 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2601 struct ceph_inode_info *ci)
2603 struct ceph_mds_client *mdsc = session->s_mdsc;
2604 struct ceph_cap *cap = ci->i_auth_cap;
2606 lockdep_assert_held(&ci->i_ceph_lock);
2608 dout("%s %p flushing %s\n", __func__, &ci->vfs_inode,
2609 ceph_cap_string(ci->i_flushing_caps));
2611 if (!list_empty(&ci->i_cap_flush_list)) {
2612 u64 oldest_flush_tid;
2613 spin_lock(&mdsc->cap_dirty_lock);
2614 list_move_tail(&ci->i_flushing_item,
2615 &cap->session->s_cap_flushing);
2616 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2617 spin_unlock(&mdsc->cap_dirty_lock);
2619 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2625 * Take references to capabilities we hold, so that we don't release
2626 * them to the MDS prematurely.
2628 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2629 bool snap_rwsem_locked)
2631 lockdep_assert_held(&ci->i_ceph_lock);
2633 if (got & CEPH_CAP_PIN)
2635 if (got & CEPH_CAP_FILE_RD)
2637 if (got & CEPH_CAP_FILE_CACHE)
2638 ci->i_rdcache_ref++;
2639 if (got & CEPH_CAP_FILE_EXCL)
2641 if (got & CEPH_CAP_FILE_WR) {
2642 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2643 BUG_ON(!snap_rwsem_locked);
2644 ci->i_head_snapc = ceph_get_snap_context(
2645 ci->i_snap_realm->cached_context);
2649 if (got & CEPH_CAP_FILE_BUFFER) {
2650 if (ci->i_wb_ref == 0)
2651 ihold(&ci->vfs_inode);
2653 dout("%s %p wb %d -> %d (?)\n", __func__,
2654 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2659 * Try to grab cap references. Specify those refs we @want, and the
2660 * minimal set we @need. Also include the larger offset we are writing
2661 * to (when applicable), and check against max_size here as well.
2662 * Note that caller is responsible for ensuring max_size increases are
2663 * requested from the MDS.
2665 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2666 * or a negative error code. There are 3 speical error codes:
2667 * -EAGAIN: need to sleep but non-blocking is specified
2668 * -EFBIG: ask caller to call check_max_size() and try again.
2669 * -EUCLEAN: ask caller to call ceph_renew_caps() and try again.
2672 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2673 NON_BLOCKING = (1 << 8),
2674 CHECK_FILELOCK = (1 << 9),
2677 static int try_get_cap_refs(struct inode *inode, int need, int want,
2678 loff_t endoff, int flags, int *got)
2680 struct ceph_inode_info *ci = ceph_inode(inode);
2681 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2683 int have, implemented;
2684 bool snap_rwsem_locked = false;
2686 dout("get_cap_refs %p need %s want %s\n", inode,
2687 ceph_cap_string(need), ceph_cap_string(want));
2690 spin_lock(&ci->i_ceph_lock);
2692 if ((flags & CHECK_FILELOCK) &&
2693 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2694 dout("try_get_cap_refs %p error filelock\n", inode);
2699 /* finish pending truncate */
2700 while (ci->i_truncate_pending) {
2701 spin_unlock(&ci->i_ceph_lock);
2702 if (snap_rwsem_locked) {
2703 up_read(&mdsc->snap_rwsem);
2704 snap_rwsem_locked = false;
2706 __ceph_do_pending_vmtruncate(inode);
2707 spin_lock(&ci->i_ceph_lock);
2710 have = __ceph_caps_issued(ci, &implemented);
2712 if (have & need & CEPH_CAP_FILE_WR) {
2713 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2714 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2715 inode, endoff, ci->i_max_size);
2716 if (endoff > ci->i_requested_max_size)
2717 ret = ci->i_auth_cap ? -EFBIG : -EUCLEAN;
2721 * If a sync write is in progress, we must wait, so that we
2722 * can get a final snapshot value for size+mtime.
2724 if (__ceph_have_pending_cap_snap(ci)) {
2725 dout("get_cap_refs %p cap_snap_pending\n", inode);
2730 if ((have & need) == need) {
2732 * Look at (implemented & ~have & not) so that we keep waiting
2733 * on transition from wanted -> needed caps. This is needed
2734 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2735 * going before a prior buffered writeback happens.
2737 int not = want & ~(have & need);
2738 int revoking = implemented & ~have;
2739 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2740 inode, ceph_cap_string(have), ceph_cap_string(not),
2741 ceph_cap_string(revoking));
2742 if ((revoking & not) == 0) {
2743 if (!snap_rwsem_locked &&
2744 !ci->i_head_snapc &&
2745 (need & CEPH_CAP_FILE_WR)) {
2746 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2748 * we can not call down_read() when
2749 * task isn't in TASK_RUNNING state
2751 if (flags & NON_BLOCKING) {
2756 spin_unlock(&ci->i_ceph_lock);
2757 down_read(&mdsc->snap_rwsem);
2758 snap_rwsem_locked = true;
2761 snap_rwsem_locked = true;
2763 if ((have & want) == want)
2767 ceph_take_cap_refs(ci, *got, true);
2771 int session_readonly = false;
2773 if (ci->i_auth_cap &&
2774 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2775 struct ceph_mds_session *s = ci->i_auth_cap->session;
2776 spin_lock(&s->s_cap_lock);
2777 session_readonly = s->s_readonly;
2778 spin_unlock(&s->s_cap_lock);
2780 if (session_readonly) {
2781 dout("get_cap_refs %p need %s but mds%d readonly\n",
2782 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2787 if (ceph_inode_is_shutdown(inode)) {
2788 dout("get_cap_refs %p inode is shutdown\n", inode);
2792 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2793 if (need & ~mds_wanted) {
2794 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2795 inode, ceph_cap_string(need),
2796 ceph_cap_string(mds_wanted));
2801 dout("get_cap_refs %p have %s need %s\n", inode,
2802 ceph_cap_string(have), ceph_cap_string(need));
2806 __ceph_touch_fmode(ci, mdsc, flags);
2808 spin_unlock(&ci->i_ceph_lock);
2809 if (snap_rwsem_locked)
2810 up_read(&mdsc->snap_rwsem);
2813 ceph_update_cap_mis(&mdsc->metric);
2815 ceph_update_cap_hit(&mdsc->metric);
2817 dout("get_cap_refs %p ret %d got %s\n", inode,
2818 ret, ceph_cap_string(*got));
2823 * Check the offset we are writing up to against our current
2824 * max_size. If necessary, tell the MDS we want to write to
2827 static void check_max_size(struct inode *inode, loff_t endoff)
2829 struct ceph_inode_info *ci = ceph_inode(inode);
2832 /* do we need to explicitly request a larger max_size? */
2833 spin_lock(&ci->i_ceph_lock);
2834 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2835 dout("write %p at large endoff %llu, req max_size\n",
2837 ci->i_wanted_max_size = endoff;
2839 /* duplicate ceph_check_caps()'s logic */
2840 if (ci->i_auth_cap &&
2841 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2842 ci->i_wanted_max_size > ci->i_max_size &&
2843 ci->i_wanted_max_size > ci->i_requested_max_size)
2845 spin_unlock(&ci->i_ceph_lock);
2847 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2850 static inline int get_used_fmode(int caps)
2853 if (caps & CEPH_CAP_FILE_RD)
2854 fmode |= CEPH_FILE_MODE_RD;
2855 if (caps & CEPH_CAP_FILE_WR)
2856 fmode |= CEPH_FILE_MODE_WR;
2860 int ceph_try_get_caps(struct inode *inode, int need, int want,
2861 bool nonblock, int *got)
2865 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2866 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2867 CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2868 CEPH_CAP_ANY_DIR_OPS));
2870 ret = ceph_pool_perm_check(inode, need);
2875 flags = get_used_fmode(need | want);
2877 flags |= NON_BLOCKING;
2879 ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2880 /* three special error codes */
2881 if (ret == -EAGAIN || ret == -EFBIG || ret == -EUCLEAN)
2887 * Wait for caps, and take cap references. If we can't get a WR cap
2888 * due to a small max_size, make sure we check_max_size (and possibly
2889 * ask the mds) so we don't get hung up indefinitely.
2891 int ceph_get_caps(struct file *filp, int need, int want, loff_t endoff, int *got)
2893 struct ceph_file_info *fi = filp->private_data;
2894 struct inode *inode = file_inode(filp);
2895 struct ceph_inode_info *ci = ceph_inode(inode);
2896 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2897 int ret, _got, flags;
2899 ret = ceph_pool_perm_check(inode, need);
2903 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2904 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2907 flags = get_used_fmode(need | want);
2910 flags &= CEPH_FILE_MODE_MASK;
2911 if (atomic_read(&fi->num_locks))
2912 flags |= CHECK_FILELOCK;
2914 ret = try_get_cap_refs(inode, need, want, endoff,
2916 WARN_ON_ONCE(ret == -EAGAIN);
2918 struct ceph_mds_client *mdsc = fsc->mdsc;
2920 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2922 cw.ino = ceph_ino(inode);
2923 cw.tgid = current->tgid;
2927 spin_lock(&mdsc->caps_list_lock);
2928 list_add(&cw.list, &mdsc->cap_wait_list);
2929 spin_unlock(&mdsc->caps_list_lock);
2931 /* make sure used fmode not timeout */
2932 ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2933 add_wait_queue(&ci->i_cap_wq, &wait);
2935 flags |= NON_BLOCKING;
2936 while (!(ret = try_get_cap_refs(inode, need, want,
2937 endoff, flags, &_got))) {
2938 if (signal_pending(current)) {
2942 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2945 remove_wait_queue(&ci->i_cap_wq, &wait);
2946 ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
2948 spin_lock(&mdsc->caps_list_lock);
2950 spin_unlock(&mdsc->caps_list_lock);
2956 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2957 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2958 if (ret >= 0 && _got)
2959 ceph_put_cap_refs(ci, _got);
2964 if (ret == -EFBIG || ret == -EUCLEAN) {
2965 int ret2 = ceph_wait_on_async_create(inode);
2969 if (ret == -EFBIG) {
2970 check_max_size(inode, endoff);
2973 if (ret == -EUCLEAN) {
2974 /* session was killed, try renew caps */
2975 ret = ceph_renew_caps(inode, flags);
2982 if (S_ISREG(ci->vfs_inode.i_mode) &&
2983 ci->i_inline_version != CEPH_INLINE_NONE &&
2984 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2985 i_size_read(inode) > 0) {
2987 find_get_page(inode->i_mapping, 0);
2989 bool uptodate = PageUptodate(page);
2996 * drop cap refs first because getattr while
2997 * holding * caps refs can cause deadlock.
2999 ceph_put_cap_refs(ci, _got);
3003 * getattr request will bring inline data into
3006 ret = __ceph_do_getattr(inode, NULL,
3007 CEPH_STAT_CAP_INLINE_DATA,
3020 * Take cap refs. Caller must already know we hold at least one ref
3021 * on the caps in question or we don't know this is safe.
3023 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
3025 spin_lock(&ci->i_ceph_lock);
3026 ceph_take_cap_refs(ci, caps, false);
3027 spin_unlock(&ci->i_ceph_lock);
3032 * drop cap_snap that is not associated with any snapshot.
3033 * we don't need to send FLUSHSNAP message for it.
3035 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
3036 struct ceph_cap_snap *capsnap)
3038 if (!capsnap->need_flush &&
3039 !capsnap->writing && !capsnap->dirty_pages) {
3040 dout("dropping cap_snap %p follows %llu\n",
3041 capsnap, capsnap->follows);
3042 BUG_ON(capsnap->cap_flush.tid > 0);
3043 ceph_put_snap_context(capsnap->context);
3044 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
3045 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3047 list_del(&capsnap->ci_item);
3048 ceph_put_cap_snap(capsnap);
3054 enum put_cap_refs_mode {
3055 PUT_CAP_REFS_SYNC = 0,
3056 PUT_CAP_REFS_NO_CHECK,
3063 * If we released the last ref on any given cap, call ceph_check_caps
3064 * to release (or schedule a release).
3066 * If we are releasing a WR cap (from a sync write), finalize any affected
3067 * cap_snap, and wake up any waiters.
3069 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
3070 enum put_cap_refs_mode mode)
3072 struct inode *inode = &ci->vfs_inode;
3073 int last = 0, put = 0, flushsnaps = 0, wake = 0;
3074 bool check_flushsnaps = false;
3076 spin_lock(&ci->i_ceph_lock);
3077 if (had & CEPH_CAP_PIN)
3079 if (had & CEPH_CAP_FILE_RD)
3080 if (--ci->i_rd_ref == 0)
3082 if (had & CEPH_CAP_FILE_CACHE)
3083 if (--ci->i_rdcache_ref == 0)
3085 if (had & CEPH_CAP_FILE_EXCL)
3086 if (--ci->i_fx_ref == 0)
3088 if (had & CEPH_CAP_FILE_BUFFER) {
3089 if (--ci->i_wb_ref == 0) {
3091 /* put the ref held by ceph_take_cap_refs() */
3093 check_flushsnaps = true;
3095 dout("put_cap_refs %p wb %d -> %d (?)\n",
3096 inode, ci->i_wb_ref+1, ci->i_wb_ref);
3098 if (had & CEPH_CAP_FILE_WR) {
3099 if (--ci->i_wr_ref == 0) {
3101 check_flushsnaps = true;
3102 if (ci->i_wrbuffer_ref_head == 0 &&
3103 ci->i_dirty_caps == 0 &&
3104 ci->i_flushing_caps == 0) {
3105 BUG_ON(!ci->i_head_snapc);
3106 ceph_put_snap_context(ci->i_head_snapc);
3107 ci->i_head_snapc = NULL;
3109 /* see comment in __ceph_remove_cap() */
3110 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3111 ceph_change_snap_realm(inode, NULL);
3114 if (check_flushsnaps && __ceph_have_pending_cap_snap(ci)) {
3115 struct ceph_cap_snap *capsnap =
3116 list_last_entry(&ci->i_cap_snaps,
3117 struct ceph_cap_snap,
3120 capsnap->writing = 0;
3121 if (ceph_try_drop_cap_snap(ci, capsnap))
3122 /* put the ref held by ceph_queue_cap_snap() */
3124 else if (__ceph_finish_cap_snap(ci, capsnap))
3128 spin_unlock(&ci->i_ceph_lock);
3130 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3131 last ? " last" : "", put ? " put" : "");
3134 case PUT_CAP_REFS_SYNC:
3136 ceph_check_caps(ci, 0, NULL);
3137 else if (flushsnaps)
3138 ceph_flush_snaps(ci, NULL);
3140 case PUT_CAP_REFS_ASYNC:
3142 ceph_queue_check_caps(inode);
3143 else if (flushsnaps)
3144 ceph_queue_flush_snaps(inode);
3150 wake_up_all(&ci->i_cap_wq);
3155 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3157 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_SYNC);
3160 void ceph_put_cap_refs_async(struct ceph_inode_info *ci, int had)
3162 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_ASYNC);
3165 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3167 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_NO_CHECK);
3171 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3172 * context. Adjust per-snap dirty page accounting as appropriate.
3173 * Once all dirty data for a cap_snap is flushed, flush snapped file
3174 * metadata back to the MDS. If we dropped the last ref, call
3177 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3178 struct ceph_snap_context *snapc)
3180 struct inode *inode = &ci->vfs_inode;
3181 struct ceph_cap_snap *capsnap = NULL;
3185 bool flush_snaps = false;
3186 bool complete_capsnap = false;
3188 spin_lock(&ci->i_ceph_lock);
3189 ci->i_wrbuffer_ref -= nr;
3190 if (ci->i_wrbuffer_ref == 0) {
3195 if (ci->i_head_snapc == snapc) {
3196 ci->i_wrbuffer_ref_head -= nr;
3197 if (ci->i_wrbuffer_ref_head == 0 &&
3198 ci->i_wr_ref == 0 &&
3199 ci->i_dirty_caps == 0 &&
3200 ci->i_flushing_caps == 0) {
3201 BUG_ON(!ci->i_head_snapc);
3202 ceph_put_snap_context(ci->i_head_snapc);
3203 ci->i_head_snapc = NULL;
3205 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3207 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3208 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3209 last ? " LAST" : "");
3211 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3212 if (capsnap->context == snapc) {
3220 * The capsnap should already be removed when removing
3221 * auth cap in the case of a forced unmount.
3223 WARN_ON_ONCE(ci->i_auth_cap);
3227 capsnap->dirty_pages -= nr;
3228 if (capsnap->dirty_pages == 0) {
3229 complete_capsnap = true;
3230 if (!capsnap->writing) {
3231 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3234 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3239 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3240 " snap %lld %d/%d -> %d/%d %s%s\n",
3241 inode, capsnap, capsnap->context->seq,
3242 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3243 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3244 last ? " (wrbuffer last)" : "",
3245 complete_capsnap ? " (complete capsnap)" : "");
3249 spin_unlock(&ci->i_ceph_lock);
3252 ceph_check_caps(ci, 0, NULL);
3253 } else if (flush_snaps) {
3254 ceph_flush_snaps(ci, NULL);
3256 if (complete_capsnap)
3257 wake_up_all(&ci->i_cap_wq);
3264 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3266 static void invalidate_aliases(struct inode *inode)
3268 struct dentry *dn, *prev = NULL;
3270 dout("invalidate_aliases inode %p\n", inode);
3271 d_prune_aliases(inode);
3273 * For non-directory inode, d_find_alias() only returns
3274 * hashed dentry. After calling d_invalidate(), the
3275 * dentry becomes unhashed.
3277 * For directory inode, d_find_alias() can return
3278 * unhashed dentry. But directory inode should have
3279 * one alias at most.
3281 while ((dn = d_find_alias(inode))) {
3295 struct cap_extra_info {
3296 struct ceph_string *pool_ns;
3306 /* currently issued */
3308 struct timespec64 btime;
3312 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3313 * actually be a revocation if it specifies a smaller cap set.)
3315 * caller holds s_mutex and i_ceph_lock, we drop both.
3317 static void handle_cap_grant(struct inode *inode,
3318 struct ceph_mds_session *session,
3319 struct ceph_cap *cap,
3320 struct ceph_mds_caps *grant,
3321 struct ceph_buffer *xattr_buf,
3322 struct cap_extra_info *extra_info)
3323 __releases(ci->i_ceph_lock)
3324 __releases(session->s_mdsc->snap_rwsem)
3326 struct ceph_inode_info *ci = ceph_inode(inode);
3327 int seq = le32_to_cpu(grant->seq);
3328 int newcaps = le32_to_cpu(grant->caps);
3329 int used, wanted, dirty;
3330 u64 size = le64_to_cpu(grant->size);
3331 u64 max_size = le64_to_cpu(grant->max_size);
3332 unsigned char check_caps = 0;
3333 bool was_stale = cap->cap_gen < atomic_read(&session->s_cap_gen);
3335 bool writeback = false;
3336 bool queue_trunc = false;
3337 bool queue_invalidate = false;
3338 bool deleted_inode = false;
3339 bool fill_inline = false;
3341 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3342 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3343 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3344 i_size_read(inode));
3348 * If CACHE is being revoked, and we have no dirty buffers,
3349 * try to invalidate (once). (If there are dirty buffers, we
3350 * will invalidate _after_ writeback.)
3352 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3353 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3354 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3355 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3356 if (try_nonblocking_invalidate(inode)) {
3357 /* there were locked pages.. invalidate later
3358 in a separate thread. */
3359 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3360 queue_invalidate = true;
3361 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3367 cap->issued = cap->implemented = CEPH_CAP_PIN;
3370 * auth mds of the inode changed. we received the cap export message,
3371 * but still haven't received the cap import message. handle_cap_export
3372 * updated the new auth MDS' cap.
3374 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3375 * that was sent before the cap import message. So don't remove caps.
3377 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3378 WARN_ON(cap != ci->i_auth_cap);
3379 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3381 newcaps |= cap->issued;
3384 /* side effects now are allowed */
3385 cap->cap_gen = atomic_read(&session->s_cap_gen);
3388 __check_cap_issue(ci, cap, newcaps);
3390 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3392 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3393 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3394 umode_t mode = le32_to_cpu(grant->mode);
3396 if (inode_wrong_type(inode, mode))
3397 pr_warn_once("inode type changed! (ino %llx.%llx is 0%o, mds says 0%o)\n",
3398 ceph_vinop(inode), inode->i_mode, mode);
3400 inode->i_mode = mode;
3401 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3402 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3403 ci->i_btime = extra_info->btime;
3404 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3405 from_kuid(&init_user_ns, inode->i_uid),
3406 from_kgid(&init_user_ns, inode->i_gid));
3409 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3410 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3411 set_nlink(inode, le32_to_cpu(grant->nlink));
3412 if (inode->i_nlink == 0)
3413 deleted_inode = true;
3416 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3418 int len = le32_to_cpu(grant->xattr_len);
3419 u64 version = le64_to_cpu(grant->xattr_version);
3421 if (version > ci->i_xattrs.version) {
3422 dout(" got new xattrs v%llu on %p len %d\n",
3423 version, inode, len);
3424 if (ci->i_xattrs.blob)
3425 ceph_buffer_put(ci->i_xattrs.blob);
3426 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3427 ci->i_xattrs.version = version;
3428 ceph_forget_all_cached_acls(inode);
3429 ceph_security_invalidate_secctx(inode);
3433 if (newcaps & CEPH_CAP_ANY_RD) {
3434 struct timespec64 mtime, atime, ctime;
3435 /* ctime/mtime/atime? */
3436 ceph_decode_timespec64(&mtime, &grant->mtime);
3437 ceph_decode_timespec64(&atime, &grant->atime);
3438 ceph_decode_timespec64(&ctime, &grant->ctime);
3439 ceph_fill_file_time(inode, extra_info->issued,
3440 le32_to_cpu(grant->time_warp_seq),
3441 &ctime, &mtime, &atime);
3444 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3445 ci->i_files = extra_info->nfiles;
3446 ci->i_subdirs = extra_info->nsubdirs;
3449 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3450 /* file layout may have changed */
3451 s64 old_pool = ci->i_layout.pool_id;
3452 struct ceph_string *old_ns;
3454 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3455 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3456 lockdep_is_held(&ci->i_ceph_lock));
3457 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3459 if (ci->i_layout.pool_id != old_pool ||
3460 extra_info->pool_ns != old_ns)
3461 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3463 extra_info->pool_ns = old_ns;
3465 /* size/truncate_seq? */
3466 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3467 le32_to_cpu(grant->truncate_seq),
3468 le64_to_cpu(grant->truncate_size),
3472 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3473 if (max_size != ci->i_max_size) {
3474 dout("max_size %lld -> %llu\n",
3475 ci->i_max_size, max_size);
3476 ci->i_max_size = max_size;
3477 if (max_size >= ci->i_wanted_max_size) {
3478 ci->i_wanted_max_size = 0; /* reset */
3479 ci->i_requested_max_size = 0;
3485 /* check cap bits */
3486 wanted = __ceph_caps_wanted(ci);
3487 used = __ceph_caps_used(ci);
3488 dirty = __ceph_caps_dirty(ci);
3489 dout(" my wanted = %s, used = %s, dirty %s\n",
3490 ceph_cap_string(wanted),
3491 ceph_cap_string(used),
3492 ceph_cap_string(dirty));
3494 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3495 (wanted & ~(cap->mds_wanted | newcaps))) {
3497 * If mds is importing cap, prior cap messages that update
3498 * 'wanted' may get dropped by mds (migrate seq mismatch).
3500 * We don't send cap message to update 'wanted' if what we
3501 * want are already issued. If mds revokes caps, cap message
3502 * that releases caps also tells mds what we want. But if
3503 * caps got revoked by mds forcedly (session stale). We may
3504 * haven't told mds what we want.
3509 /* revocation, grant, or no-op? */
3510 if (cap->issued & ~newcaps) {
3511 int revoking = cap->issued & ~newcaps;
3513 dout("revocation: %s -> %s (revoking %s)\n",
3514 ceph_cap_string(cap->issued),
3515 ceph_cap_string(newcaps),
3516 ceph_cap_string(revoking));
3517 if (S_ISREG(inode->i_mode) &&
3518 (revoking & used & CEPH_CAP_FILE_BUFFER))
3519 writeback = true; /* initiate writeback; will delay ack */
3520 else if (queue_invalidate &&
3521 revoking == CEPH_CAP_FILE_CACHE &&
3522 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3523 ; /* do nothing yet, invalidation will be queued */
3524 else if (cap == ci->i_auth_cap)
3525 check_caps = 1; /* check auth cap only */
3527 check_caps = 2; /* check all caps */
3528 cap->issued = newcaps;
3529 cap->implemented |= newcaps;
3530 } else if (cap->issued == newcaps) {
3531 dout("caps unchanged: %s -> %s\n",
3532 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3534 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3535 ceph_cap_string(newcaps));
3536 /* non-auth MDS is revoking the newly grant caps ? */
3537 if (cap == ci->i_auth_cap &&
3538 __ceph_caps_revoking_other(ci, cap, newcaps))
3541 cap->issued = newcaps;
3542 cap->implemented |= newcaps; /* add bits only, to
3543 * avoid stepping on a
3544 * pending revocation */
3547 BUG_ON(cap->issued & ~cap->implemented);
3549 if (extra_info->inline_version > 0 &&
3550 extra_info->inline_version >= ci->i_inline_version) {
3551 ci->i_inline_version = extra_info->inline_version;
3552 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3553 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3557 if (ci->i_auth_cap == cap &&
3558 le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3559 if (newcaps & ~extra_info->issued)
3562 if (ci->i_requested_max_size > max_size ||
3563 !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3564 /* re-request max_size if necessary */
3565 ci->i_requested_max_size = 0;
3569 ceph_kick_flushing_inode_caps(session, ci);
3570 spin_unlock(&ci->i_ceph_lock);
3571 up_read(&session->s_mdsc->snap_rwsem);
3573 spin_unlock(&ci->i_ceph_lock);
3577 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3578 extra_info->inline_len);
3581 ceph_queue_vmtruncate(inode);
3585 * queue inode for writeback: we can't actually call
3586 * filemap_write_and_wait, etc. from message handler
3589 ceph_queue_writeback(inode);
3590 if (queue_invalidate)
3591 ceph_queue_invalidate(inode);
3593 invalidate_aliases(inode);
3595 wake_up_all(&ci->i_cap_wq);
3597 mutex_unlock(&session->s_mutex);
3598 if (check_caps == 1)
3599 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3601 else if (check_caps == 2)
3602 ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3606 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3607 * MDS has been safely committed.
3609 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3610 struct ceph_mds_caps *m,
3611 struct ceph_mds_session *session,
3612 struct ceph_cap *cap)
3613 __releases(ci->i_ceph_lock)
3615 struct ceph_inode_info *ci = ceph_inode(inode);
3616 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3617 struct ceph_cap_flush *cf, *tmp_cf;
3618 LIST_HEAD(to_remove);
3619 unsigned seq = le32_to_cpu(m->seq);
3620 int dirty = le32_to_cpu(m->dirty);
3623 bool wake_ci = false;
3624 bool wake_mdsc = false;
3626 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3627 /* Is this the one that was flushed? */
3628 if (cf->tid == flush_tid)
3631 /* Is this a capsnap? */
3635 if (cf->tid <= flush_tid) {
3637 * An earlier or current tid. The FLUSH_ACK should
3638 * represent a superset of this flush's caps.
3640 wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3641 list_add_tail(&cf->i_list, &to_remove);
3644 * This is a later one. Any caps in it are still dirty
3645 * so don't count them as cleaned.
3647 cleaned &= ~cf->caps;
3653 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3654 " flushing %s -> %s\n",
3655 inode, session->s_mds, seq, ceph_cap_string(dirty),
3656 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3657 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3659 if (list_empty(&to_remove) && !cleaned)
3662 ci->i_flushing_caps &= ~cleaned;
3664 spin_lock(&mdsc->cap_dirty_lock);
3666 list_for_each_entry(cf, &to_remove, i_list)
3667 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3669 if (ci->i_flushing_caps == 0) {
3670 if (list_empty(&ci->i_cap_flush_list)) {
3671 list_del_init(&ci->i_flushing_item);
3672 if (!list_empty(&session->s_cap_flushing)) {
3673 dout(" mds%d still flushing cap on %p\n",
3675 &list_first_entry(&session->s_cap_flushing,
3676 struct ceph_inode_info,
3677 i_flushing_item)->vfs_inode);
3680 mdsc->num_cap_flushing--;
3681 dout(" inode %p now !flushing\n", inode);
3683 if (ci->i_dirty_caps == 0) {
3684 dout(" inode %p now clean\n", inode);
3685 BUG_ON(!list_empty(&ci->i_dirty_item));
3687 if (ci->i_wr_ref == 0 &&
3688 ci->i_wrbuffer_ref_head == 0) {
3689 BUG_ON(!ci->i_head_snapc);
3690 ceph_put_snap_context(ci->i_head_snapc);
3691 ci->i_head_snapc = NULL;
3694 BUG_ON(list_empty(&ci->i_dirty_item));
3697 spin_unlock(&mdsc->cap_dirty_lock);
3700 spin_unlock(&ci->i_ceph_lock);
3702 while (!list_empty(&to_remove)) {
3703 cf = list_first_entry(&to_remove,
3704 struct ceph_cap_flush, i_list);
3705 list_del_init(&cf->i_list);
3706 if (!cf->is_capsnap)
3707 ceph_free_cap_flush(cf);
3711 wake_up_all(&ci->i_cap_wq);
3713 wake_up_all(&mdsc->cap_flushing_wq);
3718 void __ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3719 bool *wake_ci, bool *wake_mdsc)
3721 struct ceph_inode_info *ci = ceph_inode(inode);
3722 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3725 lockdep_assert_held(&ci->i_ceph_lock);
3727 dout("removing capsnap %p, inode %p ci %p\n", capsnap, inode, ci);
3729 list_del_init(&capsnap->ci_item);
3730 ret = __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3734 spin_lock(&mdsc->cap_dirty_lock);
3735 if (list_empty(&ci->i_cap_flush_list))
3736 list_del_init(&ci->i_flushing_item);
3738 ret = __detach_cap_flush_from_mdsc(mdsc, &capsnap->cap_flush);
3741 spin_unlock(&mdsc->cap_dirty_lock);
3744 void ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3745 bool *wake_ci, bool *wake_mdsc)
3747 struct ceph_inode_info *ci = ceph_inode(inode);
3749 lockdep_assert_held(&ci->i_ceph_lock);
3751 WARN_ON_ONCE(capsnap->dirty_pages || capsnap->writing);
3752 __ceph_remove_capsnap(inode, capsnap, wake_ci, wake_mdsc);
3756 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3757 * throw away our cap_snap.
3759 * Caller hold s_mutex.
3761 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3762 struct ceph_mds_caps *m,
3763 struct ceph_mds_session *session)
3765 struct ceph_inode_info *ci = ceph_inode(inode);
3766 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3767 u64 follows = le64_to_cpu(m->snap_follows);
3768 struct ceph_cap_snap *capsnap;
3769 bool flushed = false;
3770 bool wake_ci = false;
3771 bool wake_mdsc = false;
3773 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3774 inode, ci, session->s_mds, follows);
3776 spin_lock(&ci->i_ceph_lock);
3777 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3778 if (capsnap->follows == follows) {
3779 if (capsnap->cap_flush.tid != flush_tid) {
3780 dout(" cap_snap %p follows %lld tid %lld !="
3781 " %lld\n", capsnap, follows,
3782 flush_tid, capsnap->cap_flush.tid);
3788 dout(" skipping cap_snap %p follows %lld\n",
3789 capsnap, capsnap->follows);
3793 ceph_remove_capsnap(inode, capsnap, &wake_ci, &wake_mdsc);
3794 spin_unlock(&ci->i_ceph_lock);
3797 ceph_put_snap_context(capsnap->context);
3798 ceph_put_cap_snap(capsnap);
3800 wake_up_all(&ci->i_cap_wq);
3802 wake_up_all(&mdsc->cap_flushing_wq);
3808 * Handle TRUNC from MDS, indicating file truncation.
3810 * caller hold s_mutex.
3812 static bool handle_cap_trunc(struct inode *inode,
3813 struct ceph_mds_caps *trunc,
3814 struct ceph_mds_session *session)
3816 struct ceph_inode_info *ci = ceph_inode(inode);
3817 int mds = session->s_mds;
3818 int seq = le32_to_cpu(trunc->seq);
3819 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3820 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3821 u64 size = le64_to_cpu(trunc->size);
3822 int implemented = 0;
3823 int dirty = __ceph_caps_dirty(ci);
3824 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3825 bool queue_trunc = false;
3827 lockdep_assert_held(&ci->i_ceph_lock);
3829 issued |= implemented | dirty;
3831 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3832 inode, mds, seq, truncate_size, truncate_seq);
3833 queue_trunc = ceph_fill_file_size(inode, issued,
3834 truncate_seq, truncate_size, size);
3839 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3840 * different one. If we are the most recent migration we've seen (as
3841 * indicated by mseq), make note of the migrating cap bits for the
3842 * duration (until we see the corresponding IMPORT).
3844 * caller holds s_mutex
3846 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3847 struct ceph_mds_cap_peer *ph,
3848 struct ceph_mds_session *session)
3850 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3851 struct ceph_mds_session *tsession = NULL;
3852 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3853 struct ceph_inode_info *ci = ceph_inode(inode);
3855 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3856 unsigned t_seq, t_mseq;
3858 int mds = session->s_mds;
3861 t_cap_id = le64_to_cpu(ph->cap_id);
3862 t_seq = le32_to_cpu(ph->seq);
3863 t_mseq = le32_to_cpu(ph->mseq);
3864 target = le32_to_cpu(ph->mds);
3866 t_cap_id = t_seq = t_mseq = 0;
3870 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3871 inode, ci, mds, mseq, target);
3873 spin_lock(&ci->i_ceph_lock);
3874 cap = __get_cap_for_mds(ci, mds);
3875 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3879 ceph_remove_cap(cap, false);
3884 * now we know we haven't received the cap import message yet
3885 * because the exported cap still exist.
3888 issued = cap->issued;
3889 if (issued != cap->implemented)
3890 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3891 "ino (%llx.%llx) mds%d seq %d mseq %d "
3892 "issued %s implemented %s\n",
3893 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3894 ceph_cap_string(issued),
3895 ceph_cap_string(cap->implemented));
3898 tcap = __get_cap_for_mds(ci, target);
3900 /* already have caps from the target */
3901 if (tcap->cap_id == t_cap_id &&
3902 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3903 dout(" updating import cap %p mds%d\n", tcap, target);
3904 tcap->cap_id = t_cap_id;
3905 tcap->seq = t_seq - 1;
3906 tcap->issue_seq = t_seq - 1;
3907 tcap->issued |= issued;
3908 tcap->implemented |= issued;
3909 if (cap == ci->i_auth_cap) {
3910 ci->i_auth_cap = tcap;
3911 change_auth_cap_ses(ci, tcap->session);
3914 ceph_remove_cap(cap, false);
3916 } else if (tsession) {
3917 /* add placeholder for the export tagert */
3918 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3920 ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3921 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3923 if (!list_empty(&ci->i_cap_flush_list) &&
3924 ci->i_auth_cap == tcap) {
3925 spin_lock(&mdsc->cap_dirty_lock);
3926 list_move_tail(&ci->i_flushing_item,
3927 &tcap->session->s_cap_flushing);
3928 spin_unlock(&mdsc->cap_dirty_lock);
3931 ceph_remove_cap(cap, false);
3935 spin_unlock(&ci->i_ceph_lock);
3936 mutex_unlock(&session->s_mutex);
3938 /* open target session */
3939 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3940 if (!IS_ERR(tsession)) {
3942 mutex_lock(&session->s_mutex);
3943 mutex_lock_nested(&tsession->s_mutex,
3944 SINGLE_DEPTH_NESTING);
3946 mutex_lock(&tsession->s_mutex);
3947 mutex_lock_nested(&session->s_mutex,
3948 SINGLE_DEPTH_NESTING);
3950 new_cap = ceph_get_cap(mdsc, NULL);
3955 mutex_lock(&session->s_mutex);
3960 spin_unlock(&ci->i_ceph_lock);
3961 mutex_unlock(&session->s_mutex);
3963 mutex_unlock(&tsession->s_mutex);
3964 ceph_put_mds_session(tsession);
3967 ceph_put_cap(mdsc, new_cap);
3971 * Handle cap IMPORT.
3973 * caller holds s_mutex. acquires i_ceph_lock
3975 static void handle_cap_import(struct ceph_mds_client *mdsc,
3976 struct inode *inode, struct ceph_mds_caps *im,
3977 struct ceph_mds_cap_peer *ph,
3978 struct ceph_mds_session *session,
3979 struct ceph_cap **target_cap, int *old_issued)
3981 struct ceph_inode_info *ci = ceph_inode(inode);
3982 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3983 int mds = session->s_mds;
3985 unsigned caps = le32_to_cpu(im->caps);
3986 unsigned wanted = le32_to_cpu(im->wanted);
3987 unsigned seq = le32_to_cpu(im->seq);
3988 unsigned mseq = le32_to_cpu(im->migrate_seq);
3989 u64 realmino = le64_to_cpu(im->realm);
3990 u64 cap_id = le64_to_cpu(im->cap_id);
3995 p_cap_id = le64_to_cpu(ph->cap_id);
3996 peer = le32_to_cpu(ph->mds);
4002 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
4003 inode, ci, mds, mseq, peer);
4005 cap = __get_cap_for_mds(ci, mds);
4008 spin_unlock(&ci->i_ceph_lock);
4009 new_cap = ceph_get_cap(mdsc, NULL);
4010 spin_lock(&ci->i_ceph_lock);
4016 ceph_put_cap(mdsc, new_cap);
4021 __ceph_caps_issued(ci, &issued);
4022 issued |= __ceph_caps_dirty(ci);
4024 ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
4025 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
4027 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
4028 if (ocap && ocap->cap_id == p_cap_id) {
4029 dout(" remove export cap %p mds%d flags %d\n",
4030 ocap, peer, ph->flags);
4031 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
4032 (ocap->seq != le32_to_cpu(ph->seq) ||
4033 ocap->mseq != le32_to_cpu(ph->mseq))) {
4034 pr_err_ratelimited("handle_cap_import: "
4035 "mismatched seq/mseq: ino (%llx.%llx) "
4036 "mds%d seq %d mseq %d importer mds%d "
4037 "has peer seq %d mseq %d\n",
4038 ceph_vinop(inode), peer, ocap->seq,
4039 ocap->mseq, mds, le32_to_cpu(ph->seq),
4040 le32_to_cpu(ph->mseq));
4042 ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
4045 *old_issued = issued;
4050 * Handle a caps message from the MDS.
4052 * Identify the appropriate session, inode, and call the right handler
4053 * based on the cap op.
4055 void ceph_handle_caps(struct ceph_mds_session *session,
4056 struct ceph_msg *msg)
4058 struct ceph_mds_client *mdsc = session->s_mdsc;
4059 struct inode *inode;
4060 struct ceph_inode_info *ci;
4061 struct ceph_cap *cap;
4062 struct ceph_mds_caps *h;
4063 struct ceph_mds_cap_peer *peer = NULL;
4064 struct ceph_snap_realm *realm = NULL;
4066 int msg_version = le16_to_cpu(msg->hdr.version);
4068 struct ceph_vino vino;
4070 size_t snaptrace_len;
4072 struct cap_extra_info extra_info = {};
4075 dout("handle_caps from mds%d\n", session->s_mds);
4078 end = msg->front.iov_base + msg->front.iov_len;
4079 if (msg->front.iov_len < sizeof(*h))
4081 h = msg->front.iov_base;
4082 op = le32_to_cpu(h->op);
4083 vino.ino = le64_to_cpu(h->ino);
4084 vino.snap = CEPH_NOSNAP;
4085 seq = le32_to_cpu(h->seq);
4086 mseq = le32_to_cpu(h->migrate_seq);
4089 snaptrace_len = le32_to_cpu(h->snap_trace_len);
4090 p = snaptrace + snaptrace_len;
4092 if (msg_version >= 2) {
4094 ceph_decode_32_safe(&p, end, flock_len, bad);
4095 if (p + flock_len > end)
4100 if (msg_version >= 3) {
4101 if (op == CEPH_CAP_OP_IMPORT) {
4102 if (p + sizeof(*peer) > end)
4106 } else if (op == CEPH_CAP_OP_EXPORT) {
4107 /* recorded in unused fields */
4108 peer = (void *)&h->size;
4112 if (msg_version >= 4) {
4113 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
4114 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
4115 if (p + extra_info.inline_len > end)
4117 extra_info.inline_data = p;
4118 p += extra_info.inline_len;
4121 if (msg_version >= 5) {
4122 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
4125 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
4126 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
4129 if (msg_version >= 8) {
4133 ceph_decode_skip_64(&p, end, bad); // flush_tid
4135 ceph_decode_skip_32(&p, end, bad); // caller_uid
4136 ceph_decode_skip_32(&p, end, bad); // caller_gid
4138 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4139 if (pool_ns_len > 0) {
4140 ceph_decode_need(&p, end, pool_ns_len, bad);
4141 extra_info.pool_ns =
4142 ceph_find_or_create_string(p, pool_ns_len);
4147 if (msg_version >= 9) {
4148 struct ceph_timespec *btime;
4150 if (p + sizeof(*btime) > end)
4153 ceph_decode_timespec64(&extra_info.btime, btime);
4154 p += sizeof(*btime);
4155 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4158 if (msg_version >= 11) {
4160 ceph_decode_skip_32(&p, end, bad); // flags
4162 extra_info.dirstat_valid = true;
4163 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4164 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4168 inode = ceph_find_inode(mdsc->fsc->sb, vino);
4169 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4172 mutex_lock(&session->s_mutex);
4173 inc_session_sequence(session);
4174 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4178 dout(" i don't have ino %llx\n", vino.ino);
4180 if (op == CEPH_CAP_OP_IMPORT) {
4181 cap = ceph_get_cap(mdsc, NULL);
4182 cap->cap_ino = vino.ino;
4183 cap->queue_release = 1;
4184 cap->cap_id = le64_to_cpu(h->cap_id);
4187 cap->issue_seq = seq;
4188 spin_lock(&session->s_cap_lock);
4189 __ceph_queue_cap_release(session, cap);
4190 spin_unlock(&session->s_cap_lock);
4192 goto flush_cap_releases;
4194 ci = ceph_inode(inode);
4196 /* these will work even if we don't have a cap yet */
4198 case CEPH_CAP_OP_FLUSHSNAP_ACK:
4199 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4203 case CEPH_CAP_OP_EXPORT:
4204 handle_cap_export(inode, h, peer, session);
4207 case CEPH_CAP_OP_IMPORT:
4209 if (snaptrace_len) {
4210 down_write(&mdsc->snap_rwsem);
4211 ceph_update_snap_trace(mdsc, snaptrace,
4212 snaptrace + snaptrace_len,
4214 downgrade_write(&mdsc->snap_rwsem);
4216 down_read(&mdsc->snap_rwsem);
4218 spin_lock(&ci->i_ceph_lock);
4219 handle_cap_import(mdsc, inode, h, peer, session,
4220 &cap, &extra_info.issued);
4221 handle_cap_grant(inode, session, cap,
4222 h, msg->middle, &extra_info);
4224 ceph_put_snap_realm(mdsc, realm);
4228 /* the rest require a cap */
4229 spin_lock(&ci->i_ceph_lock);
4230 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4232 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4233 inode, ceph_ino(inode), ceph_snap(inode),
4235 spin_unlock(&ci->i_ceph_lock);
4236 goto flush_cap_releases;
4239 /* note that each of these drops i_ceph_lock for us */
4241 case CEPH_CAP_OP_REVOKE:
4242 case CEPH_CAP_OP_GRANT:
4243 __ceph_caps_issued(ci, &extra_info.issued);
4244 extra_info.issued |= __ceph_caps_dirty(ci);
4245 handle_cap_grant(inode, session, cap,
4246 h, msg->middle, &extra_info);
4249 case CEPH_CAP_OP_FLUSH_ACK:
4250 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4254 case CEPH_CAP_OP_TRUNC:
4255 queue_trunc = handle_cap_trunc(inode, h, session);
4256 spin_unlock(&ci->i_ceph_lock);
4258 ceph_queue_vmtruncate(inode);
4262 spin_unlock(&ci->i_ceph_lock);
4263 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4264 ceph_cap_op_name(op));
4268 mutex_unlock(&session->s_mutex);
4272 ceph_put_string(extra_info.pool_ns);
4277 * send any cap release message to try to move things
4278 * along for the mds (who clearly thinks we still have this
4281 ceph_flush_cap_releases(mdsc, session);
4285 pr_err("ceph_handle_caps: corrupt message\n");
4291 * Delayed work handler to process end of delayed cap release LRU list.
4293 * If new caps are added to the list while processing it, these won't get
4294 * processed in this run. In this case, the ci->i_hold_caps_max will be
4295 * returned so that the work can be scheduled accordingly.
4297 unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4299 struct inode *inode;
4300 struct ceph_inode_info *ci;
4301 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
4302 unsigned long delay_max = opt->caps_wanted_delay_max * HZ;
4303 unsigned long loop_start = jiffies;
4304 unsigned long delay = 0;
4306 dout("check_delayed_caps\n");
4307 spin_lock(&mdsc->cap_delay_lock);
4308 while (!list_empty(&mdsc->cap_delay_list)) {
4309 ci = list_first_entry(&mdsc->cap_delay_list,
4310 struct ceph_inode_info,
4312 if (time_before(loop_start, ci->i_hold_caps_max - delay_max)) {
4313 dout("%s caps added recently. Exiting loop", __func__);
4314 delay = ci->i_hold_caps_max;
4317 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4318 time_before(jiffies, ci->i_hold_caps_max))
4320 list_del_init(&ci->i_cap_delay_list);
4322 inode = igrab(&ci->vfs_inode);
4324 spin_unlock(&mdsc->cap_delay_lock);
4325 dout("check_delayed_caps on %p\n", inode);
4326 ceph_check_caps(ci, 0, NULL);
4328 spin_lock(&mdsc->cap_delay_lock);
4331 spin_unlock(&mdsc->cap_delay_lock);
4337 * Flush all dirty caps to the mds
4339 static void flush_dirty_session_caps(struct ceph_mds_session *s)
4341 struct ceph_mds_client *mdsc = s->s_mdsc;
4342 struct ceph_inode_info *ci;
4343 struct inode *inode;
4345 dout("flush_dirty_caps\n");
4346 spin_lock(&mdsc->cap_dirty_lock);
4347 while (!list_empty(&s->s_cap_dirty)) {
4348 ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4350 inode = &ci->vfs_inode;
4352 dout("flush_dirty_caps %llx.%llx\n", ceph_vinop(inode));
4353 spin_unlock(&mdsc->cap_dirty_lock);
4354 ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4356 spin_lock(&mdsc->cap_dirty_lock);
4358 spin_unlock(&mdsc->cap_dirty_lock);
4359 dout("flush_dirty_caps done\n");
4362 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4364 ceph_mdsc_iterate_sessions(mdsc, flush_dirty_session_caps, true);
4367 void __ceph_touch_fmode(struct ceph_inode_info *ci,
4368 struct ceph_mds_client *mdsc, int fmode)
4370 unsigned long now = jiffies;
4371 if (fmode & CEPH_FILE_MODE_RD)
4372 ci->i_last_rd = now;
4373 if (fmode & CEPH_FILE_MODE_WR)
4374 ci->i_last_wr = now;
4375 /* queue periodic check */
4377 __ceph_is_any_real_caps(ci) &&
4378 list_empty(&ci->i_cap_delay_list))
4379 __cap_delay_requeue(mdsc, ci);
4382 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4384 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4385 int bits = (fmode << 1) | 1;
4386 bool already_opened = false;
4390 atomic64_inc(&mdsc->metric.opened_files);
4392 spin_lock(&ci->i_ceph_lock);
4393 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4395 * If any of the mode ref is larger than 0,
4396 * that means it has been already opened by
4397 * others. Just skip checking the PIN ref.
4399 if (i && ci->i_nr_by_mode[i])
4400 already_opened = true;
4402 if (bits & (1 << i))
4403 ci->i_nr_by_mode[i] += count;
4406 if (!already_opened)
4407 percpu_counter_inc(&mdsc->metric.opened_inodes);
4408 spin_unlock(&ci->i_ceph_lock);
4412 * Drop open file reference. If we were the last open file,
4413 * we may need to release capabilities to the MDS (or schedule
4414 * their delayed release).
4416 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4418 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4419 int bits = (fmode << 1) | 1;
4420 bool is_closed = true;
4424 atomic64_dec(&mdsc->metric.opened_files);
4426 spin_lock(&ci->i_ceph_lock);
4427 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4428 if (bits & (1 << i)) {
4429 BUG_ON(ci->i_nr_by_mode[i] < count);
4430 ci->i_nr_by_mode[i] -= count;
4434 * If any of the mode ref is not 0 after
4435 * decreased, that means it is still opened
4436 * by others. Just skip checking the PIN ref.
4438 if (i && ci->i_nr_by_mode[i])
4443 percpu_counter_dec(&mdsc->metric.opened_inodes);
4444 spin_unlock(&ci->i_ceph_lock);
4448 * For a soon-to-be unlinked file, drop the LINK caps. If it
4449 * looks like the link count will hit 0, drop any other caps (other
4450 * than PIN) we don't specifically want (due to the file still being
4453 int ceph_drop_caps_for_unlink(struct inode *inode)
4455 struct ceph_inode_info *ci = ceph_inode(inode);
4456 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4458 spin_lock(&ci->i_ceph_lock);
4459 if (inode->i_nlink == 1) {
4460 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4462 if (__ceph_caps_dirty(ci)) {
4463 struct ceph_mds_client *mdsc =
4464 ceph_inode_to_client(inode)->mdsc;
4465 __cap_delay_requeue_front(mdsc, ci);
4468 spin_unlock(&ci->i_ceph_lock);
4473 * Helpers for embedding cap and dentry lease releases into mds
4476 * @force is used by dentry_release (below) to force inclusion of a
4477 * record for the directory inode, even when there aren't any caps to
4480 int ceph_encode_inode_release(void **p, struct inode *inode,
4481 int mds, int drop, int unless, int force)
4483 struct ceph_inode_info *ci = ceph_inode(inode);
4484 struct ceph_cap *cap;
4485 struct ceph_mds_request_release *rel = *p;
4489 spin_lock(&ci->i_ceph_lock);
4490 used = __ceph_caps_used(ci);
4491 dirty = __ceph_caps_dirty(ci);
4493 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4494 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4495 ceph_cap_string(unless));
4497 /* only drop unused, clean caps */
4498 drop &= ~(used | dirty);
4500 cap = __get_cap_for_mds(ci, mds);
4501 if (cap && __cap_is_valid(cap)) {
4502 unless &= cap->issued;
4504 if (unless & CEPH_CAP_AUTH_EXCL)
4505 drop &= ~CEPH_CAP_AUTH_SHARED;
4506 if (unless & CEPH_CAP_LINK_EXCL)
4507 drop &= ~CEPH_CAP_LINK_SHARED;
4508 if (unless & CEPH_CAP_XATTR_EXCL)
4509 drop &= ~CEPH_CAP_XATTR_SHARED;
4510 if (unless & CEPH_CAP_FILE_EXCL)
4511 drop &= ~CEPH_CAP_FILE_SHARED;
4514 if (force || (cap->issued & drop)) {
4515 if (cap->issued & drop) {
4516 int wanted = __ceph_caps_wanted(ci);
4517 dout("encode_inode_release %p cap %p "
4518 "%s -> %s, wanted %s -> %s\n", inode, cap,
4519 ceph_cap_string(cap->issued),
4520 ceph_cap_string(cap->issued & ~drop),
4521 ceph_cap_string(cap->mds_wanted),
4522 ceph_cap_string(wanted));
4524 cap->issued &= ~drop;
4525 cap->implemented &= ~drop;
4526 cap->mds_wanted = wanted;
4527 if (cap == ci->i_auth_cap &&
4528 !(wanted & CEPH_CAP_ANY_FILE_WR))
4529 ci->i_requested_max_size = 0;
4531 dout("encode_inode_release %p cap %p %s"
4532 " (force)\n", inode, cap,
4533 ceph_cap_string(cap->issued));
4536 rel->ino = cpu_to_le64(ceph_ino(inode));
4537 rel->cap_id = cpu_to_le64(cap->cap_id);
4538 rel->seq = cpu_to_le32(cap->seq);
4539 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4540 rel->mseq = cpu_to_le32(cap->mseq);
4541 rel->caps = cpu_to_le32(cap->implemented);
4542 rel->wanted = cpu_to_le32(cap->mds_wanted);
4548 dout("encode_inode_release %p cap %p %s (noop)\n",
4549 inode, cap, ceph_cap_string(cap->issued));
4552 spin_unlock(&ci->i_ceph_lock);
4556 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4558 int mds, int drop, int unless)
4560 struct dentry *parent = NULL;
4561 struct ceph_mds_request_release *rel = *p;
4562 struct ceph_dentry_info *di = ceph_dentry(dentry);
4567 * force an record for the directory caps if we have a dentry lease.
4568 * this is racy (can't take i_ceph_lock and d_lock together), but it
4569 * doesn't have to be perfect; the mds will revoke anything we don't
4572 spin_lock(&dentry->d_lock);
4573 if (di->lease_session && di->lease_session->s_mds == mds)
4576 parent = dget(dentry->d_parent);
4577 dir = d_inode(parent);
4579 spin_unlock(&dentry->d_lock);
4581 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4584 spin_lock(&dentry->d_lock);
4585 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4586 dout("encode_dentry_release %p mds%d seq %d\n",
4587 dentry, mds, (int)di->lease_seq);
4588 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4589 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4590 *p += dentry->d_name.len;
4591 rel->dname_seq = cpu_to_le32(di->lease_seq);
4592 __ceph_mdsc_drop_dentry_lease(dentry);
4594 spin_unlock(&dentry->d_lock);
4598 static int remove_capsnaps(struct ceph_mds_client *mdsc, struct inode *inode)
4600 struct ceph_inode_info *ci = ceph_inode(inode);
4601 struct ceph_cap_snap *capsnap;
4602 int capsnap_release = 0;
4604 lockdep_assert_held(&ci->i_ceph_lock);
4606 dout("removing capsnaps, ci is %p, inode is %p\n", ci, inode);
4608 while (!list_empty(&ci->i_cap_snaps)) {
4609 capsnap = list_first_entry(&ci->i_cap_snaps,
4610 struct ceph_cap_snap, ci_item);
4611 __ceph_remove_capsnap(inode, capsnap, NULL, NULL);
4612 ceph_put_snap_context(capsnap->context);
4613 ceph_put_cap_snap(capsnap);
4616 wake_up_all(&ci->i_cap_wq);
4617 wake_up_all(&mdsc->cap_flushing_wq);
4618 return capsnap_release;
4621 int ceph_purge_inode_cap(struct inode *inode, struct ceph_cap *cap, bool *invalidate)
4623 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
4624 struct ceph_mds_client *mdsc = fsc->mdsc;
4625 struct ceph_inode_info *ci = ceph_inode(inode);
4627 bool dirty_dropped = false;
4630 lockdep_assert_held(&ci->i_ceph_lock);
4632 dout("removing cap %p, ci is %p, inode is %p\n",
4633 cap, ci, &ci->vfs_inode);
4635 is_auth = (cap == ci->i_auth_cap);
4636 __ceph_remove_cap(cap, false);
4638 struct ceph_cap_flush *cf;
4640 if (ceph_inode_is_shutdown(inode)) {
4641 if (inode->i_data.nrpages > 0)
4643 if (ci->i_wrbuffer_ref > 0)
4644 mapping_set_error(&inode->i_data, -EIO);
4647 spin_lock(&mdsc->cap_dirty_lock);
4649 /* trash all of the cap flushes for this inode */
4650 while (!list_empty(&ci->i_cap_flush_list)) {
4651 cf = list_first_entry(&ci->i_cap_flush_list,
4652 struct ceph_cap_flush, i_list);
4653 list_del_init(&cf->g_list);
4654 list_del_init(&cf->i_list);
4655 if (!cf->is_capsnap)
4656 ceph_free_cap_flush(cf);
4659 if (!list_empty(&ci->i_dirty_item)) {
4660 pr_warn_ratelimited(
4661 " dropping dirty %s state for %p %lld\n",
4662 ceph_cap_string(ci->i_dirty_caps),
4663 inode, ceph_ino(inode));
4664 ci->i_dirty_caps = 0;
4665 list_del_init(&ci->i_dirty_item);
4666 dirty_dropped = true;
4668 if (!list_empty(&ci->i_flushing_item)) {
4669 pr_warn_ratelimited(
4670 " dropping dirty+flushing %s state for %p %lld\n",
4671 ceph_cap_string(ci->i_flushing_caps),
4672 inode, ceph_ino(inode));
4673 ci->i_flushing_caps = 0;
4674 list_del_init(&ci->i_flushing_item);
4675 mdsc->num_cap_flushing--;
4676 dirty_dropped = true;
4678 spin_unlock(&mdsc->cap_dirty_lock);
4680 if (dirty_dropped) {
4681 mapping_set_error(inode->i_mapping, -EIO);
4683 if (ci->i_wrbuffer_ref_head == 0 &&
4684 ci->i_wr_ref == 0 &&
4685 ci->i_dirty_caps == 0 &&
4686 ci->i_flushing_caps == 0) {
4687 ceph_put_snap_context(ci->i_head_snapc);
4688 ci->i_head_snapc = NULL;
4692 if (atomic_read(&ci->i_filelock_ref) > 0) {
4693 /* make further file lock syscall return -EIO */
4694 ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
4695 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
4696 inode, ceph_ino(inode));
4699 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
4700 cf = ci->i_prealloc_cap_flush;
4701 ci->i_prealloc_cap_flush = NULL;
4702 if (!cf->is_capsnap)
4703 ceph_free_cap_flush(cf);
4706 if (!list_empty(&ci->i_cap_snaps))
4707 iputs = remove_capsnaps(mdsc, inode);