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 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1119 mdsc = ceph_inode_to_client(&ci->vfs_inode)->mdsc;
1121 /* remove from inode's cap rbtree, and clear auth cap */
1122 rb_erase(&cap->ci_node, &ci->i_caps);
1123 if (ci->i_auth_cap == cap) {
1124 WARN_ON_ONCE(!list_empty(&ci->i_dirty_item) &&
1125 !mdsc->fsc->blocklisted);
1126 ci->i_auth_cap = NULL;
1129 /* remove from session list */
1130 spin_lock(&session->s_cap_lock);
1131 if (session->s_cap_iterator == cap) {
1132 /* not yet, we are iterating over this very cap */
1133 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1136 list_del_init(&cap->session_caps);
1137 session->s_nr_caps--;
1138 atomic64_dec(&mdsc->metric.total_caps);
1139 cap->session = NULL;
1142 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1146 * s_cap_reconnect is protected by s_cap_lock. no one changes
1147 * s_cap_gen while session is in the reconnect state.
1149 if (queue_release &&
1150 (!session->s_cap_reconnect ||
1151 cap->cap_gen == atomic_read(&session->s_cap_gen))) {
1152 cap->queue_release = 1;
1154 __ceph_queue_cap_release(session, cap);
1158 cap->queue_release = 0;
1160 cap->cap_ino = ci->i_vino.ino;
1162 spin_unlock(&session->s_cap_lock);
1165 ceph_put_cap(mdsc, cap);
1167 if (!__ceph_is_any_real_caps(ci)) {
1168 /* when reconnect denied, we remove session caps forcibly,
1169 * i_wr_ref can be non-zero. If there are ongoing write,
1170 * keep i_snap_realm.
1172 if (ci->i_wr_ref == 0 && ci->i_snap_realm)
1173 ceph_change_snap_realm(&ci->vfs_inode, NULL);
1175 __cap_delay_cancel(mdsc, ci);
1179 struct cap_msg_args {
1180 struct ceph_mds_session *session;
1181 u64 ino, cid, follows;
1182 u64 flush_tid, oldest_flush_tid, size, max_size;
1185 struct ceph_buffer *xattr_buf;
1186 struct ceph_buffer *old_xattr_buf;
1187 struct timespec64 atime, mtime, ctime, btime;
1188 int op, caps, wanted, dirty;
1189 u32 seq, issue_seq, mseq, time_warp_seq;
1199 * cap struct size + flock buffer size + inline version + inline data size +
1200 * osd_epoch_barrier + oldest_flush_tid
1202 #define CAP_MSG_SIZE (sizeof(struct ceph_mds_caps) + \
1203 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4)
1205 /* Marshal up the cap msg to the MDS */
1206 static void encode_cap_msg(struct ceph_msg *msg, struct cap_msg_args *arg)
1208 struct ceph_mds_caps *fc;
1210 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1212 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",
1213 __func__, ceph_cap_op_name(arg->op), arg->cid, arg->ino,
1214 ceph_cap_string(arg->caps), ceph_cap_string(arg->wanted),
1215 ceph_cap_string(arg->dirty), arg->seq, arg->issue_seq,
1216 arg->flush_tid, arg->oldest_flush_tid, arg->mseq, arg->follows,
1217 arg->size, arg->max_size, arg->xattr_version,
1218 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1220 msg->hdr.version = cpu_to_le16(10);
1221 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1223 fc = msg->front.iov_base;
1224 memset(fc, 0, sizeof(*fc));
1226 fc->cap_id = cpu_to_le64(arg->cid);
1227 fc->op = cpu_to_le32(arg->op);
1228 fc->seq = cpu_to_le32(arg->seq);
1229 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1230 fc->migrate_seq = cpu_to_le32(arg->mseq);
1231 fc->caps = cpu_to_le32(arg->caps);
1232 fc->wanted = cpu_to_le32(arg->wanted);
1233 fc->dirty = cpu_to_le32(arg->dirty);
1234 fc->ino = cpu_to_le64(arg->ino);
1235 fc->snap_follows = cpu_to_le64(arg->follows);
1237 fc->size = cpu_to_le64(arg->size);
1238 fc->max_size = cpu_to_le64(arg->max_size);
1239 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1240 ceph_encode_timespec64(&fc->atime, &arg->atime);
1241 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1242 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1244 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1245 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1246 fc->mode = cpu_to_le32(arg->mode);
1248 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1249 if (arg->xattr_buf) {
1250 msg->middle = ceph_buffer_get(arg->xattr_buf);
1251 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1252 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1256 /* flock buffer size (version 2) */
1257 ceph_encode_32(&p, 0);
1258 /* inline version (version 4) */
1259 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1260 /* inline data size */
1261 ceph_encode_32(&p, 0);
1263 * osd_epoch_barrier (version 5)
1264 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1265 * case it was recently changed
1267 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1268 /* oldest_flush_tid (version 6) */
1269 ceph_encode_64(&p, arg->oldest_flush_tid);
1272 * caller_uid/caller_gid (version 7)
1274 * Currently, we don't properly track which caller dirtied the caps
1275 * last, and force a flush of them when there is a conflict. For now,
1276 * just set this to 0:0, to emulate how the MDS has worked up to now.
1278 ceph_encode_32(&p, 0);
1279 ceph_encode_32(&p, 0);
1281 /* pool namespace (version 8) (mds always ignores this) */
1282 ceph_encode_32(&p, 0);
1284 /* btime and change_attr (version 9) */
1285 ceph_encode_timespec64(p, &arg->btime);
1286 p += sizeof(struct ceph_timespec);
1287 ceph_encode_64(&p, arg->change_attr);
1289 /* Advisory flags (version 10) */
1290 ceph_encode_32(&p, arg->flags);
1294 * Queue cap releases when an inode is dropped from our cache.
1296 void __ceph_remove_caps(struct ceph_inode_info *ci)
1300 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1301 * may call __ceph_caps_issued_mask() on a freeing inode. */
1302 spin_lock(&ci->i_ceph_lock);
1303 p = rb_first(&ci->i_caps);
1305 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1307 __ceph_remove_cap(cap, true);
1309 spin_unlock(&ci->i_ceph_lock);
1313 * Prepare to send a cap message to an MDS. Update the cap state, and populate
1314 * the arg struct with the parameters that will need to be sent. This should
1315 * be done under the i_ceph_lock to guard against changes to cap state.
1317 * Make note of max_size reported/requested from mds, revoked caps
1318 * that have now been implemented.
1320 static void __prep_cap(struct cap_msg_args *arg, struct ceph_cap *cap,
1321 int op, int flags, int used, int want, int retain,
1322 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1324 struct ceph_inode_info *ci = cap->ci;
1325 struct inode *inode = &ci->vfs_inode;
1328 lockdep_assert_held(&ci->i_ceph_lock);
1330 held = cap->issued | cap->implemented;
1331 revoking = cap->implemented & ~cap->issued;
1332 retain &= ~revoking;
1334 dout("%s %p cap %p session %p %s -> %s (revoking %s)\n",
1335 __func__, inode, cap, cap->session,
1336 ceph_cap_string(held), ceph_cap_string(held & retain),
1337 ceph_cap_string(revoking));
1338 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1340 ci->i_ceph_flags &= ~CEPH_I_FLUSH;
1342 cap->issued &= retain; /* drop bits we don't want */
1344 * Wake up any waiters on wanted -> needed transition. This is due to
1345 * the weird transition from buffered to sync IO... we need to flush
1346 * dirty pages _before_ allowing sync writes to avoid reordering.
1348 arg->wake = cap->implemented & ~cap->issued;
1349 cap->implemented &= cap->issued | used;
1350 cap->mds_wanted = want;
1352 arg->session = cap->session;
1353 arg->ino = ceph_vino(inode).ino;
1354 arg->cid = cap->cap_id;
1355 arg->follows = flushing ? ci->i_head_snapc->seq : 0;
1356 arg->flush_tid = flush_tid;
1357 arg->oldest_flush_tid = oldest_flush_tid;
1359 arg->size = i_size_read(inode);
1360 ci->i_reported_size = arg->size;
1361 arg->max_size = ci->i_wanted_max_size;
1362 if (cap == ci->i_auth_cap) {
1363 if (want & CEPH_CAP_ANY_FILE_WR)
1364 ci->i_requested_max_size = arg->max_size;
1366 ci->i_requested_max_size = 0;
1369 if (flushing & CEPH_CAP_XATTR_EXCL) {
1370 arg->old_xattr_buf = __ceph_build_xattrs_blob(ci);
1371 arg->xattr_version = ci->i_xattrs.version;
1372 arg->xattr_buf = ci->i_xattrs.blob;
1374 arg->xattr_buf = NULL;
1375 arg->old_xattr_buf = NULL;
1378 arg->mtime = inode->i_mtime;
1379 arg->atime = inode->i_atime;
1380 arg->ctime = inode->i_ctime;
1381 arg->btime = ci->i_btime;
1382 arg->change_attr = inode_peek_iversion_raw(inode);
1385 arg->caps = cap->implemented;
1387 arg->dirty = flushing;
1389 arg->seq = cap->seq;
1390 arg->issue_seq = cap->issue_seq;
1391 arg->mseq = cap->mseq;
1392 arg->time_warp_seq = ci->i_time_warp_seq;
1394 arg->uid = inode->i_uid;
1395 arg->gid = inode->i_gid;
1396 arg->mode = inode->i_mode;
1398 arg->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1399 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1400 !list_empty(&ci->i_cap_snaps)) {
1401 struct ceph_cap_snap *capsnap;
1402 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1403 if (capsnap->cap_flush.tid)
1405 if (capsnap->need_flush) {
1406 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1415 * Send a cap msg on the given inode.
1417 * Caller should hold snap_rwsem (read), s_mutex.
1419 static void __send_cap(struct cap_msg_args *arg, struct ceph_inode_info *ci)
1421 struct ceph_msg *msg;
1422 struct inode *inode = &ci->vfs_inode;
1424 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1426 pr_err("error allocating cap msg: ino (%llx.%llx) flushing %s tid %llu, requeuing cap.\n",
1427 ceph_vinop(inode), ceph_cap_string(arg->dirty),
1429 spin_lock(&ci->i_ceph_lock);
1430 __cap_delay_requeue(arg->session->s_mdsc, ci);
1431 spin_unlock(&ci->i_ceph_lock);
1435 encode_cap_msg(msg, arg);
1436 ceph_con_send(&arg->session->s_con, msg);
1437 ceph_buffer_put(arg->old_xattr_buf);
1439 wake_up_all(&ci->i_cap_wq);
1442 static inline int __send_flush_snap(struct inode *inode,
1443 struct ceph_mds_session *session,
1444 struct ceph_cap_snap *capsnap,
1445 u32 mseq, u64 oldest_flush_tid)
1447 struct cap_msg_args arg;
1448 struct ceph_msg *msg;
1450 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1454 arg.session = session;
1455 arg.ino = ceph_vino(inode).ino;
1457 arg.follows = capsnap->follows;
1458 arg.flush_tid = capsnap->cap_flush.tid;
1459 arg.oldest_flush_tid = oldest_flush_tid;
1461 arg.size = capsnap->size;
1463 arg.xattr_version = capsnap->xattr_version;
1464 arg.xattr_buf = capsnap->xattr_blob;
1465 arg.old_xattr_buf = NULL;
1467 arg.atime = capsnap->atime;
1468 arg.mtime = capsnap->mtime;
1469 arg.ctime = capsnap->ctime;
1470 arg.btime = capsnap->btime;
1471 arg.change_attr = capsnap->change_attr;
1473 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1474 arg.caps = capsnap->issued;
1476 arg.dirty = capsnap->dirty;
1481 arg.time_warp_seq = capsnap->time_warp_seq;
1483 arg.uid = capsnap->uid;
1484 arg.gid = capsnap->gid;
1485 arg.mode = capsnap->mode;
1487 arg.inline_data = capsnap->inline_data;
1491 encode_cap_msg(msg, &arg);
1492 ceph_con_send(&arg.session->s_con, msg);
1497 * When a snapshot is taken, clients accumulate dirty metadata on
1498 * inodes with capabilities in ceph_cap_snaps to describe the file
1499 * state at the time the snapshot was taken. This must be flushed
1500 * asynchronously back to the MDS once sync writes complete and dirty
1501 * data is written out.
1503 * Called under i_ceph_lock.
1505 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1506 struct ceph_mds_session *session)
1507 __releases(ci->i_ceph_lock)
1508 __acquires(ci->i_ceph_lock)
1510 struct inode *inode = &ci->vfs_inode;
1511 struct ceph_mds_client *mdsc = session->s_mdsc;
1512 struct ceph_cap_snap *capsnap;
1513 u64 oldest_flush_tid = 0;
1514 u64 first_tid = 1, last_tid = 0;
1516 dout("__flush_snaps %p session %p\n", inode, session);
1518 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1520 * we need to wait for sync writes to complete and for dirty
1521 * pages to be written out.
1523 if (capsnap->dirty_pages || capsnap->writing)
1526 /* should be removed by ceph_try_drop_cap_snap() */
1527 BUG_ON(!capsnap->need_flush);
1529 /* only flush each capsnap once */
1530 if (capsnap->cap_flush.tid > 0) {
1531 dout(" already flushed %p, skipping\n", capsnap);
1535 spin_lock(&mdsc->cap_dirty_lock);
1536 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1537 list_add_tail(&capsnap->cap_flush.g_list,
1538 &mdsc->cap_flush_list);
1539 if (oldest_flush_tid == 0)
1540 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1541 if (list_empty(&ci->i_flushing_item)) {
1542 list_add_tail(&ci->i_flushing_item,
1543 &session->s_cap_flushing);
1545 spin_unlock(&mdsc->cap_dirty_lock);
1547 list_add_tail(&capsnap->cap_flush.i_list,
1548 &ci->i_cap_flush_list);
1551 first_tid = capsnap->cap_flush.tid;
1552 last_tid = capsnap->cap_flush.tid;
1555 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1557 while (first_tid <= last_tid) {
1558 struct ceph_cap *cap = ci->i_auth_cap;
1559 struct ceph_cap_flush *cf;
1562 if (!(cap && cap->session == session)) {
1563 dout("__flush_snaps %p auth cap %p not mds%d, "
1564 "stop\n", inode, cap, session->s_mds);
1569 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1570 if (cf->tid >= first_tid) {
1578 first_tid = cf->tid + 1;
1580 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1581 refcount_inc(&capsnap->nref);
1582 spin_unlock(&ci->i_ceph_lock);
1584 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1585 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1587 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1590 pr_err("__flush_snaps: error sending cap flushsnap, "
1591 "ino (%llx.%llx) tid %llu follows %llu\n",
1592 ceph_vinop(inode), cf->tid, capsnap->follows);
1595 ceph_put_cap_snap(capsnap);
1596 spin_lock(&ci->i_ceph_lock);
1600 void ceph_flush_snaps(struct ceph_inode_info *ci,
1601 struct ceph_mds_session **psession)
1603 struct inode *inode = &ci->vfs_inode;
1604 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1605 struct ceph_mds_session *session = NULL;
1608 dout("ceph_flush_snaps %p\n", inode);
1610 session = *psession;
1612 spin_lock(&ci->i_ceph_lock);
1613 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1614 dout(" no capsnap needs flush, doing nothing\n");
1617 if (!ci->i_auth_cap) {
1618 dout(" no auth cap (migrating?), doing nothing\n");
1622 mds = ci->i_auth_cap->session->s_mds;
1623 if (session && session->s_mds != mds) {
1624 dout(" oops, wrong session %p mutex\n", session);
1625 ceph_put_mds_session(session);
1629 spin_unlock(&ci->i_ceph_lock);
1630 mutex_lock(&mdsc->mutex);
1631 session = __ceph_lookup_mds_session(mdsc, mds);
1632 mutex_unlock(&mdsc->mutex);
1636 // make sure flushsnap messages are sent in proper order.
1637 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1638 __kick_flushing_caps(mdsc, session, ci, 0);
1640 __ceph_flush_snaps(ci, session);
1642 spin_unlock(&ci->i_ceph_lock);
1645 *psession = session;
1647 ceph_put_mds_session(session);
1648 /* we flushed them all; remove this inode from the queue */
1649 spin_lock(&mdsc->snap_flush_lock);
1650 list_del_init(&ci->i_snap_flush_item);
1651 spin_unlock(&mdsc->snap_flush_lock);
1655 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1656 * Caller is then responsible for calling __mark_inode_dirty with the
1657 * returned flags value.
1659 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1660 struct ceph_cap_flush **pcf)
1662 struct ceph_mds_client *mdsc =
1663 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1664 struct inode *inode = &ci->vfs_inode;
1665 int was = ci->i_dirty_caps;
1668 lockdep_assert_held(&ci->i_ceph_lock);
1670 if (!ci->i_auth_cap) {
1671 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1672 "but no auth cap (session was closed?)\n",
1673 inode, ceph_ino(inode), ceph_cap_string(mask));
1677 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1678 ceph_cap_string(mask), ceph_cap_string(was),
1679 ceph_cap_string(was | mask));
1680 ci->i_dirty_caps |= mask;
1682 struct ceph_mds_session *session = ci->i_auth_cap->session;
1684 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1685 swap(ci->i_prealloc_cap_flush, *pcf);
1687 if (!ci->i_head_snapc) {
1688 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1689 ci->i_head_snapc = ceph_get_snap_context(
1690 ci->i_snap_realm->cached_context);
1692 dout(" inode %p now dirty snapc %p auth cap %p\n",
1693 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1694 BUG_ON(!list_empty(&ci->i_dirty_item));
1695 spin_lock(&mdsc->cap_dirty_lock);
1696 list_add(&ci->i_dirty_item, &session->s_cap_dirty);
1697 spin_unlock(&mdsc->cap_dirty_lock);
1698 if (ci->i_flushing_caps == 0) {
1700 dirty |= I_DIRTY_SYNC;
1703 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1705 BUG_ON(list_empty(&ci->i_dirty_item));
1706 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1707 (mask & CEPH_CAP_FILE_BUFFER))
1708 dirty |= I_DIRTY_DATASYNC;
1709 __cap_delay_requeue(mdsc, ci);
1713 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1715 struct ceph_cap_flush *cf;
1717 cf = kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1718 cf->is_capsnap = false;
1722 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1725 kmem_cache_free(ceph_cap_flush_cachep, cf);
1728 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1730 if (!list_empty(&mdsc->cap_flush_list)) {
1731 struct ceph_cap_flush *cf =
1732 list_first_entry(&mdsc->cap_flush_list,
1733 struct ceph_cap_flush, g_list);
1740 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1741 * Return true if caller needs to wake up flush waiters.
1743 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1744 struct ceph_cap_flush *cf)
1746 struct ceph_cap_flush *prev;
1747 bool wake = cf->wake;
1749 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1750 prev = list_prev_entry(cf, g_list);
1754 list_del_init(&cf->g_list);
1758 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1759 struct ceph_cap_flush *cf)
1761 struct ceph_cap_flush *prev;
1762 bool wake = cf->wake;
1764 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1765 prev = list_prev_entry(cf, i_list);
1769 list_del_init(&cf->i_list);
1774 * Add dirty inode to the flushing list. Assigned a seq number so we
1775 * can wait for caps to flush without starving.
1777 * Called under i_ceph_lock. Returns the flush tid.
1779 static u64 __mark_caps_flushing(struct inode *inode,
1780 struct ceph_mds_session *session, bool wake,
1781 u64 *oldest_flush_tid)
1783 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1784 struct ceph_inode_info *ci = ceph_inode(inode);
1785 struct ceph_cap_flush *cf = NULL;
1788 lockdep_assert_held(&ci->i_ceph_lock);
1789 BUG_ON(ci->i_dirty_caps == 0);
1790 BUG_ON(list_empty(&ci->i_dirty_item));
1791 BUG_ON(!ci->i_prealloc_cap_flush);
1793 flushing = ci->i_dirty_caps;
1794 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1795 ceph_cap_string(flushing),
1796 ceph_cap_string(ci->i_flushing_caps),
1797 ceph_cap_string(ci->i_flushing_caps | flushing));
1798 ci->i_flushing_caps |= flushing;
1799 ci->i_dirty_caps = 0;
1800 dout(" inode %p now !dirty\n", inode);
1802 swap(cf, ci->i_prealloc_cap_flush);
1803 cf->caps = flushing;
1806 spin_lock(&mdsc->cap_dirty_lock);
1807 list_del_init(&ci->i_dirty_item);
1809 cf->tid = ++mdsc->last_cap_flush_tid;
1810 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1811 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1813 if (list_empty(&ci->i_flushing_item)) {
1814 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1815 mdsc->num_cap_flushing++;
1817 spin_unlock(&mdsc->cap_dirty_lock);
1819 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1825 * try to invalidate mapping pages without blocking.
1827 static int try_nonblocking_invalidate(struct inode *inode)
1829 struct ceph_inode_info *ci = ceph_inode(inode);
1830 u32 invalidating_gen = ci->i_rdcache_gen;
1832 spin_unlock(&ci->i_ceph_lock);
1833 ceph_fscache_invalidate(inode);
1834 invalidate_mapping_pages(&inode->i_data, 0, -1);
1835 spin_lock(&ci->i_ceph_lock);
1837 if (inode->i_data.nrpages == 0 &&
1838 invalidating_gen == ci->i_rdcache_gen) {
1840 dout("try_nonblocking_invalidate %p success\n", inode);
1841 /* save any racing async invalidate some trouble */
1842 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1845 dout("try_nonblocking_invalidate %p failed\n", inode);
1849 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1851 loff_t size = i_size_read(&ci->vfs_inode);
1852 /* mds will adjust max size according to the reported size */
1853 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1855 if (size >= ci->i_max_size)
1857 /* half of previous max_size increment has been used */
1858 if (ci->i_max_size > ci->i_reported_size &&
1859 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1865 * Swiss army knife function to examine currently used and wanted
1866 * versus held caps. Release, flush, ack revoked caps to mds as
1869 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1870 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1873 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1874 struct ceph_mds_session *session)
1876 struct inode *inode = &ci->vfs_inode;
1877 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1878 struct ceph_cap *cap;
1879 u64 flush_tid, oldest_flush_tid;
1880 int file_wanted, used, cap_used;
1881 int issued, implemented, want, retain, revoking, flushing = 0;
1882 int mds = -1; /* keep track of how far we've gone through i_caps list
1883 to avoid an infinite loop on retry */
1885 bool queue_invalidate = false;
1886 bool tried_invalidate = false;
1889 ceph_get_mds_session(session);
1891 spin_lock(&ci->i_ceph_lock);
1892 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1893 flags |= CHECK_CAPS_FLUSH;
1895 /* Caps wanted by virtue of active open files. */
1896 file_wanted = __ceph_caps_file_wanted(ci);
1898 /* Caps which have active references against them */
1899 used = __ceph_caps_used(ci);
1902 * "issued" represents the current caps that the MDS wants us to have.
1903 * "implemented" is the set that we have been granted, and includes the
1904 * ones that have not yet been returned to the MDS (the "revoking" set,
1905 * usually because they have outstanding references).
1907 issued = __ceph_caps_issued(ci, &implemented);
1908 revoking = implemented & ~issued;
1912 /* The ones we currently want to retain (may be adjusted below) */
1913 retain = file_wanted | used | CEPH_CAP_PIN;
1914 if (!mdsc->stopping && inode->i_nlink > 0) {
1916 retain |= CEPH_CAP_ANY; /* be greedy */
1917 } else if (S_ISDIR(inode->i_mode) &&
1918 (issued & CEPH_CAP_FILE_SHARED) &&
1919 __ceph_dir_is_complete(ci)) {
1921 * If a directory is complete, we want to keep
1922 * the exclusive cap. So that MDS does not end up
1923 * revoking the shared cap on every create/unlink
1926 if (IS_RDONLY(inode)) {
1927 want = CEPH_CAP_ANY_SHARED;
1929 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1934 retain |= CEPH_CAP_ANY_SHARED;
1936 * keep RD only if we didn't have the file open RW,
1937 * because then the mds would revoke it anyway to
1938 * journal max_size=0.
1940 if (ci->i_max_size == 0)
1941 retain |= CEPH_CAP_ANY_RD;
1945 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1946 " issued %s revoking %s retain %s %s%s\n", inode,
1947 ceph_cap_string(file_wanted),
1948 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1949 ceph_cap_string(ci->i_flushing_caps),
1950 ceph_cap_string(issued), ceph_cap_string(revoking),
1951 ceph_cap_string(retain),
1952 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1953 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1956 * If we no longer need to hold onto old our caps, and we may
1957 * have cached pages, but don't want them, then try to invalidate.
1958 * If we fail, it's because pages are locked.... try again later.
1960 if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
1961 S_ISREG(inode->i_mode) &&
1962 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1963 inode->i_data.nrpages && /* have cached pages */
1964 (revoking & (CEPH_CAP_FILE_CACHE|
1965 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1966 !tried_invalidate) {
1967 dout("check_caps trying to invalidate on %p\n", inode);
1968 if (try_nonblocking_invalidate(inode) < 0) {
1969 dout("check_caps queuing invalidate\n");
1970 queue_invalidate = true;
1971 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1973 tried_invalidate = true;
1977 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1979 struct cap_msg_args arg;
1981 cap = rb_entry(p, struct ceph_cap, ci_node);
1983 /* avoid looping forever */
1984 if (mds >= cap->mds ||
1985 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1989 * If we have an auth cap, we don't need to consider any
1990 * overlapping caps as used.
1993 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1994 cap_used &= ~ci->i_auth_cap->issued;
1996 revoking = cap->implemented & ~cap->issued;
1997 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1998 cap->mds, cap, ceph_cap_string(cap_used),
1999 ceph_cap_string(cap->issued),
2000 ceph_cap_string(cap->implemented),
2001 ceph_cap_string(revoking));
2003 if (cap == ci->i_auth_cap &&
2004 (cap->issued & CEPH_CAP_FILE_WR)) {
2005 /* request larger max_size from MDS? */
2006 if (ci->i_wanted_max_size > ci->i_max_size &&
2007 ci->i_wanted_max_size > ci->i_requested_max_size) {
2008 dout("requesting new max_size\n");
2012 /* approaching file_max? */
2013 if (__ceph_should_report_size(ci)) {
2014 dout("i_size approaching max_size\n");
2018 /* flush anything dirty? */
2019 if (cap == ci->i_auth_cap) {
2020 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
2021 dout("flushing dirty caps\n");
2024 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
2025 dout("flushing snap caps\n");
2030 /* completed revocation? going down and there are no caps? */
2031 if (revoking && (revoking & cap_used) == 0) {
2032 dout("completed revocation of %s\n",
2033 ceph_cap_string(cap->implemented & ~cap->issued));
2037 /* want more caps from mds? */
2038 if (want & ~cap->mds_wanted) {
2039 if (want & ~(cap->mds_wanted | cap->issued))
2041 if (!__cap_is_valid(cap))
2045 /* things we might delay */
2046 if ((cap->issued & ~retain) == 0)
2047 continue; /* nope, all good */
2050 ceph_put_mds_session(session);
2051 session = ceph_get_mds_session(cap->session);
2053 /* kick flushing and flush snaps before sending normal
2055 if (cap == ci->i_auth_cap &&
2057 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2058 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2059 __kick_flushing_caps(mdsc, session, ci, 0);
2060 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2061 __ceph_flush_snaps(ci, session);
2066 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2067 flushing = ci->i_dirty_caps;
2068 flush_tid = __mark_caps_flushing(inode, session, false,
2070 if (flags & CHECK_CAPS_FLUSH &&
2071 list_empty(&session->s_cap_dirty))
2072 mflags |= CEPH_CLIENT_CAPS_SYNC;
2076 spin_lock(&mdsc->cap_dirty_lock);
2077 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2078 spin_unlock(&mdsc->cap_dirty_lock);
2081 mds = cap->mds; /* remember mds, so we don't repeat */
2083 __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2084 want, retain, flushing, flush_tid, oldest_flush_tid);
2086 spin_unlock(&ci->i_ceph_lock);
2087 __send_cap(&arg, ci);
2088 spin_lock(&ci->i_ceph_lock);
2090 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2093 /* periodically re-calculate caps wanted by open files */
2094 if (__ceph_is_any_real_caps(ci) &&
2095 list_empty(&ci->i_cap_delay_list) &&
2096 (file_wanted & ~CEPH_CAP_PIN) &&
2097 !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2098 __cap_delay_requeue(mdsc, ci);
2101 spin_unlock(&ci->i_ceph_lock);
2103 ceph_put_mds_session(session);
2104 if (queue_invalidate)
2105 ceph_queue_invalidate(inode);
2109 * Try to flush dirty caps back to the auth mds.
2111 static int try_flush_caps(struct inode *inode, u64 *ptid)
2113 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2114 struct ceph_inode_info *ci = ceph_inode(inode);
2116 u64 flush_tid = 0, oldest_flush_tid = 0;
2118 spin_lock(&ci->i_ceph_lock);
2120 if (ci->i_dirty_caps && ci->i_auth_cap) {
2121 struct ceph_cap *cap = ci->i_auth_cap;
2122 struct cap_msg_args arg;
2123 struct ceph_mds_session *session = cap->session;
2125 if (session->s_state < CEPH_MDS_SESSION_OPEN) {
2126 spin_unlock(&ci->i_ceph_lock);
2130 if (ci->i_ceph_flags &
2131 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2132 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2133 __kick_flushing_caps(mdsc, session, ci, 0);
2134 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2135 __ceph_flush_snaps(ci, session);
2139 flushing = ci->i_dirty_caps;
2140 flush_tid = __mark_caps_flushing(inode, session, true,
2143 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2144 __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2145 (cap->issued | cap->implemented),
2146 flushing, flush_tid, oldest_flush_tid);
2147 spin_unlock(&ci->i_ceph_lock);
2149 __send_cap(&arg, ci);
2151 if (!list_empty(&ci->i_cap_flush_list)) {
2152 struct ceph_cap_flush *cf =
2153 list_last_entry(&ci->i_cap_flush_list,
2154 struct ceph_cap_flush, i_list);
2156 flush_tid = cf->tid;
2158 flushing = ci->i_flushing_caps;
2159 spin_unlock(&ci->i_ceph_lock);
2167 * Return true if we've flushed caps through the given flush_tid.
2169 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2171 struct ceph_inode_info *ci = ceph_inode(inode);
2174 spin_lock(&ci->i_ceph_lock);
2175 if (!list_empty(&ci->i_cap_flush_list)) {
2176 struct ceph_cap_flush * cf =
2177 list_first_entry(&ci->i_cap_flush_list,
2178 struct ceph_cap_flush, i_list);
2179 if (cf->tid <= flush_tid)
2182 spin_unlock(&ci->i_ceph_lock);
2187 * wait for any unsafe requests to complete.
2189 static int unsafe_request_wait(struct inode *inode)
2191 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2192 struct ceph_inode_info *ci = ceph_inode(inode);
2193 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2196 spin_lock(&ci->i_unsafe_lock);
2197 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2198 req1 = list_last_entry(&ci->i_unsafe_dirops,
2199 struct ceph_mds_request,
2201 ceph_mdsc_get_request(req1);
2203 if (!list_empty(&ci->i_unsafe_iops)) {
2204 req2 = list_last_entry(&ci->i_unsafe_iops,
2205 struct ceph_mds_request,
2206 r_unsafe_target_item);
2207 ceph_mdsc_get_request(req2);
2209 spin_unlock(&ci->i_unsafe_lock);
2212 * Trigger to flush the journal logs in all the relevant MDSes
2213 * manually, or in the worst case we must wait at most 5 seconds
2214 * to wait the journal logs to be flushed by the MDSes periodically.
2217 struct ceph_mds_session **sessions = NULL;
2218 struct ceph_mds_session *s;
2219 struct ceph_mds_request *req;
2224 * The mdsc->max_sessions is unlikely to be changed
2225 * mostly, here we will retry it by reallocating the
2226 * sessions arrary memory to get rid of the mdsc->mutex
2230 max = mdsc->max_sessions;
2231 sessions = krealloc(sessions, max * sizeof(s), __GFP_ZERO);
2235 spin_lock(&ci->i_unsafe_lock);
2237 list_for_each_entry(req, &ci->i_unsafe_dirops,
2238 r_unsafe_dir_item) {
2240 if (unlikely(s->s_mds > max)) {
2241 spin_unlock(&ci->i_unsafe_lock);
2244 if (!sessions[s->s_mds]) {
2245 s = ceph_get_mds_session(s);
2246 sessions[s->s_mds] = s;
2251 list_for_each_entry(req, &ci->i_unsafe_iops,
2252 r_unsafe_target_item) {
2254 if (unlikely(s->s_mds > max)) {
2255 spin_unlock(&ci->i_unsafe_lock);
2258 if (!sessions[s->s_mds]) {
2259 s = ceph_get_mds_session(s);
2260 sessions[s->s_mds] = s;
2264 spin_unlock(&ci->i_unsafe_lock);
2267 spin_lock(&ci->i_ceph_lock);
2268 if (ci->i_auth_cap) {
2269 s = ci->i_auth_cap->session;
2270 if (!sessions[s->s_mds])
2271 sessions[s->s_mds] = ceph_get_mds_session(s);
2273 spin_unlock(&ci->i_ceph_lock);
2275 /* send flush mdlog request to MDSes */
2276 for (i = 0; i < max; i++) {
2279 send_flush_mdlog(s);
2280 ceph_put_mds_session(s);
2286 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2287 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2289 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2290 ceph_timeout_jiffies(req1->r_timeout));
2293 ceph_mdsc_put_request(req1);
2296 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2297 ceph_timeout_jiffies(req2->r_timeout));
2300 ceph_mdsc_put_request(req2);
2305 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2307 struct ceph_file_info *fi = file->private_data;
2308 struct inode *inode = file->f_mapping->host;
2309 struct ceph_inode_info *ci = ceph_inode(inode);
2314 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2316 ret = file_write_and_wait_range(file, start, end);
2320 ret = ceph_wait_on_async_create(inode);
2324 dirty = try_flush_caps(inode, &flush_tid);
2325 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2327 err = unsafe_request_wait(inode);
2330 * only wait on non-file metadata writeback (the mds
2331 * can recover size and mtime, so we don't need to
2334 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2335 err = wait_event_interruptible(ci->i_cap_wq,
2336 caps_are_flushed(inode, flush_tid));
2342 if (errseq_check(&ci->i_meta_err, READ_ONCE(fi->meta_err))) {
2343 spin_lock(&file->f_lock);
2344 err = errseq_check_and_advance(&ci->i_meta_err,
2346 spin_unlock(&file->f_lock);
2351 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2356 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2357 * queue inode for flush but don't do so immediately, because we can
2358 * get by with fewer MDS messages if we wait for data writeback to
2361 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2363 struct ceph_inode_info *ci = ceph_inode(inode);
2367 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2369 dout("write_inode %p wait=%d\n", inode, wait);
2371 dirty = try_flush_caps(inode, &flush_tid);
2373 err = wait_event_interruptible(ci->i_cap_wq,
2374 caps_are_flushed(inode, flush_tid));
2376 struct ceph_mds_client *mdsc =
2377 ceph_sb_to_client(inode->i_sb)->mdsc;
2379 spin_lock(&ci->i_ceph_lock);
2380 if (__ceph_caps_dirty(ci))
2381 __cap_delay_requeue_front(mdsc, ci);
2382 spin_unlock(&ci->i_ceph_lock);
2387 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2388 struct ceph_mds_session *session,
2389 struct ceph_inode_info *ci,
2390 u64 oldest_flush_tid)
2391 __releases(ci->i_ceph_lock)
2392 __acquires(ci->i_ceph_lock)
2394 struct inode *inode = &ci->vfs_inode;
2395 struct ceph_cap *cap;
2396 struct ceph_cap_flush *cf;
2399 u64 last_snap_flush = 0;
2401 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2403 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2404 if (cf->is_capsnap) {
2405 last_snap_flush = cf->tid;
2410 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2411 if (cf->tid < first_tid)
2414 cap = ci->i_auth_cap;
2415 if (!(cap && cap->session == session)) {
2416 pr_err("%p auth cap %p not mds%d ???\n",
2417 inode, cap, session->s_mds);
2421 first_tid = cf->tid + 1;
2423 if (!cf->is_capsnap) {
2424 struct cap_msg_args arg;
2426 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2427 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2428 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2429 (cf->tid < last_snap_flush ?
2430 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2431 __ceph_caps_used(ci),
2432 __ceph_caps_wanted(ci),
2433 (cap->issued | cap->implemented),
2434 cf->caps, cf->tid, oldest_flush_tid);
2435 spin_unlock(&ci->i_ceph_lock);
2436 __send_cap(&arg, ci);
2438 struct ceph_cap_snap *capsnap =
2439 container_of(cf, struct ceph_cap_snap,
2441 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2442 inode, capsnap, cf->tid,
2443 ceph_cap_string(capsnap->dirty));
2445 refcount_inc(&capsnap->nref);
2446 spin_unlock(&ci->i_ceph_lock);
2448 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2451 pr_err("kick_flushing_caps: error sending "
2452 "cap flushsnap, ino (%llx.%llx) "
2453 "tid %llu follows %llu\n",
2454 ceph_vinop(inode), cf->tid,
2458 ceph_put_cap_snap(capsnap);
2461 spin_lock(&ci->i_ceph_lock);
2465 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2466 struct ceph_mds_session *session)
2468 struct ceph_inode_info *ci;
2469 struct ceph_cap *cap;
2470 u64 oldest_flush_tid;
2472 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2474 spin_lock(&mdsc->cap_dirty_lock);
2475 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2476 spin_unlock(&mdsc->cap_dirty_lock);
2478 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2479 spin_lock(&ci->i_ceph_lock);
2480 cap = ci->i_auth_cap;
2481 if (!(cap && cap->session == session)) {
2482 pr_err("%p auth cap %p not mds%d ???\n",
2483 &ci->vfs_inode, cap, session->s_mds);
2484 spin_unlock(&ci->i_ceph_lock);
2490 * if flushing caps were revoked, we re-send the cap flush
2491 * in client reconnect stage. This guarantees MDS * processes
2492 * the cap flush message before issuing the flushing caps to
2495 if ((cap->issued & ci->i_flushing_caps) !=
2496 ci->i_flushing_caps) {
2497 /* encode_caps_cb() also will reset these sequence
2498 * numbers. make sure sequence numbers in cap flush
2499 * message match later reconnect message */
2503 __kick_flushing_caps(mdsc, session, ci,
2506 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2509 spin_unlock(&ci->i_ceph_lock);
2513 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2514 struct ceph_mds_session *session)
2516 struct ceph_inode_info *ci;
2517 struct ceph_cap *cap;
2518 u64 oldest_flush_tid;
2520 lockdep_assert_held(&session->s_mutex);
2522 dout("kick_flushing_caps mds%d\n", session->s_mds);
2524 spin_lock(&mdsc->cap_dirty_lock);
2525 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2526 spin_unlock(&mdsc->cap_dirty_lock);
2528 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2529 spin_lock(&ci->i_ceph_lock);
2530 cap = ci->i_auth_cap;
2531 if (!(cap && cap->session == session)) {
2532 pr_err("%p auth cap %p not mds%d ???\n",
2533 &ci->vfs_inode, cap, session->s_mds);
2534 spin_unlock(&ci->i_ceph_lock);
2537 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2538 __kick_flushing_caps(mdsc, session, ci,
2541 spin_unlock(&ci->i_ceph_lock);
2545 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2546 struct ceph_inode_info *ci)
2548 struct ceph_mds_client *mdsc = session->s_mdsc;
2549 struct ceph_cap *cap = ci->i_auth_cap;
2551 lockdep_assert_held(&ci->i_ceph_lock);
2553 dout("%s %p flushing %s\n", __func__, &ci->vfs_inode,
2554 ceph_cap_string(ci->i_flushing_caps));
2556 if (!list_empty(&ci->i_cap_flush_list)) {
2557 u64 oldest_flush_tid;
2558 spin_lock(&mdsc->cap_dirty_lock);
2559 list_move_tail(&ci->i_flushing_item,
2560 &cap->session->s_cap_flushing);
2561 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2562 spin_unlock(&mdsc->cap_dirty_lock);
2564 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2570 * Take references to capabilities we hold, so that we don't release
2571 * them to the MDS prematurely.
2573 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2574 bool snap_rwsem_locked)
2576 lockdep_assert_held(&ci->i_ceph_lock);
2578 if (got & CEPH_CAP_PIN)
2580 if (got & CEPH_CAP_FILE_RD)
2582 if (got & CEPH_CAP_FILE_CACHE)
2583 ci->i_rdcache_ref++;
2584 if (got & CEPH_CAP_FILE_EXCL)
2586 if (got & CEPH_CAP_FILE_WR) {
2587 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2588 BUG_ON(!snap_rwsem_locked);
2589 ci->i_head_snapc = ceph_get_snap_context(
2590 ci->i_snap_realm->cached_context);
2594 if (got & CEPH_CAP_FILE_BUFFER) {
2595 if (ci->i_wb_ref == 0)
2596 ihold(&ci->vfs_inode);
2598 dout("%s %p wb %d -> %d (?)\n", __func__,
2599 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2604 * Try to grab cap references. Specify those refs we @want, and the
2605 * minimal set we @need. Also include the larger offset we are writing
2606 * to (when applicable), and check against max_size here as well.
2607 * Note that caller is responsible for ensuring max_size increases are
2608 * requested from the MDS.
2610 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2611 * or a negative error code. There are 3 speical error codes:
2612 * -EAGAIN: need to sleep but non-blocking is specified
2613 * -EFBIG: ask caller to call check_max_size() and try again.
2614 * -ESTALE: ask caller to call ceph_renew_caps() and try again.
2617 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2618 NON_BLOCKING = (1 << 8),
2619 CHECK_FILELOCK = (1 << 9),
2622 static int try_get_cap_refs(struct inode *inode, int need, int want,
2623 loff_t endoff, int flags, int *got)
2625 struct ceph_inode_info *ci = ceph_inode(inode);
2626 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2628 int have, implemented;
2629 bool snap_rwsem_locked = false;
2631 dout("get_cap_refs %p need %s want %s\n", inode,
2632 ceph_cap_string(need), ceph_cap_string(want));
2635 spin_lock(&ci->i_ceph_lock);
2637 if ((flags & CHECK_FILELOCK) &&
2638 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2639 dout("try_get_cap_refs %p error filelock\n", inode);
2644 /* finish pending truncate */
2645 while (ci->i_truncate_pending) {
2646 spin_unlock(&ci->i_ceph_lock);
2647 if (snap_rwsem_locked) {
2648 up_read(&mdsc->snap_rwsem);
2649 snap_rwsem_locked = false;
2651 __ceph_do_pending_vmtruncate(inode);
2652 spin_lock(&ci->i_ceph_lock);
2655 have = __ceph_caps_issued(ci, &implemented);
2657 if (have & need & CEPH_CAP_FILE_WR) {
2658 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2659 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2660 inode, endoff, ci->i_max_size);
2661 if (endoff > ci->i_requested_max_size)
2662 ret = ci->i_auth_cap ? -EFBIG : -ESTALE;
2666 * If a sync write is in progress, we must wait, so that we
2667 * can get a final snapshot value for size+mtime.
2669 if (__ceph_have_pending_cap_snap(ci)) {
2670 dout("get_cap_refs %p cap_snap_pending\n", inode);
2675 if ((have & need) == need) {
2677 * Look at (implemented & ~have & not) so that we keep waiting
2678 * on transition from wanted -> needed caps. This is needed
2679 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2680 * going before a prior buffered writeback happens.
2682 int not = want & ~(have & need);
2683 int revoking = implemented & ~have;
2684 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2685 inode, ceph_cap_string(have), ceph_cap_string(not),
2686 ceph_cap_string(revoking));
2687 if ((revoking & not) == 0) {
2688 if (!snap_rwsem_locked &&
2689 !ci->i_head_snapc &&
2690 (need & CEPH_CAP_FILE_WR)) {
2691 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2693 * we can not call down_read() when
2694 * task isn't in TASK_RUNNING state
2696 if (flags & NON_BLOCKING) {
2701 spin_unlock(&ci->i_ceph_lock);
2702 down_read(&mdsc->snap_rwsem);
2703 snap_rwsem_locked = true;
2706 snap_rwsem_locked = true;
2708 if ((have & want) == want)
2712 ceph_take_cap_refs(ci, *got, true);
2716 int session_readonly = false;
2718 if (ci->i_auth_cap &&
2719 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2720 struct ceph_mds_session *s = ci->i_auth_cap->session;
2721 spin_lock(&s->s_cap_lock);
2722 session_readonly = s->s_readonly;
2723 spin_unlock(&s->s_cap_lock);
2725 if (session_readonly) {
2726 dout("get_cap_refs %p need %s but mds%d readonly\n",
2727 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2732 if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
2733 dout("get_cap_refs %p forced umount\n", inode);
2737 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2738 if (need & ~mds_wanted) {
2739 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2740 inode, ceph_cap_string(need),
2741 ceph_cap_string(mds_wanted));
2746 dout("get_cap_refs %p have %s need %s\n", inode,
2747 ceph_cap_string(have), ceph_cap_string(need));
2751 __ceph_touch_fmode(ci, mdsc, flags);
2753 spin_unlock(&ci->i_ceph_lock);
2754 if (snap_rwsem_locked)
2755 up_read(&mdsc->snap_rwsem);
2758 ceph_update_cap_mis(&mdsc->metric);
2760 ceph_update_cap_hit(&mdsc->metric);
2762 dout("get_cap_refs %p ret %d got %s\n", inode,
2763 ret, ceph_cap_string(*got));
2768 * Check the offset we are writing up to against our current
2769 * max_size. If necessary, tell the MDS we want to write to
2772 static void check_max_size(struct inode *inode, loff_t endoff)
2774 struct ceph_inode_info *ci = ceph_inode(inode);
2777 /* do we need to explicitly request a larger max_size? */
2778 spin_lock(&ci->i_ceph_lock);
2779 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2780 dout("write %p at large endoff %llu, req max_size\n",
2782 ci->i_wanted_max_size = endoff;
2784 /* duplicate ceph_check_caps()'s logic */
2785 if (ci->i_auth_cap &&
2786 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2787 ci->i_wanted_max_size > ci->i_max_size &&
2788 ci->i_wanted_max_size > ci->i_requested_max_size)
2790 spin_unlock(&ci->i_ceph_lock);
2792 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2795 static inline int get_used_fmode(int caps)
2798 if (caps & CEPH_CAP_FILE_RD)
2799 fmode |= CEPH_FILE_MODE_RD;
2800 if (caps & CEPH_CAP_FILE_WR)
2801 fmode |= CEPH_FILE_MODE_WR;
2805 int ceph_try_get_caps(struct inode *inode, int need, int want,
2806 bool nonblock, int *got)
2810 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2811 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2812 CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2813 CEPH_CAP_ANY_DIR_OPS));
2815 ret = ceph_pool_perm_check(inode, need);
2820 flags = get_used_fmode(need | want);
2822 flags |= NON_BLOCKING;
2824 ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2825 /* three special error codes */
2826 if (ret == -EAGAIN || ret == -EFBIG || ret == -ESTALE)
2832 * Wait for caps, and take cap references. If we can't get a WR cap
2833 * due to a small max_size, make sure we check_max_size (and possibly
2834 * ask the mds) so we don't get hung up indefinitely.
2836 int ceph_get_caps(struct file *filp, int need, int want, loff_t endoff, int *got)
2838 struct ceph_file_info *fi = filp->private_data;
2839 struct inode *inode = file_inode(filp);
2840 struct ceph_inode_info *ci = ceph_inode(inode);
2841 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2842 int ret, _got, flags;
2844 ret = ceph_pool_perm_check(inode, need);
2848 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2849 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2852 flags = get_used_fmode(need | want);
2855 flags &= CEPH_FILE_MODE_MASK;
2856 if (atomic_read(&fi->num_locks))
2857 flags |= CHECK_FILELOCK;
2859 ret = try_get_cap_refs(inode, need, want, endoff,
2861 WARN_ON_ONCE(ret == -EAGAIN);
2863 struct ceph_mds_client *mdsc = fsc->mdsc;
2865 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2867 cw.ino = ceph_ino(inode);
2868 cw.tgid = current->tgid;
2872 spin_lock(&mdsc->caps_list_lock);
2873 list_add(&cw.list, &mdsc->cap_wait_list);
2874 spin_unlock(&mdsc->caps_list_lock);
2876 /* make sure used fmode not timeout */
2877 ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2878 add_wait_queue(&ci->i_cap_wq, &wait);
2880 flags |= NON_BLOCKING;
2881 while (!(ret = try_get_cap_refs(inode, need, want,
2882 endoff, flags, &_got))) {
2883 if (signal_pending(current)) {
2887 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2890 remove_wait_queue(&ci->i_cap_wq, &wait);
2891 ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
2893 spin_lock(&mdsc->caps_list_lock);
2895 spin_unlock(&mdsc->caps_list_lock);
2901 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2902 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2903 if (ret >= 0 && _got)
2904 ceph_put_cap_refs(ci, _got);
2909 if (ret == -EFBIG || ret == -ESTALE) {
2910 int ret2 = ceph_wait_on_async_create(inode);
2914 if (ret == -EFBIG) {
2915 check_max_size(inode, endoff);
2918 if (ret == -ESTALE) {
2919 /* session was killed, try renew caps */
2920 ret = ceph_renew_caps(inode, flags);
2927 if (S_ISREG(ci->vfs_inode.i_mode) &&
2928 ci->i_inline_version != CEPH_INLINE_NONE &&
2929 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2930 i_size_read(inode) > 0) {
2932 find_get_page(inode->i_mapping, 0);
2934 bool uptodate = PageUptodate(page);
2941 * drop cap refs first because getattr while
2942 * holding * caps refs can cause deadlock.
2944 ceph_put_cap_refs(ci, _got);
2948 * getattr request will bring inline data into
2951 ret = __ceph_do_getattr(inode, NULL,
2952 CEPH_STAT_CAP_INLINE_DATA,
2965 * Take cap refs. Caller must already know we hold at least one ref
2966 * on the caps in question or we don't know this is safe.
2968 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2970 spin_lock(&ci->i_ceph_lock);
2971 ceph_take_cap_refs(ci, caps, false);
2972 spin_unlock(&ci->i_ceph_lock);
2977 * drop cap_snap that is not associated with any snapshot.
2978 * we don't need to send FLUSHSNAP message for it.
2980 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2981 struct ceph_cap_snap *capsnap)
2983 if (!capsnap->need_flush &&
2984 !capsnap->writing && !capsnap->dirty_pages) {
2985 dout("dropping cap_snap %p follows %llu\n",
2986 capsnap, capsnap->follows);
2987 BUG_ON(capsnap->cap_flush.tid > 0);
2988 ceph_put_snap_context(capsnap->context);
2989 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2990 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2992 list_del(&capsnap->ci_item);
2993 ceph_put_cap_snap(capsnap);
2999 enum put_cap_refs_mode {
3000 PUT_CAP_REFS_SYNC = 0,
3001 PUT_CAP_REFS_NO_CHECK,
3008 * If we released the last ref on any given cap, call ceph_check_caps
3009 * to release (or schedule a release).
3011 * If we are releasing a WR cap (from a sync write), finalize any affected
3012 * cap_snap, and wake up any waiters.
3014 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
3015 enum put_cap_refs_mode mode)
3017 struct inode *inode = &ci->vfs_inode;
3018 int last = 0, put = 0, flushsnaps = 0, wake = 0;
3019 bool check_flushsnaps = false;
3021 spin_lock(&ci->i_ceph_lock);
3022 if (had & CEPH_CAP_PIN)
3024 if (had & CEPH_CAP_FILE_RD)
3025 if (--ci->i_rd_ref == 0)
3027 if (had & CEPH_CAP_FILE_CACHE)
3028 if (--ci->i_rdcache_ref == 0)
3030 if (had & CEPH_CAP_FILE_EXCL)
3031 if (--ci->i_fx_ref == 0)
3033 if (had & CEPH_CAP_FILE_BUFFER) {
3034 if (--ci->i_wb_ref == 0) {
3036 /* put the ref held by ceph_take_cap_refs() */
3038 check_flushsnaps = true;
3040 dout("put_cap_refs %p wb %d -> %d (?)\n",
3041 inode, ci->i_wb_ref+1, ci->i_wb_ref);
3043 if (had & CEPH_CAP_FILE_WR) {
3044 if (--ci->i_wr_ref == 0) {
3046 check_flushsnaps = true;
3047 if (ci->i_wrbuffer_ref_head == 0 &&
3048 ci->i_dirty_caps == 0 &&
3049 ci->i_flushing_caps == 0) {
3050 BUG_ON(!ci->i_head_snapc);
3051 ceph_put_snap_context(ci->i_head_snapc);
3052 ci->i_head_snapc = NULL;
3054 /* see comment in __ceph_remove_cap() */
3055 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3056 ceph_change_snap_realm(inode, NULL);
3059 if (check_flushsnaps && __ceph_have_pending_cap_snap(ci)) {
3060 struct ceph_cap_snap *capsnap =
3061 list_last_entry(&ci->i_cap_snaps,
3062 struct ceph_cap_snap,
3065 capsnap->writing = 0;
3066 if (ceph_try_drop_cap_snap(ci, capsnap))
3067 /* put the ref held by ceph_queue_cap_snap() */
3069 else if (__ceph_finish_cap_snap(ci, capsnap))
3073 spin_unlock(&ci->i_ceph_lock);
3075 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3076 last ? " last" : "", put ? " put" : "");
3079 case PUT_CAP_REFS_SYNC:
3081 ceph_check_caps(ci, 0, NULL);
3082 else if (flushsnaps)
3083 ceph_flush_snaps(ci, NULL);
3085 case PUT_CAP_REFS_ASYNC:
3087 ceph_queue_check_caps(inode);
3088 else if (flushsnaps)
3089 ceph_queue_flush_snaps(inode);
3095 wake_up_all(&ci->i_cap_wq);
3100 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3102 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_SYNC);
3105 void ceph_put_cap_refs_async(struct ceph_inode_info *ci, int had)
3107 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_ASYNC);
3110 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3112 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_NO_CHECK);
3116 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3117 * context. Adjust per-snap dirty page accounting as appropriate.
3118 * Once all dirty data for a cap_snap is flushed, flush snapped file
3119 * metadata back to the MDS. If we dropped the last ref, call
3122 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3123 struct ceph_snap_context *snapc)
3125 struct inode *inode = &ci->vfs_inode;
3126 struct ceph_cap_snap *capsnap = NULL;
3130 bool flush_snaps = false;
3131 bool complete_capsnap = false;
3133 spin_lock(&ci->i_ceph_lock);
3134 ci->i_wrbuffer_ref -= nr;
3135 if (ci->i_wrbuffer_ref == 0) {
3140 if (ci->i_head_snapc == snapc) {
3141 ci->i_wrbuffer_ref_head -= nr;
3142 if (ci->i_wrbuffer_ref_head == 0 &&
3143 ci->i_wr_ref == 0 &&
3144 ci->i_dirty_caps == 0 &&
3145 ci->i_flushing_caps == 0) {
3146 BUG_ON(!ci->i_head_snapc);
3147 ceph_put_snap_context(ci->i_head_snapc);
3148 ci->i_head_snapc = NULL;
3150 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3152 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3153 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3154 last ? " LAST" : "");
3156 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3157 if (capsnap->context == snapc) {
3163 capsnap->dirty_pages -= nr;
3164 if (capsnap->dirty_pages == 0) {
3165 complete_capsnap = true;
3166 if (!capsnap->writing) {
3167 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3170 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3175 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3176 " snap %lld %d/%d -> %d/%d %s%s\n",
3177 inode, capsnap, capsnap->context->seq,
3178 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3179 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3180 last ? " (wrbuffer last)" : "",
3181 complete_capsnap ? " (complete capsnap)" : "");
3184 spin_unlock(&ci->i_ceph_lock);
3187 ceph_check_caps(ci, 0, NULL);
3188 } else if (flush_snaps) {
3189 ceph_flush_snaps(ci, NULL);
3191 if (complete_capsnap)
3192 wake_up_all(&ci->i_cap_wq);
3199 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3201 static void invalidate_aliases(struct inode *inode)
3203 struct dentry *dn, *prev = NULL;
3205 dout("invalidate_aliases inode %p\n", inode);
3206 d_prune_aliases(inode);
3208 * For non-directory inode, d_find_alias() only returns
3209 * hashed dentry. After calling d_invalidate(), the
3210 * dentry becomes unhashed.
3212 * For directory inode, d_find_alias() can return
3213 * unhashed dentry. But directory inode should have
3214 * one alias at most.
3216 while ((dn = d_find_alias(inode))) {
3230 struct cap_extra_info {
3231 struct ceph_string *pool_ns;
3241 /* currently issued */
3243 struct timespec64 btime;
3247 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3248 * actually be a revocation if it specifies a smaller cap set.)
3250 * caller holds s_mutex and i_ceph_lock, we drop both.
3252 static void handle_cap_grant(struct inode *inode,
3253 struct ceph_mds_session *session,
3254 struct ceph_cap *cap,
3255 struct ceph_mds_caps *grant,
3256 struct ceph_buffer *xattr_buf,
3257 struct cap_extra_info *extra_info)
3258 __releases(ci->i_ceph_lock)
3259 __releases(session->s_mdsc->snap_rwsem)
3261 struct ceph_inode_info *ci = ceph_inode(inode);
3262 int seq = le32_to_cpu(grant->seq);
3263 int newcaps = le32_to_cpu(grant->caps);
3264 int used, wanted, dirty;
3265 u64 size = le64_to_cpu(grant->size);
3266 u64 max_size = le64_to_cpu(grant->max_size);
3267 unsigned char check_caps = 0;
3268 bool was_stale = cap->cap_gen < atomic_read(&session->s_cap_gen);
3270 bool writeback = false;
3271 bool queue_trunc = false;
3272 bool queue_invalidate = false;
3273 bool deleted_inode = false;
3274 bool fill_inline = false;
3276 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3277 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3278 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3279 i_size_read(inode));
3283 * If CACHE is being revoked, and we have no dirty buffers,
3284 * try to invalidate (once). (If there are dirty buffers, we
3285 * will invalidate _after_ writeback.)
3287 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3288 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3289 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3290 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3291 if (try_nonblocking_invalidate(inode)) {
3292 /* there were locked pages.. invalidate later
3293 in a separate thread. */
3294 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3295 queue_invalidate = true;
3296 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3302 cap->issued = cap->implemented = CEPH_CAP_PIN;
3305 * auth mds of the inode changed. we received the cap export message,
3306 * but still haven't received the cap import message. handle_cap_export
3307 * updated the new auth MDS' cap.
3309 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3310 * that was sent before the cap import message. So don't remove caps.
3312 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3313 WARN_ON(cap != ci->i_auth_cap);
3314 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3316 newcaps |= cap->issued;
3319 /* side effects now are allowed */
3320 cap->cap_gen = atomic_read(&session->s_cap_gen);
3323 __check_cap_issue(ci, cap, newcaps);
3325 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3327 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3328 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3329 umode_t mode = le32_to_cpu(grant->mode);
3331 if (inode_wrong_type(inode, mode))
3332 pr_warn_once("inode type changed! (ino %llx.%llx is 0%o, mds says 0%o)\n",
3333 ceph_vinop(inode), inode->i_mode, mode);
3335 inode->i_mode = mode;
3336 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3337 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3338 ci->i_btime = extra_info->btime;
3339 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3340 from_kuid(&init_user_ns, inode->i_uid),
3341 from_kgid(&init_user_ns, inode->i_gid));
3344 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3345 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3346 set_nlink(inode, le32_to_cpu(grant->nlink));
3347 if (inode->i_nlink == 0 &&
3348 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3349 deleted_inode = true;
3352 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3354 int len = le32_to_cpu(grant->xattr_len);
3355 u64 version = le64_to_cpu(grant->xattr_version);
3357 if (version > ci->i_xattrs.version) {
3358 dout(" got new xattrs v%llu on %p len %d\n",
3359 version, inode, len);
3360 if (ci->i_xattrs.blob)
3361 ceph_buffer_put(ci->i_xattrs.blob);
3362 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3363 ci->i_xattrs.version = version;
3364 ceph_forget_all_cached_acls(inode);
3365 ceph_security_invalidate_secctx(inode);
3369 if (newcaps & CEPH_CAP_ANY_RD) {
3370 struct timespec64 mtime, atime, ctime;
3371 /* ctime/mtime/atime? */
3372 ceph_decode_timespec64(&mtime, &grant->mtime);
3373 ceph_decode_timespec64(&atime, &grant->atime);
3374 ceph_decode_timespec64(&ctime, &grant->ctime);
3375 ceph_fill_file_time(inode, extra_info->issued,
3376 le32_to_cpu(grant->time_warp_seq),
3377 &ctime, &mtime, &atime);
3380 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3381 ci->i_files = extra_info->nfiles;
3382 ci->i_subdirs = extra_info->nsubdirs;
3385 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3386 /* file layout may have changed */
3387 s64 old_pool = ci->i_layout.pool_id;
3388 struct ceph_string *old_ns;
3390 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3391 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3392 lockdep_is_held(&ci->i_ceph_lock));
3393 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3395 if (ci->i_layout.pool_id != old_pool ||
3396 extra_info->pool_ns != old_ns)
3397 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3399 extra_info->pool_ns = old_ns;
3401 /* size/truncate_seq? */
3402 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3403 le32_to_cpu(grant->truncate_seq),
3404 le64_to_cpu(grant->truncate_size),
3408 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3409 if (max_size != ci->i_max_size) {
3410 dout("max_size %lld -> %llu\n",
3411 ci->i_max_size, max_size);
3412 ci->i_max_size = max_size;
3413 if (max_size >= ci->i_wanted_max_size) {
3414 ci->i_wanted_max_size = 0; /* reset */
3415 ci->i_requested_max_size = 0;
3421 /* check cap bits */
3422 wanted = __ceph_caps_wanted(ci);
3423 used = __ceph_caps_used(ci);
3424 dirty = __ceph_caps_dirty(ci);
3425 dout(" my wanted = %s, used = %s, dirty %s\n",
3426 ceph_cap_string(wanted),
3427 ceph_cap_string(used),
3428 ceph_cap_string(dirty));
3430 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3431 (wanted & ~(cap->mds_wanted | newcaps))) {
3433 * If mds is importing cap, prior cap messages that update
3434 * 'wanted' may get dropped by mds (migrate seq mismatch).
3436 * We don't send cap message to update 'wanted' if what we
3437 * want are already issued. If mds revokes caps, cap message
3438 * that releases caps also tells mds what we want. But if
3439 * caps got revoked by mds forcedly (session stale). We may
3440 * haven't told mds what we want.
3445 /* revocation, grant, or no-op? */
3446 if (cap->issued & ~newcaps) {
3447 int revoking = cap->issued & ~newcaps;
3449 dout("revocation: %s -> %s (revoking %s)\n",
3450 ceph_cap_string(cap->issued),
3451 ceph_cap_string(newcaps),
3452 ceph_cap_string(revoking));
3453 if (S_ISREG(inode->i_mode) &&
3454 (revoking & used & CEPH_CAP_FILE_BUFFER))
3455 writeback = true; /* initiate writeback; will delay ack */
3456 else if (queue_invalidate &&
3457 revoking == CEPH_CAP_FILE_CACHE &&
3458 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3459 ; /* do nothing yet, invalidation will be queued */
3460 else if (cap == ci->i_auth_cap)
3461 check_caps = 1; /* check auth cap only */
3463 check_caps = 2; /* check all caps */
3464 cap->issued = newcaps;
3465 cap->implemented |= newcaps;
3466 } else if (cap->issued == newcaps) {
3467 dout("caps unchanged: %s -> %s\n",
3468 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3470 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3471 ceph_cap_string(newcaps));
3472 /* non-auth MDS is revoking the newly grant caps ? */
3473 if (cap == ci->i_auth_cap &&
3474 __ceph_caps_revoking_other(ci, cap, newcaps))
3477 cap->issued = newcaps;
3478 cap->implemented |= newcaps; /* add bits only, to
3479 * avoid stepping on a
3480 * pending revocation */
3483 BUG_ON(cap->issued & ~cap->implemented);
3485 if (extra_info->inline_version > 0 &&
3486 extra_info->inline_version >= ci->i_inline_version) {
3487 ci->i_inline_version = extra_info->inline_version;
3488 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3489 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3493 if (ci->i_auth_cap == cap &&
3494 le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3495 if (newcaps & ~extra_info->issued)
3498 if (ci->i_requested_max_size > max_size ||
3499 !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3500 /* re-request max_size if necessary */
3501 ci->i_requested_max_size = 0;
3505 ceph_kick_flushing_inode_caps(session, ci);
3506 spin_unlock(&ci->i_ceph_lock);
3507 up_read(&session->s_mdsc->snap_rwsem);
3509 spin_unlock(&ci->i_ceph_lock);
3513 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3514 extra_info->inline_len);
3517 ceph_queue_vmtruncate(inode);
3521 * queue inode for writeback: we can't actually call
3522 * filemap_write_and_wait, etc. from message handler
3525 ceph_queue_writeback(inode);
3526 if (queue_invalidate)
3527 ceph_queue_invalidate(inode);
3529 invalidate_aliases(inode);
3531 wake_up_all(&ci->i_cap_wq);
3533 mutex_unlock(&session->s_mutex);
3534 if (check_caps == 1)
3535 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3537 else if (check_caps == 2)
3538 ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3542 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3543 * MDS has been safely committed.
3545 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3546 struct ceph_mds_caps *m,
3547 struct ceph_mds_session *session,
3548 struct ceph_cap *cap)
3549 __releases(ci->i_ceph_lock)
3551 struct ceph_inode_info *ci = ceph_inode(inode);
3552 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3553 struct ceph_cap_flush *cf, *tmp_cf;
3554 LIST_HEAD(to_remove);
3555 unsigned seq = le32_to_cpu(m->seq);
3556 int dirty = le32_to_cpu(m->dirty);
3559 bool wake_ci = false;
3560 bool wake_mdsc = false;
3562 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3563 /* Is this the one that was flushed? */
3564 if (cf->tid == flush_tid)
3567 /* Is this a capsnap? */
3571 if (cf->tid <= flush_tid) {
3573 * An earlier or current tid. The FLUSH_ACK should
3574 * represent a superset of this flush's caps.
3576 wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3577 list_add_tail(&cf->i_list, &to_remove);
3580 * This is a later one. Any caps in it are still dirty
3581 * so don't count them as cleaned.
3583 cleaned &= ~cf->caps;
3589 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3590 " flushing %s -> %s\n",
3591 inode, session->s_mds, seq, ceph_cap_string(dirty),
3592 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3593 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3595 if (list_empty(&to_remove) && !cleaned)
3598 ci->i_flushing_caps &= ~cleaned;
3600 spin_lock(&mdsc->cap_dirty_lock);
3602 list_for_each_entry(cf, &to_remove, i_list)
3603 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3605 if (ci->i_flushing_caps == 0) {
3606 if (list_empty(&ci->i_cap_flush_list)) {
3607 list_del_init(&ci->i_flushing_item);
3608 if (!list_empty(&session->s_cap_flushing)) {
3609 dout(" mds%d still flushing cap on %p\n",
3611 &list_first_entry(&session->s_cap_flushing,
3612 struct ceph_inode_info,
3613 i_flushing_item)->vfs_inode);
3616 mdsc->num_cap_flushing--;
3617 dout(" inode %p now !flushing\n", inode);
3619 if (ci->i_dirty_caps == 0) {
3620 dout(" inode %p now clean\n", inode);
3621 BUG_ON(!list_empty(&ci->i_dirty_item));
3623 if (ci->i_wr_ref == 0 &&
3624 ci->i_wrbuffer_ref_head == 0) {
3625 BUG_ON(!ci->i_head_snapc);
3626 ceph_put_snap_context(ci->i_head_snapc);
3627 ci->i_head_snapc = NULL;
3630 BUG_ON(list_empty(&ci->i_dirty_item));
3633 spin_unlock(&mdsc->cap_dirty_lock);
3636 spin_unlock(&ci->i_ceph_lock);
3638 while (!list_empty(&to_remove)) {
3639 cf = list_first_entry(&to_remove,
3640 struct ceph_cap_flush, i_list);
3641 list_del_init(&cf->i_list);
3642 if (!cf->is_capsnap)
3643 ceph_free_cap_flush(cf);
3647 wake_up_all(&ci->i_cap_wq);
3649 wake_up_all(&mdsc->cap_flushing_wq);
3655 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3656 * throw away our cap_snap.
3658 * Caller hold s_mutex.
3660 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3661 struct ceph_mds_caps *m,
3662 struct ceph_mds_session *session)
3664 struct ceph_inode_info *ci = ceph_inode(inode);
3665 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3666 u64 follows = le64_to_cpu(m->snap_follows);
3667 struct ceph_cap_snap *capsnap;
3668 bool flushed = false;
3669 bool wake_ci = false;
3670 bool wake_mdsc = false;
3672 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3673 inode, ci, session->s_mds, follows);
3675 spin_lock(&ci->i_ceph_lock);
3676 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3677 if (capsnap->follows == follows) {
3678 if (capsnap->cap_flush.tid != flush_tid) {
3679 dout(" cap_snap %p follows %lld tid %lld !="
3680 " %lld\n", capsnap, follows,
3681 flush_tid, capsnap->cap_flush.tid);
3687 dout(" skipping cap_snap %p follows %lld\n",
3688 capsnap, capsnap->follows);
3692 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3693 dout(" removing %p cap_snap %p follows %lld\n",
3694 inode, capsnap, follows);
3695 list_del(&capsnap->ci_item);
3696 wake_ci |= __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3698 spin_lock(&mdsc->cap_dirty_lock);
3700 if (list_empty(&ci->i_cap_flush_list))
3701 list_del_init(&ci->i_flushing_item);
3703 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc,
3704 &capsnap->cap_flush);
3705 spin_unlock(&mdsc->cap_dirty_lock);
3707 spin_unlock(&ci->i_ceph_lock);
3709 ceph_put_snap_context(capsnap->context);
3710 ceph_put_cap_snap(capsnap);
3712 wake_up_all(&ci->i_cap_wq);
3714 wake_up_all(&mdsc->cap_flushing_wq);
3720 * Handle TRUNC from MDS, indicating file truncation.
3722 * caller hold s_mutex.
3724 static bool handle_cap_trunc(struct inode *inode,
3725 struct ceph_mds_caps *trunc,
3726 struct ceph_mds_session *session)
3728 struct ceph_inode_info *ci = ceph_inode(inode);
3729 int mds = session->s_mds;
3730 int seq = le32_to_cpu(trunc->seq);
3731 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3732 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3733 u64 size = le64_to_cpu(trunc->size);
3734 int implemented = 0;
3735 int dirty = __ceph_caps_dirty(ci);
3736 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3737 bool queue_trunc = false;
3739 lockdep_assert_held(&ci->i_ceph_lock);
3741 issued |= implemented | dirty;
3743 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3744 inode, mds, seq, truncate_size, truncate_seq);
3745 queue_trunc = ceph_fill_file_size(inode, issued,
3746 truncate_seq, truncate_size, size);
3751 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3752 * different one. If we are the most recent migration we've seen (as
3753 * indicated by mseq), make note of the migrating cap bits for the
3754 * duration (until we see the corresponding IMPORT).
3756 * caller holds s_mutex
3758 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3759 struct ceph_mds_cap_peer *ph,
3760 struct ceph_mds_session *session)
3762 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3763 struct ceph_mds_session *tsession = NULL;
3764 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3765 struct ceph_inode_info *ci = ceph_inode(inode);
3767 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3768 unsigned t_seq, t_mseq;
3770 int mds = session->s_mds;
3773 t_cap_id = le64_to_cpu(ph->cap_id);
3774 t_seq = le32_to_cpu(ph->seq);
3775 t_mseq = le32_to_cpu(ph->mseq);
3776 target = le32_to_cpu(ph->mds);
3778 t_cap_id = t_seq = t_mseq = 0;
3782 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3783 inode, ci, mds, mseq, target);
3785 spin_lock(&ci->i_ceph_lock);
3786 cap = __get_cap_for_mds(ci, mds);
3787 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3791 __ceph_remove_cap(cap, false);
3796 * now we know we haven't received the cap import message yet
3797 * because the exported cap still exist.
3800 issued = cap->issued;
3801 if (issued != cap->implemented)
3802 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3803 "ino (%llx.%llx) mds%d seq %d mseq %d "
3804 "issued %s implemented %s\n",
3805 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3806 ceph_cap_string(issued),
3807 ceph_cap_string(cap->implemented));
3810 tcap = __get_cap_for_mds(ci, target);
3812 /* already have caps from the target */
3813 if (tcap->cap_id == t_cap_id &&
3814 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3815 dout(" updating import cap %p mds%d\n", tcap, target);
3816 tcap->cap_id = t_cap_id;
3817 tcap->seq = t_seq - 1;
3818 tcap->issue_seq = t_seq - 1;
3819 tcap->issued |= issued;
3820 tcap->implemented |= issued;
3821 if (cap == ci->i_auth_cap) {
3822 ci->i_auth_cap = tcap;
3823 change_auth_cap_ses(ci, tcap->session);
3826 __ceph_remove_cap(cap, false);
3828 } else if (tsession) {
3829 /* add placeholder for the export tagert */
3830 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3832 ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3833 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3835 if (!list_empty(&ci->i_cap_flush_list) &&
3836 ci->i_auth_cap == tcap) {
3837 spin_lock(&mdsc->cap_dirty_lock);
3838 list_move_tail(&ci->i_flushing_item,
3839 &tcap->session->s_cap_flushing);
3840 spin_unlock(&mdsc->cap_dirty_lock);
3843 __ceph_remove_cap(cap, false);
3847 spin_unlock(&ci->i_ceph_lock);
3848 mutex_unlock(&session->s_mutex);
3850 /* open target session */
3851 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3852 if (!IS_ERR(tsession)) {
3854 mutex_lock(&session->s_mutex);
3855 mutex_lock_nested(&tsession->s_mutex,
3856 SINGLE_DEPTH_NESTING);
3858 mutex_lock(&tsession->s_mutex);
3859 mutex_lock_nested(&session->s_mutex,
3860 SINGLE_DEPTH_NESTING);
3862 new_cap = ceph_get_cap(mdsc, NULL);
3867 mutex_lock(&session->s_mutex);
3872 spin_unlock(&ci->i_ceph_lock);
3873 mutex_unlock(&session->s_mutex);
3875 mutex_unlock(&tsession->s_mutex);
3876 ceph_put_mds_session(tsession);
3879 ceph_put_cap(mdsc, new_cap);
3883 * Handle cap IMPORT.
3885 * caller holds s_mutex. acquires i_ceph_lock
3887 static void handle_cap_import(struct ceph_mds_client *mdsc,
3888 struct inode *inode, struct ceph_mds_caps *im,
3889 struct ceph_mds_cap_peer *ph,
3890 struct ceph_mds_session *session,
3891 struct ceph_cap **target_cap, int *old_issued)
3893 struct ceph_inode_info *ci = ceph_inode(inode);
3894 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3895 int mds = session->s_mds;
3897 unsigned caps = le32_to_cpu(im->caps);
3898 unsigned wanted = le32_to_cpu(im->wanted);
3899 unsigned seq = le32_to_cpu(im->seq);
3900 unsigned mseq = le32_to_cpu(im->migrate_seq);
3901 u64 realmino = le64_to_cpu(im->realm);
3902 u64 cap_id = le64_to_cpu(im->cap_id);
3907 p_cap_id = le64_to_cpu(ph->cap_id);
3908 peer = le32_to_cpu(ph->mds);
3914 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3915 inode, ci, mds, mseq, peer);
3917 cap = __get_cap_for_mds(ci, mds);
3920 spin_unlock(&ci->i_ceph_lock);
3921 new_cap = ceph_get_cap(mdsc, NULL);
3922 spin_lock(&ci->i_ceph_lock);
3928 ceph_put_cap(mdsc, new_cap);
3933 __ceph_caps_issued(ci, &issued);
3934 issued |= __ceph_caps_dirty(ci);
3936 ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
3937 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3939 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3940 if (ocap && ocap->cap_id == p_cap_id) {
3941 dout(" remove export cap %p mds%d flags %d\n",
3942 ocap, peer, ph->flags);
3943 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3944 (ocap->seq != le32_to_cpu(ph->seq) ||
3945 ocap->mseq != le32_to_cpu(ph->mseq))) {
3946 pr_err_ratelimited("handle_cap_import: "
3947 "mismatched seq/mseq: ino (%llx.%llx) "
3948 "mds%d seq %d mseq %d importer mds%d "
3949 "has peer seq %d mseq %d\n",
3950 ceph_vinop(inode), peer, ocap->seq,
3951 ocap->mseq, mds, le32_to_cpu(ph->seq),
3952 le32_to_cpu(ph->mseq));
3954 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3957 *old_issued = issued;
3962 * Handle a caps message from the MDS.
3964 * Identify the appropriate session, inode, and call the right handler
3965 * based on the cap op.
3967 void ceph_handle_caps(struct ceph_mds_session *session,
3968 struct ceph_msg *msg)
3970 struct ceph_mds_client *mdsc = session->s_mdsc;
3971 struct inode *inode;
3972 struct ceph_inode_info *ci;
3973 struct ceph_cap *cap;
3974 struct ceph_mds_caps *h;
3975 struct ceph_mds_cap_peer *peer = NULL;
3976 struct ceph_snap_realm *realm = NULL;
3978 int msg_version = le16_to_cpu(msg->hdr.version);
3980 struct ceph_vino vino;
3982 size_t snaptrace_len;
3984 struct cap_extra_info extra_info = {};
3987 dout("handle_caps from mds%d\n", session->s_mds);
3990 end = msg->front.iov_base + msg->front.iov_len;
3991 if (msg->front.iov_len < sizeof(*h))
3993 h = msg->front.iov_base;
3994 op = le32_to_cpu(h->op);
3995 vino.ino = le64_to_cpu(h->ino);
3996 vino.snap = CEPH_NOSNAP;
3997 seq = le32_to_cpu(h->seq);
3998 mseq = le32_to_cpu(h->migrate_seq);
4001 snaptrace_len = le32_to_cpu(h->snap_trace_len);
4002 p = snaptrace + snaptrace_len;
4004 if (msg_version >= 2) {
4006 ceph_decode_32_safe(&p, end, flock_len, bad);
4007 if (p + flock_len > end)
4012 if (msg_version >= 3) {
4013 if (op == CEPH_CAP_OP_IMPORT) {
4014 if (p + sizeof(*peer) > end)
4018 } else if (op == CEPH_CAP_OP_EXPORT) {
4019 /* recorded in unused fields */
4020 peer = (void *)&h->size;
4024 if (msg_version >= 4) {
4025 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
4026 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
4027 if (p + extra_info.inline_len > end)
4029 extra_info.inline_data = p;
4030 p += extra_info.inline_len;
4033 if (msg_version >= 5) {
4034 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
4037 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
4038 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
4041 if (msg_version >= 8) {
4045 ceph_decode_skip_64(&p, end, bad); // flush_tid
4047 ceph_decode_skip_32(&p, end, bad); // caller_uid
4048 ceph_decode_skip_32(&p, end, bad); // caller_gid
4050 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4051 if (pool_ns_len > 0) {
4052 ceph_decode_need(&p, end, pool_ns_len, bad);
4053 extra_info.pool_ns =
4054 ceph_find_or_create_string(p, pool_ns_len);
4059 if (msg_version >= 9) {
4060 struct ceph_timespec *btime;
4062 if (p + sizeof(*btime) > end)
4065 ceph_decode_timespec64(&extra_info.btime, btime);
4066 p += sizeof(*btime);
4067 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4070 if (msg_version >= 11) {
4072 ceph_decode_skip_32(&p, end, bad); // flags
4074 extra_info.dirstat_valid = true;
4075 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4076 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4080 inode = ceph_find_inode(mdsc->fsc->sb, vino);
4081 ci = ceph_inode(inode);
4082 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4085 mutex_lock(&session->s_mutex);
4086 inc_session_sequence(session);
4087 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4091 dout(" i don't have ino %llx\n", vino.ino);
4093 if (op == CEPH_CAP_OP_IMPORT) {
4094 cap = ceph_get_cap(mdsc, NULL);
4095 cap->cap_ino = vino.ino;
4096 cap->queue_release = 1;
4097 cap->cap_id = le64_to_cpu(h->cap_id);
4100 cap->issue_seq = seq;
4101 spin_lock(&session->s_cap_lock);
4102 __ceph_queue_cap_release(session, cap);
4103 spin_unlock(&session->s_cap_lock);
4105 goto flush_cap_releases;
4108 /* these will work even if we don't have a cap yet */
4110 case CEPH_CAP_OP_FLUSHSNAP_ACK:
4111 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4115 case CEPH_CAP_OP_EXPORT:
4116 handle_cap_export(inode, h, peer, session);
4119 case CEPH_CAP_OP_IMPORT:
4121 if (snaptrace_len) {
4122 down_write(&mdsc->snap_rwsem);
4123 ceph_update_snap_trace(mdsc, snaptrace,
4124 snaptrace + snaptrace_len,
4126 downgrade_write(&mdsc->snap_rwsem);
4128 down_read(&mdsc->snap_rwsem);
4130 spin_lock(&ci->i_ceph_lock);
4131 handle_cap_import(mdsc, inode, h, peer, session,
4132 &cap, &extra_info.issued);
4133 handle_cap_grant(inode, session, cap,
4134 h, msg->middle, &extra_info);
4136 ceph_put_snap_realm(mdsc, realm);
4140 /* the rest require a cap */
4141 spin_lock(&ci->i_ceph_lock);
4142 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4144 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4145 inode, ceph_ino(inode), ceph_snap(inode),
4147 spin_unlock(&ci->i_ceph_lock);
4148 goto flush_cap_releases;
4151 /* note that each of these drops i_ceph_lock for us */
4153 case CEPH_CAP_OP_REVOKE:
4154 case CEPH_CAP_OP_GRANT:
4155 __ceph_caps_issued(ci, &extra_info.issued);
4156 extra_info.issued |= __ceph_caps_dirty(ci);
4157 handle_cap_grant(inode, session, cap,
4158 h, msg->middle, &extra_info);
4161 case CEPH_CAP_OP_FLUSH_ACK:
4162 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4166 case CEPH_CAP_OP_TRUNC:
4167 queue_trunc = handle_cap_trunc(inode, h, session);
4168 spin_unlock(&ci->i_ceph_lock);
4170 ceph_queue_vmtruncate(inode);
4174 spin_unlock(&ci->i_ceph_lock);
4175 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4176 ceph_cap_op_name(op));
4180 mutex_unlock(&session->s_mutex);
4184 ceph_put_string(extra_info.pool_ns);
4189 * send any cap release message to try to move things
4190 * along for the mds (who clearly thinks we still have this
4193 ceph_flush_cap_releases(mdsc, session);
4197 pr_err("ceph_handle_caps: corrupt message\n");
4203 * Delayed work handler to process end of delayed cap release LRU list.
4205 * If new caps are added to the list while processing it, these won't get
4206 * processed in this run. In this case, the ci->i_hold_caps_max will be
4207 * returned so that the work can be scheduled accordingly.
4209 unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4211 struct inode *inode;
4212 struct ceph_inode_info *ci;
4213 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
4214 unsigned long delay_max = opt->caps_wanted_delay_max * HZ;
4215 unsigned long loop_start = jiffies;
4216 unsigned long delay = 0;
4218 dout("check_delayed_caps\n");
4219 spin_lock(&mdsc->cap_delay_lock);
4220 while (!list_empty(&mdsc->cap_delay_list)) {
4221 ci = list_first_entry(&mdsc->cap_delay_list,
4222 struct ceph_inode_info,
4224 if (time_before(loop_start, ci->i_hold_caps_max - delay_max)) {
4225 dout("%s caps added recently. Exiting loop", __func__);
4226 delay = ci->i_hold_caps_max;
4229 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4230 time_before(jiffies, ci->i_hold_caps_max))
4232 list_del_init(&ci->i_cap_delay_list);
4234 inode = igrab(&ci->vfs_inode);
4236 spin_unlock(&mdsc->cap_delay_lock);
4237 dout("check_delayed_caps on %p\n", inode);
4238 ceph_check_caps(ci, 0, NULL);
4240 spin_lock(&mdsc->cap_delay_lock);
4243 spin_unlock(&mdsc->cap_delay_lock);
4249 * Flush all dirty caps to the mds
4251 static void flush_dirty_session_caps(struct ceph_mds_session *s)
4253 struct ceph_mds_client *mdsc = s->s_mdsc;
4254 struct ceph_inode_info *ci;
4255 struct inode *inode;
4257 dout("flush_dirty_caps\n");
4258 spin_lock(&mdsc->cap_dirty_lock);
4259 while (!list_empty(&s->s_cap_dirty)) {
4260 ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4262 inode = &ci->vfs_inode;
4264 dout("flush_dirty_caps %p\n", inode);
4265 spin_unlock(&mdsc->cap_dirty_lock);
4266 ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4268 spin_lock(&mdsc->cap_dirty_lock);
4270 spin_unlock(&mdsc->cap_dirty_lock);
4271 dout("flush_dirty_caps done\n");
4274 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4276 ceph_mdsc_iterate_sessions(mdsc, flush_dirty_session_caps, true);
4279 void __ceph_touch_fmode(struct ceph_inode_info *ci,
4280 struct ceph_mds_client *mdsc, int fmode)
4282 unsigned long now = jiffies;
4283 if (fmode & CEPH_FILE_MODE_RD)
4284 ci->i_last_rd = now;
4285 if (fmode & CEPH_FILE_MODE_WR)
4286 ci->i_last_wr = now;
4287 /* queue periodic check */
4289 __ceph_is_any_real_caps(ci) &&
4290 list_empty(&ci->i_cap_delay_list))
4291 __cap_delay_requeue(mdsc, ci);
4294 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4296 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4297 int bits = (fmode << 1) | 1;
4298 bool is_opened = false;
4302 atomic64_inc(&mdsc->metric.opened_files);
4304 spin_lock(&ci->i_ceph_lock);
4305 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4306 if (bits & (1 << i))
4307 ci->i_nr_by_mode[i] += count;
4310 * If any of the mode ref is larger than 1,
4311 * that means it has been already opened by
4312 * others. Just skip checking the PIN ref.
4314 if (i && ci->i_nr_by_mode[i] > 1)
4319 percpu_counter_inc(&mdsc->metric.opened_inodes);
4320 spin_unlock(&ci->i_ceph_lock);
4324 * Drop open file reference. If we were the last open file,
4325 * we may need to release capabilities to the MDS (or schedule
4326 * their delayed release).
4328 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4330 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4331 int bits = (fmode << 1) | 1;
4332 bool is_closed = true;
4336 atomic64_dec(&mdsc->metric.opened_files);
4338 spin_lock(&ci->i_ceph_lock);
4339 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4340 if (bits & (1 << i)) {
4341 BUG_ON(ci->i_nr_by_mode[i] < count);
4342 ci->i_nr_by_mode[i] -= count;
4346 * If any of the mode ref is not 0 after
4347 * decreased, that means it is still opened
4348 * by others. Just skip checking the PIN ref.
4350 if (i && ci->i_nr_by_mode[i])
4355 percpu_counter_dec(&mdsc->metric.opened_inodes);
4356 spin_unlock(&ci->i_ceph_lock);
4360 * For a soon-to-be unlinked file, drop the LINK caps. If it
4361 * looks like the link count will hit 0, drop any other caps (other
4362 * than PIN) we don't specifically want (due to the file still being
4365 int ceph_drop_caps_for_unlink(struct inode *inode)
4367 struct ceph_inode_info *ci = ceph_inode(inode);
4368 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4370 spin_lock(&ci->i_ceph_lock);
4371 if (inode->i_nlink == 1) {
4372 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4374 if (__ceph_caps_dirty(ci)) {
4375 struct ceph_mds_client *mdsc =
4376 ceph_inode_to_client(inode)->mdsc;
4377 __cap_delay_requeue_front(mdsc, ci);
4380 spin_unlock(&ci->i_ceph_lock);
4385 * Helpers for embedding cap and dentry lease releases into mds
4388 * @force is used by dentry_release (below) to force inclusion of a
4389 * record for the directory inode, even when there aren't any caps to
4392 int ceph_encode_inode_release(void **p, struct inode *inode,
4393 int mds, int drop, int unless, int force)
4395 struct ceph_inode_info *ci = ceph_inode(inode);
4396 struct ceph_cap *cap;
4397 struct ceph_mds_request_release *rel = *p;
4401 spin_lock(&ci->i_ceph_lock);
4402 used = __ceph_caps_used(ci);
4403 dirty = __ceph_caps_dirty(ci);
4405 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4406 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4407 ceph_cap_string(unless));
4409 /* only drop unused, clean caps */
4410 drop &= ~(used | dirty);
4412 cap = __get_cap_for_mds(ci, mds);
4413 if (cap && __cap_is_valid(cap)) {
4414 unless &= cap->issued;
4416 if (unless & CEPH_CAP_AUTH_EXCL)
4417 drop &= ~CEPH_CAP_AUTH_SHARED;
4418 if (unless & CEPH_CAP_LINK_EXCL)
4419 drop &= ~CEPH_CAP_LINK_SHARED;
4420 if (unless & CEPH_CAP_XATTR_EXCL)
4421 drop &= ~CEPH_CAP_XATTR_SHARED;
4422 if (unless & CEPH_CAP_FILE_EXCL)
4423 drop &= ~CEPH_CAP_FILE_SHARED;
4426 if (force || (cap->issued & drop)) {
4427 if (cap->issued & drop) {
4428 int wanted = __ceph_caps_wanted(ci);
4429 dout("encode_inode_release %p cap %p "
4430 "%s -> %s, wanted %s -> %s\n", inode, cap,
4431 ceph_cap_string(cap->issued),
4432 ceph_cap_string(cap->issued & ~drop),
4433 ceph_cap_string(cap->mds_wanted),
4434 ceph_cap_string(wanted));
4436 cap->issued &= ~drop;
4437 cap->implemented &= ~drop;
4438 cap->mds_wanted = wanted;
4439 if (cap == ci->i_auth_cap &&
4440 !(wanted & CEPH_CAP_ANY_FILE_WR))
4441 ci->i_requested_max_size = 0;
4443 dout("encode_inode_release %p cap %p %s"
4444 " (force)\n", inode, cap,
4445 ceph_cap_string(cap->issued));
4448 rel->ino = cpu_to_le64(ceph_ino(inode));
4449 rel->cap_id = cpu_to_le64(cap->cap_id);
4450 rel->seq = cpu_to_le32(cap->seq);
4451 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4452 rel->mseq = cpu_to_le32(cap->mseq);
4453 rel->caps = cpu_to_le32(cap->implemented);
4454 rel->wanted = cpu_to_le32(cap->mds_wanted);
4460 dout("encode_inode_release %p cap %p %s (noop)\n",
4461 inode, cap, ceph_cap_string(cap->issued));
4464 spin_unlock(&ci->i_ceph_lock);
4468 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4470 int mds, int drop, int unless)
4472 struct dentry *parent = NULL;
4473 struct ceph_mds_request_release *rel = *p;
4474 struct ceph_dentry_info *di = ceph_dentry(dentry);
4479 * force an record for the directory caps if we have a dentry lease.
4480 * this is racy (can't take i_ceph_lock and d_lock together), but it
4481 * doesn't have to be perfect; the mds will revoke anything we don't
4484 spin_lock(&dentry->d_lock);
4485 if (di->lease_session && di->lease_session->s_mds == mds)
4488 parent = dget(dentry->d_parent);
4489 dir = d_inode(parent);
4491 spin_unlock(&dentry->d_lock);
4493 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4496 spin_lock(&dentry->d_lock);
4497 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4498 dout("encode_dentry_release %p mds%d seq %d\n",
4499 dentry, mds, (int)di->lease_seq);
4500 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4501 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4502 *p += dentry->d_name.len;
4503 rel->dname_seq = cpu_to_le32(di->lease_seq);
4504 __ceph_mdsc_drop_dentry_lease(dentry);
4506 spin_unlock(&dentry->d_lock);