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 struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
709 spin_lock(&oldrealm->inodes_with_caps_lock);
710 list_del_init(&ci->i_snap_realm_item);
711 spin_unlock(&oldrealm->inodes_with_caps_lock);
714 spin_lock(&realm->inodes_with_caps_lock);
715 list_add(&ci->i_snap_realm_item,
716 &realm->inodes_with_caps);
717 ci->i_snap_realm = realm;
718 if (realm->ino == ci->i_vino.ino)
719 realm->inode = inode;
720 spin_unlock(&realm->inodes_with_caps_lock);
723 ceph_put_snap_realm(mdsc, oldrealm);
725 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
731 __check_cap_issue(ci, cap, issued);
734 * If we are issued caps we don't want, or the mds' wanted
735 * value appears to be off, queue a check so we'll release
736 * later and/or update the mds wanted value.
738 actual_wanted = __ceph_caps_wanted(ci);
739 if ((wanted & ~actual_wanted) ||
740 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
741 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
742 ceph_cap_string(issued), ceph_cap_string(wanted),
743 ceph_cap_string(actual_wanted));
744 __cap_delay_requeue(mdsc, ci);
747 if (flags & CEPH_CAP_FLAG_AUTH) {
748 if (!ci->i_auth_cap ||
749 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
750 if (ci->i_auth_cap &&
751 ci->i_auth_cap->session != cap->session)
752 change_auth_cap_ses(ci, cap->session);
753 ci->i_auth_cap = cap;
754 cap->mds_wanted = wanted;
757 WARN_ON(ci->i_auth_cap == cap);
760 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
761 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
762 ceph_cap_string(issued|cap->issued), seq, mds);
763 cap->cap_id = cap_id;
764 cap->issued = issued;
765 cap->implemented |= issued;
766 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
767 cap->mds_wanted = wanted;
769 cap->mds_wanted |= wanted;
771 cap->issue_seq = seq;
777 * Return true if cap has not timed out and belongs to the current
778 * generation of the MDS session (i.e. has not gone 'stale' due to
779 * us losing touch with the mds).
781 static int __cap_is_valid(struct ceph_cap *cap)
786 gen = atomic_read(&cap->session->s_cap_gen);
787 ttl = cap->session->s_cap_ttl;
789 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
790 dout("__cap_is_valid %p cap %p issued %s "
791 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
792 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
800 * Return set of valid cap bits issued to us. Note that caps time
801 * out, and may be invalidated in bulk if the client session times out
802 * and session->s_cap_gen is bumped.
804 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
806 int have = ci->i_snap_caps;
807 struct ceph_cap *cap;
812 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
813 cap = rb_entry(p, struct ceph_cap, ci_node);
814 if (!__cap_is_valid(cap))
816 dout("__ceph_caps_issued %p cap %p issued %s\n",
817 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
820 *implemented |= cap->implemented;
823 * exclude caps issued by non-auth MDS, but are been revoking
824 * by the auth MDS. The non-auth MDS should be revoking/exporting
825 * these caps, but the message is delayed.
827 if (ci->i_auth_cap) {
828 cap = ci->i_auth_cap;
829 have &= ~cap->implemented | cap->issued;
835 * Get cap bits issued by caps other than @ocap
837 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
839 int have = ci->i_snap_caps;
840 struct ceph_cap *cap;
843 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
844 cap = rb_entry(p, struct ceph_cap, ci_node);
847 if (!__cap_is_valid(cap))
855 * Move a cap to the end of the LRU (oldest caps at list head, newest
858 static void __touch_cap(struct ceph_cap *cap)
860 struct ceph_mds_session *s = cap->session;
862 spin_lock(&s->s_cap_lock);
863 if (!s->s_cap_iterator) {
864 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
866 list_move_tail(&cap->session_caps, &s->s_caps);
868 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
869 &cap->ci->vfs_inode, cap, s->s_mds);
871 spin_unlock(&s->s_cap_lock);
875 * Check if we hold the given mask. If so, move the cap(s) to the
876 * front of their respective LRUs. (This is the preferred way for
877 * callers to check for caps they want.)
879 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
881 struct ceph_cap *cap;
883 int have = ci->i_snap_caps;
885 if ((have & mask) == mask) {
886 dout("__ceph_caps_issued_mask ino 0x%llx snap issued %s"
887 " (mask %s)\n", ceph_ino(&ci->vfs_inode),
888 ceph_cap_string(have),
889 ceph_cap_string(mask));
893 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
894 cap = rb_entry(p, struct ceph_cap, ci_node);
895 if (!__cap_is_valid(cap))
897 if ((cap->issued & mask) == mask) {
898 dout("__ceph_caps_issued_mask ino 0x%llx cap %p issued %s"
899 " (mask %s)\n", ceph_ino(&ci->vfs_inode), cap,
900 ceph_cap_string(cap->issued),
901 ceph_cap_string(mask));
907 /* does a combination of caps satisfy mask? */
909 if ((have & mask) == mask) {
910 dout("__ceph_caps_issued_mask ino 0x%llx combo issued %s"
911 " (mask %s)\n", ceph_ino(&ci->vfs_inode),
912 ceph_cap_string(cap->issued),
913 ceph_cap_string(mask));
917 /* touch this + preceding caps */
919 for (q = rb_first(&ci->i_caps); q != p;
921 cap = rb_entry(q, struct ceph_cap,
923 if (!__cap_is_valid(cap))
925 if (cap->issued & mask)
936 int __ceph_caps_issued_mask_metric(struct ceph_inode_info *ci, int mask,
939 struct ceph_fs_client *fsc = ceph_sb_to_client(ci->vfs_inode.i_sb);
942 r = __ceph_caps_issued_mask(ci, mask, touch);
944 ceph_update_cap_hit(&fsc->mdsc->metric);
946 ceph_update_cap_mis(&fsc->mdsc->metric);
951 * Return true if mask caps are currently being revoked by an MDS.
953 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
954 struct ceph_cap *ocap, int mask)
956 struct ceph_cap *cap;
959 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
960 cap = rb_entry(p, struct ceph_cap, ci_node);
962 (cap->implemented & ~cap->issued & mask))
968 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
970 struct inode *inode = &ci->vfs_inode;
973 spin_lock(&ci->i_ceph_lock);
974 ret = __ceph_caps_revoking_other(ci, NULL, mask);
975 spin_unlock(&ci->i_ceph_lock);
976 dout("ceph_caps_revoking %p %s = %d\n", inode,
977 ceph_cap_string(mask), ret);
981 int __ceph_caps_used(struct ceph_inode_info *ci)
985 used |= CEPH_CAP_PIN;
987 used |= CEPH_CAP_FILE_RD;
988 if (ci->i_rdcache_ref ||
989 (S_ISREG(ci->vfs_inode.i_mode) &&
990 ci->vfs_inode.i_data.nrpages))
991 used |= CEPH_CAP_FILE_CACHE;
993 used |= CEPH_CAP_FILE_WR;
994 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
995 used |= CEPH_CAP_FILE_BUFFER;
997 used |= CEPH_CAP_FILE_EXCL;
1001 #define FMODE_WAIT_BIAS 1000
1004 * wanted, by virtue of open file modes
1006 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
1008 const int PIN_SHIFT = ffs(CEPH_FILE_MODE_PIN);
1009 const int RD_SHIFT = ffs(CEPH_FILE_MODE_RD);
1010 const int WR_SHIFT = ffs(CEPH_FILE_MODE_WR);
1011 const int LAZY_SHIFT = ffs(CEPH_FILE_MODE_LAZY);
1012 struct ceph_mount_options *opt =
1013 ceph_inode_to_client(&ci->vfs_inode)->mount_options;
1014 unsigned long used_cutoff = jiffies - opt->caps_wanted_delay_max * HZ;
1015 unsigned long idle_cutoff = jiffies - opt->caps_wanted_delay_min * HZ;
1017 if (S_ISDIR(ci->vfs_inode.i_mode)) {
1020 /* use used_cutoff here, to keep dir's wanted caps longer */
1021 if (ci->i_nr_by_mode[RD_SHIFT] > 0 ||
1022 time_after(ci->i_last_rd, used_cutoff))
1023 want |= CEPH_CAP_ANY_SHARED;
1025 if (ci->i_nr_by_mode[WR_SHIFT] > 0 ||
1026 time_after(ci->i_last_wr, used_cutoff)) {
1027 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1028 if (opt->flags & CEPH_MOUNT_OPT_ASYNC_DIROPS)
1029 want |= CEPH_CAP_ANY_DIR_OPS;
1032 if (want || ci->i_nr_by_mode[PIN_SHIFT] > 0)
1033 want |= CEPH_CAP_PIN;
1039 if (ci->i_nr_by_mode[RD_SHIFT] > 0) {
1040 if (ci->i_nr_by_mode[RD_SHIFT] >= FMODE_WAIT_BIAS ||
1041 time_after(ci->i_last_rd, used_cutoff))
1042 bits |= 1 << RD_SHIFT;
1043 } else if (time_after(ci->i_last_rd, idle_cutoff)) {
1044 bits |= 1 << RD_SHIFT;
1047 if (ci->i_nr_by_mode[WR_SHIFT] > 0) {
1048 if (ci->i_nr_by_mode[WR_SHIFT] >= FMODE_WAIT_BIAS ||
1049 time_after(ci->i_last_wr, used_cutoff))
1050 bits |= 1 << WR_SHIFT;
1051 } else if (time_after(ci->i_last_wr, idle_cutoff)) {
1052 bits |= 1 << WR_SHIFT;
1055 /* check lazyio only when read/write is wanted */
1056 if ((bits & (CEPH_FILE_MODE_RDWR << 1)) &&
1057 ci->i_nr_by_mode[LAZY_SHIFT] > 0)
1058 bits |= 1 << LAZY_SHIFT;
1060 return bits ? ceph_caps_for_mode(bits >> 1) : 0;
1065 * wanted, by virtue of open file modes AND cap refs (buffered/cached data)
1067 int __ceph_caps_wanted(struct ceph_inode_info *ci)
1069 int w = __ceph_caps_file_wanted(ci) | __ceph_caps_used(ci);
1070 if (S_ISDIR(ci->vfs_inode.i_mode)) {
1071 /* we want EXCL if holding caps of dir ops */
1072 if (w & CEPH_CAP_ANY_DIR_OPS)
1073 w |= CEPH_CAP_FILE_EXCL;
1075 /* we want EXCL if dirty data */
1076 if (w & CEPH_CAP_FILE_BUFFER)
1077 w |= CEPH_CAP_FILE_EXCL;
1083 * Return caps we have registered with the MDS(s) as 'wanted'.
1085 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
1087 struct ceph_cap *cap;
1091 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1092 cap = rb_entry(p, struct ceph_cap, ci_node);
1093 if (check && !__cap_is_valid(cap))
1095 if (cap == ci->i_auth_cap)
1096 mds_wanted |= cap->mds_wanted;
1098 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
1103 int ceph_is_any_caps(struct inode *inode)
1105 struct ceph_inode_info *ci = ceph_inode(inode);
1108 spin_lock(&ci->i_ceph_lock);
1109 ret = __ceph_is_any_real_caps(ci);
1110 spin_unlock(&ci->i_ceph_lock);
1115 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
1117 struct ceph_snap_realm *realm = ci->i_snap_realm;
1118 spin_lock(&realm->inodes_with_caps_lock);
1119 list_del_init(&ci->i_snap_realm_item);
1120 ci->i_snap_realm_counter++;
1121 ci->i_snap_realm = NULL;
1122 if (realm->ino == ci->i_vino.ino)
1123 realm->inode = NULL;
1124 spin_unlock(&realm->inodes_with_caps_lock);
1125 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
1130 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
1132 * caller should hold i_ceph_lock.
1133 * caller will not hold session s_mutex if called from destroy_inode.
1135 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1137 struct ceph_mds_session *session = cap->session;
1138 struct ceph_inode_info *ci = cap->ci;
1139 struct ceph_mds_client *mdsc;
1142 /* 'ci' being NULL means the remove have already occurred */
1144 dout("%s: cap inode is NULL\n", __func__);
1148 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1150 mdsc = ceph_inode_to_client(&ci->vfs_inode)->mdsc;
1152 /* remove from inode's cap rbtree, and clear auth cap */
1153 rb_erase(&cap->ci_node, &ci->i_caps);
1154 if (ci->i_auth_cap == cap) {
1155 WARN_ON_ONCE(!list_empty(&ci->i_dirty_item) &&
1156 !mdsc->fsc->blocklisted);
1157 ci->i_auth_cap = NULL;
1160 /* remove from session list */
1161 spin_lock(&session->s_cap_lock);
1162 if (session->s_cap_iterator == cap) {
1163 /* not yet, we are iterating over this very cap */
1164 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1167 list_del_init(&cap->session_caps);
1168 session->s_nr_caps--;
1169 atomic64_dec(&mdsc->metric.total_caps);
1170 cap->session = NULL;
1173 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1177 * s_cap_reconnect is protected by s_cap_lock. no one changes
1178 * s_cap_gen while session is in the reconnect state.
1180 if (queue_release &&
1181 (!session->s_cap_reconnect ||
1182 cap->cap_gen == atomic_read(&session->s_cap_gen))) {
1183 cap->queue_release = 1;
1185 __ceph_queue_cap_release(session, cap);
1189 cap->queue_release = 0;
1191 cap->cap_ino = ci->i_vino.ino;
1193 spin_unlock(&session->s_cap_lock);
1196 ceph_put_cap(mdsc, cap);
1198 if (!__ceph_is_any_real_caps(ci)) {
1199 /* when reconnect denied, we remove session caps forcibly,
1200 * i_wr_ref can be non-zero. If there are ongoing write,
1201 * keep i_snap_realm.
1203 if (ci->i_wr_ref == 0 && ci->i_snap_realm)
1204 drop_inode_snap_realm(ci);
1206 __cap_delay_cancel(mdsc, ci);
1210 struct cap_msg_args {
1211 struct ceph_mds_session *session;
1212 u64 ino, cid, follows;
1213 u64 flush_tid, oldest_flush_tid, size, max_size;
1216 struct ceph_buffer *xattr_buf;
1217 struct ceph_buffer *old_xattr_buf;
1218 struct timespec64 atime, mtime, ctime, btime;
1219 int op, caps, wanted, dirty;
1220 u32 seq, issue_seq, mseq, time_warp_seq;
1230 * cap struct size + flock buffer size + inline version + inline data size +
1231 * osd_epoch_barrier + oldest_flush_tid
1233 #define CAP_MSG_SIZE (sizeof(struct ceph_mds_caps) + \
1234 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4)
1236 /* Marshal up the cap msg to the MDS */
1237 static void encode_cap_msg(struct ceph_msg *msg, struct cap_msg_args *arg)
1239 struct ceph_mds_caps *fc;
1241 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1243 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",
1244 __func__, ceph_cap_op_name(arg->op), arg->cid, arg->ino,
1245 ceph_cap_string(arg->caps), ceph_cap_string(arg->wanted),
1246 ceph_cap_string(arg->dirty), arg->seq, arg->issue_seq,
1247 arg->flush_tid, arg->oldest_flush_tid, arg->mseq, arg->follows,
1248 arg->size, arg->max_size, arg->xattr_version,
1249 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1251 msg->hdr.version = cpu_to_le16(10);
1252 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1254 fc = msg->front.iov_base;
1255 memset(fc, 0, sizeof(*fc));
1257 fc->cap_id = cpu_to_le64(arg->cid);
1258 fc->op = cpu_to_le32(arg->op);
1259 fc->seq = cpu_to_le32(arg->seq);
1260 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1261 fc->migrate_seq = cpu_to_le32(arg->mseq);
1262 fc->caps = cpu_to_le32(arg->caps);
1263 fc->wanted = cpu_to_le32(arg->wanted);
1264 fc->dirty = cpu_to_le32(arg->dirty);
1265 fc->ino = cpu_to_le64(arg->ino);
1266 fc->snap_follows = cpu_to_le64(arg->follows);
1268 fc->size = cpu_to_le64(arg->size);
1269 fc->max_size = cpu_to_le64(arg->max_size);
1270 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1271 ceph_encode_timespec64(&fc->atime, &arg->atime);
1272 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1273 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1275 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1276 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1277 fc->mode = cpu_to_le32(arg->mode);
1279 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1280 if (arg->xattr_buf) {
1281 msg->middle = ceph_buffer_get(arg->xattr_buf);
1282 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1283 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1287 /* flock buffer size (version 2) */
1288 ceph_encode_32(&p, 0);
1289 /* inline version (version 4) */
1290 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1291 /* inline data size */
1292 ceph_encode_32(&p, 0);
1294 * osd_epoch_barrier (version 5)
1295 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1296 * case it was recently changed
1298 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1299 /* oldest_flush_tid (version 6) */
1300 ceph_encode_64(&p, arg->oldest_flush_tid);
1303 * caller_uid/caller_gid (version 7)
1305 * Currently, we don't properly track which caller dirtied the caps
1306 * last, and force a flush of them when there is a conflict. For now,
1307 * just set this to 0:0, to emulate how the MDS has worked up to now.
1309 ceph_encode_32(&p, 0);
1310 ceph_encode_32(&p, 0);
1312 /* pool namespace (version 8) (mds always ignores this) */
1313 ceph_encode_32(&p, 0);
1315 /* btime and change_attr (version 9) */
1316 ceph_encode_timespec64(p, &arg->btime);
1317 p += sizeof(struct ceph_timespec);
1318 ceph_encode_64(&p, arg->change_attr);
1320 /* Advisory flags (version 10) */
1321 ceph_encode_32(&p, arg->flags);
1325 * Queue cap releases when an inode is dropped from our cache.
1327 void __ceph_remove_caps(struct ceph_inode_info *ci)
1331 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1332 * may call __ceph_caps_issued_mask() on a freeing inode. */
1333 spin_lock(&ci->i_ceph_lock);
1334 p = rb_first(&ci->i_caps);
1336 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1338 __ceph_remove_cap(cap, true);
1340 spin_unlock(&ci->i_ceph_lock);
1344 * Prepare to send a cap message to an MDS. Update the cap state, and populate
1345 * the arg struct with the parameters that will need to be sent. This should
1346 * be done under the i_ceph_lock to guard against changes to cap state.
1348 * Make note of max_size reported/requested from mds, revoked caps
1349 * that have now been implemented.
1351 static void __prep_cap(struct cap_msg_args *arg, struct ceph_cap *cap,
1352 int op, int flags, int used, int want, int retain,
1353 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1355 struct ceph_inode_info *ci = cap->ci;
1356 struct inode *inode = &ci->vfs_inode;
1359 lockdep_assert_held(&ci->i_ceph_lock);
1361 held = cap->issued | cap->implemented;
1362 revoking = cap->implemented & ~cap->issued;
1363 retain &= ~revoking;
1365 dout("%s %p cap %p session %p %s -> %s (revoking %s)\n",
1366 __func__, inode, cap, cap->session,
1367 ceph_cap_string(held), ceph_cap_string(held & retain),
1368 ceph_cap_string(revoking));
1369 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1371 ci->i_ceph_flags &= ~CEPH_I_FLUSH;
1373 cap->issued &= retain; /* drop bits we don't want */
1375 * Wake up any waiters on wanted -> needed transition. This is due to
1376 * the weird transition from buffered to sync IO... we need to flush
1377 * dirty pages _before_ allowing sync writes to avoid reordering.
1379 arg->wake = cap->implemented & ~cap->issued;
1380 cap->implemented &= cap->issued | used;
1381 cap->mds_wanted = want;
1383 arg->session = cap->session;
1384 arg->ino = ceph_vino(inode).ino;
1385 arg->cid = cap->cap_id;
1386 arg->follows = flushing ? ci->i_head_snapc->seq : 0;
1387 arg->flush_tid = flush_tid;
1388 arg->oldest_flush_tid = oldest_flush_tid;
1390 arg->size = i_size_read(inode);
1391 ci->i_reported_size = arg->size;
1392 arg->max_size = ci->i_wanted_max_size;
1393 if (cap == ci->i_auth_cap) {
1394 if (want & CEPH_CAP_ANY_FILE_WR)
1395 ci->i_requested_max_size = arg->max_size;
1397 ci->i_requested_max_size = 0;
1400 if (flushing & CEPH_CAP_XATTR_EXCL) {
1401 arg->old_xattr_buf = __ceph_build_xattrs_blob(ci);
1402 arg->xattr_version = ci->i_xattrs.version;
1403 arg->xattr_buf = ci->i_xattrs.blob;
1405 arg->xattr_buf = NULL;
1406 arg->old_xattr_buf = NULL;
1409 arg->mtime = inode->i_mtime;
1410 arg->atime = inode->i_atime;
1411 arg->ctime = inode->i_ctime;
1412 arg->btime = ci->i_btime;
1413 arg->change_attr = inode_peek_iversion_raw(inode);
1416 arg->caps = cap->implemented;
1418 arg->dirty = flushing;
1420 arg->seq = cap->seq;
1421 arg->issue_seq = cap->issue_seq;
1422 arg->mseq = cap->mseq;
1423 arg->time_warp_seq = ci->i_time_warp_seq;
1425 arg->uid = inode->i_uid;
1426 arg->gid = inode->i_gid;
1427 arg->mode = inode->i_mode;
1429 arg->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1430 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1431 !list_empty(&ci->i_cap_snaps)) {
1432 struct ceph_cap_snap *capsnap;
1433 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1434 if (capsnap->cap_flush.tid)
1436 if (capsnap->need_flush) {
1437 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1446 * Send a cap msg on the given inode.
1448 * Caller should hold snap_rwsem (read), s_mutex.
1450 static void __send_cap(struct cap_msg_args *arg, struct ceph_inode_info *ci)
1452 struct ceph_msg *msg;
1453 struct inode *inode = &ci->vfs_inode;
1455 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1457 pr_err("error allocating cap msg: ino (%llx.%llx) flushing %s tid %llu, requeuing cap.\n",
1458 ceph_vinop(inode), ceph_cap_string(arg->dirty),
1460 spin_lock(&ci->i_ceph_lock);
1461 __cap_delay_requeue(arg->session->s_mdsc, ci);
1462 spin_unlock(&ci->i_ceph_lock);
1466 encode_cap_msg(msg, arg);
1467 ceph_con_send(&arg->session->s_con, msg);
1468 ceph_buffer_put(arg->old_xattr_buf);
1470 wake_up_all(&ci->i_cap_wq);
1473 static inline int __send_flush_snap(struct inode *inode,
1474 struct ceph_mds_session *session,
1475 struct ceph_cap_snap *capsnap,
1476 u32 mseq, u64 oldest_flush_tid)
1478 struct cap_msg_args arg;
1479 struct ceph_msg *msg;
1481 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1485 arg.session = session;
1486 arg.ino = ceph_vino(inode).ino;
1488 arg.follows = capsnap->follows;
1489 arg.flush_tid = capsnap->cap_flush.tid;
1490 arg.oldest_flush_tid = oldest_flush_tid;
1492 arg.size = capsnap->size;
1494 arg.xattr_version = capsnap->xattr_version;
1495 arg.xattr_buf = capsnap->xattr_blob;
1496 arg.old_xattr_buf = NULL;
1498 arg.atime = capsnap->atime;
1499 arg.mtime = capsnap->mtime;
1500 arg.ctime = capsnap->ctime;
1501 arg.btime = capsnap->btime;
1502 arg.change_attr = capsnap->change_attr;
1504 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1505 arg.caps = capsnap->issued;
1507 arg.dirty = capsnap->dirty;
1512 arg.time_warp_seq = capsnap->time_warp_seq;
1514 arg.uid = capsnap->uid;
1515 arg.gid = capsnap->gid;
1516 arg.mode = capsnap->mode;
1518 arg.inline_data = capsnap->inline_data;
1522 encode_cap_msg(msg, &arg);
1523 ceph_con_send(&arg.session->s_con, msg);
1528 * When a snapshot is taken, clients accumulate dirty metadata on
1529 * inodes with capabilities in ceph_cap_snaps to describe the file
1530 * state at the time the snapshot was taken. This must be flushed
1531 * asynchronously back to the MDS once sync writes complete and dirty
1532 * data is written out.
1534 * Called under i_ceph_lock.
1536 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1537 struct ceph_mds_session *session)
1538 __releases(ci->i_ceph_lock)
1539 __acquires(ci->i_ceph_lock)
1541 struct inode *inode = &ci->vfs_inode;
1542 struct ceph_mds_client *mdsc = session->s_mdsc;
1543 struct ceph_cap_snap *capsnap;
1544 u64 oldest_flush_tid = 0;
1545 u64 first_tid = 1, last_tid = 0;
1547 dout("__flush_snaps %p session %p\n", inode, session);
1549 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1551 * we need to wait for sync writes to complete and for dirty
1552 * pages to be written out.
1554 if (capsnap->dirty_pages || capsnap->writing)
1557 /* should be removed by ceph_try_drop_cap_snap() */
1558 BUG_ON(!capsnap->need_flush);
1560 /* only flush each capsnap once */
1561 if (capsnap->cap_flush.tid > 0) {
1562 dout(" already flushed %p, skipping\n", capsnap);
1566 spin_lock(&mdsc->cap_dirty_lock);
1567 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1568 list_add_tail(&capsnap->cap_flush.g_list,
1569 &mdsc->cap_flush_list);
1570 if (oldest_flush_tid == 0)
1571 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1572 if (list_empty(&ci->i_flushing_item)) {
1573 list_add_tail(&ci->i_flushing_item,
1574 &session->s_cap_flushing);
1576 spin_unlock(&mdsc->cap_dirty_lock);
1578 list_add_tail(&capsnap->cap_flush.i_list,
1579 &ci->i_cap_flush_list);
1582 first_tid = capsnap->cap_flush.tid;
1583 last_tid = capsnap->cap_flush.tid;
1586 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1588 while (first_tid <= last_tid) {
1589 struct ceph_cap *cap = ci->i_auth_cap;
1590 struct ceph_cap_flush *cf;
1593 if (!(cap && cap->session == session)) {
1594 dout("__flush_snaps %p auth cap %p not mds%d, "
1595 "stop\n", inode, cap, session->s_mds);
1600 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1601 if (cf->tid >= first_tid) {
1609 first_tid = cf->tid + 1;
1611 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1612 refcount_inc(&capsnap->nref);
1613 spin_unlock(&ci->i_ceph_lock);
1615 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1616 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1618 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1621 pr_err("__flush_snaps: error sending cap flushsnap, "
1622 "ino (%llx.%llx) tid %llu follows %llu\n",
1623 ceph_vinop(inode), cf->tid, capsnap->follows);
1626 ceph_put_cap_snap(capsnap);
1627 spin_lock(&ci->i_ceph_lock);
1631 void ceph_flush_snaps(struct ceph_inode_info *ci,
1632 struct ceph_mds_session **psession)
1634 struct inode *inode = &ci->vfs_inode;
1635 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1636 struct ceph_mds_session *session = NULL;
1639 dout("ceph_flush_snaps %p\n", inode);
1641 session = *psession;
1643 spin_lock(&ci->i_ceph_lock);
1644 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1645 dout(" no capsnap needs flush, doing nothing\n");
1648 if (!ci->i_auth_cap) {
1649 dout(" no auth cap (migrating?), doing nothing\n");
1653 mds = ci->i_auth_cap->session->s_mds;
1654 if (session && session->s_mds != mds) {
1655 dout(" oops, wrong session %p mutex\n", session);
1656 ceph_put_mds_session(session);
1660 spin_unlock(&ci->i_ceph_lock);
1661 mutex_lock(&mdsc->mutex);
1662 session = __ceph_lookup_mds_session(mdsc, mds);
1663 mutex_unlock(&mdsc->mutex);
1667 // make sure flushsnap messages are sent in proper order.
1668 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1669 __kick_flushing_caps(mdsc, session, ci, 0);
1671 __ceph_flush_snaps(ci, session);
1673 spin_unlock(&ci->i_ceph_lock);
1676 *psession = session;
1678 ceph_put_mds_session(session);
1679 /* we flushed them all; remove this inode from the queue */
1680 spin_lock(&mdsc->snap_flush_lock);
1681 list_del_init(&ci->i_snap_flush_item);
1682 spin_unlock(&mdsc->snap_flush_lock);
1686 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1687 * Caller is then responsible for calling __mark_inode_dirty with the
1688 * returned flags value.
1690 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1691 struct ceph_cap_flush **pcf)
1693 struct ceph_mds_client *mdsc =
1694 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1695 struct inode *inode = &ci->vfs_inode;
1696 int was = ci->i_dirty_caps;
1699 lockdep_assert_held(&ci->i_ceph_lock);
1701 if (!ci->i_auth_cap) {
1702 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1703 "but no auth cap (session was closed?)\n",
1704 inode, ceph_ino(inode), ceph_cap_string(mask));
1708 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1709 ceph_cap_string(mask), ceph_cap_string(was),
1710 ceph_cap_string(was | mask));
1711 ci->i_dirty_caps |= mask;
1713 struct ceph_mds_session *session = ci->i_auth_cap->session;
1715 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1716 swap(ci->i_prealloc_cap_flush, *pcf);
1718 if (!ci->i_head_snapc) {
1719 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1720 ci->i_head_snapc = ceph_get_snap_context(
1721 ci->i_snap_realm->cached_context);
1723 dout(" inode %p now dirty snapc %p auth cap %p\n",
1724 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1725 BUG_ON(!list_empty(&ci->i_dirty_item));
1726 spin_lock(&mdsc->cap_dirty_lock);
1727 list_add(&ci->i_dirty_item, &session->s_cap_dirty);
1728 spin_unlock(&mdsc->cap_dirty_lock);
1729 if (ci->i_flushing_caps == 0) {
1731 dirty |= I_DIRTY_SYNC;
1734 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1736 BUG_ON(list_empty(&ci->i_dirty_item));
1737 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1738 (mask & CEPH_CAP_FILE_BUFFER))
1739 dirty |= I_DIRTY_DATASYNC;
1740 __cap_delay_requeue(mdsc, ci);
1744 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1746 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1749 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1752 kmem_cache_free(ceph_cap_flush_cachep, cf);
1755 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1757 if (!list_empty(&mdsc->cap_flush_list)) {
1758 struct ceph_cap_flush *cf =
1759 list_first_entry(&mdsc->cap_flush_list,
1760 struct ceph_cap_flush, g_list);
1767 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1768 * Return true if caller needs to wake up flush waiters.
1770 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1771 struct ceph_cap_flush *cf)
1773 struct ceph_cap_flush *prev;
1774 bool wake = cf->wake;
1776 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1777 prev = list_prev_entry(cf, g_list);
1781 list_del(&cf->g_list);
1785 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1786 struct ceph_cap_flush *cf)
1788 struct ceph_cap_flush *prev;
1789 bool wake = cf->wake;
1791 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1792 prev = list_prev_entry(cf, i_list);
1796 list_del(&cf->i_list);
1801 * Add dirty inode to the flushing list. Assigned a seq number so we
1802 * can wait for caps to flush without starving.
1804 * Called under i_ceph_lock. Returns the flush tid.
1806 static u64 __mark_caps_flushing(struct inode *inode,
1807 struct ceph_mds_session *session, bool wake,
1808 u64 *oldest_flush_tid)
1810 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1811 struct ceph_inode_info *ci = ceph_inode(inode);
1812 struct ceph_cap_flush *cf = NULL;
1815 lockdep_assert_held(&ci->i_ceph_lock);
1816 BUG_ON(ci->i_dirty_caps == 0);
1817 BUG_ON(list_empty(&ci->i_dirty_item));
1818 BUG_ON(!ci->i_prealloc_cap_flush);
1820 flushing = ci->i_dirty_caps;
1821 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1822 ceph_cap_string(flushing),
1823 ceph_cap_string(ci->i_flushing_caps),
1824 ceph_cap_string(ci->i_flushing_caps | flushing));
1825 ci->i_flushing_caps |= flushing;
1826 ci->i_dirty_caps = 0;
1827 dout(" inode %p now !dirty\n", inode);
1829 swap(cf, ci->i_prealloc_cap_flush);
1830 cf->caps = flushing;
1833 spin_lock(&mdsc->cap_dirty_lock);
1834 list_del_init(&ci->i_dirty_item);
1836 cf->tid = ++mdsc->last_cap_flush_tid;
1837 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1838 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1840 if (list_empty(&ci->i_flushing_item)) {
1841 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1842 mdsc->num_cap_flushing++;
1844 spin_unlock(&mdsc->cap_dirty_lock);
1846 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1852 * try to invalidate mapping pages without blocking.
1854 static int try_nonblocking_invalidate(struct inode *inode)
1856 struct ceph_inode_info *ci = ceph_inode(inode);
1857 u32 invalidating_gen = ci->i_rdcache_gen;
1859 spin_unlock(&ci->i_ceph_lock);
1860 ceph_fscache_invalidate(inode);
1861 invalidate_mapping_pages(&inode->i_data, 0, -1);
1862 spin_lock(&ci->i_ceph_lock);
1864 if (inode->i_data.nrpages == 0 &&
1865 invalidating_gen == ci->i_rdcache_gen) {
1867 dout("try_nonblocking_invalidate %p success\n", inode);
1868 /* save any racing async invalidate some trouble */
1869 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1872 dout("try_nonblocking_invalidate %p failed\n", inode);
1876 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1878 loff_t size = i_size_read(&ci->vfs_inode);
1879 /* mds will adjust max size according to the reported size */
1880 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1882 if (size >= ci->i_max_size)
1884 /* half of previous max_size increment has been used */
1885 if (ci->i_max_size > ci->i_reported_size &&
1886 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1892 * Swiss army knife function to examine currently used and wanted
1893 * versus held caps. Release, flush, ack revoked caps to mds as
1896 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1897 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1900 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1901 struct ceph_mds_session *session)
1903 struct inode *inode = &ci->vfs_inode;
1904 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1905 struct ceph_cap *cap;
1906 u64 flush_tid, oldest_flush_tid;
1907 int file_wanted, used, cap_used;
1908 int issued, implemented, want, retain, revoking, flushing = 0;
1909 int mds = -1; /* keep track of how far we've gone through i_caps list
1910 to avoid an infinite loop on retry */
1912 bool queue_invalidate = false;
1913 bool tried_invalidate = false;
1916 ceph_get_mds_session(session);
1918 spin_lock(&ci->i_ceph_lock);
1919 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1920 flags |= CHECK_CAPS_FLUSH;
1922 /* Caps wanted by virtue of active open files. */
1923 file_wanted = __ceph_caps_file_wanted(ci);
1925 /* Caps which have active references against them */
1926 used = __ceph_caps_used(ci);
1929 * "issued" represents the current caps that the MDS wants us to have.
1930 * "implemented" is the set that we have been granted, and includes the
1931 * ones that have not yet been returned to the MDS (the "revoking" set,
1932 * usually because they have outstanding references).
1934 issued = __ceph_caps_issued(ci, &implemented);
1935 revoking = implemented & ~issued;
1939 /* The ones we currently want to retain (may be adjusted below) */
1940 retain = file_wanted | used | CEPH_CAP_PIN;
1941 if (!mdsc->stopping && inode->i_nlink > 0) {
1943 retain |= CEPH_CAP_ANY; /* be greedy */
1944 } else if (S_ISDIR(inode->i_mode) &&
1945 (issued & CEPH_CAP_FILE_SHARED) &&
1946 __ceph_dir_is_complete(ci)) {
1948 * If a directory is complete, we want to keep
1949 * the exclusive cap. So that MDS does not end up
1950 * revoking the shared cap on every create/unlink
1953 if (IS_RDONLY(inode)) {
1954 want = CEPH_CAP_ANY_SHARED;
1956 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1961 retain |= CEPH_CAP_ANY_SHARED;
1963 * keep RD only if we didn't have the file open RW,
1964 * because then the mds would revoke it anyway to
1965 * journal max_size=0.
1967 if (ci->i_max_size == 0)
1968 retain |= CEPH_CAP_ANY_RD;
1972 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1973 " issued %s revoking %s retain %s %s%s\n", inode,
1974 ceph_cap_string(file_wanted),
1975 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1976 ceph_cap_string(ci->i_flushing_caps),
1977 ceph_cap_string(issued), ceph_cap_string(revoking),
1978 ceph_cap_string(retain),
1979 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1980 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1983 * If we no longer need to hold onto old our caps, and we may
1984 * have cached pages, but don't want them, then try to invalidate.
1985 * If we fail, it's because pages are locked.... try again later.
1987 if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
1988 S_ISREG(inode->i_mode) &&
1989 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1990 inode->i_data.nrpages && /* have cached pages */
1991 (revoking & (CEPH_CAP_FILE_CACHE|
1992 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1993 !tried_invalidate) {
1994 dout("check_caps trying to invalidate on %p\n", inode);
1995 if (try_nonblocking_invalidate(inode) < 0) {
1996 dout("check_caps queuing invalidate\n");
1997 queue_invalidate = true;
1998 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2000 tried_invalidate = true;
2004 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2006 struct cap_msg_args arg;
2008 cap = rb_entry(p, struct ceph_cap, ci_node);
2010 /* avoid looping forever */
2011 if (mds >= cap->mds ||
2012 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
2016 * If we have an auth cap, we don't need to consider any
2017 * overlapping caps as used.
2020 if (ci->i_auth_cap && cap != ci->i_auth_cap)
2021 cap_used &= ~ci->i_auth_cap->issued;
2023 revoking = cap->implemented & ~cap->issued;
2024 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
2025 cap->mds, cap, ceph_cap_string(cap_used),
2026 ceph_cap_string(cap->issued),
2027 ceph_cap_string(cap->implemented),
2028 ceph_cap_string(revoking));
2030 if (cap == ci->i_auth_cap &&
2031 (cap->issued & CEPH_CAP_FILE_WR)) {
2032 /* request larger max_size from MDS? */
2033 if (ci->i_wanted_max_size > ci->i_max_size &&
2034 ci->i_wanted_max_size > ci->i_requested_max_size) {
2035 dout("requesting new max_size\n");
2039 /* approaching file_max? */
2040 if (__ceph_should_report_size(ci)) {
2041 dout("i_size approaching max_size\n");
2045 /* flush anything dirty? */
2046 if (cap == ci->i_auth_cap) {
2047 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
2048 dout("flushing dirty caps\n");
2051 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
2052 dout("flushing snap caps\n");
2057 /* completed revocation? going down and there are no caps? */
2058 if (revoking && (revoking & cap_used) == 0) {
2059 dout("completed revocation of %s\n",
2060 ceph_cap_string(cap->implemented & ~cap->issued));
2064 /* want more caps from mds? */
2065 if (want & ~cap->mds_wanted) {
2066 if (want & ~(cap->mds_wanted | cap->issued))
2068 if (!__cap_is_valid(cap))
2072 /* things we might delay */
2073 if ((cap->issued & ~retain) == 0)
2074 continue; /* nope, all good */
2077 ceph_put_mds_session(session);
2078 session = ceph_get_mds_session(cap->session);
2080 /* kick flushing and flush snaps before sending normal
2082 if (cap == ci->i_auth_cap &&
2084 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2085 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2086 __kick_flushing_caps(mdsc, session, ci, 0);
2087 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2088 __ceph_flush_snaps(ci, session);
2093 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2094 flushing = ci->i_dirty_caps;
2095 flush_tid = __mark_caps_flushing(inode, session, false,
2097 if (flags & CHECK_CAPS_FLUSH &&
2098 list_empty(&session->s_cap_dirty))
2099 mflags |= CEPH_CLIENT_CAPS_SYNC;
2103 spin_lock(&mdsc->cap_dirty_lock);
2104 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2105 spin_unlock(&mdsc->cap_dirty_lock);
2108 mds = cap->mds; /* remember mds, so we don't repeat */
2110 __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2111 want, retain, flushing, flush_tid, oldest_flush_tid);
2113 spin_unlock(&ci->i_ceph_lock);
2114 __send_cap(&arg, ci);
2115 spin_lock(&ci->i_ceph_lock);
2117 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2120 /* periodically re-calculate caps wanted by open files */
2121 if (__ceph_is_any_real_caps(ci) &&
2122 list_empty(&ci->i_cap_delay_list) &&
2123 (file_wanted & ~CEPH_CAP_PIN) &&
2124 !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2125 __cap_delay_requeue(mdsc, ci);
2128 spin_unlock(&ci->i_ceph_lock);
2130 ceph_put_mds_session(session);
2131 if (queue_invalidate)
2132 ceph_queue_invalidate(inode);
2136 * Try to flush dirty caps back to the auth mds.
2138 static int try_flush_caps(struct inode *inode, u64 *ptid)
2140 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2141 struct ceph_inode_info *ci = ceph_inode(inode);
2143 u64 flush_tid = 0, oldest_flush_tid = 0;
2145 spin_lock(&ci->i_ceph_lock);
2147 if (ci->i_dirty_caps && ci->i_auth_cap) {
2148 struct ceph_cap *cap = ci->i_auth_cap;
2149 struct cap_msg_args arg;
2150 struct ceph_mds_session *session = cap->session;
2152 if (session->s_state < CEPH_MDS_SESSION_OPEN) {
2153 spin_unlock(&ci->i_ceph_lock);
2157 if (ci->i_ceph_flags &
2158 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2159 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2160 __kick_flushing_caps(mdsc, session, ci, 0);
2161 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2162 __ceph_flush_snaps(ci, session);
2166 flushing = ci->i_dirty_caps;
2167 flush_tid = __mark_caps_flushing(inode, session, true,
2170 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2171 __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2172 (cap->issued | cap->implemented),
2173 flushing, flush_tid, oldest_flush_tid);
2174 spin_unlock(&ci->i_ceph_lock);
2176 __send_cap(&arg, ci);
2178 if (!list_empty(&ci->i_cap_flush_list)) {
2179 struct ceph_cap_flush *cf =
2180 list_last_entry(&ci->i_cap_flush_list,
2181 struct ceph_cap_flush, i_list);
2183 flush_tid = cf->tid;
2185 flushing = ci->i_flushing_caps;
2186 spin_unlock(&ci->i_ceph_lock);
2194 * Return true if we've flushed caps through the given flush_tid.
2196 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2198 struct ceph_inode_info *ci = ceph_inode(inode);
2201 spin_lock(&ci->i_ceph_lock);
2202 if (!list_empty(&ci->i_cap_flush_list)) {
2203 struct ceph_cap_flush * cf =
2204 list_first_entry(&ci->i_cap_flush_list,
2205 struct ceph_cap_flush, i_list);
2206 if (cf->tid <= flush_tid)
2209 spin_unlock(&ci->i_ceph_lock);
2214 * wait for any unsafe requests to complete.
2216 static int unsafe_request_wait(struct inode *inode)
2218 struct ceph_inode_info *ci = ceph_inode(inode);
2219 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2222 spin_lock(&ci->i_unsafe_lock);
2223 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2224 req1 = list_last_entry(&ci->i_unsafe_dirops,
2225 struct ceph_mds_request,
2227 ceph_mdsc_get_request(req1);
2229 if (!list_empty(&ci->i_unsafe_iops)) {
2230 req2 = list_last_entry(&ci->i_unsafe_iops,
2231 struct ceph_mds_request,
2232 r_unsafe_target_item);
2233 ceph_mdsc_get_request(req2);
2235 spin_unlock(&ci->i_unsafe_lock);
2237 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2238 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2240 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2241 ceph_timeout_jiffies(req1->r_timeout));
2244 ceph_mdsc_put_request(req1);
2247 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2248 ceph_timeout_jiffies(req2->r_timeout));
2251 ceph_mdsc_put_request(req2);
2256 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2258 struct ceph_file_info *fi = file->private_data;
2259 struct inode *inode = file->f_mapping->host;
2260 struct ceph_inode_info *ci = ceph_inode(inode);
2265 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2267 ret = file_write_and_wait_range(file, start, end);
2271 ret = ceph_wait_on_async_create(inode);
2275 dirty = try_flush_caps(inode, &flush_tid);
2276 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2278 err = unsafe_request_wait(inode);
2281 * only wait on non-file metadata writeback (the mds
2282 * can recover size and mtime, so we don't need to
2285 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2286 err = wait_event_interruptible(ci->i_cap_wq,
2287 caps_are_flushed(inode, flush_tid));
2293 if (errseq_check(&ci->i_meta_err, READ_ONCE(fi->meta_err))) {
2294 spin_lock(&file->f_lock);
2295 err = errseq_check_and_advance(&ci->i_meta_err,
2297 spin_unlock(&file->f_lock);
2302 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2307 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2308 * queue inode for flush but don't do so immediately, because we can
2309 * get by with fewer MDS messages if we wait for data writeback to
2312 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2314 struct ceph_inode_info *ci = ceph_inode(inode);
2318 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2320 dout("write_inode %p wait=%d\n", inode, wait);
2322 dirty = try_flush_caps(inode, &flush_tid);
2324 err = wait_event_interruptible(ci->i_cap_wq,
2325 caps_are_flushed(inode, flush_tid));
2327 struct ceph_mds_client *mdsc =
2328 ceph_sb_to_client(inode->i_sb)->mdsc;
2330 spin_lock(&ci->i_ceph_lock);
2331 if (__ceph_caps_dirty(ci))
2332 __cap_delay_requeue_front(mdsc, ci);
2333 spin_unlock(&ci->i_ceph_lock);
2338 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2339 struct ceph_mds_session *session,
2340 struct ceph_inode_info *ci,
2341 u64 oldest_flush_tid)
2342 __releases(ci->i_ceph_lock)
2343 __acquires(ci->i_ceph_lock)
2345 struct inode *inode = &ci->vfs_inode;
2346 struct ceph_cap *cap;
2347 struct ceph_cap_flush *cf;
2350 u64 last_snap_flush = 0;
2352 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2354 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2356 last_snap_flush = cf->tid;
2361 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2362 if (cf->tid < first_tid)
2365 cap = ci->i_auth_cap;
2366 if (!(cap && cap->session == session)) {
2367 pr_err("%p auth cap %p not mds%d ???\n",
2368 inode, cap, session->s_mds);
2372 first_tid = cf->tid + 1;
2375 struct cap_msg_args arg;
2377 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2378 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2379 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2380 (cf->tid < last_snap_flush ?
2381 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2382 __ceph_caps_used(ci),
2383 __ceph_caps_wanted(ci),
2384 (cap->issued | cap->implemented),
2385 cf->caps, cf->tid, oldest_flush_tid);
2386 spin_unlock(&ci->i_ceph_lock);
2387 __send_cap(&arg, ci);
2389 struct ceph_cap_snap *capsnap =
2390 container_of(cf, struct ceph_cap_snap,
2392 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2393 inode, capsnap, cf->tid,
2394 ceph_cap_string(capsnap->dirty));
2396 refcount_inc(&capsnap->nref);
2397 spin_unlock(&ci->i_ceph_lock);
2399 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2402 pr_err("kick_flushing_caps: error sending "
2403 "cap flushsnap, ino (%llx.%llx) "
2404 "tid %llu follows %llu\n",
2405 ceph_vinop(inode), cf->tid,
2409 ceph_put_cap_snap(capsnap);
2412 spin_lock(&ci->i_ceph_lock);
2416 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2417 struct ceph_mds_session *session)
2419 struct ceph_inode_info *ci;
2420 struct ceph_cap *cap;
2421 u64 oldest_flush_tid;
2423 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2425 spin_lock(&mdsc->cap_dirty_lock);
2426 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2427 spin_unlock(&mdsc->cap_dirty_lock);
2429 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2430 spin_lock(&ci->i_ceph_lock);
2431 cap = ci->i_auth_cap;
2432 if (!(cap && cap->session == session)) {
2433 pr_err("%p auth cap %p not mds%d ???\n",
2434 &ci->vfs_inode, cap, session->s_mds);
2435 spin_unlock(&ci->i_ceph_lock);
2441 * if flushing caps were revoked, we re-send the cap flush
2442 * in client reconnect stage. This guarantees MDS * processes
2443 * the cap flush message before issuing the flushing caps to
2446 if ((cap->issued & ci->i_flushing_caps) !=
2447 ci->i_flushing_caps) {
2448 /* encode_caps_cb() also will reset these sequence
2449 * numbers. make sure sequence numbers in cap flush
2450 * message match later reconnect message */
2454 __kick_flushing_caps(mdsc, session, ci,
2457 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2460 spin_unlock(&ci->i_ceph_lock);
2464 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2465 struct ceph_mds_session *session)
2467 struct ceph_inode_info *ci;
2468 struct ceph_cap *cap;
2469 u64 oldest_flush_tid;
2471 lockdep_assert_held(&session->s_mutex);
2473 dout("kick_flushing_caps mds%d\n", session->s_mds);
2475 spin_lock(&mdsc->cap_dirty_lock);
2476 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2477 spin_unlock(&mdsc->cap_dirty_lock);
2479 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2480 spin_lock(&ci->i_ceph_lock);
2481 cap = ci->i_auth_cap;
2482 if (!(cap && cap->session == session)) {
2483 pr_err("%p auth cap %p not mds%d ???\n",
2484 &ci->vfs_inode, cap, session->s_mds);
2485 spin_unlock(&ci->i_ceph_lock);
2488 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2489 __kick_flushing_caps(mdsc, session, ci,
2492 spin_unlock(&ci->i_ceph_lock);
2496 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2497 struct ceph_inode_info *ci)
2499 struct ceph_mds_client *mdsc = session->s_mdsc;
2500 struct ceph_cap *cap = ci->i_auth_cap;
2502 lockdep_assert_held(&ci->i_ceph_lock);
2504 dout("%s %p flushing %s\n", __func__, &ci->vfs_inode,
2505 ceph_cap_string(ci->i_flushing_caps));
2507 if (!list_empty(&ci->i_cap_flush_list)) {
2508 u64 oldest_flush_tid;
2509 spin_lock(&mdsc->cap_dirty_lock);
2510 list_move_tail(&ci->i_flushing_item,
2511 &cap->session->s_cap_flushing);
2512 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2513 spin_unlock(&mdsc->cap_dirty_lock);
2515 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2521 * Take references to capabilities we hold, so that we don't release
2522 * them to the MDS prematurely.
2524 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2525 bool snap_rwsem_locked)
2527 lockdep_assert_held(&ci->i_ceph_lock);
2529 if (got & CEPH_CAP_PIN)
2531 if (got & CEPH_CAP_FILE_RD)
2533 if (got & CEPH_CAP_FILE_CACHE)
2534 ci->i_rdcache_ref++;
2535 if (got & CEPH_CAP_FILE_EXCL)
2537 if (got & CEPH_CAP_FILE_WR) {
2538 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2539 BUG_ON(!snap_rwsem_locked);
2540 ci->i_head_snapc = ceph_get_snap_context(
2541 ci->i_snap_realm->cached_context);
2545 if (got & CEPH_CAP_FILE_BUFFER) {
2546 if (ci->i_wb_ref == 0)
2547 ihold(&ci->vfs_inode);
2549 dout("%s %p wb %d -> %d (?)\n", __func__,
2550 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2555 * Try to grab cap references. Specify those refs we @want, and the
2556 * minimal set we @need. Also include the larger offset we are writing
2557 * to (when applicable), and check against max_size here as well.
2558 * Note that caller is responsible for ensuring max_size increases are
2559 * requested from the MDS.
2561 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2562 * or a negative error code. There are 3 speical error codes:
2563 * -EAGAIN: need to sleep but non-blocking is specified
2564 * -EFBIG: ask caller to call check_max_size() and try again.
2565 * -ESTALE: ask caller to call ceph_renew_caps() and try again.
2568 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2569 NON_BLOCKING = (1 << 8),
2570 CHECK_FILELOCK = (1 << 9),
2573 static int try_get_cap_refs(struct inode *inode, int need, int want,
2574 loff_t endoff, int flags, int *got)
2576 struct ceph_inode_info *ci = ceph_inode(inode);
2577 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2579 int have, implemented;
2580 bool snap_rwsem_locked = false;
2582 dout("get_cap_refs %p need %s want %s\n", inode,
2583 ceph_cap_string(need), ceph_cap_string(want));
2586 spin_lock(&ci->i_ceph_lock);
2588 if ((flags & CHECK_FILELOCK) &&
2589 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2590 dout("try_get_cap_refs %p error filelock\n", inode);
2595 /* finish pending truncate */
2596 while (ci->i_truncate_pending) {
2597 spin_unlock(&ci->i_ceph_lock);
2598 if (snap_rwsem_locked) {
2599 up_read(&mdsc->snap_rwsem);
2600 snap_rwsem_locked = false;
2602 __ceph_do_pending_vmtruncate(inode);
2603 spin_lock(&ci->i_ceph_lock);
2606 have = __ceph_caps_issued(ci, &implemented);
2608 if (have & need & CEPH_CAP_FILE_WR) {
2609 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2610 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2611 inode, endoff, ci->i_max_size);
2612 if (endoff > ci->i_requested_max_size)
2613 ret = ci->i_auth_cap ? -EFBIG : -ESTALE;
2617 * If a sync write is in progress, we must wait, so that we
2618 * can get a final snapshot value for size+mtime.
2620 if (__ceph_have_pending_cap_snap(ci)) {
2621 dout("get_cap_refs %p cap_snap_pending\n", inode);
2626 if ((have & need) == need) {
2628 * Look at (implemented & ~have & not) so that we keep waiting
2629 * on transition from wanted -> needed caps. This is needed
2630 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2631 * going before a prior buffered writeback happens.
2633 int not = want & ~(have & need);
2634 int revoking = implemented & ~have;
2635 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2636 inode, ceph_cap_string(have), ceph_cap_string(not),
2637 ceph_cap_string(revoking));
2638 if ((revoking & not) == 0) {
2639 if (!snap_rwsem_locked &&
2640 !ci->i_head_snapc &&
2641 (need & CEPH_CAP_FILE_WR)) {
2642 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2644 * we can not call down_read() when
2645 * task isn't in TASK_RUNNING state
2647 if (flags & NON_BLOCKING) {
2652 spin_unlock(&ci->i_ceph_lock);
2653 down_read(&mdsc->snap_rwsem);
2654 snap_rwsem_locked = true;
2657 snap_rwsem_locked = true;
2659 if ((have & want) == want)
2663 ceph_take_cap_refs(ci, *got, true);
2667 int session_readonly = false;
2669 if (ci->i_auth_cap &&
2670 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2671 struct ceph_mds_session *s = ci->i_auth_cap->session;
2672 spin_lock(&s->s_cap_lock);
2673 session_readonly = s->s_readonly;
2674 spin_unlock(&s->s_cap_lock);
2676 if (session_readonly) {
2677 dout("get_cap_refs %p need %s but mds%d readonly\n",
2678 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2683 if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
2684 dout("get_cap_refs %p forced umount\n", inode);
2688 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2689 if (need & ~mds_wanted) {
2690 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2691 inode, ceph_cap_string(need),
2692 ceph_cap_string(mds_wanted));
2697 dout("get_cap_refs %p have %s need %s\n", inode,
2698 ceph_cap_string(have), ceph_cap_string(need));
2702 __ceph_touch_fmode(ci, mdsc, flags);
2704 spin_unlock(&ci->i_ceph_lock);
2705 if (snap_rwsem_locked)
2706 up_read(&mdsc->snap_rwsem);
2709 ceph_update_cap_mis(&mdsc->metric);
2711 ceph_update_cap_hit(&mdsc->metric);
2713 dout("get_cap_refs %p ret %d got %s\n", inode,
2714 ret, ceph_cap_string(*got));
2719 * Check the offset we are writing up to against our current
2720 * max_size. If necessary, tell the MDS we want to write to
2723 static void check_max_size(struct inode *inode, loff_t endoff)
2725 struct ceph_inode_info *ci = ceph_inode(inode);
2728 /* do we need to explicitly request a larger max_size? */
2729 spin_lock(&ci->i_ceph_lock);
2730 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2731 dout("write %p at large endoff %llu, req max_size\n",
2733 ci->i_wanted_max_size = endoff;
2735 /* duplicate ceph_check_caps()'s logic */
2736 if (ci->i_auth_cap &&
2737 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2738 ci->i_wanted_max_size > ci->i_max_size &&
2739 ci->i_wanted_max_size > ci->i_requested_max_size)
2741 spin_unlock(&ci->i_ceph_lock);
2743 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2746 static inline int get_used_fmode(int caps)
2749 if (caps & CEPH_CAP_FILE_RD)
2750 fmode |= CEPH_FILE_MODE_RD;
2751 if (caps & CEPH_CAP_FILE_WR)
2752 fmode |= CEPH_FILE_MODE_WR;
2756 int ceph_try_get_caps(struct inode *inode, int need, int want,
2757 bool nonblock, int *got)
2761 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2762 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2763 CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2764 CEPH_CAP_ANY_DIR_OPS));
2766 ret = ceph_pool_perm_check(inode, need);
2771 flags = get_used_fmode(need | want);
2773 flags |= NON_BLOCKING;
2775 ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2776 /* three special error codes */
2777 if (ret == -EAGAIN || ret == -EFBIG || ret == -ESTALE)
2783 * Wait for caps, and take cap references. If we can't get a WR cap
2784 * due to a small max_size, make sure we check_max_size (and possibly
2785 * ask the mds) so we don't get hung up indefinitely.
2787 int ceph_get_caps(struct file *filp, int need, int want, loff_t endoff, int *got)
2789 struct ceph_file_info *fi = filp->private_data;
2790 struct inode *inode = file_inode(filp);
2791 struct ceph_inode_info *ci = ceph_inode(inode);
2792 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2793 int ret, _got, flags;
2795 ret = ceph_pool_perm_check(inode, need);
2799 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2800 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2803 flags = get_used_fmode(need | want);
2806 flags &= CEPH_FILE_MODE_MASK;
2807 if (atomic_read(&fi->num_locks))
2808 flags |= CHECK_FILELOCK;
2810 ret = try_get_cap_refs(inode, need, want, endoff,
2812 WARN_ON_ONCE(ret == -EAGAIN);
2814 struct ceph_mds_client *mdsc = fsc->mdsc;
2816 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2818 cw.ino = ceph_ino(inode);
2819 cw.tgid = current->tgid;
2823 spin_lock(&mdsc->caps_list_lock);
2824 list_add(&cw.list, &mdsc->cap_wait_list);
2825 spin_unlock(&mdsc->caps_list_lock);
2827 /* make sure used fmode not timeout */
2828 ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2829 add_wait_queue(&ci->i_cap_wq, &wait);
2831 flags |= NON_BLOCKING;
2832 while (!(ret = try_get_cap_refs(inode, need, want,
2833 endoff, flags, &_got))) {
2834 if (signal_pending(current)) {
2838 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2841 remove_wait_queue(&ci->i_cap_wq, &wait);
2842 ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
2844 spin_lock(&mdsc->caps_list_lock);
2846 spin_unlock(&mdsc->caps_list_lock);
2852 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2853 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2854 if (ret >= 0 && _got)
2855 ceph_put_cap_refs(ci, _got);
2860 if (ret == -EFBIG || ret == -ESTALE) {
2861 int ret2 = ceph_wait_on_async_create(inode);
2865 if (ret == -EFBIG) {
2866 check_max_size(inode, endoff);
2869 if (ret == -ESTALE) {
2870 /* session was killed, try renew caps */
2871 ret = ceph_renew_caps(inode, flags);
2878 if (S_ISREG(ci->vfs_inode.i_mode) &&
2879 ci->i_inline_version != CEPH_INLINE_NONE &&
2880 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2881 i_size_read(inode) > 0) {
2883 find_get_page(inode->i_mapping, 0);
2885 bool uptodate = PageUptodate(page);
2892 * drop cap refs first because getattr while
2893 * holding * caps refs can cause deadlock.
2895 ceph_put_cap_refs(ci, _got);
2899 * getattr request will bring inline data into
2902 ret = __ceph_do_getattr(inode, NULL,
2903 CEPH_STAT_CAP_INLINE_DATA,
2916 * Take cap refs. Caller must already know we hold at least one ref
2917 * on the caps in question or we don't know this is safe.
2919 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2921 spin_lock(&ci->i_ceph_lock);
2922 ceph_take_cap_refs(ci, caps, false);
2923 spin_unlock(&ci->i_ceph_lock);
2928 * drop cap_snap that is not associated with any snapshot.
2929 * we don't need to send FLUSHSNAP message for it.
2931 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2932 struct ceph_cap_snap *capsnap)
2934 if (!capsnap->need_flush &&
2935 !capsnap->writing && !capsnap->dirty_pages) {
2936 dout("dropping cap_snap %p follows %llu\n",
2937 capsnap, capsnap->follows);
2938 BUG_ON(capsnap->cap_flush.tid > 0);
2939 ceph_put_snap_context(capsnap->context);
2940 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2941 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2943 list_del(&capsnap->ci_item);
2944 ceph_put_cap_snap(capsnap);
2950 enum put_cap_refs_mode {
2951 PUT_CAP_REFS_SYNC = 0,
2952 PUT_CAP_REFS_NO_CHECK,
2959 * If we released the last ref on any given cap, call ceph_check_caps
2960 * to release (or schedule a release).
2962 * If we are releasing a WR cap (from a sync write), finalize any affected
2963 * cap_snap, and wake up any waiters.
2965 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
2966 enum put_cap_refs_mode mode)
2968 struct inode *inode = &ci->vfs_inode;
2969 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2970 bool check_flushsnaps = false;
2972 spin_lock(&ci->i_ceph_lock);
2973 if (had & CEPH_CAP_PIN)
2975 if (had & CEPH_CAP_FILE_RD)
2976 if (--ci->i_rd_ref == 0)
2978 if (had & CEPH_CAP_FILE_CACHE)
2979 if (--ci->i_rdcache_ref == 0)
2981 if (had & CEPH_CAP_FILE_EXCL)
2982 if (--ci->i_fx_ref == 0)
2984 if (had & CEPH_CAP_FILE_BUFFER) {
2985 if (--ci->i_wb_ref == 0) {
2987 /* put the ref held by ceph_take_cap_refs() */
2989 check_flushsnaps = true;
2991 dout("put_cap_refs %p wb %d -> %d (?)\n",
2992 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2994 if (had & CEPH_CAP_FILE_WR) {
2995 if (--ci->i_wr_ref == 0) {
2997 check_flushsnaps = true;
2998 if (ci->i_wrbuffer_ref_head == 0 &&
2999 ci->i_dirty_caps == 0 &&
3000 ci->i_flushing_caps == 0) {
3001 BUG_ON(!ci->i_head_snapc);
3002 ceph_put_snap_context(ci->i_head_snapc);
3003 ci->i_head_snapc = NULL;
3005 /* see comment in __ceph_remove_cap() */
3006 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3007 drop_inode_snap_realm(ci);
3010 if (check_flushsnaps && __ceph_have_pending_cap_snap(ci)) {
3011 struct ceph_cap_snap *capsnap =
3012 list_last_entry(&ci->i_cap_snaps,
3013 struct ceph_cap_snap,
3016 capsnap->writing = 0;
3017 if (ceph_try_drop_cap_snap(ci, capsnap))
3018 /* put the ref held by ceph_queue_cap_snap() */
3020 else if (__ceph_finish_cap_snap(ci, capsnap))
3024 spin_unlock(&ci->i_ceph_lock);
3026 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3027 last ? " last" : "", put ? " put" : "");
3030 case PUT_CAP_REFS_SYNC:
3032 ceph_check_caps(ci, 0, NULL);
3033 else if (flushsnaps)
3034 ceph_flush_snaps(ci, NULL);
3036 case PUT_CAP_REFS_ASYNC:
3038 ceph_queue_check_caps(inode);
3039 else if (flushsnaps)
3040 ceph_queue_flush_snaps(inode);
3046 wake_up_all(&ci->i_cap_wq);
3051 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3053 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_SYNC);
3056 void ceph_put_cap_refs_async(struct ceph_inode_info *ci, int had)
3058 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_ASYNC);
3061 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3063 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_NO_CHECK);
3067 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3068 * context. Adjust per-snap dirty page accounting as appropriate.
3069 * Once all dirty data for a cap_snap is flushed, flush snapped file
3070 * metadata back to the MDS. If we dropped the last ref, call
3073 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3074 struct ceph_snap_context *snapc)
3076 struct inode *inode = &ci->vfs_inode;
3077 struct ceph_cap_snap *capsnap = NULL;
3081 bool flush_snaps = false;
3082 bool complete_capsnap = false;
3084 spin_lock(&ci->i_ceph_lock);
3085 ci->i_wrbuffer_ref -= nr;
3086 if (ci->i_wrbuffer_ref == 0) {
3091 if (ci->i_head_snapc == snapc) {
3092 ci->i_wrbuffer_ref_head -= nr;
3093 if (ci->i_wrbuffer_ref_head == 0 &&
3094 ci->i_wr_ref == 0 &&
3095 ci->i_dirty_caps == 0 &&
3096 ci->i_flushing_caps == 0) {
3097 BUG_ON(!ci->i_head_snapc);
3098 ceph_put_snap_context(ci->i_head_snapc);
3099 ci->i_head_snapc = NULL;
3101 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3103 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3104 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3105 last ? " LAST" : "");
3107 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3108 if (capsnap->context == snapc) {
3114 capsnap->dirty_pages -= nr;
3115 if (capsnap->dirty_pages == 0) {
3116 complete_capsnap = true;
3117 if (!capsnap->writing) {
3118 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3121 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3126 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3127 " snap %lld %d/%d -> %d/%d %s%s\n",
3128 inode, capsnap, capsnap->context->seq,
3129 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3130 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3131 last ? " (wrbuffer last)" : "",
3132 complete_capsnap ? " (complete capsnap)" : "");
3135 spin_unlock(&ci->i_ceph_lock);
3138 ceph_check_caps(ci, 0, NULL);
3139 } else if (flush_snaps) {
3140 ceph_flush_snaps(ci, NULL);
3142 if (complete_capsnap)
3143 wake_up_all(&ci->i_cap_wq);
3150 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3152 static void invalidate_aliases(struct inode *inode)
3154 struct dentry *dn, *prev = NULL;
3156 dout("invalidate_aliases inode %p\n", inode);
3157 d_prune_aliases(inode);
3159 * For non-directory inode, d_find_alias() only returns
3160 * hashed dentry. After calling d_invalidate(), the
3161 * dentry becomes unhashed.
3163 * For directory inode, d_find_alias() can return
3164 * unhashed dentry. But directory inode should have
3165 * one alias at most.
3167 while ((dn = d_find_alias(inode))) {
3181 struct cap_extra_info {
3182 struct ceph_string *pool_ns;
3192 /* currently issued */
3194 struct timespec64 btime;
3198 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3199 * actually be a revocation if it specifies a smaller cap set.)
3201 * caller holds s_mutex and i_ceph_lock, we drop both.
3203 static void handle_cap_grant(struct inode *inode,
3204 struct ceph_mds_session *session,
3205 struct ceph_cap *cap,
3206 struct ceph_mds_caps *grant,
3207 struct ceph_buffer *xattr_buf,
3208 struct cap_extra_info *extra_info)
3209 __releases(ci->i_ceph_lock)
3210 __releases(session->s_mdsc->snap_rwsem)
3212 struct ceph_inode_info *ci = ceph_inode(inode);
3213 int seq = le32_to_cpu(grant->seq);
3214 int newcaps = le32_to_cpu(grant->caps);
3215 int used, wanted, dirty;
3216 u64 size = le64_to_cpu(grant->size);
3217 u64 max_size = le64_to_cpu(grant->max_size);
3218 unsigned char check_caps = 0;
3219 bool was_stale = cap->cap_gen < atomic_read(&session->s_cap_gen);
3221 bool writeback = false;
3222 bool queue_trunc = false;
3223 bool queue_invalidate = false;
3224 bool deleted_inode = false;
3225 bool fill_inline = false;
3227 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3228 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3229 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3230 i_size_read(inode));
3234 * If CACHE is being revoked, and we have no dirty buffers,
3235 * try to invalidate (once). (If there are dirty buffers, we
3236 * will invalidate _after_ writeback.)
3238 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3239 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3240 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3241 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3242 if (try_nonblocking_invalidate(inode)) {
3243 /* there were locked pages.. invalidate later
3244 in a separate thread. */
3245 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3246 queue_invalidate = true;
3247 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3253 cap->issued = cap->implemented = CEPH_CAP_PIN;
3256 * auth mds of the inode changed. we received the cap export message,
3257 * but still haven't received the cap import message. handle_cap_export
3258 * updated the new auth MDS' cap.
3260 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3261 * that was sent before the cap import message. So don't remove caps.
3263 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3264 WARN_ON(cap != ci->i_auth_cap);
3265 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3267 newcaps |= cap->issued;
3270 /* side effects now are allowed */
3271 cap->cap_gen = atomic_read(&session->s_cap_gen);
3274 __check_cap_issue(ci, cap, newcaps);
3276 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3278 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3279 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3280 umode_t mode = le32_to_cpu(grant->mode);
3282 if (inode_wrong_type(inode, mode))
3283 pr_warn_once("inode type changed! (ino %llx.%llx is 0%o, mds says 0%o)\n",
3284 ceph_vinop(inode), inode->i_mode, mode);
3286 inode->i_mode = mode;
3287 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3288 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3289 ci->i_btime = extra_info->btime;
3290 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3291 from_kuid(&init_user_ns, inode->i_uid),
3292 from_kgid(&init_user_ns, inode->i_gid));
3295 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3296 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3297 set_nlink(inode, le32_to_cpu(grant->nlink));
3298 if (inode->i_nlink == 0 &&
3299 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3300 deleted_inode = true;
3303 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3305 int len = le32_to_cpu(grant->xattr_len);
3306 u64 version = le64_to_cpu(grant->xattr_version);
3308 if (version > ci->i_xattrs.version) {
3309 dout(" got new xattrs v%llu on %p len %d\n",
3310 version, inode, len);
3311 if (ci->i_xattrs.blob)
3312 ceph_buffer_put(ci->i_xattrs.blob);
3313 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3314 ci->i_xattrs.version = version;
3315 ceph_forget_all_cached_acls(inode);
3316 ceph_security_invalidate_secctx(inode);
3320 if (newcaps & CEPH_CAP_ANY_RD) {
3321 struct timespec64 mtime, atime, ctime;
3322 /* ctime/mtime/atime? */
3323 ceph_decode_timespec64(&mtime, &grant->mtime);
3324 ceph_decode_timespec64(&atime, &grant->atime);
3325 ceph_decode_timespec64(&ctime, &grant->ctime);
3326 ceph_fill_file_time(inode, extra_info->issued,
3327 le32_to_cpu(grant->time_warp_seq),
3328 &ctime, &mtime, &atime);
3331 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3332 ci->i_files = extra_info->nfiles;
3333 ci->i_subdirs = extra_info->nsubdirs;
3336 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3337 /* file layout may have changed */
3338 s64 old_pool = ci->i_layout.pool_id;
3339 struct ceph_string *old_ns;
3341 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3342 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3343 lockdep_is_held(&ci->i_ceph_lock));
3344 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3346 if (ci->i_layout.pool_id != old_pool ||
3347 extra_info->pool_ns != old_ns)
3348 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3350 extra_info->pool_ns = old_ns;
3352 /* size/truncate_seq? */
3353 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3354 le32_to_cpu(grant->truncate_seq),
3355 le64_to_cpu(grant->truncate_size),
3359 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3360 if (max_size != ci->i_max_size) {
3361 dout("max_size %lld -> %llu\n",
3362 ci->i_max_size, max_size);
3363 ci->i_max_size = max_size;
3364 if (max_size >= ci->i_wanted_max_size) {
3365 ci->i_wanted_max_size = 0; /* reset */
3366 ci->i_requested_max_size = 0;
3372 /* check cap bits */
3373 wanted = __ceph_caps_wanted(ci);
3374 used = __ceph_caps_used(ci);
3375 dirty = __ceph_caps_dirty(ci);
3376 dout(" my wanted = %s, used = %s, dirty %s\n",
3377 ceph_cap_string(wanted),
3378 ceph_cap_string(used),
3379 ceph_cap_string(dirty));
3381 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3382 (wanted & ~(cap->mds_wanted | newcaps))) {
3384 * If mds is importing cap, prior cap messages that update
3385 * 'wanted' may get dropped by mds (migrate seq mismatch).
3387 * We don't send cap message to update 'wanted' if what we
3388 * want are already issued. If mds revokes caps, cap message
3389 * that releases caps also tells mds what we want. But if
3390 * caps got revoked by mds forcedly (session stale). We may
3391 * haven't told mds what we want.
3396 /* revocation, grant, or no-op? */
3397 if (cap->issued & ~newcaps) {
3398 int revoking = cap->issued & ~newcaps;
3400 dout("revocation: %s -> %s (revoking %s)\n",
3401 ceph_cap_string(cap->issued),
3402 ceph_cap_string(newcaps),
3403 ceph_cap_string(revoking));
3404 if (S_ISREG(inode->i_mode) &&
3405 (revoking & used & CEPH_CAP_FILE_BUFFER))
3406 writeback = true; /* initiate writeback; will delay ack */
3407 else if (queue_invalidate &&
3408 revoking == CEPH_CAP_FILE_CACHE &&
3409 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3410 ; /* do nothing yet, invalidation will be queued */
3411 else if (cap == ci->i_auth_cap)
3412 check_caps = 1; /* check auth cap only */
3414 check_caps = 2; /* check all caps */
3415 cap->issued = newcaps;
3416 cap->implemented |= newcaps;
3417 } else if (cap->issued == newcaps) {
3418 dout("caps unchanged: %s -> %s\n",
3419 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3421 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3422 ceph_cap_string(newcaps));
3423 /* non-auth MDS is revoking the newly grant caps ? */
3424 if (cap == ci->i_auth_cap &&
3425 __ceph_caps_revoking_other(ci, cap, newcaps))
3428 cap->issued = newcaps;
3429 cap->implemented |= newcaps; /* add bits only, to
3430 * avoid stepping on a
3431 * pending revocation */
3434 BUG_ON(cap->issued & ~cap->implemented);
3436 if (extra_info->inline_version > 0 &&
3437 extra_info->inline_version >= ci->i_inline_version) {
3438 ci->i_inline_version = extra_info->inline_version;
3439 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3440 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3444 if (ci->i_auth_cap == cap &&
3445 le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3446 if (newcaps & ~extra_info->issued)
3449 if (ci->i_requested_max_size > max_size ||
3450 !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3451 /* re-request max_size if necessary */
3452 ci->i_requested_max_size = 0;
3456 ceph_kick_flushing_inode_caps(session, ci);
3457 spin_unlock(&ci->i_ceph_lock);
3458 up_read(&session->s_mdsc->snap_rwsem);
3460 spin_unlock(&ci->i_ceph_lock);
3464 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3465 extra_info->inline_len);
3468 ceph_queue_vmtruncate(inode);
3472 * queue inode for writeback: we can't actually call
3473 * filemap_write_and_wait, etc. from message handler
3476 ceph_queue_writeback(inode);
3477 if (queue_invalidate)
3478 ceph_queue_invalidate(inode);
3480 invalidate_aliases(inode);
3482 wake_up_all(&ci->i_cap_wq);
3484 mutex_unlock(&session->s_mutex);
3485 if (check_caps == 1)
3486 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3488 else if (check_caps == 2)
3489 ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3493 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3494 * MDS has been safely committed.
3496 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3497 struct ceph_mds_caps *m,
3498 struct ceph_mds_session *session,
3499 struct ceph_cap *cap)
3500 __releases(ci->i_ceph_lock)
3502 struct ceph_inode_info *ci = ceph_inode(inode);
3503 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3504 struct ceph_cap_flush *cf, *tmp_cf;
3505 LIST_HEAD(to_remove);
3506 unsigned seq = le32_to_cpu(m->seq);
3507 int dirty = le32_to_cpu(m->dirty);
3510 bool wake_ci = false;
3511 bool wake_mdsc = false;
3513 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3514 /* Is this the one that was flushed? */
3515 if (cf->tid == flush_tid)
3518 /* Is this a capsnap? */
3522 if (cf->tid <= flush_tid) {
3524 * An earlier or current tid. The FLUSH_ACK should
3525 * represent a superset of this flush's caps.
3527 wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3528 list_add_tail(&cf->i_list, &to_remove);
3531 * This is a later one. Any caps in it are still dirty
3532 * so don't count them as cleaned.
3534 cleaned &= ~cf->caps;
3540 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3541 " flushing %s -> %s\n",
3542 inode, session->s_mds, seq, ceph_cap_string(dirty),
3543 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3544 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3546 if (list_empty(&to_remove) && !cleaned)
3549 ci->i_flushing_caps &= ~cleaned;
3551 spin_lock(&mdsc->cap_dirty_lock);
3553 list_for_each_entry(cf, &to_remove, i_list)
3554 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3556 if (ci->i_flushing_caps == 0) {
3557 if (list_empty(&ci->i_cap_flush_list)) {
3558 list_del_init(&ci->i_flushing_item);
3559 if (!list_empty(&session->s_cap_flushing)) {
3560 dout(" mds%d still flushing cap on %p\n",
3562 &list_first_entry(&session->s_cap_flushing,
3563 struct ceph_inode_info,
3564 i_flushing_item)->vfs_inode);
3567 mdsc->num_cap_flushing--;
3568 dout(" inode %p now !flushing\n", inode);
3570 if (ci->i_dirty_caps == 0) {
3571 dout(" inode %p now clean\n", inode);
3572 BUG_ON(!list_empty(&ci->i_dirty_item));
3574 if (ci->i_wr_ref == 0 &&
3575 ci->i_wrbuffer_ref_head == 0) {
3576 BUG_ON(!ci->i_head_snapc);
3577 ceph_put_snap_context(ci->i_head_snapc);
3578 ci->i_head_snapc = NULL;
3581 BUG_ON(list_empty(&ci->i_dirty_item));
3584 spin_unlock(&mdsc->cap_dirty_lock);
3587 spin_unlock(&ci->i_ceph_lock);
3589 while (!list_empty(&to_remove)) {
3590 cf = list_first_entry(&to_remove,
3591 struct ceph_cap_flush, i_list);
3592 list_del(&cf->i_list);
3593 ceph_free_cap_flush(cf);
3597 wake_up_all(&ci->i_cap_wq);
3599 wake_up_all(&mdsc->cap_flushing_wq);
3605 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3606 * throw away our cap_snap.
3608 * Caller hold s_mutex.
3610 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3611 struct ceph_mds_caps *m,
3612 struct ceph_mds_session *session)
3614 struct ceph_inode_info *ci = ceph_inode(inode);
3615 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3616 u64 follows = le64_to_cpu(m->snap_follows);
3617 struct ceph_cap_snap *capsnap;
3618 bool flushed = false;
3619 bool wake_ci = false;
3620 bool wake_mdsc = false;
3622 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3623 inode, ci, session->s_mds, follows);
3625 spin_lock(&ci->i_ceph_lock);
3626 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3627 if (capsnap->follows == follows) {
3628 if (capsnap->cap_flush.tid != flush_tid) {
3629 dout(" cap_snap %p follows %lld tid %lld !="
3630 " %lld\n", capsnap, follows,
3631 flush_tid, capsnap->cap_flush.tid);
3637 dout(" skipping cap_snap %p follows %lld\n",
3638 capsnap, capsnap->follows);
3642 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3643 dout(" removing %p cap_snap %p follows %lld\n",
3644 inode, capsnap, follows);
3645 list_del(&capsnap->ci_item);
3646 wake_ci |= __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3648 spin_lock(&mdsc->cap_dirty_lock);
3650 if (list_empty(&ci->i_cap_flush_list))
3651 list_del_init(&ci->i_flushing_item);
3653 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc,
3654 &capsnap->cap_flush);
3655 spin_unlock(&mdsc->cap_dirty_lock);
3657 spin_unlock(&ci->i_ceph_lock);
3659 ceph_put_snap_context(capsnap->context);
3660 ceph_put_cap_snap(capsnap);
3662 wake_up_all(&ci->i_cap_wq);
3664 wake_up_all(&mdsc->cap_flushing_wq);
3670 * Handle TRUNC from MDS, indicating file truncation.
3672 * caller hold s_mutex.
3674 static bool handle_cap_trunc(struct inode *inode,
3675 struct ceph_mds_caps *trunc,
3676 struct ceph_mds_session *session)
3678 struct ceph_inode_info *ci = ceph_inode(inode);
3679 int mds = session->s_mds;
3680 int seq = le32_to_cpu(trunc->seq);
3681 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3682 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3683 u64 size = le64_to_cpu(trunc->size);
3684 int implemented = 0;
3685 int dirty = __ceph_caps_dirty(ci);
3686 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3687 bool queue_trunc = false;
3689 lockdep_assert_held(&ci->i_ceph_lock);
3691 issued |= implemented | dirty;
3693 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3694 inode, mds, seq, truncate_size, truncate_seq);
3695 queue_trunc = ceph_fill_file_size(inode, issued,
3696 truncate_seq, truncate_size, size);
3701 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3702 * different one. If we are the most recent migration we've seen (as
3703 * indicated by mseq), make note of the migrating cap bits for the
3704 * duration (until we see the corresponding IMPORT).
3706 * caller holds s_mutex
3708 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3709 struct ceph_mds_cap_peer *ph,
3710 struct ceph_mds_session *session)
3712 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3713 struct ceph_mds_session *tsession = NULL;
3714 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3715 struct ceph_inode_info *ci = ceph_inode(inode);
3717 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3718 unsigned t_seq, t_mseq;
3720 int mds = session->s_mds;
3723 t_cap_id = le64_to_cpu(ph->cap_id);
3724 t_seq = le32_to_cpu(ph->seq);
3725 t_mseq = le32_to_cpu(ph->mseq);
3726 target = le32_to_cpu(ph->mds);
3728 t_cap_id = t_seq = t_mseq = 0;
3732 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3733 inode, ci, mds, mseq, target);
3735 spin_lock(&ci->i_ceph_lock);
3736 cap = __get_cap_for_mds(ci, mds);
3737 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3741 __ceph_remove_cap(cap, false);
3746 * now we know we haven't received the cap import message yet
3747 * because the exported cap still exist.
3750 issued = cap->issued;
3751 if (issued != cap->implemented)
3752 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3753 "ino (%llx.%llx) mds%d seq %d mseq %d "
3754 "issued %s implemented %s\n",
3755 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3756 ceph_cap_string(issued),
3757 ceph_cap_string(cap->implemented));
3760 tcap = __get_cap_for_mds(ci, target);
3762 /* already have caps from the target */
3763 if (tcap->cap_id == t_cap_id &&
3764 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3765 dout(" updating import cap %p mds%d\n", tcap, target);
3766 tcap->cap_id = t_cap_id;
3767 tcap->seq = t_seq - 1;
3768 tcap->issue_seq = t_seq - 1;
3769 tcap->issued |= issued;
3770 tcap->implemented |= issued;
3771 if (cap == ci->i_auth_cap) {
3772 ci->i_auth_cap = tcap;
3773 change_auth_cap_ses(ci, tcap->session);
3776 __ceph_remove_cap(cap, false);
3778 } else if (tsession) {
3779 /* add placeholder for the export tagert */
3780 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3782 ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3783 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3785 if (!list_empty(&ci->i_cap_flush_list) &&
3786 ci->i_auth_cap == tcap) {
3787 spin_lock(&mdsc->cap_dirty_lock);
3788 list_move_tail(&ci->i_flushing_item,
3789 &tcap->session->s_cap_flushing);
3790 spin_unlock(&mdsc->cap_dirty_lock);
3793 __ceph_remove_cap(cap, false);
3797 spin_unlock(&ci->i_ceph_lock);
3798 mutex_unlock(&session->s_mutex);
3800 /* open target session */
3801 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3802 if (!IS_ERR(tsession)) {
3804 mutex_lock(&session->s_mutex);
3805 mutex_lock_nested(&tsession->s_mutex,
3806 SINGLE_DEPTH_NESTING);
3808 mutex_lock(&tsession->s_mutex);
3809 mutex_lock_nested(&session->s_mutex,
3810 SINGLE_DEPTH_NESTING);
3812 new_cap = ceph_get_cap(mdsc, NULL);
3817 mutex_lock(&session->s_mutex);
3822 spin_unlock(&ci->i_ceph_lock);
3823 mutex_unlock(&session->s_mutex);
3825 mutex_unlock(&tsession->s_mutex);
3826 ceph_put_mds_session(tsession);
3829 ceph_put_cap(mdsc, new_cap);
3833 * Handle cap IMPORT.
3835 * caller holds s_mutex. acquires i_ceph_lock
3837 static void handle_cap_import(struct ceph_mds_client *mdsc,
3838 struct inode *inode, struct ceph_mds_caps *im,
3839 struct ceph_mds_cap_peer *ph,
3840 struct ceph_mds_session *session,
3841 struct ceph_cap **target_cap, int *old_issued)
3843 struct ceph_inode_info *ci = ceph_inode(inode);
3844 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3845 int mds = session->s_mds;
3847 unsigned caps = le32_to_cpu(im->caps);
3848 unsigned wanted = le32_to_cpu(im->wanted);
3849 unsigned seq = le32_to_cpu(im->seq);
3850 unsigned mseq = le32_to_cpu(im->migrate_seq);
3851 u64 realmino = le64_to_cpu(im->realm);
3852 u64 cap_id = le64_to_cpu(im->cap_id);
3857 p_cap_id = le64_to_cpu(ph->cap_id);
3858 peer = le32_to_cpu(ph->mds);
3864 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3865 inode, ci, mds, mseq, peer);
3867 cap = __get_cap_for_mds(ci, mds);
3870 spin_unlock(&ci->i_ceph_lock);
3871 new_cap = ceph_get_cap(mdsc, NULL);
3872 spin_lock(&ci->i_ceph_lock);
3878 ceph_put_cap(mdsc, new_cap);
3883 __ceph_caps_issued(ci, &issued);
3884 issued |= __ceph_caps_dirty(ci);
3886 ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
3887 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3889 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3890 if (ocap && ocap->cap_id == p_cap_id) {
3891 dout(" remove export cap %p mds%d flags %d\n",
3892 ocap, peer, ph->flags);
3893 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3894 (ocap->seq != le32_to_cpu(ph->seq) ||
3895 ocap->mseq != le32_to_cpu(ph->mseq))) {
3896 pr_err_ratelimited("handle_cap_import: "
3897 "mismatched seq/mseq: ino (%llx.%llx) "
3898 "mds%d seq %d mseq %d importer mds%d "
3899 "has peer seq %d mseq %d\n",
3900 ceph_vinop(inode), peer, ocap->seq,
3901 ocap->mseq, mds, le32_to_cpu(ph->seq),
3902 le32_to_cpu(ph->mseq));
3904 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3907 *old_issued = issued;
3912 * Handle a caps message from the MDS.
3914 * Identify the appropriate session, inode, and call the right handler
3915 * based on the cap op.
3917 void ceph_handle_caps(struct ceph_mds_session *session,
3918 struct ceph_msg *msg)
3920 struct ceph_mds_client *mdsc = session->s_mdsc;
3921 struct inode *inode;
3922 struct ceph_inode_info *ci;
3923 struct ceph_cap *cap;
3924 struct ceph_mds_caps *h;
3925 struct ceph_mds_cap_peer *peer = NULL;
3926 struct ceph_snap_realm *realm = NULL;
3928 int msg_version = le16_to_cpu(msg->hdr.version);
3930 struct ceph_vino vino;
3932 size_t snaptrace_len;
3934 struct cap_extra_info extra_info = {};
3937 dout("handle_caps from mds%d\n", session->s_mds);
3940 end = msg->front.iov_base + msg->front.iov_len;
3941 if (msg->front.iov_len < sizeof(*h))
3943 h = msg->front.iov_base;
3944 op = le32_to_cpu(h->op);
3945 vino.ino = le64_to_cpu(h->ino);
3946 vino.snap = CEPH_NOSNAP;
3947 seq = le32_to_cpu(h->seq);
3948 mseq = le32_to_cpu(h->migrate_seq);
3951 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3952 p = snaptrace + snaptrace_len;
3954 if (msg_version >= 2) {
3956 ceph_decode_32_safe(&p, end, flock_len, bad);
3957 if (p + flock_len > end)
3962 if (msg_version >= 3) {
3963 if (op == CEPH_CAP_OP_IMPORT) {
3964 if (p + sizeof(*peer) > end)
3968 } else if (op == CEPH_CAP_OP_EXPORT) {
3969 /* recorded in unused fields */
3970 peer = (void *)&h->size;
3974 if (msg_version >= 4) {
3975 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
3976 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
3977 if (p + extra_info.inline_len > end)
3979 extra_info.inline_data = p;
3980 p += extra_info.inline_len;
3983 if (msg_version >= 5) {
3984 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
3987 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
3988 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
3991 if (msg_version >= 8) {
3995 ceph_decode_skip_64(&p, end, bad); // flush_tid
3997 ceph_decode_skip_32(&p, end, bad); // caller_uid
3998 ceph_decode_skip_32(&p, end, bad); // caller_gid
4000 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4001 if (pool_ns_len > 0) {
4002 ceph_decode_need(&p, end, pool_ns_len, bad);
4003 extra_info.pool_ns =
4004 ceph_find_or_create_string(p, pool_ns_len);
4009 if (msg_version >= 9) {
4010 struct ceph_timespec *btime;
4012 if (p + sizeof(*btime) > end)
4015 ceph_decode_timespec64(&extra_info.btime, btime);
4016 p += sizeof(*btime);
4017 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4020 if (msg_version >= 11) {
4022 ceph_decode_skip_32(&p, end, bad); // flags
4024 extra_info.dirstat_valid = true;
4025 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4026 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4030 inode = ceph_find_inode(mdsc->fsc->sb, vino);
4031 ci = ceph_inode(inode);
4032 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4035 mutex_lock(&session->s_mutex);
4036 inc_session_sequence(session);
4037 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4041 dout(" i don't have ino %llx\n", vino.ino);
4043 if (op == CEPH_CAP_OP_IMPORT) {
4044 cap = ceph_get_cap(mdsc, NULL);
4045 cap->cap_ino = vino.ino;
4046 cap->queue_release = 1;
4047 cap->cap_id = le64_to_cpu(h->cap_id);
4050 cap->issue_seq = seq;
4051 spin_lock(&session->s_cap_lock);
4052 __ceph_queue_cap_release(session, cap);
4053 spin_unlock(&session->s_cap_lock);
4055 goto flush_cap_releases;
4058 /* these will work even if we don't have a cap yet */
4060 case CEPH_CAP_OP_FLUSHSNAP_ACK:
4061 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4065 case CEPH_CAP_OP_EXPORT:
4066 handle_cap_export(inode, h, peer, session);
4069 case CEPH_CAP_OP_IMPORT:
4071 if (snaptrace_len) {
4072 down_write(&mdsc->snap_rwsem);
4073 ceph_update_snap_trace(mdsc, snaptrace,
4074 snaptrace + snaptrace_len,
4076 downgrade_write(&mdsc->snap_rwsem);
4078 down_read(&mdsc->snap_rwsem);
4080 spin_lock(&ci->i_ceph_lock);
4081 handle_cap_import(mdsc, inode, h, peer, session,
4082 &cap, &extra_info.issued);
4083 handle_cap_grant(inode, session, cap,
4084 h, msg->middle, &extra_info);
4086 ceph_put_snap_realm(mdsc, realm);
4090 /* the rest require a cap */
4091 spin_lock(&ci->i_ceph_lock);
4092 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4094 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4095 inode, ceph_ino(inode), ceph_snap(inode),
4097 spin_unlock(&ci->i_ceph_lock);
4098 goto flush_cap_releases;
4101 /* note that each of these drops i_ceph_lock for us */
4103 case CEPH_CAP_OP_REVOKE:
4104 case CEPH_CAP_OP_GRANT:
4105 __ceph_caps_issued(ci, &extra_info.issued);
4106 extra_info.issued |= __ceph_caps_dirty(ci);
4107 handle_cap_grant(inode, session, cap,
4108 h, msg->middle, &extra_info);
4111 case CEPH_CAP_OP_FLUSH_ACK:
4112 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4116 case CEPH_CAP_OP_TRUNC:
4117 queue_trunc = handle_cap_trunc(inode, h, session);
4118 spin_unlock(&ci->i_ceph_lock);
4120 ceph_queue_vmtruncate(inode);
4124 spin_unlock(&ci->i_ceph_lock);
4125 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4126 ceph_cap_op_name(op));
4130 mutex_unlock(&session->s_mutex);
4132 ceph_put_string(extra_info.pool_ns);
4138 * send any cap release message to try to move things
4139 * along for the mds (who clearly thinks we still have this
4142 ceph_flush_cap_releases(mdsc, session);
4146 pr_err("ceph_handle_caps: corrupt message\n");
4152 * Delayed work handler to process end of delayed cap release LRU list.
4154 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4156 struct inode *inode;
4157 struct ceph_inode_info *ci;
4159 dout("check_delayed_caps\n");
4160 spin_lock(&mdsc->cap_delay_lock);
4161 while (!list_empty(&mdsc->cap_delay_list)) {
4162 ci = list_first_entry(&mdsc->cap_delay_list,
4163 struct ceph_inode_info,
4165 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4166 time_before(jiffies, ci->i_hold_caps_max))
4168 list_del_init(&ci->i_cap_delay_list);
4170 inode = igrab(&ci->vfs_inode);
4172 spin_unlock(&mdsc->cap_delay_lock);
4173 dout("check_delayed_caps on %p\n", inode);
4174 ceph_check_caps(ci, 0, NULL);
4176 spin_lock(&mdsc->cap_delay_lock);
4179 spin_unlock(&mdsc->cap_delay_lock);
4183 * Flush all dirty caps to the mds
4185 static void flush_dirty_session_caps(struct ceph_mds_session *s)
4187 struct ceph_mds_client *mdsc = s->s_mdsc;
4188 struct ceph_inode_info *ci;
4189 struct inode *inode;
4191 dout("flush_dirty_caps\n");
4192 spin_lock(&mdsc->cap_dirty_lock);
4193 while (!list_empty(&s->s_cap_dirty)) {
4194 ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4196 inode = &ci->vfs_inode;
4198 dout("flush_dirty_caps %p\n", inode);
4199 spin_unlock(&mdsc->cap_dirty_lock);
4200 ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4202 spin_lock(&mdsc->cap_dirty_lock);
4204 spin_unlock(&mdsc->cap_dirty_lock);
4205 dout("flush_dirty_caps done\n");
4208 static void iterate_sessions(struct ceph_mds_client *mdsc,
4209 void (*cb)(struct ceph_mds_session *))
4213 mutex_lock(&mdsc->mutex);
4214 for (mds = 0; mds < mdsc->max_sessions; ++mds) {
4215 struct ceph_mds_session *s;
4217 if (!mdsc->sessions[mds])
4220 s = ceph_get_mds_session(mdsc->sessions[mds]);
4224 mutex_unlock(&mdsc->mutex);
4226 ceph_put_mds_session(s);
4227 mutex_lock(&mdsc->mutex);
4229 mutex_unlock(&mdsc->mutex);
4232 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4234 iterate_sessions(mdsc, flush_dirty_session_caps);
4237 void __ceph_touch_fmode(struct ceph_inode_info *ci,
4238 struct ceph_mds_client *mdsc, int fmode)
4240 unsigned long now = jiffies;
4241 if (fmode & CEPH_FILE_MODE_RD)
4242 ci->i_last_rd = now;
4243 if (fmode & CEPH_FILE_MODE_WR)
4244 ci->i_last_wr = now;
4245 /* queue periodic check */
4247 __ceph_is_any_real_caps(ci) &&
4248 list_empty(&ci->i_cap_delay_list))
4249 __cap_delay_requeue(mdsc, ci);
4252 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4254 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4255 int bits = (fmode << 1) | 1;
4256 bool is_opened = false;
4260 atomic64_inc(&mdsc->metric.opened_files);
4262 spin_lock(&ci->i_ceph_lock);
4263 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4264 if (bits & (1 << i))
4265 ci->i_nr_by_mode[i] += count;
4268 * If any of the mode ref is larger than 1,
4269 * that means it has been already opened by
4270 * others. Just skip checking the PIN ref.
4272 if (i && ci->i_nr_by_mode[i] > 1)
4277 percpu_counter_inc(&mdsc->metric.opened_inodes);
4278 spin_unlock(&ci->i_ceph_lock);
4282 * Drop open file reference. If we were the last open file,
4283 * we may need to release capabilities to the MDS (or schedule
4284 * their delayed release).
4286 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4288 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4289 int bits = (fmode << 1) | 1;
4290 bool is_closed = true;
4294 atomic64_dec(&mdsc->metric.opened_files);
4296 spin_lock(&ci->i_ceph_lock);
4297 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4298 if (bits & (1 << i)) {
4299 BUG_ON(ci->i_nr_by_mode[i] < count);
4300 ci->i_nr_by_mode[i] -= count;
4304 * If any of the mode ref is not 0 after
4305 * decreased, that means it is still opened
4306 * by others. Just skip checking the PIN ref.
4308 if (i && ci->i_nr_by_mode[i])
4313 percpu_counter_dec(&mdsc->metric.opened_inodes);
4314 spin_unlock(&ci->i_ceph_lock);
4318 * For a soon-to-be unlinked file, drop the LINK caps. If it
4319 * looks like the link count will hit 0, drop any other caps (other
4320 * than PIN) we don't specifically want (due to the file still being
4323 int ceph_drop_caps_for_unlink(struct inode *inode)
4325 struct ceph_inode_info *ci = ceph_inode(inode);
4326 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4328 spin_lock(&ci->i_ceph_lock);
4329 if (inode->i_nlink == 1) {
4330 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4332 if (__ceph_caps_dirty(ci)) {
4333 struct ceph_mds_client *mdsc =
4334 ceph_inode_to_client(inode)->mdsc;
4335 __cap_delay_requeue_front(mdsc, ci);
4338 spin_unlock(&ci->i_ceph_lock);
4343 * Helpers for embedding cap and dentry lease releases into mds
4346 * @force is used by dentry_release (below) to force inclusion of a
4347 * record for the directory inode, even when there aren't any caps to
4350 int ceph_encode_inode_release(void **p, struct inode *inode,
4351 int mds, int drop, int unless, int force)
4353 struct ceph_inode_info *ci = ceph_inode(inode);
4354 struct ceph_cap *cap;
4355 struct ceph_mds_request_release *rel = *p;
4359 spin_lock(&ci->i_ceph_lock);
4360 used = __ceph_caps_used(ci);
4361 dirty = __ceph_caps_dirty(ci);
4363 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4364 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4365 ceph_cap_string(unless));
4367 /* only drop unused, clean caps */
4368 drop &= ~(used | dirty);
4370 cap = __get_cap_for_mds(ci, mds);
4371 if (cap && __cap_is_valid(cap)) {
4372 unless &= cap->issued;
4374 if (unless & CEPH_CAP_AUTH_EXCL)
4375 drop &= ~CEPH_CAP_AUTH_SHARED;
4376 if (unless & CEPH_CAP_LINK_EXCL)
4377 drop &= ~CEPH_CAP_LINK_SHARED;
4378 if (unless & CEPH_CAP_XATTR_EXCL)
4379 drop &= ~CEPH_CAP_XATTR_SHARED;
4380 if (unless & CEPH_CAP_FILE_EXCL)
4381 drop &= ~CEPH_CAP_FILE_SHARED;
4384 if (force || (cap->issued & drop)) {
4385 if (cap->issued & drop) {
4386 int wanted = __ceph_caps_wanted(ci);
4387 dout("encode_inode_release %p cap %p "
4388 "%s -> %s, wanted %s -> %s\n", inode, cap,
4389 ceph_cap_string(cap->issued),
4390 ceph_cap_string(cap->issued & ~drop),
4391 ceph_cap_string(cap->mds_wanted),
4392 ceph_cap_string(wanted));
4394 cap->issued &= ~drop;
4395 cap->implemented &= ~drop;
4396 cap->mds_wanted = wanted;
4397 if (cap == ci->i_auth_cap &&
4398 !(wanted & CEPH_CAP_ANY_FILE_WR))
4399 ci->i_requested_max_size = 0;
4401 dout("encode_inode_release %p cap %p %s"
4402 " (force)\n", inode, cap,
4403 ceph_cap_string(cap->issued));
4406 rel->ino = cpu_to_le64(ceph_ino(inode));
4407 rel->cap_id = cpu_to_le64(cap->cap_id);
4408 rel->seq = cpu_to_le32(cap->seq);
4409 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4410 rel->mseq = cpu_to_le32(cap->mseq);
4411 rel->caps = cpu_to_le32(cap->implemented);
4412 rel->wanted = cpu_to_le32(cap->mds_wanted);
4418 dout("encode_inode_release %p cap %p %s (noop)\n",
4419 inode, cap, ceph_cap_string(cap->issued));
4422 spin_unlock(&ci->i_ceph_lock);
4426 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4428 int mds, int drop, int unless)
4430 struct dentry *parent = NULL;
4431 struct ceph_mds_request_release *rel = *p;
4432 struct ceph_dentry_info *di = ceph_dentry(dentry);
4437 * force an record for the directory caps if we have a dentry lease.
4438 * this is racy (can't take i_ceph_lock and d_lock together), but it
4439 * doesn't have to be perfect; the mds will revoke anything we don't
4442 spin_lock(&dentry->d_lock);
4443 if (di->lease_session && di->lease_session->s_mds == mds)
4446 parent = dget(dentry->d_parent);
4447 dir = d_inode(parent);
4449 spin_unlock(&dentry->d_lock);
4451 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4454 spin_lock(&dentry->d_lock);
4455 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4456 dout("encode_dentry_release %p mds%d seq %d\n",
4457 dentry, mds, (int)di->lease_seq);
4458 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4459 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4460 *p += dentry->d_name.len;
4461 rel->dname_seq = cpu_to_le32(di->lease_seq);
4462 __ceph_mdsc_drop_dentry_lease(dentry);
4464 spin_unlock(&dentry->d_lock);