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. Takes s_mutex as needed.
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 mutex_unlock(&session->s_mutex);
1657 ceph_put_mds_session(session);
1661 spin_unlock(&ci->i_ceph_lock);
1662 mutex_lock(&mdsc->mutex);
1663 session = __ceph_lookup_mds_session(mdsc, mds);
1664 mutex_unlock(&mdsc->mutex);
1666 dout(" inverting session/ino locks on %p\n", session);
1667 mutex_lock(&session->s_mutex);
1672 // make sure flushsnap messages are sent in proper order.
1673 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1674 __kick_flushing_caps(mdsc, session, ci, 0);
1676 __ceph_flush_snaps(ci, session);
1678 spin_unlock(&ci->i_ceph_lock);
1681 *psession = session;
1682 } else if (session) {
1683 mutex_unlock(&session->s_mutex);
1684 ceph_put_mds_session(session);
1686 /* we flushed them all; remove this inode from the queue */
1687 spin_lock(&mdsc->snap_flush_lock);
1688 list_del_init(&ci->i_snap_flush_item);
1689 spin_unlock(&mdsc->snap_flush_lock);
1693 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1694 * Caller is then responsible for calling __mark_inode_dirty with the
1695 * returned flags value.
1697 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1698 struct ceph_cap_flush **pcf)
1700 struct ceph_mds_client *mdsc =
1701 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1702 struct inode *inode = &ci->vfs_inode;
1703 int was = ci->i_dirty_caps;
1706 lockdep_assert_held(&ci->i_ceph_lock);
1708 if (!ci->i_auth_cap) {
1709 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1710 "but no auth cap (session was closed?)\n",
1711 inode, ceph_ino(inode), ceph_cap_string(mask));
1715 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1716 ceph_cap_string(mask), ceph_cap_string(was),
1717 ceph_cap_string(was | mask));
1718 ci->i_dirty_caps |= mask;
1720 struct ceph_mds_session *session = ci->i_auth_cap->session;
1722 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1723 swap(ci->i_prealloc_cap_flush, *pcf);
1725 if (!ci->i_head_snapc) {
1726 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1727 ci->i_head_snapc = ceph_get_snap_context(
1728 ci->i_snap_realm->cached_context);
1730 dout(" inode %p now dirty snapc %p auth cap %p\n",
1731 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1732 BUG_ON(!list_empty(&ci->i_dirty_item));
1733 spin_lock(&mdsc->cap_dirty_lock);
1734 list_add(&ci->i_dirty_item, &session->s_cap_dirty);
1735 spin_unlock(&mdsc->cap_dirty_lock);
1736 if (ci->i_flushing_caps == 0) {
1738 dirty |= I_DIRTY_SYNC;
1741 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1743 BUG_ON(list_empty(&ci->i_dirty_item));
1744 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1745 (mask & CEPH_CAP_FILE_BUFFER))
1746 dirty |= I_DIRTY_DATASYNC;
1747 __cap_delay_requeue(mdsc, ci);
1751 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1753 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1756 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1759 kmem_cache_free(ceph_cap_flush_cachep, cf);
1762 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1764 if (!list_empty(&mdsc->cap_flush_list)) {
1765 struct ceph_cap_flush *cf =
1766 list_first_entry(&mdsc->cap_flush_list,
1767 struct ceph_cap_flush, g_list);
1774 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1775 * Return true if caller needs to wake up flush waiters.
1777 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1778 struct ceph_cap_flush *cf)
1780 struct ceph_cap_flush *prev;
1781 bool wake = cf->wake;
1783 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1784 prev = list_prev_entry(cf, g_list);
1788 list_del(&cf->g_list);
1792 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1793 struct ceph_cap_flush *cf)
1795 struct ceph_cap_flush *prev;
1796 bool wake = cf->wake;
1798 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1799 prev = list_prev_entry(cf, i_list);
1803 list_del(&cf->i_list);
1808 * Add dirty inode to the flushing list. Assigned a seq number so we
1809 * can wait for caps to flush without starving.
1811 * Called under i_ceph_lock. Returns the flush tid.
1813 static u64 __mark_caps_flushing(struct inode *inode,
1814 struct ceph_mds_session *session, bool wake,
1815 u64 *oldest_flush_tid)
1817 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1818 struct ceph_inode_info *ci = ceph_inode(inode);
1819 struct ceph_cap_flush *cf = NULL;
1822 lockdep_assert_held(&ci->i_ceph_lock);
1823 BUG_ON(ci->i_dirty_caps == 0);
1824 BUG_ON(list_empty(&ci->i_dirty_item));
1825 BUG_ON(!ci->i_prealloc_cap_flush);
1827 flushing = ci->i_dirty_caps;
1828 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1829 ceph_cap_string(flushing),
1830 ceph_cap_string(ci->i_flushing_caps),
1831 ceph_cap_string(ci->i_flushing_caps | flushing));
1832 ci->i_flushing_caps |= flushing;
1833 ci->i_dirty_caps = 0;
1834 dout(" inode %p now !dirty\n", inode);
1836 swap(cf, ci->i_prealloc_cap_flush);
1837 cf->caps = flushing;
1840 spin_lock(&mdsc->cap_dirty_lock);
1841 list_del_init(&ci->i_dirty_item);
1843 cf->tid = ++mdsc->last_cap_flush_tid;
1844 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1845 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1847 if (list_empty(&ci->i_flushing_item)) {
1848 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1849 mdsc->num_cap_flushing++;
1851 spin_unlock(&mdsc->cap_dirty_lock);
1853 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1859 * try to invalidate mapping pages without blocking.
1861 static int try_nonblocking_invalidate(struct inode *inode)
1863 struct ceph_inode_info *ci = ceph_inode(inode);
1864 u32 invalidating_gen = ci->i_rdcache_gen;
1866 spin_unlock(&ci->i_ceph_lock);
1867 ceph_fscache_invalidate(inode);
1868 invalidate_mapping_pages(&inode->i_data, 0, -1);
1869 spin_lock(&ci->i_ceph_lock);
1871 if (inode->i_data.nrpages == 0 &&
1872 invalidating_gen == ci->i_rdcache_gen) {
1874 dout("try_nonblocking_invalidate %p success\n", inode);
1875 /* save any racing async invalidate some trouble */
1876 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1879 dout("try_nonblocking_invalidate %p failed\n", inode);
1883 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1885 loff_t size = i_size_read(&ci->vfs_inode);
1886 /* mds will adjust max size according to the reported size */
1887 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1889 if (size >= ci->i_max_size)
1891 /* half of previous max_size increment has been used */
1892 if (ci->i_max_size > ci->i_reported_size &&
1893 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1899 * Swiss army knife function to examine currently used and wanted
1900 * versus held caps. Release, flush, ack revoked caps to mds as
1903 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1904 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1907 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1908 struct ceph_mds_session *session)
1910 struct inode *inode = &ci->vfs_inode;
1911 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1912 struct ceph_cap *cap;
1913 u64 flush_tid, oldest_flush_tid;
1914 int file_wanted, used, cap_used;
1915 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1916 int issued, implemented, want, retain, revoking, flushing = 0;
1917 int mds = -1; /* keep track of how far we've gone through i_caps list
1918 to avoid an infinite loop on retry */
1920 bool queue_invalidate = false;
1921 bool tried_invalidate = false;
1923 spin_lock(&ci->i_ceph_lock);
1924 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1925 flags |= CHECK_CAPS_FLUSH;
1929 spin_lock(&ci->i_ceph_lock);
1931 /* Caps wanted by virtue of active open files. */
1932 file_wanted = __ceph_caps_file_wanted(ci);
1934 /* Caps which have active references against them */
1935 used = __ceph_caps_used(ci);
1938 * "issued" represents the current caps that the MDS wants us to have.
1939 * "implemented" is the set that we have been granted, and includes the
1940 * ones that have not yet been returned to the MDS (the "revoking" set,
1941 * usually because they have outstanding references).
1943 issued = __ceph_caps_issued(ci, &implemented);
1944 revoking = implemented & ~issued;
1948 /* The ones we currently want to retain (may be adjusted below) */
1949 retain = file_wanted | used | CEPH_CAP_PIN;
1950 if (!mdsc->stopping && inode->i_nlink > 0) {
1952 retain |= CEPH_CAP_ANY; /* be greedy */
1953 } else if (S_ISDIR(inode->i_mode) &&
1954 (issued & CEPH_CAP_FILE_SHARED) &&
1955 __ceph_dir_is_complete(ci)) {
1957 * If a directory is complete, we want to keep
1958 * the exclusive cap. So that MDS does not end up
1959 * revoking the shared cap on every create/unlink
1962 if (IS_RDONLY(inode)) {
1963 want = CEPH_CAP_ANY_SHARED;
1965 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1970 retain |= CEPH_CAP_ANY_SHARED;
1972 * keep RD only if we didn't have the file open RW,
1973 * because then the mds would revoke it anyway to
1974 * journal max_size=0.
1976 if (ci->i_max_size == 0)
1977 retain |= CEPH_CAP_ANY_RD;
1981 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1982 " issued %s revoking %s retain %s %s%s\n", inode,
1983 ceph_cap_string(file_wanted),
1984 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1985 ceph_cap_string(ci->i_flushing_caps),
1986 ceph_cap_string(issued), ceph_cap_string(revoking),
1987 ceph_cap_string(retain),
1988 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1989 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1992 * If we no longer need to hold onto old our caps, and we may
1993 * have cached pages, but don't want them, then try to invalidate.
1994 * If we fail, it's because pages are locked.... try again later.
1996 if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
1997 S_ISREG(inode->i_mode) &&
1998 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1999 inode->i_data.nrpages && /* have cached pages */
2000 (revoking & (CEPH_CAP_FILE_CACHE|
2001 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
2002 !tried_invalidate) {
2003 dout("check_caps trying to invalidate on %p\n", inode);
2004 if (try_nonblocking_invalidate(inode) < 0) {
2005 dout("check_caps queuing invalidate\n");
2006 queue_invalidate = true;
2007 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2009 tried_invalidate = true;
2013 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2015 struct cap_msg_args arg;
2017 cap = rb_entry(p, struct ceph_cap, ci_node);
2019 /* avoid looping forever */
2020 if (mds >= cap->mds ||
2021 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
2024 /* NOTE: no side-effects allowed, until we take s_mutex */
2027 * If we have an auth cap, we don't need to consider any
2028 * overlapping caps as used.
2031 if (ci->i_auth_cap && cap != ci->i_auth_cap)
2032 cap_used &= ~ci->i_auth_cap->issued;
2034 revoking = cap->implemented & ~cap->issued;
2035 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
2036 cap->mds, cap, ceph_cap_string(cap_used),
2037 ceph_cap_string(cap->issued),
2038 ceph_cap_string(cap->implemented),
2039 ceph_cap_string(revoking));
2041 if (cap == ci->i_auth_cap &&
2042 (cap->issued & CEPH_CAP_FILE_WR)) {
2043 /* request larger max_size from MDS? */
2044 if (ci->i_wanted_max_size > ci->i_max_size &&
2045 ci->i_wanted_max_size > ci->i_requested_max_size) {
2046 dout("requesting new max_size\n");
2050 /* approaching file_max? */
2051 if (__ceph_should_report_size(ci)) {
2052 dout("i_size approaching max_size\n");
2056 /* flush anything dirty? */
2057 if (cap == ci->i_auth_cap) {
2058 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
2059 dout("flushing dirty caps\n");
2062 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
2063 dout("flushing snap caps\n");
2068 /* completed revocation? going down and there are no caps? */
2069 if (revoking && (revoking & cap_used) == 0) {
2070 dout("completed revocation of %s\n",
2071 ceph_cap_string(cap->implemented & ~cap->issued));
2075 /* want more caps from mds? */
2076 if (want & ~cap->mds_wanted) {
2077 if (want & ~(cap->mds_wanted | cap->issued))
2079 if (!__cap_is_valid(cap))
2083 /* things we might delay */
2084 if ((cap->issued & ~retain) == 0)
2085 continue; /* nope, all good */
2088 if (session && session != cap->session) {
2089 dout("oops, wrong session %p mutex\n", session);
2090 mutex_unlock(&session->s_mutex);
2094 session = cap->session;
2095 if (mutex_trylock(&session->s_mutex) == 0) {
2096 dout("inverting session/ino locks on %p\n",
2098 session = ceph_get_mds_session(session);
2099 spin_unlock(&ci->i_ceph_lock);
2100 if (took_snap_rwsem) {
2101 up_read(&mdsc->snap_rwsem);
2102 took_snap_rwsem = 0;
2105 mutex_lock(&session->s_mutex);
2106 ceph_put_mds_session(session);
2109 * Because we take the reference while
2110 * holding the i_ceph_lock, it should
2111 * never be NULL. Throw a warning if it
2120 /* kick flushing and flush snaps before sending normal
2122 if (cap == ci->i_auth_cap &&
2124 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2125 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2126 __kick_flushing_caps(mdsc, session, ci, 0);
2127 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2128 __ceph_flush_snaps(ci, session);
2133 /* take snap_rwsem after session mutex */
2134 if (!took_snap_rwsem) {
2135 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
2136 dout("inverting snap/in locks on %p\n",
2138 spin_unlock(&ci->i_ceph_lock);
2139 down_read(&mdsc->snap_rwsem);
2140 took_snap_rwsem = 1;
2143 took_snap_rwsem = 1;
2146 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2147 flushing = ci->i_dirty_caps;
2148 flush_tid = __mark_caps_flushing(inode, session, false,
2150 if (flags & CHECK_CAPS_FLUSH &&
2151 list_empty(&session->s_cap_dirty))
2152 mflags |= CEPH_CLIENT_CAPS_SYNC;
2156 spin_lock(&mdsc->cap_dirty_lock);
2157 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2158 spin_unlock(&mdsc->cap_dirty_lock);
2161 mds = cap->mds; /* remember mds, so we don't repeat */
2163 __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2164 want, retain, flushing, flush_tid, oldest_flush_tid);
2165 spin_unlock(&ci->i_ceph_lock);
2167 __send_cap(&arg, ci);
2169 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2172 /* periodically re-calculate caps wanted by open files */
2173 if (__ceph_is_any_real_caps(ci) &&
2174 list_empty(&ci->i_cap_delay_list) &&
2175 (file_wanted & ~CEPH_CAP_PIN) &&
2176 !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2177 __cap_delay_requeue(mdsc, ci);
2180 spin_unlock(&ci->i_ceph_lock);
2182 if (queue_invalidate)
2183 ceph_queue_invalidate(inode);
2186 mutex_unlock(&session->s_mutex);
2187 if (took_snap_rwsem)
2188 up_read(&mdsc->snap_rwsem);
2192 * Try to flush dirty caps back to the auth mds.
2194 static int try_flush_caps(struct inode *inode, u64 *ptid)
2196 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2197 struct ceph_inode_info *ci = ceph_inode(inode);
2198 struct ceph_mds_session *session = NULL;
2200 u64 flush_tid = 0, oldest_flush_tid = 0;
2203 spin_lock(&ci->i_ceph_lock);
2205 if (ci->i_dirty_caps && ci->i_auth_cap) {
2206 struct ceph_cap *cap = ci->i_auth_cap;
2207 struct cap_msg_args arg;
2209 if (session != cap->session) {
2210 spin_unlock(&ci->i_ceph_lock);
2212 mutex_unlock(&session->s_mutex);
2213 session = cap->session;
2214 mutex_lock(&session->s_mutex);
2217 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) {
2218 spin_unlock(&ci->i_ceph_lock);
2222 if (ci->i_ceph_flags &
2223 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2224 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2225 __kick_flushing_caps(mdsc, session, ci, 0);
2226 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2227 __ceph_flush_snaps(ci, session);
2231 flushing = ci->i_dirty_caps;
2232 flush_tid = __mark_caps_flushing(inode, session, true,
2235 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2236 __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2237 (cap->issued | cap->implemented),
2238 flushing, flush_tid, oldest_flush_tid);
2239 spin_unlock(&ci->i_ceph_lock);
2241 __send_cap(&arg, ci);
2243 if (!list_empty(&ci->i_cap_flush_list)) {
2244 struct ceph_cap_flush *cf =
2245 list_last_entry(&ci->i_cap_flush_list,
2246 struct ceph_cap_flush, i_list);
2248 flush_tid = cf->tid;
2250 flushing = ci->i_flushing_caps;
2251 spin_unlock(&ci->i_ceph_lock);
2255 mutex_unlock(&session->s_mutex);
2262 * Return true if we've flushed caps through the given flush_tid.
2264 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2266 struct ceph_inode_info *ci = ceph_inode(inode);
2269 spin_lock(&ci->i_ceph_lock);
2270 if (!list_empty(&ci->i_cap_flush_list)) {
2271 struct ceph_cap_flush * cf =
2272 list_first_entry(&ci->i_cap_flush_list,
2273 struct ceph_cap_flush, i_list);
2274 if (cf->tid <= flush_tid)
2277 spin_unlock(&ci->i_ceph_lock);
2282 * wait for any unsafe requests to complete.
2284 static int unsafe_request_wait(struct inode *inode)
2286 struct ceph_inode_info *ci = ceph_inode(inode);
2287 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2290 spin_lock(&ci->i_unsafe_lock);
2291 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2292 req1 = list_last_entry(&ci->i_unsafe_dirops,
2293 struct ceph_mds_request,
2295 ceph_mdsc_get_request(req1);
2297 if (!list_empty(&ci->i_unsafe_iops)) {
2298 req2 = list_last_entry(&ci->i_unsafe_iops,
2299 struct ceph_mds_request,
2300 r_unsafe_target_item);
2301 ceph_mdsc_get_request(req2);
2303 spin_unlock(&ci->i_unsafe_lock);
2305 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2306 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2308 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2309 ceph_timeout_jiffies(req1->r_timeout));
2312 ceph_mdsc_put_request(req1);
2315 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2316 ceph_timeout_jiffies(req2->r_timeout));
2319 ceph_mdsc_put_request(req2);
2324 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2326 struct ceph_file_info *fi = file->private_data;
2327 struct inode *inode = file->f_mapping->host;
2328 struct ceph_inode_info *ci = ceph_inode(inode);
2333 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2335 ret = file_write_and_wait_range(file, start, end);
2339 ret = ceph_wait_on_async_create(inode);
2343 dirty = try_flush_caps(inode, &flush_tid);
2344 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2346 err = unsafe_request_wait(inode);
2349 * only wait on non-file metadata writeback (the mds
2350 * can recover size and mtime, so we don't need to
2353 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2354 err = wait_event_interruptible(ci->i_cap_wq,
2355 caps_are_flushed(inode, flush_tid));
2361 if (errseq_check(&ci->i_meta_err, READ_ONCE(fi->meta_err))) {
2362 spin_lock(&file->f_lock);
2363 err = errseq_check_and_advance(&ci->i_meta_err,
2365 spin_unlock(&file->f_lock);
2370 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2375 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2376 * queue inode for flush but don't do so immediately, because we can
2377 * get by with fewer MDS messages if we wait for data writeback to
2380 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2382 struct ceph_inode_info *ci = ceph_inode(inode);
2386 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2388 dout("write_inode %p wait=%d\n", inode, wait);
2390 dirty = try_flush_caps(inode, &flush_tid);
2392 err = wait_event_interruptible(ci->i_cap_wq,
2393 caps_are_flushed(inode, flush_tid));
2395 struct ceph_mds_client *mdsc =
2396 ceph_sb_to_client(inode->i_sb)->mdsc;
2398 spin_lock(&ci->i_ceph_lock);
2399 if (__ceph_caps_dirty(ci))
2400 __cap_delay_requeue_front(mdsc, ci);
2401 spin_unlock(&ci->i_ceph_lock);
2406 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2407 struct ceph_mds_session *session,
2408 struct ceph_inode_info *ci,
2409 u64 oldest_flush_tid)
2410 __releases(ci->i_ceph_lock)
2411 __acquires(ci->i_ceph_lock)
2413 struct inode *inode = &ci->vfs_inode;
2414 struct ceph_cap *cap;
2415 struct ceph_cap_flush *cf;
2418 u64 last_snap_flush = 0;
2420 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2422 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2424 last_snap_flush = cf->tid;
2429 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2430 if (cf->tid < first_tid)
2433 cap = ci->i_auth_cap;
2434 if (!(cap && cap->session == session)) {
2435 pr_err("%p auth cap %p not mds%d ???\n",
2436 inode, cap, session->s_mds);
2440 first_tid = cf->tid + 1;
2443 struct cap_msg_args arg;
2445 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2446 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2447 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2448 (cf->tid < last_snap_flush ?
2449 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2450 __ceph_caps_used(ci),
2451 __ceph_caps_wanted(ci),
2452 (cap->issued | cap->implemented),
2453 cf->caps, cf->tid, oldest_flush_tid);
2454 spin_unlock(&ci->i_ceph_lock);
2455 __send_cap(&arg, ci);
2457 struct ceph_cap_snap *capsnap =
2458 container_of(cf, struct ceph_cap_snap,
2460 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2461 inode, capsnap, cf->tid,
2462 ceph_cap_string(capsnap->dirty));
2464 refcount_inc(&capsnap->nref);
2465 spin_unlock(&ci->i_ceph_lock);
2467 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2470 pr_err("kick_flushing_caps: error sending "
2471 "cap flushsnap, ino (%llx.%llx) "
2472 "tid %llu follows %llu\n",
2473 ceph_vinop(inode), cf->tid,
2477 ceph_put_cap_snap(capsnap);
2480 spin_lock(&ci->i_ceph_lock);
2484 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2485 struct ceph_mds_session *session)
2487 struct ceph_inode_info *ci;
2488 struct ceph_cap *cap;
2489 u64 oldest_flush_tid;
2491 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2493 spin_lock(&mdsc->cap_dirty_lock);
2494 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2495 spin_unlock(&mdsc->cap_dirty_lock);
2497 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2498 spin_lock(&ci->i_ceph_lock);
2499 cap = ci->i_auth_cap;
2500 if (!(cap && cap->session == session)) {
2501 pr_err("%p auth cap %p not mds%d ???\n",
2502 &ci->vfs_inode, cap, session->s_mds);
2503 spin_unlock(&ci->i_ceph_lock);
2509 * if flushing caps were revoked, we re-send the cap flush
2510 * in client reconnect stage. This guarantees MDS * processes
2511 * the cap flush message before issuing the flushing caps to
2514 if ((cap->issued & ci->i_flushing_caps) !=
2515 ci->i_flushing_caps) {
2516 /* encode_caps_cb() also will reset these sequence
2517 * numbers. make sure sequence numbers in cap flush
2518 * message match later reconnect message */
2522 __kick_flushing_caps(mdsc, session, ci,
2525 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2528 spin_unlock(&ci->i_ceph_lock);
2532 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2533 struct ceph_mds_session *session)
2535 struct ceph_inode_info *ci;
2536 struct ceph_cap *cap;
2537 u64 oldest_flush_tid;
2539 lockdep_assert_held(&session->s_mutex);
2541 dout("kick_flushing_caps mds%d\n", session->s_mds);
2543 spin_lock(&mdsc->cap_dirty_lock);
2544 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2545 spin_unlock(&mdsc->cap_dirty_lock);
2547 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2548 spin_lock(&ci->i_ceph_lock);
2549 cap = ci->i_auth_cap;
2550 if (!(cap && cap->session == session)) {
2551 pr_err("%p auth cap %p not mds%d ???\n",
2552 &ci->vfs_inode, cap, session->s_mds);
2553 spin_unlock(&ci->i_ceph_lock);
2556 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2557 __kick_flushing_caps(mdsc, session, ci,
2560 spin_unlock(&ci->i_ceph_lock);
2564 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2565 struct ceph_inode_info *ci)
2567 struct ceph_mds_client *mdsc = session->s_mdsc;
2568 struct ceph_cap *cap = ci->i_auth_cap;
2570 lockdep_assert_held(&ci->i_ceph_lock);
2572 dout("%s %p flushing %s\n", __func__, &ci->vfs_inode,
2573 ceph_cap_string(ci->i_flushing_caps));
2575 if (!list_empty(&ci->i_cap_flush_list)) {
2576 u64 oldest_flush_tid;
2577 spin_lock(&mdsc->cap_dirty_lock);
2578 list_move_tail(&ci->i_flushing_item,
2579 &cap->session->s_cap_flushing);
2580 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2581 spin_unlock(&mdsc->cap_dirty_lock);
2583 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2589 * Take references to capabilities we hold, so that we don't release
2590 * them to the MDS prematurely.
2592 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2593 bool snap_rwsem_locked)
2595 lockdep_assert_held(&ci->i_ceph_lock);
2597 if (got & CEPH_CAP_PIN)
2599 if (got & CEPH_CAP_FILE_RD)
2601 if (got & CEPH_CAP_FILE_CACHE)
2602 ci->i_rdcache_ref++;
2603 if (got & CEPH_CAP_FILE_EXCL)
2605 if (got & CEPH_CAP_FILE_WR) {
2606 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2607 BUG_ON(!snap_rwsem_locked);
2608 ci->i_head_snapc = ceph_get_snap_context(
2609 ci->i_snap_realm->cached_context);
2613 if (got & CEPH_CAP_FILE_BUFFER) {
2614 if (ci->i_wb_ref == 0)
2615 ihold(&ci->vfs_inode);
2617 dout("%s %p wb %d -> %d (?)\n", __func__,
2618 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2623 * Try to grab cap references. Specify those refs we @want, and the
2624 * minimal set we @need. Also include the larger offset we are writing
2625 * to (when applicable), and check against max_size here as well.
2626 * Note that caller is responsible for ensuring max_size increases are
2627 * requested from the MDS.
2629 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2630 * or a negative error code. There are 3 speical error codes:
2631 * -EAGAIN: need to sleep but non-blocking is specified
2632 * -EFBIG: ask caller to call check_max_size() and try again.
2633 * -ESTALE: ask caller to call ceph_renew_caps() and try again.
2636 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2637 NON_BLOCKING = (1 << 8),
2638 CHECK_FILELOCK = (1 << 9),
2641 static int try_get_cap_refs(struct inode *inode, int need, int want,
2642 loff_t endoff, int flags, int *got)
2644 struct ceph_inode_info *ci = ceph_inode(inode);
2645 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2647 int have, implemented;
2648 bool snap_rwsem_locked = false;
2650 dout("get_cap_refs %p need %s want %s\n", inode,
2651 ceph_cap_string(need), ceph_cap_string(want));
2654 spin_lock(&ci->i_ceph_lock);
2656 if ((flags & CHECK_FILELOCK) &&
2657 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2658 dout("try_get_cap_refs %p error filelock\n", inode);
2663 /* finish pending truncate */
2664 while (ci->i_truncate_pending) {
2665 spin_unlock(&ci->i_ceph_lock);
2666 if (snap_rwsem_locked) {
2667 up_read(&mdsc->snap_rwsem);
2668 snap_rwsem_locked = false;
2670 __ceph_do_pending_vmtruncate(inode);
2671 spin_lock(&ci->i_ceph_lock);
2674 have = __ceph_caps_issued(ci, &implemented);
2676 if (have & need & CEPH_CAP_FILE_WR) {
2677 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2678 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2679 inode, endoff, ci->i_max_size);
2680 if (endoff > ci->i_requested_max_size)
2681 ret = ci->i_auth_cap ? -EFBIG : -ESTALE;
2685 * If a sync write is in progress, we must wait, so that we
2686 * can get a final snapshot value for size+mtime.
2688 if (__ceph_have_pending_cap_snap(ci)) {
2689 dout("get_cap_refs %p cap_snap_pending\n", inode);
2694 if ((have & need) == need) {
2696 * Look at (implemented & ~have & not) so that we keep waiting
2697 * on transition from wanted -> needed caps. This is needed
2698 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2699 * going before a prior buffered writeback happens.
2701 int not = want & ~(have & need);
2702 int revoking = implemented & ~have;
2703 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2704 inode, ceph_cap_string(have), ceph_cap_string(not),
2705 ceph_cap_string(revoking));
2706 if ((revoking & not) == 0) {
2707 if (!snap_rwsem_locked &&
2708 !ci->i_head_snapc &&
2709 (need & CEPH_CAP_FILE_WR)) {
2710 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2712 * we can not call down_read() when
2713 * task isn't in TASK_RUNNING state
2715 if (flags & NON_BLOCKING) {
2720 spin_unlock(&ci->i_ceph_lock);
2721 down_read(&mdsc->snap_rwsem);
2722 snap_rwsem_locked = true;
2725 snap_rwsem_locked = true;
2727 if ((have & want) == want)
2731 ceph_take_cap_refs(ci, *got, true);
2735 int session_readonly = false;
2737 if (ci->i_auth_cap &&
2738 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2739 struct ceph_mds_session *s = ci->i_auth_cap->session;
2740 spin_lock(&s->s_cap_lock);
2741 session_readonly = s->s_readonly;
2742 spin_unlock(&s->s_cap_lock);
2744 if (session_readonly) {
2745 dout("get_cap_refs %p need %s but mds%d readonly\n",
2746 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2751 if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
2752 dout("get_cap_refs %p forced umount\n", inode);
2756 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2757 if (need & ~mds_wanted) {
2758 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2759 inode, ceph_cap_string(need),
2760 ceph_cap_string(mds_wanted));
2765 dout("get_cap_refs %p have %s need %s\n", inode,
2766 ceph_cap_string(have), ceph_cap_string(need));
2770 __ceph_touch_fmode(ci, mdsc, flags);
2772 spin_unlock(&ci->i_ceph_lock);
2773 if (snap_rwsem_locked)
2774 up_read(&mdsc->snap_rwsem);
2777 ceph_update_cap_mis(&mdsc->metric);
2779 ceph_update_cap_hit(&mdsc->metric);
2781 dout("get_cap_refs %p ret %d got %s\n", inode,
2782 ret, ceph_cap_string(*got));
2787 * Check the offset we are writing up to against our current
2788 * max_size. If necessary, tell the MDS we want to write to
2791 static void check_max_size(struct inode *inode, loff_t endoff)
2793 struct ceph_inode_info *ci = ceph_inode(inode);
2796 /* do we need to explicitly request a larger max_size? */
2797 spin_lock(&ci->i_ceph_lock);
2798 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2799 dout("write %p at large endoff %llu, req max_size\n",
2801 ci->i_wanted_max_size = endoff;
2803 /* duplicate ceph_check_caps()'s logic */
2804 if (ci->i_auth_cap &&
2805 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2806 ci->i_wanted_max_size > ci->i_max_size &&
2807 ci->i_wanted_max_size > ci->i_requested_max_size)
2809 spin_unlock(&ci->i_ceph_lock);
2811 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2814 static inline int get_used_fmode(int caps)
2817 if (caps & CEPH_CAP_FILE_RD)
2818 fmode |= CEPH_FILE_MODE_RD;
2819 if (caps & CEPH_CAP_FILE_WR)
2820 fmode |= CEPH_FILE_MODE_WR;
2824 int ceph_try_get_caps(struct inode *inode, int need, int want,
2825 bool nonblock, int *got)
2829 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2830 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2831 CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2832 CEPH_CAP_ANY_DIR_OPS));
2834 ret = ceph_pool_perm_check(inode, need);
2839 flags = get_used_fmode(need | want);
2841 flags |= NON_BLOCKING;
2843 ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2844 /* three special error codes */
2845 if (ret == -EAGAIN || ret == -EFBIG || ret == -ESTALE)
2851 * Wait for caps, and take cap references. If we can't get a WR cap
2852 * due to a small max_size, make sure we check_max_size (and possibly
2853 * ask the mds) so we don't get hung up indefinitely.
2855 int ceph_get_caps(struct file *filp, int need, int want, loff_t endoff, int *got)
2857 struct ceph_file_info *fi = filp->private_data;
2858 struct inode *inode = file_inode(filp);
2859 struct ceph_inode_info *ci = ceph_inode(inode);
2860 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2861 int ret, _got, flags;
2863 ret = ceph_pool_perm_check(inode, need);
2867 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2868 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2871 flags = get_used_fmode(need | want);
2874 flags &= CEPH_FILE_MODE_MASK;
2875 if (atomic_read(&fi->num_locks))
2876 flags |= CHECK_FILELOCK;
2878 ret = try_get_cap_refs(inode, need, want, endoff,
2880 WARN_ON_ONCE(ret == -EAGAIN);
2882 struct ceph_mds_client *mdsc = fsc->mdsc;
2884 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2886 cw.ino = ceph_ino(inode);
2887 cw.tgid = current->tgid;
2891 spin_lock(&mdsc->caps_list_lock);
2892 list_add(&cw.list, &mdsc->cap_wait_list);
2893 spin_unlock(&mdsc->caps_list_lock);
2895 /* make sure used fmode not timeout */
2896 ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2897 add_wait_queue(&ci->i_cap_wq, &wait);
2899 flags |= NON_BLOCKING;
2900 while (!(ret = try_get_cap_refs(inode, need, want,
2901 endoff, flags, &_got))) {
2902 if (signal_pending(current)) {
2906 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2909 remove_wait_queue(&ci->i_cap_wq, &wait);
2910 ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
2912 spin_lock(&mdsc->caps_list_lock);
2914 spin_unlock(&mdsc->caps_list_lock);
2920 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2921 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2922 if (ret >= 0 && _got)
2923 ceph_put_cap_refs(ci, _got);
2928 if (ret == -EFBIG || ret == -ESTALE) {
2929 int ret2 = ceph_wait_on_async_create(inode);
2933 if (ret == -EFBIG) {
2934 check_max_size(inode, endoff);
2937 if (ret == -ESTALE) {
2938 /* session was killed, try renew caps */
2939 ret = ceph_renew_caps(inode, flags);
2946 if (S_ISREG(ci->vfs_inode.i_mode) &&
2947 ci->i_inline_version != CEPH_INLINE_NONE &&
2948 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2949 i_size_read(inode) > 0) {
2951 find_get_page(inode->i_mapping, 0);
2953 bool uptodate = PageUptodate(page);
2960 * drop cap refs first because getattr while
2961 * holding * caps refs can cause deadlock.
2963 ceph_put_cap_refs(ci, _got);
2967 * getattr request will bring inline data into
2970 ret = __ceph_do_getattr(inode, NULL,
2971 CEPH_STAT_CAP_INLINE_DATA,
2984 * Take cap refs. Caller must already know we hold at least one ref
2985 * on the caps in question or we don't know this is safe.
2987 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2989 spin_lock(&ci->i_ceph_lock);
2990 ceph_take_cap_refs(ci, caps, false);
2991 spin_unlock(&ci->i_ceph_lock);
2996 * drop cap_snap that is not associated with any snapshot.
2997 * we don't need to send FLUSHSNAP message for it.
2999 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
3000 struct ceph_cap_snap *capsnap)
3002 if (!capsnap->need_flush &&
3003 !capsnap->writing && !capsnap->dirty_pages) {
3004 dout("dropping cap_snap %p follows %llu\n",
3005 capsnap, capsnap->follows);
3006 BUG_ON(capsnap->cap_flush.tid > 0);
3007 ceph_put_snap_context(capsnap->context);
3008 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
3009 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3011 list_del(&capsnap->ci_item);
3012 ceph_put_cap_snap(capsnap);
3018 enum put_cap_refs_mode {
3019 PUT_CAP_REFS_SYNC = 0,
3020 PUT_CAP_REFS_NO_CHECK,
3027 * If we released the last ref on any given cap, call ceph_check_caps
3028 * to release (or schedule a release).
3030 * If we are releasing a WR cap (from a sync write), finalize any affected
3031 * cap_snap, and wake up any waiters.
3033 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
3034 enum put_cap_refs_mode mode)
3036 struct inode *inode = &ci->vfs_inode;
3037 int last = 0, put = 0, flushsnaps = 0, wake = 0;
3038 bool check_flushsnaps = false;
3040 spin_lock(&ci->i_ceph_lock);
3041 if (had & CEPH_CAP_PIN)
3043 if (had & CEPH_CAP_FILE_RD)
3044 if (--ci->i_rd_ref == 0)
3046 if (had & CEPH_CAP_FILE_CACHE)
3047 if (--ci->i_rdcache_ref == 0)
3049 if (had & CEPH_CAP_FILE_EXCL)
3050 if (--ci->i_fx_ref == 0)
3052 if (had & CEPH_CAP_FILE_BUFFER) {
3053 if (--ci->i_wb_ref == 0) {
3055 /* put the ref held by ceph_take_cap_refs() */
3057 check_flushsnaps = true;
3059 dout("put_cap_refs %p wb %d -> %d (?)\n",
3060 inode, ci->i_wb_ref+1, ci->i_wb_ref);
3062 if (had & CEPH_CAP_FILE_WR) {
3063 if (--ci->i_wr_ref == 0) {
3065 check_flushsnaps = true;
3066 if (ci->i_wrbuffer_ref_head == 0 &&
3067 ci->i_dirty_caps == 0 &&
3068 ci->i_flushing_caps == 0) {
3069 BUG_ON(!ci->i_head_snapc);
3070 ceph_put_snap_context(ci->i_head_snapc);
3071 ci->i_head_snapc = NULL;
3073 /* see comment in __ceph_remove_cap() */
3074 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3075 drop_inode_snap_realm(ci);
3078 if (check_flushsnaps && __ceph_have_pending_cap_snap(ci)) {
3079 struct ceph_cap_snap *capsnap =
3080 list_last_entry(&ci->i_cap_snaps,
3081 struct ceph_cap_snap,
3084 capsnap->writing = 0;
3085 if (ceph_try_drop_cap_snap(ci, capsnap))
3086 /* put the ref held by ceph_queue_cap_snap() */
3088 else if (__ceph_finish_cap_snap(ci, capsnap))
3092 spin_unlock(&ci->i_ceph_lock);
3094 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3095 last ? " last" : "", put ? " put" : "");
3098 case PUT_CAP_REFS_SYNC:
3100 ceph_check_caps(ci, 0, NULL);
3101 else if (flushsnaps)
3102 ceph_flush_snaps(ci, NULL);
3104 case PUT_CAP_REFS_ASYNC:
3106 ceph_queue_check_caps(inode);
3107 else if (flushsnaps)
3108 ceph_queue_flush_snaps(inode);
3114 wake_up_all(&ci->i_cap_wq);
3119 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3121 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_SYNC);
3124 void ceph_put_cap_refs_async(struct ceph_inode_info *ci, int had)
3126 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_ASYNC);
3129 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3131 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_NO_CHECK);
3135 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3136 * context. Adjust per-snap dirty page accounting as appropriate.
3137 * Once all dirty data for a cap_snap is flushed, flush snapped file
3138 * metadata back to the MDS. If we dropped the last ref, call
3141 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3142 struct ceph_snap_context *snapc)
3144 struct inode *inode = &ci->vfs_inode;
3145 struct ceph_cap_snap *capsnap = NULL;
3149 bool flush_snaps = false;
3150 bool complete_capsnap = false;
3152 spin_lock(&ci->i_ceph_lock);
3153 ci->i_wrbuffer_ref -= nr;
3154 if (ci->i_wrbuffer_ref == 0) {
3159 if (ci->i_head_snapc == snapc) {
3160 ci->i_wrbuffer_ref_head -= nr;
3161 if (ci->i_wrbuffer_ref_head == 0 &&
3162 ci->i_wr_ref == 0 &&
3163 ci->i_dirty_caps == 0 &&
3164 ci->i_flushing_caps == 0) {
3165 BUG_ON(!ci->i_head_snapc);
3166 ceph_put_snap_context(ci->i_head_snapc);
3167 ci->i_head_snapc = NULL;
3169 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3171 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3172 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3173 last ? " LAST" : "");
3175 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3176 if (capsnap->context == snapc) {
3182 capsnap->dirty_pages -= nr;
3183 if (capsnap->dirty_pages == 0) {
3184 complete_capsnap = true;
3185 if (!capsnap->writing) {
3186 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3189 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3194 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3195 " snap %lld %d/%d -> %d/%d %s%s\n",
3196 inode, capsnap, capsnap->context->seq,
3197 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3198 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3199 last ? " (wrbuffer last)" : "",
3200 complete_capsnap ? " (complete capsnap)" : "");
3203 spin_unlock(&ci->i_ceph_lock);
3206 ceph_check_caps(ci, 0, NULL);
3207 } else if (flush_snaps) {
3208 ceph_flush_snaps(ci, NULL);
3210 if (complete_capsnap)
3211 wake_up_all(&ci->i_cap_wq);
3213 /* avoid calling iput_final() in osd dispatch threads */
3214 ceph_async_iput(inode);
3219 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3221 static void invalidate_aliases(struct inode *inode)
3223 struct dentry *dn, *prev = NULL;
3225 dout("invalidate_aliases inode %p\n", inode);
3226 d_prune_aliases(inode);
3228 * For non-directory inode, d_find_alias() only returns
3229 * hashed dentry. After calling d_invalidate(), the
3230 * dentry becomes unhashed.
3232 * For directory inode, d_find_alias() can return
3233 * unhashed dentry. But directory inode should have
3234 * one alias at most.
3236 while ((dn = d_find_alias(inode))) {
3250 struct cap_extra_info {
3251 struct ceph_string *pool_ns;
3261 /* currently issued */
3263 struct timespec64 btime;
3267 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3268 * actually be a revocation if it specifies a smaller cap set.)
3270 * caller holds s_mutex and i_ceph_lock, we drop both.
3272 static void handle_cap_grant(struct inode *inode,
3273 struct ceph_mds_session *session,
3274 struct ceph_cap *cap,
3275 struct ceph_mds_caps *grant,
3276 struct ceph_buffer *xattr_buf,
3277 struct cap_extra_info *extra_info)
3278 __releases(ci->i_ceph_lock)
3279 __releases(session->s_mdsc->snap_rwsem)
3281 struct ceph_inode_info *ci = ceph_inode(inode);
3282 int seq = le32_to_cpu(grant->seq);
3283 int newcaps = le32_to_cpu(grant->caps);
3284 int used, wanted, dirty;
3285 u64 size = le64_to_cpu(grant->size);
3286 u64 max_size = le64_to_cpu(grant->max_size);
3287 unsigned char check_caps = 0;
3288 bool was_stale = cap->cap_gen < atomic_read(&session->s_cap_gen);
3290 bool writeback = false;
3291 bool queue_trunc = false;
3292 bool queue_invalidate = false;
3293 bool deleted_inode = false;
3294 bool fill_inline = false;
3296 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3297 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3298 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3299 i_size_read(inode));
3303 * If CACHE is being revoked, and we have no dirty buffers,
3304 * try to invalidate (once). (If there are dirty buffers, we
3305 * will invalidate _after_ writeback.)
3307 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3308 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3309 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3310 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3311 if (try_nonblocking_invalidate(inode)) {
3312 /* there were locked pages.. invalidate later
3313 in a separate thread. */
3314 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3315 queue_invalidate = true;
3316 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3322 cap->issued = cap->implemented = CEPH_CAP_PIN;
3325 * auth mds of the inode changed. we received the cap export message,
3326 * but still haven't received the cap import message. handle_cap_export
3327 * updated the new auth MDS' cap.
3329 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3330 * that was sent before the cap import message. So don't remove caps.
3332 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3333 WARN_ON(cap != ci->i_auth_cap);
3334 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3336 newcaps |= cap->issued;
3339 /* side effects now are allowed */
3340 cap->cap_gen = atomic_read(&session->s_cap_gen);
3343 __check_cap_issue(ci, cap, newcaps);
3345 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3347 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3348 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3349 umode_t mode = le32_to_cpu(grant->mode);
3351 if (inode_wrong_type(inode, mode))
3352 pr_warn_once("inode type changed! (ino %llx.%llx is 0%o, mds says 0%o)\n",
3353 ceph_vinop(inode), inode->i_mode, mode);
3355 inode->i_mode = mode;
3356 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3357 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3358 ci->i_btime = extra_info->btime;
3359 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3360 from_kuid(&init_user_ns, inode->i_uid),
3361 from_kgid(&init_user_ns, inode->i_gid));
3364 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3365 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3366 set_nlink(inode, le32_to_cpu(grant->nlink));
3367 if (inode->i_nlink == 0 &&
3368 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3369 deleted_inode = true;
3372 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3374 int len = le32_to_cpu(grant->xattr_len);
3375 u64 version = le64_to_cpu(grant->xattr_version);
3377 if (version > ci->i_xattrs.version) {
3378 dout(" got new xattrs v%llu on %p len %d\n",
3379 version, inode, len);
3380 if (ci->i_xattrs.blob)
3381 ceph_buffer_put(ci->i_xattrs.blob);
3382 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3383 ci->i_xattrs.version = version;
3384 ceph_forget_all_cached_acls(inode);
3385 ceph_security_invalidate_secctx(inode);
3389 if (newcaps & CEPH_CAP_ANY_RD) {
3390 struct timespec64 mtime, atime, ctime;
3391 /* ctime/mtime/atime? */
3392 ceph_decode_timespec64(&mtime, &grant->mtime);
3393 ceph_decode_timespec64(&atime, &grant->atime);
3394 ceph_decode_timespec64(&ctime, &grant->ctime);
3395 ceph_fill_file_time(inode, extra_info->issued,
3396 le32_to_cpu(grant->time_warp_seq),
3397 &ctime, &mtime, &atime);
3400 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3401 ci->i_files = extra_info->nfiles;
3402 ci->i_subdirs = extra_info->nsubdirs;
3405 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3406 /* file layout may have changed */
3407 s64 old_pool = ci->i_layout.pool_id;
3408 struct ceph_string *old_ns;
3410 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3411 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3412 lockdep_is_held(&ci->i_ceph_lock));
3413 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3415 if (ci->i_layout.pool_id != old_pool ||
3416 extra_info->pool_ns != old_ns)
3417 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3419 extra_info->pool_ns = old_ns;
3421 /* size/truncate_seq? */
3422 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3423 le32_to_cpu(grant->truncate_seq),
3424 le64_to_cpu(grant->truncate_size),
3428 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3429 if (max_size != ci->i_max_size) {
3430 dout("max_size %lld -> %llu\n",
3431 ci->i_max_size, max_size);
3432 ci->i_max_size = max_size;
3433 if (max_size >= ci->i_wanted_max_size) {
3434 ci->i_wanted_max_size = 0; /* reset */
3435 ci->i_requested_max_size = 0;
3441 /* check cap bits */
3442 wanted = __ceph_caps_wanted(ci);
3443 used = __ceph_caps_used(ci);
3444 dirty = __ceph_caps_dirty(ci);
3445 dout(" my wanted = %s, used = %s, dirty %s\n",
3446 ceph_cap_string(wanted),
3447 ceph_cap_string(used),
3448 ceph_cap_string(dirty));
3450 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3451 (wanted & ~(cap->mds_wanted | newcaps))) {
3453 * If mds is importing cap, prior cap messages that update
3454 * 'wanted' may get dropped by mds (migrate seq mismatch).
3456 * We don't send cap message to update 'wanted' if what we
3457 * want are already issued. If mds revokes caps, cap message
3458 * that releases caps also tells mds what we want. But if
3459 * caps got revoked by mds forcedly (session stale). We may
3460 * haven't told mds what we want.
3465 /* revocation, grant, or no-op? */
3466 if (cap->issued & ~newcaps) {
3467 int revoking = cap->issued & ~newcaps;
3469 dout("revocation: %s -> %s (revoking %s)\n",
3470 ceph_cap_string(cap->issued),
3471 ceph_cap_string(newcaps),
3472 ceph_cap_string(revoking));
3473 if (S_ISREG(inode->i_mode) &&
3474 (revoking & used & CEPH_CAP_FILE_BUFFER))
3475 writeback = true; /* initiate writeback; will delay ack */
3476 else if (queue_invalidate &&
3477 revoking == CEPH_CAP_FILE_CACHE &&
3478 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3479 ; /* do nothing yet, invalidation will be queued */
3480 else if (cap == ci->i_auth_cap)
3481 check_caps = 1; /* check auth cap only */
3483 check_caps = 2; /* check all caps */
3484 cap->issued = newcaps;
3485 cap->implemented |= newcaps;
3486 } else if (cap->issued == newcaps) {
3487 dout("caps unchanged: %s -> %s\n",
3488 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3490 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3491 ceph_cap_string(newcaps));
3492 /* non-auth MDS is revoking the newly grant caps ? */
3493 if (cap == ci->i_auth_cap &&
3494 __ceph_caps_revoking_other(ci, cap, newcaps))
3497 cap->issued = newcaps;
3498 cap->implemented |= newcaps; /* add bits only, to
3499 * avoid stepping on a
3500 * pending revocation */
3503 BUG_ON(cap->issued & ~cap->implemented);
3505 if (extra_info->inline_version > 0 &&
3506 extra_info->inline_version >= ci->i_inline_version) {
3507 ci->i_inline_version = extra_info->inline_version;
3508 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3509 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3513 if (ci->i_auth_cap == cap &&
3514 le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3515 if (newcaps & ~extra_info->issued)
3518 if (ci->i_requested_max_size > max_size ||
3519 !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3520 /* re-request max_size if necessary */
3521 ci->i_requested_max_size = 0;
3525 ceph_kick_flushing_inode_caps(session, ci);
3526 spin_unlock(&ci->i_ceph_lock);
3527 up_read(&session->s_mdsc->snap_rwsem);
3529 spin_unlock(&ci->i_ceph_lock);
3533 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3534 extra_info->inline_len);
3537 ceph_queue_vmtruncate(inode);
3541 * queue inode for writeback: we can't actually call
3542 * filemap_write_and_wait, etc. from message handler
3545 ceph_queue_writeback(inode);
3546 if (queue_invalidate)
3547 ceph_queue_invalidate(inode);
3549 invalidate_aliases(inode);
3551 wake_up_all(&ci->i_cap_wq);
3553 if (check_caps == 1)
3554 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3556 else if (check_caps == 2)
3557 ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3559 mutex_unlock(&session->s_mutex);
3563 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3564 * MDS has been safely committed.
3566 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3567 struct ceph_mds_caps *m,
3568 struct ceph_mds_session *session,
3569 struct ceph_cap *cap)
3570 __releases(ci->i_ceph_lock)
3572 struct ceph_inode_info *ci = ceph_inode(inode);
3573 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3574 struct ceph_cap_flush *cf, *tmp_cf;
3575 LIST_HEAD(to_remove);
3576 unsigned seq = le32_to_cpu(m->seq);
3577 int dirty = le32_to_cpu(m->dirty);
3580 bool wake_ci = false;
3581 bool wake_mdsc = false;
3583 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3584 /* Is this the one that was flushed? */
3585 if (cf->tid == flush_tid)
3588 /* Is this a capsnap? */
3592 if (cf->tid <= flush_tid) {
3594 * An earlier or current tid. The FLUSH_ACK should
3595 * represent a superset of this flush's caps.
3597 wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3598 list_add_tail(&cf->i_list, &to_remove);
3601 * This is a later one. Any caps in it are still dirty
3602 * so don't count them as cleaned.
3604 cleaned &= ~cf->caps;
3610 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3611 " flushing %s -> %s\n",
3612 inode, session->s_mds, seq, ceph_cap_string(dirty),
3613 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3614 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3616 if (list_empty(&to_remove) && !cleaned)
3619 ci->i_flushing_caps &= ~cleaned;
3621 spin_lock(&mdsc->cap_dirty_lock);
3623 list_for_each_entry(cf, &to_remove, i_list)
3624 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3626 if (ci->i_flushing_caps == 0) {
3627 if (list_empty(&ci->i_cap_flush_list)) {
3628 list_del_init(&ci->i_flushing_item);
3629 if (!list_empty(&session->s_cap_flushing)) {
3630 dout(" mds%d still flushing cap on %p\n",
3632 &list_first_entry(&session->s_cap_flushing,
3633 struct ceph_inode_info,
3634 i_flushing_item)->vfs_inode);
3637 mdsc->num_cap_flushing--;
3638 dout(" inode %p now !flushing\n", inode);
3640 if (ci->i_dirty_caps == 0) {
3641 dout(" inode %p now clean\n", inode);
3642 BUG_ON(!list_empty(&ci->i_dirty_item));
3644 if (ci->i_wr_ref == 0 &&
3645 ci->i_wrbuffer_ref_head == 0) {
3646 BUG_ON(!ci->i_head_snapc);
3647 ceph_put_snap_context(ci->i_head_snapc);
3648 ci->i_head_snapc = NULL;
3651 BUG_ON(list_empty(&ci->i_dirty_item));
3654 spin_unlock(&mdsc->cap_dirty_lock);
3657 spin_unlock(&ci->i_ceph_lock);
3659 while (!list_empty(&to_remove)) {
3660 cf = list_first_entry(&to_remove,
3661 struct ceph_cap_flush, i_list);
3662 list_del(&cf->i_list);
3663 ceph_free_cap_flush(cf);
3667 wake_up_all(&ci->i_cap_wq);
3669 wake_up_all(&mdsc->cap_flushing_wq);
3675 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3676 * throw away our cap_snap.
3678 * Caller hold s_mutex.
3680 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3681 struct ceph_mds_caps *m,
3682 struct ceph_mds_session *session)
3684 struct ceph_inode_info *ci = ceph_inode(inode);
3685 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3686 u64 follows = le64_to_cpu(m->snap_follows);
3687 struct ceph_cap_snap *capsnap;
3688 bool flushed = false;
3689 bool wake_ci = false;
3690 bool wake_mdsc = false;
3692 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3693 inode, ci, session->s_mds, follows);
3695 spin_lock(&ci->i_ceph_lock);
3696 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3697 if (capsnap->follows == follows) {
3698 if (capsnap->cap_flush.tid != flush_tid) {
3699 dout(" cap_snap %p follows %lld tid %lld !="
3700 " %lld\n", capsnap, follows,
3701 flush_tid, capsnap->cap_flush.tid);
3707 dout(" skipping cap_snap %p follows %lld\n",
3708 capsnap, capsnap->follows);
3712 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3713 dout(" removing %p cap_snap %p follows %lld\n",
3714 inode, capsnap, follows);
3715 list_del(&capsnap->ci_item);
3716 wake_ci |= __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3718 spin_lock(&mdsc->cap_dirty_lock);
3720 if (list_empty(&ci->i_cap_flush_list))
3721 list_del_init(&ci->i_flushing_item);
3723 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc,
3724 &capsnap->cap_flush);
3725 spin_unlock(&mdsc->cap_dirty_lock);
3727 spin_unlock(&ci->i_ceph_lock);
3729 ceph_put_snap_context(capsnap->context);
3730 ceph_put_cap_snap(capsnap);
3732 wake_up_all(&ci->i_cap_wq);
3734 wake_up_all(&mdsc->cap_flushing_wq);
3740 * Handle TRUNC from MDS, indicating file truncation.
3742 * caller hold s_mutex.
3744 static bool handle_cap_trunc(struct inode *inode,
3745 struct ceph_mds_caps *trunc,
3746 struct ceph_mds_session *session)
3748 struct ceph_inode_info *ci = ceph_inode(inode);
3749 int mds = session->s_mds;
3750 int seq = le32_to_cpu(trunc->seq);
3751 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3752 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3753 u64 size = le64_to_cpu(trunc->size);
3754 int implemented = 0;
3755 int dirty = __ceph_caps_dirty(ci);
3756 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3757 bool queue_trunc = false;
3759 lockdep_assert_held(&ci->i_ceph_lock);
3761 issued |= implemented | dirty;
3763 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3764 inode, mds, seq, truncate_size, truncate_seq);
3765 queue_trunc = ceph_fill_file_size(inode, issued,
3766 truncate_seq, truncate_size, size);
3771 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3772 * different one. If we are the most recent migration we've seen (as
3773 * indicated by mseq), make note of the migrating cap bits for the
3774 * duration (until we see the corresponding IMPORT).
3776 * caller holds s_mutex
3778 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3779 struct ceph_mds_cap_peer *ph,
3780 struct ceph_mds_session *session)
3782 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3783 struct ceph_mds_session *tsession = NULL;
3784 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3785 struct ceph_inode_info *ci = ceph_inode(inode);
3787 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3788 unsigned t_seq, t_mseq;
3790 int mds = session->s_mds;
3793 t_cap_id = le64_to_cpu(ph->cap_id);
3794 t_seq = le32_to_cpu(ph->seq);
3795 t_mseq = le32_to_cpu(ph->mseq);
3796 target = le32_to_cpu(ph->mds);
3798 t_cap_id = t_seq = t_mseq = 0;
3802 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3803 inode, ci, mds, mseq, target);
3805 spin_lock(&ci->i_ceph_lock);
3806 cap = __get_cap_for_mds(ci, mds);
3807 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3811 __ceph_remove_cap(cap, false);
3816 * now we know we haven't received the cap import message yet
3817 * because the exported cap still exist.
3820 issued = cap->issued;
3821 if (issued != cap->implemented)
3822 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3823 "ino (%llx.%llx) mds%d seq %d mseq %d "
3824 "issued %s implemented %s\n",
3825 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3826 ceph_cap_string(issued),
3827 ceph_cap_string(cap->implemented));
3830 tcap = __get_cap_for_mds(ci, target);
3832 /* already have caps from the target */
3833 if (tcap->cap_id == t_cap_id &&
3834 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3835 dout(" updating import cap %p mds%d\n", tcap, target);
3836 tcap->cap_id = t_cap_id;
3837 tcap->seq = t_seq - 1;
3838 tcap->issue_seq = t_seq - 1;
3839 tcap->issued |= issued;
3840 tcap->implemented |= issued;
3841 if (cap == ci->i_auth_cap) {
3842 ci->i_auth_cap = tcap;
3843 change_auth_cap_ses(ci, tcap->session);
3846 __ceph_remove_cap(cap, false);
3848 } else if (tsession) {
3849 /* add placeholder for the export tagert */
3850 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3852 ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3853 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3855 if (!list_empty(&ci->i_cap_flush_list) &&
3856 ci->i_auth_cap == tcap) {
3857 spin_lock(&mdsc->cap_dirty_lock);
3858 list_move_tail(&ci->i_flushing_item,
3859 &tcap->session->s_cap_flushing);
3860 spin_unlock(&mdsc->cap_dirty_lock);
3863 __ceph_remove_cap(cap, false);
3867 spin_unlock(&ci->i_ceph_lock);
3868 mutex_unlock(&session->s_mutex);
3870 /* open target session */
3871 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3872 if (!IS_ERR(tsession)) {
3874 mutex_lock(&session->s_mutex);
3875 mutex_lock_nested(&tsession->s_mutex,
3876 SINGLE_DEPTH_NESTING);
3878 mutex_lock(&tsession->s_mutex);
3879 mutex_lock_nested(&session->s_mutex,
3880 SINGLE_DEPTH_NESTING);
3882 new_cap = ceph_get_cap(mdsc, NULL);
3887 mutex_lock(&session->s_mutex);
3892 spin_unlock(&ci->i_ceph_lock);
3893 mutex_unlock(&session->s_mutex);
3895 mutex_unlock(&tsession->s_mutex);
3896 ceph_put_mds_session(tsession);
3899 ceph_put_cap(mdsc, new_cap);
3903 * Handle cap IMPORT.
3905 * caller holds s_mutex. acquires i_ceph_lock
3907 static void handle_cap_import(struct ceph_mds_client *mdsc,
3908 struct inode *inode, struct ceph_mds_caps *im,
3909 struct ceph_mds_cap_peer *ph,
3910 struct ceph_mds_session *session,
3911 struct ceph_cap **target_cap, int *old_issued)
3913 struct ceph_inode_info *ci = ceph_inode(inode);
3914 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3915 int mds = session->s_mds;
3917 unsigned caps = le32_to_cpu(im->caps);
3918 unsigned wanted = le32_to_cpu(im->wanted);
3919 unsigned seq = le32_to_cpu(im->seq);
3920 unsigned mseq = le32_to_cpu(im->migrate_seq);
3921 u64 realmino = le64_to_cpu(im->realm);
3922 u64 cap_id = le64_to_cpu(im->cap_id);
3927 p_cap_id = le64_to_cpu(ph->cap_id);
3928 peer = le32_to_cpu(ph->mds);
3934 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3935 inode, ci, mds, mseq, peer);
3937 cap = __get_cap_for_mds(ci, mds);
3940 spin_unlock(&ci->i_ceph_lock);
3941 new_cap = ceph_get_cap(mdsc, NULL);
3942 spin_lock(&ci->i_ceph_lock);
3948 ceph_put_cap(mdsc, new_cap);
3953 __ceph_caps_issued(ci, &issued);
3954 issued |= __ceph_caps_dirty(ci);
3956 ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
3957 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3959 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3960 if (ocap && ocap->cap_id == p_cap_id) {
3961 dout(" remove export cap %p mds%d flags %d\n",
3962 ocap, peer, ph->flags);
3963 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3964 (ocap->seq != le32_to_cpu(ph->seq) ||
3965 ocap->mseq != le32_to_cpu(ph->mseq))) {
3966 pr_err_ratelimited("handle_cap_import: "
3967 "mismatched seq/mseq: ino (%llx.%llx) "
3968 "mds%d seq %d mseq %d importer mds%d "
3969 "has peer seq %d mseq %d\n",
3970 ceph_vinop(inode), peer, ocap->seq,
3971 ocap->mseq, mds, le32_to_cpu(ph->seq),
3972 le32_to_cpu(ph->mseq));
3974 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3977 *old_issued = issued;
3982 * Handle a caps message from the MDS.
3984 * Identify the appropriate session, inode, and call the right handler
3985 * based on the cap op.
3987 void ceph_handle_caps(struct ceph_mds_session *session,
3988 struct ceph_msg *msg)
3990 struct ceph_mds_client *mdsc = session->s_mdsc;
3991 struct inode *inode;
3992 struct ceph_inode_info *ci;
3993 struct ceph_cap *cap;
3994 struct ceph_mds_caps *h;
3995 struct ceph_mds_cap_peer *peer = NULL;
3996 struct ceph_snap_realm *realm = NULL;
3998 int msg_version = le16_to_cpu(msg->hdr.version);
4000 struct ceph_vino vino;
4002 size_t snaptrace_len;
4004 struct cap_extra_info extra_info = {};
4007 dout("handle_caps from mds%d\n", session->s_mds);
4010 end = msg->front.iov_base + msg->front.iov_len;
4011 if (msg->front.iov_len < sizeof(*h))
4013 h = msg->front.iov_base;
4014 op = le32_to_cpu(h->op);
4015 vino.ino = le64_to_cpu(h->ino);
4016 vino.snap = CEPH_NOSNAP;
4017 seq = le32_to_cpu(h->seq);
4018 mseq = le32_to_cpu(h->migrate_seq);
4021 snaptrace_len = le32_to_cpu(h->snap_trace_len);
4022 p = snaptrace + snaptrace_len;
4024 if (msg_version >= 2) {
4026 ceph_decode_32_safe(&p, end, flock_len, bad);
4027 if (p + flock_len > end)
4032 if (msg_version >= 3) {
4033 if (op == CEPH_CAP_OP_IMPORT) {
4034 if (p + sizeof(*peer) > end)
4038 } else if (op == CEPH_CAP_OP_EXPORT) {
4039 /* recorded in unused fields */
4040 peer = (void *)&h->size;
4044 if (msg_version >= 4) {
4045 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
4046 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
4047 if (p + extra_info.inline_len > end)
4049 extra_info.inline_data = p;
4050 p += extra_info.inline_len;
4053 if (msg_version >= 5) {
4054 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
4057 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
4058 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
4061 if (msg_version >= 8) {
4065 ceph_decode_skip_64(&p, end, bad); // flush_tid
4067 ceph_decode_skip_32(&p, end, bad); // caller_uid
4068 ceph_decode_skip_32(&p, end, bad); // caller_gid
4070 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4071 if (pool_ns_len > 0) {
4072 ceph_decode_need(&p, end, pool_ns_len, bad);
4073 extra_info.pool_ns =
4074 ceph_find_or_create_string(p, pool_ns_len);
4079 if (msg_version >= 9) {
4080 struct ceph_timespec *btime;
4082 if (p + sizeof(*btime) > end)
4085 ceph_decode_timespec64(&extra_info.btime, btime);
4086 p += sizeof(*btime);
4087 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4090 if (msg_version >= 11) {
4092 ceph_decode_skip_32(&p, end, bad); // flags
4094 extra_info.dirstat_valid = true;
4095 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4096 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4100 inode = ceph_find_inode(mdsc->fsc->sb, vino);
4101 ci = ceph_inode(inode);
4102 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4105 mutex_lock(&session->s_mutex);
4106 inc_session_sequence(session);
4107 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4111 dout(" i don't have ino %llx\n", vino.ino);
4113 if (op == CEPH_CAP_OP_IMPORT) {
4114 cap = ceph_get_cap(mdsc, NULL);
4115 cap->cap_ino = vino.ino;
4116 cap->queue_release = 1;
4117 cap->cap_id = le64_to_cpu(h->cap_id);
4120 cap->issue_seq = seq;
4121 spin_lock(&session->s_cap_lock);
4122 __ceph_queue_cap_release(session, cap);
4123 spin_unlock(&session->s_cap_lock);
4125 goto flush_cap_releases;
4128 /* these will work even if we don't have a cap yet */
4130 case CEPH_CAP_OP_FLUSHSNAP_ACK:
4131 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4135 case CEPH_CAP_OP_EXPORT:
4136 handle_cap_export(inode, h, peer, session);
4139 case CEPH_CAP_OP_IMPORT:
4141 if (snaptrace_len) {
4142 down_write(&mdsc->snap_rwsem);
4143 ceph_update_snap_trace(mdsc, snaptrace,
4144 snaptrace + snaptrace_len,
4146 downgrade_write(&mdsc->snap_rwsem);
4148 down_read(&mdsc->snap_rwsem);
4150 spin_lock(&ci->i_ceph_lock);
4151 handle_cap_import(mdsc, inode, h, peer, session,
4152 &cap, &extra_info.issued);
4153 handle_cap_grant(inode, session, cap,
4154 h, msg->middle, &extra_info);
4156 ceph_put_snap_realm(mdsc, realm);
4160 /* the rest require a cap */
4161 spin_lock(&ci->i_ceph_lock);
4162 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4164 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4165 inode, ceph_ino(inode), ceph_snap(inode),
4167 spin_unlock(&ci->i_ceph_lock);
4168 goto flush_cap_releases;
4171 /* note that each of these drops i_ceph_lock for us */
4173 case CEPH_CAP_OP_REVOKE:
4174 case CEPH_CAP_OP_GRANT:
4175 __ceph_caps_issued(ci, &extra_info.issued);
4176 extra_info.issued |= __ceph_caps_dirty(ci);
4177 handle_cap_grant(inode, session, cap,
4178 h, msg->middle, &extra_info);
4181 case CEPH_CAP_OP_FLUSH_ACK:
4182 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4186 case CEPH_CAP_OP_TRUNC:
4187 queue_trunc = handle_cap_trunc(inode, h, session);
4188 spin_unlock(&ci->i_ceph_lock);
4190 ceph_queue_vmtruncate(inode);
4194 spin_unlock(&ci->i_ceph_lock);
4195 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4196 ceph_cap_op_name(op));
4200 mutex_unlock(&session->s_mutex);
4202 ceph_put_string(extra_info.pool_ns);
4203 /* avoid calling iput_final() in mds dispatch threads */
4204 ceph_async_iput(inode);
4209 * send any cap release message to try to move things
4210 * along for the mds (who clearly thinks we still have this
4213 ceph_flush_cap_releases(mdsc, session);
4217 pr_err("ceph_handle_caps: corrupt message\n");
4223 * Delayed work handler to process end of delayed cap release LRU list.
4225 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4227 struct inode *inode;
4228 struct ceph_inode_info *ci;
4230 dout("check_delayed_caps\n");
4231 spin_lock(&mdsc->cap_delay_lock);
4232 while (!list_empty(&mdsc->cap_delay_list)) {
4233 ci = list_first_entry(&mdsc->cap_delay_list,
4234 struct ceph_inode_info,
4236 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4237 time_before(jiffies, ci->i_hold_caps_max))
4239 list_del_init(&ci->i_cap_delay_list);
4241 inode = igrab(&ci->vfs_inode);
4243 spin_unlock(&mdsc->cap_delay_lock);
4244 dout("check_delayed_caps on %p\n", inode);
4245 ceph_check_caps(ci, 0, NULL);
4246 /* avoid calling iput_final() in tick thread */
4247 ceph_async_iput(inode);
4248 spin_lock(&mdsc->cap_delay_lock);
4251 spin_unlock(&mdsc->cap_delay_lock);
4255 * Flush all dirty caps to the mds
4257 static void flush_dirty_session_caps(struct ceph_mds_session *s)
4259 struct ceph_mds_client *mdsc = s->s_mdsc;
4260 struct ceph_inode_info *ci;
4261 struct inode *inode;
4263 dout("flush_dirty_caps\n");
4264 spin_lock(&mdsc->cap_dirty_lock);
4265 while (!list_empty(&s->s_cap_dirty)) {
4266 ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4268 inode = &ci->vfs_inode;
4270 dout("flush_dirty_caps %p\n", inode);
4271 spin_unlock(&mdsc->cap_dirty_lock);
4272 ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4274 spin_lock(&mdsc->cap_dirty_lock);
4276 spin_unlock(&mdsc->cap_dirty_lock);
4277 dout("flush_dirty_caps done\n");
4280 static void iterate_sessions(struct ceph_mds_client *mdsc,
4281 void (*cb)(struct ceph_mds_session *))
4285 mutex_lock(&mdsc->mutex);
4286 for (mds = 0; mds < mdsc->max_sessions; ++mds) {
4287 struct ceph_mds_session *s;
4289 if (!mdsc->sessions[mds])
4292 s = ceph_get_mds_session(mdsc->sessions[mds]);
4296 mutex_unlock(&mdsc->mutex);
4298 ceph_put_mds_session(s);
4299 mutex_lock(&mdsc->mutex);
4301 mutex_unlock(&mdsc->mutex);
4304 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4306 iterate_sessions(mdsc, flush_dirty_session_caps);
4309 void __ceph_touch_fmode(struct ceph_inode_info *ci,
4310 struct ceph_mds_client *mdsc, int fmode)
4312 unsigned long now = jiffies;
4313 if (fmode & CEPH_FILE_MODE_RD)
4314 ci->i_last_rd = now;
4315 if (fmode & CEPH_FILE_MODE_WR)
4316 ci->i_last_wr = now;
4317 /* queue periodic check */
4319 __ceph_is_any_real_caps(ci) &&
4320 list_empty(&ci->i_cap_delay_list))
4321 __cap_delay_requeue(mdsc, ci);
4324 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4326 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4327 int bits = (fmode << 1) | 1;
4328 bool is_opened = false;
4332 atomic64_inc(&mdsc->metric.opened_files);
4334 spin_lock(&ci->i_ceph_lock);
4335 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4336 if (bits & (1 << i))
4337 ci->i_nr_by_mode[i] += count;
4340 * If any of the mode ref is larger than 1,
4341 * that means it has been already opened by
4342 * others. Just skip checking the PIN ref.
4344 if (i && ci->i_nr_by_mode[i] > 1)
4349 percpu_counter_inc(&mdsc->metric.opened_inodes);
4350 spin_unlock(&ci->i_ceph_lock);
4354 * Drop open file reference. If we were the last open file,
4355 * we may need to release capabilities to the MDS (or schedule
4356 * their delayed release).
4358 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4360 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4361 int bits = (fmode << 1) | 1;
4362 bool is_closed = true;
4366 atomic64_dec(&mdsc->metric.opened_files);
4368 spin_lock(&ci->i_ceph_lock);
4369 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4370 if (bits & (1 << i)) {
4371 BUG_ON(ci->i_nr_by_mode[i] < count);
4372 ci->i_nr_by_mode[i] -= count;
4376 * If any of the mode ref is not 0 after
4377 * decreased, that means it is still opened
4378 * by others. Just skip checking the PIN ref.
4380 if (i && ci->i_nr_by_mode[i])
4385 percpu_counter_dec(&mdsc->metric.opened_inodes);
4386 spin_unlock(&ci->i_ceph_lock);
4390 * For a soon-to-be unlinked file, drop the LINK caps. If it
4391 * looks like the link count will hit 0, drop any other caps (other
4392 * than PIN) we don't specifically want (due to the file still being
4395 int ceph_drop_caps_for_unlink(struct inode *inode)
4397 struct ceph_inode_info *ci = ceph_inode(inode);
4398 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4400 spin_lock(&ci->i_ceph_lock);
4401 if (inode->i_nlink == 1) {
4402 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4404 if (__ceph_caps_dirty(ci)) {
4405 struct ceph_mds_client *mdsc =
4406 ceph_inode_to_client(inode)->mdsc;
4407 __cap_delay_requeue_front(mdsc, ci);
4410 spin_unlock(&ci->i_ceph_lock);
4415 * Helpers for embedding cap and dentry lease releases into mds
4418 * @force is used by dentry_release (below) to force inclusion of a
4419 * record for the directory inode, even when there aren't any caps to
4422 int ceph_encode_inode_release(void **p, struct inode *inode,
4423 int mds, int drop, int unless, int force)
4425 struct ceph_inode_info *ci = ceph_inode(inode);
4426 struct ceph_cap *cap;
4427 struct ceph_mds_request_release *rel = *p;
4431 spin_lock(&ci->i_ceph_lock);
4432 used = __ceph_caps_used(ci);
4433 dirty = __ceph_caps_dirty(ci);
4435 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4436 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4437 ceph_cap_string(unless));
4439 /* only drop unused, clean caps */
4440 drop &= ~(used | dirty);
4442 cap = __get_cap_for_mds(ci, mds);
4443 if (cap && __cap_is_valid(cap)) {
4444 unless &= cap->issued;
4446 if (unless & CEPH_CAP_AUTH_EXCL)
4447 drop &= ~CEPH_CAP_AUTH_SHARED;
4448 if (unless & CEPH_CAP_LINK_EXCL)
4449 drop &= ~CEPH_CAP_LINK_SHARED;
4450 if (unless & CEPH_CAP_XATTR_EXCL)
4451 drop &= ~CEPH_CAP_XATTR_SHARED;
4452 if (unless & CEPH_CAP_FILE_EXCL)
4453 drop &= ~CEPH_CAP_FILE_SHARED;
4456 if (force || (cap->issued & drop)) {
4457 if (cap->issued & drop) {
4458 int wanted = __ceph_caps_wanted(ci);
4459 dout("encode_inode_release %p cap %p "
4460 "%s -> %s, wanted %s -> %s\n", inode, cap,
4461 ceph_cap_string(cap->issued),
4462 ceph_cap_string(cap->issued & ~drop),
4463 ceph_cap_string(cap->mds_wanted),
4464 ceph_cap_string(wanted));
4466 cap->issued &= ~drop;
4467 cap->implemented &= ~drop;
4468 cap->mds_wanted = wanted;
4469 if (cap == ci->i_auth_cap &&
4470 !(wanted & CEPH_CAP_ANY_FILE_WR))
4471 ci->i_requested_max_size = 0;
4473 dout("encode_inode_release %p cap %p %s"
4474 " (force)\n", inode, cap,
4475 ceph_cap_string(cap->issued));
4478 rel->ino = cpu_to_le64(ceph_ino(inode));
4479 rel->cap_id = cpu_to_le64(cap->cap_id);
4480 rel->seq = cpu_to_le32(cap->seq);
4481 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4482 rel->mseq = cpu_to_le32(cap->mseq);
4483 rel->caps = cpu_to_le32(cap->implemented);
4484 rel->wanted = cpu_to_le32(cap->mds_wanted);
4490 dout("encode_inode_release %p cap %p %s (noop)\n",
4491 inode, cap, ceph_cap_string(cap->issued));
4494 spin_unlock(&ci->i_ceph_lock);
4498 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4500 int mds, int drop, int unless)
4502 struct dentry *parent = NULL;
4503 struct ceph_mds_request_release *rel = *p;
4504 struct ceph_dentry_info *di = ceph_dentry(dentry);
4509 * force an record for the directory caps if we have a dentry lease.
4510 * this is racy (can't take i_ceph_lock and d_lock together), but it
4511 * doesn't have to be perfect; the mds will revoke anything we don't
4514 spin_lock(&dentry->d_lock);
4515 if (di->lease_session && di->lease_session->s_mds == mds)
4518 parent = dget(dentry->d_parent);
4519 dir = d_inode(parent);
4521 spin_unlock(&dentry->d_lock);
4523 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4526 spin_lock(&dentry->d_lock);
4527 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4528 dout("encode_dentry_release %p mds%d seq %d\n",
4529 dentry, mds, (int)di->lease_seq);
4530 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4531 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4532 *p += dentry->d_name.len;
4533 rel->dname_seq = cpu_to_le32(di->lease_seq);
4534 __ceph_mdsc_drop_dentry_lease(dentry);
4536 spin_unlock(&dentry->d_lock);