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->netfs.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->netfs.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->netfs.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->netfs.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->netfs.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->netfs.inode.i_mode)) {
587 dout(" marking %p NOT complete\n", &ci->netfs.inode);
588 __ceph_dir_clear_complete(ci);
592 /* Wipe saved layout if we're losing DIR_CREATE caps */
593 if (S_ISDIR(ci->netfs.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 void change_auth_cap_ses(struct ceph_inode_info *ci,
606 struct ceph_mds_session *session)
608 lockdep_assert_held(&ci->i_ceph_lock);
610 if (list_empty(&ci->i_dirty_item) && list_empty(&ci->i_flushing_item))
613 spin_lock(&session->s_mdsc->cap_dirty_lock);
614 if (!list_empty(&ci->i_dirty_item))
615 list_move(&ci->i_dirty_item, &session->s_cap_dirty);
616 if (!list_empty(&ci->i_flushing_item))
617 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
618 spin_unlock(&session->s_mdsc->cap_dirty_lock);
622 * Add a capability under the given MDS session.
624 * Caller should hold session snap_rwsem (read) and ci->i_ceph_lock
626 * @fmode is the open file mode, if we are opening a file, otherwise
627 * it is < 0. (This is so we can atomically add the cap and add an
628 * open file reference to it.)
630 void ceph_add_cap(struct inode *inode,
631 struct ceph_mds_session *session, u64 cap_id,
632 unsigned issued, unsigned wanted,
633 unsigned seq, unsigned mseq, u64 realmino, int flags,
634 struct ceph_cap **new_cap)
636 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
637 struct ceph_inode_info *ci = ceph_inode(inode);
638 struct ceph_cap *cap;
639 int mds = session->s_mds;
643 lockdep_assert_held(&ci->i_ceph_lock);
645 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
646 session->s_mds, cap_id, ceph_cap_string(issued), seq);
648 gen = atomic_read(&session->s_cap_gen);
650 cap = __get_cap_for_mds(ci, mds);
656 cap->implemented = 0;
662 __insert_cap_node(ci, cap);
664 /* add to session cap list */
665 cap->session = session;
666 spin_lock(&session->s_cap_lock);
667 list_add_tail(&cap->session_caps, &session->s_caps);
668 session->s_nr_caps++;
669 atomic64_inc(&mdsc->metric.total_caps);
670 spin_unlock(&session->s_cap_lock);
672 spin_lock(&session->s_cap_lock);
673 list_move_tail(&cap->session_caps, &session->s_caps);
674 spin_unlock(&session->s_cap_lock);
676 if (cap->cap_gen < gen)
677 cap->issued = cap->implemented = CEPH_CAP_PIN;
680 * auth mds of the inode changed. we received the cap export
681 * message, but still haven't received the cap import message.
682 * handle_cap_export() updated the new auth MDS' cap.
684 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
685 * a message that was send before the cap import message. So
688 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
689 WARN_ON(cap != ci->i_auth_cap);
690 WARN_ON(cap->cap_id != cap_id);
693 issued |= cap->issued;
694 flags |= CEPH_CAP_FLAG_AUTH;
698 if (!ci->i_snap_realm ||
699 ((flags & CEPH_CAP_FLAG_AUTH) &&
700 realmino != (u64)-1 && ci->i_snap_realm->ino != realmino)) {
702 * add this inode to the appropriate snap realm
704 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
707 ceph_change_snap_realm(inode, realm);
709 WARN(1, "%s: couldn't find snap realm 0x%llx (ino 0x%llx oldrealm 0x%llx)\n",
710 __func__, realmino, ci->i_vino.ino,
711 ci->i_snap_realm ? ci->i_snap_realm->ino : 0);
714 __check_cap_issue(ci, cap, issued);
717 * If we are issued caps we don't want, or the mds' wanted
718 * value appears to be off, queue a check so we'll release
719 * later and/or update the mds wanted value.
721 actual_wanted = __ceph_caps_wanted(ci);
722 if ((wanted & ~actual_wanted) ||
723 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
724 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
725 ceph_cap_string(issued), ceph_cap_string(wanted),
726 ceph_cap_string(actual_wanted));
727 __cap_delay_requeue(mdsc, ci);
730 if (flags & CEPH_CAP_FLAG_AUTH) {
731 if (!ci->i_auth_cap ||
732 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
733 if (ci->i_auth_cap &&
734 ci->i_auth_cap->session != cap->session)
735 change_auth_cap_ses(ci, cap->session);
736 ci->i_auth_cap = cap;
737 cap->mds_wanted = wanted;
740 WARN_ON(ci->i_auth_cap == cap);
743 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
744 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
745 ceph_cap_string(issued|cap->issued), seq, mds);
746 cap->cap_id = cap_id;
747 cap->issued = issued;
748 cap->implemented |= issued;
749 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
750 cap->mds_wanted = wanted;
752 cap->mds_wanted |= wanted;
754 cap->issue_seq = seq;
757 wake_up_all(&ci->i_cap_wq);
761 * Return true if cap has not timed out and belongs to the current
762 * generation of the MDS session (i.e. has not gone 'stale' due to
763 * us losing touch with the mds).
765 static int __cap_is_valid(struct ceph_cap *cap)
770 gen = atomic_read(&cap->session->s_cap_gen);
771 ttl = cap->session->s_cap_ttl;
773 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
774 dout("__cap_is_valid %p cap %p issued %s "
775 "but STALE (gen %u vs %u)\n", &cap->ci->netfs.inode,
776 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
784 * Return set of valid cap bits issued to us. Note that caps time
785 * out, and may be invalidated in bulk if the client session times out
786 * and session->s_cap_gen is bumped.
788 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
790 int have = ci->i_snap_caps;
791 struct ceph_cap *cap;
796 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
797 cap = rb_entry(p, struct ceph_cap, ci_node);
798 if (!__cap_is_valid(cap))
800 dout("__ceph_caps_issued %p cap %p issued %s\n",
801 &ci->netfs.inode, cap, ceph_cap_string(cap->issued));
804 *implemented |= cap->implemented;
807 * exclude caps issued by non-auth MDS, but are been revoking
808 * by the auth MDS. The non-auth MDS should be revoking/exporting
809 * these caps, but the message is delayed.
811 if (ci->i_auth_cap) {
812 cap = ci->i_auth_cap;
813 have &= ~cap->implemented | cap->issued;
819 * Get cap bits issued by caps other than @ocap
821 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
823 int have = ci->i_snap_caps;
824 struct ceph_cap *cap;
827 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
828 cap = rb_entry(p, struct ceph_cap, ci_node);
831 if (!__cap_is_valid(cap))
839 * Move a cap to the end of the LRU (oldest caps at list head, newest
842 static void __touch_cap(struct ceph_cap *cap)
844 struct ceph_mds_session *s = cap->session;
846 spin_lock(&s->s_cap_lock);
847 if (!s->s_cap_iterator) {
848 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->netfs.inode, cap,
850 list_move_tail(&cap->session_caps, &s->s_caps);
852 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
853 &cap->ci->netfs.inode, cap, s->s_mds);
855 spin_unlock(&s->s_cap_lock);
859 * Check if we hold the given mask. If so, move the cap(s) to the
860 * front of their respective LRUs. (This is the preferred way for
861 * callers to check for caps they want.)
863 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
865 struct ceph_cap *cap;
867 int have = ci->i_snap_caps;
869 if ((have & mask) == mask) {
870 dout("__ceph_caps_issued_mask ino 0x%llx snap issued %s"
871 " (mask %s)\n", ceph_ino(&ci->netfs.inode),
872 ceph_cap_string(have),
873 ceph_cap_string(mask));
877 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
878 cap = rb_entry(p, struct ceph_cap, ci_node);
879 if (!__cap_is_valid(cap))
881 if ((cap->issued & mask) == mask) {
882 dout("__ceph_caps_issued_mask ino 0x%llx cap %p issued %s"
883 " (mask %s)\n", ceph_ino(&ci->netfs.inode), cap,
884 ceph_cap_string(cap->issued),
885 ceph_cap_string(mask));
891 /* does a combination of caps satisfy mask? */
893 if ((have & mask) == mask) {
894 dout("__ceph_caps_issued_mask ino 0x%llx combo issued %s"
895 " (mask %s)\n", ceph_ino(&ci->netfs.inode),
896 ceph_cap_string(cap->issued),
897 ceph_cap_string(mask));
901 /* touch this + preceding caps */
903 for (q = rb_first(&ci->i_caps); q != p;
905 cap = rb_entry(q, struct ceph_cap,
907 if (!__cap_is_valid(cap))
909 if (cap->issued & mask)
920 int __ceph_caps_issued_mask_metric(struct ceph_inode_info *ci, int mask,
923 struct ceph_fs_client *fsc = ceph_sb_to_client(ci->netfs.inode.i_sb);
926 r = __ceph_caps_issued_mask(ci, mask, touch);
928 ceph_update_cap_hit(&fsc->mdsc->metric);
930 ceph_update_cap_mis(&fsc->mdsc->metric);
935 * Return true if mask caps are currently being revoked by an MDS.
937 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
938 struct ceph_cap *ocap, int mask)
940 struct ceph_cap *cap;
943 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
944 cap = rb_entry(p, struct ceph_cap, ci_node);
946 (cap->implemented & ~cap->issued & mask))
952 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
954 struct inode *inode = &ci->netfs.inode;
957 spin_lock(&ci->i_ceph_lock);
958 ret = __ceph_caps_revoking_other(ci, NULL, mask);
959 spin_unlock(&ci->i_ceph_lock);
960 dout("ceph_caps_revoking %p %s = %d\n", inode,
961 ceph_cap_string(mask), ret);
965 int __ceph_caps_used(struct ceph_inode_info *ci)
969 used |= CEPH_CAP_PIN;
971 used |= CEPH_CAP_FILE_RD;
972 if (ci->i_rdcache_ref ||
973 (S_ISREG(ci->netfs.inode.i_mode) &&
974 ci->netfs.inode.i_data.nrpages))
975 used |= CEPH_CAP_FILE_CACHE;
977 used |= CEPH_CAP_FILE_WR;
978 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
979 used |= CEPH_CAP_FILE_BUFFER;
981 used |= CEPH_CAP_FILE_EXCL;
985 #define FMODE_WAIT_BIAS 1000
988 * wanted, by virtue of open file modes
990 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
992 const int PIN_SHIFT = ffs(CEPH_FILE_MODE_PIN);
993 const int RD_SHIFT = ffs(CEPH_FILE_MODE_RD);
994 const int WR_SHIFT = ffs(CEPH_FILE_MODE_WR);
995 const int LAZY_SHIFT = ffs(CEPH_FILE_MODE_LAZY);
996 struct ceph_mount_options *opt =
997 ceph_inode_to_client(&ci->netfs.inode)->mount_options;
998 unsigned long used_cutoff = jiffies - opt->caps_wanted_delay_max * HZ;
999 unsigned long idle_cutoff = jiffies - opt->caps_wanted_delay_min * HZ;
1001 if (S_ISDIR(ci->netfs.inode.i_mode)) {
1004 /* use used_cutoff here, to keep dir's wanted caps longer */
1005 if (ci->i_nr_by_mode[RD_SHIFT] > 0 ||
1006 time_after(ci->i_last_rd, used_cutoff))
1007 want |= CEPH_CAP_ANY_SHARED;
1009 if (ci->i_nr_by_mode[WR_SHIFT] > 0 ||
1010 time_after(ci->i_last_wr, used_cutoff)) {
1011 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1012 if (opt->flags & CEPH_MOUNT_OPT_ASYNC_DIROPS)
1013 want |= CEPH_CAP_ANY_DIR_OPS;
1016 if (want || ci->i_nr_by_mode[PIN_SHIFT] > 0)
1017 want |= CEPH_CAP_PIN;
1023 if (ci->i_nr_by_mode[RD_SHIFT] > 0) {
1024 if (ci->i_nr_by_mode[RD_SHIFT] >= FMODE_WAIT_BIAS ||
1025 time_after(ci->i_last_rd, used_cutoff))
1026 bits |= 1 << RD_SHIFT;
1027 } else if (time_after(ci->i_last_rd, idle_cutoff)) {
1028 bits |= 1 << RD_SHIFT;
1031 if (ci->i_nr_by_mode[WR_SHIFT] > 0) {
1032 if (ci->i_nr_by_mode[WR_SHIFT] >= FMODE_WAIT_BIAS ||
1033 time_after(ci->i_last_wr, used_cutoff))
1034 bits |= 1 << WR_SHIFT;
1035 } else if (time_after(ci->i_last_wr, idle_cutoff)) {
1036 bits |= 1 << WR_SHIFT;
1039 /* check lazyio only when read/write is wanted */
1040 if ((bits & (CEPH_FILE_MODE_RDWR << 1)) &&
1041 ci->i_nr_by_mode[LAZY_SHIFT] > 0)
1042 bits |= 1 << LAZY_SHIFT;
1044 return bits ? ceph_caps_for_mode(bits >> 1) : 0;
1049 * wanted, by virtue of open file modes AND cap refs (buffered/cached data)
1051 int __ceph_caps_wanted(struct ceph_inode_info *ci)
1053 int w = __ceph_caps_file_wanted(ci) | __ceph_caps_used(ci);
1054 if (S_ISDIR(ci->netfs.inode.i_mode)) {
1055 /* we want EXCL if holding caps of dir ops */
1056 if (w & CEPH_CAP_ANY_DIR_OPS)
1057 w |= CEPH_CAP_FILE_EXCL;
1059 /* we want EXCL if dirty data */
1060 if (w & CEPH_CAP_FILE_BUFFER)
1061 w |= CEPH_CAP_FILE_EXCL;
1067 * Return caps we have registered with the MDS(s) as 'wanted'.
1069 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
1071 struct ceph_cap *cap;
1075 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1076 cap = rb_entry(p, struct ceph_cap, ci_node);
1077 if (check && !__cap_is_valid(cap))
1079 if (cap == ci->i_auth_cap)
1080 mds_wanted |= cap->mds_wanted;
1082 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
1087 int ceph_is_any_caps(struct inode *inode)
1089 struct ceph_inode_info *ci = ceph_inode(inode);
1092 spin_lock(&ci->i_ceph_lock);
1093 ret = __ceph_is_any_real_caps(ci);
1094 spin_unlock(&ci->i_ceph_lock);
1100 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
1102 * caller should hold i_ceph_lock.
1103 * caller will not hold session s_mutex if called from destroy_inode.
1105 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1107 struct ceph_mds_session *session = cap->session;
1108 struct ceph_inode_info *ci = cap->ci;
1109 struct ceph_mds_client *mdsc;
1112 /* 'ci' being NULL means the remove have already occurred */
1114 dout("%s: cap inode is NULL\n", __func__);
1118 lockdep_assert_held(&ci->i_ceph_lock);
1120 dout("__ceph_remove_cap %p from %p\n", cap, &ci->netfs.inode);
1122 mdsc = ceph_inode_to_client(&ci->netfs.inode)->mdsc;
1124 /* remove from inode's cap rbtree, and clear auth cap */
1125 rb_erase(&cap->ci_node, &ci->i_caps);
1126 if (ci->i_auth_cap == cap)
1127 ci->i_auth_cap = NULL;
1129 /* remove from session list */
1130 spin_lock(&session->s_cap_lock);
1131 if (session->s_cap_iterator == cap) {
1132 /* not yet, we are iterating over this very cap */
1133 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1136 list_del_init(&cap->session_caps);
1137 session->s_nr_caps--;
1138 atomic64_dec(&mdsc->metric.total_caps);
1139 cap->session = NULL;
1142 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1146 * s_cap_reconnect is protected by s_cap_lock. no one changes
1147 * s_cap_gen while session is in the reconnect state.
1149 if (queue_release &&
1150 (!session->s_cap_reconnect ||
1151 cap->cap_gen == atomic_read(&session->s_cap_gen))) {
1152 cap->queue_release = 1;
1154 __ceph_queue_cap_release(session, cap);
1158 cap->queue_release = 0;
1160 cap->cap_ino = ci->i_vino.ino;
1162 spin_unlock(&session->s_cap_lock);
1165 ceph_put_cap(mdsc, cap);
1167 if (!__ceph_is_any_real_caps(ci)) {
1168 /* when reconnect denied, we remove session caps forcibly,
1169 * i_wr_ref can be non-zero. If there are ongoing write,
1170 * keep i_snap_realm.
1172 if (ci->i_wr_ref == 0 && ci->i_snap_realm)
1173 ceph_change_snap_realm(&ci->netfs.inode, NULL);
1175 __cap_delay_cancel(mdsc, ci);
1179 void ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1181 struct ceph_inode_info *ci = cap->ci;
1182 struct ceph_fs_client *fsc;
1184 /* 'ci' being NULL means the remove have already occurred */
1186 dout("%s: cap inode is NULL\n", __func__);
1190 lockdep_assert_held(&ci->i_ceph_lock);
1192 fsc = ceph_inode_to_client(&ci->netfs.inode);
1193 WARN_ON_ONCE(ci->i_auth_cap == cap &&
1194 !list_empty(&ci->i_dirty_item) &&
1195 !fsc->blocklisted &&
1196 !ceph_inode_is_shutdown(&ci->netfs.inode));
1198 __ceph_remove_cap(cap, queue_release);
1201 struct cap_msg_args {
1202 struct ceph_mds_session *session;
1203 u64 ino, cid, follows;
1204 u64 flush_tid, oldest_flush_tid, size, max_size;
1207 struct ceph_buffer *xattr_buf;
1208 struct ceph_buffer *old_xattr_buf;
1209 struct timespec64 atime, mtime, ctime, btime;
1210 int op, caps, wanted, dirty;
1211 u32 seq, issue_seq, mseq, time_warp_seq;
1221 * cap struct size + flock buffer size + inline version + inline data size +
1222 * osd_epoch_barrier + oldest_flush_tid
1224 #define CAP_MSG_SIZE (sizeof(struct ceph_mds_caps) + \
1225 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4)
1227 /* Marshal up the cap msg to the MDS */
1228 static void encode_cap_msg(struct ceph_msg *msg, struct cap_msg_args *arg)
1230 struct ceph_mds_caps *fc;
1232 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1234 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",
1235 __func__, ceph_cap_op_name(arg->op), arg->cid, arg->ino,
1236 ceph_cap_string(arg->caps), ceph_cap_string(arg->wanted),
1237 ceph_cap_string(arg->dirty), arg->seq, arg->issue_seq,
1238 arg->flush_tid, arg->oldest_flush_tid, arg->mseq, arg->follows,
1239 arg->size, arg->max_size, arg->xattr_version,
1240 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1242 msg->hdr.version = cpu_to_le16(10);
1243 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1245 fc = msg->front.iov_base;
1246 memset(fc, 0, sizeof(*fc));
1248 fc->cap_id = cpu_to_le64(arg->cid);
1249 fc->op = cpu_to_le32(arg->op);
1250 fc->seq = cpu_to_le32(arg->seq);
1251 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1252 fc->migrate_seq = cpu_to_le32(arg->mseq);
1253 fc->caps = cpu_to_le32(arg->caps);
1254 fc->wanted = cpu_to_le32(arg->wanted);
1255 fc->dirty = cpu_to_le32(arg->dirty);
1256 fc->ino = cpu_to_le64(arg->ino);
1257 fc->snap_follows = cpu_to_le64(arg->follows);
1259 fc->size = cpu_to_le64(arg->size);
1260 fc->max_size = cpu_to_le64(arg->max_size);
1261 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1262 ceph_encode_timespec64(&fc->atime, &arg->atime);
1263 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1264 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1266 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1267 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1268 fc->mode = cpu_to_le32(arg->mode);
1270 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1271 if (arg->xattr_buf) {
1272 msg->middle = ceph_buffer_get(arg->xattr_buf);
1273 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1274 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1278 /* flock buffer size (version 2) */
1279 ceph_encode_32(&p, 0);
1280 /* inline version (version 4) */
1281 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1282 /* inline data size */
1283 ceph_encode_32(&p, 0);
1285 * osd_epoch_barrier (version 5)
1286 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1287 * case it was recently changed
1289 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1290 /* oldest_flush_tid (version 6) */
1291 ceph_encode_64(&p, arg->oldest_flush_tid);
1294 * caller_uid/caller_gid (version 7)
1296 * Currently, we don't properly track which caller dirtied the caps
1297 * last, and force a flush of them when there is a conflict. For now,
1298 * just set this to 0:0, to emulate how the MDS has worked up to now.
1300 ceph_encode_32(&p, 0);
1301 ceph_encode_32(&p, 0);
1303 /* pool namespace (version 8) (mds always ignores this) */
1304 ceph_encode_32(&p, 0);
1306 /* btime and change_attr (version 9) */
1307 ceph_encode_timespec64(p, &arg->btime);
1308 p += sizeof(struct ceph_timespec);
1309 ceph_encode_64(&p, arg->change_attr);
1311 /* Advisory flags (version 10) */
1312 ceph_encode_32(&p, arg->flags);
1316 * Queue cap releases when an inode is dropped from our cache.
1318 void __ceph_remove_caps(struct ceph_inode_info *ci)
1322 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1323 * may call __ceph_caps_issued_mask() on a freeing inode. */
1324 spin_lock(&ci->i_ceph_lock);
1325 p = rb_first(&ci->i_caps);
1327 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1329 ceph_remove_cap(cap, true);
1331 spin_unlock(&ci->i_ceph_lock);
1335 * Prepare to send a cap message to an MDS. Update the cap state, and populate
1336 * the arg struct with the parameters that will need to be sent. This should
1337 * be done under the i_ceph_lock to guard against changes to cap state.
1339 * Make note of max_size reported/requested from mds, revoked caps
1340 * that have now been implemented.
1342 static void __prep_cap(struct cap_msg_args *arg, struct ceph_cap *cap,
1343 int op, int flags, int used, int want, int retain,
1344 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1346 struct ceph_inode_info *ci = cap->ci;
1347 struct inode *inode = &ci->netfs.inode;
1350 lockdep_assert_held(&ci->i_ceph_lock);
1352 held = cap->issued | cap->implemented;
1353 revoking = cap->implemented & ~cap->issued;
1354 retain &= ~revoking;
1356 dout("%s %p cap %p session %p %s -> %s (revoking %s)\n",
1357 __func__, inode, cap, cap->session,
1358 ceph_cap_string(held), ceph_cap_string(held & retain),
1359 ceph_cap_string(revoking));
1360 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1362 ci->i_ceph_flags &= ~CEPH_I_FLUSH;
1364 cap->issued &= retain; /* drop bits we don't want */
1366 * Wake up any waiters on wanted -> needed transition. This is due to
1367 * the weird transition from buffered to sync IO... we need to flush
1368 * dirty pages _before_ allowing sync writes to avoid reordering.
1370 arg->wake = cap->implemented & ~cap->issued;
1371 cap->implemented &= cap->issued | used;
1372 cap->mds_wanted = want;
1374 arg->session = cap->session;
1375 arg->ino = ceph_vino(inode).ino;
1376 arg->cid = cap->cap_id;
1377 arg->follows = flushing ? ci->i_head_snapc->seq : 0;
1378 arg->flush_tid = flush_tid;
1379 arg->oldest_flush_tid = oldest_flush_tid;
1381 arg->size = i_size_read(inode);
1382 ci->i_reported_size = arg->size;
1383 arg->max_size = ci->i_wanted_max_size;
1384 if (cap == ci->i_auth_cap) {
1385 if (want & CEPH_CAP_ANY_FILE_WR)
1386 ci->i_requested_max_size = arg->max_size;
1388 ci->i_requested_max_size = 0;
1391 if (flushing & CEPH_CAP_XATTR_EXCL) {
1392 arg->old_xattr_buf = __ceph_build_xattrs_blob(ci);
1393 arg->xattr_version = ci->i_xattrs.version;
1394 arg->xattr_buf = ci->i_xattrs.blob;
1396 arg->xattr_buf = NULL;
1397 arg->old_xattr_buf = NULL;
1400 arg->mtime = inode->i_mtime;
1401 arg->atime = inode->i_atime;
1402 arg->ctime = inode->i_ctime;
1403 arg->btime = ci->i_btime;
1404 arg->change_attr = inode_peek_iversion_raw(inode);
1407 arg->caps = cap->implemented;
1409 arg->dirty = flushing;
1411 arg->seq = cap->seq;
1412 arg->issue_seq = cap->issue_seq;
1413 arg->mseq = cap->mseq;
1414 arg->time_warp_seq = ci->i_time_warp_seq;
1416 arg->uid = inode->i_uid;
1417 arg->gid = inode->i_gid;
1418 arg->mode = inode->i_mode;
1420 arg->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1421 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1422 !list_empty(&ci->i_cap_snaps)) {
1423 struct ceph_cap_snap *capsnap;
1424 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1425 if (capsnap->cap_flush.tid)
1427 if (capsnap->need_flush) {
1428 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1437 * Send a cap msg on the given inode.
1439 * Caller should hold snap_rwsem (read), s_mutex.
1441 static void __send_cap(struct cap_msg_args *arg, struct ceph_inode_info *ci)
1443 struct ceph_msg *msg;
1444 struct inode *inode = &ci->netfs.inode;
1446 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1448 pr_err("error allocating cap msg: ino (%llx.%llx) flushing %s tid %llu, requeuing cap.\n",
1449 ceph_vinop(inode), ceph_cap_string(arg->dirty),
1451 spin_lock(&ci->i_ceph_lock);
1452 __cap_delay_requeue(arg->session->s_mdsc, ci);
1453 spin_unlock(&ci->i_ceph_lock);
1457 encode_cap_msg(msg, arg);
1458 ceph_con_send(&arg->session->s_con, msg);
1459 ceph_buffer_put(arg->old_xattr_buf);
1461 wake_up_all(&ci->i_cap_wq);
1464 static inline int __send_flush_snap(struct inode *inode,
1465 struct ceph_mds_session *session,
1466 struct ceph_cap_snap *capsnap,
1467 u32 mseq, u64 oldest_flush_tid)
1469 struct cap_msg_args arg;
1470 struct ceph_msg *msg;
1472 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1476 arg.session = session;
1477 arg.ino = ceph_vino(inode).ino;
1479 arg.follows = capsnap->follows;
1480 arg.flush_tid = capsnap->cap_flush.tid;
1481 arg.oldest_flush_tid = oldest_flush_tid;
1483 arg.size = capsnap->size;
1485 arg.xattr_version = capsnap->xattr_version;
1486 arg.xattr_buf = capsnap->xattr_blob;
1487 arg.old_xattr_buf = NULL;
1489 arg.atime = capsnap->atime;
1490 arg.mtime = capsnap->mtime;
1491 arg.ctime = capsnap->ctime;
1492 arg.btime = capsnap->btime;
1493 arg.change_attr = capsnap->change_attr;
1495 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1496 arg.caps = capsnap->issued;
1498 arg.dirty = capsnap->dirty;
1503 arg.time_warp_seq = capsnap->time_warp_seq;
1505 arg.uid = capsnap->uid;
1506 arg.gid = capsnap->gid;
1507 arg.mode = capsnap->mode;
1509 arg.inline_data = capsnap->inline_data;
1513 encode_cap_msg(msg, &arg);
1514 ceph_con_send(&arg.session->s_con, msg);
1519 * When a snapshot is taken, clients accumulate dirty metadata on
1520 * inodes with capabilities in ceph_cap_snaps to describe the file
1521 * state at the time the snapshot was taken. This must be flushed
1522 * asynchronously back to the MDS once sync writes complete and dirty
1523 * data is written out.
1525 * Called under i_ceph_lock.
1527 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1528 struct ceph_mds_session *session)
1529 __releases(ci->i_ceph_lock)
1530 __acquires(ci->i_ceph_lock)
1532 struct inode *inode = &ci->netfs.inode;
1533 struct ceph_mds_client *mdsc = session->s_mdsc;
1534 struct ceph_cap_snap *capsnap;
1535 u64 oldest_flush_tid = 0;
1536 u64 first_tid = 1, last_tid = 0;
1538 dout("__flush_snaps %p session %p\n", inode, session);
1540 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1542 * we need to wait for sync writes to complete and for dirty
1543 * pages to be written out.
1545 if (capsnap->dirty_pages || capsnap->writing)
1548 /* should be removed by ceph_try_drop_cap_snap() */
1549 BUG_ON(!capsnap->need_flush);
1551 /* only flush each capsnap once */
1552 if (capsnap->cap_flush.tid > 0) {
1553 dout(" already flushed %p, skipping\n", capsnap);
1557 spin_lock(&mdsc->cap_dirty_lock);
1558 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1559 list_add_tail(&capsnap->cap_flush.g_list,
1560 &mdsc->cap_flush_list);
1561 if (oldest_flush_tid == 0)
1562 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1563 if (list_empty(&ci->i_flushing_item)) {
1564 list_add_tail(&ci->i_flushing_item,
1565 &session->s_cap_flushing);
1567 spin_unlock(&mdsc->cap_dirty_lock);
1569 list_add_tail(&capsnap->cap_flush.i_list,
1570 &ci->i_cap_flush_list);
1573 first_tid = capsnap->cap_flush.tid;
1574 last_tid = capsnap->cap_flush.tid;
1577 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1579 while (first_tid <= last_tid) {
1580 struct ceph_cap *cap = ci->i_auth_cap;
1581 struct ceph_cap_flush *cf = NULL, *iter;
1584 if (!(cap && cap->session == session)) {
1585 dout("__flush_snaps %p auth cap %p not mds%d, "
1586 "stop\n", inode, cap, session->s_mds);
1591 list_for_each_entry(iter, &ci->i_cap_flush_list, i_list) {
1592 if (iter->tid >= first_tid) {
1601 first_tid = cf->tid + 1;
1603 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1604 refcount_inc(&capsnap->nref);
1605 spin_unlock(&ci->i_ceph_lock);
1607 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1608 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1610 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1613 pr_err("__flush_snaps: error sending cap flushsnap, "
1614 "ino (%llx.%llx) tid %llu follows %llu\n",
1615 ceph_vinop(inode), cf->tid, capsnap->follows);
1618 ceph_put_cap_snap(capsnap);
1619 spin_lock(&ci->i_ceph_lock);
1623 void ceph_flush_snaps(struct ceph_inode_info *ci,
1624 struct ceph_mds_session **psession)
1626 struct inode *inode = &ci->netfs.inode;
1627 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1628 struct ceph_mds_session *session = NULL;
1631 dout("ceph_flush_snaps %p\n", inode);
1633 session = *psession;
1635 spin_lock(&ci->i_ceph_lock);
1636 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1637 dout(" no capsnap needs flush, doing nothing\n");
1640 if (!ci->i_auth_cap) {
1641 dout(" no auth cap (migrating?), doing nothing\n");
1645 mds = ci->i_auth_cap->session->s_mds;
1646 if (session && session->s_mds != mds) {
1647 dout(" oops, wrong session %p mutex\n", session);
1648 ceph_put_mds_session(session);
1652 spin_unlock(&ci->i_ceph_lock);
1653 mutex_lock(&mdsc->mutex);
1654 session = __ceph_lookup_mds_session(mdsc, mds);
1655 mutex_unlock(&mdsc->mutex);
1659 // make sure flushsnap messages are sent in proper order.
1660 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1661 __kick_flushing_caps(mdsc, session, ci, 0);
1663 __ceph_flush_snaps(ci, session);
1665 spin_unlock(&ci->i_ceph_lock);
1668 *psession = session;
1670 ceph_put_mds_session(session);
1671 /* we flushed them all; remove this inode from the queue */
1672 spin_lock(&mdsc->snap_flush_lock);
1673 list_del_init(&ci->i_snap_flush_item);
1674 spin_unlock(&mdsc->snap_flush_lock);
1678 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1679 * Caller is then responsible for calling __mark_inode_dirty with the
1680 * returned flags value.
1682 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1683 struct ceph_cap_flush **pcf)
1685 struct ceph_mds_client *mdsc =
1686 ceph_sb_to_client(ci->netfs.inode.i_sb)->mdsc;
1687 struct inode *inode = &ci->netfs.inode;
1688 int was = ci->i_dirty_caps;
1691 lockdep_assert_held(&ci->i_ceph_lock);
1693 if (!ci->i_auth_cap) {
1694 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1695 "but no auth cap (session was closed?)\n",
1696 inode, ceph_ino(inode), ceph_cap_string(mask));
1700 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->netfs.inode,
1701 ceph_cap_string(mask), ceph_cap_string(was),
1702 ceph_cap_string(was | mask));
1703 ci->i_dirty_caps |= mask;
1705 struct ceph_mds_session *session = ci->i_auth_cap->session;
1707 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1708 swap(ci->i_prealloc_cap_flush, *pcf);
1710 if (!ci->i_head_snapc) {
1711 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1712 ci->i_head_snapc = ceph_get_snap_context(
1713 ci->i_snap_realm->cached_context);
1715 dout(" inode %p now dirty snapc %p auth cap %p\n",
1716 &ci->netfs.inode, ci->i_head_snapc, ci->i_auth_cap);
1717 BUG_ON(!list_empty(&ci->i_dirty_item));
1718 spin_lock(&mdsc->cap_dirty_lock);
1719 list_add(&ci->i_dirty_item, &session->s_cap_dirty);
1720 spin_unlock(&mdsc->cap_dirty_lock);
1721 if (ci->i_flushing_caps == 0) {
1723 dirty |= I_DIRTY_SYNC;
1726 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1728 BUG_ON(list_empty(&ci->i_dirty_item));
1729 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1730 (mask & CEPH_CAP_FILE_BUFFER))
1731 dirty |= I_DIRTY_DATASYNC;
1732 __cap_delay_requeue(mdsc, ci);
1736 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1738 struct ceph_cap_flush *cf;
1740 cf = kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1744 cf->is_capsnap = false;
1748 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1751 kmem_cache_free(ceph_cap_flush_cachep, cf);
1754 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1756 if (!list_empty(&mdsc->cap_flush_list)) {
1757 struct ceph_cap_flush *cf =
1758 list_first_entry(&mdsc->cap_flush_list,
1759 struct ceph_cap_flush, g_list);
1766 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1767 * Return true if caller needs to wake up flush waiters.
1769 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1770 struct ceph_cap_flush *cf)
1772 struct ceph_cap_flush *prev;
1773 bool wake = cf->wake;
1775 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1776 prev = list_prev_entry(cf, g_list);
1780 list_del_init(&cf->g_list);
1784 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1785 struct ceph_cap_flush *cf)
1787 struct ceph_cap_flush *prev;
1788 bool wake = cf->wake;
1790 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1791 prev = list_prev_entry(cf, i_list);
1795 list_del_init(&cf->i_list);
1800 * Add dirty inode to the flushing list. Assigned a seq number so we
1801 * can wait for caps to flush without starving.
1803 * Called under i_ceph_lock. Returns the flush tid.
1805 static u64 __mark_caps_flushing(struct inode *inode,
1806 struct ceph_mds_session *session, bool wake,
1807 u64 *oldest_flush_tid)
1809 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1810 struct ceph_inode_info *ci = ceph_inode(inode);
1811 struct ceph_cap_flush *cf = NULL;
1814 lockdep_assert_held(&ci->i_ceph_lock);
1815 BUG_ON(ci->i_dirty_caps == 0);
1816 BUG_ON(list_empty(&ci->i_dirty_item));
1817 BUG_ON(!ci->i_prealloc_cap_flush);
1819 flushing = ci->i_dirty_caps;
1820 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1821 ceph_cap_string(flushing),
1822 ceph_cap_string(ci->i_flushing_caps),
1823 ceph_cap_string(ci->i_flushing_caps | flushing));
1824 ci->i_flushing_caps |= flushing;
1825 ci->i_dirty_caps = 0;
1826 dout(" inode %p now !dirty\n", inode);
1828 swap(cf, ci->i_prealloc_cap_flush);
1829 cf->caps = flushing;
1832 spin_lock(&mdsc->cap_dirty_lock);
1833 list_del_init(&ci->i_dirty_item);
1835 cf->tid = ++mdsc->last_cap_flush_tid;
1836 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1837 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1839 if (list_empty(&ci->i_flushing_item)) {
1840 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1841 mdsc->num_cap_flushing++;
1843 spin_unlock(&mdsc->cap_dirty_lock);
1845 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1851 * try to invalidate mapping pages without blocking.
1853 static int try_nonblocking_invalidate(struct inode *inode)
1854 __releases(ci->i_ceph_lock)
1855 __acquires(ci->i_ceph_lock)
1857 struct ceph_inode_info *ci = ceph_inode(inode);
1858 u32 invalidating_gen = ci->i_rdcache_gen;
1860 spin_unlock(&ci->i_ceph_lock);
1861 ceph_fscache_invalidate(inode, false);
1862 invalidate_mapping_pages(&inode->i_data, 0, -1);
1863 spin_lock(&ci->i_ceph_lock);
1865 if (inode->i_data.nrpages == 0 &&
1866 invalidating_gen == ci->i_rdcache_gen) {
1868 dout("try_nonblocking_invalidate %p success\n", inode);
1869 /* save any racing async invalidate some trouble */
1870 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1873 dout("try_nonblocking_invalidate %p failed\n", inode);
1877 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1879 loff_t size = i_size_read(&ci->netfs.inode);
1880 /* mds will adjust max size according to the reported size */
1881 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1883 if (size >= ci->i_max_size)
1885 /* half of previous max_size increment has been used */
1886 if (ci->i_max_size > ci->i_reported_size &&
1887 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1893 * Swiss army knife function to examine currently used and wanted
1894 * versus held caps. Release, flush, ack revoked caps to mds as
1897 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1898 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1901 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1902 struct ceph_mds_session *session)
1904 struct inode *inode = &ci->netfs.inode;
1905 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1906 struct ceph_cap *cap;
1907 u64 flush_tid, oldest_flush_tid;
1908 int file_wanted, used, cap_used;
1909 int issued, implemented, want, retain, revoking, flushing = 0;
1910 int mds = -1; /* keep track of how far we've gone through i_caps list
1911 to avoid an infinite loop on retry */
1913 bool queue_invalidate = false;
1914 bool tried_invalidate = false;
1915 bool queue_writeback = false;
1918 ceph_get_mds_session(session);
1920 spin_lock(&ci->i_ceph_lock);
1921 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
1922 /* Don't send messages until we get async create reply */
1923 spin_unlock(&ci->i_ceph_lock);
1924 ceph_put_mds_session(session);
1928 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1929 flags |= CHECK_CAPS_FLUSH;
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 %llx.%llx file_want %s used %s dirty %s flushing %s"
1982 " issued %s revoking %s retain %s %s%s%s\n", ceph_vinop(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" : "",
1990 (flags & CHECK_CAPS_NOINVAL) ? " NOINVAL" : "");
1993 * If we no longer need to hold onto old our caps, and we may
1994 * have cached pages, but don't want them, then try to invalidate.
1995 * If we fail, it's because pages are locked.... try again later.
1997 if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
1998 S_ISREG(inode->i_mode) &&
1999 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
2000 inode->i_data.nrpages && /* have cached pages */
2001 (revoking & (CEPH_CAP_FILE_CACHE|
2002 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
2003 !tried_invalidate) {
2004 dout("check_caps trying to invalidate on %llx.%llx\n",
2006 if (try_nonblocking_invalidate(inode) < 0) {
2007 dout("check_caps queuing invalidate\n");
2008 queue_invalidate = true;
2009 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2011 tried_invalidate = true;
2015 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2017 struct cap_msg_args arg;
2019 cap = rb_entry(p, struct ceph_cap, ci_node);
2021 /* avoid looping forever */
2022 if (mds >= cap->mds ||
2023 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
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? */
2070 if ((revoking & cap_used) == 0) {
2071 dout("completed revocation of %s\n",
2072 ceph_cap_string(cap->implemented & ~cap->issued));
2077 * If the "i_wrbuffer_ref" was increased by mmap or generic
2078 * cache write just before the ceph_check_caps() is called,
2079 * the Fb capability revoking will fail this time. Then we
2080 * must wait for the BDI's delayed work to flush the dirty
2081 * pages and to release the "i_wrbuffer_ref", which will cost
2082 * at most 5 seconds. That means the MDS needs to wait at
2083 * most 5 seconds to finished the Fb capability's revocation.
2085 * Let's queue a writeback for it.
2087 if (S_ISREG(inode->i_mode) && ci->i_wrbuffer_ref &&
2088 (revoking & CEPH_CAP_FILE_BUFFER))
2089 queue_writeback = true;
2092 /* want more caps from mds? */
2093 if (want & ~cap->mds_wanted) {
2094 if (want & ~(cap->mds_wanted | cap->issued))
2096 if (!__cap_is_valid(cap))
2100 /* things we might delay */
2101 if ((cap->issued & ~retain) == 0)
2102 continue; /* nope, all good */
2105 ceph_put_mds_session(session);
2106 session = ceph_get_mds_session(cap->session);
2108 /* kick flushing and flush snaps before sending normal
2110 if (cap == ci->i_auth_cap &&
2112 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2113 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2114 __kick_flushing_caps(mdsc, session, ci, 0);
2115 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2116 __ceph_flush_snaps(ci, session);
2121 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2122 flushing = ci->i_dirty_caps;
2123 flush_tid = __mark_caps_flushing(inode, session, false,
2125 if (flags & CHECK_CAPS_FLUSH &&
2126 list_empty(&session->s_cap_dirty))
2127 mflags |= CEPH_CLIENT_CAPS_SYNC;
2131 spin_lock(&mdsc->cap_dirty_lock);
2132 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2133 spin_unlock(&mdsc->cap_dirty_lock);
2136 mds = cap->mds; /* remember mds, so we don't repeat */
2138 __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2139 want, retain, flushing, flush_tid, oldest_flush_tid);
2141 spin_unlock(&ci->i_ceph_lock);
2142 __send_cap(&arg, ci);
2143 spin_lock(&ci->i_ceph_lock);
2145 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2148 /* periodically re-calculate caps wanted by open files */
2149 if (__ceph_is_any_real_caps(ci) &&
2150 list_empty(&ci->i_cap_delay_list) &&
2151 (file_wanted & ~CEPH_CAP_PIN) &&
2152 !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2153 __cap_delay_requeue(mdsc, ci);
2156 spin_unlock(&ci->i_ceph_lock);
2158 ceph_put_mds_session(session);
2159 if (queue_writeback)
2160 ceph_queue_writeback(inode);
2161 if (queue_invalidate)
2162 ceph_queue_invalidate(inode);
2166 * Try to flush dirty caps back to the auth mds.
2168 static int try_flush_caps(struct inode *inode, u64 *ptid)
2170 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2171 struct ceph_inode_info *ci = ceph_inode(inode);
2173 u64 flush_tid = 0, oldest_flush_tid = 0;
2175 spin_lock(&ci->i_ceph_lock);
2177 if (ci->i_dirty_caps && ci->i_auth_cap) {
2178 struct ceph_cap *cap = ci->i_auth_cap;
2179 struct cap_msg_args arg;
2180 struct ceph_mds_session *session = cap->session;
2182 if (session->s_state < CEPH_MDS_SESSION_OPEN) {
2183 spin_unlock(&ci->i_ceph_lock);
2187 if (ci->i_ceph_flags &
2188 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2189 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2190 __kick_flushing_caps(mdsc, session, ci, 0);
2191 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2192 __ceph_flush_snaps(ci, session);
2196 flushing = ci->i_dirty_caps;
2197 flush_tid = __mark_caps_flushing(inode, session, true,
2200 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2201 __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2202 (cap->issued | cap->implemented),
2203 flushing, flush_tid, oldest_flush_tid);
2204 spin_unlock(&ci->i_ceph_lock);
2206 __send_cap(&arg, ci);
2208 if (!list_empty(&ci->i_cap_flush_list)) {
2209 struct ceph_cap_flush *cf =
2210 list_last_entry(&ci->i_cap_flush_list,
2211 struct ceph_cap_flush, i_list);
2213 flush_tid = cf->tid;
2215 flushing = ci->i_flushing_caps;
2216 spin_unlock(&ci->i_ceph_lock);
2224 * Return true if we've flushed caps through the given flush_tid.
2226 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2228 struct ceph_inode_info *ci = ceph_inode(inode);
2231 spin_lock(&ci->i_ceph_lock);
2232 if (!list_empty(&ci->i_cap_flush_list)) {
2233 struct ceph_cap_flush * cf =
2234 list_first_entry(&ci->i_cap_flush_list,
2235 struct ceph_cap_flush, i_list);
2236 if (cf->tid <= flush_tid)
2239 spin_unlock(&ci->i_ceph_lock);
2244 * flush the mdlog and wait for any unsafe requests to complete.
2246 static int flush_mdlog_and_wait_inode_unsafe_requests(struct inode *inode)
2248 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2249 struct ceph_inode_info *ci = ceph_inode(inode);
2250 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2251 unsigned int max_sessions;
2254 spin_lock(&ci->i_unsafe_lock);
2255 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2256 req1 = list_last_entry(&ci->i_unsafe_dirops,
2257 struct ceph_mds_request,
2259 ceph_mdsc_get_request(req1);
2261 if (!list_empty(&ci->i_unsafe_iops)) {
2262 req2 = list_last_entry(&ci->i_unsafe_iops,
2263 struct ceph_mds_request,
2264 r_unsafe_target_item);
2265 ceph_mdsc_get_request(req2);
2267 spin_unlock(&ci->i_unsafe_lock);
2270 * The mdsc->max_sessions is unlikely to be changed
2271 * mostly, here we will retry it by reallocating the
2272 * sessions array memory to get rid of the mdsc->mutex
2276 max_sessions = mdsc->max_sessions;
2279 * Trigger to flush the journal logs in all the relevant MDSes
2280 * manually, or in the worst case we must wait at most 5 seconds
2281 * to wait the journal logs to be flushed by the MDSes periodically.
2283 if ((req1 || req2) && likely(max_sessions)) {
2284 struct ceph_mds_session **sessions = NULL;
2285 struct ceph_mds_session *s;
2286 struct ceph_mds_request *req;
2289 sessions = kcalloc(max_sessions, sizeof(s), GFP_KERNEL);
2295 spin_lock(&ci->i_unsafe_lock);
2297 list_for_each_entry(req, &ci->i_unsafe_dirops,
2298 r_unsafe_dir_item) {
2302 if (unlikely(s->s_mds >= max_sessions)) {
2303 spin_unlock(&ci->i_unsafe_lock);
2304 for (i = 0; i < max_sessions; i++) {
2307 ceph_put_mds_session(s);
2312 if (!sessions[s->s_mds]) {
2313 s = ceph_get_mds_session(s);
2314 sessions[s->s_mds] = s;
2319 list_for_each_entry(req, &ci->i_unsafe_iops,
2320 r_unsafe_target_item) {
2324 if (unlikely(s->s_mds >= max_sessions)) {
2325 spin_unlock(&ci->i_unsafe_lock);
2326 for (i = 0; i < max_sessions; i++) {
2329 ceph_put_mds_session(s);
2334 if (!sessions[s->s_mds]) {
2335 s = ceph_get_mds_session(s);
2336 sessions[s->s_mds] = s;
2340 spin_unlock(&ci->i_unsafe_lock);
2343 spin_lock(&ci->i_ceph_lock);
2344 if (ci->i_auth_cap) {
2345 s = ci->i_auth_cap->session;
2346 if (!sessions[s->s_mds])
2347 sessions[s->s_mds] = ceph_get_mds_session(s);
2349 spin_unlock(&ci->i_ceph_lock);
2351 /* send flush mdlog request to MDSes */
2352 for (i = 0; i < max_sessions; i++) {
2355 send_flush_mdlog(s);
2356 ceph_put_mds_session(s);
2362 dout("%s %p wait on tid %llu %llu\n", __func__,
2363 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2365 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2366 ceph_timeout_jiffies(req1->r_timeout));
2371 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2372 ceph_timeout_jiffies(req2->r_timeout));
2379 ceph_mdsc_put_request(req1);
2381 ceph_mdsc_put_request(req2);
2385 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2387 struct inode *inode = file->f_mapping->host;
2388 struct ceph_inode_info *ci = ceph_inode(inode);
2393 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2395 ret = file_write_and_wait_range(file, start, end);
2399 ret = ceph_wait_on_async_create(inode);
2403 dirty = try_flush_caps(inode, &flush_tid);
2404 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2406 err = flush_mdlog_and_wait_inode_unsafe_requests(inode);
2409 * only wait on non-file metadata writeback (the mds
2410 * can recover size and mtime, so we don't need to
2413 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2414 err = wait_event_interruptible(ci->i_cap_wq,
2415 caps_are_flushed(inode, flush_tid));
2421 err = file_check_and_advance_wb_err(file);
2425 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2430 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2431 * queue inode for flush but don't do so immediately, because we can
2432 * get by with fewer MDS messages if we wait for data writeback to
2435 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2437 struct ceph_inode_info *ci = ceph_inode(inode);
2441 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2443 dout("write_inode %p wait=%d\n", inode, wait);
2444 ceph_fscache_unpin_writeback(inode, wbc);
2446 err = ceph_wait_on_async_create(inode);
2449 dirty = try_flush_caps(inode, &flush_tid);
2451 err = wait_event_interruptible(ci->i_cap_wq,
2452 caps_are_flushed(inode, flush_tid));
2454 struct ceph_mds_client *mdsc =
2455 ceph_sb_to_client(inode->i_sb)->mdsc;
2457 spin_lock(&ci->i_ceph_lock);
2458 if (__ceph_caps_dirty(ci))
2459 __cap_delay_requeue_front(mdsc, ci);
2460 spin_unlock(&ci->i_ceph_lock);
2465 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2466 struct ceph_mds_session *session,
2467 struct ceph_inode_info *ci,
2468 u64 oldest_flush_tid)
2469 __releases(ci->i_ceph_lock)
2470 __acquires(ci->i_ceph_lock)
2472 struct inode *inode = &ci->netfs.inode;
2473 struct ceph_cap *cap;
2474 struct ceph_cap_flush *cf;
2477 u64 last_snap_flush = 0;
2479 /* Don't do anything until create reply comes in */
2480 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE)
2483 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2485 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2486 if (cf->is_capsnap) {
2487 last_snap_flush = cf->tid;
2492 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2493 if (cf->tid < first_tid)
2496 cap = ci->i_auth_cap;
2497 if (!(cap && cap->session == session)) {
2498 pr_err("%p auth cap %p not mds%d ???\n",
2499 inode, cap, session->s_mds);
2503 first_tid = cf->tid + 1;
2505 if (!cf->is_capsnap) {
2506 struct cap_msg_args arg;
2508 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2509 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2510 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2511 (cf->tid < last_snap_flush ?
2512 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2513 __ceph_caps_used(ci),
2514 __ceph_caps_wanted(ci),
2515 (cap->issued | cap->implemented),
2516 cf->caps, cf->tid, oldest_flush_tid);
2517 spin_unlock(&ci->i_ceph_lock);
2518 __send_cap(&arg, ci);
2520 struct ceph_cap_snap *capsnap =
2521 container_of(cf, struct ceph_cap_snap,
2523 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2524 inode, capsnap, cf->tid,
2525 ceph_cap_string(capsnap->dirty));
2527 refcount_inc(&capsnap->nref);
2528 spin_unlock(&ci->i_ceph_lock);
2530 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2533 pr_err("kick_flushing_caps: error sending "
2534 "cap flushsnap, ino (%llx.%llx) "
2535 "tid %llu follows %llu\n",
2536 ceph_vinop(inode), cf->tid,
2540 ceph_put_cap_snap(capsnap);
2543 spin_lock(&ci->i_ceph_lock);
2547 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2548 struct ceph_mds_session *session)
2550 struct ceph_inode_info *ci;
2551 struct ceph_cap *cap;
2552 u64 oldest_flush_tid;
2554 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2556 spin_lock(&mdsc->cap_dirty_lock);
2557 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2558 spin_unlock(&mdsc->cap_dirty_lock);
2560 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2561 spin_lock(&ci->i_ceph_lock);
2562 cap = ci->i_auth_cap;
2563 if (!(cap && cap->session == session)) {
2564 pr_err("%p auth cap %p not mds%d ???\n",
2565 &ci->netfs.inode, cap, session->s_mds);
2566 spin_unlock(&ci->i_ceph_lock);
2572 * if flushing caps were revoked, we re-send the cap flush
2573 * in client reconnect stage. This guarantees MDS * processes
2574 * the cap flush message before issuing the flushing caps to
2577 if ((cap->issued & ci->i_flushing_caps) !=
2578 ci->i_flushing_caps) {
2579 /* encode_caps_cb() also will reset these sequence
2580 * numbers. make sure sequence numbers in cap flush
2581 * message match later reconnect message */
2585 __kick_flushing_caps(mdsc, session, ci,
2588 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2591 spin_unlock(&ci->i_ceph_lock);
2595 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2596 struct ceph_mds_session *session)
2598 struct ceph_inode_info *ci;
2599 struct ceph_cap *cap;
2600 u64 oldest_flush_tid;
2602 lockdep_assert_held(&session->s_mutex);
2604 dout("kick_flushing_caps mds%d\n", session->s_mds);
2606 spin_lock(&mdsc->cap_dirty_lock);
2607 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2608 spin_unlock(&mdsc->cap_dirty_lock);
2610 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2611 spin_lock(&ci->i_ceph_lock);
2612 cap = ci->i_auth_cap;
2613 if (!(cap && cap->session == session)) {
2614 pr_err("%p auth cap %p not mds%d ???\n",
2615 &ci->netfs.inode, cap, session->s_mds);
2616 spin_unlock(&ci->i_ceph_lock);
2619 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2620 __kick_flushing_caps(mdsc, session, ci,
2623 spin_unlock(&ci->i_ceph_lock);
2627 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2628 struct ceph_inode_info *ci)
2630 struct ceph_mds_client *mdsc = session->s_mdsc;
2631 struct ceph_cap *cap = ci->i_auth_cap;
2633 lockdep_assert_held(&ci->i_ceph_lock);
2635 dout("%s %p flushing %s\n", __func__, &ci->netfs.inode,
2636 ceph_cap_string(ci->i_flushing_caps));
2638 if (!list_empty(&ci->i_cap_flush_list)) {
2639 u64 oldest_flush_tid;
2640 spin_lock(&mdsc->cap_dirty_lock);
2641 list_move_tail(&ci->i_flushing_item,
2642 &cap->session->s_cap_flushing);
2643 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2644 spin_unlock(&mdsc->cap_dirty_lock);
2646 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2652 * Take references to capabilities we hold, so that we don't release
2653 * them to the MDS prematurely.
2655 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2656 bool snap_rwsem_locked)
2658 lockdep_assert_held(&ci->i_ceph_lock);
2660 if (got & CEPH_CAP_PIN)
2662 if (got & CEPH_CAP_FILE_RD)
2664 if (got & CEPH_CAP_FILE_CACHE)
2665 ci->i_rdcache_ref++;
2666 if (got & CEPH_CAP_FILE_EXCL)
2668 if (got & CEPH_CAP_FILE_WR) {
2669 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2670 BUG_ON(!snap_rwsem_locked);
2671 ci->i_head_snapc = ceph_get_snap_context(
2672 ci->i_snap_realm->cached_context);
2676 if (got & CEPH_CAP_FILE_BUFFER) {
2677 if (ci->i_wb_ref == 0)
2678 ihold(&ci->netfs.inode);
2680 dout("%s %p wb %d -> %d (?)\n", __func__,
2681 &ci->netfs.inode, ci->i_wb_ref-1, ci->i_wb_ref);
2686 * Try to grab cap references. Specify those refs we @want, and the
2687 * minimal set we @need. Also include the larger offset we are writing
2688 * to (when applicable), and check against max_size here as well.
2689 * Note that caller is responsible for ensuring max_size increases are
2690 * requested from the MDS.
2692 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2693 * or a negative error code. There are 3 speical error codes:
2694 * -EAGAIN: need to sleep but non-blocking is specified
2695 * -EFBIG: ask caller to call check_max_size() and try again.
2696 * -EUCLEAN: ask caller to call ceph_renew_caps() and try again.
2699 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2700 NON_BLOCKING = (1 << 8),
2701 CHECK_FILELOCK = (1 << 9),
2704 static int try_get_cap_refs(struct inode *inode, int need, int want,
2705 loff_t endoff, int flags, int *got)
2707 struct ceph_inode_info *ci = ceph_inode(inode);
2708 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2710 int have, implemented;
2711 bool snap_rwsem_locked = false;
2713 dout("get_cap_refs %p need %s want %s\n", inode,
2714 ceph_cap_string(need), ceph_cap_string(want));
2717 spin_lock(&ci->i_ceph_lock);
2719 if ((flags & CHECK_FILELOCK) &&
2720 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2721 dout("try_get_cap_refs %p error filelock\n", inode);
2726 /* finish pending truncate */
2727 while (ci->i_truncate_pending) {
2728 spin_unlock(&ci->i_ceph_lock);
2729 if (snap_rwsem_locked) {
2730 up_read(&mdsc->snap_rwsem);
2731 snap_rwsem_locked = false;
2733 __ceph_do_pending_vmtruncate(inode);
2734 spin_lock(&ci->i_ceph_lock);
2737 have = __ceph_caps_issued(ci, &implemented);
2739 if (have & need & CEPH_CAP_FILE_WR) {
2740 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2741 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2742 inode, endoff, ci->i_max_size);
2743 if (endoff > ci->i_requested_max_size)
2744 ret = ci->i_auth_cap ? -EFBIG : -EUCLEAN;
2748 * If a sync write is in progress, we must wait, so that we
2749 * can get a final snapshot value for size+mtime.
2751 if (__ceph_have_pending_cap_snap(ci)) {
2752 dout("get_cap_refs %p cap_snap_pending\n", inode);
2757 if ((have & need) == need) {
2759 * Look at (implemented & ~have & not) so that we keep waiting
2760 * on transition from wanted -> needed caps. This is needed
2761 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2762 * going before a prior buffered writeback happens.
2764 * For RDCACHE|RD -> RD, there is not need to wait and we can
2765 * just exclude the revoking caps and force to sync read.
2767 int not = want & ~(have & need);
2768 int revoking = implemented & ~have;
2769 int exclude = revoking & not;
2770 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2771 inode, ceph_cap_string(have), ceph_cap_string(not),
2772 ceph_cap_string(revoking));
2773 if (!exclude || !(exclude & CEPH_CAP_FILE_BUFFER)) {
2774 if (!snap_rwsem_locked &&
2775 !ci->i_head_snapc &&
2776 (need & CEPH_CAP_FILE_WR)) {
2777 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2779 * we can not call down_read() when
2780 * task isn't in TASK_RUNNING state
2782 if (flags & NON_BLOCKING) {
2787 spin_unlock(&ci->i_ceph_lock);
2788 down_read(&mdsc->snap_rwsem);
2789 snap_rwsem_locked = true;
2792 snap_rwsem_locked = true;
2794 if ((have & want) == want)
2795 *got = need | (want & ~exclude);
2798 ceph_take_cap_refs(ci, *got, true);
2802 int session_readonly = false;
2804 if (ci->i_auth_cap &&
2805 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2806 struct ceph_mds_session *s = ci->i_auth_cap->session;
2807 spin_lock(&s->s_cap_lock);
2808 session_readonly = s->s_readonly;
2809 spin_unlock(&s->s_cap_lock);
2811 if (session_readonly) {
2812 dout("get_cap_refs %p need %s but mds%d readonly\n",
2813 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2818 if (ceph_inode_is_shutdown(inode)) {
2819 dout("get_cap_refs %p inode is shutdown\n", inode);
2823 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2824 if (need & ~mds_wanted) {
2825 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2826 inode, ceph_cap_string(need),
2827 ceph_cap_string(mds_wanted));
2832 dout("get_cap_refs %p have %s need %s\n", inode,
2833 ceph_cap_string(have), ceph_cap_string(need));
2837 __ceph_touch_fmode(ci, mdsc, flags);
2839 spin_unlock(&ci->i_ceph_lock);
2840 if (snap_rwsem_locked)
2841 up_read(&mdsc->snap_rwsem);
2844 ceph_update_cap_mis(&mdsc->metric);
2846 ceph_update_cap_hit(&mdsc->metric);
2848 dout("get_cap_refs %p ret %d got %s\n", inode,
2849 ret, ceph_cap_string(*got));
2854 * Check the offset we are writing up to against our current
2855 * max_size. If necessary, tell the MDS we want to write to
2858 static void check_max_size(struct inode *inode, loff_t endoff)
2860 struct ceph_inode_info *ci = ceph_inode(inode);
2863 /* do we need to explicitly request a larger max_size? */
2864 spin_lock(&ci->i_ceph_lock);
2865 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2866 dout("write %p at large endoff %llu, req max_size\n",
2868 ci->i_wanted_max_size = endoff;
2870 /* duplicate ceph_check_caps()'s logic */
2871 if (ci->i_auth_cap &&
2872 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2873 ci->i_wanted_max_size > ci->i_max_size &&
2874 ci->i_wanted_max_size > ci->i_requested_max_size)
2876 spin_unlock(&ci->i_ceph_lock);
2878 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2881 static inline int get_used_fmode(int caps)
2884 if (caps & CEPH_CAP_FILE_RD)
2885 fmode |= CEPH_FILE_MODE_RD;
2886 if (caps & CEPH_CAP_FILE_WR)
2887 fmode |= CEPH_FILE_MODE_WR;
2891 int ceph_try_get_caps(struct inode *inode, int need, int want,
2892 bool nonblock, int *got)
2896 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2897 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2898 CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2899 CEPH_CAP_ANY_DIR_OPS));
2901 ret = ceph_pool_perm_check(inode, need);
2906 flags = get_used_fmode(need | want);
2908 flags |= NON_BLOCKING;
2910 ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2911 /* three special error codes */
2912 if (ret == -EAGAIN || ret == -EFBIG || ret == -EUCLEAN)
2918 * Wait for caps, and take cap references. If we can't get a WR cap
2919 * due to a small max_size, make sure we check_max_size (and possibly
2920 * ask the mds) so we don't get hung up indefinitely.
2922 int ceph_get_caps(struct file *filp, int need, int want, loff_t endoff, int *got)
2924 struct ceph_file_info *fi = filp->private_data;
2925 struct inode *inode = file_inode(filp);
2926 struct ceph_inode_info *ci = ceph_inode(inode);
2927 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2928 int ret, _got, flags;
2930 ret = ceph_pool_perm_check(inode, need);
2934 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2935 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2938 flags = get_used_fmode(need | want);
2941 flags &= CEPH_FILE_MODE_MASK;
2942 if (atomic_read(&fi->num_locks))
2943 flags |= CHECK_FILELOCK;
2945 ret = try_get_cap_refs(inode, need, want, endoff,
2947 WARN_ON_ONCE(ret == -EAGAIN);
2949 struct ceph_mds_client *mdsc = fsc->mdsc;
2951 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2953 cw.ino = ceph_ino(inode);
2954 cw.tgid = current->tgid;
2958 spin_lock(&mdsc->caps_list_lock);
2959 list_add(&cw.list, &mdsc->cap_wait_list);
2960 spin_unlock(&mdsc->caps_list_lock);
2962 /* make sure used fmode not timeout */
2963 ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2964 add_wait_queue(&ci->i_cap_wq, &wait);
2966 flags |= NON_BLOCKING;
2967 while (!(ret = try_get_cap_refs(inode, need, want,
2968 endoff, flags, &_got))) {
2969 if (signal_pending(current)) {
2973 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2976 remove_wait_queue(&ci->i_cap_wq, &wait);
2977 ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
2979 spin_lock(&mdsc->caps_list_lock);
2981 spin_unlock(&mdsc->caps_list_lock);
2987 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2988 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2989 if (ret >= 0 && _got)
2990 ceph_put_cap_refs(ci, _got);
2995 if (ret == -EFBIG || ret == -EUCLEAN) {
2996 int ret2 = ceph_wait_on_async_create(inode);
3000 if (ret == -EFBIG) {
3001 check_max_size(inode, endoff);
3004 if (ret == -EUCLEAN) {
3005 /* session was killed, try renew caps */
3006 ret = ceph_renew_caps(inode, flags);
3013 if (S_ISREG(ci->netfs.inode.i_mode) &&
3014 ceph_has_inline_data(ci) &&
3015 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
3016 i_size_read(inode) > 0) {
3018 find_get_page(inode->i_mapping, 0);
3020 bool uptodate = PageUptodate(page);
3027 * drop cap refs first because getattr while
3028 * holding * caps refs can cause deadlock.
3030 ceph_put_cap_refs(ci, _got);
3034 * getattr request will bring inline data into
3037 ret = __ceph_do_getattr(inode, NULL,
3038 CEPH_STAT_CAP_INLINE_DATA,
3051 * Take cap refs. Caller must already know we hold at least one ref
3052 * on the caps in question or we don't know this is safe.
3054 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
3056 spin_lock(&ci->i_ceph_lock);
3057 ceph_take_cap_refs(ci, caps, false);
3058 spin_unlock(&ci->i_ceph_lock);
3063 * drop cap_snap that is not associated with any snapshot.
3064 * we don't need to send FLUSHSNAP message for it.
3066 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
3067 struct ceph_cap_snap *capsnap)
3069 if (!capsnap->need_flush &&
3070 !capsnap->writing && !capsnap->dirty_pages) {
3071 dout("dropping cap_snap %p follows %llu\n",
3072 capsnap, capsnap->follows);
3073 BUG_ON(capsnap->cap_flush.tid > 0);
3074 ceph_put_snap_context(capsnap->context);
3075 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
3076 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3078 list_del(&capsnap->ci_item);
3079 ceph_put_cap_snap(capsnap);
3085 enum put_cap_refs_mode {
3086 PUT_CAP_REFS_SYNC = 0,
3087 PUT_CAP_REFS_NO_CHECK,
3094 * If we released the last ref on any given cap, call ceph_check_caps
3095 * to release (or schedule a release).
3097 * If we are releasing a WR cap (from a sync write), finalize any affected
3098 * cap_snap, and wake up any waiters.
3100 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
3101 enum put_cap_refs_mode mode)
3103 struct inode *inode = &ci->netfs.inode;
3104 int last = 0, put = 0, flushsnaps = 0, wake = 0;
3105 bool check_flushsnaps = false;
3107 spin_lock(&ci->i_ceph_lock);
3108 if (had & CEPH_CAP_PIN)
3110 if (had & CEPH_CAP_FILE_RD)
3111 if (--ci->i_rd_ref == 0)
3113 if (had & CEPH_CAP_FILE_CACHE)
3114 if (--ci->i_rdcache_ref == 0)
3116 if (had & CEPH_CAP_FILE_EXCL)
3117 if (--ci->i_fx_ref == 0)
3119 if (had & CEPH_CAP_FILE_BUFFER) {
3120 if (--ci->i_wb_ref == 0) {
3122 /* put the ref held by ceph_take_cap_refs() */
3124 check_flushsnaps = true;
3126 dout("put_cap_refs %p wb %d -> %d (?)\n",
3127 inode, ci->i_wb_ref+1, ci->i_wb_ref);
3129 if (had & CEPH_CAP_FILE_WR) {
3130 if (--ci->i_wr_ref == 0) {
3132 check_flushsnaps = true;
3133 if (ci->i_wrbuffer_ref_head == 0 &&
3134 ci->i_dirty_caps == 0 &&
3135 ci->i_flushing_caps == 0) {
3136 BUG_ON(!ci->i_head_snapc);
3137 ceph_put_snap_context(ci->i_head_snapc);
3138 ci->i_head_snapc = NULL;
3140 /* see comment in __ceph_remove_cap() */
3141 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3142 ceph_change_snap_realm(inode, NULL);
3145 if (check_flushsnaps && __ceph_have_pending_cap_snap(ci)) {
3146 struct ceph_cap_snap *capsnap =
3147 list_last_entry(&ci->i_cap_snaps,
3148 struct ceph_cap_snap,
3151 capsnap->writing = 0;
3152 if (ceph_try_drop_cap_snap(ci, capsnap))
3153 /* put the ref held by ceph_queue_cap_snap() */
3155 else if (__ceph_finish_cap_snap(ci, capsnap))
3159 spin_unlock(&ci->i_ceph_lock);
3161 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3162 last ? " last" : "", put ? " put" : "");
3165 case PUT_CAP_REFS_SYNC:
3167 ceph_check_caps(ci, 0, NULL);
3168 else if (flushsnaps)
3169 ceph_flush_snaps(ci, NULL);
3171 case PUT_CAP_REFS_ASYNC:
3173 ceph_queue_check_caps(inode);
3174 else if (flushsnaps)
3175 ceph_queue_flush_snaps(inode);
3181 wake_up_all(&ci->i_cap_wq);
3186 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3188 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_SYNC);
3191 void ceph_put_cap_refs_async(struct ceph_inode_info *ci, int had)
3193 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_ASYNC);
3196 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3198 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_NO_CHECK);
3202 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3203 * context. Adjust per-snap dirty page accounting as appropriate.
3204 * Once all dirty data for a cap_snap is flushed, flush snapped file
3205 * metadata back to the MDS. If we dropped the last ref, call
3208 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3209 struct ceph_snap_context *snapc)
3211 struct inode *inode = &ci->netfs.inode;
3212 struct ceph_cap_snap *capsnap = NULL, *iter;
3215 bool flush_snaps = false;
3216 bool complete_capsnap = false;
3218 spin_lock(&ci->i_ceph_lock);
3219 ci->i_wrbuffer_ref -= nr;
3220 if (ci->i_wrbuffer_ref == 0) {
3225 if (ci->i_head_snapc == snapc) {
3226 ci->i_wrbuffer_ref_head -= nr;
3227 if (ci->i_wrbuffer_ref_head == 0 &&
3228 ci->i_wr_ref == 0 &&
3229 ci->i_dirty_caps == 0 &&
3230 ci->i_flushing_caps == 0) {
3231 BUG_ON(!ci->i_head_snapc);
3232 ceph_put_snap_context(ci->i_head_snapc);
3233 ci->i_head_snapc = NULL;
3235 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3237 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3238 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3239 last ? " LAST" : "");
3241 list_for_each_entry(iter, &ci->i_cap_snaps, ci_item) {
3242 if (iter->context == snapc) {
3250 * The capsnap should already be removed when removing
3251 * auth cap in the case of a forced unmount.
3253 WARN_ON_ONCE(ci->i_auth_cap);
3257 capsnap->dirty_pages -= nr;
3258 if (capsnap->dirty_pages == 0) {
3259 complete_capsnap = true;
3260 if (!capsnap->writing) {
3261 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3264 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3269 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3270 " snap %lld %d/%d -> %d/%d %s%s\n",
3271 inode, capsnap, capsnap->context->seq,
3272 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3273 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3274 last ? " (wrbuffer last)" : "",
3275 complete_capsnap ? " (complete capsnap)" : "");
3279 spin_unlock(&ci->i_ceph_lock);
3282 ceph_check_caps(ci, 0, NULL);
3283 } else if (flush_snaps) {
3284 ceph_flush_snaps(ci, NULL);
3286 if (complete_capsnap)
3287 wake_up_all(&ci->i_cap_wq);
3294 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3296 static void invalidate_aliases(struct inode *inode)
3298 struct dentry *dn, *prev = NULL;
3300 dout("invalidate_aliases inode %p\n", inode);
3301 d_prune_aliases(inode);
3303 * For non-directory inode, d_find_alias() only returns
3304 * hashed dentry. After calling d_invalidate(), the
3305 * dentry becomes unhashed.
3307 * For directory inode, d_find_alias() can return
3308 * unhashed dentry. But directory inode should have
3309 * one alias at most.
3311 while ((dn = d_find_alias(inode))) {
3325 struct cap_extra_info {
3326 struct ceph_string *pool_ns;
3336 /* currently issued */
3338 struct timespec64 btime;
3342 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3343 * actually be a revocation if it specifies a smaller cap set.)
3345 * caller holds s_mutex and i_ceph_lock, we drop both.
3347 static void handle_cap_grant(struct inode *inode,
3348 struct ceph_mds_session *session,
3349 struct ceph_cap *cap,
3350 struct ceph_mds_caps *grant,
3351 struct ceph_buffer *xattr_buf,
3352 struct cap_extra_info *extra_info)
3353 __releases(ci->i_ceph_lock)
3354 __releases(session->s_mdsc->snap_rwsem)
3356 struct ceph_inode_info *ci = ceph_inode(inode);
3357 int seq = le32_to_cpu(grant->seq);
3358 int newcaps = le32_to_cpu(grant->caps);
3359 int used, wanted, dirty;
3360 u64 size = le64_to_cpu(grant->size);
3361 u64 max_size = le64_to_cpu(grant->max_size);
3362 unsigned char check_caps = 0;
3363 bool was_stale = cap->cap_gen < atomic_read(&session->s_cap_gen);
3365 bool writeback = false;
3366 bool queue_trunc = false;
3367 bool queue_invalidate = false;
3368 bool deleted_inode = false;
3369 bool fill_inline = false;
3371 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3372 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3373 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3374 i_size_read(inode));
3378 * If CACHE is being revoked, and we have no dirty buffers,
3379 * try to invalidate (once). (If there are dirty buffers, we
3380 * will invalidate _after_ writeback.)
3382 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3383 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3384 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3385 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3386 if (try_nonblocking_invalidate(inode)) {
3387 /* there were locked pages.. invalidate later
3388 in a separate thread. */
3389 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3390 queue_invalidate = true;
3391 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3397 cap->issued = cap->implemented = CEPH_CAP_PIN;
3400 * auth mds of the inode changed. we received the cap export message,
3401 * but still haven't received the cap import message. handle_cap_export
3402 * updated the new auth MDS' cap.
3404 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3405 * that was sent before the cap import message. So don't remove caps.
3407 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3408 WARN_ON(cap != ci->i_auth_cap);
3409 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3411 newcaps |= cap->issued;
3414 /* side effects now are allowed */
3415 cap->cap_gen = atomic_read(&session->s_cap_gen);
3418 __check_cap_issue(ci, cap, newcaps);
3420 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3422 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3423 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3424 umode_t mode = le32_to_cpu(grant->mode);
3426 if (inode_wrong_type(inode, mode))
3427 pr_warn_once("inode type changed! (ino %llx.%llx is 0%o, mds says 0%o)\n",
3428 ceph_vinop(inode), inode->i_mode, mode);
3430 inode->i_mode = mode;
3431 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3432 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3433 ci->i_btime = extra_info->btime;
3434 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3435 from_kuid(&init_user_ns, inode->i_uid),
3436 from_kgid(&init_user_ns, inode->i_gid));
3439 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3440 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3441 set_nlink(inode, le32_to_cpu(grant->nlink));
3442 if (inode->i_nlink == 0)
3443 deleted_inode = true;
3446 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3448 int len = le32_to_cpu(grant->xattr_len);
3449 u64 version = le64_to_cpu(grant->xattr_version);
3451 if (version > ci->i_xattrs.version) {
3452 dout(" got new xattrs v%llu on %p len %d\n",
3453 version, inode, len);
3454 if (ci->i_xattrs.blob)
3455 ceph_buffer_put(ci->i_xattrs.blob);
3456 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3457 ci->i_xattrs.version = version;
3458 ceph_forget_all_cached_acls(inode);
3459 ceph_security_invalidate_secctx(inode);
3463 if (newcaps & CEPH_CAP_ANY_RD) {
3464 struct timespec64 mtime, atime, ctime;
3465 /* ctime/mtime/atime? */
3466 ceph_decode_timespec64(&mtime, &grant->mtime);
3467 ceph_decode_timespec64(&atime, &grant->atime);
3468 ceph_decode_timespec64(&ctime, &grant->ctime);
3469 ceph_fill_file_time(inode, extra_info->issued,
3470 le32_to_cpu(grant->time_warp_seq),
3471 &ctime, &mtime, &atime);
3474 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3475 ci->i_files = extra_info->nfiles;
3476 ci->i_subdirs = extra_info->nsubdirs;
3479 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3480 /* file layout may have changed */
3481 s64 old_pool = ci->i_layout.pool_id;
3482 struct ceph_string *old_ns;
3484 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3485 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3486 lockdep_is_held(&ci->i_ceph_lock));
3487 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3489 if (ci->i_layout.pool_id != old_pool ||
3490 extra_info->pool_ns != old_ns)
3491 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3493 extra_info->pool_ns = old_ns;
3495 /* size/truncate_seq? */
3496 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3497 le32_to_cpu(grant->truncate_seq),
3498 le64_to_cpu(grant->truncate_size),
3502 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3503 if (max_size != ci->i_max_size) {
3504 dout("max_size %lld -> %llu\n",
3505 ci->i_max_size, max_size);
3506 ci->i_max_size = max_size;
3507 if (max_size >= ci->i_wanted_max_size) {
3508 ci->i_wanted_max_size = 0; /* reset */
3509 ci->i_requested_max_size = 0;
3515 /* check cap bits */
3516 wanted = __ceph_caps_wanted(ci);
3517 used = __ceph_caps_used(ci);
3518 dirty = __ceph_caps_dirty(ci);
3519 dout(" my wanted = %s, used = %s, dirty %s\n",
3520 ceph_cap_string(wanted),
3521 ceph_cap_string(used),
3522 ceph_cap_string(dirty));
3524 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3525 (wanted & ~(cap->mds_wanted | newcaps))) {
3527 * If mds is importing cap, prior cap messages that update
3528 * 'wanted' may get dropped by mds (migrate seq mismatch).
3530 * We don't send cap message to update 'wanted' if what we
3531 * want are already issued. If mds revokes caps, cap message
3532 * that releases caps also tells mds what we want. But if
3533 * caps got revoked by mds forcedly (session stale). We may
3534 * haven't told mds what we want.
3539 /* revocation, grant, or no-op? */
3540 if (cap->issued & ~newcaps) {
3541 int revoking = cap->issued & ~newcaps;
3543 dout("revocation: %s -> %s (revoking %s)\n",
3544 ceph_cap_string(cap->issued),
3545 ceph_cap_string(newcaps),
3546 ceph_cap_string(revoking));
3547 if (S_ISREG(inode->i_mode) &&
3548 (revoking & used & CEPH_CAP_FILE_BUFFER))
3549 writeback = true; /* initiate writeback; will delay ack */
3550 else if (queue_invalidate &&
3551 revoking == CEPH_CAP_FILE_CACHE &&
3552 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3553 ; /* do nothing yet, invalidation will be queued */
3554 else if (cap == ci->i_auth_cap)
3555 check_caps = 1; /* check auth cap only */
3557 check_caps = 2; /* check all caps */
3558 /* If there is new caps, try to wake up the waiters */
3559 if (~cap->issued & newcaps)
3561 cap->issued = newcaps;
3562 cap->implemented |= newcaps;
3563 } else if (cap->issued == newcaps) {
3564 dout("caps unchanged: %s -> %s\n",
3565 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3567 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3568 ceph_cap_string(newcaps));
3569 /* non-auth MDS is revoking the newly grant caps ? */
3570 if (cap == ci->i_auth_cap &&
3571 __ceph_caps_revoking_other(ci, cap, newcaps))
3574 cap->issued = newcaps;
3575 cap->implemented |= newcaps; /* add bits only, to
3576 * avoid stepping on a
3577 * pending revocation */
3580 BUG_ON(cap->issued & ~cap->implemented);
3582 if (extra_info->inline_version > 0 &&
3583 extra_info->inline_version >= ci->i_inline_version) {
3584 ci->i_inline_version = extra_info->inline_version;
3585 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3586 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3590 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3591 if (ci->i_auth_cap == cap) {
3592 if (newcaps & ~extra_info->issued)
3595 if (ci->i_requested_max_size > max_size ||
3596 !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3597 /* re-request max_size if necessary */
3598 ci->i_requested_max_size = 0;
3602 ceph_kick_flushing_inode_caps(session, ci);
3604 up_read(&session->s_mdsc->snap_rwsem);
3606 spin_unlock(&ci->i_ceph_lock);
3609 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3610 extra_info->inline_len);
3613 ceph_queue_vmtruncate(inode);
3617 * queue inode for writeback: we can't actually call
3618 * filemap_write_and_wait, etc. from message handler
3621 ceph_queue_writeback(inode);
3622 if (queue_invalidate)
3623 ceph_queue_invalidate(inode);
3625 invalidate_aliases(inode);
3627 wake_up_all(&ci->i_cap_wq);
3629 mutex_unlock(&session->s_mutex);
3630 if (check_caps == 1)
3631 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3633 else if (check_caps == 2)
3634 ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3638 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3639 * MDS has been safely committed.
3641 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3642 struct ceph_mds_caps *m,
3643 struct ceph_mds_session *session,
3644 struct ceph_cap *cap)
3645 __releases(ci->i_ceph_lock)
3647 struct ceph_inode_info *ci = ceph_inode(inode);
3648 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3649 struct ceph_cap_flush *cf, *tmp_cf;
3650 LIST_HEAD(to_remove);
3651 unsigned seq = le32_to_cpu(m->seq);
3652 int dirty = le32_to_cpu(m->dirty);
3655 bool wake_ci = false;
3656 bool wake_mdsc = false;
3658 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3659 /* Is this the one that was flushed? */
3660 if (cf->tid == flush_tid)
3663 /* Is this a capsnap? */
3667 if (cf->tid <= flush_tid) {
3669 * An earlier or current tid. The FLUSH_ACK should
3670 * represent a superset of this flush's caps.
3672 wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3673 list_add_tail(&cf->i_list, &to_remove);
3676 * This is a later one. Any caps in it are still dirty
3677 * so don't count them as cleaned.
3679 cleaned &= ~cf->caps;
3685 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3686 " flushing %s -> %s\n",
3687 inode, session->s_mds, seq, ceph_cap_string(dirty),
3688 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3689 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3691 if (list_empty(&to_remove) && !cleaned)
3694 ci->i_flushing_caps &= ~cleaned;
3696 spin_lock(&mdsc->cap_dirty_lock);
3698 list_for_each_entry(cf, &to_remove, i_list)
3699 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3701 if (ci->i_flushing_caps == 0) {
3702 if (list_empty(&ci->i_cap_flush_list)) {
3703 list_del_init(&ci->i_flushing_item);
3704 if (!list_empty(&session->s_cap_flushing)) {
3705 dout(" mds%d still flushing cap on %p\n",
3707 &list_first_entry(&session->s_cap_flushing,
3708 struct ceph_inode_info,
3709 i_flushing_item)->netfs.inode);
3712 mdsc->num_cap_flushing--;
3713 dout(" inode %p now !flushing\n", inode);
3715 if (ci->i_dirty_caps == 0) {
3716 dout(" inode %p now clean\n", inode);
3717 BUG_ON(!list_empty(&ci->i_dirty_item));
3719 if (ci->i_wr_ref == 0 &&
3720 ci->i_wrbuffer_ref_head == 0) {
3721 BUG_ON(!ci->i_head_snapc);
3722 ceph_put_snap_context(ci->i_head_snapc);
3723 ci->i_head_snapc = NULL;
3726 BUG_ON(list_empty(&ci->i_dirty_item));
3729 spin_unlock(&mdsc->cap_dirty_lock);
3732 spin_unlock(&ci->i_ceph_lock);
3734 while (!list_empty(&to_remove)) {
3735 cf = list_first_entry(&to_remove,
3736 struct ceph_cap_flush, i_list);
3737 list_del_init(&cf->i_list);
3738 if (!cf->is_capsnap)
3739 ceph_free_cap_flush(cf);
3743 wake_up_all(&ci->i_cap_wq);
3745 wake_up_all(&mdsc->cap_flushing_wq);
3750 void __ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3751 bool *wake_ci, bool *wake_mdsc)
3753 struct ceph_inode_info *ci = ceph_inode(inode);
3754 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3757 lockdep_assert_held(&ci->i_ceph_lock);
3759 dout("removing capsnap %p, inode %p ci %p\n", capsnap, inode, ci);
3761 list_del_init(&capsnap->ci_item);
3762 ret = __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3766 spin_lock(&mdsc->cap_dirty_lock);
3767 if (list_empty(&ci->i_cap_flush_list))
3768 list_del_init(&ci->i_flushing_item);
3770 ret = __detach_cap_flush_from_mdsc(mdsc, &capsnap->cap_flush);
3773 spin_unlock(&mdsc->cap_dirty_lock);
3776 void ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3777 bool *wake_ci, bool *wake_mdsc)
3779 struct ceph_inode_info *ci = ceph_inode(inode);
3781 lockdep_assert_held(&ci->i_ceph_lock);
3783 WARN_ON_ONCE(capsnap->dirty_pages || capsnap->writing);
3784 __ceph_remove_capsnap(inode, capsnap, wake_ci, wake_mdsc);
3788 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3789 * throw away our cap_snap.
3791 * Caller hold s_mutex.
3793 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3794 struct ceph_mds_caps *m,
3795 struct ceph_mds_session *session)
3797 struct ceph_inode_info *ci = ceph_inode(inode);
3798 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3799 u64 follows = le64_to_cpu(m->snap_follows);
3800 struct ceph_cap_snap *capsnap = NULL, *iter;
3801 bool wake_ci = false;
3802 bool wake_mdsc = false;
3804 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3805 inode, ci, session->s_mds, follows);
3807 spin_lock(&ci->i_ceph_lock);
3808 list_for_each_entry(iter, &ci->i_cap_snaps, ci_item) {
3809 if (iter->follows == follows) {
3810 if (iter->cap_flush.tid != flush_tid) {
3811 dout(" cap_snap %p follows %lld tid %lld !="
3812 " %lld\n", iter, follows,
3813 flush_tid, iter->cap_flush.tid);
3819 dout(" skipping cap_snap %p follows %lld\n",
3820 iter, iter->follows);
3824 ceph_remove_capsnap(inode, capsnap, &wake_ci, &wake_mdsc);
3825 spin_unlock(&ci->i_ceph_lock);
3828 ceph_put_snap_context(capsnap->context);
3829 ceph_put_cap_snap(capsnap);
3831 wake_up_all(&ci->i_cap_wq);
3833 wake_up_all(&mdsc->cap_flushing_wq);
3839 * Handle TRUNC from MDS, indicating file truncation.
3841 * caller hold s_mutex.
3843 static bool handle_cap_trunc(struct inode *inode,
3844 struct ceph_mds_caps *trunc,
3845 struct ceph_mds_session *session)
3847 struct ceph_inode_info *ci = ceph_inode(inode);
3848 int mds = session->s_mds;
3849 int seq = le32_to_cpu(trunc->seq);
3850 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3851 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3852 u64 size = le64_to_cpu(trunc->size);
3853 int implemented = 0;
3854 int dirty = __ceph_caps_dirty(ci);
3855 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3856 bool queue_trunc = false;
3858 lockdep_assert_held(&ci->i_ceph_lock);
3860 issued |= implemented | dirty;
3862 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3863 inode, mds, seq, truncate_size, truncate_seq);
3864 queue_trunc = ceph_fill_file_size(inode, issued,
3865 truncate_seq, truncate_size, size);
3870 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3871 * different one. If we are the most recent migration we've seen (as
3872 * indicated by mseq), make note of the migrating cap bits for the
3873 * duration (until we see the corresponding IMPORT).
3875 * caller holds s_mutex
3877 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3878 struct ceph_mds_cap_peer *ph,
3879 struct ceph_mds_session *session)
3881 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3882 struct ceph_mds_session *tsession = NULL;
3883 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3884 struct ceph_inode_info *ci = ceph_inode(inode);
3886 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3887 unsigned t_seq, t_mseq;
3889 int mds = session->s_mds;
3892 t_cap_id = le64_to_cpu(ph->cap_id);
3893 t_seq = le32_to_cpu(ph->seq);
3894 t_mseq = le32_to_cpu(ph->mseq);
3895 target = le32_to_cpu(ph->mds);
3897 t_cap_id = t_seq = t_mseq = 0;
3901 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3902 inode, ci, mds, mseq, target);
3904 down_read(&mdsc->snap_rwsem);
3905 spin_lock(&ci->i_ceph_lock);
3906 cap = __get_cap_for_mds(ci, mds);
3907 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3911 ceph_remove_cap(cap, false);
3916 * now we know we haven't received the cap import message yet
3917 * because the exported cap still exist.
3920 issued = cap->issued;
3921 if (issued != cap->implemented)
3922 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3923 "ino (%llx.%llx) mds%d seq %d mseq %d "
3924 "issued %s implemented %s\n",
3925 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3926 ceph_cap_string(issued),
3927 ceph_cap_string(cap->implemented));
3930 tcap = __get_cap_for_mds(ci, target);
3932 /* already have caps from the target */
3933 if (tcap->cap_id == t_cap_id &&
3934 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3935 dout(" updating import cap %p mds%d\n", tcap, target);
3936 tcap->cap_id = t_cap_id;
3937 tcap->seq = t_seq - 1;
3938 tcap->issue_seq = t_seq - 1;
3939 tcap->issued |= issued;
3940 tcap->implemented |= issued;
3941 if (cap == ci->i_auth_cap) {
3942 ci->i_auth_cap = tcap;
3943 change_auth_cap_ses(ci, tcap->session);
3946 ceph_remove_cap(cap, false);
3948 } else if (tsession) {
3949 /* add placeholder for the export tagert */
3950 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3952 ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3953 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3955 if (!list_empty(&ci->i_cap_flush_list) &&
3956 ci->i_auth_cap == tcap) {
3957 spin_lock(&mdsc->cap_dirty_lock);
3958 list_move_tail(&ci->i_flushing_item,
3959 &tcap->session->s_cap_flushing);
3960 spin_unlock(&mdsc->cap_dirty_lock);
3963 ceph_remove_cap(cap, false);
3967 spin_unlock(&ci->i_ceph_lock);
3968 up_read(&mdsc->snap_rwsem);
3969 mutex_unlock(&session->s_mutex);
3971 /* open target session */
3972 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3973 if (!IS_ERR(tsession)) {
3975 mutex_lock(&session->s_mutex);
3976 mutex_lock_nested(&tsession->s_mutex,
3977 SINGLE_DEPTH_NESTING);
3979 mutex_lock(&tsession->s_mutex);
3980 mutex_lock_nested(&session->s_mutex,
3981 SINGLE_DEPTH_NESTING);
3983 new_cap = ceph_get_cap(mdsc, NULL);
3988 mutex_lock(&session->s_mutex);
3993 spin_unlock(&ci->i_ceph_lock);
3994 up_read(&mdsc->snap_rwsem);
3995 mutex_unlock(&session->s_mutex);
3997 mutex_unlock(&tsession->s_mutex);
3998 ceph_put_mds_session(tsession);
4001 ceph_put_cap(mdsc, new_cap);
4005 * Handle cap IMPORT.
4007 * caller holds s_mutex. acquires i_ceph_lock
4009 static void handle_cap_import(struct ceph_mds_client *mdsc,
4010 struct inode *inode, struct ceph_mds_caps *im,
4011 struct ceph_mds_cap_peer *ph,
4012 struct ceph_mds_session *session,
4013 struct ceph_cap **target_cap, int *old_issued)
4015 struct ceph_inode_info *ci = ceph_inode(inode);
4016 struct ceph_cap *cap, *ocap, *new_cap = NULL;
4017 int mds = session->s_mds;
4019 unsigned caps = le32_to_cpu(im->caps);
4020 unsigned wanted = le32_to_cpu(im->wanted);
4021 unsigned seq = le32_to_cpu(im->seq);
4022 unsigned mseq = le32_to_cpu(im->migrate_seq);
4023 u64 realmino = le64_to_cpu(im->realm);
4024 u64 cap_id = le64_to_cpu(im->cap_id);
4029 p_cap_id = le64_to_cpu(ph->cap_id);
4030 peer = le32_to_cpu(ph->mds);
4036 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
4037 inode, ci, mds, mseq, peer);
4039 cap = __get_cap_for_mds(ci, mds);
4042 spin_unlock(&ci->i_ceph_lock);
4043 new_cap = ceph_get_cap(mdsc, NULL);
4044 spin_lock(&ci->i_ceph_lock);
4050 ceph_put_cap(mdsc, new_cap);
4055 __ceph_caps_issued(ci, &issued);
4056 issued |= __ceph_caps_dirty(ci);
4058 ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
4059 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
4061 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
4062 if (ocap && ocap->cap_id == p_cap_id) {
4063 dout(" remove export cap %p mds%d flags %d\n",
4064 ocap, peer, ph->flags);
4065 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
4066 (ocap->seq != le32_to_cpu(ph->seq) ||
4067 ocap->mseq != le32_to_cpu(ph->mseq))) {
4068 pr_err_ratelimited("handle_cap_import: "
4069 "mismatched seq/mseq: ino (%llx.%llx) "
4070 "mds%d seq %d mseq %d importer mds%d "
4071 "has peer seq %d mseq %d\n",
4072 ceph_vinop(inode), peer, ocap->seq,
4073 ocap->mseq, mds, le32_to_cpu(ph->seq),
4074 le32_to_cpu(ph->mseq));
4076 ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
4079 *old_issued = issued;
4084 * Handle a caps message from the MDS.
4086 * Identify the appropriate session, inode, and call the right handler
4087 * based on the cap op.
4089 void ceph_handle_caps(struct ceph_mds_session *session,
4090 struct ceph_msg *msg)
4092 struct ceph_mds_client *mdsc = session->s_mdsc;
4093 struct inode *inode;
4094 struct ceph_inode_info *ci;
4095 struct ceph_cap *cap;
4096 struct ceph_mds_caps *h;
4097 struct ceph_mds_cap_peer *peer = NULL;
4098 struct ceph_snap_realm *realm = NULL;
4100 int msg_version = le16_to_cpu(msg->hdr.version);
4102 struct ceph_vino vino;
4104 size_t snaptrace_len;
4106 struct cap_extra_info extra_info = {};
4109 dout("handle_caps from mds%d\n", session->s_mds);
4112 end = msg->front.iov_base + msg->front.iov_len;
4113 if (msg->front.iov_len < sizeof(*h))
4115 h = msg->front.iov_base;
4116 op = le32_to_cpu(h->op);
4117 vino.ino = le64_to_cpu(h->ino);
4118 vino.snap = CEPH_NOSNAP;
4119 seq = le32_to_cpu(h->seq);
4120 mseq = le32_to_cpu(h->migrate_seq);
4123 snaptrace_len = le32_to_cpu(h->snap_trace_len);
4124 p = snaptrace + snaptrace_len;
4126 if (msg_version >= 2) {
4128 ceph_decode_32_safe(&p, end, flock_len, bad);
4129 if (p + flock_len > end)
4134 if (msg_version >= 3) {
4135 if (op == CEPH_CAP_OP_IMPORT) {
4136 if (p + sizeof(*peer) > end)
4140 } else if (op == CEPH_CAP_OP_EXPORT) {
4141 /* recorded in unused fields */
4142 peer = (void *)&h->size;
4146 if (msg_version >= 4) {
4147 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
4148 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
4149 if (p + extra_info.inline_len > end)
4151 extra_info.inline_data = p;
4152 p += extra_info.inline_len;
4155 if (msg_version >= 5) {
4156 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
4159 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
4160 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
4163 if (msg_version >= 8) {
4167 ceph_decode_skip_64(&p, end, bad); // flush_tid
4169 ceph_decode_skip_32(&p, end, bad); // caller_uid
4170 ceph_decode_skip_32(&p, end, bad); // caller_gid
4172 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4173 if (pool_ns_len > 0) {
4174 ceph_decode_need(&p, end, pool_ns_len, bad);
4175 extra_info.pool_ns =
4176 ceph_find_or_create_string(p, pool_ns_len);
4181 if (msg_version >= 9) {
4182 struct ceph_timespec *btime;
4184 if (p + sizeof(*btime) > end)
4187 ceph_decode_timespec64(&extra_info.btime, btime);
4188 p += sizeof(*btime);
4189 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4192 if (msg_version >= 11) {
4194 ceph_decode_skip_32(&p, end, bad); // flags
4196 extra_info.dirstat_valid = true;
4197 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4198 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4202 inode = ceph_find_inode(mdsc->fsc->sb, vino);
4203 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4206 mutex_lock(&session->s_mutex);
4207 inc_session_sequence(session);
4208 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4212 dout(" i don't have ino %llx\n", vino.ino);
4214 if (op == CEPH_CAP_OP_IMPORT) {
4215 cap = ceph_get_cap(mdsc, NULL);
4216 cap->cap_ino = vino.ino;
4217 cap->queue_release = 1;
4218 cap->cap_id = le64_to_cpu(h->cap_id);
4221 cap->issue_seq = seq;
4222 spin_lock(&session->s_cap_lock);
4223 __ceph_queue_cap_release(session, cap);
4224 spin_unlock(&session->s_cap_lock);
4226 goto flush_cap_releases;
4228 ci = ceph_inode(inode);
4230 /* these will work even if we don't have a cap yet */
4232 case CEPH_CAP_OP_FLUSHSNAP_ACK:
4233 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4237 case CEPH_CAP_OP_EXPORT:
4238 handle_cap_export(inode, h, peer, session);
4241 case CEPH_CAP_OP_IMPORT:
4243 if (snaptrace_len) {
4244 down_write(&mdsc->snap_rwsem);
4245 ceph_update_snap_trace(mdsc, snaptrace,
4246 snaptrace + snaptrace_len,
4248 downgrade_write(&mdsc->snap_rwsem);
4250 down_read(&mdsc->snap_rwsem);
4252 spin_lock(&ci->i_ceph_lock);
4253 handle_cap_import(mdsc, inode, h, peer, session,
4254 &cap, &extra_info.issued);
4255 handle_cap_grant(inode, session, cap,
4256 h, msg->middle, &extra_info);
4258 ceph_put_snap_realm(mdsc, realm);
4262 /* the rest require a cap */
4263 spin_lock(&ci->i_ceph_lock);
4264 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4266 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4267 inode, ceph_ino(inode), ceph_snap(inode),
4269 spin_unlock(&ci->i_ceph_lock);
4270 goto flush_cap_releases;
4273 /* note that each of these drops i_ceph_lock for us */
4275 case CEPH_CAP_OP_REVOKE:
4276 case CEPH_CAP_OP_GRANT:
4277 __ceph_caps_issued(ci, &extra_info.issued);
4278 extra_info.issued |= __ceph_caps_dirty(ci);
4279 handle_cap_grant(inode, session, cap,
4280 h, msg->middle, &extra_info);
4283 case CEPH_CAP_OP_FLUSH_ACK:
4284 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4288 case CEPH_CAP_OP_TRUNC:
4289 queue_trunc = handle_cap_trunc(inode, h, session);
4290 spin_unlock(&ci->i_ceph_lock);
4292 ceph_queue_vmtruncate(inode);
4296 spin_unlock(&ci->i_ceph_lock);
4297 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4298 ceph_cap_op_name(op));
4302 mutex_unlock(&session->s_mutex);
4306 ceph_put_string(extra_info.pool_ns);
4311 * send any cap release message to try to move things
4312 * along for the mds (who clearly thinks we still have this
4315 ceph_flush_cap_releases(mdsc, session);
4319 pr_err("ceph_handle_caps: corrupt message\n");
4325 * Delayed work handler to process end of delayed cap release LRU list.
4327 * If new caps are added to the list while processing it, these won't get
4328 * processed in this run. In this case, the ci->i_hold_caps_max will be
4329 * returned so that the work can be scheduled accordingly.
4331 unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4333 struct inode *inode;
4334 struct ceph_inode_info *ci;
4335 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
4336 unsigned long delay_max = opt->caps_wanted_delay_max * HZ;
4337 unsigned long loop_start = jiffies;
4338 unsigned long delay = 0;
4340 dout("check_delayed_caps\n");
4341 spin_lock(&mdsc->cap_delay_lock);
4342 while (!list_empty(&mdsc->cap_delay_list)) {
4343 ci = list_first_entry(&mdsc->cap_delay_list,
4344 struct ceph_inode_info,
4346 if (time_before(loop_start, ci->i_hold_caps_max - delay_max)) {
4347 dout("%s caps added recently. Exiting loop", __func__);
4348 delay = ci->i_hold_caps_max;
4351 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4352 time_before(jiffies, ci->i_hold_caps_max))
4354 list_del_init(&ci->i_cap_delay_list);
4356 inode = igrab(&ci->netfs.inode);
4358 spin_unlock(&mdsc->cap_delay_lock);
4359 dout("check_delayed_caps on %p\n", inode);
4360 ceph_check_caps(ci, 0, NULL);
4362 spin_lock(&mdsc->cap_delay_lock);
4365 spin_unlock(&mdsc->cap_delay_lock);
4371 * Flush all dirty caps to the mds
4373 static void flush_dirty_session_caps(struct ceph_mds_session *s)
4375 struct ceph_mds_client *mdsc = s->s_mdsc;
4376 struct ceph_inode_info *ci;
4377 struct inode *inode;
4379 dout("flush_dirty_caps\n");
4380 spin_lock(&mdsc->cap_dirty_lock);
4381 while (!list_empty(&s->s_cap_dirty)) {
4382 ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4384 inode = &ci->netfs.inode;
4386 dout("flush_dirty_caps %llx.%llx\n", ceph_vinop(inode));
4387 spin_unlock(&mdsc->cap_dirty_lock);
4388 ceph_wait_on_async_create(inode);
4389 ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4391 spin_lock(&mdsc->cap_dirty_lock);
4393 spin_unlock(&mdsc->cap_dirty_lock);
4394 dout("flush_dirty_caps done\n");
4397 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4399 ceph_mdsc_iterate_sessions(mdsc, flush_dirty_session_caps, true);
4402 void __ceph_touch_fmode(struct ceph_inode_info *ci,
4403 struct ceph_mds_client *mdsc, int fmode)
4405 unsigned long now = jiffies;
4406 if (fmode & CEPH_FILE_MODE_RD)
4407 ci->i_last_rd = now;
4408 if (fmode & CEPH_FILE_MODE_WR)
4409 ci->i_last_wr = now;
4410 /* queue periodic check */
4412 __ceph_is_any_real_caps(ci) &&
4413 list_empty(&ci->i_cap_delay_list))
4414 __cap_delay_requeue(mdsc, ci);
4417 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4419 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->netfs.inode.i_sb);
4420 int bits = (fmode << 1) | 1;
4421 bool already_opened = false;
4425 atomic64_inc(&mdsc->metric.opened_files);
4427 spin_lock(&ci->i_ceph_lock);
4428 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4430 * If any of the mode ref is larger than 0,
4431 * that means it has been already opened by
4432 * others. Just skip checking the PIN ref.
4434 if (i && ci->i_nr_by_mode[i])
4435 already_opened = true;
4437 if (bits & (1 << i))
4438 ci->i_nr_by_mode[i] += count;
4441 if (!already_opened)
4442 percpu_counter_inc(&mdsc->metric.opened_inodes);
4443 spin_unlock(&ci->i_ceph_lock);
4447 * Drop open file reference. If we were the last open file,
4448 * we may need to release capabilities to the MDS (or schedule
4449 * their delayed release).
4451 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4453 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->netfs.inode.i_sb);
4454 int bits = (fmode << 1) | 1;
4455 bool is_closed = true;
4459 atomic64_dec(&mdsc->metric.opened_files);
4461 spin_lock(&ci->i_ceph_lock);
4462 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4463 if (bits & (1 << i)) {
4464 BUG_ON(ci->i_nr_by_mode[i] < count);
4465 ci->i_nr_by_mode[i] -= count;
4469 * If any of the mode ref is not 0 after
4470 * decreased, that means it is still opened
4471 * by others. Just skip checking the PIN ref.
4473 if (i && ci->i_nr_by_mode[i])
4478 percpu_counter_dec(&mdsc->metric.opened_inodes);
4479 spin_unlock(&ci->i_ceph_lock);
4483 * For a soon-to-be unlinked file, drop the LINK caps. If it
4484 * looks like the link count will hit 0, drop any other caps (other
4485 * than PIN) we don't specifically want (due to the file still being
4488 int ceph_drop_caps_for_unlink(struct inode *inode)
4490 struct ceph_inode_info *ci = ceph_inode(inode);
4491 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4493 spin_lock(&ci->i_ceph_lock);
4494 if (inode->i_nlink == 1) {
4495 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4497 if (__ceph_caps_dirty(ci)) {
4498 struct ceph_mds_client *mdsc =
4499 ceph_inode_to_client(inode)->mdsc;
4500 __cap_delay_requeue_front(mdsc, ci);
4503 spin_unlock(&ci->i_ceph_lock);
4508 * Helpers for embedding cap and dentry lease releases into mds
4511 * @force is used by dentry_release (below) to force inclusion of a
4512 * record for the directory inode, even when there aren't any caps to
4515 int ceph_encode_inode_release(void **p, struct inode *inode,
4516 int mds, int drop, int unless, int force)
4518 struct ceph_inode_info *ci = ceph_inode(inode);
4519 struct ceph_cap *cap;
4520 struct ceph_mds_request_release *rel = *p;
4524 spin_lock(&ci->i_ceph_lock);
4525 used = __ceph_caps_used(ci);
4526 dirty = __ceph_caps_dirty(ci);
4528 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4529 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4530 ceph_cap_string(unless));
4532 /* only drop unused, clean caps */
4533 drop &= ~(used | dirty);
4535 cap = __get_cap_for_mds(ci, mds);
4536 if (cap && __cap_is_valid(cap)) {
4537 unless &= cap->issued;
4539 if (unless & CEPH_CAP_AUTH_EXCL)
4540 drop &= ~CEPH_CAP_AUTH_SHARED;
4541 if (unless & CEPH_CAP_LINK_EXCL)
4542 drop &= ~CEPH_CAP_LINK_SHARED;
4543 if (unless & CEPH_CAP_XATTR_EXCL)
4544 drop &= ~CEPH_CAP_XATTR_SHARED;
4545 if (unless & CEPH_CAP_FILE_EXCL)
4546 drop &= ~CEPH_CAP_FILE_SHARED;
4549 if (force || (cap->issued & drop)) {
4550 if (cap->issued & drop) {
4551 int wanted = __ceph_caps_wanted(ci);
4552 dout("encode_inode_release %p cap %p "
4553 "%s -> %s, wanted %s -> %s\n", inode, cap,
4554 ceph_cap_string(cap->issued),
4555 ceph_cap_string(cap->issued & ~drop),
4556 ceph_cap_string(cap->mds_wanted),
4557 ceph_cap_string(wanted));
4559 cap->issued &= ~drop;
4560 cap->implemented &= ~drop;
4561 cap->mds_wanted = wanted;
4562 if (cap == ci->i_auth_cap &&
4563 !(wanted & CEPH_CAP_ANY_FILE_WR))
4564 ci->i_requested_max_size = 0;
4566 dout("encode_inode_release %p cap %p %s"
4567 " (force)\n", inode, cap,
4568 ceph_cap_string(cap->issued));
4571 rel->ino = cpu_to_le64(ceph_ino(inode));
4572 rel->cap_id = cpu_to_le64(cap->cap_id);
4573 rel->seq = cpu_to_le32(cap->seq);
4574 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4575 rel->mseq = cpu_to_le32(cap->mseq);
4576 rel->caps = cpu_to_le32(cap->implemented);
4577 rel->wanted = cpu_to_le32(cap->mds_wanted);
4583 dout("encode_inode_release %p cap %p %s (noop)\n",
4584 inode, cap, ceph_cap_string(cap->issued));
4587 spin_unlock(&ci->i_ceph_lock);
4591 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4593 int mds, int drop, int unless)
4595 struct dentry *parent = NULL;
4596 struct ceph_mds_request_release *rel = *p;
4597 struct ceph_dentry_info *di = ceph_dentry(dentry);
4602 * force an record for the directory caps if we have a dentry lease.
4603 * this is racy (can't take i_ceph_lock and d_lock together), but it
4604 * doesn't have to be perfect; the mds will revoke anything we don't
4607 spin_lock(&dentry->d_lock);
4608 if (di->lease_session && di->lease_session->s_mds == mds)
4611 parent = dget(dentry->d_parent);
4612 dir = d_inode(parent);
4614 spin_unlock(&dentry->d_lock);
4616 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4619 spin_lock(&dentry->d_lock);
4620 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4621 dout("encode_dentry_release %p mds%d seq %d\n",
4622 dentry, mds, (int)di->lease_seq);
4623 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4624 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4625 *p += dentry->d_name.len;
4626 rel->dname_seq = cpu_to_le32(di->lease_seq);
4627 __ceph_mdsc_drop_dentry_lease(dentry);
4629 spin_unlock(&dentry->d_lock);
4633 static int remove_capsnaps(struct ceph_mds_client *mdsc, struct inode *inode)
4635 struct ceph_inode_info *ci = ceph_inode(inode);
4636 struct ceph_cap_snap *capsnap;
4637 int capsnap_release = 0;
4639 lockdep_assert_held(&ci->i_ceph_lock);
4641 dout("removing capsnaps, ci is %p, inode is %p\n", ci, inode);
4643 while (!list_empty(&ci->i_cap_snaps)) {
4644 capsnap = list_first_entry(&ci->i_cap_snaps,
4645 struct ceph_cap_snap, ci_item);
4646 __ceph_remove_capsnap(inode, capsnap, NULL, NULL);
4647 ceph_put_snap_context(capsnap->context);
4648 ceph_put_cap_snap(capsnap);
4651 wake_up_all(&ci->i_cap_wq);
4652 wake_up_all(&mdsc->cap_flushing_wq);
4653 return capsnap_release;
4656 int ceph_purge_inode_cap(struct inode *inode, struct ceph_cap *cap, bool *invalidate)
4658 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
4659 struct ceph_mds_client *mdsc = fsc->mdsc;
4660 struct ceph_inode_info *ci = ceph_inode(inode);
4662 bool dirty_dropped = false;
4665 lockdep_assert_held(&ci->i_ceph_lock);
4667 dout("removing cap %p, ci is %p, inode is %p\n",
4668 cap, ci, &ci->netfs.inode);
4670 is_auth = (cap == ci->i_auth_cap);
4671 __ceph_remove_cap(cap, false);
4673 struct ceph_cap_flush *cf;
4675 if (ceph_inode_is_shutdown(inode)) {
4676 if (inode->i_data.nrpages > 0)
4678 if (ci->i_wrbuffer_ref > 0)
4679 mapping_set_error(&inode->i_data, -EIO);
4682 spin_lock(&mdsc->cap_dirty_lock);
4684 /* trash all of the cap flushes for this inode */
4685 while (!list_empty(&ci->i_cap_flush_list)) {
4686 cf = list_first_entry(&ci->i_cap_flush_list,
4687 struct ceph_cap_flush, i_list);
4688 list_del_init(&cf->g_list);
4689 list_del_init(&cf->i_list);
4690 if (!cf->is_capsnap)
4691 ceph_free_cap_flush(cf);
4694 if (!list_empty(&ci->i_dirty_item)) {
4695 pr_warn_ratelimited(
4696 " dropping dirty %s state for %p %lld\n",
4697 ceph_cap_string(ci->i_dirty_caps),
4698 inode, ceph_ino(inode));
4699 ci->i_dirty_caps = 0;
4700 list_del_init(&ci->i_dirty_item);
4701 dirty_dropped = true;
4703 if (!list_empty(&ci->i_flushing_item)) {
4704 pr_warn_ratelimited(
4705 " dropping dirty+flushing %s state for %p %lld\n",
4706 ceph_cap_string(ci->i_flushing_caps),
4707 inode, ceph_ino(inode));
4708 ci->i_flushing_caps = 0;
4709 list_del_init(&ci->i_flushing_item);
4710 mdsc->num_cap_flushing--;
4711 dirty_dropped = true;
4713 spin_unlock(&mdsc->cap_dirty_lock);
4715 if (dirty_dropped) {
4716 mapping_set_error(inode->i_mapping, -EIO);
4718 if (ci->i_wrbuffer_ref_head == 0 &&
4719 ci->i_wr_ref == 0 &&
4720 ci->i_dirty_caps == 0 &&
4721 ci->i_flushing_caps == 0) {
4722 ceph_put_snap_context(ci->i_head_snapc);
4723 ci->i_head_snapc = NULL;
4727 if (atomic_read(&ci->i_filelock_ref) > 0) {
4728 /* make further file lock syscall return -EIO */
4729 ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
4730 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
4731 inode, ceph_ino(inode));
4734 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
4735 cf = ci->i_prealloc_cap_flush;
4736 ci->i_prealloc_cap_flush = NULL;
4737 if (!cf->is_capsnap)
4738 ceph_free_cap_flush(cf);
4741 if (!list_empty(&ci->i_cap_snaps))
4742 iputs = remove_capsnaps(mdsc, inode);