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 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 = ceph_buffer_get(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);
1460 ceph_buffer_put(arg->xattr_buf);
1462 wake_up_all(&ci->i_cap_wq);
1465 static inline int __send_flush_snap(struct inode *inode,
1466 struct ceph_mds_session *session,
1467 struct ceph_cap_snap *capsnap,
1468 u32 mseq, u64 oldest_flush_tid)
1470 struct cap_msg_args arg;
1471 struct ceph_msg *msg;
1473 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1477 arg.session = session;
1478 arg.ino = ceph_vino(inode).ino;
1480 arg.follows = capsnap->follows;
1481 arg.flush_tid = capsnap->cap_flush.tid;
1482 arg.oldest_flush_tid = oldest_flush_tid;
1484 arg.size = capsnap->size;
1486 arg.xattr_version = capsnap->xattr_version;
1487 arg.xattr_buf = capsnap->xattr_blob;
1488 arg.old_xattr_buf = NULL;
1490 arg.atime = capsnap->atime;
1491 arg.mtime = capsnap->mtime;
1492 arg.ctime = capsnap->ctime;
1493 arg.btime = capsnap->btime;
1494 arg.change_attr = capsnap->change_attr;
1496 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1497 arg.caps = capsnap->issued;
1499 arg.dirty = capsnap->dirty;
1504 arg.time_warp_seq = capsnap->time_warp_seq;
1506 arg.uid = capsnap->uid;
1507 arg.gid = capsnap->gid;
1508 arg.mode = capsnap->mode;
1510 arg.inline_data = capsnap->inline_data;
1514 encode_cap_msg(msg, &arg);
1515 ceph_con_send(&arg.session->s_con, msg);
1520 * When a snapshot is taken, clients accumulate dirty metadata on
1521 * inodes with capabilities in ceph_cap_snaps to describe the file
1522 * state at the time the snapshot was taken. This must be flushed
1523 * asynchronously back to the MDS once sync writes complete and dirty
1524 * data is written out.
1526 * Called under i_ceph_lock.
1528 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1529 struct ceph_mds_session *session)
1530 __releases(ci->i_ceph_lock)
1531 __acquires(ci->i_ceph_lock)
1533 struct inode *inode = &ci->netfs.inode;
1534 struct ceph_mds_client *mdsc = session->s_mdsc;
1535 struct ceph_cap_snap *capsnap;
1536 u64 oldest_flush_tid = 0;
1537 u64 first_tid = 1, last_tid = 0;
1539 dout("__flush_snaps %p session %p\n", inode, session);
1541 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1543 * we need to wait for sync writes to complete and for dirty
1544 * pages to be written out.
1546 if (capsnap->dirty_pages || capsnap->writing)
1549 /* should be removed by ceph_try_drop_cap_snap() */
1550 BUG_ON(!capsnap->need_flush);
1552 /* only flush each capsnap once */
1553 if (capsnap->cap_flush.tid > 0) {
1554 dout(" already flushed %p, skipping\n", capsnap);
1558 spin_lock(&mdsc->cap_dirty_lock);
1559 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1560 list_add_tail(&capsnap->cap_flush.g_list,
1561 &mdsc->cap_flush_list);
1562 if (oldest_flush_tid == 0)
1563 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1564 if (list_empty(&ci->i_flushing_item)) {
1565 list_add_tail(&ci->i_flushing_item,
1566 &session->s_cap_flushing);
1568 spin_unlock(&mdsc->cap_dirty_lock);
1570 list_add_tail(&capsnap->cap_flush.i_list,
1571 &ci->i_cap_flush_list);
1574 first_tid = capsnap->cap_flush.tid;
1575 last_tid = capsnap->cap_flush.tid;
1578 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1580 while (first_tid <= last_tid) {
1581 struct ceph_cap *cap = ci->i_auth_cap;
1582 struct ceph_cap_flush *cf = NULL, *iter;
1585 if (!(cap && cap->session == session)) {
1586 dout("__flush_snaps %p auth cap %p not mds%d, "
1587 "stop\n", inode, cap, session->s_mds);
1592 list_for_each_entry(iter, &ci->i_cap_flush_list, i_list) {
1593 if (iter->tid >= first_tid) {
1602 first_tid = cf->tid + 1;
1604 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1605 refcount_inc(&capsnap->nref);
1606 spin_unlock(&ci->i_ceph_lock);
1608 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1609 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1611 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1614 pr_err("__flush_snaps: error sending cap flushsnap, "
1615 "ino (%llx.%llx) tid %llu follows %llu\n",
1616 ceph_vinop(inode), cf->tid, capsnap->follows);
1619 ceph_put_cap_snap(capsnap);
1620 spin_lock(&ci->i_ceph_lock);
1624 void ceph_flush_snaps(struct ceph_inode_info *ci,
1625 struct ceph_mds_session **psession)
1627 struct inode *inode = &ci->netfs.inode;
1628 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1629 struct ceph_mds_session *session = NULL;
1630 bool need_put = false;
1633 dout("ceph_flush_snaps %p\n", inode);
1635 session = *psession;
1637 spin_lock(&ci->i_ceph_lock);
1638 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1639 dout(" no capsnap needs flush, doing nothing\n");
1642 if (!ci->i_auth_cap) {
1643 dout(" no auth cap (migrating?), doing nothing\n");
1647 mds = ci->i_auth_cap->session->s_mds;
1648 if (session && session->s_mds != mds) {
1649 dout(" oops, wrong session %p mutex\n", session);
1650 ceph_put_mds_session(session);
1654 spin_unlock(&ci->i_ceph_lock);
1655 mutex_lock(&mdsc->mutex);
1656 session = __ceph_lookup_mds_session(mdsc, mds);
1657 mutex_unlock(&mdsc->mutex);
1661 // make sure flushsnap messages are sent in proper order.
1662 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1663 __kick_flushing_caps(mdsc, session, ci, 0);
1665 __ceph_flush_snaps(ci, session);
1667 spin_unlock(&ci->i_ceph_lock);
1670 *psession = session;
1672 ceph_put_mds_session(session);
1673 /* we flushed them all; remove this inode from the queue */
1674 spin_lock(&mdsc->snap_flush_lock);
1675 if (!list_empty(&ci->i_snap_flush_item))
1677 list_del_init(&ci->i_snap_flush_item);
1678 spin_unlock(&mdsc->snap_flush_lock);
1685 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1686 * Caller is then responsible for calling __mark_inode_dirty with the
1687 * returned flags value.
1689 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1690 struct ceph_cap_flush **pcf)
1692 struct ceph_mds_client *mdsc =
1693 ceph_sb_to_client(ci->netfs.inode.i_sb)->mdsc;
1694 struct inode *inode = &ci->netfs.inode;
1695 int was = ci->i_dirty_caps;
1698 lockdep_assert_held(&ci->i_ceph_lock);
1700 if (!ci->i_auth_cap) {
1701 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1702 "but no auth cap (session was closed?)\n",
1703 inode, ceph_ino(inode), ceph_cap_string(mask));
1707 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->netfs.inode,
1708 ceph_cap_string(mask), ceph_cap_string(was),
1709 ceph_cap_string(was | mask));
1710 ci->i_dirty_caps |= mask;
1712 struct ceph_mds_session *session = ci->i_auth_cap->session;
1714 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1715 swap(ci->i_prealloc_cap_flush, *pcf);
1717 if (!ci->i_head_snapc) {
1718 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1719 ci->i_head_snapc = ceph_get_snap_context(
1720 ci->i_snap_realm->cached_context);
1722 dout(" inode %p now dirty snapc %p auth cap %p\n",
1723 &ci->netfs.inode, ci->i_head_snapc, ci->i_auth_cap);
1724 BUG_ON(!list_empty(&ci->i_dirty_item));
1725 spin_lock(&mdsc->cap_dirty_lock);
1726 list_add(&ci->i_dirty_item, &session->s_cap_dirty);
1727 spin_unlock(&mdsc->cap_dirty_lock);
1728 if (ci->i_flushing_caps == 0) {
1730 dirty |= I_DIRTY_SYNC;
1733 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1735 BUG_ON(list_empty(&ci->i_dirty_item));
1736 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1737 (mask & CEPH_CAP_FILE_BUFFER))
1738 dirty |= I_DIRTY_DATASYNC;
1739 __cap_delay_requeue(mdsc, ci);
1743 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1745 struct ceph_cap_flush *cf;
1747 cf = kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1751 cf->is_capsnap = false;
1755 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1758 kmem_cache_free(ceph_cap_flush_cachep, cf);
1761 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1763 if (!list_empty(&mdsc->cap_flush_list)) {
1764 struct ceph_cap_flush *cf =
1765 list_first_entry(&mdsc->cap_flush_list,
1766 struct ceph_cap_flush, g_list);
1773 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1774 * Return true if caller needs to wake up flush waiters.
1776 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1777 struct ceph_cap_flush *cf)
1779 struct ceph_cap_flush *prev;
1780 bool wake = cf->wake;
1782 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1783 prev = list_prev_entry(cf, g_list);
1787 list_del_init(&cf->g_list);
1791 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1792 struct ceph_cap_flush *cf)
1794 struct ceph_cap_flush *prev;
1795 bool wake = cf->wake;
1797 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1798 prev = list_prev_entry(cf, i_list);
1802 list_del_init(&cf->i_list);
1807 * Add dirty inode to the flushing list. Assigned a seq number so we
1808 * can wait for caps to flush without starving.
1810 * Called under i_ceph_lock. Returns the flush tid.
1812 static u64 __mark_caps_flushing(struct inode *inode,
1813 struct ceph_mds_session *session, bool wake,
1814 u64 *oldest_flush_tid)
1816 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1817 struct ceph_inode_info *ci = ceph_inode(inode);
1818 struct ceph_cap_flush *cf = NULL;
1821 lockdep_assert_held(&ci->i_ceph_lock);
1822 BUG_ON(ci->i_dirty_caps == 0);
1823 BUG_ON(list_empty(&ci->i_dirty_item));
1824 BUG_ON(!ci->i_prealloc_cap_flush);
1826 flushing = ci->i_dirty_caps;
1827 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1828 ceph_cap_string(flushing),
1829 ceph_cap_string(ci->i_flushing_caps),
1830 ceph_cap_string(ci->i_flushing_caps | flushing));
1831 ci->i_flushing_caps |= flushing;
1832 ci->i_dirty_caps = 0;
1833 dout(" inode %p now !dirty\n", inode);
1835 swap(cf, ci->i_prealloc_cap_flush);
1836 cf->caps = flushing;
1839 spin_lock(&mdsc->cap_dirty_lock);
1840 list_del_init(&ci->i_dirty_item);
1842 cf->tid = ++mdsc->last_cap_flush_tid;
1843 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1844 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1846 if (list_empty(&ci->i_flushing_item)) {
1847 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1848 mdsc->num_cap_flushing++;
1850 spin_unlock(&mdsc->cap_dirty_lock);
1852 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1858 * try to invalidate mapping pages without blocking.
1860 static int try_nonblocking_invalidate(struct inode *inode)
1861 __releases(ci->i_ceph_lock)
1862 __acquires(ci->i_ceph_lock)
1864 struct ceph_inode_info *ci = ceph_inode(inode);
1865 u32 invalidating_gen = ci->i_rdcache_gen;
1867 spin_unlock(&ci->i_ceph_lock);
1868 ceph_fscache_invalidate(inode, false);
1869 invalidate_mapping_pages(&inode->i_data, 0, -1);
1870 spin_lock(&ci->i_ceph_lock);
1872 if (inode->i_data.nrpages == 0 &&
1873 invalidating_gen == ci->i_rdcache_gen) {
1875 dout("try_nonblocking_invalidate %p success\n", inode);
1876 /* save any racing async invalidate some trouble */
1877 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1880 dout("try_nonblocking_invalidate %p failed\n", inode);
1884 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1886 loff_t size = i_size_read(&ci->netfs.inode);
1887 /* mds will adjust max size according to the reported size */
1888 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1890 if (size >= ci->i_max_size)
1892 /* half of previous max_size increment has been used */
1893 if (ci->i_max_size > ci->i_reported_size &&
1894 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1900 * Swiss army knife function to examine currently used and wanted
1901 * versus held caps. Release, flush, ack revoked caps to mds as
1904 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1905 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1908 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1909 struct ceph_mds_session *session)
1911 struct inode *inode = &ci->netfs.inode;
1912 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1913 struct ceph_cap *cap;
1914 u64 flush_tid, oldest_flush_tid;
1915 int file_wanted, used, cap_used;
1916 int issued, implemented, want, retain, revoking, flushing = 0;
1917 int mds = -1; /* keep track of how far we've gone through i_caps list
1918 to avoid an infinite loop on retry */
1920 bool queue_invalidate = false;
1921 bool tried_invalidate = false;
1922 bool queue_writeback = false;
1925 ceph_get_mds_session(session);
1927 spin_lock(&ci->i_ceph_lock);
1928 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
1929 /* Don't send messages until we get async create reply */
1930 spin_unlock(&ci->i_ceph_lock);
1931 ceph_put_mds_session(session);
1935 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1936 flags |= CHECK_CAPS_FLUSH;
1938 /* Caps wanted by virtue of active open files. */
1939 file_wanted = __ceph_caps_file_wanted(ci);
1941 /* Caps which have active references against them */
1942 used = __ceph_caps_used(ci);
1945 * "issued" represents the current caps that the MDS wants us to have.
1946 * "implemented" is the set that we have been granted, and includes the
1947 * ones that have not yet been returned to the MDS (the "revoking" set,
1948 * usually because they have outstanding references).
1950 issued = __ceph_caps_issued(ci, &implemented);
1951 revoking = implemented & ~issued;
1955 /* The ones we currently want to retain (may be adjusted below) */
1956 retain = file_wanted | used | CEPH_CAP_PIN;
1957 if (!mdsc->stopping && inode->i_nlink > 0) {
1959 retain |= CEPH_CAP_ANY; /* be greedy */
1960 } else if (S_ISDIR(inode->i_mode) &&
1961 (issued & CEPH_CAP_FILE_SHARED) &&
1962 __ceph_dir_is_complete(ci)) {
1964 * If a directory is complete, we want to keep
1965 * the exclusive cap. So that MDS does not end up
1966 * revoking the shared cap on every create/unlink
1969 if (IS_RDONLY(inode)) {
1970 want = CEPH_CAP_ANY_SHARED;
1972 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1977 retain |= CEPH_CAP_ANY_SHARED;
1979 * keep RD only if we didn't have the file open RW,
1980 * because then the mds would revoke it anyway to
1981 * journal max_size=0.
1983 if (ci->i_max_size == 0)
1984 retain |= CEPH_CAP_ANY_RD;
1988 dout("check_caps %llx.%llx file_want %s used %s dirty %s flushing %s"
1989 " issued %s revoking %s retain %s %s%s%s\n", ceph_vinop(inode),
1990 ceph_cap_string(file_wanted),
1991 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1992 ceph_cap_string(ci->i_flushing_caps),
1993 ceph_cap_string(issued), ceph_cap_string(revoking),
1994 ceph_cap_string(retain),
1995 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1996 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "",
1997 (flags & CHECK_CAPS_NOINVAL) ? " NOINVAL" : "");
2000 * If we no longer need to hold onto old our caps, and we may
2001 * have cached pages, but don't want them, then try to invalidate.
2002 * If we fail, it's because pages are locked.... try again later.
2004 if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
2005 S_ISREG(inode->i_mode) &&
2006 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
2007 inode->i_data.nrpages && /* have cached pages */
2008 (revoking & (CEPH_CAP_FILE_CACHE|
2009 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
2010 !tried_invalidate) {
2011 dout("check_caps trying to invalidate on %llx.%llx\n",
2013 if (try_nonblocking_invalidate(inode) < 0) {
2014 dout("check_caps queuing invalidate\n");
2015 queue_invalidate = true;
2016 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2018 tried_invalidate = true;
2022 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2024 struct cap_msg_args arg;
2026 cap = rb_entry(p, struct ceph_cap, ci_node);
2028 /* avoid looping forever */
2029 if (mds >= cap->mds ||
2030 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
2034 * If we have an auth cap, we don't need to consider any
2035 * overlapping caps as used.
2038 if (ci->i_auth_cap && cap != ci->i_auth_cap)
2039 cap_used &= ~ci->i_auth_cap->issued;
2041 revoking = cap->implemented & ~cap->issued;
2042 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
2043 cap->mds, cap, ceph_cap_string(cap_used),
2044 ceph_cap_string(cap->issued),
2045 ceph_cap_string(cap->implemented),
2046 ceph_cap_string(revoking));
2048 if (cap == ci->i_auth_cap &&
2049 (cap->issued & CEPH_CAP_FILE_WR)) {
2050 /* request larger max_size from MDS? */
2051 if (ci->i_wanted_max_size > ci->i_max_size &&
2052 ci->i_wanted_max_size > ci->i_requested_max_size) {
2053 dout("requesting new max_size\n");
2057 /* approaching file_max? */
2058 if (__ceph_should_report_size(ci)) {
2059 dout("i_size approaching max_size\n");
2063 /* flush anything dirty? */
2064 if (cap == ci->i_auth_cap) {
2065 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
2066 dout("flushing dirty caps\n");
2069 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
2070 dout("flushing snap caps\n");
2075 /* completed revocation? going down and there are no caps? */
2077 if ((revoking & cap_used) == 0) {
2078 dout("completed revocation of %s\n",
2079 ceph_cap_string(cap->implemented & ~cap->issued));
2084 * If the "i_wrbuffer_ref" was increased by mmap or generic
2085 * cache write just before the ceph_check_caps() is called,
2086 * the Fb capability revoking will fail this time. Then we
2087 * must wait for the BDI's delayed work to flush the dirty
2088 * pages and to release the "i_wrbuffer_ref", which will cost
2089 * at most 5 seconds. That means the MDS needs to wait at
2090 * most 5 seconds to finished the Fb capability's revocation.
2092 * Let's queue a writeback for it.
2094 if (S_ISREG(inode->i_mode) && ci->i_wrbuffer_ref &&
2095 (revoking & CEPH_CAP_FILE_BUFFER))
2096 queue_writeback = true;
2099 /* want more caps from mds? */
2100 if (want & ~cap->mds_wanted) {
2101 if (want & ~(cap->mds_wanted | cap->issued))
2103 if (!__cap_is_valid(cap))
2107 /* things we might delay */
2108 if ((cap->issued & ~retain) == 0)
2109 continue; /* nope, all good */
2112 ceph_put_mds_session(session);
2113 session = ceph_get_mds_session(cap->session);
2115 /* kick flushing and flush snaps before sending normal
2117 if (cap == ci->i_auth_cap &&
2119 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2120 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2121 __kick_flushing_caps(mdsc, session, ci, 0);
2122 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2123 __ceph_flush_snaps(ci, session);
2128 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2129 flushing = ci->i_dirty_caps;
2130 flush_tid = __mark_caps_flushing(inode, session, false,
2132 if (flags & CHECK_CAPS_FLUSH &&
2133 list_empty(&session->s_cap_dirty))
2134 mflags |= CEPH_CLIENT_CAPS_SYNC;
2138 spin_lock(&mdsc->cap_dirty_lock);
2139 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2140 spin_unlock(&mdsc->cap_dirty_lock);
2143 mds = cap->mds; /* remember mds, so we don't repeat */
2145 __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2146 want, retain, flushing, flush_tid, oldest_flush_tid);
2148 spin_unlock(&ci->i_ceph_lock);
2149 __send_cap(&arg, ci);
2150 spin_lock(&ci->i_ceph_lock);
2152 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2155 /* periodically re-calculate caps wanted by open files */
2156 if (__ceph_is_any_real_caps(ci) &&
2157 list_empty(&ci->i_cap_delay_list) &&
2158 (file_wanted & ~CEPH_CAP_PIN) &&
2159 !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2160 __cap_delay_requeue(mdsc, ci);
2163 spin_unlock(&ci->i_ceph_lock);
2165 ceph_put_mds_session(session);
2166 if (queue_writeback)
2167 ceph_queue_writeback(inode);
2168 if (queue_invalidate)
2169 ceph_queue_invalidate(inode);
2173 * Try to flush dirty caps back to the auth mds.
2175 static int try_flush_caps(struct inode *inode, u64 *ptid)
2177 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2178 struct ceph_inode_info *ci = ceph_inode(inode);
2180 u64 flush_tid = 0, oldest_flush_tid = 0;
2182 spin_lock(&ci->i_ceph_lock);
2184 if (ci->i_dirty_caps && ci->i_auth_cap) {
2185 struct ceph_cap *cap = ci->i_auth_cap;
2186 struct cap_msg_args arg;
2187 struct ceph_mds_session *session = cap->session;
2189 if (session->s_state < CEPH_MDS_SESSION_OPEN) {
2190 spin_unlock(&ci->i_ceph_lock);
2194 if (ci->i_ceph_flags &
2195 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2196 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2197 __kick_flushing_caps(mdsc, session, ci, 0);
2198 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2199 __ceph_flush_snaps(ci, session);
2203 flushing = ci->i_dirty_caps;
2204 flush_tid = __mark_caps_flushing(inode, session, true,
2207 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2208 __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2209 (cap->issued | cap->implemented),
2210 flushing, flush_tid, oldest_flush_tid);
2211 spin_unlock(&ci->i_ceph_lock);
2213 __send_cap(&arg, ci);
2215 if (!list_empty(&ci->i_cap_flush_list)) {
2216 struct ceph_cap_flush *cf =
2217 list_last_entry(&ci->i_cap_flush_list,
2218 struct ceph_cap_flush, i_list);
2220 flush_tid = cf->tid;
2222 flushing = ci->i_flushing_caps;
2223 spin_unlock(&ci->i_ceph_lock);
2231 * Return true if we've flushed caps through the given flush_tid.
2233 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2235 struct ceph_inode_info *ci = ceph_inode(inode);
2238 spin_lock(&ci->i_ceph_lock);
2239 if (!list_empty(&ci->i_cap_flush_list)) {
2240 struct ceph_cap_flush * cf =
2241 list_first_entry(&ci->i_cap_flush_list,
2242 struct ceph_cap_flush, i_list);
2243 if (cf->tid <= flush_tid)
2246 spin_unlock(&ci->i_ceph_lock);
2251 * flush the mdlog and wait for any unsafe requests to complete.
2253 static int flush_mdlog_and_wait_inode_unsafe_requests(struct inode *inode)
2255 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2256 struct ceph_inode_info *ci = ceph_inode(inode);
2257 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2260 spin_lock(&ci->i_unsafe_lock);
2261 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2262 req1 = list_last_entry(&ci->i_unsafe_dirops,
2263 struct ceph_mds_request,
2265 ceph_mdsc_get_request(req1);
2267 if (!list_empty(&ci->i_unsafe_iops)) {
2268 req2 = list_last_entry(&ci->i_unsafe_iops,
2269 struct ceph_mds_request,
2270 r_unsafe_target_item);
2271 ceph_mdsc_get_request(req2);
2273 spin_unlock(&ci->i_unsafe_lock);
2276 * Trigger to flush the journal logs in all the relevant MDSes
2277 * manually, or in the worst case we must wait at most 5 seconds
2278 * to wait the journal logs to be flushed by the MDSes periodically.
2281 struct ceph_mds_request *req;
2282 struct ceph_mds_session **sessions;
2283 struct ceph_mds_session *s;
2284 unsigned int max_sessions;
2287 mutex_lock(&mdsc->mutex);
2288 max_sessions = mdsc->max_sessions;
2290 sessions = kcalloc(max_sessions, sizeof(s), GFP_KERNEL);
2292 mutex_unlock(&mdsc->mutex);
2297 spin_lock(&ci->i_unsafe_lock);
2299 list_for_each_entry(req, &ci->i_unsafe_dirops,
2300 r_unsafe_dir_item) {
2304 if (!sessions[s->s_mds]) {
2305 s = ceph_get_mds_session(s);
2306 sessions[s->s_mds] = s;
2311 list_for_each_entry(req, &ci->i_unsafe_iops,
2312 r_unsafe_target_item) {
2316 if (!sessions[s->s_mds]) {
2317 s = ceph_get_mds_session(s);
2318 sessions[s->s_mds] = s;
2322 spin_unlock(&ci->i_unsafe_lock);
2325 spin_lock(&ci->i_ceph_lock);
2326 if (ci->i_auth_cap) {
2327 s = ci->i_auth_cap->session;
2328 if (!sessions[s->s_mds])
2329 sessions[s->s_mds] = ceph_get_mds_session(s);
2331 spin_unlock(&ci->i_ceph_lock);
2332 mutex_unlock(&mdsc->mutex);
2334 /* send flush mdlog request to MDSes */
2335 for (i = 0; i < max_sessions; i++) {
2338 send_flush_mdlog(s);
2339 ceph_put_mds_session(s);
2345 dout("%s %p wait on tid %llu %llu\n", __func__,
2346 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2348 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2349 ceph_timeout_jiffies(req1->r_timeout));
2354 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2355 ceph_timeout_jiffies(req2->r_timeout));
2362 ceph_mdsc_put_request(req1);
2364 ceph_mdsc_put_request(req2);
2368 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2370 struct inode *inode = file->f_mapping->host;
2371 struct ceph_inode_info *ci = ceph_inode(inode);
2376 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2378 ret = file_write_and_wait_range(file, start, end);
2382 ret = ceph_wait_on_async_create(inode);
2386 dirty = try_flush_caps(inode, &flush_tid);
2387 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2389 err = flush_mdlog_and_wait_inode_unsafe_requests(inode);
2392 * only wait on non-file metadata writeback (the mds
2393 * can recover size and mtime, so we don't need to
2396 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2397 err = wait_event_interruptible(ci->i_cap_wq,
2398 caps_are_flushed(inode, flush_tid));
2404 err = file_check_and_advance_wb_err(file);
2408 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2413 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2414 * queue inode for flush but don't do so immediately, because we can
2415 * get by with fewer MDS messages if we wait for data writeback to
2418 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2420 struct ceph_inode_info *ci = ceph_inode(inode);
2424 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2426 dout("write_inode %p wait=%d\n", inode, wait);
2427 ceph_fscache_unpin_writeback(inode, wbc);
2429 err = ceph_wait_on_async_create(inode);
2432 dirty = try_flush_caps(inode, &flush_tid);
2434 err = wait_event_interruptible(ci->i_cap_wq,
2435 caps_are_flushed(inode, flush_tid));
2437 struct ceph_mds_client *mdsc =
2438 ceph_sb_to_client(inode->i_sb)->mdsc;
2440 spin_lock(&ci->i_ceph_lock);
2441 if (__ceph_caps_dirty(ci))
2442 __cap_delay_requeue_front(mdsc, ci);
2443 spin_unlock(&ci->i_ceph_lock);
2448 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2449 struct ceph_mds_session *session,
2450 struct ceph_inode_info *ci,
2451 u64 oldest_flush_tid)
2452 __releases(ci->i_ceph_lock)
2453 __acquires(ci->i_ceph_lock)
2455 struct inode *inode = &ci->netfs.inode;
2456 struct ceph_cap *cap;
2457 struct ceph_cap_flush *cf;
2460 u64 last_snap_flush = 0;
2462 /* Don't do anything until create reply comes in */
2463 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE)
2466 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2468 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2469 if (cf->is_capsnap) {
2470 last_snap_flush = cf->tid;
2475 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2476 if (cf->tid < first_tid)
2479 cap = ci->i_auth_cap;
2480 if (!(cap && cap->session == session)) {
2481 pr_err("%p auth cap %p not mds%d ???\n",
2482 inode, cap, session->s_mds);
2486 first_tid = cf->tid + 1;
2488 if (!cf->is_capsnap) {
2489 struct cap_msg_args arg;
2491 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2492 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2493 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2494 (cf->tid < last_snap_flush ?
2495 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2496 __ceph_caps_used(ci),
2497 __ceph_caps_wanted(ci),
2498 (cap->issued | cap->implemented),
2499 cf->caps, cf->tid, oldest_flush_tid);
2500 spin_unlock(&ci->i_ceph_lock);
2501 __send_cap(&arg, ci);
2503 struct ceph_cap_snap *capsnap =
2504 container_of(cf, struct ceph_cap_snap,
2506 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2507 inode, capsnap, cf->tid,
2508 ceph_cap_string(capsnap->dirty));
2510 refcount_inc(&capsnap->nref);
2511 spin_unlock(&ci->i_ceph_lock);
2513 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2516 pr_err("kick_flushing_caps: error sending "
2517 "cap flushsnap, ino (%llx.%llx) "
2518 "tid %llu follows %llu\n",
2519 ceph_vinop(inode), cf->tid,
2523 ceph_put_cap_snap(capsnap);
2526 spin_lock(&ci->i_ceph_lock);
2530 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2531 struct ceph_mds_session *session)
2533 struct ceph_inode_info *ci;
2534 struct ceph_cap *cap;
2535 u64 oldest_flush_tid;
2537 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2539 spin_lock(&mdsc->cap_dirty_lock);
2540 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2541 spin_unlock(&mdsc->cap_dirty_lock);
2543 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2544 spin_lock(&ci->i_ceph_lock);
2545 cap = ci->i_auth_cap;
2546 if (!(cap && cap->session == session)) {
2547 pr_err("%p auth cap %p not mds%d ???\n",
2548 &ci->netfs.inode, cap, session->s_mds);
2549 spin_unlock(&ci->i_ceph_lock);
2555 * if flushing caps were revoked, we re-send the cap flush
2556 * in client reconnect stage. This guarantees MDS * processes
2557 * the cap flush message before issuing the flushing caps to
2560 if ((cap->issued & ci->i_flushing_caps) !=
2561 ci->i_flushing_caps) {
2562 /* encode_caps_cb() also will reset these sequence
2563 * numbers. make sure sequence numbers in cap flush
2564 * message match later reconnect message */
2568 __kick_flushing_caps(mdsc, session, ci,
2571 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2574 spin_unlock(&ci->i_ceph_lock);
2578 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2579 struct ceph_mds_session *session)
2581 struct ceph_inode_info *ci;
2582 struct ceph_cap *cap;
2583 u64 oldest_flush_tid;
2585 lockdep_assert_held(&session->s_mutex);
2587 dout("kick_flushing_caps mds%d\n", session->s_mds);
2589 spin_lock(&mdsc->cap_dirty_lock);
2590 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2591 spin_unlock(&mdsc->cap_dirty_lock);
2593 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2594 spin_lock(&ci->i_ceph_lock);
2595 cap = ci->i_auth_cap;
2596 if (!(cap && cap->session == session)) {
2597 pr_err("%p auth cap %p not mds%d ???\n",
2598 &ci->netfs.inode, cap, session->s_mds);
2599 spin_unlock(&ci->i_ceph_lock);
2602 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2603 __kick_flushing_caps(mdsc, session, ci,
2606 spin_unlock(&ci->i_ceph_lock);
2610 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2611 struct ceph_inode_info *ci)
2613 struct ceph_mds_client *mdsc = session->s_mdsc;
2614 struct ceph_cap *cap = ci->i_auth_cap;
2616 lockdep_assert_held(&ci->i_ceph_lock);
2618 dout("%s %p flushing %s\n", __func__, &ci->netfs.inode,
2619 ceph_cap_string(ci->i_flushing_caps));
2621 if (!list_empty(&ci->i_cap_flush_list)) {
2622 u64 oldest_flush_tid;
2623 spin_lock(&mdsc->cap_dirty_lock);
2624 list_move_tail(&ci->i_flushing_item,
2625 &cap->session->s_cap_flushing);
2626 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2627 spin_unlock(&mdsc->cap_dirty_lock);
2629 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2635 * Take references to capabilities we hold, so that we don't release
2636 * them to the MDS prematurely.
2638 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2639 bool snap_rwsem_locked)
2641 lockdep_assert_held(&ci->i_ceph_lock);
2643 if (got & CEPH_CAP_PIN)
2645 if (got & CEPH_CAP_FILE_RD)
2647 if (got & CEPH_CAP_FILE_CACHE)
2648 ci->i_rdcache_ref++;
2649 if (got & CEPH_CAP_FILE_EXCL)
2651 if (got & CEPH_CAP_FILE_WR) {
2652 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2653 BUG_ON(!snap_rwsem_locked);
2654 ci->i_head_snapc = ceph_get_snap_context(
2655 ci->i_snap_realm->cached_context);
2659 if (got & CEPH_CAP_FILE_BUFFER) {
2660 if (ci->i_wb_ref == 0)
2661 ihold(&ci->netfs.inode);
2663 dout("%s %p wb %d -> %d (?)\n", __func__,
2664 &ci->netfs.inode, ci->i_wb_ref-1, ci->i_wb_ref);
2669 * Try to grab cap references. Specify those refs we @want, and the
2670 * minimal set we @need. Also include the larger offset we are writing
2671 * to (when applicable), and check against max_size here as well.
2672 * Note that caller is responsible for ensuring max_size increases are
2673 * requested from the MDS.
2675 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2676 * or a negative error code. There are 3 speical error codes:
2677 * -EAGAIN: need to sleep but non-blocking is specified
2678 * -EFBIG: ask caller to call check_max_size() and try again.
2679 * -EUCLEAN: ask caller to call ceph_renew_caps() and try again.
2682 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2683 NON_BLOCKING = (1 << 8),
2684 CHECK_FILELOCK = (1 << 9),
2687 static int try_get_cap_refs(struct inode *inode, int need, int want,
2688 loff_t endoff, int flags, int *got)
2690 struct ceph_inode_info *ci = ceph_inode(inode);
2691 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2693 int have, implemented;
2694 bool snap_rwsem_locked = false;
2696 dout("get_cap_refs %p need %s want %s\n", inode,
2697 ceph_cap_string(need), ceph_cap_string(want));
2700 spin_lock(&ci->i_ceph_lock);
2702 if ((flags & CHECK_FILELOCK) &&
2703 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2704 dout("try_get_cap_refs %p error filelock\n", inode);
2709 /* finish pending truncate */
2710 while (ci->i_truncate_pending) {
2711 spin_unlock(&ci->i_ceph_lock);
2712 if (snap_rwsem_locked) {
2713 up_read(&mdsc->snap_rwsem);
2714 snap_rwsem_locked = false;
2716 __ceph_do_pending_vmtruncate(inode);
2717 spin_lock(&ci->i_ceph_lock);
2720 have = __ceph_caps_issued(ci, &implemented);
2722 if (have & need & CEPH_CAP_FILE_WR) {
2723 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2724 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2725 inode, endoff, ci->i_max_size);
2726 if (endoff > ci->i_requested_max_size)
2727 ret = ci->i_auth_cap ? -EFBIG : -EUCLEAN;
2731 * If a sync write is in progress, we must wait, so that we
2732 * can get a final snapshot value for size+mtime.
2734 if (__ceph_have_pending_cap_snap(ci)) {
2735 dout("get_cap_refs %p cap_snap_pending\n", inode);
2740 if ((have & need) == need) {
2742 * Look at (implemented & ~have & not) so that we keep waiting
2743 * on transition from wanted -> needed caps. This is needed
2744 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2745 * going before a prior buffered writeback happens.
2747 * For RDCACHE|RD -> RD, there is not need to wait and we can
2748 * just exclude the revoking caps and force to sync read.
2750 int not = want & ~(have & need);
2751 int revoking = implemented & ~have;
2752 int exclude = revoking & not;
2753 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2754 inode, ceph_cap_string(have), ceph_cap_string(not),
2755 ceph_cap_string(revoking));
2756 if (!exclude || !(exclude & CEPH_CAP_FILE_BUFFER)) {
2757 if (!snap_rwsem_locked &&
2758 !ci->i_head_snapc &&
2759 (need & CEPH_CAP_FILE_WR)) {
2760 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2762 * we can not call down_read() when
2763 * task isn't in TASK_RUNNING state
2765 if (flags & NON_BLOCKING) {
2770 spin_unlock(&ci->i_ceph_lock);
2771 down_read(&mdsc->snap_rwsem);
2772 snap_rwsem_locked = true;
2775 snap_rwsem_locked = true;
2777 if ((have & want) == want)
2778 *got = need | (want & ~exclude);
2781 ceph_take_cap_refs(ci, *got, true);
2785 int session_readonly = false;
2787 if (ci->i_auth_cap &&
2788 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2789 struct ceph_mds_session *s = ci->i_auth_cap->session;
2790 spin_lock(&s->s_cap_lock);
2791 session_readonly = s->s_readonly;
2792 spin_unlock(&s->s_cap_lock);
2794 if (session_readonly) {
2795 dout("get_cap_refs %p need %s but mds%d readonly\n",
2796 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2801 if (ceph_inode_is_shutdown(inode)) {
2802 dout("get_cap_refs %p inode is shutdown\n", inode);
2806 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2807 if (need & ~mds_wanted) {
2808 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2809 inode, ceph_cap_string(need),
2810 ceph_cap_string(mds_wanted));
2815 dout("get_cap_refs %p have %s need %s\n", inode,
2816 ceph_cap_string(have), ceph_cap_string(need));
2820 __ceph_touch_fmode(ci, mdsc, flags);
2822 spin_unlock(&ci->i_ceph_lock);
2823 if (snap_rwsem_locked)
2824 up_read(&mdsc->snap_rwsem);
2827 ceph_update_cap_mis(&mdsc->metric);
2829 ceph_update_cap_hit(&mdsc->metric);
2831 dout("get_cap_refs %p ret %d got %s\n", inode,
2832 ret, ceph_cap_string(*got));
2837 * Check the offset we are writing up to against our current
2838 * max_size. If necessary, tell the MDS we want to write to
2841 static void check_max_size(struct inode *inode, loff_t endoff)
2843 struct ceph_inode_info *ci = ceph_inode(inode);
2846 /* do we need to explicitly request a larger max_size? */
2847 spin_lock(&ci->i_ceph_lock);
2848 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2849 dout("write %p at large endoff %llu, req max_size\n",
2851 ci->i_wanted_max_size = endoff;
2853 /* duplicate ceph_check_caps()'s logic */
2854 if (ci->i_auth_cap &&
2855 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2856 ci->i_wanted_max_size > ci->i_max_size &&
2857 ci->i_wanted_max_size > ci->i_requested_max_size)
2859 spin_unlock(&ci->i_ceph_lock);
2861 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2864 static inline int get_used_fmode(int caps)
2867 if (caps & CEPH_CAP_FILE_RD)
2868 fmode |= CEPH_FILE_MODE_RD;
2869 if (caps & CEPH_CAP_FILE_WR)
2870 fmode |= CEPH_FILE_MODE_WR;
2874 int ceph_try_get_caps(struct inode *inode, int need, int want,
2875 bool nonblock, int *got)
2879 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2880 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2881 CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2882 CEPH_CAP_ANY_DIR_OPS));
2884 ret = ceph_pool_perm_check(inode, need);
2889 flags = get_used_fmode(need | want);
2891 flags |= NON_BLOCKING;
2893 ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2894 /* three special error codes */
2895 if (ret == -EAGAIN || ret == -EFBIG || ret == -EUCLEAN)
2901 * Wait for caps, and take cap references. If we can't get a WR cap
2902 * due to a small max_size, make sure we check_max_size (and possibly
2903 * ask the mds) so we don't get hung up indefinitely.
2905 int ceph_get_caps(struct file *filp, int need, int want, loff_t endoff, int *got)
2907 struct ceph_file_info *fi = filp->private_data;
2908 struct inode *inode = file_inode(filp);
2909 struct ceph_inode_info *ci = ceph_inode(inode);
2910 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2911 int ret, _got, flags;
2913 ret = ceph_pool_perm_check(inode, need);
2917 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2918 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2921 flags = get_used_fmode(need | want);
2924 flags &= CEPH_FILE_MODE_MASK;
2925 if (vfs_inode_has_locks(inode))
2926 flags |= CHECK_FILELOCK;
2928 ret = try_get_cap_refs(inode, need, want, endoff,
2930 WARN_ON_ONCE(ret == -EAGAIN);
2932 struct ceph_mds_client *mdsc = fsc->mdsc;
2934 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2936 cw.ino = ceph_ino(inode);
2937 cw.tgid = current->tgid;
2941 spin_lock(&mdsc->caps_list_lock);
2942 list_add(&cw.list, &mdsc->cap_wait_list);
2943 spin_unlock(&mdsc->caps_list_lock);
2945 /* make sure used fmode not timeout */
2946 ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2947 add_wait_queue(&ci->i_cap_wq, &wait);
2949 flags |= NON_BLOCKING;
2950 while (!(ret = try_get_cap_refs(inode, need, want,
2951 endoff, flags, &_got))) {
2952 if (signal_pending(current)) {
2956 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2959 remove_wait_queue(&ci->i_cap_wq, &wait);
2960 ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
2962 spin_lock(&mdsc->caps_list_lock);
2964 spin_unlock(&mdsc->caps_list_lock);
2970 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2971 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2972 if (ret >= 0 && _got)
2973 ceph_put_cap_refs(ci, _got);
2978 if (ret == -EFBIG || ret == -EUCLEAN) {
2979 int ret2 = ceph_wait_on_async_create(inode);
2983 if (ret == -EFBIG) {
2984 check_max_size(inode, endoff);
2987 if (ret == -EUCLEAN) {
2988 /* session was killed, try renew caps */
2989 ret = ceph_renew_caps(inode, flags);
2996 if (S_ISREG(ci->netfs.inode.i_mode) &&
2997 ceph_has_inline_data(ci) &&
2998 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2999 i_size_read(inode) > 0) {
3001 find_get_page(inode->i_mapping, 0);
3003 bool uptodate = PageUptodate(page);
3010 * drop cap refs first because getattr while
3011 * holding * caps refs can cause deadlock.
3013 ceph_put_cap_refs(ci, _got);
3017 * getattr request will bring inline data into
3020 ret = __ceph_do_getattr(inode, NULL,
3021 CEPH_STAT_CAP_INLINE_DATA,
3034 * Take cap refs. Caller must already know we hold at least one ref
3035 * on the caps in question or we don't know this is safe.
3037 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
3039 spin_lock(&ci->i_ceph_lock);
3040 ceph_take_cap_refs(ci, caps, false);
3041 spin_unlock(&ci->i_ceph_lock);
3046 * drop cap_snap that is not associated with any snapshot.
3047 * we don't need to send FLUSHSNAP message for it.
3049 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
3050 struct ceph_cap_snap *capsnap)
3052 if (!capsnap->need_flush &&
3053 !capsnap->writing && !capsnap->dirty_pages) {
3054 dout("dropping cap_snap %p follows %llu\n",
3055 capsnap, capsnap->follows);
3056 BUG_ON(capsnap->cap_flush.tid > 0);
3057 ceph_put_snap_context(capsnap->context);
3058 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
3059 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3061 list_del(&capsnap->ci_item);
3062 ceph_put_cap_snap(capsnap);
3068 enum put_cap_refs_mode {
3069 PUT_CAP_REFS_SYNC = 0,
3070 PUT_CAP_REFS_NO_CHECK,
3077 * If we released the last ref on any given cap, call ceph_check_caps
3078 * to release (or schedule a release).
3080 * If we are releasing a WR cap (from a sync write), finalize any affected
3081 * cap_snap, and wake up any waiters.
3083 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
3084 enum put_cap_refs_mode mode)
3086 struct inode *inode = &ci->netfs.inode;
3087 int last = 0, put = 0, flushsnaps = 0, wake = 0;
3088 bool check_flushsnaps = false;
3090 spin_lock(&ci->i_ceph_lock);
3091 if (had & CEPH_CAP_PIN)
3093 if (had & CEPH_CAP_FILE_RD)
3094 if (--ci->i_rd_ref == 0)
3096 if (had & CEPH_CAP_FILE_CACHE)
3097 if (--ci->i_rdcache_ref == 0)
3099 if (had & CEPH_CAP_FILE_EXCL)
3100 if (--ci->i_fx_ref == 0)
3102 if (had & CEPH_CAP_FILE_BUFFER) {
3103 if (--ci->i_wb_ref == 0) {
3105 /* put the ref held by ceph_take_cap_refs() */
3107 check_flushsnaps = true;
3109 dout("put_cap_refs %p wb %d -> %d (?)\n",
3110 inode, ci->i_wb_ref+1, ci->i_wb_ref);
3112 if (had & CEPH_CAP_FILE_WR) {
3113 if (--ci->i_wr_ref == 0) {
3115 check_flushsnaps = true;
3116 if (ci->i_wrbuffer_ref_head == 0 &&
3117 ci->i_dirty_caps == 0 &&
3118 ci->i_flushing_caps == 0) {
3119 BUG_ON(!ci->i_head_snapc);
3120 ceph_put_snap_context(ci->i_head_snapc);
3121 ci->i_head_snapc = NULL;
3123 /* see comment in __ceph_remove_cap() */
3124 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3125 ceph_change_snap_realm(inode, NULL);
3128 if (check_flushsnaps && __ceph_have_pending_cap_snap(ci)) {
3129 struct ceph_cap_snap *capsnap =
3130 list_last_entry(&ci->i_cap_snaps,
3131 struct ceph_cap_snap,
3134 capsnap->writing = 0;
3135 if (ceph_try_drop_cap_snap(ci, capsnap))
3136 /* put the ref held by ceph_queue_cap_snap() */
3138 else if (__ceph_finish_cap_snap(ci, capsnap))
3142 spin_unlock(&ci->i_ceph_lock);
3144 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3145 last ? " last" : "", put ? " put" : "");
3148 case PUT_CAP_REFS_SYNC:
3150 ceph_check_caps(ci, 0, NULL);
3151 else if (flushsnaps)
3152 ceph_flush_snaps(ci, NULL);
3154 case PUT_CAP_REFS_ASYNC:
3156 ceph_queue_check_caps(inode);
3157 else if (flushsnaps)
3158 ceph_queue_flush_snaps(inode);
3164 wake_up_all(&ci->i_cap_wq);
3169 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3171 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_SYNC);
3174 void ceph_put_cap_refs_async(struct ceph_inode_info *ci, int had)
3176 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_ASYNC);
3179 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3181 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_NO_CHECK);
3185 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3186 * context. Adjust per-snap dirty page accounting as appropriate.
3187 * Once all dirty data for a cap_snap is flushed, flush snapped file
3188 * metadata back to the MDS. If we dropped the last ref, call
3191 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3192 struct ceph_snap_context *snapc)
3194 struct inode *inode = &ci->netfs.inode;
3195 struct ceph_cap_snap *capsnap = NULL, *iter;
3198 bool flush_snaps = false;
3199 bool complete_capsnap = false;
3201 spin_lock(&ci->i_ceph_lock);
3202 ci->i_wrbuffer_ref -= nr;
3203 if (ci->i_wrbuffer_ref == 0) {
3208 if (ci->i_head_snapc == snapc) {
3209 ci->i_wrbuffer_ref_head -= nr;
3210 if (ci->i_wrbuffer_ref_head == 0 &&
3211 ci->i_wr_ref == 0 &&
3212 ci->i_dirty_caps == 0 &&
3213 ci->i_flushing_caps == 0) {
3214 BUG_ON(!ci->i_head_snapc);
3215 ceph_put_snap_context(ci->i_head_snapc);
3216 ci->i_head_snapc = NULL;
3218 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3220 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3221 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3222 last ? " LAST" : "");
3224 list_for_each_entry(iter, &ci->i_cap_snaps, ci_item) {
3225 if (iter->context == snapc) {
3233 * The capsnap should already be removed when removing
3234 * auth cap in the case of a forced unmount.
3236 WARN_ON_ONCE(ci->i_auth_cap);
3240 capsnap->dirty_pages -= nr;
3241 if (capsnap->dirty_pages == 0) {
3242 complete_capsnap = true;
3243 if (!capsnap->writing) {
3244 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3247 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3252 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3253 " snap %lld %d/%d -> %d/%d %s%s\n",
3254 inode, capsnap, capsnap->context->seq,
3255 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3256 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3257 last ? " (wrbuffer last)" : "",
3258 complete_capsnap ? " (complete capsnap)" : "");
3262 spin_unlock(&ci->i_ceph_lock);
3265 ceph_check_caps(ci, 0, NULL);
3266 } else if (flush_snaps) {
3267 ceph_flush_snaps(ci, NULL);
3269 if (complete_capsnap)
3270 wake_up_all(&ci->i_cap_wq);
3277 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3279 static void invalidate_aliases(struct inode *inode)
3281 struct dentry *dn, *prev = NULL;
3283 dout("invalidate_aliases inode %p\n", inode);
3284 d_prune_aliases(inode);
3286 * For non-directory inode, d_find_alias() only returns
3287 * hashed dentry. After calling d_invalidate(), the
3288 * dentry becomes unhashed.
3290 * For directory inode, d_find_alias() can return
3291 * unhashed dentry. But directory inode should have
3292 * one alias at most.
3294 while ((dn = d_find_alias(inode))) {
3308 struct cap_extra_info {
3309 struct ceph_string *pool_ns;
3319 /* currently issued */
3321 struct timespec64 btime;
3325 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3326 * actually be a revocation if it specifies a smaller cap set.)
3328 * caller holds s_mutex and i_ceph_lock, we drop both.
3330 static void handle_cap_grant(struct inode *inode,
3331 struct ceph_mds_session *session,
3332 struct ceph_cap *cap,
3333 struct ceph_mds_caps *grant,
3334 struct ceph_buffer *xattr_buf,
3335 struct cap_extra_info *extra_info)
3336 __releases(ci->i_ceph_lock)
3337 __releases(session->s_mdsc->snap_rwsem)
3339 struct ceph_inode_info *ci = ceph_inode(inode);
3340 int seq = le32_to_cpu(grant->seq);
3341 int newcaps = le32_to_cpu(grant->caps);
3342 int used, wanted, dirty;
3343 u64 size = le64_to_cpu(grant->size);
3344 u64 max_size = le64_to_cpu(grant->max_size);
3345 unsigned char check_caps = 0;
3346 bool was_stale = cap->cap_gen < atomic_read(&session->s_cap_gen);
3348 bool writeback = false;
3349 bool queue_trunc = false;
3350 bool queue_invalidate = false;
3351 bool deleted_inode = false;
3352 bool fill_inline = false;
3354 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3355 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3356 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3357 i_size_read(inode));
3361 * If CACHE is being revoked, and we have no dirty buffers,
3362 * try to invalidate (once). (If there are dirty buffers, we
3363 * will invalidate _after_ writeback.)
3365 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3366 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3367 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3368 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3369 if (try_nonblocking_invalidate(inode)) {
3370 /* there were locked pages.. invalidate later
3371 in a separate thread. */
3372 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3373 queue_invalidate = true;
3374 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3380 cap->issued = cap->implemented = CEPH_CAP_PIN;
3383 * auth mds of the inode changed. we received the cap export message,
3384 * but still haven't received the cap import message. handle_cap_export
3385 * updated the new auth MDS' cap.
3387 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3388 * that was sent before the cap import message. So don't remove caps.
3390 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3391 WARN_ON(cap != ci->i_auth_cap);
3392 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3394 newcaps |= cap->issued;
3397 /* side effects now are allowed */
3398 cap->cap_gen = atomic_read(&session->s_cap_gen);
3401 __check_cap_issue(ci, cap, newcaps);
3403 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3405 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3406 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3407 umode_t mode = le32_to_cpu(grant->mode);
3409 if (inode_wrong_type(inode, mode))
3410 pr_warn_once("inode type changed! (ino %llx.%llx is 0%o, mds says 0%o)\n",
3411 ceph_vinop(inode), inode->i_mode, mode);
3413 inode->i_mode = mode;
3414 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3415 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3416 ci->i_btime = extra_info->btime;
3417 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3418 from_kuid(&init_user_ns, inode->i_uid),
3419 from_kgid(&init_user_ns, inode->i_gid));
3422 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3423 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3424 set_nlink(inode, le32_to_cpu(grant->nlink));
3425 if (inode->i_nlink == 0)
3426 deleted_inode = true;
3429 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3431 int len = le32_to_cpu(grant->xattr_len);
3432 u64 version = le64_to_cpu(grant->xattr_version);
3434 if (version > ci->i_xattrs.version) {
3435 dout(" got new xattrs v%llu on %p len %d\n",
3436 version, inode, len);
3437 if (ci->i_xattrs.blob)
3438 ceph_buffer_put(ci->i_xattrs.blob);
3439 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3440 ci->i_xattrs.version = version;
3441 ceph_forget_all_cached_acls(inode);
3442 ceph_security_invalidate_secctx(inode);
3446 if (newcaps & CEPH_CAP_ANY_RD) {
3447 struct timespec64 mtime, atime, ctime;
3448 /* ctime/mtime/atime? */
3449 ceph_decode_timespec64(&mtime, &grant->mtime);
3450 ceph_decode_timespec64(&atime, &grant->atime);
3451 ceph_decode_timespec64(&ctime, &grant->ctime);
3452 ceph_fill_file_time(inode, extra_info->issued,
3453 le32_to_cpu(grant->time_warp_seq),
3454 &ctime, &mtime, &atime);
3457 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3458 ci->i_files = extra_info->nfiles;
3459 ci->i_subdirs = extra_info->nsubdirs;
3462 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3463 /* file layout may have changed */
3464 s64 old_pool = ci->i_layout.pool_id;
3465 struct ceph_string *old_ns;
3467 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3468 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3469 lockdep_is_held(&ci->i_ceph_lock));
3470 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3472 if (ci->i_layout.pool_id != old_pool ||
3473 extra_info->pool_ns != old_ns)
3474 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3476 extra_info->pool_ns = old_ns;
3478 /* size/truncate_seq? */
3479 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3480 le32_to_cpu(grant->truncate_seq),
3481 le64_to_cpu(grant->truncate_size),
3485 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3486 if (max_size != ci->i_max_size) {
3487 dout("max_size %lld -> %llu\n",
3488 ci->i_max_size, max_size);
3489 ci->i_max_size = max_size;
3490 if (max_size >= ci->i_wanted_max_size) {
3491 ci->i_wanted_max_size = 0; /* reset */
3492 ci->i_requested_max_size = 0;
3498 /* check cap bits */
3499 wanted = __ceph_caps_wanted(ci);
3500 used = __ceph_caps_used(ci);
3501 dirty = __ceph_caps_dirty(ci);
3502 dout(" my wanted = %s, used = %s, dirty %s\n",
3503 ceph_cap_string(wanted),
3504 ceph_cap_string(used),
3505 ceph_cap_string(dirty));
3507 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3508 (wanted & ~(cap->mds_wanted | newcaps))) {
3510 * If mds is importing cap, prior cap messages that update
3511 * 'wanted' may get dropped by mds (migrate seq mismatch).
3513 * We don't send cap message to update 'wanted' if what we
3514 * want are already issued. If mds revokes caps, cap message
3515 * that releases caps also tells mds what we want. But if
3516 * caps got revoked by mds forcedly (session stale). We may
3517 * haven't told mds what we want.
3522 /* revocation, grant, or no-op? */
3523 if (cap->issued & ~newcaps) {
3524 int revoking = cap->issued & ~newcaps;
3526 dout("revocation: %s -> %s (revoking %s)\n",
3527 ceph_cap_string(cap->issued),
3528 ceph_cap_string(newcaps),
3529 ceph_cap_string(revoking));
3530 if (S_ISREG(inode->i_mode) &&
3531 (revoking & used & CEPH_CAP_FILE_BUFFER))
3532 writeback = true; /* initiate writeback; will delay ack */
3533 else if (queue_invalidate &&
3534 revoking == CEPH_CAP_FILE_CACHE &&
3535 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3536 ; /* do nothing yet, invalidation will be queued */
3537 else if (cap == ci->i_auth_cap)
3538 check_caps = 1; /* check auth cap only */
3540 check_caps = 2; /* check all caps */
3541 /* If there is new caps, try to wake up the waiters */
3542 if (~cap->issued & newcaps)
3544 cap->issued = newcaps;
3545 cap->implemented |= newcaps;
3546 } else if (cap->issued == newcaps) {
3547 dout("caps unchanged: %s -> %s\n",
3548 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3550 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3551 ceph_cap_string(newcaps));
3552 /* non-auth MDS is revoking the newly grant caps ? */
3553 if (cap == ci->i_auth_cap &&
3554 __ceph_caps_revoking_other(ci, cap, newcaps))
3557 cap->issued = newcaps;
3558 cap->implemented |= newcaps; /* add bits only, to
3559 * avoid stepping on a
3560 * pending revocation */
3563 BUG_ON(cap->issued & ~cap->implemented);
3565 /* don't let check_caps skip sending a response to MDS for revoke msgs */
3566 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_REVOKE) {
3567 cap->mds_wanted = 0;
3568 if (cap == ci->i_auth_cap)
3569 check_caps = 1; /* check auth cap only */
3571 check_caps = 2; /* check all caps */
3574 if (extra_info->inline_version > 0 &&
3575 extra_info->inline_version >= ci->i_inline_version) {
3576 ci->i_inline_version = extra_info->inline_version;
3577 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3578 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3582 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3583 if (ci->i_auth_cap == cap) {
3584 if (newcaps & ~extra_info->issued)
3587 if (ci->i_requested_max_size > max_size ||
3588 !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3589 /* re-request max_size if necessary */
3590 ci->i_requested_max_size = 0;
3594 ceph_kick_flushing_inode_caps(session, ci);
3596 up_read(&session->s_mdsc->snap_rwsem);
3598 spin_unlock(&ci->i_ceph_lock);
3601 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3602 extra_info->inline_len);
3605 ceph_queue_vmtruncate(inode);
3609 * queue inode for writeback: we can't actually call
3610 * filemap_write_and_wait, etc. from message handler
3613 ceph_queue_writeback(inode);
3614 if (queue_invalidate)
3615 ceph_queue_invalidate(inode);
3617 invalidate_aliases(inode);
3619 wake_up_all(&ci->i_cap_wq);
3621 mutex_unlock(&session->s_mutex);
3622 if (check_caps == 1)
3623 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3625 else if (check_caps == 2)
3626 ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3630 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3631 * MDS has been safely committed.
3633 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3634 struct ceph_mds_caps *m,
3635 struct ceph_mds_session *session,
3636 struct ceph_cap *cap)
3637 __releases(ci->i_ceph_lock)
3639 struct ceph_inode_info *ci = ceph_inode(inode);
3640 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3641 struct ceph_cap_flush *cf, *tmp_cf;
3642 LIST_HEAD(to_remove);
3643 unsigned seq = le32_to_cpu(m->seq);
3644 int dirty = le32_to_cpu(m->dirty);
3647 bool wake_ci = false;
3648 bool wake_mdsc = false;
3650 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3651 /* Is this the one that was flushed? */
3652 if (cf->tid == flush_tid)
3655 /* Is this a capsnap? */
3659 if (cf->tid <= flush_tid) {
3661 * An earlier or current tid. The FLUSH_ACK should
3662 * represent a superset of this flush's caps.
3664 wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3665 list_add_tail(&cf->i_list, &to_remove);
3668 * This is a later one. Any caps in it are still dirty
3669 * so don't count them as cleaned.
3671 cleaned &= ~cf->caps;
3677 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3678 " flushing %s -> %s\n",
3679 inode, session->s_mds, seq, ceph_cap_string(dirty),
3680 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3681 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3683 if (list_empty(&to_remove) && !cleaned)
3686 ci->i_flushing_caps &= ~cleaned;
3688 spin_lock(&mdsc->cap_dirty_lock);
3690 list_for_each_entry(cf, &to_remove, i_list)
3691 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3693 if (ci->i_flushing_caps == 0) {
3694 if (list_empty(&ci->i_cap_flush_list)) {
3695 list_del_init(&ci->i_flushing_item);
3696 if (!list_empty(&session->s_cap_flushing)) {
3697 dout(" mds%d still flushing cap on %p\n",
3699 &list_first_entry(&session->s_cap_flushing,
3700 struct ceph_inode_info,
3701 i_flushing_item)->netfs.inode);
3704 mdsc->num_cap_flushing--;
3705 dout(" inode %p now !flushing\n", inode);
3707 if (ci->i_dirty_caps == 0) {
3708 dout(" inode %p now clean\n", inode);
3709 BUG_ON(!list_empty(&ci->i_dirty_item));
3711 if (ci->i_wr_ref == 0 &&
3712 ci->i_wrbuffer_ref_head == 0) {
3713 BUG_ON(!ci->i_head_snapc);
3714 ceph_put_snap_context(ci->i_head_snapc);
3715 ci->i_head_snapc = NULL;
3718 BUG_ON(list_empty(&ci->i_dirty_item));
3721 spin_unlock(&mdsc->cap_dirty_lock);
3724 spin_unlock(&ci->i_ceph_lock);
3726 while (!list_empty(&to_remove)) {
3727 cf = list_first_entry(&to_remove,
3728 struct ceph_cap_flush, i_list);
3729 list_del_init(&cf->i_list);
3730 if (!cf->is_capsnap)
3731 ceph_free_cap_flush(cf);
3735 wake_up_all(&ci->i_cap_wq);
3737 wake_up_all(&mdsc->cap_flushing_wq);
3742 void __ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3743 bool *wake_ci, bool *wake_mdsc)
3745 struct ceph_inode_info *ci = ceph_inode(inode);
3746 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3749 lockdep_assert_held(&ci->i_ceph_lock);
3751 dout("removing capsnap %p, inode %p ci %p\n", capsnap, inode, ci);
3753 list_del_init(&capsnap->ci_item);
3754 ret = __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3758 spin_lock(&mdsc->cap_dirty_lock);
3759 if (list_empty(&ci->i_cap_flush_list))
3760 list_del_init(&ci->i_flushing_item);
3762 ret = __detach_cap_flush_from_mdsc(mdsc, &capsnap->cap_flush);
3765 spin_unlock(&mdsc->cap_dirty_lock);
3768 void ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3769 bool *wake_ci, bool *wake_mdsc)
3771 struct ceph_inode_info *ci = ceph_inode(inode);
3773 lockdep_assert_held(&ci->i_ceph_lock);
3775 WARN_ON_ONCE(capsnap->dirty_pages || capsnap->writing);
3776 __ceph_remove_capsnap(inode, capsnap, wake_ci, wake_mdsc);
3780 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3781 * throw away our cap_snap.
3783 * Caller hold s_mutex.
3785 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3786 struct ceph_mds_caps *m,
3787 struct ceph_mds_session *session)
3789 struct ceph_inode_info *ci = ceph_inode(inode);
3790 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3791 u64 follows = le64_to_cpu(m->snap_follows);
3792 struct ceph_cap_snap *capsnap = NULL, *iter;
3793 bool wake_ci = false;
3794 bool wake_mdsc = false;
3796 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3797 inode, ci, session->s_mds, follows);
3799 spin_lock(&ci->i_ceph_lock);
3800 list_for_each_entry(iter, &ci->i_cap_snaps, ci_item) {
3801 if (iter->follows == follows) {
3802 if (iter->cap_flush.tid != flush_tid) {
3803 dout(" cap_snap %p follows %lld tid %lld !="
3804 " %lld\n", iter, follows,
3805 flush_tid, iter->cap_flush.tid);
3811 dout(" skipping cap_snap %p follows %lld\n",
3812 iter, iter->follows);
3816 ceph_remove_capsnap(inode, capsnap, &wake_ci, &wake_mdsc);
3817 spin_unlock(&ci->i_ceph_lock);
3820 ceph_put_snap_context(capsnap->context);
3821 ceph_put_cap_snap(capsnap);
3823 wake_up_all(&ci->i_cap_wq);
3825 wake_up_all(&mdsc->cap_flushing_wq);
3831 * Handle TRUNC from MDS, indicating file truncation.
3833 * caller hold s_mutex.
3835 static bool handle_cap_trunc(struct inode *inode,
3836 struct ceph_mds_caps *trunc,
3837 struct ceph_mds_session *session)
3839 struct ceph_inode_info *ci = ceph_inode(inode);
3840 int mds = session->s_mds;
3841 int seq = le32_to_cpu(trunc->seq);
3842 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3843 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3844 u64 size = le64_to_cpu(trunc->size);
3845 int implemented = 0;
3846 int dirty = __ceph_caps_dirty(ci);
3847 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3848 bool queue_trunc = false;
3850 lockdep_assert_held(&ci->i_ceph_lock);
3852 issued |= implemented | dirty;
3854 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3855 inode, mds, seq, truncate_size, truncate_seq);
3856 queue_trunc = ceph_fill_file_size(inode, issued,
3857 truncate_seq, truncate_size, size);
3862 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3863 * different one. If we are the most recent migration we've seen (as
3864 * indicated by mseq), make note of the migrating cap bits for the
3865 * duration (until we see the corresponding IMPORT).
3867 * caller holds s_mutex
3869 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3870 struct ceph_mds_cap_peer *ph,
3871 struct ceph_mds_session *session)
3873 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3874 struct ceph_mds_session *tsession = NULL;
3875 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3876 struct ceph_inode_info *ci = ceph_inode(inode);
3878 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3879 unsigned t_seq, t_mseq;
3881 int mds = session->s_mds;
3884 t_cap_id = le64_to_cpu(ph->cap_id);
3885 t_seq = le32_to_cpu(ph->seq);
3886 t_mseq = le32_to_cpu(ph->mseq);
3887 target = le32_to_cpu(ph->mds);
3889 t_cap_id = t_seq = t_mseq = 0;
3893 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3894 inode, ci, mds, mseq, target);
3896 down_read(&mdsc->snap_rwsem);
3897 spin_lock(&ci->i_ceph_lock);
3898 cap = __get_cap_for_mds(ci, mds);
3899 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3903 ceph_remove_cap(cap, false);
3908 * now we know we haven't received the cap import message yet
3909 * because the exported cap still exist.
3912 issued = cap->issued;
3913 if (issued != cap->implemented)
3914 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3915 "ino (%llx.%llx) mds%d seq %d mseq %d "
3916 "issued %s implemented %s\n",
3917 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3918 ceph_cap_string(issued),
3919 ceph_cap_string(cap->implemented));
3922 tcap = __get_cap_for_mds(ci, target);
3924 /* already have caps from the target */
3925 if (tcap->cap_id == t_cap_id &&
3926 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3927 dout(" updating import cap %p mds%d\n", tcap, target);
3928 tcap->cap_id = t_cap_id;
3929 tcap->seq = t_seq - 1;
3930 tcap->issue_seq = t_seq - 1;
3931 tcap->issued |= issued;
3932 tcap->implemented |= issued;
3933 if (cap == ci->i_auth_cap) {
3934 ci->i_auth_cap = tcap;
3935 change_auth_cap_ses(ci, tcap->session);
3938 ceph_remove_cap(cap, false);
3940 } else if (tsession) {
3941 /* add placeholder for the export tagert */
3942 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3944 ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3945 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3947 if (!list_empty(&ci->i_cap_flush_list) &&
3948 ci->i_auth_cap == tcap) {
3949 spin_lock(&mdsc->cap_dirty_lock);
3950 list_move_tail(&ci->i_flushing_item,
3951 &tcap->session->s_cap_flushing);
3952 spin_unlock(&mdsc->cap_dirty_lock);
3955 ceph_remove_cap(cap, false);
3959 spin_unlock(&ci->i_ceph_lock);
3960 up_read(&mdsc->snap_rwsem);
3961 mutex_unlock(&session->s_mutex);
3963 /* open target session */
3964 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3965 if (!IS_ERR(tsession)) {
3967 mutex_lock(&session->s_mutex);
3968 mutex_lock_nested(&tsession->s_mutex,
3969 SINGLE_DEPTH_NESTING);
3971 mutex_lock(&tsession->s_mutex);
3972 mutex_lock_nested(&session->s_mutex,
3973 SINGLE_DEPTH_NESTING);
3975 new_cap = ceph_get_cap(mdsc, NULL);
3980 mutex_lock(&session->s_mutex);
3985 spin_unlock(&ci->i_ceph_lock);
3986 up_read(&mdsc->snap_rwsem);
3987 mutex_unlock(&session->s_mutex);
3989 mutex_unlock(&tsession->s_mutex);
3990 ceph_put_mds_session(tsession);
3993 ceph_put_cap(mdsc, new_cap);
3997 * Handle cap IMPORT.
3999 * caller holds s_mutex. acquires i_ceph_lock
4001 static void handle_cap_import(struct ceph_mds_client *mdsc,
4002 struct inode *inode, struct ceph_mds_caps *im,
4003 struct ceph_mds_cap_peer *ph,
4004 struct ceph_mds_session *session,
4005 struct ceph_cap **target_cap, int *old_issued)
4007 struct ceph_inode_info *ci = ceph_inode(inode);
4008 struct ceph_cap *cap, *ocap, *new_cap = NULL;
4009 int mds = session->s_mds;
4011 unsigned caps = le32_to_cpu(im->caps);
4012 unsigned wanted = le32_to_cpu(im->wanted);
4013 unsigned seq = le32_to_cpu(im->seq);
4014 unsigned mseq = le32_to_cpu(im->migrate_seq);
4015 u64 realmino = le64_to_cpu(im->realm);
4016 u64 cap_id = le64_to_cpu(im->cap_id);
4021 p_cap_id = le64_to_cpu(ph->cap_id);
4022 peer = le32_to_cpu(ph->mds);
4028 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
4029 inode, ci, mds, mseq, peer);
4031 cap = __get_cap_for_mds(ci, mds);
4034 spin_unlock(&ci->i_ceph_lock);
4035 new_cap = ceph_get_cap(mdsc, NULL);
4036 spin_lock(&ci->i_ceph_lock);
4042 ceph_put_cap(mdsc, new_cap);
4047 __ceph_caps_issued(ci, &issued);
4048 issued |= __ceph_caps_dirty(ci);
4050 ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
4051 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
4053 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
4054 if (ocap && ocap->cap_id == p_cap_id) {
4055 dout(" remove export cap %p mds%d flags %d\n",
4056 ocap, peer, ph->flags);
4057 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
4058 (ocap->seq != le32_to_cpu(ph->seq) ||
4059 ocap->mseq != le32_to_cpu(ph->mseq))) {
4060 pr_err_ratelimited("handle_cap_import: "
4061 "mismatched seq/mseq: ino (%llx.%llx) "
4062 "mds%d seq %d mseq %d importer mds%d "
4063 "has peer seq %d mseq %d\n",
4064 ceph_vinop(inode), peer, ocap->seq,
4065 ocap->mseq, mds, le32_to_cpu(ph->seq),
4066 le32_to_cpu(ph->mseq));
4068 ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
4071 *old_issued = issued;
4076 * Handle a caps message from the MDS.
4078 * Identify the appropriate session, inode, and call the right handler
4079 * based on the cap op.
4081 void ceph_handle_caps(struct ceph_mds_session *session,
4082 struct ceph_msg *msg)
4084 struct ceph_mds_client *mdsc = session->s_mdsc;
4085 struct inode *inode;
4086 struct ceph_inode_info *ci;
4087 struct ceph_cap *cap;
4088 struct ceph_mds_caps *h;
4089 struct ceph_mds_cap_peer *peer = NULL;
4090 struct ceph_snap_realm *realm = NULL;
4092 int msg_version = le16_to_cpu(msg->hdr.version);
4094 struct ceph_vino vino;
4096 size_t snaptrace_len;
4098 struct cap_extra_info extra_info = {};
4100 bool close_sessions = false;
4102 dout("handle_caps from mds%d\n", session->s_mds);
4104 if (!ceph_inc_mds_stopping_blocker(mdsc, session))
4108 end = msg->front.iov_base + msg->front.iov_len;
4109 if (msg->front.iov_len < sizeof(*h))
4111 h = msg->front.iov_base;
4112 op = le32_to_cpu(h->op);
4113 vino.ino = le64_to_cpu(h->ino);
4114 vino.snap = CEPH_NOSNAP;
4115 seq = le32_to_cpu(h->seq);
4116 mseq = le32_to_cpu(h->migrate_seq);
4119 snaptrace_len = le32_to_cpu(h->snap_trace_len);
4120 p = snaptrace + snaptrace_len;
4122 if (msg_version >= 2) {
4124 ceph_decode_32_safe(&p, end, flock_len, bad);
4125 if (p + flock_len > end)
4130 if (msg_version >= 3) {
4131 if (op == CEPH_CAP_OP_IMPORT) {
4132 if (p + sizeof(*peer) > end)
4136 } else if (op == CEPH_CAP_OP_EXPORT) {
4137 /* recorded in unused fields */
4138 peer = (void *)&h->size;
4142 if (msg_version >= 4) {
4143 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
4144 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
4145 if (p + extra_info.inline_len > end)
4147 extra_info.inline_data = p;
4148 p += extra_info.inline_len;
4151 if (msg_version >= 5) {
4152 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
4155 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
4156 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
4159 if (msg_version >= 8) {
4163 ceph_decode_skip_64(&p, end, bad); // flush_tid
4165 ceph_decode_skip_32(&p, end, bad); // caller_uid
4166 ceph_decode_skip_32(&p, end, bad); // caller_gid
4168 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4169 if (pool_ns_len > 0) {
4170 ceph_decode_need(&p, end, pool_ns_len, bad);
4171 extra_info.pool_ns =
4172 ceph_find_or_create_string(p, pool_ns_len);
4177 if (msg_version >= 9) {
4178 struct ceph_timespec *btime;
4180 if (p + sizeof(*btime) > end)
4183 ceph_decode_timespec64(&extra_info.btime, btime);
4184 p += sizeof(*btime);
4185 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4188 if (msg_version >= 11) {
4190 ceph_decode_skip_32(&p, end, bad); // flags
4192 extra_info.dirstat_valid = true;
4193 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4194 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4198 inode = ceph_find_inode(mdsc->fsc->sb, vino);
4199 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4202 mutex_lock(&session->s_mutex);
4203 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4207 dout(" i don't have ino %llx\n", vino.ino);
4209 if (op == CEPH_CAP_OP_IMPORT) {
4210 cap = ceph_get_cap(mdsc, NULL);
4211 cap->cap_ino = vino.ino;
4212 cap->queue_release = 1;
4213 cap->cap_id = le64_to_cpu(h->cap_id);
4216 cap->issue_seq = seq;
4217 spin_lock(&session->s_cap_lock);
4218 __ceph_queue_cap_release(session, cap);
4219 spin_unlock(&session->s_cap_lock);
4221 goto flush_cap_releases;
4223 ci = ceph_inode(inode);
4225 /* these will work even if we don't have a cap yet */
4227 case CEPH_CAP_OP_FLUSHSNAP_ACK:
4228 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4232 case CEPH_CAP_OP_EXPORT:
4233 handle_cap_export(inode, h, peer, session);
4236 case CEPH_CAP_OP_IMPORT:
4238 if (snaptrace_len) {
4239 down_write(&mdsc->snap_rwsem);
4240 if (ceph_update_snap_trace(mdsc, snaptrace,
4241 snaptrace + snaptrace_len,
4243 up_write(&mdsc->snap_rwsem);
4244 close_sessions = true;
4247 downgrade_write(&mdsc->snap_rwsem);
4249 down_read(&mdsc->snap_rwsem);
4251 spin_lock(&ci->i_ceph_lock);
4252 handle_cap_import(mdsc, inode, h, peer, session,
4253 &cap, &extra_info.issued);
4254 handle_cap_grant(inode, session, cap,
4255 h, msg->middle, &extra_info);
4257 ceph_put_snap_realm(mdsc, realm);
4261 /* the rest require a cap */
4262 spin_lock(&ci->i_ceph_lock);
4263 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4265 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4266 inode, ceph_ino(inode), ceph_snap(inode),
4268 spin_unlock(&ci->i_ceph_lock);
4269 goto flush_cap_releases;
4272 /* note that each of these drops i_ceph_lock for us */
4274 case CEPH_CAP_OP_REVOKE:
4275 case CEPH_CAP_OP_GRANT:
4276 __ceph_caps_issued(ci, &extra_info.issued);
4277 extra_info.issued |= __ceph_caps_dirty(ci);
4278 handle_cap_grant(inode, session, cap,
4279 h, msg->middle, &extra_info);
4282 case CEPH_CAP_OP_FLUSH_ACK:
4283 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4287 case CEPH_CAP_OP_TRUNC:
4288 queue_trunc = handle_cap_trunc(inode, h, session);
4289 spin_unlock(&ci->i_ceph_lock);
4291 ceph_queue_vmtruncate(inode);
4295 spin_unlock(&ci->i_ceph_lock);
4296 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4297 ceph_cap_op_name(op));
4301 mutex_unlock(&session->s_mutex);
4305 ceph_dec_mds_stopping_blocker(mdsc);
4307 ceph_put_string(extra_info.pool_ns);
4309 /* Defer closing the sessions after s_mutex lock being released */
4311 ceph_mdsc_close_sessions(mdsc);
4317 * send any cap release message to try to move things
4318 * along for the mds (who clearly thinks we still have this
4321 ceph_flush_cap_releases(mdsc, session);
4325 pr_err("ceph_handle_caps: corrupt message\n");
4331 * Delayed work handler to process end of delayed cap release LRU list.
4333 * If new caps are added to the list while processing it, these won't get
4334 * processed in this run. In this case, the ci->i_hold_caps_max will be
4335 * returned so that the work can be scheduled accordingly.
4337 unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4339 struct inode *inode;
4340 struct ceph_inode_info *ci;
4341 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
4342 unsigned long delay_max = opt->caps_wanted_delay_max * HZ;
4343 unsigned long loop_start = jiffies;
4344 unsigned long delay = 0;
4346 dout("check_delayed_caps\n");
4347 spin_lock(&mdsc->cap_delay_lock);
4348 while (!list_empty(&mdsc->cap_delay_list)) {
4349 ci = list_first_entry(&mdsc->cap_delay_list,
4350 struct ceph_inode_info,
4352 if (time_before(loop_start, ci->i_hold_caps_max - delay_max)) {
4353 dout("%s caps added recently. Exiting loop", __func__);
4354 delay = ci->i_hold_caps_max;
4357 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4358 time_before(jiffies, ci->i_hold_caps_max))
4360 list_del_init(&ci->i_cap_delay_list);
4362 inode = igrab(&ci->netfs.inode);
4364 spin_unlock(&mdsc->cap_delay_lock);
4365 dout("check_delayed_caps on %p\n", inode);
4366 ceph_check_caps(ci, 0, NULL);
4368 spin_lock(&mdsc->cap_delay_lock);
4371 spin_unlock(&mdsc->cap_delay_lock);
4377 * Flush all dirty caps to the mds
4379 static void flush_dirty_session_caps(struct ceph_mds_session *s)
4381 struct ceph_mds_client *mdsc = s->s_mdsc;
4382 struct ceph_inode_info *ci;
4383 struct inode *inode;
4385 dout("flush_dirty_caps\n");
4386 spin_lock(&mdsc->cap_dirty_lock);
4387 while (!list_empty(&s->s_cap_dirty)) {
4388 ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4390 inode = &ci->netfs.inode;
4392 dout("flush_dirty_caps %llx.%llx\n", ceph_vinop(inode));
4393 spin_unlock(&mdsc->cap_dirty_lock);
4394 ceph_wait_on_async_create(inode);
4395 ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4397 spin_lock(&mdsc->cap_dirty_lock);
4399 spin_unlock(&mdsc->cap_dirty_lock);
4400 dout("flush_dirty_caps done\n");
4403 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4405 ceph_mdsc_iterate_sessions(mdsc, flush_dirty_session_caps, true);
4408 void __ceph_touch_fmode(struct ceph_inode_info *ci,
4409 struct ceph_mds_client *mdsc, int fmode)
4411 unsigned long now = jiffies;
4412 if (fmode & CEPH_FILE_MODE_RD)
4413 ci->i_last_rd = now;
4414 if (fmode & CEPH_FILE_MODE_WR)
4415 ci->i_last_wr = now;
4416 /* queue periodic check */
4418 __ceph_is_any_real_caps(ci) &&
4419 list_empty(&ci->i_cap_delay_list))
4420 __cap_delay_requeue(mdsc, ci);
4423 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4425 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->netfs.inode.i_sb);
4426 int bits = (fmode << 1) | 1;
4427 bool already_opened = false;
4431 atomic64_inc(&mdsc->metric.opened_files);
4433 spin_lock(&ci->i_ceph_lock);
4434 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4436 * If any of the mode ref is larger than 0,
4437 * that means it has been already opened by
4438 * others. Just skip checking the PIN ref.
4440 if (i && ci->i_nr_by_mode[i])
4441 already_opened = true;
4443 if (bits & (1 << i))
4444 ci->i_nr_by_mode[i] += count;
4447 if (!already_opened)
4448 percpu_counter_inc(&mdsc->metric.opened_inodes);
4449 spin_unlock(&ci->i_ceph_lock);
4453 * Drop open file reference. If we were the last open file,
4454 * we may need to release capabilities to the MDS (or schedule
4455 * their delayed release).
4457 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4459 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->netfs.inode.i_sb);
4460 int bits = (fmode << 1) | 1;
4461 bool is_closed = true;
4465 atomic64_dec(&mdsc->metric.opened_files);
4467 spin_lock(&ci->i_ceph_lock);
4468 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4469 if (bits & (1 << i)) {
4470 BUG_ON(ci->i_nr_by_mode[i] < count);
4471 ci->i_nr_by_mode[i] -= count;
4475 * If any of the mode ref is not 0 after
4476 * decreased, that means it is still opened
4477 * by others. Just skip checking the PIN ref.
4479 if (i && ci->i_nr_by_mode[i])
4484 percpu_counter_dec(&mdsc->metric.opened_inodes);
4485 spin_unlock(&ci->i_ceph_lock);
4489 * For a soon-to-be unlinked file, drop the LINK caps. If it
4490 * looks like the link count will hit 0, drop any other caps (other
4491 * than PIN) we don't specifically want (due to the file still being
4494 int ceph_drop_caps_for_unlink(struct inode *inode)
4496 struct ceph_inode_info *ci = ceph_inode(inode);
4497 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4499 spin_lock(&ci->i_ceph_lock);
4500 if (inode->i_nlink == 1) {
4501 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4503 if (__ceph_caps_dirty(ci)) {
4504 struct ceph_mds_client *mdsc =
4505 ceph_inode_to_client(inode)->mdsc;
4506 __cap_delay_requeue_front(mdsc, ci);
4509 spin_unlock(&ci->i_ceph_lock);
4514 * Helpers for embedding cap and dentry lease releases into mds
4517 * @force is used by dentry_release (below) to force inclusion of a
4518 * record for the directory inode, even when there aren't any caps to
4521 int ceph_encode_inode_release(void **p, struct inode *inode,
4522 int mds, int drop, int unless, int force)
4524 struct ceph_inode_info *ci = ceph_inode(inode);
4525 struct ceph_cap *cap;
4526 struct ceph_mds_request_release *rel = *p;
4530 spin_lock(&ci->i_ceph_lock);
4531 used = __ceph_caps_used(ci);
4532 dirty = __ceph_caps_dirty(ci);
4534 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4535 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4536 ceph_cap_string(unless));
4538 /* only drop unused, clean caps */
4539 drop &= ~(used | dirty);
4541 cap = __get_cap_for_mds(ci, mds);
4542 if (cap && __cap_is_valid(cap)) {
4543 unless &= cap->issued;
4545 if (unless & CEPH_CAP_AUTH_EXCL)
4546 drop &= ~CEPH_CAP_AUTH_SHARED;
4547 if (unless & CEPH_CAP_LINK_EXCL)
4548 drop &= ~CEPH_CAP_LINK_SHARED;
4549 if (unless & CEPH_CAP_XATTR_EXCL)
4550 drop &= ~CEPH_CAP_XATTR_SHARED;
4551 if (unless & CEPH_CAP_FILE_EXCL)
4552 drop &= ~CEPH_CAP_FILE_SHARED;
4555 if (force || (cap->issued & drop)) {
4556 if (cap->issued & drop) {
4557 int wanted = __ceph_caps_wanted(ci);
4558 dout("encode_inode_release %p cap %p "
4559 "%s -> %s, wanted %s -> %s\n", inode, cap,
4560 ceph_cap_string(cap->issued),
4561 ceph_cap_string(cap->issued & ~drop),
4562 ceph_cap_string(cap->mds_wanted),
4563 ceph_cap_string(wanted));
4565 cap->issued &= ~drop;
4566 cap->implemented &= ~drop;
4567 cap->mds_wanted = wanted;
4568 if (cap == ci->i_auth_cap &&
4569 !(wanted & CEPH_CAP_ANY_FILE_WR))
4570 ci->i_requested_max_size = 0;
4572 dout("encode_inode_release %p cap %p %s"
4573 " (force)\n", inode, cap,
4574 ceph_cap_string(cap->issued));
4577 rel->ino = cpu_to_le64(ceph_ino(inode));
4578 rel->cap_id = cpu_to_le64(cap->cap_id);
4579 rel->seq = cpu_to_le32(cap->seq);
4580 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4581 rel->mseq = cpu_to_le32(cap->mseq);
4582 rel->caps = cpu_to_le32(cap->implemented);
4583 rel->wanted = cpu_to_le32(cap->mds_wanted);
4589 dout("encode_inode_release %p cap %p %s (noop)\n",
4590 inode, cap, ceph_cap_string(cap->issued));
4593 spin_unlock(&ci->i_ceph_lock);
4597 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4599 int mds, int drop, int unless)
4601 struct ceph_mds_request_release *rel = *p;
4602 struct ceph_dentry_info *di = ceph_dentry(dentry);
4606 /* This shouldn't happen */
4610 * force an record for the directory caps if we have a dentry lease.
4611 * this is racy (can't take i_ceph_lock and d_lock together), but it
4612 * doesn't have to be perfect; the mds will revoke anything we don't
4615 spin_lock(&dentry->d_lock);
4616 if (di->lease_session && di->lease_session->s_mds == mds)
4618 spin_unlock(&dentry->d_lock);
4620 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4622 spin_lock(&dentry->d_lock);
4623 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4624 dout("encode_dentry_release %p mds%d seq %d\n",
4625 dentry, mds, (int)di->lease_seq);
4626 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4627 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4628 *p += dentry->d_name.len;
4629 rel->dname_seq = cpu_to_le32(di->lease_seq);
4630 __ceph_mdsc_drop_dentry_lease(dentry);
4632 spin_unlock(&dentry->d_lock);
4636 static int remove_capsnaps(struct ceph_mds_client *mdsc, struct inode *inode)
4638 struct ceph_inode_info *ci = ceph_inode(inode);
4639 struct ceph_cap_snap *capsnap;
4640 int capsnap_release = 0;
4642 lockdep_assert_held(&ci->i_ceph_lock);
4644 dout("removing capsnaps, ci is %p, inode is %p\n", ci, inode);
4646 while (!list_empty(&ci->i_cap_snaps)) {
4647 capsnap = list_first_entry(&ci->i_cap_snaps,
4648 struct ceph_cap_snap, ci_item);
4649 __ceph_remove_capsnap(inode, capsnap, NULL, NULL);
4650 ceph_put_snap_context(capsnap->context);
4651 ceph_put_cap_snap(capsnap);
4654 wake_up_all(&ci->i_cap_wq);
4655 wake_up_all(&mdsc->cap_flushing_wq);
4656 return capsnap_release;
4659 int ceph_purge_inode_cap(struct inode *inode, struct ceph_cap *cap, bool *invalidate)
4661 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
4662 struct ceph_mds_client *mdsc = fsc->mdsc;
4663 struct ceph_inode_info *ci = ceph_inode(inode);
4665 bool dirty_dropped = false;
4668 lockdep_assert_held(&ci->i_ceph_lock);
4670 dout("removing cap %p, ci is %p, inode is %p\n",
4671 cap, ci, &ci->netfs.inode);
4673 is_auth = (cap == ci->i_auth_cap);
4674 __ceph_remove_cap(cap, false);
4676 struct ceph_cap_flush *cf;
4678 if (ceph_inode_is_shutdown(inode)) {
4679 if (inode->i_data.nrpages > 0)
4681 if (ci->i_wrbuffer_ref > 0)
4682 mapping_set_error(&inode->i_data, -EIO);
4685 spin_lock(&mdsc->cap_dirty_lock);
4687 /* trash all of the cap flushes for this inode */
4688 while (!list_empty(&ci->i_cap_flush_list)) {
4689 cf = list_first_entry(&ci->i_cap_flush_list,
4690 struct ceph_cap_flush, i_list);
4691 list_del_init(&cf->g_list);
4692 list_del_init(&cf->i_list);
4693 if (!cf->is_capsnap)
4694 ceph_free_cap_flush(cf);
4697 if (!list_empty(&ci->i_dirty_item)) {
4698 pr_warn_ratelimited(
4699 " dropping dirty %s state for %p %lld\n",
4700 ceph_cap_string(ci->i_dirty_caps),
4701 inode, ceph_ino(inode));
4702 ci->i_dirty_caps = 0;
4703 list_del_init(&ci->i_dirty_item);
4704 dirty_dropped = true;
4706 if (!list_empty(&ci->i_flushing_item)) {
4707 pr_warn_ratelimited(
4708 " dropping dirty+flushing %s state for %p %lld\n",
4709 ceph_cap_string(ci->i_flushing_caps),
4710 inode, ceph_ino(inode));
4711 ci->i_flushing_caps = 0;
4712 list_del_init(&ci->i_flushing_item);
4713 mdsc->num_cap_flushing--;
4714 dirty_dropped = true;
4716 spin_unlock(&mdsc->cap_dirty_lock);
4718 if (dirty_dropped) {
4719 mapping_set_error(inode->i_mapping, -EIO);
4721 if (ci->i_wrbuffer_ref_head == 0 &&
4722 ci->i_wr_ref == 0 &&
4723 ci->i_dirty_caps == 0 &&
4724 ci->i_flushing_caps == 0) {
4725 ceph_put_snap_context(ci->i_head_snapc);
4726 ci->i_head_snapc = NULL;
4730 if (atomic_read(&ci->i_filelock_ref) > 0) {
4731 /* make further file lock syscall return -EIO */
4732 ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
4733 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
4734 inode, ceph_ino(inode));
4737 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
4738 cf = ci->i_prealloc_cap_flush;
4739 ci->i_prealloc_cap_flush = NULL;
4740 if (!cf->is_capsnap)
4741 ceph_free_cap_flush(cf);
4744 if (!list_empty(&ci->i_cap_snaps))
4745 iputs = remove_capsnaps(mdsc, inode);