1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
5 #include <linux/kernel.h>
6 #include <linux/sched/signal.h>
7 #include <linux/slab.h>
8 #include <linux/vmalloc.h>
9 #include <linux/wait.h>
10 #include <linux/writeback.h>
11 #include <linux/iversion.h>
14 #include "mds_client.h"
16 #include <linux/ceph/decode.h>
17 #include <linux/ceph/messenger.h>
20 * Capability management
22 * The Ceph metadata servers control client access to inode metadata
23 * and file data by issuing capabilities, granting clients permission
24 * to read and/or write both inode field and file data to OSDs
25 * (storage nodes). Each capability consists of a set of bits
26 * indicating which operations are allowed.
28 * If the client holds a *_SHARED cap, the client has a coherent value
29 * that can be safely read from the cached inode.
31 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
32 * client is allowed to change inode attributes (e.g., file size,
33 * mtime), note its dirty state in the ceph_cap, and asynchronously
34 * flush that metadata change to the MDS.
36 * In the event of a conflicting operation (perhaps by another
37 * client), the MDS will revoke the conflicting client capabilities.
39 * In order for a client to cache an inode, it must hold a capability
40 * with at least one MDS server. When inodes are released, release
41 * notifications are batched and periodically sent en masse to the MDS
42 * cluster to release server state.
45 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc);
46 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
47 struct ceph_mds_session *session,
48 struct ceph_inode_info *ci,
49 u64 oldest_flush_tid);
52 * Generate readable cap strings for debugging output.
54 #define MAX_CAP_STR 20
55 static char cap_str[MAX_CAP_STR][40];
56 static DEFINE_SPINLOCK(cap_str_lock);
57 static int last_cap_str;
59 static char *gcap_string(char *s, int c)
61 if (c & CEPH_CAP_GSHARED)
63 if (c & CEPH_CAP_GEXCL)
65 if (c & CEPH_CAP_GCACHE)
71 if (c & CEPH_CAP_GBUFFER)
73 if (c & CEPH_CAP_GWREXTEND)
75 if (c & CEPH_CAP_GLAZYIO)
80 const char *ceph_cap_string(int caps)
86 spin_lock(&cap_str_lock);
88 if (last_cap_str == MAX_CAP_STR)
90 spin_unlock(&cap_str_lock);
94 if (caps & CEPH_CAP_PIN)
97 c = (caps >> CEPH_CAP_SAUTH) & 3;
100 s = gcap_string(s, c);
103 c = (caps >> CEPH_CAP_SLINK) & 3;
106 s = gcap_string(s, c);
109 c = (caps >> CEPH_CAP_SXATTR) & 3;
112 s = gcap_string(s, c);
115 c = caps >> CEPH_CAP_SFILE;
118 s = gcap_string(s, c);
127 void ceph_caps_init(struct ceph_mds_client *mdsc)
129 INIT_LIST_HEAD(&mdsc->caps_list);
130 spin_lock_init(&mdsc->caps_list_lock);
133 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
135 struct ceph_cap *cap;
137 spin_lock(&mdsc->caps_list_lock);
138 while (!list_empty(&mdsc->caps_list)) {
139 cap = list_first_entry(&mdsc->caps_list,
140 struct ceph_cap, caps_item);
141 list_del(&cap->caps_item);
142 kmem_cache_free(ceph_cap_cachep, cap);
144 mdsc->caps_total_count = 0;
145 mdsc->caps_avail_count = 0;
146 mdsc->caps_use_count = 0;
147 mdsc->caps_reserve_count = 0;
148 mdsc->caps_min_count = 0;
149 spin_unlock(&mdsc->caps_list_lock);
152 void ceph_adjust_caps_max_min(struct ceph_mds_client *mdsc,
153 struct ceph_mount_options *fsopt)
155 spin_lock(&mdsc->caps_list_lock);
156 mdsc->caps_min_count = fsopt->max_readdir;
157 if (mdsc->caps_min_count < 1024)
158 mdsc->caps_min_count = 1024;
159 mdsc->caps_use_max = fsopt->caps_max;
160 if (mdsc->caps_use_max > 0 &&
161 mdsc->caps_use_max < mdsc->caps_min_count)
162 mdsc->caps_use_max = mdsc->caps_min_count;
163 spin_unlock(&mdsc->caps_list_lock);
166 static void __ceph_unreserve_caps(struct ceph_mds_client *mdsc, int nr_caps)
168 struct ceph_cap *cap;
172 BUG_ON(mdsc->caps_reserve_count < nr_caps);
173 mdsc->caps_reserve_count -= nr_caps;
174 if (mdsc->caps_avail_count >=
175 mdsc->caps_reserve_count + mdsc->caps_min_count) {
176 mdsc->caps_total_count -= nr_caps;
177 for (i = 0; i < nr_caps; i++) {
178 cap = list_first_entry(&mdsc->caps_list,
179 struct ceph_cap, caps_item);
180 list_del(&cap->caps_item);
181 kmem_cache_free(ceph_cap_cachep, cap);
184 mdsc->caps_avail_count += nr_caps;
187 dout("%s: caps %d = %d used + %d resv + %d avail\n",
189 mdsc->caps_total_count, mdsc->caps_use_count,
190 mdsc->caps_reserve_count, mdsc->caps_avail_count);
191 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
192 mdsc->caps_reserve_count +
193 mdsc->caps_avail_count);
198 * Called under mdsc->mutex.
200 int ceph_reserve_caps(struct ceph_mds_client *mdsc,
201 struct ceph_cap_reservation *ctx, int need)
204 struct ceph_cap *cap;
209 bool trimmed = false;
210 struct ceph_mds_session *s;
213 dout("reserve caps ctx=%p need=%d\n", ctx, need);
215 /* first reserve any caps that are already allocated */
216 spin_lock(&mdsc->caps_list_lock);
217 if (mdsc->caps_avail_count >= need)
220 have = mdsc->caps_avail_count;
221 mdsc->caps_avail_count -= have;
222 mdsc->caps_reserve_count += have;
223 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
224 mdsc->caps_reserve_count +
225 mdsc->caps_avail_count);
226 spin_unlock(&mdsc->caps_list_lock);
228 for (i = have; i < need; ) {
229 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
231 list_add(&cap->caps_item, &newcaps);
238 for (j = 0; j < mdsc->max_sessions; j++) {
239 s = __ceph_lookup_mds_session(mdsc, j);
242 mutex_unlock(&mdsc->mutex);
244 mutex_lock(&s->s_mutex);
245 max_caps = s->s_nr_caps - (need - i);
246 ceph_trim_caps(mdsc, s, max_caps);
247 mutex_unlock(&s->s_mutex);
249 ceph_put_mds_session(s);
250 mutex_lock(&mdsc->mutex);
254 spin_lock(&mdsc->caps_list_lock);
255 if (mdsc->caps_avail_count) {
257 if (mdsc->caps_avail_count >= need - i)
258 more_have = need - i;
260 more_have = mdsc->caps_avail_count;
264 mdsc->caps_avail_count -= more_have;
265 mdsc->caps_reserve_count += more_have;
268 spin_unlock(&mdsc->caps_list_lock);
273 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
274 ctx, need, have + alloc);
280 BUG_ON(have + alloc != need);
285 spin_lock(&mdsc->caps_list_lock);
286 mdsc->caps_total_count += alloc;
287 mdsc->caps_reserve_count += alloc;
288 list_splice(&newcaps, &mdsc->caps_list);
290 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
291 mdsc->caps_reserve_count +
292 mdsc->caps_avail_count);
295 __ceph_unreserve_caps(mdsc, have + alloc);
297 spin_unlock(&mdsc->caps_list_lock);
299 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
300 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
301 mdsc->caps_reserve_count, mdsc->caps_avail_count);
305 void ceph_unreserve_caps(struct ceph_mds_client *mdsc,
306 struct ceph_cap_reservation *ctx)
308 bool reclaim = false;
312 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
313 spin_lock(&mdsc->caps_list_lock);
314 __ceph_unreserve_caps(mdsc, ctx->count);
317 if (mdsc->caps_use_max > 0 &&
318 mdsc->caps_use_count > mdsc->caps_use_max)
320 spin_unlock(&mdsc->caps_list_lock);
323 ceph_reclaim_caps_nr(mdsc, ctx->used);
326 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
327 struct ceph_cap_reservation *ctx)
329 struct ceph_cap *cap = NULL;
331 /* temporary, until we do something about cap import/export */
333 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
335 spin_lock(&mdsc->caps_list_lock);
336 mdsc->caps_use_count++;
337 mdsc->caps_total_count++;
338 spin_unlock(&mdsc->caps_list_lock);
340 spin_lock(&mdsc->caps_list_lock);
341 if (mdsc->caps_avail_count) {
342 BUG_ON(list_empty(&mdsc->caps_list));
344 mdsc->caps_avail_count--;
345 mdsc->caps_use_count++;
346 cap = list_first_entry(&mdsc->caps_list,
347 struct ceph_cap, caps_item);
348 list_del(&cap->caps_item);
350 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
351 mdsc->caps_reserve_count + mdsc->caps_avail_count);
353 spin_unlock(&mdsc->caps_list_lock);
359 spin_lock(&mdsc->caps_list_lock);
360 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
361 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
362 mdsc->caps_reserve_count, mdsc->caps_avail_count);
364 BUG_ON(ctx->count > mdsc->caps_reserve_count);
365 BUG_ON(list_empty(&mdsc->caps_list));
369 mdsc->caps_reserve_count--;
370 mdsc->caps_use_count++;
372 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
373 list_del(&cap->caps_item);
375 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
376 mdsc->caps_reserve_count + mdsc->caps_avail_count);
377 spin_unlock(&mdsc->caps_list_lock);
381 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
383 spin_lock(&mdsc->caps_list_lock);
384 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
385 cap, mdsc->caps_total_count, mdsc->caps_use_count,
386 mdsc->caps_reserve_count, mdsc->caps_avail_count);
387 mdsc->caps_use_count--;
389 * Keep some preallocated caps around (ceph_min_count), to
390 * avoid lots of free/alloc churn.
392 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
393 mdsc->caps_min_count) {
394 mdsc->caps_total_count--;
395 kmem_cache_free(ceph_cap_cachep, cap);
397 mdsc->caps_avail_count++;
398 list_add(&cap->caps_item, &mdsc->caps_list);
401 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
402 mdsc->caps_reserve_count + mdsc->caps_avail_count);
403 spin_unlock(&mdsc->caps_list_lock);
406 void ceph_reservation_status(struct ceph_fs_client *fsc,
407 int *total, int *avail, int *used, int *reserved,
410 struct ceph_mds_client *mdsc = fsc->mdsc;
412 spin_lock(&mdsc->caps_list_lock);
415 *total = mdsc->caps_total_count;
417 *avail = mdsc->caps_avail_count;
419 *used = mdsc->caps_use_count;
421 *reserved = mdsc->caps_reserve_count;
423 *min = mdsc->caps_min_count;
425 spin_unlock(&mdsc->caps_list_lock);
429 * Find ceph_cap for given mds, if any.
431 * Called with i_ceph_lock held.
433 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
435 struct ceph_cap *cap;
436 struct rb_node *n = ci->i_caps.rb_node;
439 cap = rb_entry(n, struct ceph_cap, ci_node);
442 else if (mds > cap->mds)
450 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
452 struct ceph_cap *cap;
454 spin_lock(&ci->i_ceph_lock);
455 cap = __get_cap_for_mds(ci, mds);
456 spin_unlock(&ci->i_ceph_lock);
461 * Called under i_ceph_lock.
463 static void __insert_cap_node(struct ceph_inode_info *ci,
464 struct ceph_cap *new)
466 struct rb_node **p = &ci->i_caps.rb_node;
467 struct rb_node *parent = NULL;
468 struct ceph_cap *cap = NULL;
472 cap = rb_entry(parent, struct ceph_cap, ci_node);
473 if (new->mds < cap->mds)
475 else if (new->mds > cap->mds)
481 rb_link_node(&new->ci_node, parent, p);
482 rb_insert_color(&new->ci_node, &ci->i_caps);
486 * (re)set cap hold timeouts, which control the delayed release
487 * of unused caps back to the MDS. Should be called on cap use.
489 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
490 struct ceph_inode_info *ci)
492 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
493 ci->i_hold_caps_max = round_jiffies(jiffies +
494 opt->caps_wanted_delay_max * HZ);
495 dout("__cap_set_timeouts %p %lu\n", &ci->netfs.inode,
496 ci->i_hold_caps_max - jiffies);
500 * (Re)queue cap at the end of the delayed cap release list.
502 * If I_FLUSH is set, leave the inode at the front of the list.
504 * Caller holds i_ceph_lock
505 * -> we take mdsc->cap_delay_lock
507 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
508 struct ceph_inode_info *ci)
510 dout("__cap_delay_requeue %p flags 0x%lx at %lu\n", &ci->netfs.inode,
511 ci->i_ceph_flags, ci->i_hold_caps_max);
512 if (!mdsc->stopping) {
513 spin_lock(&mdsc->cap_delay_lock);
514 if (!list_empty(&ci->i_cap_delay_list)) {
515 if (ci->i_ceph_flags & CEPH_I_FLUSH)
517 list_del_init(&ci->i_cap_delay_list);
519 __cap_set_timeouts(mdsc, ci);
520 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
522 spin_unlock(&mdsc->cap_delay_lock);
527 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
528 * indicating we should send a cap message to flush dirty metadata
529 * asap, and move to the front of the delayed cap list.
531 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
532 struct ceph_inode_info *ci)
534 dout("__cap_delay_requeue_front %p\n", &ci->netfs.inode);
535 spin_lock(&mdsc->cap_delay_lock);
536 ci->i_ceph_flags |= CEPH_I_FLUSH;
537 if (!list_empty(&ci->i_cap_delay_list))
538 list_del_init(&ci->i_cap_delay_list);
539 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
540 spin_unlock(&mdsc->cap_delay_lock);
544 * Cancel delayed work on cap.
546 * Caller must hold i_ceph_lock.
548 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
549 struct ceph_inode_info *ci)
551 dout("__cap_delay_cancel %p\n", &ci->netfs.inode);
552 if (list_empty(&ci->i_cap_delay_list))
554 spin_lock(&mdsc->cap_delay_lock);
555 list_del_init(&ci->i_cap_delay_list);
556 spin_unlock(&mdsc->cap_delay_lock);
559 /* Common issue checks for add_cap, handle_cap_grant. */
560 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
563 unsigned had = __ceph_caps_issued(ci, NULL);
565 lockdep_assert_held(&ci->i_ceph_lock);
568 * Each time we receive FILE_CACHE anew, we increment
571 if (S_ISREG(ci->netfs.inode.i_mode) &&
572 (issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
573 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
578 * If FILE_SHARED is newly issued, mark dir not complete. We don't
579 * know what happened to this directory while we didn't have the cap.
580 * If FILE_SHARED is being revoked, also mark dir not complete. It
581 * stops on-going cached readdir.
583 if ((issued & CEPH_CAP_FILE_SHARED) != (had & CEPH_CAP_FILE_SHARED)) {
584 if (issued & CEPH_CAP_FILE_SHARED)
585 atomic_inc(&ci->i_shared_gen);
586 if (S_ISDIR(ci->netfs.inode.i_mode)) {
587 dout(" marking %p NOT complete\n", &ci->netfs.inode);
588 __ceph_dir_clear_complete(ci);
592 /* Wipe saved layout if we're losing DIR_CREATE caps */
593 if (S_ISDIR(ci->netfs.inode.i_mode) && (had & CEPH_CAP_DIR_CREATE) &&
594 !(issued & CEPH_CAP_DIR_CREATE)) {
595 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
596 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
601 * change_auth_cap_ses - move inode to appropriate lists when auth caps change
602 * @ci: inode to be moved
603 * @session: new auth caps session
605 void change_auth_cap_ses(struct ceph_inode_info *ci,
606 struct ceph_mds_session *session)
608 lockdep_assert_held(&ci->i_ceph_lock);
610 if (list_empty(&ci->i_dirty_item) && list_empty(&ci->i_flushing_item))
613 spin_lock(&session->s_mdsc->cap_dirty_lock);
614 if (!list_empty(&ci->i_dirty_item))
615 list_move(&ci->i_dirty_item, &session->s_cap_dirty);
616 if (!list_empty(&ci->i_flushing_item))
617 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
618 spin_unlock(&session->s_mdsc->cap_dirty_lock);
622 * Add a capability under the given MDS session.
624 * Caller should hold session snap_rwsem (read) and ci->i_ceph_lock
626 * @fmode is the open file mode, if we are opening a file, otherwise
627 * it is < 0. (This is so we can atomically add the cap and add an
628 * open file reference to it.)
630 void ceph_add_cap(struct inode *inode,
631 struct ceph_mds_session *session, u64 cap_id,
632 unsigned issued, unsigned wanted,
633 unsigned seq, unsigned mseq, u64 realmino, int flags,
634 struct ceph_cap **new_cap)
636 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
637 struct ceph_inode_info *ci = ceph_inode(inode);
638 struct ceph_cap *cap;
639 int mds = session->s_mds;
643 lockdep_assert_held(&ci->i_ceph_lock);
645 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
646 session->s_mds, cap_id, ceph_cap_string(issued), seq);
648 gen = atomic_read(&session->s_cap_gen);
650 cap = __get_cap_for_mds(ci, mds);
656 cap->implemented = 0;
662 __insert_cap_node(ci, cap);
664 /* add to session cap list */
665 cap->session = session;
666 spin_lock(&session->s_cap_lock);
667 list_add_tail(&cap->session_caps, &session->s_caps);
668 session->s_nr_caps++;
669 atomic64_inc(&mdsc->metric.total_caps);
670 spin_unlock(&session->s_cap_lock);
672 spin_lock(&session->s_cap_lock);
673 list_move_tail(&cap->session_caps, &session->s_caps);
674 spin_unlock(&session->s_cap_lock);
676 if (cap->cap_gen < gen)
677 cap->issued = cap->implemented = CEPH_CAP_PIN;
680 * auth mds of the inode changed. we received the cap export
681 * message, but still haven't received the cap import message.
682 * handle_cap_export() updated the new auth MDS' cap.
684 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
685 * a message that was send before the cap import message. So
688 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
689 WARN_ON(cap != ci->i_auth_cap);
690 WARN_ON(cap->cap_id != cap_id);
693 issued |= cap->issued;
694 flags |= CEPH_CAP_FLAG_AUTH;
698 if (!ci->i_snap_realm ||
699 ((flags & CEPH_CAP_FLAG_AUTH) &&
700 realmino != (u64)-1 && ci->i_snap_realm->ino != realmino)) {
702 * add this inode to the appropriate snap realm
704 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
707 ceph_change_snap_realm(inode, realm);
709 WARN(1, "%s: couldn't find snap realm 0x%llx (ino 0x%llx oldrealm 0x%llx)\n",
710 __func__, realmino, ci->i_vino.ino,
711 ci->i_snap_realm ? ci->i_snap_realm->ino : 0);
714 __check_cap_issue(ci, cap, issued);
717 * If we are issued caps we don't want, or the mds' wanted
718 * value appears to be off, queue a check so we'll release
719 * later and/or update the mds wanted value.
721 actual_wanted = __ceph_caps_wanted(ci);
722 if ((wanted & ~actual_wanted) ||
723 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
724 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
725 ceph_cap_string(issued), ceph_cap_string(wanted),
726 ceph_cap_string(actual_wanted));
727 __cap_delay_requeue(mdsc, ci);
730 if (flags & CEPH_CAP_FLAG_AUTH) {
731 if (!ci->i_auth_cap ||
732 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
733 if (ci->i_auth_cap &&
734 ci->i_auth_cap->session != cap->session)
735 change_auth_cap_ses(ci, cap->session);
736 ci->i_auth_cap = cap;
737 cap->mds_wanted = wanted;
740 WARN_ON(ci->i_auth_cap == cap);
743 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
744 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
745 ceph_cap_string(issued|cap->issued), seq, mds);
746 cap->cap_id = cap_id;
747 cap->issued = issued;
748 cap->implemented |= issued;
749 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
750 cap->mds_wanted = wanted;
752 cap->mds_wanted |= wanted;
754 cap->issue_seq = seq;
757 wake_up_all(&ci->i_cap_wq);
761 * Return true if cap has not timed out and belongs to the current
762 * generation of the MDS session (i.e. has not gone 'stale' due to
763 * us losing touch with the mds).
765 static int __cap_is_valid(struct ceph_cap *cap)
770 gen = atomic_read(&cap->session->s_cap_gen);
771 ttl = cap->session->s_cap_ttl;
773 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
774 dout("__cap_is_valid %p cap %p issued %s "
775 "but STALE (gen %u vs %u)\n", &cap->ci->netfs.inode,
776 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
784 * Return set of valid cap bits issued to us. Note that caps time
785 * out, and may be invalidated in bulk if the client session times out
786 * and session->s_cap_gen is bumped.
788 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
790 int have = ci->i_snap_caps;
791 struct ceph_cap *cap;
796 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
797 cap = rb_entry(p, struct ceph_cap, ci_node);
798 if (!__cap_is_valid(cap))
800 dout("__ceph_caps_issued %p cap %p issued %s\n",
801 &ci->netfs.inode, cap, ceph_cap_string(cap->issued));
804 *implemented |= cap->implemented;
807 * exclude caps issued by non-auth MDS, but are been revoking
808 * by the auth MDS. The non-auth MDS should be revoking/exporting
809 * these caps, but the message is delayed.
811 if (ci->i_auth_cap) {
812 cap = ci->i_auth_cap;
813 have &= ~cap->implemented | cap->issued;
819 * Get cap bits issued by caps other than @ocap
821 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
823 int have = ci->i_snap_caps;
824 struct ceph_cap *cap;
827 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
828 cap = rb_entry(p, struct ceph_cap, ci_node);
831 if (!__cap_is_valid(cap))
839 * Move a cap to the end of the LRU (oldest caps at list head, newest
842 static void __touch_cap(struct ceph_cap *cap)
844 struct ceph_mds_session *s = cap->session;
846 spin_lock(&s->s_cap_lock);
847 if (!s->s_cap_iterator) {
848 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->netfs.inode, cap,
850 list_move_tail(&cap->session_caps, &s->s_caps);
852 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
853 &cap->ci->netfs.inode, cap, s->s_mds);
855 spin_unlock(&s->s_cap_lock);
859 * Check if we hold the given mask. If so, move the cap(s) to the
860 * front of their respective LRUs. (This is the preferred way for
861 * callers to check for caps they want.)
863 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
865 struct ceph_cap *cap;
867 int have = ci->i_snap_caps;
869 if ((have & mask) == mask) {
870 dout("__ceph_caps_issued_mask ino 0x%llx snap issued %s"
871 " (mask %s)\n", ceph_ino(&ci->netfs.inode),
872 ceph_cap_string(have),
873 ceph_cap_string(mask));
877 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
878 cap = rb_entry(p, struct ceph_cap, ci_node);
879 if (!__cap_is_valid(cap))
881 if ((cap->issued & mask) == mask) {
882 dout("__ceph_caps_issued_mask ino 0x%llx cap %p issued %s"
883 " (mask %s)\n", ceph_ino(&ci->netfs.inode), cap,
884 ceph_cap_string(cap->issued),
885 ceph_cap_string(mask));
891 /* does a combination of caps satisfy mask? */
893 if ((have & mask) == mask) {
894 dout("__ceph_caps_issued_mask ino 0x%llx combo issued %s"
895 " (mask %s)\n", ceph_ino(&ci->netfs.inode),
896 ceph_cap_string(cap->issued),
897 ceph_cap_string(mask));
901 /* touch this + preceding caps */
903 for (q = rb_first(&ci->i_caps); q != p;
905 cap = rb_entry(q, struct ceph_cap,
907 if (!__cap_is_valid(cap))
909 if (cap->issued & mask)
920 int __ceph_caps_issued_mask_metric(struct ceph_inode_info *ci, int mask,
923 struct ceph_fs_client *fsc = ceph_sb_to_client(ci->netfs.inode.i_sb);
926 r = __ceph_caps_issued_mask(ci, mask, touch);
928 ceph_update_cap_hit(&fsc->mdsc->metric);
930 ceph_update_cap_mis(&fsc->mdsc->metric);
935 * Return true if mask caps are currently being revoked by an MDS.
937 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
938 struct ceph_cap *ocap, int mask)
940 struct ceph_cap *cap;
943 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
944 cap = rb_entry(p, struct ceph_cap, ci_node);
946 (cap->implemented & ~cap->issued & mask))
952 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
954 struct inode *inode = &ci->netfs.inode;
957 spin_lock(&ci->i_ceph_lock);
958 ret = __ceph_caps_revoking_other(ci, NULL, mask);
959 spin_unlock(&ci->i_ceph_lock);
960 dout("ceph_caps_revoking %p %s = %d\n", inode,
961 ceph_cap_string(mask), ret);
965 int __ceph_caps_used(struct ceph_inode_info *ci)
969 used |= CEPH_CAP_PIN;
971 used |= CEPH_CAP_FILE_RD;
972 if (ci->i_rdcache_ref ||
973 (S_ISREG(ci->netfs.inode.i_mode) &&
974 ci->netfs.inode.i_data.nrpages))
975 used |= CEPH_CAP_FILE_CACHE;
977 used |= CEPH_CAP_FILE_WR;
978 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
979 used |= CEPH_CAP_FILE_BUFFER;
981 used |= CEPH_CAP_FILE_EXCL;
985 #define FMODE_WAIT_BIAS 1000
988 * wanted, by virtue of open file modes
990 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
992 const int PIN_SHIFT = ffs(CEPH_FILE_MODE_PIN);
993 const int RD_SHIFT = ffs(CEPH_FILE_MODE_RD);
994 const int WR_SHIFT = ffs(CEPH_FILE_MODE_WR);
995 const int LAZY_SHIFT = ffs(CEPH_FILE_MODE_LAZY);
996 struct ceph_mount_options *opt =
997 ceph_inode_to_client(&ci->netfs.inode)->mount_options;
998 unsigned long used_cutoff = jiffies - opt->caps_wanted_delay_max * HZ;
999 unsigned long idle_cutoff = jiffies - opt->caps_wanted_delay_min * HZ;
1001 if (S_ISDIR(ci->netfs.inode.i_mode)) {
1004 /* use used_cutoff here, to keep dir's wanted caps longer */
1005 if (ci->i_nr_by_mode[RD_SHIFT] > 0 ||
1006 time_after(ci->i_last_rd, used_cutoff))
1007 want |= CEPH_CAP_ANY_SHARED;
1009 if (ci->i_nr_by_mode[WR_SHIFT] > 0 ||
1010 time_after(ci->i_last_wr, used_cutoff)) {
1011 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1012 if (opt->flags & CEPH_MOUNT_OPT_ASYNC_DIROPS)
1013 want |= CEPH_CAP_ANY_DIR_OPS;
1016 if (want || ci->i_nr_by_mode[PIN_SHIFT] > 0)
1017 want |= CEPH_CAP_PIN;
1023 if (ci->i_nr_by_mode[RD_SHIFT] > 0) {
1024 if (ci->i_nr_by_mode[RD_SHIFT] >= FMODE_WAIT_BIAS ||
1025 time_after(ci->i_last_rd, used_cutoff))
1026 bits |= 1 << RD_SHIFT;
1027 } else if (time_after(ci->i_last_rd, idle_cutoff)) {
1028 bits |= 1 << RD_SHIFT;
1031 if (ci->i_nr_by_mode[WR_SHIFT] > 0) {
1032 if (ci->i_nr_by_mode[WR_SHIFT] >= FMODE_WAIT_BIAS ||
1033 time_after(ci->i_last_wr, used_cutoff))
1034 bits |= 1 << WR_SHIFT;
1035 } else if (time_after(ci->i_last_wr, idle_cutoff)) {
1036 bits |= 1 << WR_SHIFT;
1039 /* check lazyio only when read/write is wanted */
1040 if ((bits & (CEPH_FILE_MODE_RDWR << 1)) &&
1041 ci->i_nr_by_mode[LAZY_SHIFT] > 0)
1042 bits |= 1 << LAZY_SHIFT;
1044 return bits ? ceph_caps_for_mode(bits >> 1) : 0;
1049 * wanted, by virtue of open file modes AND cap refs (buffered/cached data)
1051 int __ceph_caps_wanted(struct ceph_inode_info *ci)
1053 int w = __ceph_caps_file_wanted(ci) | __ceph_caps_used(ci);
1054 if (S_ISDIR(ci->netfs.inode.i_mode)) {
1055 /* we want EXCL if holding caps of dir ops */
1056 if (w & CEPH_CAP_ANY_DIR_OPS)
1057 w |= CEPH_CAP_FILE_EXCL;
1059 /* we want EXCL if dirty data */
1060 if (w & CEPH_CAP_FILE_BUFFER)
1061 w |= CEPH_CAP_FILE_EXCL;
1067 * Return caps we have registered with the MDS(s) as 'wanted'.
1069 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
1071 struct ceph_cap *cap;
1075 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1076 cap = rb_entry(p, struct ceph_cap, ci_node);
1077 if (check && !__cap_is_valid(cap))
1079 if (cap == ci->i_auth_cap)
1080 mds_wanted |= cap->mds_wanted;
1082 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
1087 int ceph_is_any_caps(struct inode *inode)
1089 struct ceph_inode_info *ci = ceph_inode(inode);
1092 spin_lock(&ci->i_ceph_lock);
1093 ret = __ceph_is_any_real_caps(ci);
1094 spin_unlock(&ci->i_ceph_lock);
1100 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
1102 * caller should hold i_ceph_lock.
1103 * caller will not hold session s_mutex if called from destroy_inode.
1105 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1107 struct ceph_mds_session *session = cap->session;
1108 struct ceph_inode_info *ci = cap->ci;
1109 struct ceph_mds_client *mdsc;
1112 /* 'ci' being NULL means the remove have already occurred */
1114 dout("%s: cap inode is NULL\n", __func__);
1118 lockdep_assert_held(&ci->i_ceph_lock);
1120 dout("__ceph_remove_cap %p from %p\n", cap, &ci->netfs.inode);
1122 mdsc = ceph_inode_to_client(&ci->netfs.inode)->mdsc;
1124 /* remove from inode's cap rbtree, and clear auth cap */
1125 rb_erase(&cap->ci_node, &ci->i_caps);
1126 if (ci->i_auth_cap == cap)
1127 ci->i_auth_cap = NULL;
1129 /* remove from session list */
1130 spin_lock(&session->s_cap_lock);
1131 if (session->s_cap_iterator == cap) {
1132 /* not yet, we are iterating over this very cap */
1133 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1136 list_del_init(&cap->session_caps);
1137 session->s_nr_caps--;
1138 atomic64_dec(&mdsc->metric.total_caps);
1139 cap->session = NULL;
1142 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1146 * s_cap_reconnect is protected by s_cap_lock. no one changes
1147 * s_cap_gen while session is in the reconnect state.
1149 if (queue_release &&
1150 (!session->s_cap_reconnect ||
1151 cap->cap_gen == atomic_read(&session->s_cap_gen))) {
1152 cap->queue_release = 1;
1154 __ceph_queue_cap_release(session, cap);
1158 cap->queue_release = 0;
1160 cap->cap_ino = ci->i_vino.ino;
1162 spin_unlock(&session->s_cap_lock);
1165 ceph_put_cap(mdsc, cap);
1167 if (!__ceph_is_any_real_caps(ci)) {
1168 /* when reconnect denied, we remove session caps forcibly,
1169 * i_wr_ref can be non-zero. If there are ongoing write,
1170 * keep i_snap_realm.
1172 if (ci->i_wr_ref == 0 && ci->i_snap_realm)
1173 ceph_change_snap_realm(&ci->netfs.inode, NULL);
1175 __cap_delay_cancel(mdsc, ci);
1179 void ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1181 struct ceph_inode_info *ci = cap->ci;
1182 struct ceph_fs_client *fsc;
1184 /* 'ci' being NULL means the remove have already occurred */
1186 dout("%s: cap inode is NULL\n", __func__);
1190 lockdep_assert_held(&ci->i_ceph_lock);
1192 fsc = ceph_inode_to_client(&ci->netfs.inode);
1193 WARN_ON_ONCE(ci->i_auth_cap == cap &&
1194 !list_empty(&ci->i_dirty_item) &&
1195 !fsc->blocklisted &&
1196 !ceph_inode_is_shutdown(&ci->netfs.inode));
1198 __ceph_remove_cap(cap, queue_release);
1201 struct cap_msg_args {
1202 struct ceph_mds_session *session;
1203 u64 ino, cid, follows;
1204 u64 flush_tid, oldest_flush_tid, size, max_size;
1207 struct ceph_buffer *xattr_buf;
1208 struct ceph_buffer *old_xattr_buf;
1209 struct timespec64 atime, mtime, ctime, btime;
1210 int op, caps, wanted, dirty;
1211 u32 seq, issue_seq, mseq, time_warp_seq;
1221 * cap struct size + flock buffer size + inline version + inline data size +
1222 * osd_epoch_barrier + oldest_flush_tid
1224 #define CAP_MSG_SIZE (sizeof(struct ceph_mds_caps) + \
1225 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4)
1227 /* Marshal up the cap msg to the MDS */
1228 static void encode_cap_msg(struct ceph_msg *msg, struct cap_msg_args *arg)
1230 struct ceph_mds_caps *fc;
1232 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1234 dout("%s %s %llx %llx caps %s wanted %s dirty %s seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu xattr_ver %llu xattr_len %d\n",
1235 __func__, ceph_cap_op_name(arg->op), arg->cid, arg->ino,
1236 ceph_cap_string(arg->caps), ceph_cap_string(arg->wanted),
1237 ceph_cap_string(arg->dirty), arg->seq, arg->issue_seq,
1238 arg->flush_tid, arg->oldest_flush_tid, arg->mseq, arg->follows,
1239 arg->size, arg->max_size, arg->xattr_version,
1240 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1242 msg->hdr.version = cpu_to_le16(10);
1243 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1245 fc = msg->front.iov_base;
1246 memset(fc, 0, sizeof(*fc));
1248 fc->cap_id = cpu_to_le64(arg->cid);
1249 fc->op = cpu_to_le32(arg->op);
1250 fc->seq = cpu_to_le32(arg->seq);
1251 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1252 fc->migrate_seq = cpu_to_le32(arg->mseq);
1253 fc->caps = cpu_to_le32(arg->caps);
1254 fc->wanted = cpu_to_le32(arg->wanted);
1255 fc->dirty = cpu_to_le32(arg->dirty);
1256 fc->ino = cpu_to_le64(arg->ino);
1257 fc->snap_follows = cpu_to_le64(arg->follows);
1259 fc->size = cpu_to_le64(arg->size);
1260 fc->max_size = cpu_to_le64(arg->max_size);
1261 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1262 ceph_encode_timespec64(&fc->atime, &arg->atime);
1263 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1264 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1266 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1267 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1268 fc->mode = cpu_to_le32(arg->mode);
1270 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1271 if (arg->xattr_buf) {
1272 msg->middle = ceph_buffer_get(arg->xattr_buf);
1273 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1274 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1278 /* flock buffer size (version 2) */
1279 ceph_encode_32(&p, 0);
1280 /* inline version (version 4) */
1281 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1282 /* inline data size */
1283 ceph_encode_32(&p, 0);
1285 * osd_epoch_barrier (version 5)
1286 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1287 * case it was recently changed
1289 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1290 /* oldest_flush_tid (version 6) */
1291 ceph_encode_64(&p, arg->oldest_flush_tid);
1294 * caller_uid/caller_gid (version 7)
1296 * Currently, we don't properly track which caller dirtied the caps
1297 * last, and force a flush of them when there is a conflict. For now,
1298 * just set this to 0:0, to emulate how the MDS has worked up to now.
1300 ceph_encode_32(&p, 0);
1301 ceph_encode_32(&p, 0);
1303 /* pool namespace (version 8) (mds always ignores this) */
1304 ceph_encode_32(&p, 0);
1306 /* btime and change_attr (version 9) */
1307 ceph_encode_timespec64(p, &arg->btime);
1308 p += sizeof(struct ceph_timespec);
1309 ceph_encode_64(&p, arg->change_attr);
1311 /* Advisory flags (version 10) */
1312 ceph_encode_32(&p, arg->flags);
1316 * Queue cap releases when an inode is dropped from our cache.
1318 void __ceph_remove_caps(struct ceph_inode_info *ci)
1322 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1323 * may call __ceph_caps_issued_mask() on a freeing inode. */
1324 spin_lock(&ci->i_ceph_lock);
1325 p = rb_first(&ci->i_caps);
1327 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1329 ceph_remove_cap(cap, true);
1331 spin_unlock(&ci->i_ceph_lock);
1335 * Prepare to send a cap message to an MDS. Update the cap state, and populate
1336 * the arg struct with the parameters that will need to be sent. This should
1337 * be done under the i_ceph_lock to guard against changes to cap state.
1339 * Make note of max_size reported/requested from mds, revoked caps
1340 * that have now been implemented.
1342 static void __prep_cap(struct cap_msg_args *arg, struct ceph_cap *cap,
1343 int op, int flags, int used, int want, int retain,
1344 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1346 struct ceph_inode_info *ci = cap->ci;
1347 struct inode *inode = &ci->netfs.inode;
1350 lockdep_assert_held(&ci->i_ceph_lock);
1352 held = cap->issued | cap->implemented;
1353 revoking = cap->implemented & ~cap->issued;
1354 retain &= ~revoking;
1356 dout("%s %p cap %p session %p %s -> %s (revoking %s)\n",
1357 __func__, inode, cap, cap->session,
1358 ceph_cap_string(held), ceph_cap_string(held & retain),
1359 ceph_cap_string(revoking));
1360 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1362 ci->i_ceph_flags &= ~CEPH_I_FLUSH;
1364 cap->issued &= retain; /* drop bits we don't want */
1366 * Wake up any waiters on wanted -> needed transition. This is due to
1367 * the weird transition from buffered to sync IO... we need to flush
1368 * dirty pages _before_ allowing sync writes to avoid reordering.
1370 arg->wake = cap->implemented & ~cap->issued;
1371 cap->implemented &= cap->issued | used;
1372 cap->mds_wanted = want;
1374 arg->session = cap->session;
1375 arg->ino = ceph_vino(inode).ino;
1376 arg->cid = cap->cap_id;
1377 arg->follows = flushing ? ci->i_head_snapc->seq : 0;
1378 arg->flush_tid = flush_tid;
1379 arg->oldest_flush_tid = oldest_flush_tid;
1381 arg->size = i_size_read(inode);
1382 ci->i_reported_size = arg->size;
1383 arg->max_size = ci->i_wanted_max_size;
1384 if (cap == ci->i_auth_cap) {
1385 if (want & CEPH_CAP_ANY_FILE_WR)
1386 ci->i_requested_max_size = arg->max_size;
1388 ci->i_requested_max_size = 0;
1391 if (flushing & CEPH_CAP_XATTR_EXCL) {
1392 arg->old_xattr_buf = __ceph_build_xattrs_blob(ci);
1393 arg->xattr_version = ci->i_xattrs.version;
1394 arg->xattr_buf = ci->i_xattrs.blob;
1396 arg->xattr_buf = NULL;
1397 arg->old_xattr_buf = NULL;
1400 arg->mtime = inode->i_mtime;
1401 arg->atime = inode->i_atime;
1402 arg->ctime = inode->i_ctime;
1403 arg->btime = ci->i_btime;
1404 arg->change_attr = inode_peek_iversion_raw(inode);
1407 arg->caps = cap->implemented;
1409 arg->dirty = flushing;
1411 arg->seq = cap->seq;
1412 arg->issue_seq = cap->issue_seq;
1413 arg->mseq = cap->mseq;
1414 arg->time_warp_seq = ci->i_time_warp_seq;
1416 arg->uid = inode->i_uid;
1417 arg->gid = inode->i_gid;
1418 arg->mode = inode->i_mode;
1420 arg->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1421 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1422 !list_empty(&ci->i_cap_snaps)) {
1423 struct ceph_cap_snap *capsnap;
1424 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1425 if (capsnap->cap_flush.tid)
1427 if (capsnap->need_flush) {
1428 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1437 * Send a cap msg on the given inode.
1439 * Caller should hold snap_rwsem (read), s_mutex.
1441 static void __send_cap(struct cap_msg_args *arg, struct ceph_inode_info *ci)
1443 struct ceph_msg *msg;
1444 struct inode *inode = &ci->netfs.inode;
1446 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1448 pr_err("error allocating cap msg: ino (%llx.%llx) flushing %s tid %llu, requeuing cap.\n",
1449 ceph_vinop(inode), ceph_cap_string(arg->dirty),
1451 spin_lock(&ci->i_ceph_lock);
1452 __cap_delay_requeue(arg->session->s_mdsc, ci);
1453 spin_unlock(&ci->i_ceph_lock);
1457 encode_cap_msg(msg, arg);
1458 ceph_con_send(&arg->session->s_con, msg);
1459 ceph_buffer_put(arg->old_xattr_buf);
1461 wake_up_all(&ci->i_cap_wq);
1464 static inline int __send_flush_snap(struct inode *inode,
1465 struct ceph_mds_session *session,
1466 struct ceph_cap_snap *capsnap,
1467 u32 mseq, u64 oldest_flush_tid)
1469 struct cap_msg_args arg;
1470 struct ceph_msg *msg;
1472 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1476 arg.session = session;
1477 arg.ino = ceph_vino(inode).ino;
1479 arg.follows = capsnap->follows;
1480 arg.flush_tid = capsnap->cap_flush.tid;
1481 arg.oldest_flush_tid = oldest_flush_tid;
1483 arg.size = capsnap->size;
1485 arg.xattr_version = capsnap->xattr_version;
1486 arg.xattr_buf = capsnap->xattr_blob;
1487 arg.old_xattr_buf = NULL;
1489 arg.atime = capsnap->atime;
1490 arg.mtime = capsnap->mtime;
1491 arg.ctime = capsnap->ctime;
1492 arg.btime = capsnap->btime;
1493 arg.change_attr = capsnap->change_attr;
1495 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1496 arg.caps = capsnap->issued;
1498 arg.dirty = capsnap->dirty;
1503 arg.time_warp_seq = capsnap->time_warp_seq;
1505 arg.uid = capsnap->uid;
1506 arg.gid = capsnap->gid;
1507 arg.mode = capsnap->mode;
1509 arg.inline_data = capsnap->inline_data;
1513 encode_cap_msg(msg, &arg);
1514 ceph_con_send(&arg.session->s_con, msg);
1519 * When a snapshot is taken, clients accumulate dirty metadata on
1520 * inodes with capabilities in ceph_cap_snaps to describe the file
1521 * state at the time the snapshot was taken. This must be flushed
1522 * asynchronously back to the MDS once sync writes complete and dirty
1523 * data is written out.
1525 * Called under i_ceph_lock.
1527 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1528 struct ceph_mds_session *session)
1529 __releases(ci->i_ceph_lock)
1530 __acquires(ci->i_ceph_lock)
1532 struct inode *inode = &ci->netfs.inode;
1533 struct ceph_mds_client *mdsc = session->s_mdsc;
1534 struct ceph_cap_snap *capsnap;
1535 u64 oldest_flush_tid = 0;
1536 u64 first_tid = 1, last_tid = 0;
1538 dout("__flush_snaps %p session %p\n", inode, session);
1540 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1542 * we need to wait for sync writes to complete and for dirty
1543 * pages to be written out.
1545 if (capsnap->dirty_pages || capsnap->writing)
1548 /* should be removed by ceph_try_drop_cap_snap() */
1549 BUG_ON(!capsnap->need_flush);
1551 /* only flush each capsnap once */
1552 if (capsnap->cap_flush.tid > 0) {
1553 dout(" already flushed %p, skipping\n", capsnap);
1557 spin_lock(&mdsc->cap_dirty_lock);
1558 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1559 list_add_tail(&capsnap->cap_flush.g_list,
1560 &mdsc->cap_flush_list);
1561 if (oldest_flush_tid == 0)
1562 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1563 if (list_empty(&ci->i_flushing_item)) {
1564 list_add_tail(&ci->i_flushing_item,
1565 &session->s_cap_flushing);
1567 spin_unlock(&mdsc->cap_dirty_lock);
1569 list_add_tail(&capsnap->cap_flush.i_list,
1570 &ci->i_cap_flush_list);
1573 first_tid = capsnap->cap_flush.tid;
1574 last_tid = capsnap->cap_flush.tid;
1577 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1579 while (first_tid <= last_tid) {
1580 struct ceph_cap *cap = ci->i_auth_cap;
1581 struct ceph_cap_flush *cf = NULL, *iter;
1584 if (!(cap && cap->session == session)) {
1585 dout("__flush_snaps %p auth cap %p not mds%d, "
1586 "stop\n", inode, cap, session->s_mds);
1591 list_for_each_entry(iter, &ci->i_cap_flush_list, i_list) {
1592 if (iter->tid >= first_tid) {
1601 first_tid = cf->tid + 1;
1603 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1604 refcount_inc(&capsnap->nref);
1605 spin_unlock(&ci->i_ceph_lock);
1607 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1608 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1610 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1613 pr_err("__flush_snaps: error sending cap flushsnap, "
1614 "ino (%llx.%llx) tid %llu follows %llu\n",
1615 ceph_vinop(inode), cf->tid, capsnap->follows);
1618 ceph_put_cap_snap(capsnap);
1619 spin_lock(&ci->i_ceph_lock);
1623 void ceph_flush_snaps(struct ceph_inode_info *ci,
1624 struct ceph_mds_session **psession)
1626 struct inode *inode = &ci->netfs.inode;
1627 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1628 struct ceph_mds_session *session = NULL;
1631 dout("ceph_flush_snaps %p\n", inode);
1633 session = *psession;
1635 spin_lock(&ci->i_ceph_lock);
1636 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1637 dout(" no capsnap needs flush, doing nothing\n");
1640 if (!ci->i_auth_cap) {
1641 dout(" no auth cap (migrating?), doing nothing\n");
1645 mds = ci->i_auth_cap->session->s_mds;
1646 if (session && session->s_mds != mds) {
1647 dout(" oops, wrong session %p mutex\n", session);
1648 ceph_put_mds_session(session);
1652 spin_unlock(&ci->i_ceph_lock);
1653 mutex_lock(&mdsc->mutex);
1654 session = __ceph_lookup_mds_session(mdsc, mds);
1655 mutex_unlock(&mdsc->mutex);
1659 // make sure flushsnap messages are sent in proper order.
1660 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1661 __kick_flushing_caps(mdsc, session, ci, 0);
1663 __ceph_flush_snaps(ci, session);
1665 spin_unlock(&ci->i_ceph_lock);
1668 *psession = session;
1670 ceph_put_mds_session(session);
1671 /* we flushed them all; remove this inode from the queue */
1672 spin_lock(&mdsc->snap_flush_lock);
1673 list_del_init(&ci->i_snap_flush_item);
1674 spin_unlock(&mdsc->snap_flush_lock);
1678 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1679 * Caller is then responsible for calling __mark_inode_dirty with the
1680 * returned flags value.
1682 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1683 struct ceph_cap_flush **pcf)
1685 struct ceph_mds_client *mdsc =
1686 ceph_sb_to_client(ci->netfs.inode.i_sb)->mdsc;
1687 struct inode *inode = &ci->netfs.inode;
1688 int was = ci->i_dirty_caps;
1691 lockdep_assert_held(&ci->i_ceph_lock);
1693 if (!ci->i_auth_cap) {
1694 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1695 "but no auth cap (session was closed?)\n",
1696 inode, ceph_ino(inode), ceph_cap_string(mask));
1700 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->netfs.inode,
1701 ceph_cap_string(mask), ceph_cap_string(was),
1702 ceph_cap_string(was | mask));
1703 ci->i_dirty_caps |= mask;
1705 struct ceph_mds_session *session = ci->i_auth_cap->session;
1707 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1708 swap(ci->i_prealloc_cap_flush, *pcf);
1710 if (!ci->i_head_snapc) {
1711 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1712 ci->i_head_snapc = ceph_get_snap_context(
1713 ci->i_snap_realm->cached_context);
1715 dout(" inode %p now dirty snapc %p auth cap %p\n",
1716 &ci->netfs.inode, ci->i_head_snapc, ci->i_auth_cap);
1717 BUG_ON(!list_empty(&ci->i_dirty_item));
1718 spin_lock(&mdsc->cap_dirty_lock);
1719 list_add(&ci->i_dirty_item, &session->s_cap_dirty);
1720 spin_unlock(&mdsc->cap_dirty_lock);
1721 if (ci->i_flushing_caps == 0) {
1723 dirty |= I_DIRTY_SYNC;
1726 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1728 BUG_ON(list_empty(&ci->i_dirty_item));
1729 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1730 (mask & CEPH_CAP_FILE_BUFFER))
1731 dirty |= I_DIRTY_DATASYNC;
1732 __cap_delay_requeue(mdsc, ci);
1736 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1738 struct ceph_cap_flush *cf;
1740 cf = kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1744 cf->is_capsnap = false;
1748 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1751 kmem_cache_free(ceph_cap_flush_cachep, cf);
1754 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1756 if (!list_empty(&mdsc->cap_flush_list)) {
1757 struct ceph_cap_flush *cf =
1758 list_first_entry(&mdsc->cap_flush_list,
1759 struct ceph_cap_flush, g_list);
1766 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1767 * Return true if caller needs to wake up flush waiters.
1769 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1770 struct ceph_cap_flush *cf)
1772 struct ceph_cap_flush *prev;
1773 bool wake = cf->wake;
1775 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1776 prev = list_prev_entry(cf, g_list);
1780 list_del_init(&cf->g_list);
1784 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1785 struct ceph_cap_flush *cf)
1787 struct ceph_cap_flush *prev;
1788 bool wake = cf->wake;
1790 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1791 prev = list_prev_entry(cf, i_list);
1795 list_del_init(&cf->i_list);
1800 * Add dirty inode to the flushing list. Assigned a seq number so we
1801 * can wait for caps to flush without starving.
1803 * Called under i_ceph_lock. Returns the flush tid.
1805 static u64 __mark_caps_flushing(struct inode *inode,
1806 struct ceph_mds_session *session, bool wake,
1807 u64 *oldest_flush_tid)
1809 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1810 struct ceph_inode_info *ci = ceph_inode(inode);
1811 struct ceph_cap_flush *cf = NULL;
1814 lockdep_assert_held(&ci->i_ceph_lock);
1815 BUG_ON(ci->i_dirty_caps == 0);
1816 BUG_ON(list_empty(&ci->i_dirty_item));
1817 BUG_ON(!ci->i_prealloc_cap_flush);
1819 flushing = ci->i_dirty_caps;
1820 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1821 ceph_cap_string(flushing),
1822 ceph_cap_string(ci->i_flushing_caps),
1823 ceph_cap_string(ci->i_flushing_caps | flushing));
1824 ci->i_flushing_caps |= flushing;
1825 ci->i_dirty_caps = 0;
1826 dout(" inode %p now !dirty\n", inode);
1828 swap(cf, ci->i_prealloc_cap_flush);
1829 cf->caps = flushing;
1832 spin_lock(&mdsc->cap_dirty_lock);
1833 list_del_init(&ci->i_dirty_item);
1835 cf->tid = ++mdsc->last_cap_flush_tid;
1836 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1837 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1839 if (list_empty(&ci->i_flushing_item)) {
1840 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1841 mdsc->num_cap_flushing++;
1843 spin_unlock(&mdsc->cap_dirty_lock);
1845 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1851 * try to invalidate mapping pages without blocking.
1853 static int try_nonblocking_invalidate(struct inode *inode)
1854 __releases(ci->i_ceph_lock)
1855 __acquires(ci->i_ceph_lock)
1857 struct ceph_inode_info *ci = ceph_inode(inode);
1858 u32 invalidating_gen = ci->i_rdcache_gen;
1860 spin_unlock(&ci->i_ceph_lock);
1861 ceph_fscache_invalidate(inode, false);
1862 invalidate_mapping_pages(&inode->i_data, 0, -1);
1863 spin_lock(&ci->i_ceph_lock);
1865 if (inode->i_data.nrpages == 0 &&
1866 invalidating_gen == ci->i_rdcache_gen) {
1868 dout("try_nonblocking_invalidate %p success\n", inode);
1869 /* save any racing async invalidate some trouble */
1870 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1873 dout("try_nonblocking_invalidate %p failed\n", inode);
1877 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1879 loff_t size = i_size_read(&ci->netfs.inode);
1880 /* mds will adjust max size according to the reported size */
1881 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1883 if (size >= ci->i_max_size)
1885 /* half of previous max_size increment has been used */
1886 if (ci->i_max_size > ci->i_reported_size &&
1887 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1893 * Swiss army knife function to examine currently used and wanted
1894 * versus held caps. Release, flush, ack revoked caps to mds as
1897 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1898 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1901 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1902 struct ceph_mds_session *session)
1904 struct inode *inode = &ci->netfs.inode;
1905 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1906 struct ceph_cap *cap;
1907 u64 flush_tid, oldest_flush_tid;
1908 int file_wanted, used, cap_used;
1909 int issued, implemented, want, retain, revoking, flushing = 0;
1910 int mds = -1; /* keep track of how far we've gone through i_caps list
1911 to avoid an infinite loop on retry */
1913 bool queue_invalidate = false;
1914 bool tried_invalidate = false;
1915 bool queue_writeback = false;
1918 ceph_get_mds_session(session);
1920 spin_lock(&ci->i_ceph_lock);
1921 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
1922 /* Don't send messages until we get async create reply */
1923 spin_unlock(&ci->i_ceph_lock);
1924 ceph_put_mds_session(session);
1928 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1929 flags |= CHECK_CAPS_FLUSH;
1931 /* Caps wanted by virtue of active open files. */
1932 file_wanted = __ceph_caps_file_wanted(ci);
1934 /* Caps which have active references against them */
1935 used = __ceph_caps_used(ci);
1938 * "issued" represents the current caps that the MDS wants us to have.
1939 * "implemented" is the set that we have been granted, and includes the
1940 * ones that have not yet been returned to the MDS (the "revoking" set,
1941 * usually because they have outstanding references).
1943 issued = __ceph_caps_issued(ci, &implemented);
1944 revoking = implemented & ~issued;
1948 /* The ones we currently want to retain (may be adjusted below) */
1949 retain = file_wanted | used | CEPH_CAP_PIN;
1950 if (!mdsc->stopping && inode->i_nlink > 0) {
1952 retain |= CEPH_CAP_ANY; /* be greedy */
1953 } else if (S_ISDIR(inode->i_mode) &&
1954 (issued & CEPH_CAP_FILE_SHARED) &&
1955 __ceph_dir_is_complete(ci)) {
1957 * If a directory is complete, we want to keep
1958 * the exclusive cap. So that MDS does not end up
1959 * revoking the shared cap on every create/unlink
1962 if (IS_RDONLY(inode)) {
1963 want = CEPH_CAP_ANY_SHARED;
1965 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1970 retain |= CEPH_CAP_ANY_SHARED;
1972 * keep RD only if we didn't have the file open RW,
1973 * because then the mds would revoke it anyway to
1974 * journal max_size=0.
1976 if (ci->i_max_size == 0)
1977 retain |= CEPH_CAP_ANY_RD;
1981 dout("check_caps %llx.%llx file_want %s used %s dirty %s flushing %s"
1982 " issued %s revoking %s retain %s %s%s%s\n", ceph_vinop(inode),
1983 ceph_cap_string(file_wanted),
1984 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1985 ceph_cap_string(ci->i_flushing_caps),
1986 ceph_cap_string(issued), ceph_cap_string(revoking),
1987 ceph_cap_string(retain),
1988 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1989 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "",
1990 (flags & CHECK_CAPS_NOINVAL) ? " NOINVAL" : "");
1993 * If we no longer need to hold onto old our caps, and we may
1994 * have cached pages, but don't want them, then try to invalidate.
1995 * If we fail, it's because pages are locked.... try again later.
1997 if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
1998 S_ISREG(inode->i_mode) &&
1999 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
2000 inode->i_data.nrpages && /* have cached pages */
2001 (revoking & (CEPH_CAP_FILE_CACHE|
2002 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
2003 !tried_invalidate) {
2004 dout("check_caps trying to invalidate on %llx.%llx\n",
2006 if (try_nonblocking_invalidate(inode) < 0) {
2007 dout("check_caps queuing invalidate\n");
2008 queue_invalidate = true;
2009 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2011 tried_invalidate = true;
2015 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
2017 struct cap_msg_args arg;
2019 cap = rb_entry(p, struct ceph_cap, ci_node);
2021 /* avoid looping forever */
2022 if (mds >= cap->mds ||
2023 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
2027 * If we have an auth cap, we don't need to consider any
2028 * overlapping caps as used.
2031 if (ci->i_auth_cap && cap != ci->i_auth_cap)
2032 cap_used &= ~ci->i_auth_cap->issued;
2034 revoking = cap->implemented & ~cap->issued;
2035 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
2036 cap->mds, cap, ceph_cap_string(cap_used),
2037 ceph_cap_string(cap->issued),
2038 ceph_cap_string(cap->implemented),
2039 ceph_cap_string(revoking));
2041 if (cap == ci->i_auth_cap &&
2042 (cap->issued & CEPH_CAP_FILE_WR)) {
2043 /* request larger max_size from MDS? */
2044 if (ci->i_wanted_max_size > ci->i_max_size &&
2045 ci->i_wanted_max_size > ci->i_requested_max_size) {
2046 dout("requesting new max_size\n");
2050 /* approaching file_max? */
2051 if (__ceph_should_report_size(ci)) {
2052 dout("i_size approaching max_size\n");
2056 /* flush anything dirty? */
2057 if (cap == ci->i_auth_cap) {
2058 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
2059 dout("flushing dirty caps\n");
2062 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
2063 dout("flushing snap caps\n");
2068 /* completed revocation? going down and there are no caps? */
2070 if ((revoking & cap_used) == 0) {
2071 dout("completed revocation of %s\n",
2072 ceph_cap_string(cap->implemented & ~cap->issued));
2077 * If the "i_wrbuffer_ref" was increased by mmap or generic
2078 * cache write just before the ceph_check_caps() is called,
2079 * the Fb capability revoking will fail this time. Then we
2080 * must wait for the BDI's delayed work to flush the dirty
2081 * pages and to release the "i_wrbuffer_ref", which will cost
2082 * at most 5 seconds. That means the MDS needs to wait at
2083 * most 5 seconds to finished the Fb capability's revocation.
2085 * Let's queue a writeback for it.
2087 if (S_ISREG(inode->i_mode) && ci->i_wrbuffer_ref &&
2088 (revoking & CEPH_CAP_FILE_BUFFER))
2089 queue_writeback = true;
2092 /* want more caps from mds? */
2093 if (want & ~cap->mds_wanted) {
2094 if (want & ~(cap->mds_wanted | cap->issued))
2096 if (!__cap_is_valid(cap))
2100 /* things we might delay */
2101 if ((cap->issued & ~retain) == 0)
2102 continue; /* nope, all good */
2105 ceph_put_mds_session(session);
2106 session = ceph_get_mds_session(cap->session);
2108 /* kick flushing and flush snaps before sending normal
2110 if (cap == ci->i_auth_cap &&
2112 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2113 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2114 __kick_flushing_caps(mdsc, session, ci, 0);
2115 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2116 __ceph_flush_snaps(ci, session);
2121 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2122 flushing = ci->i_dirty_caps;
2123 flush_tid = __mark_caps_flushing(inode, session, false,
2125 if (flags & CHECK_CAPS_FLUSH &&
2126 list_empty(&session->s_cap_dirty))
2127 mflags |= CEPH_CLIENT_CAPS_SYNC;
2131 spin_lock(&mdsc->cap_dirty_lock);
2132 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2133 spin_unlock(&mdsc->cap_dirty_lock);
2136 mds = cap->mds; /* remember mds, so we don't repeat */
2138 __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2139 want, retain, flushing, flush_tid, oldest_flush_tid);
2141 spin_unlock(&ci->i_ceph_lock);
2142 __send_cap(&arg, ci);
2143 spin_lock(&ci->i_ceph_lock);
2145 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2148 /* periodically re-calculate caps wanted by open files */
2149 if (__ceph_is_any_real_caps(ci) &&
2150 list_empty(&ci->i_cap_delay_list) &&
2151 (file_wanted & ~CEPH_CAP_PIN) &&
2152 !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2153 __cap_delay_requeue(mdsc, ci);
2156 spin_unlock(&ci->i_ceph_lock);
2158 ceph_put_mds_session(session);
2159 if (queue_writeback)
2160 ceph_queue_writeback(inode);
2161 if (queue_invalidate)
2162 ceph_queue_invalidate(inode);
2166 * Try to flush dirty caps back to the auth mds.
2168 static int try_flush_caps(struct inode *inode, u64 *ptid)
2170 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2171 struct ceph_inode_info *ci = ceph_inode(inode);
2173 u64 flush_tid = 0, oldest_flush_tid = 0;
2175 spin_lock(&ci->i_ceph_lock);
2177 if (ci->i_dirty_caps && ci->i_auth_cap) {
2178 struct ceph_cap *cap = ci->i_auth_cap;
2179 struct cap_msg_args arg;
2180 struct ceph_mds_session *session = cap->session;
2182 if (session->s_state < CEPH_MDS_SESSION_OPEN) {
2183 spin_unlock(&ci->i_ceph_lock);
2187 if (ci->i_ceph_flags &
2188 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2189 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2190 __kick_flushing_caps(mdsc, session, ci, 0);
2191 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2192 __ceph_flush_snaps(ci, session);
2196 flushing = ci->i_dirty_caps;
2197 flush_tid = __mark_caps_flushing(inode, session, true,
2200 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2201 __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2202 (cap->issued | cap->implemented),
2203 flushing, flush_tid, oldest_flush_tid);
2204 spin_unlock(&ci->i_ceph_lock);
2206 __send_cap(&arg, ci);
2208 if (!list_empty(&ci->i_cap_flush_list)) {
2209 struct ceph_cap_flush *cf =
2210 list_last_entry(&ci->i_cap_flush_list,
2211 struct ceph_cap_flush, i_list);
2213 flush_tid = cf->tid;
2215 flushing = ci->i_flushing_caps;
2216 spin_unlock(&ci->i_ceph_lock);
2224 * Return true if we've flushed caps through the given flush_tid.
2226 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2228 struct ceph_inode_info *ci = ceph_inode(inode);
2231 spin_lock(&ci->i_ceph_lock);
2232 if (!list_empty(&ci->i_cap_flush_list)) {
2233 struct ceph_cap_flush * cf =
2234 list_first_entry(&ci->i_cap_flush_list,
2235 struct ceph_cap_flush, i_list);
2236 if (cf->tid <= flush_tid)
2239 spin_unlock(&ci->i_ceph_lock);
2244 * flush the mdlog and wait for any unsafe requests to complete.
2246 static int flush_mdlog_and_wait_inode_unsafe_requests(struct inode *inode)
2248 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2249 struct ceph_inode_info *ci = ceph_inode(inode);
2250 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2253 spin_lock(&ci->i_unsafe_lock);
2254 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2255 req1 = list_last_entry(&ci->i_unsafe_dirops,
2256 struct ceph_mds_request,
2258 ceph_mdsc_get_request(req1);
2260 if (!list_empty(&ci->i_unsafe_iops)) {
2261 req2 = list_last_entry(&ci->i_unsafe_iops,
2262 struct ceph_mds_request,
2263 r_unsafe_target_item);
2264 ceph_mdsc_get_request(req2);
2266 spin_unlock(&ci->i_unsafe_lock);
2269 * Trigger to flush the journal logs in all the relevant MDSes
2270 * manually, or in the worst case we must wait at most 5 seconds
2271 * to wait the journal logs to be flushed by the MDSes periodically.
2274 struct ceph_mds_request *req;
2275 struct ceph_mds_session **sessions;
2276 struct ceph_mds_session *s;
2277 unsigned int max_sessions;
2280 mutex_lock(&mdsc->mutex);
2281 max_sessions = mdsc->max_sessions;
2283 sessions = kcalloc(max_sessions, sizeof(s), GFP_KERNEL);
2285 mutex_unlock(&mdsc->mutex);
2290 spin_lock(&ci->i_unsafe_lock);
2292 list_for_each_entry(req, &ci->i_unsafe_dirops,
2293 r_unsafe_dir_item) {
2297 if (!sessions[s->s_mds]) {
2298 s = ceph_get_mds_session(s);
2299 sessions[s->s_mds] = s;
2304 list_for_each_entry(req, &ci->i_unsafe_iops,
2305 r_unsafe_target_item) {
2309 if (!sessions[s->s_mds]) {
2310 s = ceph_get_mds_session(s);
2311 sessions[s->s_mds] = s;
2315 spin_unlock(&ci->i_unsafe_lock);
2318 spin_lock(&ci->i_ceph_lock);
2319 if (ci->i_auth_cap) {
2320 s = ci->i_auth_cap->session;
2321 if (!sessions[s->s_mds])
2322 sessions[s->s_mds] = ceph_get_mds_session(s);
2324 spin_unlock(&ci->i_ceph_lock);
2325 mutex_unlock(&mdsc->mutex);
2327 /* send flush mdlog request to MDSes */
2328 for (i = 0; i < max_sessions; i++) {
2331 send_flush_mdlog(s);
2332 ceph_put_mds_session(s);
2338 dout("%s %p wait on tid %llu %llu\n", __func__,
2339 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2341 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2342 ceph_timeout_jiffies(req1->r_timeout));
2347 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2348 ceph_timeout_jiffies(req2->r_timeout));
2355 ceph_mdsc_put_request(req1);
2357 ceph_mdsc_put_request(req2);
2361 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2363 struct inode *inode = file->f_mapping->host;
2364 struct ceph_inode_info *ci = ceph_inode(inode);
2369 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2371 ret = file_write_and_wait_range(file, start, end);
2375 ret = ceph_wait_on_async_create(inode);
2379 dirty = try_flush_caps(inode, &flush_tid);
2380 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2382 err = flush_mdlog_and_wait_inode_unsafe_requests(inode);
2385 * only wait on non-file metadata writeback (the mds
2386 * can recover size and mtime, so we don't need to
2389 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2390 err = wait_event_interruptible(ci->i_cap_wq,
2391 caps_are_flushed(inode, flush_tid));
2397 err = file_check_and_advance_wb_err(file);
2401 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2406 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2407 * queue inode for flush but don't do so immediately, because we can
2408 * get by with fewer MDS messages if we wait for data writeback to
2411 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2413 struct ceph_inode_info *ci = ceph_inode(inode);
2417 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2419 dout("write_inode %p wait=%d\n", inode, wait);
2420 ceph_fscache_unpin_writeback(inode, wbc);
2422 err = ceph_wait_on_async_create(inode);
2425 dirty = try_flush_caps(inode, &flush_tid);
2427 err = wait_event_interruptible(ci->i_cap_wq,
2428 caps_are_flushed(inode, flush_tid));
2430 struct ceph_mds_client *mdsc =
2431 ceph_sb_to_client(inode->i_sb)->mdsc;
2433 spin_lock(&ci->i_ceph_lock);
2434 if (__ceph_caps_dirty(ci))
2435 __cap_delay_requeue_front(mdsc, ci);
2436 spin_unlock(&ci->i_ceph_lock);
2441 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2442 struct ceph_mds_session *session,
2443 struct ceph_inode_info *ci,
2444 u64 oldest_flush_tid)
2445 __releases(ci->i_ceph_lock)
2446 __acquires(ci->i_ceph_lock)
2448 struct inode *inode = &ci->netfs.inode;
2449 struct ceph_cap *cap;
2450 struct ceph_cap_flush *cf;
2453 u64 last_snap_flush = 0;
2455 /* Don't do anything until create reply comes in */
2456 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE)
2459 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2461 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2462 if (cf->is_capsnap) {
2463 last_snap_flush = cf->tid;
2468 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2469 if (cf->tid < first_tid)
2472 cap = ci->i_auth_cap;
2473 if (!(cap && cap->session == session)) {
2474 pr_err("%p auth cap %p not mds%d ???\n",
2475 inode, cap, session->s_mds);
2479 first_tid = cf->tid + 1;
2481 if (!cf->is_capsnap) {
2482 struct cap_msg_args arg;
2484 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2485 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2486 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2487 (cf->tid < last_snap_flush ?
2488 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2489 __ceph_caps_used(ci),
2490 __ceph_caps_wanted(ci),
2491 (cap->issued | cap->implemented),
2492 cf->caps, cf->tid, oldest_flush_tid);
2493 spin_unlock(&ci->i_ceph_lock);
2494 __send_cap(&arg, ci);
2496 struct ceph_cap_snap *capsnap =
2497 container_of(cf, struct ceph_cap_snap,
2499 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2500 inode, capsnap, cf->tid,
2501 ceph_cap_string(capsnap->dirty));
2503 refcount_inc(&capsnap->nref);
2504 spin_unlock(&ci->i_ceph_lock);
2506 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2509 pr_err("kick_flushing_caps: error sending "
2510 "cap flushsnap, ino (%llx.%llx) "
2511 "tid %llu follows %llu\n",
2512 ceph_vinop(inode), cf->tid,
2516 ceph_put_cap_snap(capsnap);
2519 spin_lock(&ci->i_ceph_lock);
2523 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2524 struct ceph_mds_session *session)
2526 struct ceph_inode_info *ci;
2527 struct ceph_cap *cap;
2528 u64 oldest_flush_tid;
2530 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2532 spin_lock(&mdsc->cap_dirty_lock);
2533 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2534 spin_unlock(&mdsc->cap_dirty_lock);
2536 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2537 spin_lock(&ci->i_ceph_lock);
2538 cap = ci->i_auth_cap;
2539 if (!(cap && cap->session == session)) {
2540 pr_err("%p auth cap %p not mds%d ???\n",
2541 &ci->netfs.inode, cap, session->s_mds);
2542 spin_unlock(&ci->i_ceph_lock);
2548 * if flushing caps were revoked, we re-send the cap flush
2549 * in client reconnect stage. This guarantees MDS * processes
2550 * the cap flush message before issuing the flushing caps to
2553 if ((cap->issued & ci->i_flushing_caps) !=
2554 ci->i_flushing_caps) {
2555 /* encode_caps_cb() also will reset these sequence
2556 * numbers. make sure sequence numbers in cap flush
2557 * message match later reconnect message */
2561 __kick_flushing_caps(mdsc, session, ci,
2564 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2567 spin_unlock(&ci->i_ceph_lock);
2571 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2572 struct ceph_mds_session *session)
2574 struct ceph_inode_info *ci;
2575 struct ceph_cap *cap;
2576 u64 oldest_flush_tid;
2578 lockdep_assert_held(&session->s_mutex);
2580 dout("kick_flushing_caps mds%d\n", session->s_mds);
2582 spin_lock(&mdsc->cap_dirty_lock);
2583 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2584 spin_unlock(&mdsc->cap_dirty_lock);
2586 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2587 spin_lock(&ci->i_ceph_lock);
2588 cap = ci->i_auth_cap;
2589 if (!(cap && cap->session == session)) {
2590 pr_err("%p auth cap %p not mds%d ???\n",
2591 &ci->netfs.inode, cap, session->s_mds);
2592 spin_unlock(&ci->i_ceph_lock);
2595 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2596 __kick_flushing_caps(mdsc, session, ci,
2599 spin_unlock(&ci->i_ceph_lock);
2603 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2604 struct ceph_inode_info *ci)
2606 struct ceph_mds_client *mdsc = session->s_mdsc;
2607 struct ceph_cap *cap = ci->i_auth_cap;
2609 lockdep_assert_held(&ci->i_ceph_lock);
2611 dout("%s %p flushing %s\n", __func__, &ci->netfs.inode,
2612 ceph_cap_string(ci->i_flushing_caps));
2614 if (!list_empty(&ci->i_cap_flush_list)) {
2615 u64 oldest_flush_tid;
2616 spin_lock(&mdsc->cap_dirty_lock);
2617 list_move_tail(&ci->i_flushing_item,
2618 &cap->session->s_cap_flushing);
2619 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2620 spin_unlock(&mdsc->cap_dirty_lock);
2622 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2628 * Take references to capabilities we hold, so that we don't release
2629 * them to the MDS prematurely.
2631 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2632 bool snap_rwsem_locked)
2634 lockdep_assert_held(&ci->i_ceph_lock);
2636 if (got & CEPH_CAP_PIN)
2638 if (got & CEPH_CAP_FILE_RD)
2640 if (got & CEPH_CAP_FILE_CACHE)
2641 ci->i_rdcache_ref++;
2642 if (got & CEPH_CAP_FILE_EXCL)
2644 if (got & CEPH_CAP_FILE_WR) {
2645 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2646 BUG_ON(!snap_rwsem_locked);
2647 ci->i_head_snapc = ceph_get_snap_context(
2648 ci->i_snap_realm->cached_context);
2652 if (got & CEPH_CAP_FILE_BUFFER) {
2653 if (ci->i_wb_ref == 0)
2654 ihold(&ci->netfs.inode);
2656 dout("%s %p wb %d -> %d (?)\n", __func__,
2657 &ci->netfs.inode, ci->i_wb_ref-1, ci->i_wb_ref);
2662 * Try to grab cap references. Specify those refs we @want, and the
2663 * minimal set we @need. Also include the larger offset we are writing
2664 * to (when applicable), and check against max_size here as well.
2665 * Note that caller is responsible for ensuring max_size increases are
2666 * requested from the MDS.
2668 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2669 * or a negative error code. There are 3 speical error codes:
2670 * -EAGAIN: need to sleep but non-blocking is specified
2671 * -EFBIG: ask caller to call check_max_size() and try again.
2672 * -EUCLEAN: ask caller to call ceph_renew_caps() and try again.
2675 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2676 NON_BLOCKING = (1 << 8),
2677 CHECK_FILELOCK = (1 << 9),
2680 static int try_get_cap_refs(struct inode *inode, int need, int want,
2681 loff_t endoff, int flags, int *got)
2683 struct ceph_inode_info *ci = ceph_inode(inode);
2684 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2686 int have, implemented;
2687 bool snap_rwsem_locked = false;
2689 dout("get_cap_refs %p need %s want %s\n", inode,
2690 ceph_cap_string(need), ceph_cap_string(want));
2693 spin_lock(&ci->i_ceph_lock);
2695 if ((flags & CHECK_FILELOCK) &&
2696 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2697 dout("try_get_cap_refs %p error filelock\n", inode);
2702 /* finish pending truncate */
2703 while (ci->i_truncate_pending) {
2704 spin_unlock(&ci->i_ceph_lock);
2705 if (snap_rwsem_locked) {
2706 up_read(&mdsc->snap_rwsem);
2707 snap_rwsem_locked = false;
2709 __ceph_do_pending_vmtruncate(inode);
2710 spin_lock(&ci->i_ceph_lock);
2713 have = __ceph_caps_issued(ci, &implemented);
2715 if (have & need & CEPH_CAP_FILE_WR) {
2716 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2717 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2718 inode, endoff, ci->i_max_size);
2719 if (endoff > ci->i_requested_max_size)
2720 ret = ci->i_auth_cap ? -EFBIG : -EUCLEAN;
2724 * If a sync write is in progress, we must wait, so that we
2725 * can get a final snapshot value for size+mtime.
2727 if (__ceph_have_pending_cap_snap(ci)) {
2728 dout("get_cap_refs %p cap_snap_pending\n", inode);
2733 if ((have & need) == need) {
2735 * Look at (implemented & ~have & not) so that we keep waiting
2736 * on transition from wanted -> needed caps. This is needed
2737 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2738 * going before a prior buffered writeback happens.
2740 * For RDCACHE|RD -> RD, there is not need to wait and we can
2741 * just exclude the revoking caps and force to sync read.
2743 int not = want & ~(have & need);
2744 int revoking = implemented & ~have;
2745 int exclude = revoking & not;
2746 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2747 inode, ceph_cap_string(have), ceph_cap_string(not),
2748 ceph_cap_string(revoking));
2749 if (!exclude || !(exclude & CEPH_CAP_FILE_BUFFER)) {
2750 if (!snap_rwsem_locked &&
2751 !ci->i_head_snapc &&
2752 (need & CEPH_CAP_FILE_WR)) {
2753 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2755 * we can not call down_read() when
2756 * task isn't in TASK_RUNNING state
2758 if (flags & NON_BLOCKING) {
2763 spin_unlock(&ci->i_ceph_lock);
2764 down_read(&mdsc->snap_rwsem);
2765 snap_rwsem_locked = true;
2768 snap_rwsem_locked = true;
2770 if ((have & want) == want)
2771 *got = need | (want & ~exclude);
2774 ceph_take_cap_refs(ci, *got, true);
2778 int session_readonly = false;
2780 if (ci->i_auth_cap &&
2781 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2782 struct ceph_mds_session *s = ci->i_auth_cap->session;
2783 spin_lock(&s->s_cap_lock);
2784 session_readonly = s->s_readonly;
2785 spin_unlock(&s->s_cap_lock);
2787 if (session_readonly) {
2788 dout("get_cap_refs %p need %s but mds%d readonly\n",
2789 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2794 if (ceph_inode_is_shutdown(inode)) {
2795 dout("get_cap_refs %p inode is shutdown\n", inode);
2799 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2800 if (need & ~mds_wanted) {
2801 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2802 inode, ceph_cap_string(need),
2803 ceph_cap_string(mds_wanted));
2808 dout("get_cap_refs %p have %s need %s\n", inode,
2809 ceph_cap_string(have), ceph_cap_string(need));
2813 __ceph_touch_fmode(ci, mdsc, flags);
2815 spin_unlock(&ci->i_ceph_lock);
2816 if (snap_rwsem_locked)
2817 up_read(&mdsc->snap_rwsem);
2820 ceph_update_cap_mis(&mdsc->metric);
2822 ceph_update_cap_hit(&mdsc->metric);
2824 dout("get_cap_refs %p ret %d got %s\n", inode,
2825 ret, ceph_cap_string(*got));
2830 * Check the offset we are writing up to against our current
2831 * max_size. If necessary, tell the MDS we want to write to
2834 static void check_max_size(struct inode *inode, loff_t endoff)
2836 struct ceph_inode_info *ci = ceph_inode(inode);
2839 /* do we need to explicitly request a larger max_size? */
2840 spin_lock(&ci->i_ceph_lock);
2841 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2842 dout("write %p at large endoff %llu, req max_size\n",
2844 ci->i_wanted_max_size = endoff;
2846 /* duplicate ceph_check_caps()'s logic */
2847 if (ci->i_auth_cap &&
2848 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2849 ci->i_wanted_max_size > ci->i_max_size &&
2850 ci->i_wanted_max_size > ci->i_requested_max_size)
2852 spin_unlock(&ci->i_ceph_lock);
2854 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2857 static inline int get_used_fmode(int caps)
2860 if (caps & CEPH_CAP_FILE_RD)
2861 fmode |= CEPH_FILE_MODE_RD;
2862 if (caps & CEPH_CAP_FILE_WR)
2863 fmode |= CEPH_FILE_MODE_WR;
2867 int ceph_try_get_caps(struct inode *inode, int need, int want,
2868 bool nonblock, int *got)
2872 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2873 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2874 CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2875 CEPH_CAP_ANY_DIR_OPS));
2877 ret = ceph_pool_perm_check(inode, need);
2882 flags = get_used_fmode(need | want);
2884 flags |= NON_BLOCKING;
2886 ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2887 /* three special error codes */
2888 if (ret == -EAGAIN || ret == -EFBIG || ret == -EUCLEAN)
2894 * Wait for caps, and take cap references. If we can't get a WR cap
2895 * due to a small max_size, make sure we check_max_size (and possibly
2896 * ask the mds) so we don't get hung up indefinitely.
2898 int ceph_get_caps(struct file *filp, int need, int want, loff_t endoff, int *got)
2900 struct ceph_file_info *fi = filp->private_data;
2901 struct inode *inode = file_inode(filp);
2902 struct ceph_inode_info *ci = ceph_inode(inode);
2903 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2904 int ret, _got, flags;
2906 ret = ceph_pool_perm_check(inode, need);
2910 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2911 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2914 flags = get_used_fmode(need | want);
2917 flags &= CEPH_FILE_MODE_MASK;
2918 if (atomic_read(&fi->num_locks))
2919 flags |= CHECK_FILELOCK;
2921 ret = try_get_cap_refs(inode, need, want, endoff,
2923 WARN_ON_ONCE(ret == -EAGAIN);
2925 struct ceph_mds_client *mdsc = fsc->mdsc;
2927 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2929 cw.ino = ceph_ino(inode);
2930 cw.tgid = current->tgid;
2934 spin_lock(&mdsc->caps_list_lock);
2935 list_add(&cw.list, &mdsc->cap_wait_list);
2936 spin_unlock(&mdsc->caps_list_lock);
2938 /* make sure used fmode not timeout */
2939 ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2940 add_wait_queue(&ci->i_cap_wq, &wait);
2942 flags |= NON_BLOCKING;
2943 while (!(ret = try_get_cap_refs(inode, need, want,
2944 endoff, flags, &_got))) {
2945 if (signal_pending(current)) {
2949 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2952 remove_wait_queue(&ci->i_cap_wq, &wait);
2953 ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
2955 spin_lock(&mdsc->caps_list_lock);
2957 spin_unlock(&mdsc->caps_list_lock);
2963 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2964 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2965 if (ret >= 0 && _got)
2966 ceph_put_cap_refs(ci, _got);
2971 if (ret == -EFBIG || ret == -EUCLEAN) {
2972 int ret2 = ceph_wait_on_async_create(inode);
2976 if (ret == -EFBIG) {
2977 check_max_size(inode, endoff);
2980 if (ret == -EUCLEAN) {
2981 /* session was killed, try renew caps */
2982 ret = ceph_renew_caps(inode, flags);
2989 if (S_ISREG(ci->netfs.inode.i_mode) &&
2990 ceph_has_inline_data(ci) &&
2991 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2992 i_size_read(inode) > 0) {
2994 find_get_page(inode->i_mapping, 0);
2996 bool uptodate = PageUptodate(page);
3003 * drop cap refs first because getattr while
3004 * holding * caps refs can cause deadlock.
3006 ceph_put_cap_refs(ci, _got);
3010 * getattr request will bring inline data into
3013 ret = __ceph_do_getattr(inode, NULL,
3014 CEPH_STAT_CAP_INLINE_DATA,
3027 * Take cap refs. Caller must already know we hold at least one ref
3028 * on the caps in question or we don't know this is safe.
3030 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
3032 spin_lock(&ci->i_ceph_lock);
3033 ceph_take_cap_refs(ci, caps, false);
3034 spin_unlock(&ci->i_ceph_lock);
3039 * drop cap_snap that is not associated with any snapshot.
3040 * we don't need to send FLUSHSNAP message for it.
3042 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
3043 struct ceph_cap_snap *capsnap)
3045 if (!capsnap->need_flush &&
3046 !capsnap->writing && !capsnap->dirty_pages) {
3047 dout("dropping cap_snap %p follows %llu\n",
3048 capsnap, capsnap->follows);
3049 BUG_ON(capsnap->cap_flush.tid > 0);
3050 ceph_put_snap_context(capsnap->context);
3051 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
3052 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3054 list_del(&capsnap->ci_item);
3055 ceph_put_cap_snap(capsnap);
3061 enum put_cap_refs_mode {
3062 PUT_CAP_REFS_SYNC = 0,
3063 PUT_CAP_REFS_NO_CHECK,
3070 * If we released the last ref on any given cap, call ceph_check_caps
3071 * to release (or schedule a release).
3073 * If we are releasing a WR cap (from a sync write), finalize any affected
3074 * cap_snap, and wake up any waiters.
3076 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
3077 enum put_cap_refs_mode mode)
3079 struct inode *inode = &ci->netfs.inode;
3080 int last = 0, put = 0, flushsnaps = 0, wake = 0;
3081 bool check_flushsnaps = false;
3083 spin_lock(&ci->i_ceph_lock);
3084 if (had & CEPH_CAP_PIN)
3086 if (had & CEPH_CAP_FILE_RD)
3087 if (--ci->i_rd_ref == 0)
3089 if (had & CEPH_CAP_FILE_CACHE)
3090 if (--ci->i_rdcache_ref == 0)
3092 if (had & CEPH_CAP_FILE_EXCL)
3093 if (--ci->i_fx_ref == 0)
3095 if (had & CEPH_CAP_FILE_BUFFER) {
3096 if (--ci->i_wb_ref == 0) {
3098 /* put the ref held by ceph_take_cap_refs() */
3100 check_flushsnaps = true;
3102 dout("put_cap_refs %p wb %d -> %d (?)\n",
3103 inode, ci->i_wb_ref+1, ci->i_wb_ref);
3105 if (had & CEPH_CAP_FILE_WR) {
3106 if (--ci->i_wr_ref == 0) {
3108 check_flushsnaps = true;
3109 if (ci->i_wrbuffer_ref_head == 0 &&
3110 ci->i_dirty_caps == 0 &&
3111 ci->i_flushing_caps == 0) {
3112 BUG_ON(!ci->i_head_snapc);
3113 ceph_put_snap_context(ci->i_head_snapc);
3114 ci->i_head_snapc = NULL;
3116 /* see comment in __ceph_remove_cap() */
3117 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3118 ceph_change_snap_realm(inode, NULL);
3121 if (check_flushsnaps && __ceph_have_pending_cap_snap(ci)) {
3122 struct ceph_cap_snap *capsnap =
3123 list_last_entry(&ci->i_cap_snaps,
3124 struct ceph_cap_snap,
3127 capsnap->writing = 0;
3128 if (ceph_try_drop_cap_snap(ci, capsnap))
3129 /* put the ref held by ceph_queue_cap_snap() */
3131 else if (__ceph_finish_cap_snap(ci, capsnap))
3135 spin_unlock(&ci->i_ceph_lock);
3137 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3138 last ? " last" : "", put ? " put" : "");
3141 case PUT_CAP_REFS_SYNC:
3143 ceph_check_caps(ci, 0, NULL);
3144 else if (flushsnaps)
3145 ceph_flush_snaps(ci, NULL);
3147 case PUT_CAP_REFS_ASYNC:
3149 ceph_queue_check_caps(inode);
3150 else if (flushsnaps)
3151 ceph_queue_flush_snaps(inode);
3157 wake_up_all(&ci->i_cap_wq);
3162 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3164 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_SYNC);
3167 void ceph_put_cap_refs_async(struct ceph_inode_info *ci, int had)
3169 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_ASYNC);
3172 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3174 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_NO_CHECK);
3178 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3179 * context. Adjust per-snap dirty page accounting as appropriate.
3180 * Once all dirty data for a cap_snap is flushed, flush snapped file
3181 * metadata back to the MDS. If we dropped the last ref, call
3184 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3185 struct ceph_snap_context *snapc)
3187 struct inode *inode = &ci->netfs.inode;
3188 struct ceph_cap_snap *capsnap = NULL, *iter;
3191 bool flush_snaps = false;
3192 bool complete_capsnap = false;
3194 spin_lock(&ci->i_ceph_lock);
3195 ci->i_wrbuffer_ref -= nr;
3196 if (ci->i_wrbuffer_ref == 0) {
3201 if (ci->i_head_snapc == snapc) {
3202 ci->i_wrbuffer_ref_head -= nr;
3203 if (ci->i_wrbuffer_ref_head == 0 &&
3204 ci->i_wr_ref == 0 &&
3205 ci->i_dirty_caps == 0 &&
3206 ci->i_flushing_caps == 0) {
3207 BUG_ON(!ci->i_head_snapc);
3208 ceph_put_snap_context(ci->i_head_snapc);
3209 ci->i_head_snapc = NULL;
3211 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3213 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3214 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3215 last ? " LAST" : "");
3217 list_for_each_entry(iter, &ci->i_cap_snaps, ci_item) {
3218 if (iter->context == snapc) {
3226 * The capsnap should already be removed when removing
3227 * auth cap in the case of a forced unmount.
3229 WARN_ON_ONCE(ci->i_auth_cap);
3233 capsnap->dirty_pages -= nr;
3234 if (capsnap->dirty_pages == 0) {
3235 complete_capsnap = true;
3236 if (!capsnap->writing) {
3237 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3240 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3245 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3246 " snap %lld %d/%d -> %d/%d %s%s\n",
3247 inode, capsnap, capsnap->context->seq,
3248 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3249 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3250 last ? " (wrbuffer last)" : "",
3251 complete_capsnap ? " (complete capsnap)" : "");
3255 spin_unlock(&ci->i_ceph_lock);
3258 ceph_check_caps(ci, 0, NULL);
3259 } else if (flush_snaps) {
3260 ceph_flush_snaps(ci, NULL);
3262 if (complete_capsnap)
3263 wake_up_all(&ci->i_cap_wq);
3270 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3272 static void invalidate_aliases(struct inode *inode)
3274 struct dentry *dn, *prev = NULL;
3276 dout("invalidate_aliases inode %p\n", inode);
3277 d_prune_aliases(inode);
3279 * For non-directory inode, d_find_alias() only returns
3280 * hashed dentry. After calling d_invalidate(), the
3281 * dentry becomes unhashed.
3283 * For directory inode, d_find_alias() can return
3284 * unhashed dentry. But directory inode should have
3285 * one alias at most.
3287 while ((dn = d_find_alias(inode))) {
3301 struct cap_extra_info {
3302 struct ceph_string *pool_ns;
3312 /* currently issued */
3314 struct timespec64 btime;
3318 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3319 * actually be a revocation if it specifies a smaller cap set.)
3321 * caller holds s_mutex and i_ceph_lock, we drop both.
3323 static void handle_cap_grant(struct inode *inode,
3324 struct ceph_mds_session *session,
3325 struct ceph_cap *cap,
3326 struct ceph_mds_caps *grant,
3327 struct ceph_buffer *xattr_buf,
3328 struct cap_extra_info *extra_info)
3329 __releases(ci->i_ceph_lock)
3330 __releases(session->s_mdsc->snap_rwsem)
3332 struct ceph_inode_info *ci = ceph_inode(inode);
3333 int seq = le32_to_cpu(grant->seq);
3334 int newcaps = le32_to_cpu(grant->caps);
3335 int used, wanted, dirty;
3336 u64 size = le64_to_cpu(grant->size);
3337 u64 max_size = le64_to_cpu(grant->max_size);
3338 unsigned char check_caps = 0;
3339 bool was_stale = cap->cap_gen < atomic_read(&session->s_cap_gen);
3341 bool writeback = false;
3342 bool queue_trunc = false;
3343 bool queue_invalidate = false;
3344 bool deleted_inode = false;
3345 bool fill_inline = false;
3347 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3348 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3349 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3350 i_size_read(inode));
3354 * If CACHE is being revoked, and we have no dirty buffers,
3355 * try to invalidate (once). (If there are dirty buffers, we
3356 * will invalidate _after_ writeback.)
3358 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3359 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3360 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3361 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3362 if (try_nonblocking_invalidate(inode)) {
3363 /* there were locked pages.. invalidate later
3364 in a separate thread. */
3365 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3366 queue_invalidate = true;
3367 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3373 cap->issued = cap->implemented = CEPH_CAP_PIN;
3376 * auth mds of the inode changed. we received the cap export message,
3377 * but still haven't received the cap import message. handle_cap_export
3378 * updated the new auth MDS' cap.
3380 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3381 * that was sent before the cap import message. So don't remove caps.
3383 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3384 WARN_ON(cap != ci->i_auth_cap);
3385 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3387 newcaps |= cap->issued;
3390 /* side effects now are allowed */
3391 cap->cap_gen = atomic_read(&session->s_cap_gen);
3394 __check_cap_issue(ci, cap, newcaps);
3396 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3398 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3399 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3400 umode_t mode = le32_to_cpu(grant->mode);
3402 if (inode_wrong_type(inode, mode))
3403 pr_warn_once("inode type changed! (ino %llx.%llx is 0%o, mds says 0%o)\n",
3404 ceph_vinop(inode), inode->i_mode, mode);
3406 inode->i_mode = mode;
3407 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3408 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3409 ci->i_btime = extra_info->btime;
3410 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3411 from_kuid(&init_user_ns, inode->i_uid),
3412 from_kgid(&init_user_ns, inode->i_gid));
3415 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3416 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3417 set_nlink(inode, le32_to_cpu(grant->nlink));
3418 if (inode->i_nlink == 0)
3419 deleted_inode = true;
3422 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3424 int len = le32_to_cpu(grant->xattr_len);
3425 u64 version = le64_to_cpu(grant->xattr_version);
3427 if (version > ci->i_xattrs.version) {
3428 dout(" got new xattrs v%llu on %p len %d\n",
3429 version, inode, len);
3430 if (ci->i_xattrs.blob)
3431 ceph_buffer_put(ci->i_xattrs.blob);
3432 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3433 ci->i_xattrs.version = version;
3434 ceph_forget_all_cached_acls(inode);
3435 ceph_security_invalidate_secctx(inode);
3439 if (newcaps & CEPH_CAP_ANY_RD) {
3440 struct timespec64 mtime, atime, ctime;
3441 /* ctime/mtime/atime? */
3442 ceph_decode_timespec64(&mtime, &grant->mtime);
3443 ceph_decode_timespec64(&atime, &grant->atime);
3444 ceph_decode_timespec64(&ctime, &grant->ctime);
3445 ceph_fill_file_time(inode, extra_info->issued,
3446 le32_to_cpu(grant->time_warp_seq),
3447 &ctime, &mtime, &atime);
3450 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3451 ci->i_files = extra_info->nfiles;
3452 ci->i_subdirs = extra_info->nsubdirs;
3455 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3456 /* file layout may have changed */
3457 s64 old_pool = ci->i_layout.pool_id;
3458 struct ceph_string *old_ns;
3460 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3461 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3462 lockdep_is_held(&ci->i_ceph_lock));
3463 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3465 if (ci->i_layout.pool_id != old_pool ||
3466 extra_info->pool_ns != old_ns)
3467 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3469 extra_info->pool_ns = old_ns;
3471 /* size/truncate_seq? */
3472 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3473 le32_to_cpu(grant->truncate_seq),
3474 le64_to_cpu(grant->truncate_size),
3478 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3479 if (max_size != ci->i_max_size) {
3480 dout("max_size %lld -> %llu\n",
3481 ci->i_max_size, max_size);
3482 ci->i_max_size = max_size;
3483 if (max_size >= ci->i_wanted_max_size) {
3484 ci->i_wanted_max_size = 0; /* reset */
3485 ci->i_requested_max_size = 0;
3491 /* check cap bits */
3492 wanted = __ceph_caps_wanted(ci);
3493 used = __ceph_caps_used(ci);
3494 dirty = __ceph_caps_dirty(ci);
3495 dout(" my wanted = %s, used = %s, dirty %s\n",
3496 ceph_cap_string(wanted),
3497 ceph_cap_string(used),
3498 ceph_cap_string(dirty));
3500 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3501 (wanted & ~(cap->mds_wanted | newcaps))) {
3503 * If mds is importing cap, prior cap messages that update
3504 * 'wanted' may get dropped by mds (migrate seq mismatch).
3506 * We don't send cap message to update 'wanted' if what we
3507 * want are already issued. If mds revokes caps, cap message
3508 * that releases caps also tells mds what we want. But if
3509 * caps got revoked by mds forcedly (session stale). We may
3510 * haven't told mds what we want.
3515 /* revocation, grant, or no-op? */
3516 if (cap->issued & ~newcaps) {
3517 int revoking = cap->issued & ~newcaps;
3519 dout("revocation: %s -> %s (revoking %s)\n",
3520 ceph_cap_string(cap->issued),
3521 ceph_cap_string(newcaps),
3522 ceph_cap_string(revoking));
3523 if (S_ISREG(inode->i_mode) &&
3524 (revoking & used & CEPH_CAP_FILE_BUFFER))
3525 writeback = true; /* initiate writeback; will delay ack */
3526 else if (queue_invalidate &&
3527 revoking == CEPH_CAP_FILE_CACHE &&
3528 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3529 ; /* do nothing yet, invalidation will be queued */
3530 else if (cap == ci->i_auth_cap)
3531 check_caps = 1; /* check auth cap only */
3533 check_caps = 2; /* check all caps */
3534 /* If there is new caps, try to wake up the waiters */
3535 if (~cap->issued & newcaps)
3537 cap->issued = newcaps;
3538 cap->implemented |= newcaps;
3539 } else if (cap->issued == newcaps) {
3540 dout("caps unchanged: %s -> %s\n",
3541 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3543 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3544 ceph_cap_string(newcaps));
3545 /* non-auth MDS is revoking the newly grant caps ? */
3546 if (cap == ci->i_auth_cap &&
3547 __ceph_caps_revoking_other(ci, cap, newcaps))
3550 cap->issued = newcaps;
3551 cap->implemented |= newcaps; /* add bits only, to
3552 * avoid stepping on a
3553 * pending revocation */
3556 BUG_ON(cap->issued & ~cap->implemented);
3558 if (extra_info->inline_version > 0 &&
3559 extra_info->inline_version >= ci->i_inline_version) {
3560 ci->i_inline_version = extra_info->inline_version;
3561 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3562 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3566 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3567 if (ci->i_auth_cap == cap) {
3568 if (newcaps & ~extra_info->issued)
3571 if (ci->i_requested_max_size > max_size ||
3572 !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3573 /* re-request max_size if necessary */
3574 ci->i_requested_max_size = 0;
3578 ceph_kick_flushing_inode_caps(session, ci);
3580 up_read(&session->s_mdsc->snap_rwsem);
3582 spin_unlock(&ci->i_ceph_lock);
3585 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3586 extra_info->inline_len);
3589 ceph_queue_vmtruncate(inode);
3593 * queue inode for writeback: we can't actually call
3594 * filemap_write_and_wait, etc. from message handler
3597 ceph_queue_writeback(inode);
3598 if (queue_invalidate)
3599 ceph_queue_invalidate(inode);
3601 invalidate_aliases(inode);
3603 wake_up_all(&ci->i_cap_wq);
3605 mutex_unlock(&session->s_mutex);
3606 if (check_caps == 1)
3607 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3609 else if (check_caps == 2)
3610 ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3614 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3615 * MDS has been safely committed.
3617 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3618 struct ceph_mds_caps *m,
3619 struct ceph_mds_session *session,
3620 struct ceph_cap *cap)
3621 __releases(ci->i_ceph_lock)
3623 struct ceph_inode_info *ci = ceph_inode(inode);
3624 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3625 struct ceph_cap_flush *cf, *tmp_cf;
3626 LIST_HEAD(to_remove);
3627 unsigned seq = le32_to_cpu(m->seq);
3628 int dirty = le32_to_cpu(m->dirty);
3631 bool wake_ci = false;
3632 bool wake_mdsc = false;
3634 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3635 /* Is this the one that was flushed? */
3636 if (cf->tid == flush_tid)
3639 /* Is this a capsnap? */
3643 if (cf->tid <= flush_tid) {
3645 * An earlier or current tid. The FLUSH_ACK should
3646 * represent a superset of this flush's caps.
3648 wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3649 list_add_tail(&cf->i_list, &to_remove);
3652 * This is a later one. Any caps in it are still dirty
3653 * so don't count them as cleaned.
3655 cleaned &= ~cf->caps;
3661 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3662 " flushing %s -> %s\n",
3663 inode, session->s_mds, seq, ceph_cap_string(dirty),
3664 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3665 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3667 if (list_empty(&to_remove) && !cleaned)
3670 ci->i_flushing_caps &= ~cleaned;
3672 spin_lock(&mdsc->cap_dirty_lock);
3674 list_for_each_entry(cf, &to_remove, i_list)
3675 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3677 if (ci->i_flushing_caps == 0) {
3678 if (list_empty(&ci->i_cap_flush_list)) {
3679 list_del_init(&ci->i_flushing_item);
3680 if (!list_empty(&session->s_cap_flushing)) {
3681 dout(" mds%d still flushing cap on %p\n",
3683 &list_first_entry(&session->s_cap_flushing,
3684 struct ceph_inode_info,
3685 i_flushing_item)->netfs.inode);
3688 mdsc->num_cap_flushing--;
3689 dout(" inode %p now !flushing\n", inode);
3691 if (ci->i_dirty_caps == 0) {
3692 dout(" inode %p now clean\n", inode);
3693 BUG_ON(!list_empty(&ci->i_dirty_item));
3695 if (ci->i_wr_ref == 0 &&
3696 ci->i_wrbuffer_ref_head == 0) {
3697 BUG_ON(!ci->i_head_snapc);
3698 ceph_put_snap_context(ci->i_head_snapc);
3699 ci->i_head_snapc = NULL;
3702 BUG_ON(list_empty(&ci->i_dirty_item));
3705 spin_unlock(&mdsc->cap_dirty_lock);
3708 spin_unlock(&ci->i_ceph_lock);
3710 while (!list_empty(&to_remove)) {
3711 cf = list_first_entry(&to_remove,
3712 struct ceph_cap_flush, i_list);
3713 list_del_init(&cf->i_list);
3714 if (!cf->is_capsnap)
3715 ceph_free_cap_flush(cf);
3719 wake_up_all(&ci->i_cap_wq);
3721 wake_up_all(&mdsc->cap_flushing_wq);
3726 void __ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3727 bool *wake_ci, bool *wake_mdsc)
3729 struct ceph_inode_info *ci = ceph_inode(inode);
3730 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3733 lockdep_assert_held(&ci->i_ceph_lock);
3735 dout("removing capsnap %p, inode %p ci %p\n", capsnap, inode, ci);
3737 list_del_init(&capsnap->ci_item);
3738 ret = __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3742 spin_lock(&mdsc->cap_dirty_lock);
3743 if (list_empty(&ci->i_cap_flush_list))
3744 list_del_init(&ci->i_flushing_item);
3746 ret = __detach_cap_flush_from_mdsc(mdsc, &capsnap->cap_flush);
3749 spin_unlock(&mdsc->cap_dirty_lock);
3752 void ceph_remove_capsnap(struct inode *inode, struct ceph_cap_snap *capsnap,
3753 bool *wake_ci, bool *wake_mdsc)
3755 struct ceph_inode_info *ci = ceph_inode(inode);
3757 lockdep_assert_held(&ci->i_ceph_lock);
3759 WARN_ON_ONCE(capsnap->dirty_pages || capsnap->writing);
3760 __ceph_remove_capsnap(inode, capsnap, wake_ci, wake_mdsc);
3764 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3765 * throw away our cap_snap.
3767 * Caller hold s_mutex.
3769 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3770 struct ceph_mds_caps *m,
3771 struct ceph_mds_session *session)
3773 struct ceph_inode_info *ci = ceph_inode(inode);
3774 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3775 u64 follows = le64_to_cpu(m->snap_follows);
3776 struct ceph_cap_snap *capsnap = NULL, *iter;
3777 bool wake_ci = false;
3778 bool wake_mdsc = false;
3780 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3781 inode, ci, session->s_mds, follows);
3783 spin_lock(&ci->i_ceph_lock);
3784 list_for_each_entry(iter, &ci->i_cap_snaps, ci_item) {
3785 if (iter->follows == follows) {
3786 if (iter->cap_flush.tid != flush_tid) {
3787 dout(" cap_snap %p follows %lld tid %lld !="
3788 " %lld\n", iter, follows,
3789 flush_tid, iter->cap_flush.tid);
3795 dout(" skipping cap_snap %p follows %lld\n",
3796 iter, iter->follows);
3800 ceph_remove_capsnap(inode, capsnap, &wake_ci, &wake_mdsc);
3801 spin_unlock(&ci->i_ceph_lock);
3804 ceph_put_snap_context(capsnap->context);
3805 ceph_put_cap_snap(capsnap);
3807 wake_up_all(&ci->i_cap_wq);
3809 wake_up_all(&mdsc->cap_flushing_wq);
3815 * Handle TRUNC from MDS, indicating file truncation.
3817 * caller hold s_mutex.
3819 static bool handle_cap_trunc(struct inode *inode,
3820 struct ceph_mds_caps *trunc,
3821 struct ceph_mds_session *session)
3823 struct ceph_inode_info *ci = ceph_inode(inode);
3824 int mds = session->s_mds;
3825 int seq = le32_to_cpu(trunc->seq);
3826 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3827 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3828 u64 size = le64_to_cpu(trunc->size);
3829 int implemented = 0;
3830 int dirty = __ceph_caps_dirty(ci);
3831 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3832 bool queue_trunc = false;
3834 lockdep_assert_held(&ci->i_ceph_lock);
3836 issued |= implemented | dirty;
3838 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3839 inode, mds, seq, truncate_size, truncate_seq);
3840 queue_trunc = ceph_fill_file_size(inode, issued,
3841 truncate_seq, truncate_size, size);
3846 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3847 * different one. If we are the most recent migration we've seen (as
3848 * indicated by mseq), make note of the migrating cap bits for the
3849 * duration (until we see the corresponding IMPORT).
3851 * caller holds s_mutex
3853 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3854 struct ceph_mds_cap_peer *ph,
3855 struct ceph_mds_session *session)
3857 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3858 struct ceph_mds_session *tsession = NULL;
3859 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3860 struct ceph_inode_info *ci = ceph_inode(inode);
3862 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3863 unsigned t_seq, t_mseq;
3865 int mds = session->s_mds;
3868 t_cap_id = le64_to_cpu(ph->cap_id);
3869 t_seq = le32_to_cpu(ph->seq);
3870 t_mseq = le32_to_cpu(ph->mseq);
3871 target = le32_to_cpu(ph->mds);
3873 t_cap_id = t_seq = t_mseq = 0;
3877 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3878 inode, ci, mds, mseq, target);
3880 down_read(&mdsc->snap_rwsem);
3881 spin_lock(&ci->i_ceph_lock);
3882 cap = __get_cap_for_mds(ci, mds);
3883 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3887 ceph_remove_cap(cap, false);
3892 * now we know we haven't received the cap import message yet
3893 * because the exported cap still exist.
3896 issued = cap->issued;
3897 if (issued != cap->implemented)
3898 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3899 "ino (%llx.%llx) mds%d seq %d mseq %d "
3900 "issued %s implemented %s\n",
3901 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3902 ceph_cap_string(issued),
3903 ceph_cap_string(cap->implemented));
3906 tcap = __get_cap_for_mds(ci, target);
3908 /* already have caps from the target */
3909 if (tcap->cap_id == t_cap_id &&
3910 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3911 dout(" updating import cap %p mds%d\n", tcap, target);
3912 tcap->cap_id = t_cap_id;
3913 tcap->seq = t_seq - 1;
3914 tcap->issue_seq = t_seq - 1;
3915 tcap->issued |= issued;
3916 tcap->implemented |= issued;
3917 if (cap == ci->i_auth_cap) {
3918 ci->i_auth_cap = tcap;
3919 change_auth_cap_ses(ci, tcap->session);
3922 ceph_remove_cap(cap, false);
3924 } else if (tsession) {
3925 /* add placeholder for the export tagert */
3926 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3928 ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3929 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3931 if (!list_empty(&ci->i_cap_flush_list) &&
3932 ci->i_auth_cap == tcap) {
3933 spin_lock(&mdsc->cap_dirty_lock);
3934 list_move_tail(&ci->i_flushing_item,
3935 &tcap->session->s_cap_flushing);
3936 spin_unlock(&mdsc->cap_dirty_lock);
3939 ceph_remove_cap(cap, false);
3943 spin_unlock(&ci->i_ceph_lock);
3944 up_read(&mdsc->snap_rwsem);
3945 mutex_unlock(&session->s_mutex);
3947 /* open target session */
3948 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3949 if (!IS_ERR(tsession)) {
3951 mutex_lock(&session->s_mutex);
3952 mutex_lock_nested(&tsession->s_mutex,
3953 SINGLE_DEPTH_NESTING);
3955 mutex_lock(&tsession->s_mutex);
3956 mutex_lock_nested(&session->s_mutex,
3957 SINGLE_DEPTH_NESTING);
3959 new_cap = ceph_get_cap(mdsc, NULL);
3964 mutex_lock(&session->s_mutex);
3969 spin_unlock(&ci->i_ceph_lock);
3970 up_read(&mdsc->snap_rwsem);
3971 mutex_unlock(&session->s_mutex);
3973 mutex_unlock(&tsession->s_mutex);
3974 ceph_put_mds_session(tsession);
3977 ceph_put_cap(mdsc, new_cap);
3981 * Handle cap IMPORT.
3983 * caller holds s_mutex. acquires i_ceph_lock
3985 static void handle_cap_import(struct ceph_mds_client *mdsc,
3986 struct inode *inode, struct ceph_mds_caps *im,
3987 struct ceph_mds_cap_peer *ph,
3988 struct ceph_mds_session *session,
3989 struct ceph_cap **target_cap, int *old_issued)
3991 struct ceph_inode_info *ci = ceph_inode(inode);
3992 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3993 int mds = session->s_mds;
3995 unsigned caps = le32_to_cpu(im->caps);
3996 unsigned wanted = le32_to_cpu(im->wanted);
3997 unsigned seq = le32_to_cpu(im->seq);
3998 unsigned mseq = le32_to_cpu(im->migrate_seq);
3999 u64 realmino = le64_to_cpu(im->realm);
4000 u64 cap_id = le64_to_cpu(im->cap_id);
4005 p_cap_id = le64_to_cpu(ph->cap_id);
4006 peer = le32_to_cpu(ph->mds);
4012 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
4013 inode, ci, mds, mseq, peer);
4015 cap = __get_cap_for_mds(ci, mds);
4018 spin_unlock(&ci->i_ceph_lock);
4019 new_cap = ceph_get_cap(mdsc, NULL);
4020 spin_lock(&ci->i_ceph_lock);
4026 ceph_put_cap(mdsc, new_cap);
4031 __ceph_caps_issued(ci, &issued);
4032 issued |= __ceph_caps_dirty(ci);
4034 ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
4035 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
4037 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
4038 if (ocap && ocap->cap_id == p_cap_id) {
4039 dout(" remove export cap %p mds%d flags %d\n",
4040 ocap, peer, ph->flags);
4041 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
4042 (ocap->seq != le32_to_cpu(ph->seq) ||
4043 ocap->mseq != le32_to_cpu(ph->mseq))) {
4044 pr_err_ratelimited("handle_cap_import: "
4045 "mismatched seq/mseq: ino (%llx.%llx) "
4046 "mds%d seq %d mseq %d importer mds%d "
4047 "has peer seq %d mseq %d\n",
4048 ceph_vinop(inode), peer, ocap->seq,
4049 ocap->mseq, mds, le32_to_cpu(ph->seq),
4050 le32_to_cpu(ph->mseq));
4052 ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
4055 *old_issued = issued;
4060 * Handle a caps message from the MDS.
4062 * Identify the appropriate session, inode, and call the right handler
4063 * based on the cap op.
4065 void ceph_handle_caps(struct ceph_mds_session *session,
4066 struct ceph_msg *msg)
4068 struct ceph_mds_client *mdsc = session->s_mdsc;
4069 struct inode *inode;
4070 struct ceph_inode_info *ci;
4071 struct ceph_cap *cap;
4072 struct ceph_mds_caps *h;
4073 struct ceph_mds_cap_peer *peer = NULL;
4074 struct ceph_snap_realm *realm = NULL;
4076 int msg_version = le16_to_cpu(msg->hdr.version);
4078 struct ceph_vino vino;
4080 size_t snaptrace_len;
4082 struct cap_extra_info extra_info = {};
4085 dout("handle_caps from mds%d\n", session->s_mds);
4088 end = msg->front.iov_base + msg->front.iov_len;
4089 if (msg->front.iov_len < sizeof(*h))
4091 h = msg->front.iov_base;
4092 op = le32_to_cpu(h->op);
4093 vino.ino = le64_to_cpu(h->ino);
4094 vino.snap = CEPH_NOSNAP;
4095 seq = le32_to_cpu(h->seq);
4096 mseq = le32_to_cpu(h->migrate_seq);
4099 snaptrace_len = le32_to_cpu(h->snap_trace_len);
4100 p = snaptrace + snaptrace_len;
4102 if (msg_version >= 2) {
4104 ceph_decode_32_safe(&p, end, flock_len, bad);
4105 if (p + flock_len > end)
4110 if (msg_version >= 3) {
4111 if (op == CEPH_CAP_OP_IMPORT) {
4112 if (p + sizeof(*peer) > end)
4116 } else if (op == CEPH_CAP_OP_EXPORT) {
4117 /* recorded in unused fields */
4118 peer = (void *)&h->size;
4122 if (msg_version >= 4) {
4123 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
4124 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
4125 if (p + extra_info.inline_len > end)
4127 extra_info.inline_data = p;
4128 p += extra_info.inline_len;
4131 if (msg_version >= 5) {
4132 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
4135 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
4136 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
4139 if (msg_version >= 8) {
4143 ceph_decode_skip_64(&p, end, bad); // flush_tid
4145 ceph_decode_skip_32(&p, end, bad); // caller_uid
4146 ceph_decode_skip_32(&p, end, bad); // caller_gid
4148 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4149 if (pool_ns_len > 0) {
4150 ceph_decode_need(&p, end, pool_ns_len, bad);
4151 extra_info.pool_ns =
4152 ceph_find_or_create_string(p, pool_ns_len);
4157 if (msg_version >= 9) {
4158 struct ceph_timespec *btime;
4160 if (p + sizeof(*btime) > end)
4163 ceph_decode_timespec64(&extra_info.btime, btime);
4164 p += sizeof(*btime);
4165 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4168 if (msg_version >= 11) {
4170 ceph_decode_skip_32(&p, end, bad); // flags
4172 extra_info.dirstat_valid = true;
4173 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4174 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4178 inode = ceph_find_inode(mdsc->fsc->sb, vino);
4179 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4182 mutex_lock(&session->s_mutex);
4183 inc_session_sequence(session);
4184 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4188 dout(" i don't have ino %llx\n", vino.ino);
4190 if (op == CEPH_CAP_OP_IMPORT) {
4191 cap = ceph_get_cap(mdsc, NULL);
4192 cap->cap_ino = vino.ino;
4193 cap->queue_release = 1;
4194 cap->cap_id = le64_to_cpu(h->cap_id);
4197 cap->issue_seq = seq;
4198 spin_lock(&session->s_cap_lock);
4199 __ceph_queue_cap_release(session, cap);
4200 spin_unlock(&session->s_cap_lock);
4202 goto flush_cap_releases;
4204 ci = ceph_inode(inode);
4206 /* these will work even if we don't have a cap yet */
4208 case CEPH_CAP_OP_FLUSHSNAP_ACK:
4209 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4213 case CEPH_CAP_OP_EXPORT:
4214 handle_cap_export(inode, h, peer, session);
4217 case CEPH_CAP_OP_IMPORT:
4219 if (snaptrace_len) {
4220 down_write(&mdsc->snap_rwsem);
4221 ceph_update_snap_trace(mdsc, snaptrace,
4222 snaptrace + snaptrace_len,
4224 downgrade_write(&mdsc->snap_rwsem);
4226 down_read(&mdsc->snap_rwsem);
4228 spin_lock(&ci->i_ceph_lock);
4229 handle_cap_import(mdsc, inode, h, peer, session,
4230 &cap, &extra_info.issued);
4231 handle_cap_grant(inode, session, cap,
4232 h, msg->middle, &extra_info);
4234 ceph_put_snap_realm(mdsc, realm);
4238 /* the rest require a cap */
4239 spin_lock(&ci->i_ceph_lock);
4240 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4242 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4243 inode, ceph_ino(inode), ceph_snap(inode),
4245 spin_unlock(&ci->i_ceph_lock);
4246 goto flush_cap_releases;
4249 /* note that each of these drops i_ceph_lock for us */
4251 case CEPH_CAP_OP_REVOKE:
4252 case CEPH_CAP_OP_GRANT:
4253 __ceph_caps_issued(ci, &extra_info.issued);
4254 extra_info.issued |= __ceph_caps_dirty(ci);
4255 handle_cap_grant(inode, session, cap,
4256 h, msg->middle, &extra_info);
4259 case CEPH_CAP_OP_FLUSH_ACK:
4260 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4264 case CEPH_CAP_OP_TRUNC:
4265 queue_trunc = handle_cap_trunc(inode, h, session);
4266 spin_unlock(&ci->i_ceph_lock);
4268 ceph_queue_vmtruncate(inode);
4272 spin_unlock(&ci->i_ceph_lock);
4273 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4274 ceph_cap_op_name(op));
4278 mutex_unlock(&session->s_mutex);
4282 ceph_put_string(extra_info.pool_ns);
4287 * send any cap release message to try to move things
4288 * along for the mds (who clearly thinks we still have this
4291 ceph_flush_cap_releases(mdsc, session);
4295 pr_err("ceph_handle_caps: corrupt message\n");
4301 * Delayed work handler to process end of delayed cap release LRU list.
4303 * If new caps are added to the list while processing it, these won't get
4304 * processed in this run. In this case, the ci->i_hold_caps_max will be
4305 * returned so that the work can be scheduled accordingly.
4307 unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4309 struct inode *inode;
4310 struct ceph_inode_info *ci;
4311 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
4312 unsigned long delay_max = opt->caps_wanted_delay_max * HZ;
4313 unsigned long loop_start = jiffies;
4314 unsigned long delay = 0;
4316 dout("check_delayed_caps\n");
4317 spin_lock(&mdsc->cap_delay_lock);
4318 while (!list_empty(&mdsc->cap_delay_list)) {
4319 ci = list_first_entry(&mdsc->cap_delay_list,
4320 struct ceph_inode_info,
4322 if (time_before(loop_start, ci->i_hold_caps_max - delay_max)) {
4323 dout("%s caps added recently. Exiting loop", __func__);
4324 delay = ci->i_hold_caps_max;
4327 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4328 time_before(jiffies, ci->i_hold_caps_max))
4330 list_del_init(&ci->i_cap_delay_list);
4332 inode = igrab(&ci->netfs.inode);
4334 spin_unlock(&mdsc->cap_delay_lock);
4335 dout("check_delayed_caps on %p\n", inode);
4336 ceph_check_caps(ci, 0, NULL);
4338 spin_lock(&mdsc->cap_delay_lock);
4341 spin_unlock(&mdsc->cap_delay_lock);
4347 * Flush all dirty caps to the mds
4349 static void flush_dirty_session_caps(struct ceph_mds_session *s)
4351 struct ceph_mds_client *mdsc = s->s_mdsc;
4352 struct ceph_inode_info *ci;
4353 struct inode *inode;
4355 dout("flush_dirty_caps\n");
4356 spin_lock(&mdsc->cap_dirty_lock);
4357 while (!list_empty(&s->s_cap_dirty)) {
4358 ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4360 inode = &ci->netfs.inode;
4362 dout("flush_dirty_caps %llx.%llx\n", ceph_vinop(inode));
4363 spin_unlock(&mdsc->cap_dirty_lock);
4364 ceph_wait_on_async_create(inode);
4365 ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4367 spin_lock(&mdsc->cap_dirty_lock);
4369 spin_unlock(&mdsc->cap_dirty_lock);
4370 dout("flush_dirty_caps done\n");
4373 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4375 ceph_mdsc_iterate_sessions(mdsc, flush_dirty_session_caps, true);
4378 void __ceph_touch_fmode(struct ceph_inode_info *ci,
4379 struct ceph_mds_client *mdsc, int fmode)
4381 unsigned long now = jiffies;
4382 if (fmode & CEPH_FILE_MODE_RD)
4383 ci->i_last_rd = now;
4384 if (fmode & CEPH_FILE_MODE_WR)
4385 ci->i_last_wr = now;
4386 /* queue periodic check */
4388 __ceph_is_any_real_caps(ci) &&
4389 list_empty(&ci->i_cap_delay_list))
4390 __cap_delay_requeue(mdsc, ci);
4393 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4395 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->netfs.inode.i_sb);
4396 int bits = (fmode << 1) | 1;
4397 bool already_opened = false;
4401 atomic64_inc(&mdsc->metric.opened_files);
4403 spin_lock(&ci->i_ceph_lock);
4404 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4406 * If any of the mode ref is larger than 0,
4407 * that means it has been already opened by
4408 * others. Just skip checking the PIN ref.
4410 if (i && ci->i_nr_by_mode[i])
4411 already_opened = true;
4413 if (bits & (1 << i))
4414 ci->i_nr_by_mode[i] += count;
4417 if (!already_opened)
4418 percpu_counter_inc(&mdsc->metric.opened_inodes);
4419 spin_unlock(&ci->i_ceph_lock);
4423 * Drop open file reference. If we were the last open file,
4424 * we may need to release capabilities to the MDS (or schedule
4425 * their delayed release).
4427 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4429 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->netfs.inode.i_sb);
4430 int bits = (fmode << 1) | 1;
4431 bool is_closed = true;
4435 atomic64_dec(&mdsc->metric.opened_files);
4437 spin_lock(&ci->i_ceph_lock);
4438 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4439 if (bits & (1 << i)) {
4440 BUG_ON(ci->i_nr_by_mode[i] < count);
4441 ci->i_nr_by_mode[i] -= count;
4445 * If any of the mode ref is not 0 after
4446 * decreased, that means it is still opened
4447 * by others. Just skip checking the PIN ref.
4449 if (i && ci->i_nr_by_mode[i])
4454 percpu_counter_dec(&mdsc->metric.opened_inodes);
4455 spin_unlock(&ci->i_ceph_lock);
4459 * For a soon-to-be unlinked file, drop the LINK caps. If it
4460 * looks like the link count will hit 0, drop any other caps (other
4461 * than PIN) we don't specifically want (due to the file still being
4464 int ceph_drop_caps_for_unlink(struct inode *inode)
4466 struct ceph_inode_info *ci = ceph_inode(inode);
4467 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4469 spin_lock(&ci->i_ceph_lock);
4470 if (inode->i_nlink == 1) {
4471 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4473 if (__ceph_caps_dirty(ci)) {
4474 struct ceph_mds_client *mdsc =
4475 ceph_inode_to_client(inode)->mdsc;
4476 __cap_delay_requeue_front(mdsc, ci);
4479 spin_unlock(&ci->i_ceph_lock);
4484 * Helpers for embedding cap and dentry lease releases into mds
4487 * @force is used by dentry_release (below) to force inclusion of a
4488 * record for the directory inode, even when there aren't any caps to
4491 int ceph_encode_inode_release(void **p, struct inode *inode,
4492 int mds, int drop, int unless, int force)
4494 struct ceph_inode_info *ci = ceph_inode(inode);
4495 struct ceph_cap *cap;
4496 struct ceph_mds_request_release *rel = *p;
4500 spin_lock(&ci->i_ceph_lock);
4501 used = __ceph_caps_used(ci);
4502 dirty = __ceph_caps_dirty(ci);
4504 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4505 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4506 ceph_cap_string(unless));
4508 /* only drop unused, clean caps */
4509 drop &= ~(used | dirty);
4511 cap = __get_cap_for_mds(ci, mds);
4512 if (cap && __cap_is_valid(cap)) {
4513 unless &= cap->issued;
4515 if (unless & CEPH_CAP_AUTH_EXCL)
4516 drop &= ~CEPH_CAP_AUTH_SHARED;
4517 if (unless & CEPH_CAP_LINK_EXCL)
4518 drop &= ~CEPH_CAP_LINK_SHARED;
4519 if (unless & CEPH_CAP_XATTR_EXCL)
4520 drop &= ~CEPH_CAP_XATTR_SHARED;
4521 if (unless & CEPH_CAP_FILE_EXCL)
4522 drop &= ~CEPH_CAP_FILE_SHARED;
4525 if (force || (cap->issued & drop)) {
4526 if (cap->issued & drop) {
4527 int wanted = __ceph_caps_wanted(ci);
4528 dout("encode_inode_release %p cap %p "
4529 "%s -> %s, wanted %s -> %s\n", inode, cap,
4530 ceph_cap_string(cap->issued),
4531 ceph_cap_string(cap->issued & ~drop),
4532 ceph_cap_string(cap->mds_wanted),
4533 ceph_cap_string(wanted));
4535 cap->issued &= ~drop;
4536 cap->implemented &= ~drop;
4537 cap->mds_wanted = wanted;
4538 if (cap == ci->i_auth_cap &&
4539 !(wanted & CEPH_CAP_ANY_FILE_WR))
4540 ci->i_requested_max_size = 0;
4542 dout("encode_inode_release %p cap %p %s"
4543 " (force)\n", inode, cap,
4544 ceph_cap_string(cap->issued));
4547 rel->ino = cpu_to_le64(ceph_ino(inode));
4548 rel->cap_id = cpu_to_le64(cap->cap_id);
4549 rel->seq = cpu_to_le32(cap->seq);
4550 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4551 rel->mseq = cpu_to_le32(cap->mseq);
4552 rel->caps = cpu_to_le32(cap->implemented);
4553 rel->wanted = cpu_to_le32(cap->mds_wanted);
4559 dout("encode_inode_release %p cap %p %s (noop)\n",
4560 inode, cap, ceph_cap_string(cap->issued));
4563 spin_unlock(&ci->i_ceph_lock);
4567 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4569 int mds, int drop, int unless)
4571 struct dentry *parent = NULL;
4572 struct ceph_mds_request_release *rel = *p;
4573 struct ceph_dentry_info *di = ceph_dentry(dentry);
4578 * force an record for the directory caps if we have a dentry lease.
4579 * this is racy (can't take i_ceph_lock and d_lock together), but it
4580 * doesn't have to be perfect; the mds will revoke anything we don't
4583 spin_lock(&dentry->d_lock);
4584 if (di->lease_session && di->lease_session->s_mds == mds)
4587 parent = dget(dentry->d_parent);
4588 dir = d_inode(parent);
4590 spin_unlock(&dentry->d_lock);
4592 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4595 spin_lock(&dentry->d_lock);
4596 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4597 dout("encode_dentry_release %p mds%d seq %d\n",
4598 dentry, mds, (int)di->lease_seq);
4599 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4600 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4601 *p += dentry->d_name.len;
4602 rel->dname_seq = cpu_to_le32(di->lease_seq);
4603 __ceph_mdsc_drop_dentry_lease(dentry);
4605 spin_unlock(&dentry->d_lock);
4609 static int remove_capsnaps(struct ceph_mds_client *mdsc, struct inode *inode)
4611 struct ceph_inode_info *ci = ceph_inode(inode);
4612 struct ceph_cap_snap *capsnap;
4613 int capsnap_release = 0;
4615 lockdep_assert_held(&ci->i_ceph_lock);
4617 dout("removing capsnaps, ci is %p, inode is %p\n", ci, inode);
4619 while (!list_empty(&ci->i_cap_snaps)) {
4620 capsnap = list_first_entry(&ci->i_cap_snaps,
4621 struct ceph_cap_snap, ci_item);
4622 __ceph_remove_capsnap(inode, capsnap, NULL, NULL);
4623 ceph_put_snap_context(capsnap->context);
4624 ceph_put_cap_snap(capsnap);
4627 wake_up_all(&ci->i_cap_wq);
4628 wake_up_all(&mdsc->cap_flushing_wq);
4629 return capsnap_release;
4632 int ceph_purge_inode_cap(struct inode *inode, struct ceph_cap *cap, bool *invalidate)
4634 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
4635 struct ceph_mds_client *mdsc = fsc->mdsc;
4636 struct ceph_inode_info *ci = ceph_inode(inode);
4638 bool dirty_dropped = false;
4641 lockdep_assert_held(&ci->i_ceph_lock);
4643 dout("removing cap %p, ci is %p, inode is %p\n",
4644 cap, ci, &ci->netfs.inode);
4646 is_auth = (cap == ci->i_auth_cap);
4647 __ceph_remove_cap(cap, false);
4649 struct ceph_cap_flush *cf;
4651 if (ceph_inode_is_shutdown(inode)) {
4652 if (inode->i_data.nrpages > 0)
4654 if (ci->i_wrbuffer_ref > 0)
4655 mapping_set_error(&inode->i_data, -EIO);
4658 spin_lock(&mdsc->cap_dirty_lock);
4660 /* trash all of the cap flushes for this inode */
4661 while (!list_empty(&ci->i_cap_flush_list)) {
4662 cf = list_first_entry(&ci->i_cap_flush_list,
4663 struct ceph_cap_flush, i_list);
4664 list_del_init(&cf->g_list);
4665 list_del_init(&cf->i_list);
4666 if (!cf->is_capsnap)
4667 ceph_free_cap_flush(cf);
4670 if (!list_empty(&ci->i_dirty_item)) {
4671 pr_warn_ratelimited(
4672 " dropping dirty %s state for %p %lld\n",
4673 ceph_cap_string(ci->i_dirty_caps),
4674 inode, ceph_ino(inode));
4675 ci->i_dirty_caps = 0;
4676 list_del_init(&ci->i_dirty_item);
4677 dirty_dropped = true;
4679 if (!list_empty(&ci->i_flushing_item)) {
4680 pr_warn_ratelimited(
4681 " dropping dirty+flushing %s state for %p %lld\n",
4682 ceph_cap_string(ci->i_flushing_caps),
4683 inode, ceph_ino(inode));
4684 ci->i_flushing_caps = 0;
4685 list_del_init(&ci->i_flushing_item);
4686 mdsc->num_cap_flushing--;
4687 dirty_dropped = true;
4689 spin_unlock(&mdsc->cap_dirty_lock);
4691 if (dirty_dropped) {
4692 mapping_set_error(inode->i_mapping, -EIO);
4694 if (ci->i_wrbuffer_ref_head == 0 &&
4695 ci->i_wr_ref == 0 &&
4696 ci->i_dirty_caps == 0 &&
4697 ci->i_flushing_caps == 0) {
4698 ceph_put_snap_context(ci->i_head_snapc);
4699 ci->i_head_snapc = NULL;
4703 if (atomic_read(&ci->i_filelock_ref) > 0) {
4704 /* make further file lock syscall return -EIO */
4705 ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
4706 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
4707 inode, ceph_ino(inode));
4710 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
4711 cf = ci->i_prealloc_cap_flush;
4712 ci->i_prealloc_cap_flush = NULL;
4713 if (!cf->is_capsnap)
4714 ceph_free_cap_flush(cf);
4717 if (!list_empty(&ci->i_cap_snaps))
4718 iputs = remove_capsnaps(mdsc, inode);