1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
5 #include <linux/kernel.h>
6 #include <linux/sched/signal.h>
7 #include <linux/slab.h>
8 #include <linux/vmalloc.h>
9 #include <linux/wait.h>
10 #include <linux/writeback.h>
11 #include <linux/iversion.h>
14 #include "mds_client.h"
16 #include <linux/ceph/decode.h>
17 #include <linux/ceph/messenger.h>
20 * Capability management
22 * The Ceph metadata servers control client access to inode metadata
23 * and file data by issuing capabilities, granting clients permission
24 * to read and/or write both inode field and file data to OSDs
25 * (storage nodes). Each capability consists of a set of bits
26 * indicating which operations are allowed.
28 * If the client holds a *_SHARED cap, the client has a coherent value
29 * that can be safely read from the cached inode.
31 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
32 * client is allowed to change inode attributes (e.g., file size,
33 * mtime), note its dirty state in the ceph_cap, and asynchronously
34 * flush that metadata change to the MDS.
36 * In the event of a conflicting operation (perhaps by another
37 * client), the MDS will revoke the conflicting client capabilities.
39 * In order for a client to cache an inode, it must hold a capability
40 * with at least one MDS server. When inodes are released, release
41 * notifications are batched and periodically sent en masse to the MDS
42 * cluster to release server state.
45 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc);
46 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
47 struct ceph_mds_session *session,
48 struct ceph_inode_info *ci,
49 u64 oldest_flush_tid);
52 * Generate readable cap strings for debugging output.
54 #define MAX_CAP_STR 20
55 static char cap_str[MAX_CAP_STR][40];
56 static DEFINE_SPINLOCK(cap_str_lock);
57 static int last_cap_str;
59 static char *gcap_string(char *s, int c)
61 if (c & CEPH_CAP_GSHARED)
63 if (c & CEPH_CAP_GEXCL)
65 if (c & CEPH_CAP_GCACHE)
71 if (c & CEPH_CAP_GBUFFER)
73 if (c & CEPH_CAP_GWREXTEND)
75 if (c & CEPH_CAP_GLAZYIO)
80 const char *ceph_cap_string(int caps)
86 spin_lock(&cap_str_lock);
88 if (last_cap_str == MAX_CAP_STR)
90 spin_unlock(&cap_str_lock);
94 if (caps & CEPH_CAP_PIN)
97 c = (caps >> CEPH_CAP_SAUTH) & 3;
100 s = gcap_string(s, c);
103 c = (caps >> CEPH_CAP_SLINK) & 3;
106 s = gcap_string(s, c);
109 c = (caps >> CEPH_CAP_SXATTR) & 3;
112 s = gcap_string(s, c);
115 c = caps >> CEPH_CAP_SFILE;
118 s = gcap_string(s, c);
127 void ceph_caps_init(struct ceph_mds_client *mdsc)
129 INIT_LIST_HEAD(&mdsc->caps_list);
130 spin_lock_init(&mdsc->caps_list_lock);
133 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
135 struct ceph_cap *cap;
137 spin_lock(&mdsc->caps_list_lock);
138 while (!list_empty(&mdsc->caps_list)) {
139 cap = list_first_entry(&mdsc->caps_list,
140 struct ceph_cap, caps_item);
141 list_del(&cap->caps_item);
142 kmem_cache_free(ceph_cap_cachep, cap);
144 mdsc->caps_total_count = 0;
145 mdsc->caps_avail_count = 0;
146 mdsc->caps_use_count = 0;
147 mdsc->caps_reserve_count = 0;
148 mdsc->caps_min_count = 0;
149 spin_unlock(&mdsc->caps_list_lock);
152 void ceph_adjust_caps_max_min(struct ceph_mds_client *mdsc,
153 struct ceph_mount_options *fsopt)
155 spin_lock(&mdsc->caps_list_lock);
156 mdsc->caps_min_count = fsopt->max_readdir;
157 if (mdsc->caps_min_count < 1024)
158 mdsc->caps_min_count = 1024;
159 mdsc->caps_use_max = fsopt->caps_max;
160 if (mdsc->caps_use_max > 0 &&
161 mdsc->caps_use_max < mdsc->caps_min_count)
162 mdsc->caps_use_max = mdsc->caps_min_count;
163 spin_unlock(&mdsc->caps_list_lock);
166 static void __ceph_unreserve_caps(struct ceph_mds_client *mdsc, int nr_caps)
168 struct ceph_cap *cap;
172 BUG_ON(mdsc->caps_reserve_count < nr_caps);
173 mdsc->caps_reserve_count -= nr_caps;
174 if (mdsc->caps_avail_count >=
175 mdsc->caps_reserve_count + mdsc->caps_min_count) {
176 mdsc->caps_total_count -= nr_caps;
177 for (i = 0; i < nr_caps; i++) {
178 cap = list_first_entry(&mdsc->caps_list,
179 struct ceph_cap, caps_item);
180 list_del(&cap->caps_item);
181 kmem_cache_free(ceph_cap_cachep, cap);
184 mdsc->caps_avail_count += nr_caps;
187 dout("%s: caps %d = %d used + %d resv + %d avail\n",
189 mdsc->caps_total_count, mdsc->caps_use_count,
190 mdsc->caps_reserve_count, mdsc->caps_avail_count);
191 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
192 mdsc->caps_reserve_count +
193 mdsc->caps_avail_count);
198 * Called under mdsc->mutex.
200 int ceph_reserve_caps(struct ceph_mds_client *mdsc,
201 struct ceph_cap_reservation *ctx, int need)
204 struct ceph_cap *cap;
209 bool trimmed = false;
210 struct ceph_mds_session *s;
213 dout("reserve caps ctx=%p need=%d\n", ctx, need);
215 /* first reserve any caps that are already allocated */
216 spin_lock(&mdsc->caps_list_lock);
217 if (mdsc->caps_avail_count >= need)
220 have = mdsc->caps_avail_count;
221 mdsc->caps_avail_count -= have;
222 mdsc->caps_reserve_count += have;
223 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
224 mdsc->caps_reserve_count +
225 mdsc->caps_avail_count);
226 spin_unlock(&mdsc->caps_list_lock);
228 for (i = have; i < need; ) {
229 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
231 list_add(&cap->caps_item, &newcaps);
238 for (j = 0; j < mdsc->max_sessions; j++) {
239 s = __ceph_lookup_mds_session(mdsc, j);
242 mutex_unlock(&mdsc->mutex);
244 mutex_lock(&s->s_mutex);
245 max_caps = s->s_nr_caps - (need - i);
246 ceph_trim_caps(mdsc, s, max_caps);
247 mutex_unlock(&s->s_mutex);
249 ceph_put_mds_session(s);
250 mutex_lock(&mdsc->mutex);
254 spin_lock(&mdsc->caps_list_lock);
255 if (mdsc->caps_avail_count) {
257 if (mdsc->caps_avail_count >= need - i)
258 more_have = need - i;
260 more_have = mdsc->caps_avail_count;
264 mdsc->caps_avail_count -= more_have;
265 mdsc->caps_reserve_count += more_have;
268 spin_unlock(&mdsc->caps_list_lock);
273 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
274 ctx, need, have + alloc);
280 BUG_ON(have + alloc != need);
285 spin_lock(&mdsc->caps_list_lock);
286 mdsc->caps_total_count += alloc;
287 mdsc->caps_reserve_count += alloc;
288 list_splice(&newcaps, &mdsc->caps_list);
290 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
291 mdsc->caps_reserve_count +
292 mdsc->caps_avail_count);
295 __ceph_unreserve_caps(mdsc, have + alloc);
297 spin_unlock(&mdsc->caps_list_lock);
299 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
300 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
301 mdsc->caps_reserve_count, mdsc->caps_avail_count);
305 void ceph_unreserve_caps(struct ceph_mds_client *mdsc,
306 struct ceph_cap_reservation *ctx)
308 bool reclaim = false;
312 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
313 spin_lock(&mdsc->caps_list_lock);
314 __ceph_unreserve_caps(mdsc, ctx->count);
317 if (mdsc->caps_use_max > 0 &&
318 mdsc->caps_use_count > mdsc->caps_use_max)
320 spin_unlock(&mdsc->caps_list_lock);
323 ceph_reclaim_caps_nr(mdsc, ctx->used);
326 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
327 struct ceph_cap_reservation *ctx)
329 struct ceph_cap *cap = NULL;
331 /* temporary, until we do something about cap import/export */
333 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
335 spin_lock(&mdsc->caps_list_lock);
336 mdsc->caps_use_count++;
337 mdsc->caps_total_count++;
338 spin_unlock(&mdsc->caps_list_lock);
340 spin_lock(&mdsc->caps_list_lock);
341 if (mdsc->caps_avail_count) {
342 BUG_ON(list_empty(&mdsc->caps_list));
344 mdsc->caps_avail_count--;
345 mdsc->caps_use_count++;
346 cap = list_first_entry(&mdsc->caps_list,
347 struct ceph_cap, caps_item);
348 list_del(&cap->caps_item);
350 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
351 mdsc->caps_reserve_count + mdsc->caps_avail_count);
353 spin_unlock(&mdsc->caps_list_lock);
359 spin_lock(&mdsc->caps_list_lock);
360 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
361 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
362 mdsc->caps_reserve_count, mdsc->caps_avail_count);
364 BUG_ON(ctx->count > mdsc->caps_reserve_count);
365 BUG_ON(list_empty(&mdsc->caps_list));
369 mdsc->caps_reserve_count--;
370 mdsc->caps_use_count++;
372 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
373 list_del(&cap->caps_item);
375 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
376 mdsc->caps_reserve_count + mdsc->caps_avail_count);
377 spin_unlock(&mdsc->caps_list_lock);
381 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
383 spin_lock(&mdsc->caps_list_lock);
384 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
385 cap, mdsc->caps_total_count, mdsc->caps_use_count,
386 mdsc->caps_reserve_count, mdsc->caps_avail_count);
387 mdsc->caps_use_count--;
389 * Keep some preallocated caps around (ceph_min_count), to
390 * avoid lots of free/alloc churn.
392 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
393 mdsc->caps_min_count) {
394 mdsc->caps_total_count--;
395 kmem_cache_free(ceph_cap_cachep, cap);
397 mdsc->caps_avail_count++;
398 list_add(&cap->caps_item, &mdsc->caps_list);
401 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
402 mdsc->caps_reserve_count + mdsc->caps_avail_count);
403 spin_unlock(&mdsc->caps_list_lock);
406 void ceph_reservation_status(struct ceph_fs_client *fsc,
407 int *total, int *avail, int *used, int *reserved,
410 struct ceph_mds_client *mdsc = fsc->mdsc;
412 spin_lock(&mdsc->caps_list_lock);
415 *total = mdsc->caps_total_count;
417 *avail = mdsc->caps_avail_count;
419 *used = mdsc->caps_use_count;
421 *reserved = mdsc->caps_reserve_count;
423 *min = mdsc->caps_min_count;
425 spin_unlock(&mdsc->caps_list_lock);
429 * Find ceph_cap for given mds, if any.
431 * Called with i_ceph_lock held.
433 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
435 struct ceph_cap *cap;
436 struct rb_node *n = ci->i_caps.rb_node;
439 cap = rb_entry(n, struct ceph_cap, ci_node);
442 else if (mds > cap->mds)
450 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
452 struct ceph_cap *cap;
454 spin_lock(&ci->i_ceph_lock);
455 cap = __get_cap_for_mds(ci, mds);
456 spin_unlock(&ci->i_ceph_lock);
461 * Called under i_ceph_lock.
463 static void __insert_cap_node(struct ceph_inode_info *ci,
464 struct ceph_cap *new)
466 struct rb_node **p = &ci->i_caps.rb_node;
467 struct rb_node *parent = NULL;
468 struct ceph_cap *cap = NULL;
472 cap = rb_entry(parent, struct ceph_cap, ci_node);
473 if (new->mds < cap->mds)
475 else if (new->mds > cap->mds)
481 rb_link_node(&new->ci_node, parent, p);
482 rb_insert_color(&new->ci_node, &ci->i_caps);
486 * (re)set cap hold timeouts, which control the delayed release
487 * of unused caps back to the MDS. Should be called on cap use.
489 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
490 struct ceph_inode_info *ci)
492 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
493 ci->i_hold_caps_max = round_jiffies(jiffies +
494 opt->caps_wanted_delay_max * HZ);
495 dout("__cap_set_timeouts %p %lu\n", &ci->vfs_inode,
496 ci->i_hold_caps_max - jiffies);
500 * (Re)queue cap at the end of the delayed cap release list.
502 * If I_FLUSH is set, leave the inode at the front of the list.
504 * Caller holds i_ceph_lock
505 * -> we take mdsc->cap_delay_lock
507 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
508 struct ceph_inode_info *ci)
510 dout("__cap_delay_requeue %p flags 0x%lx at %lu\n", &ci->vfs_inode,
511 ci->i_ceph_flags, ci->i_hold_caps_max);
512 if (!mdsc->stopping) {
513 spin_lock(&mdsc->cap_delay_lock);
514 if (!list_empty(&ci->i_cap_delay_list)) {
515 if (ci->i_ceph_flags & CEPH_I_FLUSH)
517 list_del_init(&ci->i_cap_delay_list);
519 __cap_set_timeouts(mdsc, ci);
520 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
522 spin_unlock(&mdsc->cap_delay_lock);
527 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
528 * indicating we should send a cap message to flush dirty metadata
529 * asap, and move to the front of the delayed cap list.
531 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
532 struct ceph_inode_info *ci)
534 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
535 spin_lock(&mdsc->cap_delay_lock);
536 ci->i_ceph_flags |= CEPH_I_FLUSH;
537 if (!list_empty(&ci->i_cap_delay_list))
538 list_del_init(&ci->i_cap_delay_list);
539 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
540 spin_unlock(&mdsc->cap_delay_lock);
544 * Cancel delayed work on cap.
546 * Caller must hold i_ceph_lock.
548 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
549 struct ceph_inode_info *ci)
551 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
552 if (list_empty(&ci->i_cap_delay_list))
554 spin_lock(&mdsc->cap_delay_lock);
555 list_del_init(&ci->i_cap_delay_list);
556 spin_unlock(&mdsc->cap_delay_lock);
559 /* Common issue checks for add_cap, handle_cap_grant. */
560 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
563 unsigned had = __ceph_caps_issued(ci, NULL);
565 lockdep_assert_held(&ci->i_ceph_lock);
568 * Each time we receive FILE_CACHE anew, we increment
571 if (S_ISREG(ci->vfs_inode.i_mode) &&
572 (issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
573 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
578 * If FILE_SHARED is newly issued, mark dir not complete. We don't
579 * know what happened to this directory while we didn't have the cap.
580 * If FILE_SHARED is being revoked, also mark dir not complete. It
581 * stops on-going cached readdir.
583 if ((issued & CEPH_CAP_FILE_SHARED) != (had & CEPH_CAP_FILE_SHARED)) {
584 if (issued & CEPH_CAP_FILE_SHARED)
585 atomic_inc(&ci->i_shared_gen);
586 if (S_ISDIR(ci->vfs_inode.i_mode)) {
587 dout(" marking %p NOT complete\n", &ci->vfs_inode);
588 __ceph_dir_clear_complete(ci);
592 /* Wipe saved layout if we're losing DIR_CREATE caps */
593 if (S_ISDIR(ci->vfs_inode.i_mode) && (had & CEPH_CAP_DIR_CREATE) &&
594 !(issued & CEPH_CAP_DIR_CREATE)) {
595 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
596 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
601 * change_auth_cap_ses - move inode to appropriate lists when auth caps change
602 * @ci: inode to be moved
603 * @session: new auth caps session
605 static void change_auth_cap_ses(struct ceph_inode_info *ci,
606 struct ceph_mds_session *session)
608 lockdep_assert_held(&ci->i_ceph_lock);
610 if (list_empty(&ci->i_dirty_item) && list_empty(&ci->i_flushing_item))
613 spin_lock(&session->s_mdsc->cap_dirty_lock);
614 if (!list_empty(&ci->i_dirty_item))
615 list_move(&ci->i_dirty_item, &session->s_cap_dirty);
616 if (!list_empty(&ci->i_flushing_item))
617 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
618 spin_unlock(&session->s_mdsc->cap_dirty_lock);
622 * Add a capability under the given MDS session.
624 * Caller should hold session snap_rwsem (read) and ci->i_ceph_lock
626 * @fmode is the open file mode, if we are opening a file, otherwise
627 * it is < 0. (This is so we can atomically add the cap and add an
628 * open file reference to it.)
630 void ceph_add_cap(struct inode *inode,
631 struct ceph_mds_session *session, u64 cap_id,
632 unsigned issued, unsigned wanted,
633 unsigned seq, unsigned mseq, u64 realmino, int flags,
634 struct ceph_cap **new_cap)
636 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
637 struct ceph_inode_info *ci = ceph_inode(inode);
638 struct ceph_cap *cap;
639 int mds = session->s_mds;
643 lockdep_assert_held(&ci->i_ceph_lock);
645 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
646 session->s_mds, cap_id, ceph_cap_string(issued), seq);
648 gen = atomic_read(&session->s_cap_gen);
650 cap = __get_cap_for_mds(ci, mds);
656 cap->implemented = 0;
662 __insert_cap_node(ci, cap);
664 /* add to session cap list */
665 cap->session = session;
666 spin_lock(&session->s_cap_lock);
667 list_add_tail(&cap->session_caps, &session->s_caps);
668 session->s_nr_caps++;
669 atomic64_inc(&mdsc->metric.total_caps);
670 spin_unlock(&session->s_cap_lock);
672 spin_lock(&session->s_cap_lock);
673 list_move_tail(&cap->session_caps, &session->s_caps);
674 spin_unlock(&session->s_cap_lock);
676 if (cap->cap_gen < gen)
677 cap->issued = cap->implemented = CEPH_CAP_PIN;
680 * auth mds of the inode changed. we received the cap export
681 * message, but still haven't received the cap import message.
682 * handle_cap_export() updated the new auth MDS' cap.
684 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
685 * a message that was send before the cap import message. So
688 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
689 WARN_ON(cap != ci->i_auth_cap);
690 WARN_ON(cap->cap_id != cap_id);
693 issued |= cap->issued;
694 flags |= CEPH_CAP_FLAG_AUTH;
698 if (!ci->i_snap_realm ||
699 ((flags & CEPH_CAP_FLAG_AUTH) &&
700 realmino != (u64)-1 && ci->i_snap_realm->ino != realmino)) {
702 * add this inode to the appropriate snap realm
704 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
707 ceph_change_snap_realm(inode, realm);
709 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
715 __check_cap_issue(ci, cap, issued);
718 * If we are issued caps we don't want, or the mds' wanted
719 * value appears to be off, queue a check so we'll release
720 * later and/or update the mds wanted value.
722 actual_wanted = __ceph_caps_wanted(ci);
723 if ((wanted & ~actual_wanted) ||
724 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
725 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
726 ceph_cap_string(issued), ceph_cap_string(wanted),
727 ceph_cap_string(actual_wanted));
728 __cap_delay_requeue(mdsc, ci);
731 if (flags & CEPH_CAP_FLAG_AUTH) {
732 if (!ci->i_auth_cap ||
733 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
734 if (ci->i_auth_cap &&
735 ci->i_auth_cap->session != cap->session)
736 change_auth_cap_ses(ci, cap->session);
737 ci->i_auth_cap = cap;
738 cap->mds_wanted = wanted;
741 WARN_ON(ci->i_auth_cap == cap);
744 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
745 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
746 ceph_cap_string(issued|cap->issued), seq, mds);
747 cap->cap_id = cap_id;
748 cap->issued = issued;
749 cap->implemented |= issued;
750 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
751 cap->mds_wanted = wanted;
753 cap->mds_wanted |= wanted;
755 cap->issue_seq = seq;
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->vfs_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->vfs_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->vfs_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->vfs_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->vfs_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->vfs_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->vfs_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->vfs_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->vfs_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->vfs_inode.i_mode) &&
974 ci->vfs_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->vfs_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->vfs_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->vfs_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 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1120 mdsc = ceph_inode_to_client(&ci->vfs_inode)->mdsc;
1122 /* remove from inode's cap rbtree, and clear auth cap */
1123 rb_erase(&cap->ci_node, &ci->i_caps);
1124 if (ci->i_auth_cap == cap) {
1125 WARN_ON_ONCE(!list_empty(&ci->i_dirty_item) &&
1126 !mdsc->fsc->blocklisted);
1127 ci->i_auth_cap = NULL;
1130 /* remove from session list */
1131 spin_lock(&session->s_cap_lock);
1132 if (session->s_cap_iterator == cap) {
1133 /* not yet, we are iterating over this very cap */
1134 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1137 list_del_init(&cap->session_caps);
1138 session->s_nr_caps--;
1139 atomic64_dec(&mdsc->metric.total_caps);
1140 cap->session = NULL;
1143 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1147 * s_cap_reconnect is protected by s_cap_lock. no one changes
1148 * s_cap_gen while session is in the reconnect state.
1150 if (queue_release &&
1151 (!session->s_cap_reconnect ||
1152 cap->cap_gen == atomic_read(&session->s_cap_gen))) {
1153 cap->queue_release = 1;
1155 __ceph_queue_cap_release(session, cap);
1159 cap->queue_release = 0;
1161 cap->cap_ino = ci->i_vino.ino;
1163 spin_unlock(&session->s_cap_lock);
1166 ceph_put_cap(mdsc, cap);
1168 if (!__ceph_is_any_real_caps(ci)) {
1169 /* when reconnect denied, we remove session caps forcibly,
1170 * i_wr_ref can be non-zero. If there are ongoing write,
1171 * keep i_snap_realm.
1173 if (ci->i_wr_ref == 0 && ci->i_snap_realm)
1174 ceph_change_snap_realm(&ci->vfs_inode, NULL);
1176 __cap_delay_cancel(mdsc, ci);
1180 struct cap_msg_args {
1181 struct ceph_mds_session *session;
1182 u64 ino, cid, follows;
1183 u64 flush_tid, oldest_flush_tid, size, max_size;
1186 struct ceph_buffer *xattr_buf;
1187 struct ceph_buffer *old_xattr_buf;
1188 struct timespec64 atime, mtime, ctime, btime;
1189 int op, caps, wanted, dirty;
1190 u32 seq, issue_seq, mseq, time_warp_seq;
1200 * cap struct size + flock buffer size + inline version + inline data size +
1201 * osd_epoch_barrier + oldest_flush_tid
1203 #define CAP_MSG_SIZE (sizeof(struct ceph_mds_caps) + \
1204 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4)
1206 /* Marshal up the cap msg to the MDS */
1207 static void encode_cap_msg(struct ceph_msg *msg, struct cap_msg_args *arg)
1209 struct ceph_mds_caps *fc;
1211 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1213 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",
1214 __func__, ceph_cap_op_name(arg->op), arg->cid, arg->ino,
1215 ceph_cap_string(arg->caps), ceph_cap_string(arg->wanted),
1216 ceph_cap_string(arg->dirty), arg->seq, arg->issue_seq,
1217 arg->flush_tid, arg->oldest_flush_tid, arg->mseq, arg->follows,
1218 arg->size, arg->max_size, arg->xattr_version,
1219 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1221 msg->hdr.version = cpu_to_le16(10);
1222 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1224 fc = msg->front.iov_base;
1225 memset(fc, 0, sizeof(*fc));
1227 fc->cap_id = cpu_to_le64(arg->cid);
1228 fc->op = cpu_to_le32(arg->op);
1229 fc->seq = cpu_to_le32(arg->seq);
1230 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1231 fc->migrate_seq = cpu_to_le32(arg->mseq);
1232 fc->caps = cpu_to_le32(arg->caps);
1233 fc->wanted = cpu_to_le32(arg->wanted);
1234 fc->dirty = cpu_to_le32(arg->dirty);
1235 fc->ino = cpu_to_le64(arg->ino);
1236 fc->snap_follows = cpu_to_le64(arg->follows);
1238 fc->size = cpu_to_le64(arg->size);
1239 fc->max_size = cpu_to_le64(arg->max_size);
1240 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1241 ceph_encode_timespec64(&fc->atime, &arg->atime);
1242 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1243 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1245 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1246 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1247 fc->mode = cpu_to_le32(arg->mode);
1249 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1250 if (arg->xattr_buf) {
1251 msg->middle = ceph_buffer_get(arg->xattr_buf);
1252 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1253 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1257 /* flock buffer size (version 2) */
1258 ceph_encode_32(&p, 0);
1259 /* inline version (version 4) */
1260 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1261 /* inline data size */
1262 ceph_encode_32(&p, 0);
1264 * osd_epoch_barrier (version 5)
1265 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1266 * case it was recently changed
1268 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1269 /* oldest_flush_tid (version 6) */
1270 ceph_encode_64(&p, arg->oldest_flush_tid);
1273 * caller_uid/caller_gid (version 7)
1275 * Currently, we don't properly track which caller dirtied the caps
1276 * last, and force a flush of them when there is a conflict. For now,
1277 * just set this to 0:0, to emulate how the MDS has worked up to now.
1279 ceph_encode_32(&p, 0);
1280 ceph_encode_32(&p, 0);
1282 /* pool namespace (version 8) (mds always ignores this) */
1283 ceph_encode_32(&p, 0);
1285 /* btime and change_attr (version 9) */
1286 ceph_encode_timespec64(p, &arg->btime);
1287 p += sizeof(struct ceph_timespec);
1288 ceph_encode_64(&p, arg->change_attr);
1290 /* Advisory flags (version 10) */
1291 ceph_encode_32(&p, arg->flags);
1295 * Queue cap releases when an inode is dropped from our cache.
1297 void __ceph_remove_caps(struct ceph_inode_info *ci)
1301 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1302 * may call __ceph_caps_issued_mask() on a freeing inode. */
1303 spin_lock(&ci->i_ceph_lock);
1304 p = rb_first(&ci->i_caps);
1306 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1308 __ceph_remove_cap(cap, true);
1310 spin_unlock(&ci->i_ceph_lock);
1314 * Prepare to send a cap message to an MDS. Update the cap state, and populate
1315 * the arg struct with the parameters that will need to be sent. This should
1316 * be done under the i_ceph_lock to guard against changes to cap state.
1318 * Make note of max_size reported/requested from mds, revoked caps
1319 * that have now been implemented.
1321 static void __prep_cap(struct cap_msg_args *arg, struct ceph_cap *cap,
1322 int op, int flags, int used, int want, int retain,
1323 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1325 struct ceph_inode_info *ci = cap->ci;
1326 struct inode *inode = &ci->vfs_inode;
1329 lockdep_assert_held(&ci->i_ceph_lock);
1331 held = cap->issued | cap->implemented;
1332 revoking = cap->implemented & ~cap->issued;
1333 retain &= ~revoking;
1335 dout("%s %p cap %p session %p %s -> %s (revoking %s)\n",
1336 __func__, inode, cap, cap->session,
1337 ceph_cap_string(held), ceph_cap_string(held & retain),
1338 ceph_cap_string(revoking));
1339 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1341 ci->i_ceph_flags &= ~CEPH_I_FLUSH;
1343 cap->issued &= retain; /* drop bits we don't want */
1345 * Wake up any waiters on wanted -> needed transition. This is due to
1346 * the weird transition from buffered to sync IO... we need to flush
1347 * dirty pages _before_ allowing sync writes to avoid reordering.
1349 arg->wake = cap->implemented & ~cap->issued;
1350 cap->implemented &= cap->issued | used;
1351 cap->mds_wanted = want;
1353 arg->session = cap->session;
1354 arg->ino = ceph_vino(inode).ino;
1355 arg->cid = cap->cap_id;
1356 arg->follows = flushing ? ci->i_head_snapc->seq : 0;
1357 arg->flush_tid = flush_tid;
1358 arg->oldest_flush_tid = oldest_flush_tid;
1360 arg->size = i_size_read(inode);
1361 ci->i_reported_size = arg->size;
1362 arg->max_size = ci->i_wanted_max_size;
1363 if (cap == ci->i_auth_cap) {
1364 if (want & CEPH_CAP_ANY_FILE_WR)
1365 ci->i_requested_max_size = arg->max_size;
1367 ci->i_requested_max_size = 0;
1370 if (flushing & CEPH_CAP_XATTR_EXCL) {
1371 arg->old_xattr_buf = __ceph_build_xattrs_blob(ci);
1372 arg->xattr_version = ci->i_xattrs.version;
1373 arg->xattr_buf = ci->i_xattrs.blob;
1375 arg->xattr_buf = NULL;
1376 arg->old_xattr_buf = NULL;
1379 arg->mtime = inode->i_mtime;
1380 arg->atime = inode->i_atime;
1381 arg->ctime = inode->i_ctime;
1382 arg->btime = ci->i_btime;
1383 arg->change_attr = inode_peek_iversion_raw(inode);
1386 arg->caps = cap->implemented;
1388 arg->dirty = flushing;
1390 arg->seq = cap->seq;
1391 arg->issue_seq = cap->issue_seq;
1392 arg->mseq = cap->mseq;
1393 arg->time_warp_seq = ci->i_time_warp_seq;
1395 arg->uid = inode->i_uid;
1396 arg->gid = inode->i_gid;
1397 arg->mode = inode->i_mode;
1399 arg->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1400 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1401 !list_empty(&ci->i_cap_snaps)) {
1402 struct ceph_cap_snap *capsnap;
1403 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1404 if (capsnap->cap_flush.tid)
1406 if (capsnap->need_flush) {
1407 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1416 * Send a cap msg on the given inode.
1418 * Caller should hold snap_rwsem (read), s_mutex.
1420 static void __send_cap(struct cap_msg_args *arg, struct ceph_inode_info *ci)
1422 struct ceph_msg *msg;
1423 struct inode *inode = &ci->vfs_inode;
1425 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1427 pr_err("error allocating cap msg: ino (%llx.%llx) flushing %s tid %llu, requeuing cap.\n",
1428 ceph_vinop(inode), ceph_cap_string(arg->dirty),
1430 spin_lock(&ci->i_ceph_lock);
1431 __cap_delay_requeue(arg->session->s_mdsc, ci);
1432 spin_unlock(&ci->i_ceph_lock);
1436 encode_cap_msg(msg, arg);
1437 ceph_con_send(&arg->session->s_con, msg);
1438 ceph_buffer_put(arg->old_xattr_buf);
1440 wake_up_all(&ci->i_cap_wq);
1443 static inline int __send_flush_snap(struct inode *inode,
1444 struct ceph_mds_session *session,
1445 struct ceph_cap_snap *capsnap,
1446 u32 mseq, u64 oldest_flush_tid)
1448 struct cap_msg_args arg;
1449 struct ceph_msg *msg;
1451 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, CAP_MSG_SIZE, GFP_NOFS, false);
1455 arg.session = session;
1456 arg.ino = ceph_vino(inode).ino;
1458 arg.follows = capsnap->follows;
1459 arg.flush_tid = capsnap->cap_flush.tid;
1460 arg.oldest_flush_tid = oldest_flush_tid;
1462 arg.size = capsnap->size;
1464 arg.xattr_version = capsnap->xattr_version;
1465 arg.xattr_buf = capsnap->xattr_blob;
1466 arg.old_xattr_buf = NULL;
1468 arg.atime = capsnap->atime;
1469 arg.mtime = capsnap->mtime;
1470 arg.ctime = capsnap->ctime;
1471 arg.btime = capsnap->btime;
1472 arg.change_attr = capsnap->change_attr;
1474 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1475 arg.caps = capsnap->issued;
1477 arg.dirty = capsnap->dirty;
1482 arg.time_warp_seq = capsnap->time_warp_seq;
1484 arg.uid = capsnap->uid;
1485 arg.gid = capsnap->gid;
1486 arg.mode = capsnap->mode;
1488 arg.inline_data = capsnap->inline_data;
1492 encode_cap_msg(msg, &arg);
1493 ceph_con_send(&arg.session->s_con, msg);
1498 * When a snapshot is taken, clients accumulate dirty metadata on
1499 * inodes with capabilities in ceph_cap_snaps to describe the file
1500 * state at the time the snapshot was taken. This must be flushed
1501 * asynchronously back to the MDS once sync writes complete and dirty
1502 * data is written out.
1504 * Called under i_ceph_lock.
1506 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1507 struct ceph_mds_session *session)
1508 __releases(ci->i_ceph_lock)
1509 __acquires(ci->i_ceph_lock)
1511 struct inode *inode = &ci->vfs_inode;
1512 struct ceph_mds_client *mdsc = session->s_mdsc;
1513 struct ceph_cap_snap *capsnap;
1514 u64 oldest_flush_tid = 0;
1515 u64 first_tid = 1, last_tid = 0;
1517 dout("__flush_snaps %p session %p\n", inode, session);
1519 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1521 * we need to wait for sync writes to complete and for dirty
1522 * pages to be written out.
1524 if (capsnap->dirty_pages || capsnap->writing)
1527 /* should be removed by ceph_try_drop_cap_snap() */
1528 BUG_ON(!capsnap->need_flush);
1530 /* only flush each capsnap once */
1531 if (capsnap->cap_flush.tid > 0) {
1532 dout(" already flushed %p, skipping\n", capsnap);
1536 spin_lock(&mdsc->cap_dirty_lock);
1537 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1538 list_add_tail(&capsnap->cap_flush.g_list,
1539 &mdsc->cap_flush_list);
1540 if (oldest_flush_tid == 0)
1541 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1542 if (list_empty(&ci->i_flushing_item)) {
1543 list_add_tail(&ci->i_flushing_item,
1544 &session->s_cap_flushing);
1546 spin_unlock(&mdsc->cap_dirty_lock);
1548 list_add_tail(&capsnap->cap_flush.i_list,
1549 &ci->i_cap_flush_list);
1552 first_tid = capsnap->cap_flush.tid;
1553 last_tid = capsnap->cap_flush.tid;
1556 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1558 while (first_tid <= last_tid) {
1559 struct ceph_cap *cap = ci->i_auth_cap;
1560 struct ceph_cap_flush *cf;
1563 if (!(cap && cap->session == session)) {
1564 dout("__flush_snaps %p auth cap %p not mds%d, "
1565 "stop\n", inode, cap, session->s_mds);
1570 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1571 if (cf->tid >= first_tid) {
1579 first_tid = cf->tid + 1;
1581 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1582 refcount_inc(&capsnap->nref);
1583 spin_unlock(&ci->i_ceph_lock);
1585 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1586 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1588 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1591 pr_err("__flush_snaps: error sending cap flushsnap, "
1592 "ino (%llx.%llx) tid %llu follows %llu\n",
1593 ceph_vinop(inode), cf->tid, capsnap->follows);
1596 ceph_put_cap_snap(capsnap);
1597 spin_lock(&ci->i_ceph_lock);
1601 void ceph_flush_snaps(struct ceph_inode_info *ci,
1602 struct ceph_mds_session **psession)
1604 struct inode *inode = &ci->vfs_inode;
1605 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1606 struct ceph_mds_session *session = NULL;
1609 dout("ceph_flush_snaps %p\n", inode);
1611 session = *psession;
1613 spin_lock(&ci->i_ceph_lock);
1614 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1615 dout(" no capsnap needs flush, doing nothing\n");
1618 if (!ci->i_auth_cap) {
1619 dout(" no auth cap (migrating?), doing nothing\n");
1623 mds = ci->i_auth_cap->session->s_mds;
1624 if (session && session->s_mds != mds) {
1625 dout(" oops, wrong session %p mutex\n", session);
1626 ceph_put_mds_session(session);
1630 spin_unlock(&ci->i_ceph_lock);
1631 mutex_lock(&mdsc->mutex);
1632 session = __ceph_lookup_mds_session(mdsc, mds);
1633 mutex_unlock(&mdsc->mutex);
1637 // make sure flushsnap messages are sent in proper order.
1638 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1639 __kick_flushing_caps(mdsc, session, ci, 0);
1641 __ceph_flush_snaps(ci, session);
1643 spin_unlock(&ci->i_ceph_lock);
1646 *psession = session;
1648 ceph_put_mds_session(session);
1649 /* we flushed them all; remove this inode from the queue */
1650 spin_lock(&mdsc->snap_flush_lock);
1651 list_del_init(&ci->i_snap_flush_item);
1652 spin_unlock(&mdsc->snap_flush_lock);
1656 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1657 * Caller is then responsible for calling __mark_inode_dirty with the
1658 * returned flags value.
1660 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1661 struct ceph_cap_flush **pcf)
1663 struct ceph_mds_client *mdsc =
1664 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1665 struct inode *inode = &ci->vfs_inode;
1666 int was = ci->i_dirty_caps;
1669 lockdep_assert_held(&ci->i_ceph_lock);
1671 if (!ci->i_auth_cap) {
1672 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1673 "but no auth cap (session was closed?)\n",
1674 inode, ceph_ino(inode), ceph_cap_string(mask));
1678 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1679 ceph_cap_string(mask), ceph_cap_string(was),
1680 ceph_cap_string(was | mask));
1681 ci->i_dirty_caps |= mask;
1683 struct ceph_mds_session *session = ci->i_auth_cap->session;
1685 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1686 swap(ci->i_prealloc_cap_flush, *pcf);
1688 if (!ci->i_head_snapc) {
1689 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1690 ci->i_head_snapc = ceph_get_snap_context(
1691 ci->i_snap_realm->cached_context);
1693 dout(" inode %p now dirty snapc %p auth cap %p\n",
1694 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1695 BUG_ON(!list_empty(&ci->i_dirty_item));
1696 spin_lock(&mdsc->cap_dirty_lock);
1697 list_add(&ci->i_dirty_item, &session->s_cap_dirty);
1698 spin_unlock(&mdsc->cap_dirty_lock);
1699 if (ci->i_flushing_caps == 0) {
1701 dirty |= I_DIRTY_SYNC;
1704 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1706 BUG_ON(list_empty(&ci->i_dirty_item));
1707 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1708 (mask & CEPH_CAP_FILE_BUFFER))
1709 dirty |= I_DIRTY_DATASYNC;
1710 __cap_delay_requeue(mdsc, ci);
1714 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1716 struct ceph_cap_flush *cf;
1718 cf = kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1719 cf->is_capsnap = false;
1723 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1726 kmem_cache_free(ceph_cap_flush_cachep, cf);
1729 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1731 if (!list_empty(&mdsc->cap_flush_list)) {
1732 struct ceph_cap_flush *cf =
1733 list_first_entry(&mdsc->cap_flush_list,
1734 struct ceph_cap_flush, g_list);
1741 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1742 * Return true if caller needs to wake up flush waiters.
1744 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
1745 struct ceph_cap_flush *cf)
1747 struct ceph_cap_flush *prev;
1748 bool wake = cf->wake;
1750 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1751 prev = list_prev_entry(cf, g_list);
1755 list_del_init(&cf->g_list);
1759 static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
1760 struct ceph_cap_flush *cf)
1762 struct ceph_cap_flush *prev;
1763 bool wake = cf->wake;
1765 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1766 prev = list_prev_entry(cf, i_list);
1770 list_del_init(&cf->i_list);
1775 * Add dirty inode to the flushing list. Assigned a seq number so we
1776 * can wait for caps to flush without starving.
1778 * Called under i_ceph_lock. Returns the flush tid.
1780 static u64 __mark_caps_flushing(struct inode *inode,
1781 struct ceph_mds_session *session, bool wake,
1782 u64 *oldest_flush_tid)
1784 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1785 struct ceph_inode_info *ci = ceph_inode(inode);
1786 struct ceph_cap_flush *cf = NULL;
1789 lockdep_assert_held(&ci->i_ceph_lock);
1790 BUG_ON(ci->i_dirty_caps == 0);
1791 BUG_ON(list_empty(&ci->i_dirty_item));
1792 BUG_ON(!ci->i_prealloc_cap_flush);
1794 flushing = ci->i_dirty_caps;
1795 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1796 ceph_cap_string(flushing),
1797 ceph_cap_string(ci->i_flushing_caps),
1798 ceph_cap_string(ci->i_flushing_caps | flushing));
1799 ci->i_flushing_caps |= flushing;
1800 ci->i_dirty_caps = 0;
1801 dout(" inode %p now !dirty\n", inode);
1803 swap(cf, ci->i_prealloc_cap_flush);
1804 cf->caps = flushing;
1807 spin_lock(&mdsc->cap_dirty_lock);
1808 list_del_init(&ci->i_dirty_item);
1810 cf->tid = ++mdsc->last_cap_flush_tid;
1811 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1812 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1814 if (list_empty(&ci->i_flushing_item)) {
1815 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1816 mdsc->num_cap_flushing++;
1818 spin_unlock(&mdsc->cap_dirty_lock);
1820 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1826 * try to invalidate mapping pages without blocking.
1828 static int try_nonblocking_invalidate(struct inode *inode)
1830 struct ceph_inode_info *ci = ceph_inode(inode);
1831 u32 invalidating_gen = ci->i_rdcache_gen;
1833 spin_unlock(&ci->i_ceph_lock);
1834 ceph_fscache_invalidate(inode);
1835 invalidate_mapping_pages(&inode->i_data, 0, -1);
1836 spin_lock(&ci->i_ceph_lock);
1838 if (inode->i_data.nrpages == 0 &&
1839 invalidating_gen == ci->i_rdcache_gen) {
1841 dout("try_nonblocking_invalidate %p success\n", inode);
1842 /* save any racing async invalidate some trouble */
1843 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1846 dout("try_nonblocking_invalidate %p failed\n", inode);
1850 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1852 loff_t size = i_size_read(&ci->vfs_inode);
1853 /* mds will adjust max size according to the reported size */
1854 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1856 if (size >= ci->i_max_size)
1858 /* half of previous max_size increment has been used */
1859 if (ci->i_max_size > ci->i_reported_size &&
1860 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1866 * Swiss army knife function to examine currently used and wanted
1867 * versus held caps. Release, flush, ack revoked caps to mds as
1870 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1871 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1874 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1875 struct ceph_mds_session *session)
1877 struct inode *inode = &ci->vfs_inode;
1878 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
1879 struct ceph_cap *cap;
1880 u64 flush_tid, oldest_flush_tid;
1881 int file_wanted, used, cap_used;
1882 int issued, implemented, want, retain, revoking, flushing = 0;
1883 int mds = -1; /* keep track of how far we've gone through i_caps list
1884 to avoid an infinite loop on retry */
1886 bool queue_invalidate = false;
1887 bool tried_invalidate = false;
1890 ceph_get_mds_session(session);
1892 spin_lock(&ci->i_ceph_lock);
1893 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1894 flags |= CHECK_CAPS_FLUSH;
1896 /* Caps wanted by virtue of active open files. */
1897 file_wanted = __ceph_caps_file_wanted(ci);
1899 /* Caps which have active references against them */
1900 used = __ceph_caps_used(ci);
1903 * "issued" represents the current caps that the MDS wants us to have.
1904 * "implemented" is the set that we have been granted, and includes the
1905 * ones that have not yet been returned to the MDS (the "revoking" set,
1906 * usually because they have outstanding references).
1908 issued = __ceph_caps_issued(ci, &implemented);
1909 revoking = implemented & ~issued;
1913 /* The ones we currently want to retain (may be adjusted below) */
1914 retain = file_wanted | used | CEPH_CAP_PIN;
1915 if (!mdsc->stopping && inode->i_nlink > 0) {
1917 retain |= CEPH_CAP_ANY; /* be greedy */
1918 } else if (S_ISDIR(inode->i_mode) &&
1919 (issued & CEPH_CAP_FILE_SHARED) &&
1920 __ceph_dir_is_complete(ci)) {
1922 * If a directory is complete, we want to keep
1923 * the exclusive cap. So that MDS does not end up
1924 * revoking the shared cap on every create/unlink
1927 if (IS_RDONLY(inode)) {
1928 want = CEPH_CAP_ANY_SHARED;
1930 want |= CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1935 retain |= CEPH_CAP_ANY_SHARED;
1937 * keep RD only if we didn't have the file open RW,
1938 * because then the mds would revoke it anyway to
1939 * journal max_size=0.
1941 if (ci->i_max_size == 0)
1942 retain |= CEPH_CAP_ANY_RD;
1946 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1947 " issued %s revoking %s retain %s %s%s\n", inode,
1948 ceph_cap_string(file_wanted),
1949 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1950 ceph_cap_string(ci->i_flushing_caps),
1951 ceph_cap_string(issued), ceph_cap_string(revoking),
1952 ceph_cap_string(retain),
1953 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1954 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1957 * If we no longer need to hold onto old our caps, and we may
1958 * have cached pages, but don't want them, then try to invalidate.
1959 * If we fail, it's because pages are locked.... try again later.
1961 if ((!(flags & CHECK_CAPS_NOINVAL) || mdsc->stopping) &&
1962 S_ISREG(inode->i_mode) &&
1963 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1964 inode->i_data.nrpages && /* have cached pages */
1965 (revoking & (CEPH_CAP_FILE_CACHE|
1966 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1967 !tried_invalidate) {
1968 dout("check_caps trying to invalidate on %p\n", inode);
1969 if (try_nonblocking_invalidate(inode) < 0) {
1970 dout("check_caps queuing invalidate\n");
1971 queue_invalidate = true;
1972 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1974 tried_invalidate = true;
1978 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1980 struct cap_msg_args arg;
1982 cap = rb_entry(p, struct ceph_cap, ci_node);
1984 /* avoid looping forever */
1985 if (mds >= cap->mds ||
1986 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1990 * If we have an auth cap, we don't need to consider any
1991 * overlapping caps as used.
1994 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1995 cap_used &= ~ci->i_auth_cap->issued;
1997 revoking = cap->implemented & ~cap->issued;
1998 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1999 cap->mds, cap, ceph_cap_string(cap_used),
2000 ceph_cap_string(cap->issued),
2001 ceph_cap_string(cap->implemented),
2002 ceph_cap_string(revoking));
2004 if (cap == ci->i_auth_cap &&
2005 (cap->issued & CEPH_CAP_FILE_WR)) {
2006 /* request larger max_size from MDS? */
2007 if (ci->i_wanted_max_size > ci->i_max_size &&
2008 ci->i_wanted_max_size > ci->i_requested_max_size) {
2009 dout("requesting new max_size\n");
2013 /* approaching file_max? */
2014 if (__ceph_should_report_size(ci)) {
2015 dout("i_size approaching max_size\n");
2019 /* flush anything dirty? */
2020 if (cap == ci->i_auth_cap) {
2021 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
2022 dout("flushing dirty caps\n");
2025 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
2026 dout("flushing snap caps\n");
2031 /* completed revocation? going down and there are no caps? */
2032 if (revoking && (revoking & cap_used) == 0) {
2033 dout("completed revocation of %s\n",
2034 ceph_cap_string(cap->implemented & ~cap->issued));
2038 /* want more caps from mds? */
2039 if (want & ~cap->mds_wanted) {
2040 if (want & ~(cap->mds_wanted | cap->issued))
2042 if (!__cap_is_valid(cap))
2046 /* things we might delay */
2047 if ((cap->issued & ~retain) == 0)
2048 continue; /* nope, all good */
2051 ceph_put_mds_session(session);
2052 session = ceph_get_mds_session(cap->session);
2054 /* kick flushing and flush snaps before sending normal
2056 if (cap == ci->i_auth_cap &&
2058 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2059 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2060 __kick_flushing_caps(mdsc, session, ci, 0);
2061 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2062 __ceph_flush_snaps(ci, session);
2067 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2068 flushing = ci->i_dirty_caps;
2069 flush_tid = __mark_caps_flushing(inode, session, false,
2071 if (flags & CHECK_CAPS_FLUSH &&
2072 list_empty(&session->s_cap_dirty))
2073 mflags |= CEPH_CLIENT_CAPS_SYNC;
2077 spin_lock(&mdsc->cap_dirty_lock);
2078 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2079 spin_unlock(&mdsc->cap_dirty_lock);
2082 mds = cap->mds; /* remember mds, so we don't repeat */
2084 __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
2085 want, retain, flushing, flush_tid, oldest_flush_tid);
2087 spin_unlock(&ci->i_ceph_lock);
2088 __send_cap(&arg, ci);
2089 spin_lock(&ci->i_ceph_lock);
2091 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2094 /* periodically re-calculate caps wanted by open files */
2095 if (__ceph_is_any_real_caps(ci) &&
2096 list_empty(&ci->i_cap_delay_list) &&
2097 (file_wanted & ~CEPH_CAP_PIN) &&
2098 !(used & (CEPH_CAP_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
2099 __cap_delay_requeue(mdsc, ci);
2102 spin_unlock(&ci->i_ceph_lock);
2104 ceph_put_mds_session(session);
2105 if (queue_invalidate)
2106 ceph_queue_invalidate(inode);
2110 * Try to flush dirty caps back to the auth mds.
2112 static int try_flush_caps(struct inode *inode, u64 *ptid)
2114 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2115 struct ceph_inode_info *ci = ceph_inode(inode);
2117 u64 flush_tid = 0, oldest_flush_tid = 0;
2119 spin_lock(&ci->i_ceph_lock);
2121 if (ci->i_dirty_caps && ci->i_auth_cap) {
2122 struct ceph_cap *cap = ci->i_auth_cap;
2123 struct cap_msg_args arg;
2124 struct ceph_mds_session *session = cap->session;
2126 if (session->s_state < CEPH_MDS_SESSION_OPEN) {
2127 spin_unlock(&ci->i_ceph_lock);
2131 if (ci->i_ceph_flags &
2132 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2133 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2134 __kick_flushing_caps(mdsc, session, ci, 0);
2135 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2136 __ceph_flush_snaps(ci, session);
2140 flushing = ci->i_dirty_caps;
2141 flush_tid = __mark_caps_flushing(inode, session, true,
2144 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH, CEPH_CLIENT_CAPS_SYNC,
2145 __ceph_caps_used(ci), __ceph_caps_wanted(ci),
2146 (cap->issued | cap->implemented),
2147 flushing, flush_tid, oldest_flush_tid);
2148 spin_unlock(&ci->i_ceph_lock);
2150 __send_cap(&arg, ci);
2152 if (!list_empty(&ci->i_cap_flush_list)) {
2153 struct ceph_cap_flush *cf =
2154 list_last_entry(&ci->i_cap_flush_list,
2155 struct ceph_cap_flush, i_list);
2157 flush_tid = cf->tid;
2159 flushing = ci->i_flushing_caps;
2160 spin_unlock(&ci->i_ceph_lock);
2168 * Return true if we've flushed caps through the given flush_tid.
2170 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2172 struct ceph_inode_info *ci = ceph_inode(inode);
2175 spin_lock(&ci->i_ceph_lock);
2176 if (!list_empty(&ci->i_cap_flush_list)) {
2177 struct ceph_cap_flush * cf =
2178 list_first_entry(&ci->i_cap_flush_list,
2179 struct ceph_cap_flush, i_list);
2180 if (cf->tid <= flush_tid)
2183 spin_unlock(&ci->i_ceph_lock);
2188 * wait for any unsafe requests to complete.
2190 static int unsafe_request_wait(struct inode *inode)
2192 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2193 struct ceph_inode_info *ci = ceph_inode(inode);
2194 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2197 spin_lock(&ci->i_unsafe_lock);
2198 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2199 req1 = list_last_entry(&ci->i_unsafe_dirops,
2200 struct ceph_mds_request,
2202 ceph_mdsc_get_request(req1);
2204 if (!list_empty(&ci->i_unsafe_iops)) {
2205 req2 = list_last_entry(&ci->i_unsafe_iops,
2206 struct ceph_mds_request,
2207 r_unsafe_target_item);
2208 ceph_mdsc_get_request(req2);
2210 spin_unlock(&ci->i_unsafe_lock);
2213 * Trigger to flush the journal logs in all the relevant MDSes
2214 * manually, or in the worst case we must wait at most 5 seconds
2215 * to wait the journal logs to be flushed by the MDSes periodically.
2218 struct ceph_mds_session **sessions = NULL;
2219 struct ceph_mds_session *s;
2220 struct ceph_mds_request *req;
2225 * The mdsc->max_sessions is unlikely to be changed
2226 * mostly, here we will retry it by reallocating the
2227 * sessions arrary memory to get rid of the mdsc->mutex
2231 max = mdsc->max_sessions;
2232 sessions = krealloc(sessions, max * sizeof(s), __GFP_ZERO);
2236 spin_lock(&ci->i_unsafe_lock);
2238 list_for_each_entry(req, &ci->i_unsafe_dirops,
2239 r_unsafe_dir_item) {
2241 if (unlikely(s->s_mds > max)) {
2242 spin_unlock(&ci->i_unsafe_lock);
2245 if (!sessions[s->s_mds]) {
2246 s = ceph_get_mds_session(s);
2247 sessions[s->s_mds] = s;
2252 list_for_each_entry(req, &ci->i_unsafe_iops,
2253 r_unsafe_target_item) {
2255 if (unlikely(s->s_mds > max)) {
2256 spin_unlock(&ci->i_unsafe_lock);
2259 if (!sessions[s->s_mds]) {
2260 s = ceph_get_mds_session(s);
2261 sessions[s->s_mds] = s;
2265 spin_unlock(&ci->i_unsafe_lock);
2268 spin_lock(&ci->i_ceph_lock);
2269 if (ci->i_auth_cap) {
2270 s = ci->i_auth_cap->session;
2271 if (!sessions[s->s_mds])
2272 sessions[s->s_mds] = ceph_get_mds_session(s);
2274 spin_unlock(&ci->i_ceph_lock);
2276 /* send flush mdlog request to MDSes */
2277 for (i = 0; i < max; i++) {
2280 send_flush_mdlog(s);
2281 ceph_put_mds_session(s);
2287 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2288 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2290 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2291 ceph_timeout_jiffies(req1->r_timeout));
2294 ceph_mdsc_put_request(req1);
2297 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2298 ceph_timeout_jiffies(req2->r_timeout));
2301 ceph_mdsc_put_request(req2);
2306 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2308 struct ceph_file_info *fi = file->private_data;
2309 struct inode *inode = file->f_mapping->host;
2310 struct ceph_inode_info *ci = ceph_inode(inode);
2315 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2317 ret = file_write_and_wait_range(file, start, end);
2321 ret = ceph_wait_on_async_create(inode);
2325 dirty = try_flush_caps(inode, &flush_tid);
2326 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2328 err = unsafe_request_wait(inode);
2331 * only wait on non-file metadata writeback (the mds
2332 * can recover size and mtime, so we don't need to
2335 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2336 err = wait_event_interruptible(ci->i_cap_wq,
2337 caps_are_flushed(inode, flush_tid));
2343 if (errseq_check(&ci->i_meta_err, READ_ONCE(fi->meta_err))) {
2344 spin_lock(&file->f_lock);
2345 err = errseq_check_and_advance(&ci->i_meta_err,
2347 spin_unlock(&file->f_lock);
2352 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2357 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2358 * queue inode for flush but don't do so immediately, because we can
2359 * get by with fewer MDS messages if we wait for data writeback to
2362 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2364 struct ceph_inode_info *ci = ceph_inode(inode);
2368 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2370 dout("write_inode %p wait=%d\n", inode, wait);
2372 dirty = try_flush_caps(inode, &flush_tid);
2374 err = wait_event_interruptible(ci->i_cap_wq,
2375 caps_are_flushed(inode, flush_tid));
2377 struct ceph_mds_client *mdsc =
2378 ceph_sb_to_client(inode->i_sb)->mdsc;
2380 spin_lock(&ci->i_ceph_lock);
2381 if (__ceph_caps_dirty(ci))
2382 __cap_delay_requeue_front(mdsc, ci);
2383 spin_unlock(&ci->i_ceph_lock);
2388 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2389 struct ceph_mds_session *session,
2390 struct ceph_inode_info *ci,
2391 u64 oldest_flush_tid)
2392 __releases(ci->i_ceph_lock)
2393 __acquires(ci->i_ceph_lock)
2395 struct inode *inode = &ci->vfs_inode;
2396 struct ceph_cap *cap;
2397 struct ceph_cap_flush *cf;
2400 u64 last_snap_flush = 0;
2402 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2404 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2405 if (cf->is_capsnap) {
2406 last_snap_flush = cf->tid;
2411 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2412 if (cf->tid < first_tid)
2415 cap = ci->i_auth_cap;
2416 if (!(cap && cap->session == session)) {
2417 pr_err("%p auth cap %p not mds%d ???\n",
2418 inode, cap, session->s_mds);
2422 first_tid = cf->tid + 1;
2424 if (!cf->is_capsnap) {
2425 struct cap_msg_args arg;
2427 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2428 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2429 __prep_cap(&arg, cap, CEPH_CAP_OP_FLUSH,
2430 (cf->tid < last_snap_flush ?
2431 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2432 __ceph_caps_used(ci),
2433 __ceph_caps_wanted(ci),
2434 (cap->issued | cap->implemented),
2435 cf->caps, cf->tid, oldest_flush_tid);
2436 spin_unlock(&ci->i_ceph_lock);
2437 __send_cap(&arg, ci);
2439 struct ceph_cap_snap *capsnap =
2440 container_of(cf, struct ceph_cap_snap,
2442 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2443 inode, capsnap, cf->tid,
2444 ceph_cap_string(capsnap->dirty));
2446 refcount_inc(&capsnap->nref);
2447 spin_unlock(&ci->i_ceph_lock);
2449 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2452 pr_err("kick_flushing_caps: error sending "
2453 "cap flushsnap, ino (%llx.%llx) "
2454 "tid %llu follows %llu\n",
2455 ceph_vinop(inode), cf->tid,
2459 ceph_put_cap_snap(capsnap);
2462 spin_lock(&ci->i_ceph_lock);
2466 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2467 struct ceph_mds_session *session)
2469 struct ceph_inode_info *ci;
2470 struct ceph_cap *cap;
2471 u64 oldest_flush_tid;
2473 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2475 spin_lock(&mdsc->cap_dirty_lock);
2476 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2477 spin_unlock(&mdsc->cap_dirty_lock);
2479 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2480 spin_lock(&ci->i_ceph_lock);
2481 cap = ci->i_auth_cap;
2482 if (!(cap && cap->session == session)) {
2483 pr_err("%p auth cap %p not mds%d ???\n",
2484 &ci->vfs_inode, cap, session->s_mds);
2485 spin_unlock(&ci->i_ceph_lock);
2491 * if flushing caps were revoked, we re-send the cap flush
2492 * in client reconnect stage. This guarantees MDS * processes
2493 * the cap flush message before issuing the flushing caps to
2496 if ((cap->issued & ci->i_flushing_caps) !=
2497 ci->i_flushing_caps) {
2498 /* encode_caps_cb() also will reset these sequence
2499 * numbers. make sure sequence numbers in cap flush
2500 * message match later reconnect message */
2504 __kick_flushing_caps(mdsc, session, ci,
2507 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2510 spin_unlock(&ci->i_ceph_lock);
2514 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2515 struct ceph_mds_session *session)
2517 struct ceph_inode_info *ci;
2518 struct ceph_cap *cap;
2519 u64 oldest_flush_tid;
2521 lockdep_assert_held(&session->s_mutex);
2523 dout("kick_flushing_caps mds%d\n", session->s_mds);
2525 spin_lock(&mdsc->cap_dirty_lock);
2526 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2527 spin_unlock(&mdsc->cap_dirty_lock);
2529 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2530 spin_lock(&ci->i_ceph_lock);
2531 cap = ci->i_auth_cap;
2532 if (!(cap && cap->session == session)) {
2533 pr_err("%p auth cap %p not mds%d ???\n",
2534 &ci->vfs_inode, cap, session->s_mds);
2535 spin_unlock(&ci->i_ceph_lock);
2538 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2539 __kick_flushing_caps(mdsc, session, ci,
2542 spin_unlock(&ci->i_ceph_lock);
2546 void ceph_kick_flushing_inode_caps(struct ceph_mds_session *session,
2547 struct ceph_inode_info *ci)
2549 struct ceph_mds_client *mdsc = session->s_mdsc;
2550 struct ceph_cap *cap = ci->i_auth_cap;
2552 lockdep_assert_held(&ci->i_ceph_lock);
2554 dout("%s %p flushing %s\n", __func__, &ci->vfs_inode,
2555 ceph_cap_string(ci->i_flushing_caps));
2557 if (!list_empty(&ci->i_cap_flush_list)) {
2558 u64 oldest_flush_tid;
2559 spin_lock(&mdsc->cap_dirty_lock);
2560 list_move_tail(&ci->i_flushing_item,
2561 &cap->session->s_cap_flushing);
2562 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2563 spin_unlock(&mdsc->cap_dirty_lock);
2565 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2571 * Take references to capabilities we hold, so that we don't release
2572 * them to the MDS prematurely.
2574 void ceph_take_cap_refs(struct ceph_inode_info *ci, int got,
2575 bool snap_rwsem_locked)
2577 lockdep_assert_held(&ci->i_ceph_lock);
2579 if (got & CEPH_CAP_PIN)
2581 if (got & CEPH_CAP_FILE_RD)
2583 if (got & CEPH_CAP_FILE_CACHE)
2584 ci->i_rdcache_ref++;
2585 if (got & CEPH_CAP_FILE_EXCL)
2587 if (got & CEPH_CAP_FILE_WR) {
2588 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2589 BUG_ON(!snap_rwsem_locked);
2590 ci->i_head_snapc = ceph_get_snap_context(
2591 ci->i_snap_realm->cached_context);
2595 if (got & CEPH_CAP_FILE_BUFFER) {
2596 if (ci->i_wb_ref == 0)
2597 ihold(&ci->vfs_inode);
2599 dout("%s %p wb %d -> %d (?)\n", __func__,
2600 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2605 * Try to grab cap references. Specify those refs we @want, and the
2606 * minimal set we @need. Also include the larger offset we are writing
2607 * to (when applicable), and check against max_size here as well.
2608 * Note that caller is responsible for ensuring max_size increases are
2609 * requested from the MDS.
2611 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2612 * or a negative error code. There are 3 speical error codes:
2613 * -EAGAIN: need to sleep but non-blocking is specified
2614 * -EFBIG: ask caller to call check_max_size() and try again.
2615 * -ESTALE: ask caller to call ceph_renew_caps() and try again.
2618 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2619 NON_BLOCKING = (1 << 8),
2620 CHECK_FILELOCK = (1 << 9),
2623 static int try_get_cap_refs(struct inode *inode, int need, int want,
2624 loff_t endoff, int flags, int *got)
2626 struct ceph_inode_info *ci = ceph_inode(inode);
2627 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2629 int have, implemented;
2630 bool snap_rwsem_locked = false;
2632 dout("get_cap_refs %p need %s want %s\n", inode,
2633 ceph_cap_string(need), ceph_cap_string(want));
2636 spin_lock(&ci->i_ceph_lock);
2638 if ((flags & CHECK_FILELOCK) &&
2639 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2640 dout("try_get_cap_refs %p error filelock\n", inode);
2645 /* finish pending truncate */
2646 while (ci->i_truncate_pending) {
2647 spin_unlock(&ci->i_ceph_lock);
2648 if (snap_rwsem_locked) {
2649 up_read(&mdsc->snap_rwsem);
2650 snap_rwsem_locked = false;
2652 __ceph_do_pending_vmtruncate(inode);
2653 spin_lock(&ci->i_ceph_lock);
2656 have = __ceph_caps_issued(ci, &implemented);
2658 if (have & need & CEPH_CAP_FILE_WR) {
2659 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2660 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2661 inode, endoff, ci->i_max_size);
2662 if (endoff > ci->i_requested_max_size)
2663 ret = ci->i_auth_cap ? -EFBIG : -ESTALE;
2667 * If a sync write is in progress, we must wait, so that we
2668 * can get a final snapshot value for size+mtime.
2670 if (__ceph_have_pending_cap_snap(ci)) {
2671 dout("get_cap_refs %p cap_snap_pending\n", inode);
2676 if ((have & need) == need) {
2678 * Look at (implemented & ~have & not) so that we keep waiting
2679 * on transition from wanted -> needed caps. This is needed
2680 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2681 * going before a prior buffered writeback happens.
2683 int not = want & ~(have & need);
2684 int revoking = implemented & ~have;
2685 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2686 inode, ceph_cap_string(have), ceph_cap_string(not),
2687 ceph_cap_string(revoking));
2688 if ((revoking & not) == 0) {
2689 if (!snap_rwsem_locked &&
2690 !ci->i_head_snapc &&
2691 (need & CEPH_CAP_FILE_WR)) {
2692 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2694 * we can not call down_read() when
2695 * task isn't in TASK_RUNNING state
2697 if (flags & NON_BLOCKING) {
2702 spin_unlock(&ci->i_ceph_lock);
2703 down_read(&mdsc->snap_rwsem);
2704 snap_rwsem_locked = true;
2707 snap_rwsem_locked = true;
2709 if ((have & want) == want)
2713 ceph_take_cap_refs(ci, *got, true);
2717 int session_readonly = false;
2719 if (ci->i_auth_cap &&
2720 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2721 struct ceph_mds_session *s = ci->i_auth_cap->session;
2722 spin_lock(&s->s_cap_lock);
2723 session_readonly = s->s_readonly;
2724 spin_unlock(&s->s_cap_lock);
2726 if (session_readonly) {
2727 dout("get_cap_refs %p need %s but mds%d readonly\n",
2728 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2733 if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN) {
2734 dout("get_cap_refs %p forced umount\n", inode);
2738 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2739 if (need & ~mds_wanted) {
2740 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2741 inode, ceph_cap_string(need),
2742 ceph_cap_string(mds_wanted));
2747 dout("get_cap_refs %p have %s need %s\n", inode,
2748 ceph_cap_string(have), ceph_cap_string(need));
2752 __ceph_touch_fmode(ci, mdsc, flags);
2754 spin_unlock(&ci->i_ceph_lock);
2755 if (snap_rwsem_locked)
2756 up_read(&mdsc->snap_rwsem);
2759 ceph_update_cap_mis(&mdsc->metric);
2761 ceph_update_cap_hit(&mdsc->metric);
2763 dout("get_cap_refs %p ret %d got %s\n", inode,
2764 ret, ceph_cap_string(*got));
2769 * Check the offset we are writing up to against our current
2770 * max_size. If necessary, tell the MDS we want to write to
2773 static void check_max_size(struct inode *inode, loff_t endoff)
2775 struct ceph_inode_info *ci = ceph_inode(inode);
2778 /* do we need to explicitly request a larger max_size? */
2779 spin_lock(&ci->i_ceph_lock);
2780 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2781 dout("write %p at large endoff %llu, req max_size\n",
2783 ci->i_wanted_max_size = endoff;
2785 /* duplicate ceph_check_caps()'s logic */
2786 if (ci->i_auth_cap &&
2787 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2788 ci->i_wanted_max_size > ci->i_max_size &&
2789 ci->i_wanted_max_size > ci->i_requested_max_size)
2791 spin_unlock(&ci->i_ceph_lock);
2793 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2796 static inline int get_used_fmode(int caps)
2799 if (caps & CEPH_CAP_FILE_RD)
2800 fmode |= CEPH_FILE_MODE_RD;
2801 if (caps & CEPH_CAP_FILE_WR)
2802 fmode |= CEPH_FILE_MODE_WR;
2806 int ceph_try_get_caps(struct inode *inode, int need, int want,
2807 bool nonblock, int *got)
2811 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2812 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO |
2813 CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL |
2814 CEPH_CAP_ANY_DIR_OPS));
2816 ret = ceph_pool_perm_check(inode, need);
2821 flags = get_used_fmode(need | want);
2823 flags |= NON_BLOCKING;
2825 ret = try_get_cap_refs(inode, need, want, 0, flags, got);
2826 /* three special error codes */
2827 if (ret == -EAGAIN || ret == -EFBIG || ret == -ESTALE)
2833 * Wait for caps, and take cap references. If we can't get a WR cap
2834 * due to a small max_size, make sure we check_max_size (and possibly
2835 * ask the mds) so we don't get hung up indefinitely.
2837 int ceph_get_caps(struct file *filp, int need, int want, loff_t endoff, int *got)
2839 struct ceph_file_info *fi = filp->private_data;
2840 struct inode *inode = file_inode(filp);
2841 struct ceph_inode_info *ci = ceph_inode(inode);
2842 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2843 int ret, _got, flags;
2845 ret = ceph_pool_perm_check(inode, need);
2849 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2850 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2853 flags = get_used_fmode(need | want);
2856 flags &= CEPH_FILE_MODE_MASK;
2857 if (atomic_read(&fi->num_locks))
2858 flags |= CHECK_FILELOCK;
2860 ret = try_get_cap_refs(inode, need, want, endoff,
2862 WARN_ON_ONCE(ret == -EAGAIN);
2864 struct ceph_mds_client *mdsc = fsc->mdsc;
2866 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2868 cw.ino = ceph_ino(inode);
2869 cw.tgid = current->tgid;
2873 spin_lock(&mdsc->caps_list_lock);
2874 list_add(&cw.list, &mdsc->cap_wait_list);
2875 spin_unlock(&mdsc->caps_list_lock);
2877 /* make sure used fmode not timeout */
2878 ceph_get_fmode(ci, flags, FMODE_WAIT_BIAS);
2879 add_wait_queue(&ci->i_cap_wq, &wait);
2881 flags |= NON_BLOCKING;
2882 while (!(ret = try_get_cap_refs(inode, need, want,
2883 endoff, flags, &_got))) {
2884 if (signal_pending(current)) {
2888 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2891 remove_wait_queue(&ci->i_cap_wq, &wait);
2892 ceph_put_fmode(ci, flags, FMODE_WAIT_BIAS);
2894 spin_lock(&mdsc->caps_list_lock);
2896 spin_unlock(&mdsc->caps_list_lock);
2902 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2903 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2904 if (ret >= 0 && _got)
2905 ceph_put_cap_refs(ci, _got);
2910 if (ret == -EFBIG || ret == -ESTALE) {
2911 int ret2 = ceph_wait_on_async_create(inode);
2915 if (ret == -EFBIG) {
2916 check_max_size(inode, endoff);
2919 if (ret == -ESTALE) {
2920 /* session was killed, try renew caps */
2921 ret = ceph_renew_caps(inode, flags);
2928 if (S_ISREG(ci->vfs_inode.i_mode) &&
2929 ci->i_inline_version != CEPH_INLINE_NONE &&
2930 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2931 i_size_read(inode) > 0) {
2933 find_get_page(inode->i_mapping, 0);
2935 bool uptodate = PageUptodate(page);
2942 * drop cap refs first because getattr while
2943 * holding * caps refs can cause deadlock.
2945 ceph_put_cap_refs(ci, _got);
2949 * getattr request will bring inline data into
2952 ret = __ceph_do_getattr(inode, NULL,
2953 CEPH_STAT_CAP_INLINE_DATA,
2966 * Take cap refs. Caller must already know we hold at least one ref
2967 * on the caps in question or we don't know this is safe.
2969 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2971 spin_lock(&ci->i_ceph_lock);
2972 ceph_take_cap_refs(ci, caps, false);
2973 spin_unlock(&ci->i_ceph_lock);
2978 * drop cap_snap that is not associated with any snapshot.
2979 * we don't need to send FLUSHSNAP message for it.
2981 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2982 struct ceph_cap_snap *capsnap)
2984 if (!capsnap->need_flush &&
2985 !capsnap->writing && !capsnap->dirty_pages) {
2986 dout("dropping cap_snap %p follows %llu\n",
2987 capsnap, capsnap->follows);
2988 BUG_ON(capsnap->cap_flush.tid > 0);
2989 ceph_put_snap_context(capsnap->context);
2990 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2991 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2993 list_del(&capsnap->ci_item);
2994 ceph_put_cap_snap(capsnap);
3000 enum put_cap_refs_mode {
3001 PUT_CAP_REFS_SYNC = 0,
3002 PUT_CAP_REFS_NO_CHECK,
3009 * If we released the last ref on any given cap, call ceph_check_caps
3010 * to release (or schedule a release).
3012 * If we are releasing a WR cap (from a sync write), finalize any affected
3013 * cap_snap, and wake up any waiters.
3015 static void __ceph_put_cap_refs(struct ceph_inode_info *ci, int had,
3016 enum put_cap_refs_mode mode)
3018 struct inode *inode = &ci->vfs_inode;
3019 int last = 0, put = 0, flushsnaps = 0, wake = 0;
3020 bool check_flushsnaps = false;
3022 spin_lock(&ci->i_ceph_lock);
3023 if (had & CEPH_CAP_PIN)
3025 if (had & CEPH_CAP_FILE_RD)
3026 if (--ci->i_rd_ref == 0)
3028 if (had & CEPH_CAP_FILE_CACHE)
3029 if (--ci->i_rdcache_ref == 0)
3031 if (had & CEPH_CAP_FILE_EXCL)
3032 if (--ci->i_fx_ref == 0)
3034 if (had & CEPH_CAP_FILE_BUFFER) {
3035 if (--ci->i_wb_ref == 0) {
3037 /* put the ref held by ceph_take_cap_refs() */
3039 check_flushsnaps = true;
3041 dout("put_cap_refs %p wb %d -> %d (?)\n",
3042 inode, ci->i_wb_ref+1, ci->i_wb_ref);
3044 if (had & CEPH_CAP_FILE_WR) {
3045 if (--ci->i_wr_ref == 0) {
3047 check_flushsnaps = true;
3048 if (ci->i_wrbuffer_ref_head == 0 &&
3049 ci->i_dirty_caps == 0 &&
3050 ci->i_flushing_caps == 0) {
3051 BUG_ON(!ci->i_head_snapc);
3052 ceph_put_snap_context(ci->i_head_snapc);
3053 ci->i_head_snapc = NULL;
3055 /* see comment in __ceph_remove_cap() */
3056 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
3057 ceph_change_snap_realm(inode, NULL);
3060 if (check_flushsnaps && __ceph_have_pending_cap_snap(ci)) {
3061 struct ceph_cap_snap *capsnap =
3062 list_last_entry(&ci->i_cap_snaps,
3063 struct ceph_cap_snap,
3066 capsnap->writing = 0;
3067 if (ceph_try_drop_cap_snap(ci, capsnap))
3068 /* put the ref held by ceph_queue_cap_snap() */
3070 else if (__ceph_finish_cap_snap(ci, capsnap))
3074 spin_unlock(&ci->i_ceph_lock);
3076 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
3077 last ? " last" : "", put ? " put" : "");
3080 case PUT_CAP_REFS_SYNC:
3082 ceph_check_caps(ci, 0, NULL);
3083 else if (flushsnaps)
3084 ceph_flush_snaps(ci, NULL);
3086 case PUT_CAP_REFS_ASYNC:
3088 ceph_queue_check_caps(inode);
3089 else if (flushsnaps)
3090 ceph_queue_flush_snaps(inode);
3096 wake_up_all(&ci->i_cap_wq);
3101 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
3103 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_SYNC);
3106 void ceph_put_cap_refs_async(struct ceph_inode_info *ci, int had)
3108 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_ASYNC);
3111 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info *ci, int had)
3113 __ceph_put_cap_refs(ci, had, PUT_CAP_REFS_NO_CHECK);
3117 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3118 * context. Adjust per-snap dirty page accounting as appropriate.
3119 * Once all dirty data for a cap_snap is flushed, flush snapped file
3120 * metadata back to the MDS. If we dropped the last ref, call
3123 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3124 struct ceph_snap_context *snapc)
3126 struct inode *inode = &ci->vfs_inode;
3127 struct ceph_cap_snap *capsnap = NULL;
3131 bool flush_snaps = false;
3132 bool complete_capsnap = false;
3134 spin_lock(&ci->i_ceph_lock);
3135 ci->i_wrbuffer_ref -= nr;
3136 if (ci->i_wrbuffer_ref == 0) {
3141 if (ci->i_head_snapc == snapc) {
3142 ci->i_wrbuffer_ref_head -= nr;
3143 if (ci->i_wrbuffer_ref_head == 0 &&
3144 ci->i_wr_ref == 0 &&
3145 ci->i_dirty_caps == 0 &&
3146 ci->i_flushing_caps == 0) {
3147 BUG_ON(!ci->i_head_snapc);
3148 ceph_put_snap_context(ci->i_head_snapc);
3149 ci->i_head_snapc = NULL;
3151 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3153 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3154 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3155 last ? " LAST" : "");
3157 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3158 if (capsnap->context == snapc) {
3164 capsnap->dirty_pages -= nr;
3165 if (capsnap->dirty_pages == 0) {
3166 complete_capsnap = true;
3167 if (!capsnap->writing) {
3168 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3171 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3176 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3177 " snap %lld %d/%d -> %d/%d %s%s\n",
3178 inode, capsnap, capsnap->context->seq,
3179 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3180 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3181 last ? " (wrbuffer last)" : "",
3182 complete_capsnap ? " (complete capsnap)" : "");
3185 spin_unlock(&ci->i_ceph_lock);
3188 ceph_check_caps(ci, 0, NULL);
3189 } else if (flush_snaps) {
3190 ceph_flush_snaps(ci, NULL);
3192 if (complete_capsnap)
3193 wake_up_all(&ci->i_cap_wq);
3200 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3202 static void invalidate_aliases(struct inode *inode)
3204 struct dentry *dn, *prev = NULL;
3206 dout("invalidate_aliases inode %p\n", inode);
3207 d_prune_aliases(inode);
3209 * For non-directory inode, d_find_alias() only returns
3210 * hashed dentry. After calling d_invalidate(), the
3211 * dentry becomes unhashed.
3213 * For directory inode, d_find_alias() can return
3214 * unhashed dentry. But directory inode should have
3215 * one alias at most.
3217 while ((dn = d_find_alias(inode))) {
3231 struct cap_extra_info {
3232 struct ceph_string *pool_ns;
3242 /* currently issued */
3244 struct timespec64 btime;
3248 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3249 * actually be a revocation if it specifies a smaller cap set.)
3251 * caller holds s_mutex and i_ceph_lock, we drop both.
3253 static void handle_cap_grant(struct inode *inode,
3254 struct ceph_mds_session *session,
3255 struct ceph_cap *cap,
3256 struct ceph_mds_caps *grant,
3257 struct ceph_buffer *xattr_buf,
3258 struct cap_extra_info *extra_info)
3259 __releases(ci->i_ceph_lock)
3260 __releases(session->s_mdsc->snap_rwsem)
3262 struct ceph_inode_info *ci = ceph_inode(inode);
3263 int seq = le32_to_cpu(grant->seq);
3264 int newcaps = le32_to_cpu(grant->caps);
3265 int used, wanted, dirty;
3266 u64 size = le64_to_cpu(grant->size);
3267 u64 max_size = le64_to_cpu(grant->max_size);
3268 unsigned char check_caps = 0;
3269 bool was_stale = cap->cap_gen < atomic_read(&session->s_cap_gen);
3271 bool writeback = false;
3272 bool queue_trunc = false;
3273 bool queue_invalidate = false;
3274 bool deleted_inode = false;
3275 bool fill_inline = false;
3277 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3278 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3279 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3280 i_size_read(inode));
3284 * If CACHE is being revoked, and we have no dirty buffers,
3285 * try to invalidate (once). (If there are dirty buffers, we
3286 * will invalidate _after_ writeback.)
3288 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3289 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3290 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3291 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3292 if (try_nonblocking_invalidate(inode)) {
3293 /* there were locked pages.. invalidate later
3294 in a separate thread. */
3295 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3296 queue_invalidate = true;
3297 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3303 cap->issued = cap->implemented = CEPH_CAP_PIN;
3306 * auth mds of the inode changed. we received the cap export message,
3307 * but still haven't received the cap import message. handle_cap_export
3308 * updated the new auth MDS' cap.
3310 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3311 * that was sent before the cap import message. So don't remove caps.
3313 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3314 WARN_ON(cap != ci->i_auth_cap);
3315 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3317 newcaps |= cap->issued;
3320 /* side effects now are allowed */
3321 cap->cap_gen = atomic_read(&session->s_cap_gen);
3324 __check_cap_issue(ci, cap, newcaps);
3326 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3328 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3329 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3330 umode_t mode = le32_to_cpu(grant->mode);
3332 if (inode_wrong_type(inode, mode))
3333 pr_warn_once("inode type changed! (ino %llx.%llx is 0%o, mds says 0%o)\n",
3334 ceph_vinop(inode), inode->i_mode, mode);
3336 inode->i_mode = mode;
3337 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3338 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3339 ci->i_btime = extra_info->btime;
3340 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3341 from_kuid(&init_user_ns, inode->i_uid),
3342 from_kgid(&init_user_ns, inode->i_gid));
3345 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3346 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3347 set_nlink(inode, le32_to_cpu(grant->nlink));
3348 if (inode->i_nlink == 0 &&
3349 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3350 deleted_inode = true;
3353 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3355 int len = le32_to_cpu(grant->xattr_len);
3356 u64 version = le64_to_cpu(grant->xattr_version);
3358 if (version > ci->i_xattrs.version) {
3359 dout(" got new xattrs v%llu on %p len %d\n",
3360 version, inode, len);
3361 if (ci->i_xattrs.blob)
3362 ceph_buffer_put(ci->i_xattrs.blob);
3363 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3364 ci->i_xattrs.version = version;
3365 ceph_forget_all_cached_acls(inode);
3366 ceph_security_invalidate_secctx(inode);
3370 if (newcaps & CEPH_CAP_ANY_RD) {
3371 struct timespec64 mtime, atime, ctime;
3372 /* ctime/mtime/atime? */
3373 ceph_decode_timespec64(&mtime, &grant->mtime);
3374 ceph_decode_timespec64(&atime, &grant->atime);
3375 ceph_decode_timespec64(&ctime, &grant->ctime);
3376 ceph_fill_file_time(inode, extra_info->issued,
3377 le32_to_cpu(grant->time_warp_seq),
3378 &ctime, &mtime, &atime);
3381 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3382 ci->i_files = extra_info->nfiles;
3383 ci->i_subdirs = extra_info->nsubdirs;
3386 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3387 /* file layout may have changed */
3388 s64 old_pool = ci->i_layout.pool_id;
3389 struct ceph_string *old_ns;
3391 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3392 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3393 lockdep_is_held(&ci->i_ceph_lock));
3394 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3396 if (ci->i_layout.pool_id != old_pool ||
3397 extra_info->pool_ns != old_ns)
3398 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3400 extra_info->pool_ns = old_ns;
3402 /* size/truncate_seq? */
3403 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3404 le32_to_cpu(grant->truncate_seq),
3405 le64_to_cpu(grant->truncate_size),
3409 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3410 if (max_size != ci->i_max_size) {
3411 dout("max_size %lld -> %llu\n",
3412 ci->i_max_size, max_size);
3413 ci->i_max_size = max_size;
3414 if (max_size >= ci->i_wanted_max_size) {
3415 ci->i_wanted_max_size = 0; /* reset */
3416 ci->i_requested_max_size = 0;
3422 /* check cap bits */
3423 wanted = __ceph_caps_wanted(ci);
3424 used = __ceph_caps_used(ci);
3425 dirty = __ceph_caps_dirty(ci);
3426 dout(" my wanted = %s, used = %s, dirty %s\n",
3427 ceph_cap_string(wanted),
3428 ceph_cap_string(used),
3429 ceph_cap_string(dirty));
3431 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3432 (wanted & ~(cap->mds_wanted | newcaps))) {
3434 * If mds is importing cap, prior cap messages that update
3435 * 'wanted' may get dropped by mds (migrate seq mismatch).
3437 * We don't send cap message to update 'wanted' if what we
3438 * want are already issued. If mds revokes caps, cap message
3439 * that releases caps also tells mds what we want. But if
3440 * caps got revoked by mds forcedly (session stale). We may
3441 * haven't told mds what we want.
3446 /* revocation, grant, or no-op? */
3447 if (cap->issued & ~newcaps) {
3448 int revoking = cap->issued & ~newcaps;
3450 dout("revocation: %s -> %s (revoking %s)\n",
3451 ceph_cap_string(cap->issued),
3452 ceph_cap_string(newcaps),
3453 ceph_cap_string(revoking));
3454 if (S_ISREG(inode->i_mode) &&
3455 (revoking & used & CEPH_CAP_FILE_BUFFER))
3456 writeback = true; /* initiate writeback; will delay ack */
3457 else if (queue_invalidate &&
3458 revoking == CEPH_CAP_FILE_CACHE &&
3459 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3460 ; /* do nothing yet, invalidation will be queued */
3461 else if (cap == ci->i_auth_cap)
3462 check_caps = 1; /* check auth cap only */
3464 check_caps = 2; /* check all caps */
3465 cap->issued = newcaps;
3466 cap->implemented |= newcaps;
3467 } else if (cap->issued == newcaps) {
3468 dout("caps unchanged: %s -> %s\n",
3469 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3471 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3472 ceph_cap_string(newcaps));
3473 /* non-auth MDS is revoking the newly grant caps ? */
3474 if (cap == ci->i_auth_cap &&
3475 __ceph_caps_revoking_other(ci, cap, newcaps))
3478 cap->issued = newcaps;
3479 cap->implemented |= newcaps; /* add bits only, to
3480 * avoid stepping on a
3481 * pending revocation */
3484 BUG_ON(cap->issued & ~cap->implemented);
3486 if (extra_info->inline_version > 0 &&
3487 extra_info->inline_version >= ci->i_inline_version) {
3488 ci->i_inline_version = extra_info->inline_version;
3489 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3490 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3494 if (ci->i_auth_cap == cap &&
3495 le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3496 if (newcaps & ~extra_info->issued)
3499 if (ci->i_requested_max_size > max_size ||
3500 !(le32_to_cpu(grant->wanted) & CEPH_CAP_ANY_FILE_WR)) {
3501 /* re-request max_size if necessary */
3502 ci->i_requested_max_size = 0;
3506 ceph_kick_flushing_inode_caps(session, ci);
3507 spin_unlock(&ci->i_ceph_lock);
3508 up_read(&session->s_mdsc->snap_rwsem);
3510 spin_unlock(&ci->i_ceph_lock);
3514 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3515 extra_info->inline_len);
3518 ceph_queue_vmtruncate(inode);
3522 * queue inode for writeback: we can't actually call
3523 * filemap_write_and_wait, etc. from message handler
3526 ceph_queue_writeback(inode);
3527 if (queue_invalidate)
3528 ceph_queue_invalidate(inode);
3530 invalidate_aliases(inode);
3532 wake_up_all(&ci->i_cap_wq);
3534 mutex_unlock(&session->s_mutex);
3535 if (check_caps == 1)
3536 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_NOINVAL,
3538 else if (check_caps == 2)
3539 ceph_check_caps(ci, CHECK_CAPS_NOINVAL, session);
3543 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3544 * MDS has been safely committed.
3546 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3547 struct ceph_mds_caps *m,
3548 struct ceph_mds_session *session,
3549 struct ceph_cap *cap)
3550 __releases(ci->i_ceph_lock)
3552 struct ceph_inode_info *ci = ceph_inode(inode);
3553 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3554 struct ceph_cap_flush *cf, *tmp_cf;
3555 LIST_HEAD(to_remove);
3556 unsigned seq = le32_to_cpu(m->seq);
3557 int dirty = le32_to_cpu(m->dirty);
3560 bool wake_ci = false;
3561 bool wake_mdsc = false;
3563 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3564 /* Is this the one that was flushed? */
3565 if (cf->tid == flush_tid)
3568 /* Is this a capsnap? */
3572 if (cf->tid <= flush_tid) {
3574 * An earlier or current tid. The FLUSH_ACK should
3575 * represent a superset of this flush's caps.
3577 wake_ci |= __detach_cap_flush_from_ci(ci, cf);
3578 list_add_tail(&cf->i_list, &to_remove);
3581 * This is a later one. Any caps in it are still dirty
3582 * so don't count them as cleaned.
3584 cleaned &= ~cf->caps;
3590 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3591 " flushing %s -> %s\n",
3592 inode, session->s_mds, seq, ceph_cap_string(dirty),
3593 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3594 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3596 if (list_empty(&to_remove) && !cleaned)
3599 ci->i_flushing_caps &= ~cleaned;
3601 spin_lock(&mdsc->cap_dirty_lock);
3603 list_for_each_entry(cf, &to_remove, i_list)
3604 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc, cf);
3606 if (ci->i_flushing_caps == 0) {
3607 if (list_empty(&ci->i_cap_flush_list)) {
3608 list_del_init(&ci->i_flushing_item);
3609 if (!list_empty(&session->s_cap_flushing)) {
3610 dout(" mds%d still flushing cap on %p\n",
3612 &list_first_entry(&session->s_cap_flushing,
3613 struct ceph_inode_info,
3614 i_flushing_item)->vfs_inode);
3617 mdsc->num_cap_flushing--;
3618 dout(" inode %p now !flushing\n", inode);
3620 if (ci->i_dirty_caps == 0) {
3621 dout(" inode %p now clean\n", inode);
3622 BUG_ON(!list_empty(&ci->i_dirty_item));
3624 if (ci->i_wr_ref == 0 &&
3625 ci->i_wrbuffer_ref_head == 0) {
3626 BUG_ON(!ci->i_head_snapc);
3627 ceph_put_snap_context(ci->i_head_snapc);
3628 ci->i_head_snapc = NULL;
3631 BUG_ON(list_empty(&ci->i_dirty_item));
3634 spin_unlock(&mdsc->cap_dirty_lock);
3637 spin_unlock(&ci->i_ceph_lock);
3639 while (!list_empty(&to_remove)) {
3640 cf = list_first_entry(&to_remove,
3641 struct ceph_cap_flush, i_list);
3642 list_del_init(&cf->i_list);
3643 if (!cf->is_capsnap)
3644 ceph_free_cap_flush(cf);
3648 wake_up_all(&ci->i_cap_wq);
3650 wake_up_all(&mdsc->cap_flushing_wq);
3656 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3657 * throw away our cap_snap.
3659 * Caller hold s_mutex.
3661 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3662 struct ceph_mds_caps *m,
3663 struct ceph_mds_session *session)
3665 struct ceph_inode_info *ci = ceph_inode(inode);
3666 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3667 u64 follows = le64_to_cpu(m->snap_follows);
3668 struct ceph_cap_snap *capsnap;
3669 bool flushed = false;
3670 bool wake_ci = false;
3671 bool wake_mdsc = false;
3673 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3674 inode, ci, session->s_mds, follows);
3676 spin_lock(&ci->i_ceph_lock);
3677 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3678 if (capsnap->follows == follows) {
3679 if (capsnap->cap_flush.tid != flush_tid) {
3680 dout(" cap_snap %p follows %lld tid %lld !="
3681 " %lld\n", capsnap, follows,
3682 flush_tid, capsnap->cap_flush.tid);
3688 dout(" skipping cap_snap %p follows %lld\n",
3689 capsnap, capsnap->follows);
3693 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3694 dout(" removing %p cap_snap %p follows %lld\n",
3695 inode, capsnap, follows);
3696 list_del(&capsnap->ci_item);
3697 wake_ci |= __detach_cap_flush_from_ci(ci, &capsnap->cap_flush);
3699 spin_lock(&mdsc->cap_dirty_lock);
3701 if (list_empty(&ci->i_cap_flush_list))
3702 list_del_init(&ci->i_flushing_item);
3704 wake_mdsc |= __detach_cap_flush_from_mdsc(mdsc,
3705 &capsnap->cap_flush);
3706 spin_unlock(&mdsc->cap_dirty_lock);
3708 spin_unlock(&ci->i_ceph_lock);
3710 ceph_put_snap_context(capsnap->context);
3711 ceph_put_cap_snap(capsnap);
3713 wake_up_all(&ci->i_cap_wq);
3715 wake_up_all(&mdsc->cap_flushing_wq);
3721 * Handle TRUNC from MDS, indicating file truncation.
3723 * caller hold s_mutex.
3725 static bool handle_cap_trunc(struct inode *inode,
3726 struct ceph_mds_caps *trunc,
3727 struct ceph_mds_session *session)
3729 struct ceph_inode_info *ci = ceph_inode(inode);
3730 int mds = session->s_mds;
3731 int seq = le32_to_cpu(trunc->seq);
3732 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3733 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3734 u64 size = le64_to_cpu(trunc->size);
3735 int implemented = 0;
3736 int dirty = __ceph_caps_dirty(ci);
3737 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3738 bool queue_trunc = false;
3740 lockdep_assert_held(&ci->i_ceph_lock);
3742 issued |= implemented | dirty;
3744 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3745 inode, mds, seq, truncate_size, truncate_seq);
3746 queue_trunc = ceph_fill_file_size(inode, issued,
3747 truncate_seq, truncate_size, size);
3752 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3753 * different one. If we are the most recent migration we've seen (as
3754 * indicated by mseq), make note of the migrating cap bits for the
3755 * duration (until we see the corresponding IMPORT).
3757 * caller holds s_mutex
3759 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3760 struct ceph_mds_cap_peer *ph,
3761 struct ceph_mds_session *session)
3763 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3764 struct ceph_mds_session *tsession = NULL;
3765 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3766 struct ceph_inode_info *ci = ceph_inode(inode);
3768 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3769 unsigned t_seq, t_mseq;
3771 int mds = session->s_mds;
3774 t_cap_id = le64_to_cpu(ph->cap_id);
3775 t_seq = le32_to_cpu(ph->seq);
3776 t_mseq = le32_to_cpu(ph->mseq);
3777 target = le32_to_cpu(ph->mds);
3779 t_cap_id = t_seq = t_mseq = 0;
3783 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3784 inode, ci, mds, mseq, target);
3786 spin_lock(&ci->i_ceph_lock);
3787 cap = __get_cap_for_mds(ci, mds);
3788 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3792 __ceph_remove_cap(cap, false);
3797 * now we know we haven't received the cap import message yet
3798 * because the exported cap still exist.
3801 issued = cap->issued;
3802 if (issued != cap->implemented)
3803 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3804 "ino (%llx.%llx) mds%d seq %d mseq %d "
3805 "issued %s implemented %s\n",
3806 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3807 ceph_cap_string(issued),
3808 ceph_cap_string(cap->implemented));
3811 tcap = __get_cap_for_mds(ci, target);
3813 /* already have caps from the target */
3814 if (tcap->cap_id == t_cap_id &&
3815 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3816 dout(" updating import cap %p mds%d\n", tcap, target);
3817 tcap->cap_id = t_cap_id;
3818 tcap->seq = t_seq - 1;
3819 tcap->issue_seq = t_seq - 1;
3820 tcap->issued |= issued;
3821 tcap->implemented |= issued;
3822 if (cap == ci->i_auth_cap) {
3823 ci->i_auth_cap = tcap;
3824 change_auth_cap_ses(ci, tcap->session);
3827 __ceph_remove_cap(cap, false);
3829 } else if (tsession) {
3830 /* add placeholder for the export tagert */
3831 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3833 ceph_add_cap(inode, tsession, t_cap_id, issued, 0,
3834 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3836 if (!list_empty(&ci->i_cap_flush_list) &&
3837 ci->i_auth_cap == tcap) {
3838 spin_lock(&mdsc->cap_dirty_lock);
3839 list_move_tail(&ci->i_flushing_item,
3840 &tcap->session->s_cap_flushing);
3841 spin_unlock(&mdsc->cap_dirty_lock);
3844 __ceph_remove_cap(cap, false);
3848 spin_unlock(&ci->i_ceph_lock);
3849 mutex_unlock(&session->s_mutex);
3851 /* open target session */
3852 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3853 if (!IS_ERR(tsession)) {
3855 mutex_lock(&session->s_mutex);
3856 mutex_lock_nested(&tsession->s_mutex,
3857 SINGLE_DEPTH_NESTING);
3859 mutex_lock(&tsession->s_mutex);
3860 mutex_lock_nested(&session->s_mutex,
3861 SINGLE_DEPTH_NESTING);
3863 new_cap = ceph_get_cap(mdsc, NULL);
3868 mutex_lock(&session->s_mutex);
3873 spin_unlock(&ci->i_ceph_lock);
3874 mutex_unlock(&session->s_mutex);
3876 mutex_unlock(&tsession->s_mutex);
3877 ceph_put_mds_session(tsession);
3880 ceph_put_cap(mdsc, new_cap);
3884 * Handle cap IMPORT.
3886 * caller holds s_mutex. acquires i_ceph_lock
3888 static void handle_cap_import(struct ceph_mds_client *mdsc,
3889 struct inode *inode, struct ceph_mds_caps *im,
3890 struct ceph_mds_cap_peer *ph,
3891 struct ceph_mds_session *session,
3892 struct ceph_cap **target_cap, int *old_issued)
3894 struct ceph_inode_info *ci = ceph_inode(inode);
3895 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3896 int mds = session->s_mds;
3898 unsigned caps = le32_to_cpu(im->caps);
3899 unsigned wanted = le32_to_cpu(im->wanted);
3900 unsigned seq = le32_to_cpu(im->seq);
3901 unsigned mseq = le32_to_cpu(im->migrate_seq);
3902 u64 realmino = le64_to_cpu(im->realm);
3903 u64 cap_id = le64_to_cpu(im->cap_id);
3908 p_cap_id = le64_to_cpu(ph->cap_id);
3909 peer = le32_to_cpu(ph->mds);
3915 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3916 inode, ci, mds, mseq, peer);
3918 cap = __get_cap_for_mds(ci, mds);
3921 spin_unlock(&ci->i_ceph_lock);
3922 new_cap = ceph_get_cap(mdsc, NULL);
3923 spin_lock(&ci->i_ceph_lock);
3929 ceph_put_cap(mdsc, new_cap);
3934 __ceph_caps_issued(ci, &issued);
3935 issued |= __ceph_caps_dirty(ci);
3937 ceph_add_cap(inode, session, cap_id, caps, wanted, seq, mseq,
3938 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3940 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3941 if (ocap && ocap->cap_id == p_cap_id) {
3942 dout(" remove export cap %p mds%d flags %d\n",
3943 ocap, peer, ph->flags);
3944 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3945 (ocap->seq != le32_to_cpu(ph->seq) ||
3946 ocap->mseq != le32_to_cpu(ph->mseq))) {
3947 pr_err_ratelimited("handle_cap_import: "
3948 "mismatched seq/mseq: ino (%llx.%llx) "
3949 "mds%d seq %d mseq %d importer mds%d "
3950 "has peer seq %d mseq %d\n",
3951 ceph_vinop(inode), peer, ocap->seq,
3952 ocap->mseq, mds, le32_to_cpu(ph->seq),
3953 le32_to_cpu(ph->mseq));
3955 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3958 *old_issued = issued;
3963 * Handle a caps message from the MDS.
3965 * Identify the appropriate session, inode, and call the right handler
3966 * based on the cap op.
3968 void ceph_handle_caps(struct ceph_mds_session *session,
3969 struct ceph_msg *msg)
3971 struct ceph_mds_client *mdsc = session->s_mdsc;
3972 struct inode *inode;
3973 struct ceph_inode_info *ci;
3974 struct ceph_cap *cap;
3975 struct ceph_mds_caps *h;
3976 struct ceph_mds_cap_peer *peer = NULL;
3977 struct ceph_snap_realm *realm = NULL;
3979 int msg_version = le16_to_cpu(msg->hdr.version);
3981 struct ceph_vino vino;
3983 size_t snaptrace_len;
3985 struct cap_extra_info extra_info = {};
3988 dout("handle_caps from mds%d\n", session->s_mds);
3991 end = msg->front.iov_base + msg->front.iov_len;
3992 if (msg->front.iov_len < sizeof(*h))
3994 h = msg->front.iov_base;
3995 op = le32_to_cpu(h->op);
3996 vino.ino = le64_to_cpu(h->ino);
3997 vino.snap = CEPH_NOSNAP;
3998 seq = le32_to_cpu(h->seq);
3999 mseq = le32_to_cpu(h->migrate_seq);
4002 snaptrace_len = le32_to_cpu(h->snap_trace_len);
4003 p = snaptrace + snaptrace_len;
4005 if (msg_version >= 2) {
4007 ceph_decode_32_safe(&p, end, flock_len, bad);
4008 if (p + flock_len > end)
4013 if (msg_version >= 3) {
4014 if (op == CEPH_CAP_OP_IMPORT) {
4015 if (p + sizeof(*peer) > end)
4019 } else if (op == CEPH_CAP_OP_EXPORT) {
4020 /* recorded in unused fields */
4021 peer = (void *)&h->size;
4025 if (msg_version >= 4) {
4026 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
4027 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
4028 if (p + extra_info.inline_len > end)
4030 extra_info.inline_data = p;
4031 p += extra_info.inline_len;
4034 if (msg_version >= 5) {
4035 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
4038 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
4039 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
4042 if (msg_version >= 8) {
4046 ceph_decode_skip_64(&p, end, bad); // flush_tid
4048 ceph_decode_skip_32(&p, end, bad); // caller_uid
4049 ceph_decode_skip_32(&p, end, bad); // caller_gid
4051 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
4052 if (pool_ns_len > 0) {
4053 ceph_decode_need(&p, end, pool_ns_len, bad);
4054 extra_info.pool_ns =
4055 ceph_find_or_create_string(p, pool_ns_len);
4060 if (msg_version >= 9) {
4061 struct ceph_timespec *btime;
4063 if (p + sizeof(*btime) > end)
4066 ceph_decode_timespec64(&extra_info.btime, btime);
4067 p += sizeof(*btime);
4068 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
4071 if (msg_version >= 11) {
4073 ceph_decode_skip_32(&p, end, bad); // flags
4075 extra_info.dirstat_valid = true;
4076 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
4077 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
4081 inode = ceph_find_inode(mdsc->fsc->sb, vino);
4082 ci = ceph_inode(inode);
4083 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
4086 mutex_lock(&session->s_mutex);
4087 inc_session_sequence(session);
4088 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
4092 dout(" i don't have ino %llx\n", vino.ino);
4094 if (op == CEPH_CAP_OP_IMPORT) {
4095 cap = ceph_get_cap(mdsc, NULL);
4096 cap->cap_ino = vino.ino;
4097 cap->queue_release = 1;
4098 cap->cap_id = le64_to_cpu(h->cap_id);
4101 cap->issue_seq = seq;
4102 spin_lock(&session->s_cap_lock);
4103 __ceph_queue_cap_release(session, cap);
4104 spin_unlock(&session->s_cap_lock);
4106 goto flush_cap_releases;
4109 /* these will work even if we don't have a cap yet */
4111 case CEPH_CAP_OP_FLUSHSNAP_ACK:
4112 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
4116 case CEPH_CAP_OP_EXPORT:
4117 handle_cap_export(inode, h, peer, session);
4120 case CEPH_CAP_OP_IMPORT:
4122 if (snaptrace_len) {
4123 down_write(&mdsc->snap_rwsem);
4124 ceph_update_snap_trace(mdsc, snaptrace,
4125 snaptrace + snaptrace_len,
4127 downgrade_write(&mdsc->snap_rwsem);
4129 down_read(&mdsc->snap_rwsem);
4131 spin_lock(&ci->i_ceph_lock);
4132 handle_cap_import(mdsc, inode, h, peer, session,
4133 &cap, &extra_info.issued);
4134 handle_cap_grant(inode, session, cap,
4135 h, msg->middle, &extra_info);
4137 ceph_put_snap_realm(mdsc, realm);
4141 /* the rest require a cap */
4142 spin_lock(&ci->i_ceph_lock);
4143 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4145 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4146 inode, ceph_ino(inode), ceph_snap(inode),
4148 spin_unlock(&ci->i_ceph_lock);
4149 goto flush_cap_releases;
4152 /* note that each of these drops i_ceph_lock for us */
4154 case CEPH_CAP_OP_REVOKE:
4155 case CEPH_CAP_OP_GRANT:
4156 __ceph_caps_issued(ci, &extra_info.issued);
4157 extra_info.issued |= __ceph_caps_dirty(ci);
4158 handle_cap_grant(inode, session, cap,
4159 h, msg->middle, &extra_info);
4162 case CEPH_CAP_OP_FLUSH_ACK:
4163 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4167 case CEPH_CAP_OP_TRUNC:
4168 queue_trunc = handle_cap_trunc(inode, h, session);
4169 spin_unlock(&ci->i_ceph_lock);
4171 ceph_queue_vmtruncate(inode);
4175 spin_unlock(&ci->i_ceph_lock);
4176 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4177 ceph_cap_op_name(op));
4181 mutex_unlock(&session->s_mutex);
4185 ceph_put_string(extra_info.pool_ns);
4190 * send any cap release message to try to move things
4191 * along for the mds (who clearly thinks we still have this
4194 ceph_flush_cap_releases(mdsc, session);
4198 pr_err("ceph_handle_caps: corrupt message\n");
4204 * Delayed work handler to process end of delayed cap release LRU list.
4206 * If new caps are added to the list while processing it, these won't get
4207 * processed in this run. In this case, the ci->i_hold_caps_max will be
4208 * returned so that the work can be scheduled accordingly.
4210 unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4212 struct inode *inode;
4213 struct ceph_inode_info *ci;
4214 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
4215 unsigned long delay_max = opt->caps_wanted_delay_max * HZ;
4216 unsigned long loop_start = jiffies;
4217 unsigned long delay = 0;
4219 dout("check_delayed_caps\n");
4220 spin_lock(&mdsc->cap_delay_lock);
4221 while (!list_empty(&mdsc->cap_delay_list)) {
4222 ci = list_first_entry(&mdsc->cap_delay_list,
4223 struct ceph_inode_info,
4225 if (time_before(loop_start, ci->i_hold_caps_max - delay_max)) {
4226 dout("%s caps added recently. Exiting loop", __func__);
4227 delay = ci->i_hold_caps_max;
4230 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4231 time_before(jiffies, ci->i_hold_caps_max))
4233 list_del_init(&ci->i_cap_delay_list);
4235 inode = igrab(&ci->vfs_inode);
4237 spin_unlock(&mdsc->cap_delay_lock);
4238 dout("check_delayed_caps on %p\n", inode);
4239 ceph_check_caps(ci, 0, NULL);
4241 spin_lock(&mdsc->cap_delay_lock);
4244 spin_unlock(&mdsc->cap_delay_lock);
4250 * Flush all dirty caps to the mds
4252 static void flush_dirty_session_caps(struct ceph_mds_session *s)
4254 struct ceph_mds_client *mdsc = s->s_mdsc;
4255 struct ceph_inode_info *ci;
4256 struct inode *inode;
4258 dout("flush_dirty_caps\n");
4259 spin_lock(&mdsc->cap_dirty_lock);
4260 while (!list_empty(&s->s_cap_dirty)) {
4261 ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
4263 inode = &ci->vfs_inode;
4265 dout("flush_dirty_caps %p\n", inode);
4266 spin_unlock(&mdsc->cap_dirty_lock);
4267 ceph_check_caps(ci, CHECK_CAPS_FLUSH, NULL);
4269 spin_lock(&mdsc->cap_dirty_lock);
4271 spin_unlock(&mdsc->cap_dirty_lock);
4272 dout("flush_dirty_caps done\n");
4275 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4277 ceph_mdsc_iterate_sessions(mdsc, flush_dirty_session_caps, true);
4280 void __ceph_touch_fmode(struct ceph_inode_info *ci,
4281 struct ceph_mds_client *mdsc, int fmode)
4283 unsigned long now = jiffies;
4284 if (fmode & CEPH_FILE_MODE_RD)
4285 ci->i_last_rd = now;
4286 if (fmode & CEPH_FILE_MODE_WR)
4287 ci->i_last_wr = now;
4288 /* queue periodic check */
4290 __ceph_is_any_real_caps(ci) &&
4291 list_empty(&ci->i_cap_delay_list))
4292 __cap_delay_requeue(mdsc, ci);
4295 void ceph_get_fmode(struct ceph_inode_info *ci, int fmode, int count)
4297 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4298 int bits = (fmode << 1) | 1;
4299 bool is_opened = false;
4303 atomic64_inc(&mdsc->metric.opened_files);
4305 spin_lock(&ci->i_ceph_lock);
4306 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4307 if (bits & (1 << i))
4308 ci->i_nr_by_mode[i] += count;
4311 * If any of the mode ref is larger than 1,
4312 * that means it has been already opened by
4313 * others. Just skip checking the PIN ref.
4315 if (i && ci->i_nr_by_mode[i] > 1)
4320 percpu_counter_inc(&mdsc->metric.opened_inodes);
4321 spin_unlock(&ci->i_ceph_lock);
4325 * Drop open file reference. If we were the last open file,
4326 * we may need to release capabilities to the MDS (or schedule
4327 * their delayed release).
4329 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode, int count)
4331 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(ci->vfs_inode.i_sb);
4332 int bits = (fmode << 1) | 1;
4333 bool is_closed = true;
4337 atomic64_dec(&mdsc->metric.opened_files);
4339 spin_lock(&ci->i_ceph_lock);
4340 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4341 if (bits & (1 << i)) {
4342 BUG_ON(ci->i_nr_by_mode[i] < count);
4343 ci->i_nr_by_mode[i] -= count;
4347 * If any of the mode ref is not 0 after
4348 * decreased, that means it is still opened
4349 * by others. Just skip checking the PIN ref.
4351 if (i && ci->i_nr_by_mode[i])
4356 percpu_counter_dec(&mdsc->metric.opened_inodes);
4357 spin_unlock(&ci->i_ceph_lock);
4361 * For a soon-to-be unlinked file, drop the LINK caps. If it
4362 * looks like the link count will hit 0, drop any other caps (other
4363 * than PIN) we don't specifically want (due to the file still being
4366 int ceph_drop_caps_for_unlink(struct inode *inode)
4368 struct ceph_inode_info *ci = ceph_inode(inode);
4369 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4371 spin_lock(&ci->i_ceph_lock);
4372 if (inode->i_nlink == 1) {
4373 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4375 if (__ceph_caps_dirty(ci)) {
4376 struct ceph_mds_client *mdsc =
4377 ceph_inode_to_client(inode)->mdsc;
4378 __cap_delay_requeue_front(mdsc, ci);
4381 spin_unlock(&ci->i_ceph_lock);
4386 * Helpers for embedding cap and dentry lease releases into mds
4389 * @force is used by dentry_release (below) to force inclusion of a
4390 * record for the directory inode, even when there aren't any caps to
4393 int ceph_encode_inode_release(void **p, struct inode *inode,
4394 int mds, int drop, int unless, int force)
4396 struct ceph_inode_info *ci = ceph_inode(inode);
4397 struct ceph_cap *cap;
4398 struct ceph_mds_request_release *rel = *p;
4402 spin_lock(&ci->i_ceph_lock);
4403 used = __ceph_caps_used(ci);
4404 dirty = __ceph_caps_dirty(ci);
4406 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4407 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4408 ceph_cap_string(unless));
4410 /* only drop unused, clean caps */
4411 drop &= ~(used | dirty);
4413 cap = __get_cap_for_mds(ci, mds);
4414 if (cap && __cap_is_valid(cap)) {
4415 unless &= cap->issued;
4417 if (unless & CEPH_CAP_AUTH_EXCL)
4418 drop &= ~CEPH_CAP_AUTH_SHARED;
4419 if (unless & CEPH_CAP_LINK_EXCL)
4420 drop &= ~CEPH_CAP_LINK_SHARED;
4421 if (unless & CEPH_CAP_XATTR_EXCL)
4422 drop &= ~CEPH_CAP_XATTR_SHARED;
4423 if (unless & CEPH_CAP_FILE_EXCL)
4424 drop &= ~CEPH_CAP_FILE_SHARED;
4427 if (force || (cap->issued & drop)) {
4428 if (cap->issued & drop) {
4429 int wanted = __ceph_caps_wanted(ci);
4430 dout("encode_inode_release %p cap %p "
4431 "%s -> %s, wanted %s -> %s\n", inode, cap,
4432 ceph_cap_string(cap->issued),
4433 ceph_cap_string(cap->issued & ~drop),
4434 ceph_cap_string(cap->mds_wanted),
4435 ceph_cap_string(wanted));
4437 cap->issued &= ~drop;
4438 cap->implemented &= ~drop;
4439 cap->mds_wanted = wanted;
4440 if (cap == ci->i_auth_cap &&
4441 !(wanted & CEPH_CAP_ANY_FILE_WR))
4442 ci->i_requested_max_size = 0;
4444 dout("encode_inode_release %p cap %p %s"
4445 " (force)\n", inode, cap,
4446 ceph_cap_string(cap->issued));
4449 rel->ino = cpu_to_le64(ceph_ino(inode));
4450 rel->cap_id = cpu_to_le64(cap->cap_id);
4451 rel->seq = cpu_to_le32(cap->seq);
4452 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4453 rel->mseq = cpu_to_le32(cap->mseq);
4454 rel->caps = cpu_to_le32(cap->implemented);
4455 rel->wanted = cpu_to_le32(cap->mds_wanted);
4461 dout("encode_inode_release %p cap %p %s (noop)\n",
4462 inode, cap, ceph_cap_string(cap->issued));
4465 spin_unlock(&ci->i_ceph_lock);
4469 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4471 int mds, int drop, int unless)
4473 struct dentry *parent = NULL;
4474 struct ceph_mds_request_release *rel = *p;
4475 struct ceph_dentry_info *di = ceph_dentry(dentry);
4480 * force an record for the directory caps if we have a dentry lease.
4481 * this is racy (can't take i_ceph_lock and d_lock together), but it
4482 * doesn't have to be perfect; the mds will revoke anything we don't
4485 spin_lock(&dentry->d_lock);
4486 if (di->lease_session && di->lease_session->s_mds == mds)
4489 parent = dget(dentry->d_parent);
4490 dir = d_inode(parent);
4492 spin_unlock(&dentry->d_lock);
4494 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4497 spin_lock(&dentry->d_lock);
4498 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4499 dout("encode_dentry_release %p mds%d seq %d\n",
4500 dentry, mds, (int)di->lease_seq);
4501 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4502 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4503 *p += dentry->d_name.len;
4504 rel->dname_seq = cpu_to_le32(di->lease_seq);
4505 __ceph_mdsc_drop_dentry_lease(dentry);
4507 spin_unlock(&dentry->d_lock);