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;
494 ci->i_hold_caps_min = round_jiffies(jiffies +
495 opt->caps_wanted_delay_min * HZ);
496 ci->i_hold_caps_max = round_jiffies(jiffies +
497 opt->caps_wanted_delay_max * HZ);
498 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
499 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
503 * (Re)queue cap at the end of the delayed cap release list.
505 * If I_FLUSH is set, leave the inode at the front of the list.
507 * Caller holds i_ceph_lock
508 * -> we take mdsc->cap_delay_lock
510 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
511 struct ceph_inode_info *ci,
514 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
515 ci->i_ceph_flags, ci->i_hold_caps_max);
516 if (!mdsc->stopping) {
517 spin_lock(&mdsc->cap_delay_lock);
518 if (!list_empty(&ci->i_cap_delay_list)) {
519 if (ci->i_ceph_flags & CEPH_I_FLUSH)
521 list_del_init(&ci->i_cap_delay_list);
524 __cap_set_timeouts(mdsc, ci);
525 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
527 spin_unlock(&mdsc->cap_delay_lock);
532 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
533 * indicating we should send a cap message to flush dirty metadata
534 * asap, and move to the front of the delayed cap list.
536 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
537 struct ceph_inode_info *ci)
539 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
540 spin_lock(&mdsc->cap_delay_lock);
541 ci->i_ceph_flags |= CEPH_I_FLUSH;
542 if (!list_empty(&ci->i_cap_delay_list))
543 list_del_init(&ci->i_cap_delay_list);
544 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
545 spin_unlock(&mdsc->cap_delay_lock);
549 * Cancel delayed work on cap.
551 * Caller must hold i_ceph_lock.
553 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
554 struct ceph_inode_info *ci)
556 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
557 if (list_empty(&ci->i_cap_delay_list))
559 spin_lock(&mdsc->cap_delay_lock);
560 list_del_init(&ci->i_cap_delay_list);
561 spin_unlock(&mdsc->cap_delay_lock);
565 * Common issue checks for add_cap, handle_cap_grant.
567 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
570 unsigned had = __ceph_caps_issued(ci, NULL);
573 * Each time we receive FILE_CACHE anew, we increment
576 if (S_ISREG(ci->vfs_inode.i_mode) &&
577 (issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
578 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
583 * If FILE_SHARED is newly issued, mark dir not complete. We don't
584 * know what happened to this directory while we didn't have the cap.
585 * If FILE_SHARED is being revoked, also mark dir not complete. It
586 * stops on-going cached readdir.
588 if ((issued & CEPH_CAP_FILE_SHARED) != (had & CEPH_CAP_FILE_SHARED)) {
589 if (issued & CEPH_CAP_FILE_SHARED)
590 atomic_inc(&ci->i_shared_gen);
591 if (S_ISDIR(ci->vfs_inode.i_mode)) {
592 dout(" marking %p NOT complete\n", &ci->vfs_inode);
593 __ceph_dir_clear_complete(ci);
599 * Add a capability under the given MDS session.
601 * Caller should hold session snap_rwsem (read) and ci->i_ceph_lock
603 * @fmode is the open file mode, if we are opening a file, otherwise
604 * it is < 0. (This is so we can atomically add the cap and add an
605 * open file reference to it.)
607 void ceph_add_cap(struct inode *inode,
608 struct ceph_mds_session *session, u64 cap_id,
609 int fmode, unsigned issued, unsigned wanted,
610 unsigned seq, unsigned mseq, u64 realmino, int flags,
611 struct ceph_cap **new_cap)
613 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
614 struct ceph_inode_info *ci = ceph_inode(inode);
615 struct ceph_cap *cap;
616 int mds = session->s_mds;
620 lockdep_assert_held(&ci->i_ceph_lock);
622 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
623 session->s_mds, cap_id, ceph_cap_string(issued), seq);
626 * If we are opening the file, include file mode wanted bits
630 wanted |= ceph_caps_for_mode(fmode);
632 spin_lock(&session->s_gen_ttl_lock);
633 gen = session->s_cap_gen;
634 spin_unlock(&session->s_gen_ttl_lock);
636 cap = __get_cap_for_mds(ci, mds);
642 cap->implemented = 0;
648 __insert_cap_node(ci, cap);
650 /* add to session cap list */
651 cap->session = session;
652 spin_lock(&session->s_cap_lock);
653 list_add_tail(&cap->session_caps, &session->s_caps);
654 session->s_nr_caps++;
655 spin_unlock(&session->s_cap_lock);
657 spin_lock(&session->s_cap_lock);
658 list_move_tail(&cap->session_caps, &session->s_caps);
659 spin_unlock(&session->s_cap_lock);
661 if (cap->cap_gen < gen)
662 cap->issued = cap->implemented = CEPH_CAP_PIN;
665 * auth mds of the inode changed. we received the cap export
666 * message, but still haven't received the cap import message.
667 * handle_cap_export() updated the new auth MDS' cap.
669 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
670 * a message that was send before the cap import message. So
673 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
674 WARN_ON(cap != ci->i_auth_cap);
675 WARN_ON(cap->cap_id != cap_id);
678 issued |= cap->issued;
679 flags |= CEPH_CAP_FLAG_AUTH;
683 if (!ci->i_snap_realm ||
684 ((flags & CEPH_CAP_FLAG_AUTH) &&
685 realmino != (u64)-1 && ci->i_snap_realm->ino != realmino)) {
687 * add this inode to the appropriate snap realm
689 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
692 struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
694 spin_lock(&oldrealm->inodes_with_caps_lock);
695 list_del_init(&ci->i_snap_realm_item);
696 spin_unlock(&oldrealm->inodes_with_caps_lock);
699 spin_lock(&realm->inodes_with_caps_lock);
700 list_add(&ci->i_snap_realm_item,
701 &realm->inodes_with_caps);
702 ci->i_snap_realm = realm;
703 if (realm->ino == ci->i_vino.ino)
704 realm->inode = inode;
705 spin_unlock(&realm->inodes_with_caps_lock);
708 ceph_put_snap_realm(mdsc, oldrealm);
710 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
716 __check_cap_issue(ci, cap, issued);
719 * If we are issued caps we don't want, or the mds' wanted
720 * value appears to be off, queue a check so we'll release
721 * later and/or update the mds wanted value.
723 actual_wanted = __ceph_caps_wanted(ci);
724 if ((wanted & ~actual_wanted) ||
725 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
726 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
727 ceph_cap_string(issued), ceph_cap_string(wanted),
728 ceph_cap_string(actual_wanted));
729 __cap_delay_requeue(mdsc, ci, true);
732 if (flags & CEPH_CAP_FLAG_AUTH) {
733 if (!ci->i_auth_cap ||
734 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
735 ci->i_auth_cap = cap;
736 cap->mds_wanted = wanted;
739 WARN_ON(ci->i_auth_cap == cap);
742 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
743 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
744 ceph_cap_string(issued|cap->issued), seq, mds);
745 cap->cap_id = cap_id;
746 cap->issued = issued;
747 cap->implemented |= issued;
748 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
749 cap->mds_wanted = wanted;
751 cap->mds_wanted |= wanted;
753 cap->issue_seq = seq;
758 __ceph_get_fmode(ci, fmode);
762 * Return true if cap has not timed out and belongs to the current
763 * generation of the MDS session (i.e. has not gone 'stale' due to
764 * us losing touch with the mds).
766 static int __cap_is_valid(struct ceph_cap *cap)
771 spin_lock(&cap->session->s_gen_ttl_lock);
772 gen = cap->session->s_cap_gen;
773 ttl = cap->session->s_cap_ttl;
774 spin_unlock(&cap->session->s_gen_ttl_lock);
776 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
777 dout("__cap_is_valid %p cap %p issued %s "
778 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
779 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
787 * Return set of valid cap bits issued to us. Note that caps time
788 * out, and may be invalidated in bulk if the client session times out
789 * and session->s_cap_gen is bumped.
791 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
793 int have = ci->i_snap_caps;
794 struct ceph_cap *cap;
799 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
800 cap = rb_entry(p, struct ceph_cap, ci_node);
801 if (!__cap_is_valid(cap))
803 dout("__ceph_caps_issued %p cap %p issued %s\n",
804 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
807 *implemented |= cap->implemented;
810 * exclude caps issued by non-auth MDS, but are been revoking
811 * by the auth MDS. The non-auth MDS should be revoking/exporting
812 * these caps, but the message is delayed.
814 if (ci->i_auth_cap) {
815 cap = ci->i_auth_cap;
816 have &= ~cap->implemented | cap->issued;
822 * Get cap bits issued by caps other than @ocap
824 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
826 int have = ci->i_snap_caps;
827 struct ceph_cap *cap;
830 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
831 cap = rb_entry(p, struct ceph_cap, ci_node);
834 if (!__cap_is_valid(cap))
842 * Move a cap to the end of the LRU (oldest caps at list head, newest
845 static void __touch_cap(struct ceph_cap *cap)
847 struct ceph_mds_session *s = cap->session;
849 spin_lock(&s->s_cap_lock);
850 if (!s->s_cap_iterator) {
851 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
853 list_move_tail(&cap->session_caps, &s->s_caps);
855 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
856 &cap->ci->vfs_inode, cap, s->s_mds);
858 spin_unlock(&s->s_cap_lock);
862 * Check if we hold the given mask. If so, move the cap(s) to the
863 * front of their respective LRUs. (This is the preferred way for
864 * callers to check for caps they want.)
866 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
868 struct ceph_cap *cap;
870 int have = ci->i_snap_caps;
872 if ((have & mask) == mask) {
873 dout("__ceph_caps_issued_mask ino 0x%lx snap issued %s"
874 " (mask %s)\n", ci->vfs_inode.i_ino,
875 ceph_cap_string(have),
876 ceph_cap_string(mask));
880 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
881 cap = rb_entry(p, struct ceph_cap, ci_node);
882 if (!__cap_is_valid(cap))
884 if ((cap->issued & mask) == mask) {
885 dout("__ceph_caps_issued_mask ino 0x%lx cap %p issued %s"
886 " (mask %s)\n", ci->vfs_inode.i_ino, cap,
887 ceph_cap_string(cap->issued),
888 ceph_cap_string(mask));
894 /* does a combination of caps satisfy mask? */
896 if ((have & mask) == mask) {
897 dout("__ceph_caps_issued_mask ino 0x%lx combo issued %s"
898 " (mask %s)\n", ci->vfs_inode.i_ino,
899 ceph_cap_string(cap->issued),
900 ceph_cap_string(mask));
904 /* touch this + preceding caps */
906 for (q = rb_first(&ci->i_caps); q != p;
908 cap = rb_entry(q, struct ceph_cap,
910 if (!__cap_is_valid(cap))
912 if (cap->issued & mask)
924 * Return true if mask caps are currently being revoked by an MDS.
926 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
927 struct ceph_cap *ocap, int mask)
929 struct ceph_cap *cap;
932 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
933 cap = rb_entry(p, struct ceph_cap, ci_node);
935 (cap->implemented & ~cap->issued & mask))
941 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
943 struct inode *inode = &ci->vfs_inode;
946 spin_lock(&ci->i_ceph_lock);
947 ret = __ceph_caps_revoking_other(ci, NULL, mask);
948 spin_unlock(&ci->i_ceph_lock);
949 dout("ceph_caps_revoking %p %s = %d\n", inode,
950 ceph_cap_string(mask), ret);
954 int __ceph_caps_used(struct ceph_inode_info *ci)
958 used |= CEPH_CAP_PIN;
960 used |= CEPH_CAP_FILE_RD;
961 if (ci->i_rdcache_ref ||
962 (S_ISREG(ci->vfs_inode.i_mode) &&
963 ci->vfs_inode.i_data.nrpages))
964 used |= CEPH_CAP_FILE_CACHE;
966 used |= CEPH_CAP_FILE_WR;
967 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
968 used |= CEPH_CAP_FILE_BUFFER;
970 used |= CEPH_CAP_FILE_EXCL;
975 * wanted, by virtue of open file modes
977 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
980 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
981 if (ci->i_nr_by_mode[i])
986 return ceph_caps_for_mode(bits >> 1);
990 * wanted, by virtue of open file modes AND cap refs (buffered/cached data)
992 int __ceph_caps_wanted(struct ceph_inode_info *ci)
994 int w = __ceph_caps_file_wanted(ci) | __ceph_caps_used(ci);
995 if (!S_ISDIR(ci->vfs_inode.i_mode)) {
996 /* we want EXCL if dirty data */
997 if (w & CEPH_CAP_FILE_BUFFER)
998 w |= CEPH_CAP_FILE_EXCL;
1004 * Return caps we have registered with the MDS(s) as 'wanted'.
1006 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
1008 struct ceph_cap *cap;
1012 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1013 cap = rb_entry(p, struct ceph_cap, ci_node);
1014 if (check && !__cap_is_valid(cap))
1016 if (cap == ci->i_auth_cap)
1017 mds_wanted |= cap->mds_wanted;
1019 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
1025 * called under i_ceph_lock
1027 static int __ceph_is_single_caps(struct ceph_inode_info *ci)
1029 return rb_first(&ci->i_caps) == rb_last(&ci->i_caps);
1032 int ceph_is_any_caps(struct inode *inode)
1034 struct ceph_inode_info *ci = ceph_inode(inode);
1037 spin_lock(&ci->i_ceph_lock);
1038 ret = __ceph_is_any_real_caps(ci);
1039 spin_unlock(&ci->i_ceph_lock);
1044 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
1046 struct ceph_snap_realm *realm = ci->i_snap_realm;
1047 spin_lock(&realm->inodes_with_caps_lock);
1048 list_del_init(&ci->i_snap_realm_item);
1049 ci->i_snap_realm_counter++;
1050 ci->i_snap_realm = NULL;
1051 if (realm->ino == ci->i_vino.ino)
1052 realm->inode = NULL;
1053 spin_unlock(&realm->inodes_with_caps_lock);
1054 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
1059 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
1061 * caller should hold i_ceph_lock.
1062 * caller will not hold session s_mutex if called from destroy_inode.
1064 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1066 struct ceph_mds_session *session = cap->session;
1067 struct ceph_inode_info *ci = cap->ci;
1068 struct ceph_mds_client *mdsc =
1069 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1072 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1074 /* remove from inode's cap rbtree, and clear auth cap */
1075 rb_erase(&cap->ci_node, &ci->i_caps);
1076 if (ci->i_auth_cap == cap)
1077 ci->i_auth_cap = NULL;
1079 /* remove from session list */
1080 spin_lock(&session->s_cap_lock);
1081 if (session->s_cap_iterator == cap) {
1082 /* not yet, we are iterating over this very cap */
1083 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1086 list_del_init(&cap->session_caps);
1087 session->s_nr_caps--;
1088 cap->session = NULL;
1091 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1095 * s_cap_reconnect is protected by s_cap_lock. no one changes
1096 * s_cap_gen while session is in the reconnect state.
1098 if (queue_release &&
1099 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
1100 cap->queue_release = 1;
1102 __ceph_queue_cap_release(session, cap);
1106 cap->queue_release = 0;
1108 cap->cap_ino = ci->i_vino.ino;
1110 spin_unlock(&session->s_cap_lock);
1113 ceph_put_cap(mdsc, cap);
1115 if (!__ceph_is_any_real_caps(ci)) {
1116 /* when reconnect denied, we remove session caps forcibly,
1117 * i_wr_ref can be non-zero. If there are ongoing write,
1118 * keep i_snap_realm.
1120 if (ci->i_wr_ref == 0 && ci->i_snap_realm)
1121 drop_inode_snap_realm(ci);
1123 __cap_delay_cancel(mdsc, ci);
1127 struct cap_msg_args {
1128 struct ceph_mds_session *session;
1129 u64 ino, cid, follows;
1130 u64 flush_tid, oldest_flush_tid, size, max_size;
1133 struct ceph_buffer *xattr_buf;
1134 struct timespec64 atime, mtime, ctime, btime;
1135 int op, caps, wanted, dirty;
1136 u32 seq, issue_seq, mseq, time_warp_seq;
1145 * Build and send a cap message to the given MDS.
1147 * Caller should be holding s_mutex.
1149 static int send_cap_msg(struct cap_msg_args *arg)
1151 struct ceph_mds_caps *fc;
1152 struct ceph_msg *msg;
1155 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1157 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1158 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1159 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(arg->op),
1160 arg->cid, arg->ino, ceph_cap_string(arg->caps),
1161 ceph_cap_string(arg->wanted), ceph_cap_string(arg->dirty),
1162 arg->seq, arg->issue_seq, arg->flush_tid, arg->oldest_flush_tid,
1163 arg->mseq, arg->follows, arg->size, arg->max_size,
1165 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1167 /* flock buffer size + inline version + inline data size +
1168 * osd_epoch_barrier + oldest_flush_tid */
1169 extra_len = 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4;
1170 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1175 msg->hdr.version = cpu_to_le16(10);
1176 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1178 fc = msg->front.iov_base;
1179 memset(fc, 0, sizeof(*fc));
1181 fc->cap_id = cpu_to_le64(arg->cid);
1182 fc->op = cpu_to_le32(arg->op);
1183 fc->seq = cpu_to_le32(arg->seq);
1184 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1185 fc->migrate_seq = cpu_to_le32(arg->mseq);
1186 fc->caps = cpu_to_le32(arg->caps);
1187 fc->wanted = cpu_to_le32(arg->wanted);
1188 fc->dirty = cpu_to_le32(arg->dirty);
1189 fc->ino = cpu_to_le64(arg->ino);
1190 fc->snap_follows = cpu_to_le64(arg->follows);
1192 fc->size = cpu_to_le64(arg->size);
1193 fc->max_size = cpu_to_le64(arg->max_size);
1194 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1195 ceph_encode_timespec64(&fc->atime, &arg->atime);
1196 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1197 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1199 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1200 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1201 fc->mode = cpu_to_le32(arg->mode);
1203 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1204 if (arg->xattr_buf) {
1205 msg->middle = ceph_buffer_get(arg->xattr_buf);
1206 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1207 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1211 /* flock buffer size (version 2) */
1212 ceph_encode_32(&p, 0);
1213 /* inline version (version 4) */
1214 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1215 /* inline data size */
1216 ceph_encode_32(&p, 0);
1218 * osd_epoch_barrier (version 5)
1219 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1220 * case it was recently changed
1222 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1223 /* oldest_flush_tid (version 6) */
1224 ceph_encode_64(&p, arg->oldest_flush_tid);
1227 * caller_uid/caller_gid (version 7)
1229 * Currently, we don't properly track which caller dirtied the caps
1230 * last, and force a flush of them when there is a conflict. For now,
1231 * just set this to 0:0, to emulate how the MDS has worked up to now.
1233 ceph_encode_32(&p, 0);
1234 ceph_encode_32(&p, 0);
1236 /* pool namespace (version 8) (mds always ignores this) */
1237 ceph_encode_32(&p, 0);
1239 /* btime and change_attr (version 9) */
1240 ceph_encode_timespec64(p, &arg->btime);
1241 p += sizeof(struct ceph_timespec);
1242 ceph_encode_64(&p, arg->change_attr);
1244 /* Advisory flags (version 10) */
1245 ceph_encode_32(&p, arg->flags);
1247 ceph_con_send(&arg->session->s_con, msg);
1252 * Queue cap releases when an inode is dropped from our cache.
1254 void __ceph_remove_caps(struct ceph_inode_info *ci)
1258 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1259 * may call __ceph_caps_issued_mask() on a freeing inode. */
1260 spin_lock(&ci->i_ceph_lock);
1261 p = rb_first(&ci->i_caps);
1263 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1265 __ceph_remove_cap(cap, true);
1267 spin_unlock(&ci->i_ceph_lock);
1271 * Send a cap msg on the given inode. Update our caps state, then
1272 * drop i_ceph_lock and send the message.
1274 * Make note of max_size reported/requested from mds, revoked caps
1275 * that have now been implemented.
1277 * Return non-zero if delayed release, or we experienced an error
1278 * such that the caller should requeue + retry later.
1280 * called with i_ceph_lock, then drops it.
1281 * caller should hold snap_rwsem (read), s_mutex.
1283 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1284 int op, int flags, int used, int want, int retain,
1285 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1286 __releases(cap->ci->i_ceph_lock)
1288 struct ceph_inode_info *ci = cap->ci;
1289 struct inode *inode = &ci->vfs_inode;
1290 struct ceph_buffer *old_blob = NULL;
1291 struct cap_msg_args arg;
1297 held = cap->issued | cap->implemented;
1298 revoking = cap->implemented & ~cap->issued;
1299 retain &= ~revoking;
1301 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1302 inode, cap, cap->session,
1303 ceph_cap_string(held), ceph_cap_string(held & retain),
1304 ceph_cap_string(revoking));
1305 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1307 arg.session = cap->session;
1309 /* don't release wanted unless we've waited a bit. */
1310 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1311 time_before(jiffies, ci->i_hold_caps_min)) {
1312 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1313 ceph_cap_string(cap->issued),
1314 ceph_cap_string(cap->issued & retain),
1315 ceph_cap_string(cap->mds_wanted),
1316 ceph_cap_string(want));
1317 want |= cap->mds_wanted;
1318 retain |= cap->issued;
1321 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1322 if (want & ~cap->mds_wanted) {
1323 /* user space may open/close single file frequently.
1324 * This avoids droping mds_wanted immediately after
1325 * requesting new mds_wanted.
1327 __cap_set_timeouts(mdsc, ci);
1330 cap->issued &= retain; /* drop bits we don't want */
1331 if (cap->implemented & ~cap->issued) {
1333 * Wake up any waiters on wanted -> needed transition.
1334 * This is due to the weird transition from buffered
1335 * to sync IO... we need to flush dirty pages _before_
1336 * allowing sync writes to avoid reordering.
1340 cap->implemented &= cap->issued | used;
1341 cap->mds_wanted = want;
1343 arg.ino = ceph_vino(inode).ino;
1344 arg.cid = cap->cap_id;
1345 arg.follows = flushing ? ci->i_head_snapc->seq : 0;
1346 arg.flush_tid = flush_tid;
1347 arg.oldest_flush_tid = oldest_flush_tid;
1349 arg.size = inode->i_size;
1350 ci->i_reported_size = arg.size;
1351 arg.max_size = ci->i_wanted_max_size;
1352 ci->i_requested_max_size = arg.max_size;
1354 if (flushing & CEPH_CAP_XATTR_EXCL) {
1355 old_blob = __ceph_build_xattrs_blob(ci);
1356 arg.xattr_version = ci->i_xattrs.version;
1357 arg.xattr_buf = ci->i_xattrs.blob;
1359 arg.xattr_buf = NULL;
1362 arg.mtime = inode->i_mtime;
1363 arg.atime = inode->i_atime;
1364 arg.ctime = inode->i_ctime;
1365 arg.btime = ci->i_btime;
1366 arg.change_attr = inode_peek_iversion_raw(inode);
1369 arg.caps = cap->implemented;
1371 arg.dirty = flushing;
1374 arg.issue_seq = cap->issue_seq;
1375 arg.mseq = cap->mseq;
1376 arg.time_warp_seq = ci->i_time_warp_seq;
1378 arg.uid = inode->i_uid;
1379 arg.gid = inode->i_gid;
1380 arg.mode = inode->i_mode;
1382 arg.inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1383 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1384 !list_empty(&ci->i_cap_snaps)) {
1385 struct ceph_cap_snap *capsnap;
1386 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1387 if (capsnap->cap_flush.tid)
1389 if (capsnap->need_flush) {
1390 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1397 spin_unlock(&ci->i_ceph_lock);
1399 ceph_buffer_put(old_blob);
1401 ret = send_cap_msg(&arg);
1403 dout("error sending cap msg, must requeue %p\n", inode);
1408 wake_up_all(&ci->i_cap_wq);
1413 static inline int __send_flush_snap(struct inode *inode,
1414 struct ceph_mds_session *session,
1415 struct ceph_cap_snap *capsnap,
1416 u32 mseq, u64 oldest_flush_tid)
1418 struct cap_msg_args arg;
1420 arg.session = session;
1421 arg.ino = ceph_vino(inode).ino;
1423 arg.follows = capsnap->follows;
1424 arg.flush_tid = capsnap->cap_flush.tid;
1425 arg.oldest_flush_tid = oldest_flush_tid;
1427 arg.size = capsnap->size;
1429 arg.xattr_version = capsnap->xattr_version;
1430 arg.xattr_buf = capsnap->xattr_blob;
1432 arg.atime = capsnap->atime;
1433 arg.mtime = capsnap->mtime;
1434 arg.ctime = capsnap->ctime;
1435 arg.btime = capsnap->btime;
1436 arg.change_attr = capsnap->change_attr;
1438 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1439 arg.caps = capsnap->issued;
1441 arg.dirty = capsnap->dirty;
1446 arg.time_warp_seq = capsnap->time_warp_seq;
1448 arg.uid = capsnap->uid;
1449 arg.gid = capsnap->gid;
1450 arg.mode = capsnap->mode;
1452 arg.inline_data = capsnap->inline_data;
1455 return send_cap_msg(&arg);
1459 * When a snapshot is taken, clients accumulate dirty metadata on
1460 * inodes with capabilities in ceph_cap_snaps to describe the file
1461 * state at the time the snapshot was taken. This must be flushed
1462 * asynchronously back to the MDS once sync writes complete and dirty
1463 * data is written out.
1465 * Called under i_ceph_lock. Takes s_mutex as needed.
1467 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1468 struct ceph_mds_session *session)
1469 __releases(ci->i_ceph_lock)
1470 __acquires(ci->i_ceph_lock)
1472 struct inode *inode = &ci->vfs_inode;
1473 struct ceph_mds_client *mdsc = session->s_mdsc;
1474 struct ceph_cap_snap *capsnap;
1475 u64 oldest_flush_tid = 0;
1476 u64 first_tid = 1, last_tid = 0;
1478 dout("__flush_snaps %p session %p\n", inode, session);
1480 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1482 * we need to wait for sync writes to complete and for dirty
1483 * pages to be written out.
1485 if (capsnap->dirty_pages || capsnap->writing)
1488 /* should be removed by ceph_try_drop_cap_snap() */
1489 BUG_ON(!capsnap->need_flush);
1491 /* only flush each capsnap once */
1492 if (capsnap->cap_flush.tid > 0) {
1493 dout(" already flushed %p, skipping\n", capsnap);
1497 spin_lock(&mdsc->cap_dirty_lock);
1498 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1499 list_add_tail(&capsnap->cap_flush.g_list,
1500 &mdsc->cap_flush_list);
1501 if (oldest_flush_tid == 0)
1502 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1503 if (list_empty(&ci->i_flushing_item)) {
1504 list_add_tail(&ci->i_flushing_item,
1505 &session->s_cap_flushing);
1507 spin_unlock(&mdsc->cap_dirty_lock);
1509 list_add_tail(&capsnap->cap_flush.i_list,
1510 &ci->i_cap_flush_list);
1513 first_tid = capsnap->cap_flush.tid;
1514 last_tid = capsnap->cap_flush.tid;
1517 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1519 while (first_tid <= last_tid) {
1520 struct ceph_cap *cap = ci->i_auth_cap;
1521 struct ceph_cap_flush *cf;
1524 if (!(cap && cap->session == session)) {
1525 dout("__flush_snaps %p auth cap %p not mds%d, "
1526 "stop\n", inode, cap, session->s_mds);
1531 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1532 if (cf->tid >= first_tid) {
1540 first_tid = cf->tid + 1;
1542 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1543 refcount_inc(&capsnap->nref);
1544 spin_unlock(&ci->i_ceph_lock);
1546 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1547 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1549 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1552 pr_err("__flush_snaps: error sending cap flushsnap, "
1553 "ino (%llx.%llx) tid %llu follows %llu\n",
1554 ceph_vinop(inode), cf->tid, capsnap->follows);
1557 ceph_put_cap_snap(capsnap);
1558 spin_lock(&ci->i_ceph_lock);
1562 void ceph_flush_snaps(struct ceph_inode_info *ci,
1563 struct ceph_mds_session **psession)
1565 struct inode *inode = &ci->vfs_inode;
1566 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1567 struct ceph_mds_session *session = NULL;
1570 dout("ceph_flush_snaps %p\n", inode);
1572 session = *psession;
1574 spin_lock(&ci->i_ceph_lock);
1575 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1576 dout(" no capsnap needs flush, doing nothing\n");
1579 if (!ci->i_auth_cap) {
1580 dout(" no auth cap (migrating?), doing nothing\n");
1584 mds = ci->i_auth_cap->session->s_mds;
1585 if (session && session->s_mds != mds) {
1586 dout(" oops, wrong session %p mutex\n", session);
1587 mutex_unlock(&session->s_mutex);
1588 ceph_put_mds_session(session);
1592 spin_unlock(&ci->i_ceph_lock);
1593 mutex_lock(&mdsc->mutex);
1594 session = __ceph_lookup_mds_session(mdsc, mds);
1595 mutex_unlock(&mdsc->mutex);
1597 dout(" inverting session/ino locks on %p\n", session);
1598 mutex_lock(&session->s_mutex);
1603 // make sure flushsnap messages are sent in proper order.
1604 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1605 __kick_flushing_caps(mdsc, session, ci, 0);
1607 __ceph_flush_snaps(ci, session);
1609 spin_unlock(&ci->i_ceph_lock);
1612 *psession = session;
1613 } else if (session) {
1614 mutex_unlock(&session->s_mutex);
1615 ceph_put_mds_session(session);
1617 /* we flushed them all; remove this inode from the queue */
1618 spin_lock(&mdsc->snap_flush_lock);
1619 list_del_init(&ci->i_snap_flush_item);
1620 spin_unlock(&mdsc->snap_flush_lock);
1624 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1625 * Caller is then responsible for calling __mark_inode_dirty with the
1626 * returned flags value.
1628 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1629 struct ceph_cap_flush **pcf)
1631 struct ceph_mds_client *mdsc =
1632 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1633 struct inode *inode = &ci->vfs_inode;
1634 int was = ci->i_dirty_caps;
1637 if (!ci->i_auth_cap) {
1638 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1639 "but no auth cap (session was closed?)\n",
1640 inode, ceph_ino(inode), ceph_cap_string(mask));
1644 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1645 ceph_cap_string(mask), ceph_cap_string(was),
1646 ceph_cap_string(was | mask));
1647 ci->i_dirty_caps |= mask;
1649 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1650 swap(ci->i_prealloc_cap_flush, *pcf);
1652 if (!ci->i_head_snapc) {
1653 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1654 ci->i_head_snapc = ceph_get_snap_context(
1655 ci->i_snap_realm->cached_context);
1657 dout(" inode %p now dirty snapc %p auth cap %p\n",
1658 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1659 BUG_ON(!list_empty(&ci->i_dirty_item));
1660 spin_lock(&mdsc->cap_dirty_lock);
1661 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1662 spin_unlock(&mdsc->cap_dirty_lock);
1663 if (ci->i_flushing_caps == 0) {
1665 dirty |= I_DIRTY_SYNC;
1668 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1670 BUG_ON(list_empty(&ci->i_dirty_item));
1671 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1672 (mask & CEPH_CAP_FILE_BUFFER))
1673 dirty |= I_DIRTY_DATASYNC;
1674 __cap_delay_requeue(mdsc, ci, true);
1678 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1680 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1683 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1686 kmem_cache_free(ceph_cap_flush_cachep, cf);
1689 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1691 if (!list_empty(&mdsc->cap_flush_list)) {
1692 struct ceph_cap_flush *cf =
1693 list_first_entry(&mdsc->cap_flush_list,
1694 struct ceph_cap_flush, g_list);
1701 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1702 * Return true if caller needs to wake up flush waiters.
1704 static bool __finish_cap_flush(struct ceph_mds_client *mdsc,
1705 struct ceph_inode_info *ci,
1706 struct ceph_cap_flush *cf)
1708 struct ceph_cap_flush *prev;
1709 bool wake = cf->wake;
1711 /* are there older pending cap flushes? */
1712 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1713 prev = list_prev_entry(cf, g_list);
1717 list_del(&cf->g_list);
1719 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1720 prev = list_prev_entry(cf, i_list);
1724 list_del(&cf->i_list);
1732 * Add dirty inode to the flushing list. Assigned a seq number so we
1733 * can wait for caps to flush without starving.
1735 * Called under i_ceph_lock. Returns the flush tid.
1737 static u64 __mark_caps_flushing(struct inode *inode,
1738 struct ceph_mds_session *session, bool wake,
1739 u64 *oldest_flush_tid)
1741 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1742 struct ceph_inode_info *ci = ceph_inode(inode);
1743 struct ceph_cap_flush *cf = NULL;
1746 BUG_ON(ci->i_dirty_caps == 0);
1747 BUG_ON(list_empty(&ci->i_dirty_item));
1748 BUG_ON(!ci->i_prealloc_cap_flush);
1750 flushing = ci->i_dirty_caps;
1751 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1752 ceph_cap_string(flushing),
1753 ceph_cap_string(ci->i_flushing_caps),
1754 ceph_cap_string(ci->i_flushing_caps | flushing));
1755 ci->i_flushing_caps |= flushing;
1756 ci->i_dirty_caps = 0;
1757 dout(" inode %p now !dirty\n", inode);
1759 swap(cf, ci->i_prealloc_cap_flush);
1760 cf->caps = flushing;
1763 spin_lock(&mdsc->cap_dirty_lock);
1764 list_del_init(&ci->i_dirty_item);
1766 cf->tid = ++mdsc->last_cap_flush_tid;
1767 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1768 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1770 if (list_empty(&ci->i_flushing_item)) {
1771 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1772 mdsc->num_cap_flushing++;
1774 spin_unlock(&mdsc->cap_dirty_lock);
1776 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1782 * try to invalidate mapping pages without blocking.
1784 static int try_nonblocking_invalidate(struct inode *inode)
1786 struct ceph_inode_info *ci = ceph_inode(inode);
1787 u32 invalidating_gen = ci->i_rdcache_gen;
1789 spin_unlock(&ci->i_ceph_lock);
1790 invalidate_mapping_pages(&inode->i_data, 0, -1);
1791 spin_lock(&ci->i_ceph_lock);
1793 if (inode->i_data.nrpages == 0 &&
1794 invalidating_gen == ci->i_rdcache_gen) {
1796 dout("try_nonblocking_invalidate %p success\n", inode);
1797 /* save any racing async invalidate some trouble */
1798 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1801 dout("try_nonblocking_invalidate %p failed\n", inode);
1805 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1807 loff_t size = ci->vfs_inode.i_size;
1808 /* mds will adjust max size according to the reported size */
1809 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1811 if (size >= ci->i_max_size)
1813 /* half of previous max_size increment has been used */
1814 if (ci->i_max_size > ci->i_reported_size &&
1815 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1821 * Swiss army knife function to examine currently used and wanted
1822 * versus held caps. Release, flush, ack revoked caps to mds as
1825 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1826 * cap release further.
1827 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1828 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1831 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1832 struct ceph_mds_session *session)
1834 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1835 struct ceph_mds_client *mdsc = fsc->mdsc;
1836 struct inode *inode = &ci->vfs_inode;
1837 struct ceph_cap *cap;
1838 u64 flush_tid, oldest_flush_tid;
1839 int file_wanted, used, cap_used;
1840 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1841 int issued, implemented, want, retain, revoking, flushing = 0;
1842 int mds = -1; /* keep track of how far we've gone through i_caps list
1843 to avoid an infinite loop on retry */
1845 int delayed = 0, sent = 0;
1846 bool no_delay = flags & CHECK_CAPS_NODELAY;
1847 bool queue_invalidate = false;
1848 bool tried_invalidate = false;
1850 /* if we are unmounting, flush any unused caps immediately. */
1854 spin_lock(&ci->i_ceph_lock);
1856 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1857 flags |= CHECK_CAPS_FLUSH;
1859 if (!(flags & CHECK_CAPS_AUTHONLY) ||
1860 (ci->i_auth_cap && __ceph_is_single_caps(ci)))
1861 __cap_delay_cancel(mdsc, ci);
1865 spin_lock(&ci->i_ceph_lock);
1867 file_wanted = __ceph_caps_file_wanted(ci);
1868 used = __ceph_caps_used(ci);
1869 issued = __ceph_caps_issued(ci, &implemented);
1870 revoking = implemented & ~issued;
1873 retain = file_wanted | used | CEPH_CAP_PIN;
1874 if (!mdsc->stopping && inode->i_nlink > 0) {
1876 retain |= CEPH_CAP_ANY; /* be greedy */
1877 } else if (S_ISDIR(inode->i_mode) &&
1878 (issued & CEPH_CAP_FILE_SHARED) &&
1879 __ceph_dir_is_complete(ci)) {
1881 * If a directory is complete, we want to keep
1882 * the exclusive cap. So that MDS does not end up
1883 * revoking the shared cap on every create/unlink
1886 if (IS_RDONLY(inode))
1887 want = CEPH_CAP_ANY_SHARED;
1889 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1893 retain |= CEPH_CAP_ANY_SHARED;
1895 * keep RD only if we didn't have the file open RW,
1896 * because then the mds would revoke it anyway to
1897 * journal max_size=0.
1899 if (ci->i_max_size == 0)
1900 retain |= CEPH_CAP_ANY_RD;
1904 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1905 " issued %s revoking %s retain %s %s%s%s\n", inode,
1906 ceph_cap_string(file_wanted),
1907 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1908 ceph_cap_string(ci->i_flushing_caps),
1909 ceph_cap_string(issued), ceph_cap_string(revoking),
1910 ceph_cap_string(retain),
1911 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1912 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1913 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1916 * If we no longer need to hold onto old our caps, and we may
1917 * have cached pages, but don't want them, then try to invalidate.
1918 * If we fail, it's because pages are locked.... try again later.
1920 if ((!no_delay || mdsc->stopping) &&
1921 S_ISREG(inode->i_mode) &&
1922 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1923 inode->i_data.nrpages && /* have cached pages */
1924 (revoking & (CEPH_CAP_FILE_CACHE|
1925 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1926 !tried_invalidate) {
1927 dout("check_caps trying to invalidate on %p\n", inode);
1928 if (try_nonblocking_invalidate(inode) < 0) {
1929 dout("check_caps queuing invalidate\n");
1930 queue_invalidate = true;
1931 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1933 tried_invalidate = true;
1937 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1938 cap = rb_entry(p, struct ceph_cap, ci_node);
1940 /* avoid looping forever */
1941 if (mds >= cap->mds ||
1942 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1945 /* NOTE: no side-effects allowed, until we take s_mutex */
1948 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1949 cap_used &= ~ci->i_auth_cap->issued;
1951 revoking = cap->implemented & ~cap->issued;
1952 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1953 cap->mds, cap, ceph_cap_string(cap_used),
1954 ceph_cap_string(cap->issued),
1955 ceph_cap_string(cap->implemented),
1956 ceph_cap_string(revoking));
1958 if (cap == ci->i_auth_cap &&
1959 (cap->issued & CEPH_CAP_FILE_WR)) {
1960 /* request larger max_size from MDS? */
1961 if (ci->i_wanted_max_size > ci->i_max_size &&
1962 ci->i_wanted_max_size > ci->i_requested_max_size) {
1963 dout("requesting new max_size\n");
1967 /* approaching file_max? */
1968 if (__ceph_should_report_size(ci)) {
1969 dout("i_size approaching max_size\n");
1973 /* flush anything dirty? */
1974 if (cap == ci->i_auth_cap) {
1975 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
1976 dout("flushing dirty caps\n");
1979 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
1980 dout("flushing snap caps\n");
1985 /* completed revocation? going down and there are no caps? */
1986 if (revoking && (revoking & cap_used) == 0) {
1987 dout("completed revocation of %s\n",
1988 ceph_cap_string(cap->implemented & ~cap->issued));
1992 /* want more caps from mds? */
1993 if (want & ~(cap->mds_wanted | cap->issued))
1996 /* things we might delay */
1997 if ((cap->issued & ~retain) == 0)
1998 continue; /* nope, all good */
2004 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
2005 time_before(jiffies, ci->i_hold_caps_max)) {
2006 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
2007 ceph_cap_string(cap->issued),
2008 ceph_cap_string(cap->issued & retain),
2009 ceph_cap_string(cap->mds_wanted),
2010 ceph_cap_string(want));
2016 if (session && session != cap->session) {
2017 dout("oops, wrong session %p mutex\n", session);
2018 mutex_unlock(&session->s_mutex);
2022 session = cap->session;
2023 if (mutex_trylock(&session->s_mutex) == 0) {
2024 dout("inverting session/ino locks on %p\n",
2026 spin_unlock(&ci->i_ceph_lock);
2027 if (took_snap_rwsem) {
2028 up_read(&mdsc->snap_rwsem);
2029 took_snap_rwsem = 0;
2031 mutex_lock(&session->s_mutex);
2036 /* kick flushing and flush snaps before sending normal
2038 if (cap == ci->i_auth_cap &&
2040 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2041 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2042 __kick_flushing_caps(mdsc, session, ci, 0);
2043 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2044 __ceph_flush_snaps(ci, session);
2049 /* take snap_rwsem after session mutex */
2050 if (!took_snap_rwsem) {
2051 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
2052 dout("inverting snap/in locks on %p\n",
2054 spin_unlock(&ci->i_ceph_lock);
2055 down_read(&mdsc->snap_rwsem);
2056 took_snap_rwsem = 1;
2059 took_snap_rwsem = 1;
2062 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2063 flushing = ci->i_dirty_caps;
2064 flush_tid = __mark_caps_flushing(inode, session, false,
2069 spin_lock(&mdsc->cap_dirty_lock);
2070 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2071 spin_unlock(&mdsc->cap_dirty_lock);
2074 mds = cap->mds; /* remember mds, so we don't repeat */
2077 /* __send_cap drops i_ceph_lock */
2078 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, 0,
2079 cap_used, want, retain, flushing,
2080 flush_tid, oldest_flush_tid);
2081 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2084 /* Reschedule delayed caps release if we delayed anything */
2086 __cap_delay_requeue(mdsc, ci, false);
2088 spin_unlock(&ci->i_ceph_lock);
2090 if (queue_invalidate)
2091 ceph_queue_invalidate(inode);
2094 mutex_unlock(&session->s_mutex);
2095 if (took_snap_rwsem)
2096 up_read(&mdsc->snap_rwsem);
2100 * Try to flush dirty caps back to the auth mds.
2102 static int try_flush_caps(struct inode *inode, u64 *ptid)
2104 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2105 struct ceph_inode_info *ci = ceph_inode(inode);
2106 struct ceph_mds_session *session = NULL;
2108 u64 flush_tid = 0, oldest_flush_tid = 0;
2111 spin_lock(&ci->i_ceph_lock);
2113 if (ci->i_dirty_caps && ci->i_auth_cap) {
2114 struct ceph_cap *cap = ci->i_auth_cap;
2117 if (session != cap->session) {
2118 spin_unlock(&ci->i_ceph_lock);
2120 mutex_unlock(&session->s_mutex);
2121 session = cap->session;
2122 mutex_lock(&session->s_mutex);
2125 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) {
2126 spin_unlock(&ci->i_ceph_lock);
2130 if (ci->i_ceph_flags &
2131 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2132 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2133 __kick_flushing_caps(mdsc, session, ci, 0);
2134 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2135 __ceph_flush_snaps(ci, session);
2139 flushing = ci->i_dirty_caps;
2140 flush_tid = __mark_caps_flushing(inode, session, true,
2143 /* __send_cap drops i_ceph_lock */
2144 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2145 CEPH_CLIENT_CAPS_SYNC,
2146 __ceph_caps_used(ci),
2147 __ceph_caps_wanted(ci),
2148 (cap->issued | cap->implemented),
2149 flushing, flush_tid, oldest_flush_tid);
2152 spin_lock(&ci->i_ceph_lock);
2153 __cap_delay_requeue(mdsc, ci, true);
2154 spin_unlock(&ci->i_ceph_lock);
2157 if (!list_empty(&ci->i_cap_flush_list)) {
2158 struct ceph_cap_flush *cf =
2159 list_last_entry(&ci->i_cap_flush_list,
2160 struct ceph_cap_flush, i_list);
2162 flush_tid = cf->tid;
2164 flushing = ci->i_flushing_caps;
2165 spin_unlock(&ci->i_ceph_lock);
2169 mutex_unlock(&session->s_mutex);
2176 * Return true if we've flushed caps through the given flush_tid.
2178 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2180 struct ceph_inode_info *ci = ceph_inode(inode);
2183 spin_lock(&ci->i_ceph_lock);
2184 if (!list_empty(&ci->i_cap_flush_list)) {
2185 struct ceph_cap_flush * cf =
2186 list_first_entry(&ci->i_cap_flush_list,
2187 struct ceph_cap_flush, i_list);
2188 if (cf->tid <= flush_tid)
2191 spin_unlock(&ci->i_ceph_lock);
2196 * wait for any unsafe requests to complete.
2198 static int unsafe_request_wait(struct inode *inode)
2200 struct ceph_inode_info *ci = ceph_inode(inode);
2201 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2204 spin_lock(&ci->i_unsafe_lock);
2205 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2206 req1 = list_last_entry(&ci->i_unsafe_dirops,
2207 struct ceph_mds_request,
2209 ceph_mdsc_get_request(req1);
2211 if (!list_empty(&ci->i_unsafe_iops)) {
2212 req2 = list_last_entry(&ci->i_unsafe_iops,
2213 struct ceph_mds_request,
2214 r_unsafe_target_item);
2215 ceph_mdsc_get_request(req2);
2217 spin_unlock(&ci->i_unsafe_lock);
2219 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2220 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2222 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2223 ceph_timeout_jiffies(req1->r_timeout));
2226 ceph_mdsc_put_request(req1);
2229 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2230 ceph_timeout_jiffies(req2->r_timeout));
2233 ceph_mdsc_put_request(req2);
2238 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2240 struct ceph_file_info *fi = file->private_data;
2241 struct inode *inode = file->f_mapping->host;
2242 struct ceph_inode_info *ci = ceph_inode(inode);
2247 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2249 ret = file_write_and_wait_range(file, start, end);
2253 dirty = try_flush_caps(inode, &flush_tid);
2254 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2256 err = unsafe_request_wait(inode);
2259 * only wait on non-file metadata writeback (the mds
2260 * can recover size and mtime, so we don't need to
2263 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2264 err = wait_event_interruptible(ci->i_cap_wq,
2265 caps_are_flushed(inode, flush_tid));
2271 if (errseq_check(&ci->i_meta_err, READ_ONCE(fi->meta_err))) {
2272 spin_lock(&file->f_lock);
2273 err = errseq_check_and_advance(&ci->i_meta_err,
2275 spin_unlock(&file->f_lock);
2280 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2285 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2286 * queue inode for flush but don't do so immediately, because we can
2287 * get by with fewer MDS messages if we wait for data writeback to
2290 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2292 struct ceph_inode_info *ci = ceph_inode(inode);
2296 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2298 dout("write_inode %p wait=%d\n", inode, wait);
2300 dirty = try_flush_caps(inode, &flush_tid);
2302 err = wait_event_interruptible(ci->i_cap_wq,
2303 caps_are_flushed(inode, flush_tid));
2305 struct ceph_mds_client *mdsc =
2306 ceph_sb_to_client(inode->i_sb)->mdsc;
2308 spin_lock(&ci->i_ceph_lock);
2309 if (__ceph_caps_dirty(ci))
2310 __cap_delay_requeue_front(mdsc, ci);
2311 spin_unlock(&ci->i_ceph_lock);
2316 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2317 struct ceph_mds_session *session,
2318 struct ceph_inode_info *ci,
2319 u64 oldest_flush_tid)
2320 __releases(ci->i_ceph_lock)
2321 __acquires(ci->i_ceph_lock)
2323 struct inode *inode = &ci->vfs_inode;
2324 struct ceph_cap *cap;
2325 struct ceph_cap_flush *cf;
2328 u64 last_snap_flush = 0;
2330 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2332 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2334 last_snap_flush = cf->tid;
2339 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2340 if (cf->tid < first_tid)
2343 cap = ci->i_auth_cap;
2344 if (!(cap && cap->session == session)) {
2345 pr_err("%p auth cap %p not mds%d ???\n",
2346 inode, cap, session->s_mds);
2350 first_tid = cf->tid + 1;
2353 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2354 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2355 ci->i_ceph_flags |= CEPH_I_NODELAY;
2357 ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2358 (cf->tid < last_snap_flush ?
2359 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2360 __ceph_caps_used(ci),
2361 __ceph_caps_wanted(ci),
2362 (cap->issued | cap->implemented),
2363 cf->caps, cf->tid, oldest_flush_tid);
2365 pr_err("kick_flushing_caps: error sending "
2366 "cap flush, ino (%llx.%llx) "
2367 "tid %llu flushing %s\n",
2368 ceph_vinop(inode), cf->tid,
2369 ceph_cap_string(cf->caps));
2372 struct ceph_cap_snap *capsnap =
2373 container_of(cf, struct ceph_cap_snap,
2375 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2376 inode, capsnap, cf->tid,
2377 ceph_cap_string(capsnap->dirty));
2379 refcount_inc(&capsnap->nref);
2380 spin_unlock(&ci->i_ceph_lock);
2382 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2385 pr_err("kick_flushing_caps: error sending "
2386 "cap flushsnap, ino (%llx.%llx) "
2387 "tid %llu follows %llu\n",
2388 ceph_vinop(inode), cf->tid,
2392 ceph_put_cap_snap(capsnap);
2395 spin_lock(&ci->i_ceph_lock);
2399 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2400 struct ceph_mds_session *session)
2402 struct ceph_inode_info *ci;
2403 struct ceph_cap *cap;
2404 u64 oldest_flush_tid;
2406 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2408 spin_lock(&mdsc->cap_dirty_lock);
2409 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2410 spin_unlock(&mdsc->cap_dirty_lock);
2412 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2413 spin_lock(&ci->i_ceph_lock);
2414 cap = ci->i_auth_cap;
2415 if (!(cap && cap->session == session)) {
2416 pr_err("%p auth cap %p not mds%d ???\n",
2417 &ci->vfs_inode, cap, session->s_mds);
2418 spin_unlock(&ci->i_ceph_lock);
2424 * if flushing caps were revoked, we re-send the cap flush
2425 * in client reconnect stage. This guarantees MDS * processes
2426 * the cap flush message before issuing the flushing caps to
2429 if ((cap->issued & ci->i_flushing_caps) !=
2430 ci->i_flushing_caps) {
2431 /* encode_caps_cb() also will reset these sequence
2432 * numbers. make sure sequence numbers in cap flush
2433 * message match later reconnect message */
2437 __kick_flushing_caps(mdsc, session, ci,
2440 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2443 spin_unlock(&ci->i_ceph_lock);
2447 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2448 struct ceph_mds_session *session)
2450 struct ceph_inode_info *ci;
2451 struct ceph_cap *cap;
2452 u64 oldest_flush_tid;
2454 dout("kick_flushing_caps mds%d\n", session->s_mds);
2456 spin_lock(&mdsc->cap_dirty_lock);
2457 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2458 spin_unlock(&mdsc->cap_dirty_lock);
2460 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2461 spin_lock(&ci->i_ceph_lock);
2462 cap = ci->i_auth_cap;
2463 if (!(cap && cap->session == session)) {
2464 pr_err("%p auth cap %p not mds%d ???\n",
2465 &ci->vfs_inode, cap, session->s_mds);
2466 spin_unlock(&ci->i_ceph_lock);
2469 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2470 __kick_flushing_caps(mdsc, session, ci,
2473 spin_unlock(&ci->i_ceph_lock);
2477 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2478 struct ceph_mds_session *session,
2479 struct inode *inode)
2480 __releases(ci->i_ceph_lock)
2482 struct ceph_inode_info *ci = ceph_inode(inode);
2483 struct ceph_cap *cap;
2485 cap = ci->i_auth_cap;
2486 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2487 ceph_cap_string(ci->i_flushing_caps));
2489 if (!list_empty(&ci->i_cap_flush_list)) {
2490 u64 oldest_flush_tid;
2491 spin_lock(&mdsc->cap_dirty_lock);
2492 list_move_tail(&ci->i_flushing_item,
2493 &cap->session->s_cap_flushing);
2494 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2495 spin_unlock(&mdsc->cap_dirty_lock);
2497 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2498 spin_unlock(&ci->i_ceph_lock);
2500 spin_unlock(&ci->i_ceph_lock);
2506 * Take references to capabilities we hold, so that we don't release
2507 * them to the MDS prematurely.
2509 * Protected by i_ceph_lock.
2511 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2512 bool snap_rwsem_locked)
2514 if (got & CEPH_CAP_PIN)
2516 if (got & CEPH_CAP_FILE_RD)
2518 if (got & CEPH_CAP_FILE_CACHE)
2519 ci->i_rdcache_ref++;
2520 if (got & CEPH_CAP_FILE_EXCL)
2522 if (got & CEPH_CAP_FILE_WR) {
2523 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2524 BUG_ON(!snap_rwsem_locked);
2525 ci->i_head_snapc = ceph_get_snap_context(
2526 ci->i_snap_realm->cached_context);
2530 if (got & CEPH_CAP_FILE_BUFFER) {
2531 if (ci->i_wb_ref == 0)
2532 ihold(&ci->vfs_inode);
2534 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2535 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2540 * Try to grab cap references. Specify those refs we @want, and the
2541 * minimal set we @need. Also include the larger offset we are writing
2542 * to (when applicable), and check against max_size here as well.
2543 * Note that caller is responsible for ensuring max_size increases are
2544 * requested from the MDS.
2546 * Returns 0 if caps were not able to be acquired (yet), a 1 if they were,
2547 * or a negative error code.
2549 * FIXME: how does a 0 return differ from -EAGAIN?
2556 static int try_get_cap_refs(struct inode *inode, int need, int want,
2557 loff_t endoff, int flags, int *got)
2559 struct ceph_inode_info *ci = ceph_inode(inode);
2560 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2562 int have, implemented;
2564 bool snap_rwsem_locked = false;
2566 dout("get_cap_refs %p need %s want %s\n", inode,
2567 ceph_cap_string(need), ceph_cap_string(want));
2570 spin_lock(&ci->i_ceph_lock);
2572 if ((flags & CHECK_FILELOCK) &&
2573 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2574 dout("try_get_cap_refs %p error filelock\n", inode);
2579 /* make sure file is actually open */
2580 file_wanted = __ceph_caps_file_wanted(ci);
2581 if ((file_wanted & need) != need) {
2582 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2583 ceph_cap_string(need), ceph_cap_string(file_wanted));
2588 /* finish pending truncate */
2589 while (ci->i_truncate_pending) {
2590 spin_unlock(&ci->i_ceph_lock);
2591 if (snap_rwsem_locked) {
2592 up_read(&mdsc->snap_rwsem);
2593 snap_rwsem_locked = false;
2595 __ceph_do_pending_vmtruncate(inode);
2596 spin_lock(&ci->i_ceph_lock);
2599 have = __ceph_caps_issued(ci, &implemented);
2601 if (have & need & CEPH_CAP_FILE_WR) {
2602 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2603 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2604 inode, endoff, ci->i_max_size);
2605 if (endoff > ci->i_requested_max_size)
2610 * If a sync write is in progress, we must wait, so that we
2611 * can get a final snapshot value for size+mtime.
2613 if (__ceph_have_pending_cap_snap(ci)) {
2614 dout("get_cap_refs %p cap_snap_pending\n", inode);
2619 if ((have & need) == need) {
2621 * Look at (implemented & ~have & not) so that we keep waiting
2622 * on transition from wanted -> needed caps. This is needed
2623 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2624 * going before a prior buffered writeback happens.
2626 int not = want & ~(have & need);
2627 int revoking = implemented & ~have;
2628 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2629 inode, ceph_cap_string(have), ceph_cap_string(not),
2630 ceph_cap_string(revoking));
2631 if ((revoking & not) == 0) {
2632 if (!snap_rwsem_locked &&
2633 !ci->i_head_snapc &&
2634 (need & CEPH_CAP_FILE_WR)) {
2635 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2637 * we can not call down_read() when
2638 * task isn't in TASK_RUNNING state
2640 if (flags & NON_BLOCKING) {
2645 spin_unlock(&ci->i_ceph_lock);
2646 down_read(&mdsc->snap_rwsem);
2647 snap_rwsem_locked = true;
2650 snap_rwsem_locked = true;
2652 *got = need | (have & want);
2653 if (S_ISREG(inode->i_mode) &&
2654 (need & CEPH_CAP_FILE_RD) &&
2655 !(*got & CEPH_CAP_FILE_CACHE))
2656 ceph_disable_fscache_readpage(ci);
2657 __take_cap_refs(ci, *got, true);
2661 int session_readonly = false;
2662 if (ci->i_auth_cap &&
2663 (need & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_EXCL))) {
2664 struct ceph_mds_session *s = ci->i_auth_cap->session;
2665 spin_lock(&s->s_cap_lock);
2666 session_readonly = s->s_readonly;
2667 spin_unlock(&s->s_cap_lock);
2669 if (session_readonly) {
2670 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2671 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2676 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) {
2678 if (READ_ONCE(mdsc->fsc->mount_state) ==
2679 CEPH_MOUNT_SHUTDOWN) {
2680 dout("get_cap_refs %p forced umount\n", inode);
2684 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2685 if (need & ~(mds_wanted & need)) {
2686 dout("get_cap_refs %p caps were dropped"
2687 " (session killed?)\n", inode);
2691 if (!(file_wanted & ~mds_wanted))
2692 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED;
2695 dout("get_cap_refs %p have %s needed %s\n", inode,
2696 ceph_cap_string(have), ceph_cap_string(need));
2699 spin_unlock(&ci->i_ceph_lock);
2700 if (snap_rwsem_locked)
2701 up_read(&mdsc->snap_rwsem);
2703 dout("get_cap_refs %p ret %d got %s\n", inode,
2704 ret, ceph_cap_string(*got));
2709 * Check the offset we are writing up to against our current
2710 * max_size. If necessary, tell the MDS we want to write to
2713 static void check_max_size(struct inode *inode, loff_t endoff)
2715 struct ceph_inode_info *ci = ceph_inode(inode);
2718 /* do we need to explicitly request a larger max_size? */
2719 spin_lock(&ci->i_ceph_lock);
2720 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2721 dout("write %p at large endoff %llu, req max_size\n",
2723 ci->i_wanted_max_size = endoff;
2725 /* duplicate ceph_check_caps()'s logic */
2726 if (ci->i_auth_cap &&
2727 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2728 ci->i_wanted_max_size > ci->i_max_size &&
2729 ci->i_wanted_max_size > ci->i_requested_max_size)
2731 spin_unlock(&ci->i_ceph_lock);
2733 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2736 int ceph_try_get_caps(struct inode *inode, int need, int want,
2737 bool nonblock, int *got)
2741 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2742 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO|CEPH_CAP_FILE_SHARED));
2743 ret = ceph_pool_perm_check(inode, need);
2747 ret = try_get_cap_refs(inode, need, want, 0,
2748 (nonblock ? NON_BLOCKING : 0), got);
2749 return ret == -EAGAIN ? 0 : ret;
2753 * Wait for caps, and take cap references. If we can't get a WR cap
2754 * due to a small max_size, make sure we check_max_size (and possibly
2755 * ask the mds) so we don't get hung up indefinitely.
2757 int ceph_get_caps(struct file *filp, int need, int want,
2758 loff_t endoff, int *got, struct page **pinned_page)
2760 struct ceph_file_info *fi = filp->private_data;
2761 struct inode *inode = file_inode(filp);
2762 struct ceph_inode_info *ci = ceph_inode(inode);
2763 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2764 int ret, _got, flags;
2766 ret = ceph_pool_perm_check(inode, need);
2770 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2771 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2776 check_max_size(inode, endoff);
2778 flags = atomic_read(&fi->num_locks) ? CHECK_FILELOCK : 0;
2780 ret = try_get_cap_refs(inode, need, want, endoff,
2785 struct ceph_mds_client *mdsc = fsc->mdsc;
2787 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2789 cw.ino = inode->i_ino;
2790 cw.tgid = current->tgid;
2794 spin_lock(&mdsc->caps_list_lock);
2795 list_add(&cw.list, &mdsc->cap_wait_list);
2796 spin_unlock(&mdsc->caps_list_lock);
2798 add_wait_queue(&ci->i_cap_wq, &wait);
2800 flags |= NON_BLOCKING;
2801 while (!(ret = try_get_cap_refs(inode, need, want,
2802 endoff, flags, &_got))) {
2803 if (signal_pending(current)) {
2807 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2810 remove_wait_queue(&ci->i_cap_wq, &wait);
2812 spin_lock(&mdsc->caps_list_lock);
2814 spin_unlock(&mdsc->caps_list_lock);
2820 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2821 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2822 if (ret >= 0 && _got)
2823 ceph_put_cap_refs(ci, _got);
2828 if (ret == -ESTALE) {
2829 /* session was killed, try renew caps */
2830 ret = ceph_renew_caps(inode);
2837 if (S_ISREG(ci->vfs_inode.i_mode) &&
2838 ci->i_inline_version != CEPH_INLINE_NONE &&
2839 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2840 i_size_read(inode) > 0) {
2842 find_get_page(inode->i_mapping, 0);
2844 if (PageUptodate(page)) {
2845 *pinned_page = page;
2851 * drop cap refs first because getattr while
2852 * holding * caps refs can cause deadlock.
2854 ceph_put_cap_refs(ci, _got);
2858 * getattr request will bring inline data into
2861 ret = __ceph_do_getattr(inode, NULL,
2862 CEPH_STAT_CAP_INLINE_DATA,
2871 if (S_ISREG(ci->vfs_inode.i_mode) &&
2872 (_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
2873 ceph_fscache_revalidate_cookie(ci);
2880 * Take cap refs. Caller must already know we hold at least one ref
2881 * on the caps in question or we don't know this is safe.
2883 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2885 spin_lock(&ci->i_ceph_lock);
2886 __take_cap_refs(ci, caps, false);
2887 spin_unlock(&ci->i_ceph_lock);
2892 * drop cap_snap that is not associated with any snapshot.
2893 * we don't need to send FLUSHSNAP message for it.
2895 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2896 struct ceph_cap_snap *capsnap)
2898 if (!capsnap->need_flush &&
2899 !capsnap->writing && !capsnap->dirty_pages) {
2900 dout("dropping cap_snap %p follows %llu\n",
2901 capsnap, capsnap->follows);
2902 BUG_ON(capsnap->cap_flush.tid > 0);
2903 ceph_put_snap_context(capsnap->context);
2904 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2905 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2907 list_del(&capsnap->ci_item);
2908 ceph_put_cap_snap(capsnap);
2917 * If we released the last ref on any given cap, call ceph_check_caps
2918 * to release (or schedule a release).
2920 * If we are releasing a WR cap (from a sync write), finalize any affected
2921 * cap_snap, and wake up any waiters.
2923 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2925 struct inode *inode = &ci->vfs_inode;
2926 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2928 spin_lock(&ci->i_ceph_lock);
2929 if (had & CEPH_CAP_PIN)
2931 if (had & CEPH_CAP_FILE_RD)
2932 if (--ci->i_rd_ref == 0)
2934 if (had & CEPH_CAP_FILE_CACHE)
2935 if (--ci->i_rdcache_ref == 0)
2937 if (had & CEPH_CAP_FILE_EXCL)
2938 if (--ci->i_fx_ref == 0)
2940 if (had & CEPH_CAP_FILE_BUFFER) {
2941 if (--ci->i_wb_ref == 0) {
2945 dout("put_cap_refs %p wb %d -> %d (?)\n",
2946 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2948 if (had & CEPH_CAP_FILE_WR)
2949 if (--ci->i_wr_ref == 0) {
2951 if (__ceph_have_pending_cap_snap(ci)) {
2952 struct ceph_cap_snap *capsnap =
2953 list_last_entry(&ci->i_cap_snaps,
2954 struct ceph_cap_snap,
2956 capsnap->writing = 0;
2957 if (ceph_try_drop_cap_snap(ci, capsnap))
2959 else if (__ceph_finish_cap_snap(ci, capsnap))
2963 if (ci->i_wrbuffer_ref_head == 0 &&
2964 ci->i_dirty_caps == 0 &&
2965 ci->i_flushing_caps == 0) {
2966 BUG_ON(!ci->i_head_snapc);
2967 ceph_put_snap_context(ci->i_head_snapc);
2968 ci->i_head_snapc = NULL;
2970 /* see comment in __ceph_remove_cap() */
2971 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
2972 drop_inode_snap_realm(ci);
2974 spin_unlock(&ci->i_ceph_lock);
2976 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2977 last ? " last" : "", put ? " put" : "");
2979 if (last && !flushsnaps)
2980 ceph_check_caps(ci, 0, NULL);
2981 else if (flushsnaps)
2982 ceph_flush_snaps(ci, NULL);
2984 wake_up_all(&ci->i_cap_wq);
2990 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2991 * context. Adjust per-snap dirty page accounting as appropriate.
2992 * Once all dirty data for a cap_snap is flushed, flush snapped file
2993 * metadata back to the MDS. If we dropped the last ref, call
2996 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2997 struct ceph_snap_context *snapc)
2999 struct inode *inode = &ci->vfs_inode;
3000 struct ceph_cap_snap *capsnap = NULL;
3004 bool flush_snaps = false;
3005 bool complete_capsnap = false;
3007 spin_lock(&ci->i_ceph_lock);
3008 ci->i_wrbuffer_ref -= nr;
3009 if (ci->i_wrbuffer_ref == 0) {
3014 if (ci->i_head_snapc == snapc) {
3015 ci->i_wrbuffer_ref_head -= nr;
3016 if (ci->i_wrbuffer_ref_head == 0 &&
3017 ci->i_wr_ref == 0 &&
3018 ci->i_dirty_caps == 0 &&
3019 ci->i_flushing_caps == 0) {
3020 BUG_ON(!ci->i_head_snapc);
3021 ceph_put_snap_context(ci->i_head_snapc);
3022 ci->i_head_snapc = NULL;
3024 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3026 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3027 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3028 last ? " LAST" : "");
3030 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3031 if (capsnap->context == snapc) {
3037 capsnap->dirty_pages -= nr;
3038 if (capsnap->dirty_pages == 0) {
3039 complete_capsnap = true;
3040 if (!capsnap->writing) {
3041 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3044 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3049 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3050 " snap %lld %d/%d -> %d/%d %s%s\n",
3051 inode, capsnap, capsnap->context->seq,
3052 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3053 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3054 last ? " (wrbuffer last)" : "",
3055 complete_capsnap ? " (complete capsnap)" : "");
3058 spin_unlock(&ci->i_ceph_lock);
3061 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
3062 } else if (flush_snaps) {
3063 ceph_flush_snaps(ci, NULL);
3065 if (complete_capsnap)
3066 wake_up_all(&ci->i_cap_wq);
3068 /* avoid calling iput_final() in osd dispatch threads */
3069 ceph_async_iput(inode);
3074 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3076 static void invalidate_aliases(struct inode *inode)
3078 struct dentry *dn, *prev = NULL;
3080 dout("invalidate_aliases inode %p\n", inode);
3081 d_prune_aliases(inode);
3083 * For non-directory inode, d_find_alias() only returns
3084 * hashed dentry. After calling d_invalidate(), the
3085 * dentry becomes unhashed.
3087 * For directory inode, d_find_alias() can return
3088 * unhashed dentry. But directory inode should have
3089 * one alias at most.
3091 while ((dn = d_find_alias(inode))) {
3105 struct cap_extra_info {
3106 struct ceph_string *pool_ns;
3116 /* currently issued */
3118 struct timespec64 btime;
3122 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3123 * actually be a revocation if it specifies a smaller cap set.)
3125 * caller holds s_mutex and i_ceph_lock, we drop both.
3127 static void handle_cap_grant(struct inode *inode,
3128 struct ceph_mds_session *session,
3129 struct ceph_cap *cap,
3130 struct ceph_mds_caps *grant,
3131 struct ceph_buffer *xattr_buf,
3132 struct cap_extra_info *extra_info)
3133 __releases(ci->i_ceph_lock)
3134 __releases(session->s_mdsc->snap_rwsem)
3136 struct ceph_inode_info *ci = ceph_inode(inode);
3137 int seq = le32_to_cpu(grant->seq);
3138 int newcaps = le32_to_cpu(grant->caps);
3139 int used, wanted, dirty;
3140 u64 size = le64_to_cpu(grant->size);
3141 u64 max_size = le64_to_cpu(grant->max_size);
3142 unsigned char check_caps = 0;
3143 bool was_stale = cap->cap_gen < session->s_cap_gen;
3145 bool writeback = false;
3146 bool queue_trunc = false;
3147 bool queue_invalidate = false;
3148 bool deleted_inode = false;
3149 bool fill_inline = false;
3151 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3152 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3153 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3158 * If CACHE is being revoked, and we have no dirty buffers,
3159 * try to invalidate (once). (If there are dirty buffers, we
3160 * will invalidate _after_ writeback.)
3162 if (S_ISREG(inode->i_mode) && /* don't invalidate readdir cache */
3163 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3164 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3165 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3166 if (try_nonblocking_invalidate(inode)) {
3167 /* there were locked pages.. invalidate later
3168 in a separate thread. */
3169 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3170 queue_invalidate = true;
3171 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3177 cap->issued = cap->implemented = CEPH_CAP_PIN;
3180 * auth mds of the inode changed. we received the cap export message,
3181 * but still haven't received the cap import message. handle_cap_export
3182 * updated the new auth MDS' cap.
3184 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3185 * that was sent before the cap import message. So don't remove caps.
3187 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3188 WARN_ON(cap != ci->i_auth_cap);
3189 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3191 newcaps |= cap->issued;
3194 /* side effects now are allowed */
3195 cap->cap_gen = session->s_cap_gen;
3198 __check_cap_issue(ci, cap, newcaps);
3200 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3202 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3203 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3204 inode->i_mode = le32_to_cpu(grant->mode);
3205 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3206 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3207 ci->i_btime = extra_info->btime;
3208 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3209 from_kuid(&init_user_ns, inode->i_uid),
3210 from_kgid(&init_user_ns, inode->i_gid));
3213 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3214 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3215 set_nlink(inode, le32_to_cpu(grant->nlink));
3216 if (inode->i_nlink == 0 &&
3217 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3218 deleted_inode = true;
3221 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3223 int len = le32_to_cpu(grant->xattr_len);
3224 u64 version = le64_to_cpu(grant->xattr_version);
3226 if (version > ci->i_xattrs.version) {
3227 dout(" got new xattrs v%llu on %p len %d\n",
3228 version, inode, len);
3229 if (ci->i_xattrs.blob)
3230 ceph_buffer_put(ci->i_xattrs.blob);
3231 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3232 ci->i_xattrs.version = version;
3233 ceph_forget_all_cached_acls(inode);
3234 ceph_security_invalidate_secctx(inode);
3238 if (newcaps & CEPH_CAP_ANY_RD) {
3239 struct timespec64 mtime, atime, ctime;
3240 /* ctime/mtime/atime? */
3241 ceph_decode_timespec64(&mtime, &grant->mtime);
3242 ceph_decode_timespec64(&atime, &grant->atime);
3243 ceph_decode_timespec64(&ctime, &grant->ctime);
3244 ceph_fill_file_time(inode, extra_info->issued,
3245 le32_to_cpu(grant->time_warp_seq),
3246 &ctime, &mtime, &atime);
3249 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3250 ci->i_files = extra_info->nfiles;
3251 ci->i_subdirs = extra_info->nsubdirs;
3254 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3255 /* file layout may have changed */
3256 s64 old_pool = ci->i_layout.pool_id;
3257 struct ceph_string *old_ns;
3259 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3260 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3261 lockdep_is_held(&ci->i_ceph_lock));
3262 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3264 if (ci->i_layout.pool_id != old_pool ||
3265 extra_info->pool_ns != old_ns)
3266 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3268 extra_info->pool_ns = old_ns;
3270 /* size/truncate_seq? */
3271 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3272 le32_to_cpu(grant->truncate_seq),
3273 le64_to_cpu(grant->truncate_size),
3277 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3278 if (max_size != ci->i_max_size) {
3279 dout("max_size %lld -> %llu\n",
3280 ci->i_max_size, max_size);
3281 ci->i_max_size = max_size;
3282 if (max_size >= ci->i_wanted_max_size) {
3283 ci->i_wanted_max_size = 0; /* reset */
3284 ci->i_requested_max_size = 0;
3287 } else if (ci->i_wanted_max_size > ci->i_max_size &&
3288 ci->i_wanted_max_size > ci->i_requested_max_size) {
3289 /* CEPH_CAP_OP_IMPORT */
3294 /* check cap bits */
3295 wanted = __ceph_caps_wanted(ci);
3296 used = __ceph_caps_used(ci);
3297 dirty = __ceph_caps_dirty(ci);
3298 dout(" my wanted = %s, used = %s, dirty %s\n",
3299 ceph_cap_string(wanted),
3300 ceph_cap_string(used),
3301 ceph_cap_string(dirty));
3303 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3304 (wanted & ~(cap->mds_wanted | newcaps))) {
3306 * If mds is importing cap, prior cap messages that update
3307 * 'wanted' may get dropped by mds (migrate seq mismatch).
3309 * We don't send cap message to update 'wanted' if what we
3310 * want are already issued. If mds revokes caps, cap message
3311 * that releases caps also tells mds what we want. But if
3312 * caps got revoked by mds forcedly (session stale). We may
3313 * haven't told mds what we want.
3318 /* revocation, grant, or no-op? */
3319 if (cap->issued & ~newcaps) {
3320 int revoking = cap->issued & ~newcaps;
3322 dout("revocation: %s -> %s (revoking %s)\n",
3323 ceph_cap_string(cap->issued),
3324 ceph_cap_string(newcaps),
3325 ceph_cap_string(revoking));
3326 if (S_ISREG(inode->i_mode) &&
3327 (revoking & used & CEPH_CAP_FILE_BUFFER))
3328 writeback = true; /* initiate writeback; will delay ack */
3329 else if (queue_invalidate &&
3330 revoking == CEPH_CAP_FILE_CACHE &&
3331 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0)
3332 ; /* do nothing yet, invalidation will be queued */
3333 else if (cap == ci->i_auth_cap)
3334 check_caps = 1; /* check auth cap only */
3336 check_caps = 2; /* check all caps */
3337 cap->issued = newcaps;
3338 cap->implemented |= newcaps;
3339 } else if (cap->issued == newcaps) {
3340 dout("caps unchanged: %s -> %s\n",
3341 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3343 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3344 ceph_cap_string(newcaps));
3345 /* non-auth MDS is revoking the newly grant caps ? */
3346 if (cap == ci->i_auth_cap &&
3347 __ceph_caps_revoking_other(ci, cap, newcaps))
3350 cap->issued = newcaps;
3351 cap->implemented |= newcaps; /* add bits only, to
3352 * avoid stepping on a
3353 * pending revocation */
3356 BUG_ON(cap->issued & ~cap->implemented);
3358 if (extra_info->inline_version > 0 &&
3359 extra_info->inline_version >= ci->i_inline_version) {
3360 ci->i_inline_version = extra_info->inline_version;
3361 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3362 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3366 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3367 if (newcaps & ~extra_info->issued)
3369 kick_flushing_inode_caps(session->s_mdsc, session, inode);
3370 up_read(&session->s_mdsc->snap_rwsem);
3372 spin_unlock(&ci->i_ceph_lock);
3376 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3377 extra_info->inline_len);
3380 ceph_queue_vmtruncate(inode);
3384 * queue inode for writeback: we can't actually call
3385 * filemap_write_and_wait, etc. from message handler
3388 ceph_queue_writeback(inode);
3389 if (queue_invalidate)
3390 ceph_queue_invalidate(inode);
3392 invalidate_aliases(inode);
3394 wake_up_all(&ci->i_cap_wq);
3396 if (check_caps == 1)
3397 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
3399 else if (check_caps == 2)
3400 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
3402 mutex_unlock(&session->s_mutex);
3406 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3407 * MDS has been safely committed.
3409 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3410 struct ceph_mds_caps *m,
3411 struct ceph_mds_session *session,
3412 struct ceph_cap *cap)
3413 __releases(ci->i_ceph_lock)
3415 struct ceph_inode_info *ci = ceph_inode(inode);
3416 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3417 struct ceph_cap_flush *cf, *tmp_cf;
3418 LIST_HEAD(to_remove);
3419 unsigned seq = le32_to_cpu(m->seq);
3420 int dirty = le32_to_cpu(m->dirty);
3423 bool wake_ci = false;
3424 bool wake_mdsc = false;
3426 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3427 if (cf->tid == flush_tid)
3429 if (cf->caps == 0) /* capsnap */
3431 if (cf->tid <= flush_tid) {
3432 if (__finish_cap_flush(NULL, ci, cf))
3434 list_add_tail(&cf->i_list, &to_remove);
3436 cleaned &= ~cf->caps;
3442 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3443 " flushing %s -> %s\n",
3444 inode, session->s_mds, seq, ceph_cap_string(dirty),
3445 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3446 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3448 if (list_empty(&to_remove) && !cleaned)
3451 ci->i_flushing_caps &= ~cleaned;
3453 spin_lock(&mdsc->cap_dirty_lock);
3455 list_for_each_entry(cf, &to_remove, i_list) {
3456 if (__finish_cap_flush(mdsc, NULL, cf))
3460 if (ci->i_flushing_caps == 0) {
3461 if (list_empty(&ci->i_cap_flush_list)) {
3462 list_del_init(&ci->i_flushing_item);
3463 if (!list_empty(&session->s_cap_flushing)) {
3464 dout(" mds%d still flushing cap on %p\n",
3466 &list_first_entry(&session->s_cap_flushing,
3467 struct ceph_inode_info,
3468 i_flushing_item)->vfs_inode);
3471 mdsc->num_cap_flushing--;
3472 dout(" inode %p now !flushing\n", inode);
3474 if (ci->i_dirty_caps == 0) {
3475 dout(" inode %p now clean\n", inode);
3476 BUG_ON(!list_empty(&ci->i_dirty_item));
3478 if (ci->i_wr_ref == 0 &&
3479 ci->i_wrbuffer_ref_head == 0) {
3480 BUG_ON(!ci->i_head_snapc);
3481 ceph_put_snap_context(ci->i_head_snapc);
3482 ci->i_head_snapc = NULL;
3485 BUG_ON(list_empty(&ci->i_dirty_item));
3488 spin_unlock(&mdsc->cap_dirty_lock);
3491 spin_unlock(&ci->i_ceph_lock);
3493 while (!list_empty(&to_remove)) {
3494 cf = list_first_entry(&to_remove,
3495 struct ceph_cap_flush, i_list);
3496 list_del(&cf->i_list);
3497 ceph_free_cap_flush(cf);
3501 wake_up_all(&ci->i_cap_wq);
3503 wake_up_all(&mdsc->cap_flushing_wq);
3509 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3510 * throw away our cap_snap.
3512 * Caller hold s_mutex.
3514 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3515 struct ceph_mds_caps *m,
3516 struct ceph_mds_session *session)
3518 struct ceph_inode_info *ci = ceph_inode(inode);
3519 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3520 u64 follows = le64_to_cpu(m->snap_follows);
3521 struct ceph_cap_snap *capsnap;
3522 bool flushed = false;
3523 bool wake_ci = false;
3524 bool wake_mdsc = false;
3526 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3527 inode, ci, session->s_mds, follows);
3529 spin_lock(&ci->i_ceph_lock);
3530 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3531 if (capsnap->follows == follows) {
3532 if (capsnap->cap_flush.tid != flush_tid) {
3533 dout(" cap_snap %p follows %lld tid %lld !="
3534 " %lld\n", capsnap, follows,
3535 flush_tid, capsnap->cap_flush.tid);
3541 dout(" skipping cap_snap %p follows %lld\n",
3542 capsnap, capsnap->follows);
3546 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3547 dout(" removing %p cap_snap %p follows %lld\n",
3548 inode, capsnap, follows);
3549 list_del(&capsnap->ci_item);
3550 if (__finish_cap_flush(NULL, ci, &capsnap->cap_flush))
3553 spin_lock(&mdsc->cap_dirty_lock);
3555 if (list_empty(&ci->i_cap_flush_list))
3556 list_del_init(&ci->i_flushing_item);
3558 if (__finish_cap_flush(mdsc, NULL, &capsnap->cap_flush))
3561 spin_unlock(&mdsc->cap_dirty_lock);
3563 spin_unlock(&ci->i_ceph_lock);
3565 ceph_put_snap_context(capsnap->context);
3566 ceph_put_cap_snap(capsnap);
3568 wake_up_all(&ci->i_cap_wq);
3570 wake_up_all(&mdsc->cap_flushing_wq);
3576 * Handle TRUNC from MDS, indicating file truncation.
3578 * caller hold s_mutex.
3580 static void handle_cap_trunc(struct inode *inode,
3581 struct ceph_mds_caps *trunc,
3582 struct ceph_mds_session *session)
3583 __releases(ci->i_ceph_lock)
3585 struct ceph_inode_info *ci = ceph_inode(inode);
3586 int mds = session->s_mds;
3587 int seq = le32_to_cpu(trunc->seq);
3588 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3589 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3590 u64 size = le64_to_cpu(trunc->size);
3591 int implemented = 0;
3592 int dirty = __ceph_caps_dirty(ci);
3593 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3594 int queue_trunc = 0;
3596 issued |= implemented | dirty;
3598 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3599 inode, mds, seq, truncate_size, truncate_seq);
3600 queue_trunc = ceph_fill_file_size(inode, issued,
3601 truncate_seq, truncate_size, size);
3602 spin_unlock(&ci->i_ceph_lock);
3605 ceph_queue_vmtruncate(inode);
3609 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3610 * different one. If we are the most recent migration we've seen (as
3611 * indicated by mseq), make note of the migrating cap bits for the
3612 * duration (until we see the corresponding IMPORT).
3614 * caller holds s_mutex
3616 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3617 struct ceph_mds_cap_peer *ph,
3618 struct ceph_mds_session *session)
3620 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3621 struct ceph_mds_session *tsession = NULL;
3622 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3623 struct ceph_inode_info *ci = ceph_inode(inode);
3625 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3626 unsigned t_seq, t_mseq;
3628 int mds = session->s_mds;
3631 t_cap_id = le64_to_cpu(ph->cap_id);
3632 t_seq = le32_to_cpu(ph->seq);
3633 t_mseq = le32_to_cpu(ph->mseq);
3634 target = le32_to_cpu(ph->mds);
3636 t_cap_id = t_seq = t_mseq = 0;
3640 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3641 inode, ci, mds, mseq, target);
3643 spin_lock(&ci->i_ceph_lock);
3644 cap = __get_cap_for_mds(ci, mds);
3645 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3649 if (cap->mds_wanted | cap->issued)
3650 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
3651 __ceph_remove_cap(cap, false);
3656 * now we know we haven't received the cap import message yet
3657 * because the exported cap still exist.
3660 issued = cap->issued;
3661 if (issued != cap->implemented)
3662 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3663 "ino (%llx.%llx) mds%d seq %d mseq %d "
3664 "issued %s implemented %s\n",
3665 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3666 ceph_cap_string(issued),
3667 ceph_cap_string(cap->implemented));
3670 tcap = __get_cap_for_mds(ci, target);
3672 /* already have caps from the target */
3673 if (tcap->cap_id == t_cap_id &&
3674 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3675 dout(" updating import cap %p mds%d\n", tcap, target);
3676 tcap->cap_id = t_cap_id;
3677 tcap->seq = t_seq - 1;
3678 tcap->issue_seq = t_seq - 1;
3679 tcap->issued |= issued;
3680 tcap->implemented |= issued;
3681 if (cap == ci->i_auth_cap)
3682 ci->i_auth_cap = tcap;
3684 if (!list_empty(&ci->i_cap_flush_list) &&
3685 ci->i_auth_cap == tcap) {
3686 spin_lock(&mdsc->cap_dirty_lock);
3687 list_move_tail(&ci->i_flushing_item,
3688 &tcap->session->s_cap_flushing);
3689 spin_unlock(&mdsc->cap_dirty_lock);
3692 __ceph_remove_cap(cap, false);
3694 } else if (tsession) {
3695 /* add placeholder for the export tagert */
3696 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3698 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3699 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3701 if (!list_empty(&ci->i_cap_flush_list) &&
3702 ci->i_auth_cap == tcap) {
3703 spin_lock(&mdsc->cap_dirty_lock);
3704 list_move_tail(&ci->i_flushing_item,
3705 &tcap->session->s_cap_flushing);
3706 spin_unlock(&mdsc->cap_dirty_lock);
3709 __ceph_remove_cap(cap, false);
3713 spin_unlock(&ci->i_ceph_lock);
3714 mutex_unlock(&session->s_mutex);
3716 /* open target session */
3717 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3718 if (!IS_ERR(tsession)) {
3720 mutex_lock(&session->s_mutex);
3721 mutex_lock_nested(&tsession->s_mutex,
3722 SINGLE_DEPTH_NESTING);
3724 mutex_lock(&tsession->s_mutex);
3725 mutex_lock_nested(&session->s_mutex,
3726 SINGLE_DEPTH_NESTING);
3728 new_cap = ceph_get_cap(mdsc, NULL);
3737 spin_unlock(&ci->i_ceph_lock);
3738 mutex_unlock(&session->s_mutex);
3740 mutex_unlock(&tsession->s_mutex);
3741 ceph_put_mds_session(tsession);
3744 ceph_put_cap(mdsc, new_cap);
3748 * Handle cap IMPORT.
3750 * caller holds s_mutex. acquires i_ceph_lock
3752 static void handle_cap_import(struct ceph_mds_client *mdsc,
3753 struct inode *inode, struct ceph_mds_caps *im,
3754 struct ceph_mds_cap_peer *ph,
3755 struct ceph_mds_session *session,
3756 struct ceph_cap **target_cap, int *old_issued)
3757 __acquires(ci->i_ceph_lock)
3759 struct ceph_inode_info *ci = ceph_inode(inode);
3760 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3761 int mds = session->s_mds;
3763 unsigned caps = le32_to_cpu(im->caps);
3764 unsigned wanted = le32_to_cpu(im->wanted);
3765 unsigned seq = le32_to_cpu(im->seq);
3766 unsigned mseq = le32_to_cpu(im->migrate_seq);
3767 u64 realmino = le64_to_cpu(im->realm);
3768 u64 cap_id = le64_to_cpu(im->cap_id);
3773 p_cap_id = le64_to_cpu(ph->cap_id);
3774 peer = le32_to_cpu(ph->mds);
3780 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3781 inode, ci, mds, mseq, peer);
3784 spin_lock(&ci->i_ceph_lock);
3785 cap = __get_cap_for_mds(ci, mds);
3788 spin_unlock(&ci->i_ceph_lock);
3789 new_cap = ceph_get_cap(mdsc, NULL);
3795 ceph_put_cap(mdsc, new_cap);
3800 __ceph_caps_issued(ci, &issued);
3801 issued |= __ceph_caps_dirty(ci);
3803 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3804 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3806 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3807 if (ocap && ocap->cap_id == p_cap_id) {
3808 dout(" remove export cap %p mds%d flags %d\n",
3809 ocap, peer, ph->flags);
3810 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3811 (ocap->seq != le32_to_cpu(ph->seq) ||
3812 ocap->mseq != le32_to_cpu(ph->mseq))) {
3813 pr_err_ratelimited("handle_cap_import: "
3814 "mismatched seq/mseq: ino (%llx.%llx) "
3815 "mds%d seq %d mseq %d importer mds%d "
3816 "has peer seq %d mseq %d\n",
3817 ceph_vinop(inode), peer, ocap->seq,
3818 ocap->mseq, mds, le32_to_cpu(ph->seq),
3819 le32_to_cpu(ph->mseq));
3821 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3824 /* make sure we re-request max_size, if necessary */
3825 ci->i_requested_max_size = 0;
3827 *old_issued = issued;
3832 * Handle a caps message from the MDS.
3834 * Identify the appropriate session, inode, and call the right handler
3835 * based on the cap op.
3837 void ceph_handle_caps(struct ceph_mds_session *session,
3838 struct ceph_msg *msg)
3840 struct ceph_mds_client *mdsc = session->s_mdsc;
3841 struct inode *inode;
3842 struct ceph_inode_info *ci;
3843 struct ceph_cap *cap;
3844 struct ceph_mds_caps *h;
3845 struct ceph_mds_cap_peer *peer = NULL;
3846 struct ceph_snap_realm *realm = NULL;
3848 int msg_version = le16_to_cpu(msg->hdr.version);
3850 struct ceph_vino vino;
3852 size_t snaptrace_len;
3854 struct cap_extra_info extra_info = {};
3856 dout("handle_caps from mds%d\n", session->s_mds);
3859 end = msg->front.iov_base + msg->front.iov_len;
3860 if (msg->front.iov_len < sizeof(*h))
3862 h = msg->front.iov_base;
3863 op = le32_to_cpu(h->op);
3864 vino.ino = le64_to_cpu(h->ino);
3865 vino.snap = CEPH_NOSNAP;
3866 seq = le32_to_cpu(h->seq);
3867 mseq = le32_to_cpu(h->migrate_seq);
3870 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3871 p = snaptrace + snaptrace_len;
3873 if (msg_version >= 2) {
3875 ceph_decode_32_safe(&p, end, flock_len, bad);
3876 if (p + flock_len > end)
3881 if (msg_version >= 3) {
3882 if (op == CEPH_CAP_OP_IMPORT) {
3883 if (p + sizeof(*peer) > end)
3887 } else if (op == CEPH_CAP_OP_EXPORT) {
3888 /* recorded in unused fields */
3889 peer = (void *)&h->size;
3893 if (msg_version >= 4) {
3894 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
3895 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
3896 if (p + extra_info.inline_len > end)
3898 extra_info.inline_data = p;
3899 p += extra_info.inline_len;
3902 if (msg_version >= 5) {
3903 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
3906 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
3907 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
3910 if (msg_version >= 8) {
3912 u32 caller_uid, caller_gid;
3916 ceph_decode_64_safe(&p, end, flush_tid, bad);
3918 ceph_decode_32_safe(&p, end, caller_uid, bad);
3919 ceph_decode_32_safe(&p, end, caller_gid, bad);
3921 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
3922 if (pool_ns_len > 0) {
3923 ceph_decode_need(&p, end, pool_ns_len, bad);
3924 extra_info.pool_ns =
3925 ceph_find_or_create_string(p, pool_ns_len);
3930 if (msg_version >= 9) {
3931 struct ceph_timespec *btime;
3933 if (p + sizeof(*btime) > end)
3936 ceph_decode_timespec64(&extra_info.btime, btime);
3937 p += sizeof(*btime);
3938 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
3941 if (msg_version >= 11) {
3944 ceph_decode_32_safe(&p, end, flags, bad);
3946 extra_info.dirstat_valid = true;
3947 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
3948 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
3952 inode = ceph_find_inode(mdsc->fsc->sb, vino);
3953 ci = ceph_inode(inode);
3954 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3957 mutex_lock(&session->s_mutex);
3959 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3963 dout(" i don't have ino %llx\n", vino.ino);
3965 if (op == CEPH_CAP_OP_IMPORT) {
3966 cap = ceph_get_cap(mdsc, NULL);
3967 cap->cap_ino = vino.ino;
3968 cap->queue_release = 1;
3969 cap->cap_id = le64_to_cpu(h->cap_id);
3972 cap->issue_seq = seq;
3973 spin_lock(&session->s_cap_lock);
3974 __ceph_queue_cap_release(session, cap);
3975 spin_unlock(&session->s_cap_lock);
3980 /* these will work even if we don't have a cap yet */
3982 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3983 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
3987 case CEPH_CAP_OP_EXPORT:
3988 handle_cap_export(inode, h, peer, session);
3991 case CEPH_CAP_OP_IMPORT:
3993 if (snaptrace_len) {
3994 down_write(&mdsc->snap_rwsem);
3995 ceph_update_snap_trace(mdsc, snaptrace,
3996 snaptrace + snaptrace_len,
3998 downgrade_write(&mdsc->snap_rwsem);
4000 down_read(&mdsc->snap_rwsem);
4002 handle_cap_import(mdsc, inode, h, peer, session,
4003 &cap, &extra_info.issued);
4004 handle_cap_grant(inode, session, cap,
4005 h, msg->middle, &extra_info);
4007 ceph_put_snap_realm(mdsc, realm);
4011 /* the rest require a cap */
4012 spin_lock(&ci->i_ceph_lock);
4013 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4015 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4016 inode, ceph_ino(inode), ceph_snap(inode),
4018 spin_unlock(&ci->i_ceph_lock);
4019 goto flush_cap_releases;
4022 /* note that each of these drops i_ceph_lock for us */
4024 case CEPH_CAP_OP_REVOKE:
4025 case CEPH_CAP_OP_GRANT:
4026 __ceph_caps_issued(ci, &extra_info.issued);
4027 extra_info.issued |= __ceph_caps_dirty(ci);
4028 handle_cap_grant(inode, session, cap,
4029 h, msg->middle, &extra_info);
4032 case CEPH_CAP_OP_FLUSH_ACK:
4033 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4037 case CEPH_CAP_OP_TRUNC:
4038 handle_cap_trunc(inode, h, session);
4042 spin_unlock(&ci->i_ceph_lock);
4043 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4044 ceph_cap_op_name(op));
4048 mutex_unlock(&session->s_mutex);
4050 ceph_put_string(extra_info.pool_ns);
4051 /* avoid calling iput_final() in mds dispatch threads */
4052 ceph_async_iput(inode);
4057 * send any cap release message to try to move things
4058 * along for the mds (who clearly thinks we still have this
4061 ceph_flush_cap_releases(mdsc, session);
4065 pr_err("ceph_handle_caps: corrupt message\n");
4071 * Delayed work handler to process end of delayed cap release LRU list.
4073 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4075 struct inode *inode;
4076 struct ceph_inode_info *ci;
4077 int flags = CHECK_CAPS_NODELAY;
4079 dout("check_delayed_caps\n");
4081 spin_lock(&mdsc->cap_delay_lock);
4082 if (list_empty(&mdsc->cap_delay_list))
4084 ci = list_first_entry(&mdsc->cap_delay_list,
4085 struct ceph_inode_info,
4087 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4088 time_before(jiffies, ci->i_hold_caps_max))
4090 list_del_init(&ci->i_cap_delay_list);
4092 inode = igrab(&ci->vfs_inode);
4093 spin_unlock(&mdsc->cap_delay_lock);
4096 dout("check_delayed_caps on %p\n", inode);
4097 ceph_check_caps(ci, flags, NULL);
4098 /* avoid calling iput_final() in tick thread */
4099 ceph_async_iput(inode);
4102 spin_unlock(&mdsc->cap_delay_lock);
4106 * Flush all dirty caps to the mds
4108 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4110 struct ceph_inode_info *ci;
4111 struct inode *inode;
4113 dout("flush_dirty_caps\n");
4114 spin_lock(&mdsc->cap_dirty_lock);
4115 while (!list_empty(&mdsc->cap_dirty)) {
4116 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
4118 inode = &ci->vfs_inode;
4120 dout("flush_dirty_caps %p\n", inode);
4121 spin_unlock(&mdsc->cap_dirty_lock);
4122 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
4124 spin_lock(&mdsc->cap_dirty_lock);
4126 spin_unlock(&mdsc->cap_dirty_lock);
4127 dout("flush_dirty_caps done\n");
4130 void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode)
4133 int bits = (fmode << 1) | 1;
4134 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4135 if (bits & (1 << i))
4136 ci->i_nr_by_mode[i]++;
4141 * Drop open file reference. If we were the last open file,
4142 * we may need to release capabilities to the MDS (or schedule
4143 * their delayed release).
4145 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
4148 int bits = (fmode << 1) | 1;
4149 spin_lock(&ci->i_ceph_lock);
4150 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4151 if (bits & (1 << i)) {
4152 BUG_ON(ci->i_nr_by_mode[i] == 0);
4153 if (--ci->i_nr_by_mode[i] == 0)
4157 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
4158 &ci->vfs_inode, fmode,
4159 ci->i_nr_by_mode[0], ci->i_nr_by_mode[1],
4160 ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]);
4161 spin_unlock(&ci->i_ceph_lock);
4163 if (last && ci->i_vino.snap == CEPH_NOSNAP)
4164 ceph_check_caps(ci, 0, NULL);
4168 * For a soon-to-be unlinked file, drop the LINK caps. If it
4169 * looks like the link count will hit 0, drop any other caps (other
4170 * than PIN) we don't specifically want (due to the file still being
4173 int ceph_drop_caps_for_unlink(struct inode *inode)
4175 struct ceph_inode_info *ci = ceph_inode(inode);
4176 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4178 spin_lock(&ci->i_ceph_lock);
4179 if (inode->i_nlink == 1) {
4180 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4182 ci->i_ceph_flags |= CEPH_I_NODELAY;
4183 if (__ceph_caps_dirty(ci)) {
4184 struct ceph_mds_client *mdsc =
4185 ceph_inode_to_client(inode)->mdsc;
4186 __cap_delay_requeue_front(mdsc, ci);
4189 spin_unlock(&ci->i_ceph_lock);
4194 * Helpers for embedding cap and dentry lease releases into mds
4197 * @force is used by dentry_release (below) to force inclusion of a
4198 * record for the directory inode, even when there aren't any caps to
4201 int ceph_encode_inode_release(void **p, struct inode *inode,
4202 int mds, int drop, int unless, int force)
4204 struct ceph_inode_info *ci = ceph_inode(inode);
4205 struct ceph_cap *cap;
4206 struct ceph_mds_request_release *rel = *p;
4210 spin_lock(&ci->i_ceph_lock);
4211 used = __ceph_caps_used(ci);
4212 dirty = __ceph_caps_dirty(ci);
4214 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4215 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4216 ceph_cap_string(unless));
4218 /* only drop unused, clean caps */
4219 drop &= ~(used | dirty);
4221 cap = __get_cap_for_mds(ci, mds);
4222 if (cap && __cap_is_valid(cap)) {
4223 unless &= cap->issued;
4225 if (unless & CEPH_CAP_AUTH_EXCL)
4226 drop &= ~CEPH_CAP_AUTH_SHARED;
4227 if (unless & CEPH_CAP_LINK_EXCL)
4228 drop &= ~CEPH_CAP_LINK_SHARED;
4229 if (unless & CEPH_CAP_XATTR_EXCL)
4230 drop &= ~CEPH_CAP_XATTR_SHARED;
4231 if (unless & CEPH_CAP_FILE_EXCL)
4232 drop &= ~CEPH_CAP_FILE_SHARED;
4235 if (force || (cap->issued & drop)) {
4236 if (cap->issued & drop) {
4237 int wanted = __ceph_caps_wanted(ci);
4238 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
4239 wanted |= cap->mds_wanted;
4240 dout("encode_inode_release %p cap %p "
4241 "%s -> %s, wanted %s -> %s\n", inode, cap,
4242 ceph_cap_string(cap->issued),
4243 ceph_cap_string(cap->issued & ~drop),
4244 ceph_cap_string(cap->mds_wanted),
4245 ceph_cap_string(wanted));
4247 cap->issued &= ~drop;
4248 cap->implemented &= ~drop;
4249 cap->mds_wanted = wanted;
4251 dout("encode_inode_release %p cap %p %s"
4252 " (force)\n", inode, cap,
4253 ceph_cap_string(cap->issued));
4256 rel->ino = cpu_to_le64(ceph_ino(inode));
4257 rel->cap_id = cpu_to_le64(cap->cap_id);
4258 rel->seq = cpu_to_le32(cap->seq);
4259 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4260 rel->mseq = cpu_to_le32(cap->mseq);
4261 rel->caps = cpu_to_le32(cap->implemented);
4262 rel->wanted = cpu_to_le32(cap->mds_wanted);
4268 dout("encode_inode_release %p cap %p %s (noop)\n",
4269 inode, cap, ceph_cap_string(cap->issued));
4272 spin_unlock(&ci->i_ceph_lock);
4276 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4278 int mds, int drop, int unless)
4280 struct dentry *parent = NULL;
4281 struct ceph_mds_request_release *rel = *p;
4282 struct ceph_dentry_info *di = ceph_dentry(dentry);
4287 * force an record for the directory caps if we have a dentry lease.
4288 * this is racy (can't take i_ceph_lock and d_lock together), but it
4289 * doesn't have to be perfect; the mds will revoke anything we don't
4292 spin_lock(&dentry->d_lock);
4293 if (di->lease_session && di->lease_session->s_mds == mds)
4296 parent = dget(dentry->d_parent);
4297 dir = d_inode(parent);
4299 spin_unlock(&dentry->d_lock);
4301 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4304 spin_lock(&dentry->d_lock);
4305 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4306 dout("encode_dentry_release %p mds%d seq %d\n",
4307 dentry, mds, (int)di->lease_seq);
4308 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4309 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4310 *p += dentry->d_name.len;
4311 rel->dname_seq = cpu_to_le32(di->lease_seq);
4312 __ceph_mdsc_drop_dentry_lease(dentry);
4314 spin_unlock(&dentry->d_lock);