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 ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
577 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
582 * If FILE_SHARED is newly issued, mark dir not complete. We don't
583 * know what happened to this directory while we didn't have the cap.
584 * If FILE_SHARED is being revoked, also mark dir not complete. It
585 * stops on-going cached readdir.
587 if ((issued & CEPH_CAP_FILE_SHARED) != (had & CEPH_CAP_FILE_SHARED)) {
588 if (issued & CEPH_CAP_FILE_SHARED)
589 atomic_inc(&ci->i_shared_gen);
590 if (S_ISDIR(ci->vfs_inode.i_mode)) {
591 dout(" marking %p NOT complete\n", &ci->vfs_inode);
592 __ceph_dir_clear_complete(ci);
598 * Add a capability under the given MDS session.
600 * Caller should hold session snap_rwsem (read) and ci->i_ceph_lock
602 * @fmode is the open file mode, if we are opening a file, otherwise
603 * it is < 0. (This is so we can atomically add the cap and add an
604 * open file reference to it.)
606 void ceph_add_cap(struct inode *inode,
607 struct ceph_mds_session *session, u64 cap_id,
608 int fmode, unsigned issued, unsigned wanted,
609 unsigned seq, unsigned mseq, u64 realmino, int flags,
610 struct ceph_cap **new_cap)
612 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
613 struct ceph_inode_info *ci = ceph_inode(inode);
614 struct ceph_cap *cap;
615 int mds = session->s_mds;
619 lockdep_assert_held(&ci->i_ceph_lock);
621 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
622 session->s_mds, cap_id, ceph_cap_string(issued), seq);
625 * If we are opening the file, include file mode wanted bits
629 wanted |= ceph_caps_for_mode(fmode);
631 spin_lock(&session->s_gen_ttl_lock);
632 gen = session->s_cap_gen;
633 spin_unlock(&session->s_gen_ttl_lock);
635 cap = __get_cap_for_mds(ci, mds);
641 cap->implemented = 0;
647 __insert_cap_node(ci, cap);
649 /* add to session cap list */
650 cap->session = session;
651 spin_lock(&session->s_cap_lock);
652 list_add_tail(&cap->session_caps, &session->s_caps);
653 session->s_nr_caps++;
654 spin_unlock(&session->s_cap_lock);
656 spin_lock(&session->s_cap_lock);
657 list_move_tail(&cap->session_caps, &session->s_caps);
658 spin_unlock(&session->s_cap_lock);
660 if (cap->cap_gen < gen)
661 cap->issued = cap->implemented = CEPH_CAP_PIN;
664 * auth mds of the inode changed. we received the cap export
665 * message, but still haven't received the cap import message.
666 * handle_cap_export() updated the new auth MDS' cap.
668 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
669 * a message that was send before the cap import message. So
672 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
673 WARN_ON(cap != ci->i_auth_cap);
674 WARN_ON(cap->cap_id != cap_id);
677 issued |= cap->issued;
678 flags |= CEPH_CAP_FLAG_AUTH;
682 if (!ci->i_snap_realm ||
683 ((flags & CEPH_CAP_FLAG_AUTH) &&
684 realmino != (u64)-1 && ci->i_snap_realm->ino != realmino)) {
686 * add this inode to the appropriate snap realm
688 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
691 struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
693 spin_lock(&oldrealm->inodes_with_caps_lock);
694 list_del_init(&ci->i_snap_realm_item);
695 spin_unlock(&oldrealm->inodes_with_caps_lock);
698 spin_lock(&realm->inodes_with_caps_lock);
699 list_add(&ci->i_snap_realm_item,
700 &realm->inodes_with_caps);
701 ci->i_snap_realm = realm;
702 if (realm->ino == ci->i_vino.ino)
703 realm->inode = inode;
704 spin_unlock(&realm->inodes_with_caps_lock);
707 ceph_put_snap_realm(mdsc, oldrealm);
709 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
715 __check_cap_issue(ci, cap, issued);
718 * If we are issued caps we don't want, or the mds' wanted
719 * value appears to be off, queue a check so we'll release
720 * later and/or update the mds wanted value.
722 actual_wanted = __ceph_caps_wanted(ci);
723 if ((wanted & ~actual_wanted) ||
724 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
725 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
726 ceph_cap_string(issued), ceph_cap_string(wanted),
727 ceph_cap_string(actual_wanted));
728 __cap_delay_requeue(mdsc, ci, true);
731 if (flags & CEPH_CAP_FLAG_AUTH) {
732 if (!ci->i_auth_cap ||
733 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
734 ci->i_auth_cap = cap;
735 cap->mds_wanted = wanted;
738 WARN_ON(ci->i_auth_cap == cap);
741 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
742 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
743 ceph_cap_string(issued|cap->issued), seq, mds);
744 cap->cap_id = cap_id;
745 cap->issued = issued;
746 cap->implemented |= issued;
747 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
748 cap->mds_wanted = wanted;
750 cap->mds_wanted |= wanted;
752 cap->issue_seq = seq;
757 __ceph_get_fmode(ci, fmode);
761 * Return true if cap has not timed out and belongs to the current
762 * generation of the MDS session (i.e. has not gone 'stale' due to
763 * us losing touch with the mds).
765 static int __cap_is_valid(struct ceph_cap *cap)
770 spin_lock(&cap->session->s_gen_ttl_lock);
771 gen = cap->session->s_cap_gen;
772 ttl = cap->session->s_cap_ttl;
773 spin_unlock(&cap->session->s_gen_ttl_lock);
775 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
776 dout("__cap_is_valid %p cap %p issued %s "
777 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
778 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
786 * Return set of valid cap bits issued to us. Note that caps time
787 * out, and may be invalidated in bulk if the client session times out
788 * and session->s_cap_gen is bumped.
790 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
792 int have = ci->i_snap_caps;
793 struct ceph_cap *cap;
798 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
799 cap = rb_entry(p, struct ceph_cap, ci_node);
800 if (!__cap_is_valid(cap))
802 dout("__ceph_caps_issued %p cap %p issued %s\n",
803 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
806 *implemented |= cap->implemented;
809 * exclude caps issued by non-auth MDS, but are been revoking
810 * by the auth MDS. The non-auth MDS should be revoking/exporting
811 * these caps, but the message is delayed.
813 if (ci->i_auth_cap) {
814 cap = ci->i_auth_cap;
815 have &= ~cap->implemented | cap->issued;
821 * Get cap bits issued by caps other than @ocap
823 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
825 int have = ci->i_snap_caps;
826 struct ceph_cap *cap;
829 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
830 cap = rb_entry(p, struct ceph_cap, ci_node);
833 if (!__cap_is_valid(cap))
841 * Move a cap to the end of the LRU (oldest caps at list head, newest
844 static void __touch_cap(struct ceph_cap *cap)
846 struct ceph_mds_session *s = cap->session;
848 spin_lock(&s->s_cap_lock);
849 if (!s->s_cap_iterator) {
850 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
852 list_move_tail(&cap->session_caps, &s->s_caps);
854 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
855 &cap->ci->vfs_inode, cap, s->s_mds);
857 spin_unlock(&s->s_cap_lock);
861 * Check if we hold the given mask. If so, move the cap(s) to the
862 * front of their respective LRUs. (This is the preferred way for
863 * callers to check for caps they want.)
865 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
867 struct ceph_cap *cap;
869 int have = ci->i_snap_caps;
871 if ((have & mask) == mask) {
872 dout("__ceph_caps_issued_mask ino 0x%lx snap issued %s"
873 " (mask %s)\n", ci->vfs_inode.i_ino,
874 ceph_cap_string(have),
875 ceph_cap_string(mask));
879 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
880 cap = rb_entry(p, struct ceph_cap, ci_node);
881 if (!__cap_is_valid(cap))
883 if ((cap->issued & mask) == mask) {
884 dout("__ceph_caps_issued_mask ino 0x%lx cap %p issued %s"
885 " (mask %s)\n", ci->vfs_inode.i_ino, cap,
886 ceph_cap_string(cap->issued),
887 ceph_cap_string(mask));
893 /* does a combination of caps satisfy mask? */
895 if ((have & mask) == mask) {
896 dout("__ceph_caps_issued_mask ino 0x%lx combo issued %s"
897 " (mask %s)\n", ci->vfs_inode.i_ino,
898 ceph_cap_string(cap->issued),
899 ceph_cap_string(mask));
903 /* touch this + preceding caps */
905 for (q = rb_first(&ci->i_caps); q != p;
907 cap = rb_entry(q, struct ceph_cap,
909 if (!__cap_is_valid(cap))
922 * Return true if mask caps are currently being revoked by an MDS.
924 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
925 struct ceph_cap *ocap, int mask)
927 struct ceph_cap *cap;
930 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
931 cap = rb_entry(p, struct ceph_cap, ci_node);
933 (cap->implemented & ~cap->issued & mask))
939 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
941 struct inode *inode = &ci->vfs_inode;
944 spin_lock(&ci->i_ceph_lock);
945 ret = __ceph_caps_revoking_other(ci, NULL, mask);
946 spin_unlock(&ci->i_ceph_lock);
947 dout("ceph_caps_revoking %p %s = %d\n", inode,
948 ceph_cap_string(mask), ret);
952 int __ceph_caps_used(struct ceph_inode_info *ci)
956 used |= CEPH_CAP_PIN;
958 used |= CEPH_CAP_FILE_RD;
959 if (ci->i_rdcache_ref ||
960 (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */
961 ci->vfs_inode.i_data.nrpages))
962 used |= CEPH_CAP_FILE_CACHE;
964 used |= CEPH_CAP_FILE_WR;
965 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
966 used |= CEPH_CAP_FILE_BUFFER;
971 * wanted, by virtue of open file modes
973 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
976 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
977 if (ci->i_nr_by_mode[i])
982 return ceph_caps_for_mode(bits >> 1);
986 * Return caps we have registered with the MDS(s) as 'wanted'.
988 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
990 struct ceph_cap *cap;
994 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
995 cap = rb_entry(p, struct ceph_cap, ci_node);
996 if (check && !__cap_is_valid(cap))
998 if (cap == ci->i_auth_cap)
999 mds_wanted |= cap->mds_wanted;
1001 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
1007 * called under i_ceph_lock
1009 static int __ceph_is_single_caps(struct ceph_inode_info *ci)
1011 return rb_first(&ci->i_caps) == rb_last(&ci->i_caps);
1014 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
1016 return !RB_EMPTY_ROOT(&ci->i_caps);
1019 int ceph_is_any_caps(struct inode *inode)
1021 struct ceph_inode_info *ci = ceph_inode(inode);
1024 spin_lock(&ci->i_ceph_lock);
1025 ret = __ceph_is_any_caps(ci);
1026 spin_unlock(&ci->i_ceph_lock);
1031 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
1033 struct ceph_snap_realm *realm = ci->i_snap_realm;
1034 spin_lock(&realm->inodes_with_caps_lock);
1035 list_del_init(&ci->i_snap_realm_item);
1036 ci->i_snap_realm_counter++;
1037 ci->i_snap_realm = NULL;
1038 if (realm->ino == ci->i_vino.ino)
1039 realm->inode = NULL;
1040 spin_unlock(&realm->inodes_with_caps_lock);
1041 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
1046 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
1048 * caller should hold i_ceph_lock.
1049 * caller will not hold session s_mutex if called from destroy_inode.
1051 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1053 struct ceph_mds_session *session = cap->session;
1054 struct ceph_inode_info *ci = cap->ci;
1055 struct ceph_mds_client *mdsc =
1056 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1059 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1061 /* remove from inode's cap rbtree, and clear auth cap */
1062 rb_erase(&cap->ci_node, &ci->i_caps);
1063 if (ci->i_auth_cap == cap)
1064 ci->i_auth_cap = NULL;
1066 /* remove from session list */
1067 spin_lock(&session->s_cap_lock);
1068 if (session->s_cap_iterator == cap) {
1069 /* not yet, we are iterating over this very cap */
1070 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1073 list_del_init(&cap->session_caps);
1074 session->s_nr_caps--;
1075 cap->session = NULL;
1078 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1082 * s_cap_reconnect is protected by s_cap_lock. no one changes
1083 * s_cap_gen while session is in the reconnect state.
1085 if (queue_release &&
1086 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
1087 cap->queue_release = 1;
1089 __ceph_queue_cap_release(session, cap);
1093 cap->queue_release = 0;
1095 cap->cap_ino = ci->i_vino.ino;
1097 spin_unlock(&session->s_cap_lock);
1100 ceph_put_cap(mdsc, cap);
1102 /* when reconnect denied, we remove session caps forcibly,
1103 * i_wr_ref can be non-zero. If there are ongoing write,
1104 * keep i_snap_realm.
1106 if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm)
1107 drop_inode_snap_realm(ci);
1109 if (!__ceph_is_any_real_caps(ci))
1110 __cap_delay_cancel(mdsc, ci);
1113 struct cap_msg_args {
1114 struct ceph_mds_session *session;
1115 u64 ino, cid, follows;
1116 u64 flush_tid, oldest_flush_tid, size, max_size;
1119 struct ceph_buffer *xattr_buf;
1120 struct timespec64 atime, mtime, ctime, btime;
1121 int op, caps, wanted, dirty;
1122 u32 seq, issue_seq, mseq, time_warp_seq;
1131 * Build and send a cap message to the given MDS.
1133 * Caller should be holding s_mutex.
1135 static int send_cap_msg(struct cap_msg_args *arg)
1137 struct ceph_mds_caps *fc;
1138 struct ceph_msg *msg;
1141 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1143 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1144 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1145 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(arg->op),
1146 arg->cid, arg->ino, ceph_cap_string(arg->caps),
1147 ceph_cap_string(arg->wanted), ceph_cap_string(arg->dirty),
1148 arg->seq, arg->issue_seq, arg->flush_tid, arg->oldest_flush_tid,
1149 arg->mseq, arg->follows, arg->size, arg->max_size,
1151 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1153 /* flock buffer size + inline version + inline data size +
1154 * osd_epoch_barrier + oldest_flush_tid */
1155 extra_len = 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4;
1156 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1161 msg->hdr.version = cpu_to_le16(10);
1162 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1164 fc = msg->front.iov_base;
1165 memset(fc, 0, sizeof(*fc));
1167 fc->cap_id = cpu_to_le64(arg->cid);
1168 fc->op = cpu_to_le32(arg->op);
1169 fc->seq = cpu_to_le32(arg->seq);
1170 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1171 fc->migrate_seq = cpu_to_le32(arg->mseq);
1172 fc->caps = cpu_to_le32(arg->caps);
1173 fc->wanted = cpu_to_le32(arg->wanted);
1174 fc->dirty = cpu_to_le32(arg->dirty);
1175 fc->ino = cpu_to_le64(arg->ino);
1176 fc->snap_follows = cpu_to_le64(arg->follows);
1178 fc->size = cpu_to_le64(arg->size);
1179 fc->max_size = cpu_to_le64(arg->max_size);
1180 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1181 ceph_encode_timespec64(&fc->atime, &arg->atime);
1182 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1183 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1185 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1186 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1187 fc->mode = cpu_to_le32(arg->mode);
1189 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1190 if (arg->xattr_buf) {
1191 msg->middle = ceph_buffer_get(arg->xattr_buf);
1192 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1193 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1197 /* flock buffer size (version 2) */
1198 ceph_encode_32(&p, 0);
1199 /* inline version (version 4) */
1200 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1201 /* inline data size */
1202 ceph_encode_32(&p, 0);
1204 * osd_epoch_barrier (version 5)
1205 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1206 * case it was recently changed
1208 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1209 /* oldest_flush_tid (version 6) */
1210 ceph_encode_64(&p, arg->oldest_flush_tid);
1213 * caller_uid/caller_gid (version 7)
1215 * Currently, we don't properly track which caller dirtied the caps
1216 * last, and force a flush of them when there is a conflict. For now,
1217 * just set this to 0:0, to emulate how the MDS has worked up to now.
1219 ceph_encode_32(&p, 0);
1220 ceph_encode_32(&p, 0);
1222 /* pool namespace (version 8) (mds always ignores this) */
1223 ceph_encode_32(&p, 0);
1225 /* btime and change_attr (version 9) */
1226 ceph_encode_timespec64(p, &arg->btime);
1227 p += sizeof(struct ceph_timespec);
1228 ceph_encode_64(&p, arg->change_attr);
1230 /* Advisory flags (version 10) */
1231 ceph_encode_32(&p, arg->flags);
1233 ceph_con_send(&arg->session->s_con, msg);
1238 * Queue cap releases when an inode is dropped from our cache.
1240 void __ceph_remove_caps(struct ceph_inode_info *ci)
1244 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1245 * may call __ceph_caps_issued_mask() on a freeing inode. */
1246 spin_lock(&ci->i_ceph_lock);
1247 p = rb_first(&ci->i_caps);
1249 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1251 __ceph_remove_cap(cap, true);
1253 spin_unlock(&ci->i_ceph_lock);
1257 * Send a cap msg on the given inode. Update our caps state, then
1258 * drop i_ceph_lock and send the message.
1260 * Make note of max_size reported/requested from mds, revoked caps
1261 * that have now been implemented.
1263 * Return non-zero if delayed release, or we experienced an error
1264 * such that the caller should requeue + retry later.
1266 * called with i_ceph_lock, then drops it.
1267 * caller should hold snap_rwsem (read), s_mutex.
1269 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1270 int op, int flags, int used, int want, int retain,
1271 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1272 __releases(cap->ci->i_ceph_lock)
1274 struct ceph_inode_info *ci = cap->ci;
1275 struct inode *inode = &ci->vfs_inode;
1276 struct ceph_buffer *old_blob = NULL;
1277 struct cap_msg_args arg;
1283 held = cap->issued | cap->implemented;
1284 revoking = cap->implemented & ~cap->issued;
1285 retain &= ~revoking;
1287 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1288 inode, cap, cap->session,
1289 ceph_cap_string(held), ceph_cap_string(held & retain),
1290 ceph_cap_string(revoking));
1291 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1293 arg.session = cap->session;
1295 /* don't release wanted unless we've waited a bit. */
1296 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1297 time_before(jiffies, ci->i_hold_caps_min)) {
1298 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1299 ceph_cap_string(cap->issued),
1300 ceph_cap_string(cap->issued & retain),
1301 ceph_cap_string(cap->mds_wanted),
1302 ceph_cap_string(want));
1303 want |= cap->mds_wanted;
1304 retain |= cap->issued;
1307 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1308 if (want & ~cap->mds_wanted) {
1309 /* user space may open/close single file frequently.
1310 * This avoids droping mds_wanted immediately after
1311 * requesting new mds_wanted.
1313 __cap_set_timeouts(mdsc, ci);
1316 cap->issued &= retain; /* drop bits we don't want */
1317 if (cap->implemented & ~cap->issued) {
1319 * Wake up any waiters on wanted -> needed transition.
1320 * This is due to the weird transition from buffered
1321 * to sync IO... we need to flush dirty pages _before_
1322 * allowing sync writes to avoid reordering.
1326 cap->implemented &= cap->issued | used;
1327 cap->mds_wanted = want;
1329 arg.ino = ceph_vino(inode).ino;
1330 arg.cid = cap->cap_id;
1331 arg.follows = flushing ? ci->i_head_snapc->seq : 0;
1332 arg.flush_tid = flush_tid;
1333 arg.oldest_flush_tid = oldest_flush_tid;
1335 arg.size = inode->i_size;
1336 ci->i_reported_size = arg.size;
1337 arg.max_size = ci->i_wanted_max_size;
1338 ci->i_requested_max_size = arg.max_size;
1340 if (flushing & CEPH_CAP_XATTR_EXCL) {
1341 old_blob = __ceph_build_xattrs_blob(ci);
1342 arg.xattr_version = ci->i_xattrs.version;
1343 arg.xattr_buf = ci->i_xattrs.blob;
1345 arg.xattr_buf = NULL;
1348 arg.mtime = inode->i_mtime;
1349 arg.atime = inode->i_atime;
1350 arg.ctime = inode->i_ctime;
1351 arg.btime = ci->i_btime;
1352 arg.change_attr = inode_peek_iversion_raw(inode);
1355 arg.caps = cap->implemented;
1357 arg.dirty = flushing;
1360 arg.issue_seq = cap->issue_seq;
1361 arg.mseq = cap->mseq;
1362 arg.time_warp_seq = ci->i_time_warp_seq;
1364 arg.uid = inode->i_uid;
1365 arg.gid = inode->i_gid;
1366 arg.mode = inode->i_mode;
1368 arg.inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1369 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1370 !list_empty(&ci->i_cap_snaps)) {
1371 struct ceph_cap_snap *capsnap;
1372 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1373 if (capsnap->cap_flush.tid)
1375 if (capsnap->need_flush) {
1376 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1383 spin_unlock(&ci->i_ceph_lock);
1385 ceph_buffer_put(old_blob);
1387 ret = send_cap_msg(&arg);
1389 dout("error sending cap msg, must requeue %p\n", inode);
1394 wake_up_all(&ci->i_cap_wq);
1399 static inline int __send_flush_snap(struct inode *inode,
1400 struct ceph_mds_session *session,
1401 struct ceph_cap_snap *capsnap,
1402 u32 mseq, u64 oldest_flush_tid)
1404 struct cap_msg_args arg;
1406 arg.session = session;
1407 arg.ino = ceph_vino(inode).ino;
1409 arg.follows = capsnap->follows;
1410 arg.flush_tid = capsnap->cap_flush.tid;
1411 arg.oldest_flush_tid = oldest_flush_tid;
1413 arg.size = capsnap->size;
1415 arg.xattr_version = capsnap->xattr_version;
1416 arg.xattr_buf = capsnap->xattr_blob;
1418 arg.atime = capsnap->atime;
1419 arg.mtime = capsnap->mtime;
1420 arg.ctime = capsnap->ctime;
1421 arg.btime = capsnap->btime;
1422 arg.change_attr = capsnap->change_attr;
1424 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1425 arg.caps = capsnap->issued;
1427 arg.dirty = capsnap->dirty;
1432 arg.time_warp_seq = capsnap->time_warp_seq;
1434 arg.uid = capsnap->uid;
1435 arg.gid = capsnap->gid;
1436 arg.mode = capsnap->mode;
1438 arg.inline_data = capsnap->inline_data;
1441 return send_cap_msg(&arg);
1445 * When a snapshot is taken, clients accumulate dirty metadata on
1446 * inodes with capabilities in ceph_cap_snaps to describe the file
1447 * state at the time the snapshot was taken. This must be flushed
1448 * asynchronously back to the MDS once sync writes complete and dirty
1449 * data is written out.
1451 * Called under i_ceph_lock. Takes s_mutex as needed.
1453 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1454 struct ceph_mds_session *session)
1455 __releases(ci->i_ceph_lock)
1456 __acquires(ci->i_ceph_lock)
1458 struct inode *inode = &ci->vfs_inode;
1459 struct ceph_mds_client *mdsc = session->s_mdsc;
1460 struct ceph_cap_snap *capsnap;
1461 u64 oldest_flush_tid = 0;
1462 u64 first_tid = 1, last_tid = 0;
1464 dout("__flush_snaps %p session %p\n", inode, session);
1466 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1468 * we need to wait for sync writes to complete and for dirty
1469 * pages to be written out.
1471 if (capsnap->dirty_pages || capsnap->writing)
1474 /* should be removed by ceph_try_drop_cap_snap() */
1475 BUG_ON(!capsnap->need_flush);
1477 /* only flush each capsnap once */
1478 if (capsnap->cap_flush.tid > 0) {
1479 dout(" already flushed %p, skipping\n", capsnap);
1483 spin_lock(&mdsc->cap_dirty_lock);
1484 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1485 list_add_tail(&capsnap->cap_flush.g_list,
1486 &mdsc->cap_flush_list);
1487 if (oldest_flush_tid == 0)
1488 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1489 if (list_empty(&ci->i_flushing_item)) {
1490 list_add_tail(&ci->i_flushing_item,
1491 &session->s_cap_flushing);
1493 spin_unlock(&mdsc->cap_dirty_lock);
1495 list_add_tail(&capsnap->cap_flush.i_list,
1496 &ci->i_cap_flush_list);
1499 first_tid = capsnap->cap_flush.tid;
1500 last_tid = capsnap->cap_flush.tid;
1503 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1505 while (first_tid <= last_tid) {
1506 struct ceph_cap *cap = ci->i_auth_cap;
1507 struct ceph_cap_flush *cf;
1510 if (!(cap && cap->session == session)) {
1511 dout("__flush_snaps %p auth cap %p not mds%d, "
1512 "stop\n", inode, cap, session->s_mds);
1517 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1518 if (cf->tid >= first_tid) {
1526 first_tid = cf->tid + 1;
1528 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1529 refcount_inc(&capsnap->nref);
1530 spin_unlock(&ci->i_ceph_lock);
1532 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1533 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1535 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1538 pr_err("__flush_snaps: error sending cap flushsnap, "
1539 "ino (%llx.%llx) tid %llu follows %llu\n",
1540 ceph_vinop(inode), cf->tid, capsnap->follows);
1543 ceph_put_cap_snap(capsnap);
1544 spin_lock(&ci->i_ceph_lock);
1548 void ceph_flush_snaps(struct ceph_inode_info *ci,
1549 struct ceph_mds_session **psession)
1551 struct inode *inode = &ci->vfs_inode;
1552 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1553 struct ceph_mds_session *session = NULL;
1556 dout("ceph_flush_snaps %p\n", inode);
1558 session = *psession;
1560 spin_lock(&ci->i_ceph_lock);
1561 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1562 dout(" no capsnap needs flush, doing nothing\n");
1565 if (!ci->i_auth_cap) {
1566 dout(" no auth cap (migrating?), doing nothing\n");
1570 mds = ci->i_auth_cap->session->s_mds;
1571 if (session && session->s_mds != mds) {
1572 dout(" oops, wrong session %p mutex\n", session);
1573 mutex_unlock(&session->s_mutex);
1574 ceph_put_mds_session(session);
1578 spin_unlock(&ci->i_ceph_lock);
1579 mutex_lock(&mdsc->mutex);
1580 session = __ceph_lookup_mds_session(mdsc, mds);
1581 mutex_unlock(&mdsc->mutex);
1583 dout(" inverting session/ino locks on %p\n", session);
1584 mutex_lock(&session->s_mutex);
1589 // make sure flushsnap messages are sent in proper order.
1590 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1591 __kick_flushing_caps(mdsc, session, ci, 0);
1593 __ceph_flush_snaps(ci, session);
1595 spin_unlock(&ci->i_ceph_lock);
1598 *psession = session;
1599 } else if (session) {
1600 mutex_unlock(&session->s_mutex);
1601 ceph_put_mds_session(session);
1603 /* we flushed them all; remove this inode from the queue */
1604 spin_lock(&mdsc->snap_flush_lock);
1605 list_del_init(&ci->i_snap_flush_item);
1606 spin_unlock(&mdsc->snap_flush_lock);
1610 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1611 * Caller is then responsible for calling __mark_inode_dirty with the
1612 * returned flags value.
1614 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1615 struct ceph_cap_flush **pcf)
1617 struct ceph_mds_client *mdsc =
1618 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1619 struct inode *inode = &ci->vfs_inode;
1620 int was = ci->i_dirty_caps;
1623 if (!ci->i_auth_cap) {
1624 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1625 "but no auth cap (session was closed?)\n",
1626 inode, ceph_ino(inode), ceph_cap_string(mask));
1630 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1631 ceph_cap_string(mask), ceph_cap_string(was),
1632 ceph_cap_string(was | mask));
1633 ci->i_dirty_caps |= mask;
1635 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1636 swap(ci->i_prealloc_cap_flush, *pcf);
1638 if (!ci->i_head_snapc) {
1639 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1640 ci->i_head_snapc = ceph_get_snap_context(
1641 ci->i_snap_realm->cached_context);
1643 dout(" inode %p now dirty snapc %p auth cap %p\n",
1644 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1645 BUG_ON(!list_empty(&ci->i_dirty_item));
1646 spin_lock(&mdsc->cap_dirty_lock);
1647 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1648 spin_unlock(&mdsc->cap_dirty_lock);
1649 if (ci->i_flushing_caps == 0) {
1651 dirty |= I_DIRTY_SYNC;
1654 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1656 BUG_ON(list_empty(&ci->i_dirty_item));
1657 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1658 (mask & CEPH_CAP_FILE_BUFFER))
1659 dirty |= I_DIRTY_DATASYNC;
1660 __cap_delay_requeue(mdsc, ci, true);
1664 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1666 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1669 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1672 kmem_cache_free(ceph_cap_flush_cachep, cf);
1675 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1677 if (!list_empty(&mdsc->cap_flush_list)) {
1678 struct ceph_cap_flush *cf =
1679 list_first_entry(&mdsc->cap_flush_list,
1680 struct ceph_cap_flush, g_list);
1687 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1688 * Return true if caller needs to wake up flush waiters.
1690 static bool __finish_cap_flush(struct ceph_mds_client *mdsc,
1691 struct ceph_inode_info *ci,
1692 struct ceph_cap_flush *cf)
1694 struct ceph_cap_flush *prev;
1695 bool wake = cf->wake;
1697 /* are there older pending cap flushes? */
1698 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1699 prev = list_prev_entry(cf, g_list);
1703 list_del(&cf->g_list);
1705 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1706 prev = list_prev_entry(cf, i_list);
1710 list_del(&cf->i_list);
1718 * Add dirty inode to the flushing list. Assigned a seq number so we
1719 * can wait for caps to flush without starving.
1721 * Called under i_ceph_lock. Returns the flush tid.
1723 static u64 __mark_caps_flushing(struct inode *inode,
1724 struct ceph_mds_session *session, bool wake,
1725 u64 *oldest_flush_tid)
1727 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1728 struct ceph_inode_info *ci = ceph_inode(inode);
1729 struct ceph_cap_flush *cf = NULL;
1732 BUG_ON(ci->i_dirty_caps == 0);
1733 BUG_ON(list_empty(&ci->i_dirty_item));
1734 BUG_ON(!ci->i_prealloc_cap_flush);
1736 flushing = ci->i_dirty_caps;
1737 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1738 ceph_cap_string(flushing),
1739 ceph_cap_string(ci->i_flushing_caps),
1740 ceph_cap_string(ci->i_flushing_caps | flushing));
1741 ci->i_flushing_caps |= flushing;
1742 ci->i_dirty_caps = 0;
1743 dout(" inode %p now !dirty\n", inode);
1745 swap(cf, ci->i_prealloc_cap_flush);
1746 cf->caps = flushing;
1749 spin_lock(&mdsc->cap_dirty_lock);
1750 list_del_init(&ci->i_dirty_item);
1752 cf->tid = ++mdsc->last_cap_flush_tid;
1753 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1754 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1756 if (list_empty(&ci->i_flushing_item)) {
1757 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1758 mdsc->num_cap_flushing++;
1760 spin_unlock(&mdsc->cap_dirty_lock);
1762 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1768 * try to invalidate mapping pages without blocking.
1770 static int try_nonblocking_invalidate(struct inode *inode)
1772 struct ceph_inode_info *ci = ceph_inode(inode);
1773 u32 invalidating_gen = ci->i_rdcache_gen;
1775 spin_unlock(&ci->i_ceph_lock);
1776 invalidate_mapping_pages(&inode->i_data, 0, -1);
1777 spin_lock(&ci->i_ceph_lock);
1779 if (inode->i_data.nrpages == 0 &&
1780 invalidating_gen == ci->i_rdcache_gen) {
1782 dout("try_nonblocking_invalidate %p success\n", inode);
1783 /* save any racing async invalidate some trouble */
1784 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1787 dout("try_nonblocking_invalidate %p failed\n", inode);
1791 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1793 loff_t size = ci->vfs_inode.i_size;
1794 /* mds will adjust max size according to the reported size */
1795 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1797 if (size >= ci->i_max_size)
1799 /* half of previous max_size increment has been used */
1800 if (ci->i_max_size > ci->i_reported_size &&
1801 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1807 * Swiss army knife function to examine currently used and wanted
1808 * versus held caps. Release, flush, ack revoked caps to mds as
1811 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1812 * cap release further.
1813 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1814 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1817 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1818 struct ceph_mds_session *session)
1820 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1821 struct ceph_mds_client *mdsc = fsc->mdsc;
1822 struct inode *inode = &ci->vfs_inode;
1823 struct ceph_cap *cap;
1824 u64 flush_tid, oldest_flush_tid;
1825 int file_wanted, used, cap_used;
1826 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1827 int issued, implemented, want, retain, revoking, flushing = 0;
1828 int mds = -1; /* keep track of how far we've gone through i_caps list
1829 to avoid an infinite loop on retry */
1831 int delayed = 0, sent = 0;
1832 bool no_delay = flags & CHECK_CAPS_NODELAY;
1833 bool queue_invalidate = false;
1834 bool tried_invalidate = false;
1836 /* if we are unmounting, flush any unused caps immediately. */
1840 spin_lock(&ci->i_ceph_lock);
1842 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1843 flags |= CHECK_CAPS_FLUSH;
1845 if (!(flags & CHECK_CAPS_AUTHONLY) ||
1846 (ci->i_auth_cap && __ceph_is_single_caps(ci)))
1847 __cap_delay_cancel(mdsc, ci);
1851 spin_lock(&ci->i_ceph_lock);
1853 file_wanted = __ceph_caps_file_wanted(ci);
1854 used = __ceph_caps_used(ci);
1855 issued = __ceph_caps_issued(ci, &implemented);
1856 revoking = implemented & ~issued;
1859 retain = file_wanted | used | CEPH_CAP_PIN;
1860 if (!mdsc->stopping && inode->i_nlink > 0) {
1862 retain |= CEPH_CAP_ANY; /* be greedy */
1863 } else if (S_ISDIR(inode->i_mode) &&
1864 (issued & CEPH_CAP_FILE_SHARED) &&
1865 __ceph_dir_is_complete(ci)) {
1867 * If a directory is complete, we want to keep
1868 * the exclusive cap. So that MDS does not end up
1869 * revoking the shared cap on every create/unlink
1872 if (IS_RDONLY(inode))
1873 want = CEPH_CAP_ANY_SHARED;
1875 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1879 retain |= CEPH_CAP_ANY_SHARED;
1881 * keep RD only if we didn't have the file open RW,
1882 * because then the mds would revoke it anyway to
1883 * journal max_size=0.
1885 if (ci->i_max_size == 0)
1886 retain |= CEPH_CAP_ANY_RD;
1890 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1891 " issued %s revoking %s retain %s %s%s%s\n", inode,
1892 ceph_cap_string(file_wanted),
1893 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1894 ceph_cap_string(ci->i_flushing_caps),
1895 ceph_cap_string(issued), ceph_cap_string(revoking),
1896 ceph_cap_string(retain),
1897 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1898 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1899 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1902 * If we no longer need to hold onto old our caps, and we may
1903 * have cached pages, but don't want them, then try to invalidate.
1904 * If we fail, it's because pages are locked.... try again later.
1906 if ((!no_delay || mdsc->stopping) &&
1907 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */
1908 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1909 inode->i_data.nrpages && /* have cached pages */
1910 (revoking & (CEPH_CAP_FILE_CACHE|
1911 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1912 !tried_invalidate) {
1913 dout("check_caps trying to invalidate on %p\n", inode);
1914 if (try_nonblocking_invalidate(inode) < 0) {
1915 dout("check_caps queuing invalidate\n");
1916 queue_invalidate = true;
1917 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1919 tried_invalidate = true;
1923 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1924 cap = rb_entry(p, struct ceph_cap, ci_node);
1926 /* avoid looping forever */
1927 if (mds >= cap->mds ||
1928 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1931 /* NOTE: no side-effects allowed, until we take s_mutex */
1934 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1935 cap_used &= ~ci->i_auth_cap->issued;
1937 revoking = cap->implemented & ~cap->issued;
1938 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1939 cap->mds, cap, ceph_cap_string(cap_used),
1940 ceph_cap_string(cap->issued),
1941 ceph_cap_string(cap->implemented),
1942 ceph_cap_string(revoking));
1944 if (cap == ci->i_auth_cap &&
1945 (cap->issued & CEPH_CAP_FILE_WR)) {
1946 /* request larger max_size from MDS? */
1947 if (ci->i_wanted_max_size > ci->i_max_size &&
1948 ci->i_wanted_max_size > ci->i_requested_max_size) {
1949 dout("requesting new max_size\n");
1953 /* approaching file_max? */
1954 if (__ceph_should_report_size(ci)) {
1955 dout("i_size approaching max_size\n");
1959 /* flush anything dirty? */
1960 if (cap == ci->i_auth_cap) {
1961 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
1962 dout("flushing dirty caps\n");
1965 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
1966 dout("flushing snap caps\n");
1971 /* completed revocation? going down and there are no caps? */
1972 if (revoking && (revoking & cap_used) == 0) {
1973 dout("completed revocation of %s\n",
1974 ceph_cap_string(cap->implemented & ~cap->issued));
1978 /* want more caps from mds? */
1979 if (want & ~(cap->mds_wanted | cap->issued))
1982 /* things we might delay */
1983 if ((cap->issued & ~retain) == 0)
1984 continue; /* nope, all good */
1990 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1991 time_before(jiffies, ci->i_hold_caps_max)) {
1992 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1993 ceph_cap_string(cap->issued),
1994 ceph_cap_string(cap->issued & retain),
1995 ceph_cap_string(cap->mds_wanted),
1996 ceph_cap_string(want));
2002 if (session && session != cap->session) {
2003 dout("oops, wrong session %p mutex\n", session);
2004 mutex_unlock(&session->s_mutex);
2008 session = cap->session;
2009 if (mutex_trylock(&session->s_mutex) == 0) {
2010 dout("inverting session/ino locks on %p\n",
2012 spin_unlock(&ci->i_ceph_lock);
2013 if (took_snap_rwsem) {
2014 up_read(&mdsc->snap_rwsem);
2015 took_snap_rwsem = 0;
2017 mutex_lock(&session->s_mutex);
2022 /* kick flushing and flush snaps before sending normal
2024 if (cap == ci->i_auth_cap &&
2026 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2027 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2028 __kick_flushing_caps(mdsc, session, ci, 0);
2029 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2030 __ceph_flush_snaps(ci, session);
2035 /* take snap_rwsem after session mutex */
2036 if (!took_snap_rwsem) {
2037 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
2038 dout("inverting snap/in locks on %p\n",
2040 spin_unlock(&ci->i_ceph_lock);
2041 down_read(&mdsc->snap_rwsem);
2042 took_snap_rwsem = 1;
2045 took_snap_rwsem = 1;
2048 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2049 flushing = ci->i_dirty_caps;
2050 flush_tid = __mark_caps_flushing(inode, session, false,
2055 spin_lock(&mdsc->cap_dirty_lock);
2056 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2057 spin_unlock(&mdsc->cap_dirty_lock);
2060 mds = cap->mds; /* remember mds, so we don't repeat */
2063 /* __send_cap drops i_ceph_lock */
2064 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, 0,
2065 cap_used, want, retain, flushing,
2066 flush_tid, oldest_flush_tid);
2067 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2070 /* Reschedule delayed caps release if we delayed anything */
2072 __cap_delay_requeue(mdsc, ci, false);
2074 spin_unlock(&ci->i_ceph_lock);
2076 if (queue_invalidate)
2077 ceph_queue_invalidate(inode);
2080 mutex_unlock(&session->s_mutex);
2081 if (took_snap_rwsem)
2082 up_read(&mdsc->snap_rwsem);
2086 * Try to flush dirty caps back to the auth mds.
2088 static int try_flush_caps(struct inode *inode, u64 *ptid)
2090 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2091 struct ceph_inode_info *ci = ceph_inode(inode);
2092 struct ceph_mds_session *session = NULL;
2094 u64 flush_tid = 0, oldest_flush_tid = 0;
2097 spin_lock(&ci->i_ceph_lock);
2099 if (ci->i_dirty_caps && ci->i_auth_cap) {
2100 struct ceph_cap *cap = ci->i_auth_cap;
2103 if (session != cap->session) {
2104 spin_unlock(&ci->i_ceph_lock);
2106 mutex_unlock(&session->s_mutex);
2107 session = cap->session;
2108 mutex_lock(&session->s_mutex);
2111 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) {
2112 spin_unlock(&ci->i_ceph_lock);
2116 if (ci->i_ceph_flags &
2117 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2118 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2119 __kick_flushing_caps(mdsc, session, ci, 0);
2120 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2121 __ceph_flush_snaps(ci, session);
2125 flushing = ci->i_dirty_caps;
2126 flush_tid = __mark_caps_flushing(inode, session, true,
2129 /* __send_cap drops i_ceph_lock */
2130 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2131 CEPH_CLIENT_CAPS_SYNC,
2132 __ceph_caps_used(ci),
2133 __ceph_caps_wanted(ci),
2134 (cap->issued | cap->implemented),
2135 flushing, flush_tid, oldest_flush_tid);
2138 spin_lock(&ci->i_ceph_lock);
2139 __cap_delay_requeue(mdsc, ci, true);
2140 spin_unlock(&ci->i_ceph_lock);
2143 if (!list_empty(&ci->i_cap_flush_list)) {
2144 struct ceph_cap_flush *cf =
2145 list_last_entry(&ci->i_cap_flush_list,
2146 struct ceph_cap_flush, i_list);
2148 flush_tid = cf->tid;
2150 flushing = ci->i_flushing_caps;
2151 spin_unlock(&ci->i_ceph_lock);
2155 mutex_unlock(&session->s_mutex);
2162 * Return true if we've flushed caps through the given flush_tid.
2164 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2166 struct ceph_inode_info *ci = ceph_inode(inode);
2169 spin_lock(&ci->i_ceph_lock);
2170 if (!list_empty(&ci->i_cap_flush_list)) {
2171 struct ceph_cap_flush * cf =
2172 list_first_entry(&ci->i_cap_flush_list,
2173 struct ceph_cap_flush, i_list);
2174 if (cf->tid <= flush_tid)
2177 spin_unlock(&ci->i_ceph_lock);
2182 * wait for any unsafe requests to complete.
2184 static int unsafe_request_wait(struct inode *inode)
2186 struct ceph_inode_info *ci = ceph_inode(inode);
2187 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2190 spin_lock(&ci->i_unsafe_lock);
2191 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2192 req1 = list_last_entry(&ci->i_unsafe_dirops,
2193 struct ceph_mds_request,
2195 ceph_mdsc_get_request(req1);
2197 if (!list_empty(&ci->i_unsafe_iops)) {
2198 req2 = list_last_entry(&ci->i_unsafe_iops,
2199 struct ceph_mds_request,
2200 r_unsafe_target_item);
2201 ceph_mdsc_get_request(req2);
2203 spin_unlock(&ci->i_unsafe_lock);
2205 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2206 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2208 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2209 ceph_timeout_jiffies(req1->r_timeout));
2212 ceph_mdsc_put_request(req1);
2215 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2216 ceph_timeout_jiffies(req2->r_timeout));
2219 ceph_mdsc_put_request(req2);
2224 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2226 struct ceph_file_info *fi = file->private_data;
2227 struct inode *inode = file->f_mapping->host;
2228 struct ceph_inode_info *ci = ceph_inode(inode);
2233 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2235 ret = file_write_and_wait_range(file, start, end);
2239 dirty = try_flush_caps(inode, &flush_tid);
2240 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2242 err = unsafe_request_wait(inode);
2245 * only wait on non-file metadata writeback (the mds
2246 * can recover size and mtime, so we don't need to
2249 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2250 err = wait_event_interruptible(ci->i_cap_wq,
2251 caps_are_flushed(inode, flush_tid));
2257 if (errseq_check(&ci->i_meta_err, READ_ONCE(fi->meta_err))) {
2258 spin_lock(&file->f_lock);
2259 err = errseq_check_and_advance(&ci->i_meta_err,
2261 spin_unlock(&file->f_lock);
2266 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2271 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2272 * queue inode for flush but don't do so immediately, because we can
2273 * get by with fewer MDS messages if we wait for data writeback to
2276 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2278 struct ceph_inode_info *ci = ceph_inode(inode);
2282 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2284 dout("write_inode %p wait=%d\n", inode, wait);
2286 dirty = try_flush_caps(inode, &flush_tid);
2288 err = wait_event_interruptible(ci->i_cap_wq,
2289 caps_are_flushed(inode, flush_tid));
2291 struct ceph_mds_client *mdsc =
2292 ceph_sb_to_client(inode->i_sb)->mdsc;
2294 spin_lock(&ci->i_ceph_lock);
2295 if (__ceph_caps_dirty(ci))
2296 __cap_delay_requeue_front(mdsc, ci);
2297 spin_unlock(&ci->i_ceph_lock);
2302 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2303 struct ceph_mds_session *session,
2304 struct ceph_inode_info *ci,
2305 u64 oldest_flush_tid)
2306 __releases(ci->i_ceph_lock)
2307 __acquires(ci->i_ceph_lock)
2309 struct inode *inode = &ci->vfs_inode;
2310 struct ceph_cap *cap;
2311 struct ceph_cap_flush *cf;
2314 u64 last_snap_flush = 0;
2316 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2318 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2320 last_snap_flush = cf->tid;
2325 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2326 if (cf->tid < first_tid)
2329 cap = ci->i_auth_cap;
2330 if (!(cap && cap->session == session)) {
2331 pr_err("%p auth cap %p not mds%d ???\n",
2332 inode, cap, session->s_mds);
2336 first_tid = cf->tid + 1;
2339 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2340 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2341 ci->i_ceph_flags |= CEPH_I_NODELAY;
2343 ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2344 (cf->tid < last_snap_flush ?
2345 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2346 __ceph_caps_used(ci),
2347 __ceph_caps_wanted(ci),
2348 (cap->issued | cap->implemented),
2349 cf->caps, cf->tid, oldest_flush_tid);
2351 pr_err("kick_flushing_caps: error sending "
2352 "cap flush, ino (%llx.%llx) "
2353 "tid %llu flushing %s\n",
2354 ceph_vinop(inode), cf->tid,
2355 ceph_cap_string(cf->caps));
2358 struct ceph_cap_snap *capsnap =
2359 container_of(cf, struct ceph_cap_snap,
2361 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2362 inode, capsnap, cf->tid,
2363 ceph_cap_string(capsnap->dirty));
2365 refcount_inc(&capsnap->nref);
2366 spin_unlock(&ci->i_ceph_lock);
2368 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2371 pr_err("kick_flushing_caps: error sending "
2372 "cap flushsnap, ino (%llx.%llx) "
2373 "tid %llu follows %llu\n",
2374 ceph_vinop(inode), cf->tid,
2378 ceph_put_cap_snap(capsnap);
2381 spin_lock(&ci->i_ceph_lock);
2385 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2386 struct ceph_mds_session *session)
2388 struct ceph_inode_info *ci;
2389 struct ceph_cap *cap;
2390 u64 oldest_flush_tid;
2392 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2394 spin_lock(&mdsc->cap_dirty_lock);
2395 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2396 spin_unlock(&mdsc->cap_dirty_lock);
2398 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2399 spin_lock(&ci->i_ceph_lock);
2400 cap = ci->i_auth_cap;
2401 if (!(cap && cap->session == session)) {
2402 pr_err("%p auth cap %p not mds%d ???\n",
2403 &ci->vfs_inode, cap, session->s_mds);
2404 spin_unlock(&ci->i_ceph_lock);
2410 * if flushing caps were revoked, we re-send the cap flush
2411 * in client reconnect stage. This guarantees MDS * processes
2412 * the cap flush message before issuing the flushing caps to
2415 if ((cap->issued & ci->i_flushing_caps) !=
2416 ci->i_flushing_caps) {
2417 /* encode_caps_cb() also will reset these sequence
2418 * numbers. make sure sequence numbers in cap flush
2419 * message match later reconnect message */
2423 __kick_flushing_caps(mdsc, session, ci,
2426 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2429 spin_unlock(&ci->i_ceph_lock);
2433 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2434 struct ceph_mds_session *session)
2436 struct ceph_inode_info *ci;
2437 struct ceph_cap *cap;
2438 u64 oldest_flush_tid;
2440 dout("kick_flushing_caps mds%d\n", session->s_mds);
2442 spin_lock(&mdsc->cap_dirty_lock);
2443 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2444 spin_unlock(&mdsc->cap_dirty_lock);
2446 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2447 spin_lock(&ci->i_ceph_lock);
2448 cap = ci->i_auth_cap;
2449 if (!(cap && cap->session == session)) {
2450 pr_err("%p auth cap %p not mds%d ???\n",
2451 &ci->vfs_inode, cap, session->s_mds);
2452 spin_unlock(&ci->i_ceph_lock);
2455 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2456 __kick_flushing_caps(mdsc, session, ci,
2459 spin_unlock(&ci->i_ceph_lock);
2463 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2464 struct ceph_mds_session *session,
2465 struct inode *inode)
2466 __releases(ci->i_ceph_lock)
2468 struct ceph_inode_info *ci = ceph_inode(inode);
2469 struct ceph_cap *cap;
2471 cap = ci->i_auth_cap;
2472 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2473 ceph_cap_string(ci->i_flushing_caps));
2475 if (!list_empty(&ci->i_cap_flush_list)) {
2476 u64 oldest_flush_tid;
2477 spin_lock(&mdsc->cap_dirty_lock);
2478 list_move_tail(&ci->i_flushing_item,
2479 &cap->session->s_cap_flushing);
2480 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2481 spin_unlock(&mdsc->cap_dirty_lock);
2483 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2484 spin_unlock(&ci->i_ceph_lock);
2486 spin_unlock(&ci->i_ceph_lock);
2492 * Take references to capabilities we hold, so that we don't release
2493 * them to the MDS prematurely.
2495 * Protected by i_ceph_lock.
2497 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2498 bool snap_rwsem_locked)
2500 if (got & CEPH_CAP_PIN)
2502 if (got & CEPH_CAP_FILE_RD)
2504 if (got & CEPH_CAP_FILE_CACHE)
2505 ci->i_rdcache_ref++;
2506 if (got & CEPH_CAP_FILE_WR) {
2507 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2508 BUG_ON(!snap_rwsem_locked);
2509 ci->i_head_snapc = ceph_get_snap_context(
2510 ci->i_snap_realm->cached_context);
2514 if (got & CEPH_CAP_FILE_BUFFER) {
2515 if (ci->i_wb_ref == 0)
2516 ihold(&ci->vfs_inode);
2518 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2519 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2524 * Try to grab cap references. Specify those refs we @want, and the
2525 * minimal set we @need. Also include the larger offset we are writing
2526 * to (when applicable), and check against max_size here as well.
2527 * Note that caller is responsible for ensuring max_size increases are
2528 * requested from the MDS.
2530 * Returns 0 if caps were not able to be acquired (yet), a 1 if they were,
2531 * or a negative error code.
2533 * FIXME: how does a 0 return differ from -EAGAIN?
2540 static int try_get_cap_refs(struct inode *inode, int need, int want,
2541 loff_t endoff, int flags, int *got)
2543 struct ceph_inode_info *ci = ceph_inode(inode);
2544 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2546 int have, implemented;
2548 bool snap_rwsem_locked = false;
2550 dout("get_cap_refs %p need %s want %s\n", inode,
2551 ceph_cap_string(need), ceph_cap_string(want));
2554 spin_lock(&ci->i_ceph_lock);
2556 if ((flags & CHECK_FILELOCK) &&
2557 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2558 dout("try_get_cap_refs %p error filelock\n", inode);
2563 /* make sure file is actually open */
2564 file_wanted = __ceph_caps_file_wanted(ci);
2565 if ((file_wanted & need) != need) {
2566 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2567 ceph_cap_string(need), ceph_cap_string(file_wanted));
2572 /* finish pending truncate */
2573 while (ci->i_truncate_pending) {
2574 spin_unlock(&ci->i_ceph_lock);
2575 if (snap_rwsem_locked) {
2576 up_read(&mdsc->snap_rwsem);
2577 snap_rwsem_locked = false;
2579 __ceph_do_pending_vmtruncate(inode);
2580 spin_lock(&ci->i_ceph_lock);
2583 have = __ceph_caps_issued(ci, &implemented);
2585 if (have & need & CEPH_CAP_FILE_WR) {
2586 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2587 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2588 inode, endoff, ci->i_max_size);
2589 if (endoff > ci->i_requested_max_size)
2594 * If a sync write is in progress, we must wait, so that we
2595 * can get a final snapshot value for size+mtime.
2597 if (__ceph_have_pending_cap_snap(ci)) {
2598 dout("get_cap_refs %p cap_snap_pending\n", inode);
2603 if ((have & need) == need) {
2605 * Look at (implemented & ~have & not) so that we keep waiting
2606 * on transition from wanted -> needed caps. This is needed
2607 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2608 * going before a prior buffered writeback happens.
2610 int not = want & ~(have & need);
2611 int revoking = implemented & ~have;
2612 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2613 inode, ceph_cap_string(have), ceph_cap_string(not),
2614 ceph_cap_string(revoking));
2615 if ((revoking & not) == 0) {
2616 if (!snap_rwsem_locked &&
2617 !ci->i_head_snapc &&
2618 (need & CEPH_CAP_FILE_WR)) {
2619 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2621 * we can not call down_read() when
2622 * task isn't in TASK_RUNNING state
2624 if (flags & NON_BLOCKING) {
2629 spin_unlock(&ci->i_ceph_lock);
2630 down_read(&mdsc->snap_rwsem);
2631 snap_rwsem_locked = true;
2634 snap_rwsem_locked = true;
2636 *got = need | (have & want);
2637 if ((need & CEPH_CAP_FILE_RD) &&
2638 !(*got & CEPH_CAP_FILE_CACHE))
2639 ceph_disable_fscache_readpage(ci);
2640 __take_cap_refs(ci, *got, true);
2644 int session_readonly = false;
2645 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2646 struct ceph_mds_session *s = ci->i_auth_cap->session;
2647 spin_lock(&s->s_cap_lock);
2648 session_readonly = s->s_readonly;
2649 spin_unlock(&s->s_cap_lock);
2651 if (session_readonly) {
2652 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2653 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2658 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) {
2660 if (READ_ONCE(mdsc->fsc->mount_state) ==
2661 CEPH_MOUNT_SHUTDOWN) {
2662 dout("get_cap_refs %p forced umount\n", inode);
2666 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2667 if (need & ~(mds_wanted & need)) {
2668 dout("get_cap_refs %p caps were dropped"
2669 " (session killed?)\n", inode);
2673 if (!(file_wanted & ~mds_wanted))
2674 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED;
2677 dout("get_cap_refs %p have %s needed %s\n", inode,
2678 ceph_cap_string(have), ceph_cap_string(need));
2681 spin_unlock(&ci->i_ceph_lock);
2682 if (snap_rwsem_locked)
2683 up_read(&mdsc->snap_rwsem);
2685 dout("get_cap_refs %p ret %d got %s\n", inode,
2686 ret, ceph_cap_string(*got));
2691 * Check the offset we are writing up to against our current
2692 * max_size. If necessary, tell the MDS we want to write to
2695 static void check_max_size(struct inode *inode, loff_t endoff)
2697 struct ceph_inode_info *ci = ceph_inode(inode);
2700 /* do we need to explicitly request a larger max_size? */
2701 spin_lock(&ci->i_ceph_lock);
2702 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2703 dout("write %p at large endoff %llu, req max_size\n",
2705 ci->i_wanted_max_size = endoff;
2707 /* duplicate ceph_check_caps()'s logic */
2708 if (ci->i_auth_cap &&
2709 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2710 ci->i_wanted_max_size > ci->i_max_size &&
2711 ci->i_wanted_max_size > ci->i_requested_max_size)
2713 spin_unlock(&ci->i_ceph_lock);
2715 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2718 int ceph_try_get_caps(struct inode *inode, int need, int want,
2719 bool nonblock, int *got)
2723 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2724 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO|CEPH_CAP_FILE_SHARED));
2725 ret = ceph_pool_perm_check(inode, need);
2729 ret = try_get_cap_refs(inode, need, want, 0,
2730 (nonblock ? NON_BLOCKING : 0), got);
2731 return ret == -EAGAIN ? 0 : ret;
2735 * Wait for caps, and take cap references. If we can't get a WR cap
2736 * due to a small max_size, make sure we check_max_size (and possibly
2737 * ask the mds) so we don't get hung up indefinitely.
2739 int ceph_get_caps(struct file *filp, int need, int want,
2740 loff_t endoff, int *got, struct page **pinned_page)
2742 struct ceph_file_info *fi = filp->private_data;
2743 struct inode *inode = file_inode(filp);
2744 struct ceph_inode_info *ci = ceph_inode(inode);
2745 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2746 int ret, _got, flags;
2748 ret = ceph_pool_perm_check(inode, need);
2752 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2753 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2758 check_max_size(inode, endoff);
2760 flags = atomic_read(&fi->num_locks) ? CHECK_FILELOCK : 0;
2762 ret = try_get_cap_refs(inode, need, want, endoff,
2767 struct ceph_mds_client *mdsc = fsc->mdsc;
2769 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2771 cw.ino = inode->i_ino;
2772 cw.tgid = current->tgid;
2776 spin_lock(&mdsc->caps_list_lock);
2777 list_add(&cw.list, &mdsc->cap_wait_list);
2778 spin_unlock(&mdsc->caps_list_lock);
2780 add_wait_queue(&ci->i_cap_wq, &wait);
2782 flags |= NON_BLOCKING;
2783 while (!(ret = try_get_cap_refs(inode, need, want,
2784 endoff, flags, &_got))) {
2785 if (signal_pending(current)) {
2789 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2792 remove_wait_queue(&ci->i_cap_wq, &wait);
2794 spin_lock(&mdsc->caps_list_lock);
2796 spin_unlock(&mdsc->caps_list_lock);
2802 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2803 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2804 if (ret >= 0 && _got)
2805 ceph_put_cap_refs(ci, _got);
2810 if (ret == -ESTALE) {
2811 /* session was killed, try renew caps */
2812 ret = ceph_renew_caps(inode);
2819 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2820 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2821 i_size_read(inode) > 0) {
2823 find_get_page(inode->i_mapping, 0);
2825 if (PageUptodate(page)) {
2826 *pinned_page = page;
2832 * drop cap refs first because getattr while
2833 * holding * caps refs can cause deadlock.
2835 ceph_put_cap_refs(ci, _got);
2839 * getattr request will bring inline data into
2842 ret = __ceph_do_getattr(inode, NULL,
2843 CEPH_STAT_CAP_INLINE_DATA,
2852 if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
2853 ceph_fscache_revalidate_cookie(ci);
2860 * Take cap refs. Caller must already know we hold at least one ref
2861 * on the caps in question or we don't know this is safe.
2863 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2865 spin_lock(&ci->i_ceph_lock);
2866 __take_cap_refs(ci, caps, false);
2867 spin_unlock(&ci->i_ceph_lock);
2872 * drop cap_snap that is not associated with any snapshot.
2873 * we don't need to send FLUSHSNAP message for it.
2875 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2876 struct ceph_cap_snap *capsnap)
2878 if (!capsnap->need_flush &&
2879 !capsnap->writing && !capsnap->dirty_pages) {
2880 dout("dropping cap_snap %p follows %llu\n",
2881 capsnap, capsnap->follows);
2882 BUG_ON(capsnap->cap_flush.tid > 0);
2883 ceph_put_snap_context(capsnap->context);
2884 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2885 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2887 list_del(&capsnap->ci_item);
2888 ceph_put_cap_snap(capsnap);
2897 * If we released the last ref on any given cap, call ceph_check_caps
2898 * to release (or schedule a release).
2900 * If we are releasing a WR cap (from a sync write), finalize any affected
2901 * cap_snap, and wake up any waiters.
2903 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2905 struct inode *inode = &ci->vfs_inode;
2906 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2908 spin_lock(&ci->i_ceph_lock);
2909 if (had & CEPH_CAP_PIN)
2911 if (had & CEPH_CAP_FILE_RD)
2912 if (--ci->i_rd_ref == 0)
2914 if (had & CEPH_CAP_FILE_CACHE)
2915 if (--ci->i_rdcache_ref == 0)
2917 if (had & CEPH_CAP_FILE_BUFFER) {
2918 if (--ci->i_wb_ref == 0) {
2922 dout("put_cap_refs %p wb %d -> %d (?)\n",
2923 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2925 if (had & CEPH_CAP_FILE_WR)
2926 if (--ci->i_wr_ref == 0) {
2928 if (__ceph_have_pending_cap_snap(ci)) {
2929 struct ceph_cap_snap *capsnap =
2930 list_last_entry(&ci->i_cap_snaps,
2931 struct ceph_cap_snap,
2933 capsnap->writing = 0;
2934 if (ceph_try_drop_cap_snap(ci, capsnap))
2936 else if (__ceph_finish_cap_snap(ci, capsnap))
2940 if (ci->i_wrbuffer_ref_head == 0 &&
2941 ci->i_dirty_caps == 0 &&
2942 ci->i_flushing_caps == 0) {
2943 BUG_ON(!ci->i_head_snapc);
2944 ceph_put_snap_context(ci->i_head_snapc);
2945 ci->i_head_snapc = NULL;
2947 /* see comment in __ceph_remove_cap() */
2948 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2949 drop_inode_snap_realm(ci);
2951 spin_unlock(&ci->i_ceph_lock);
2953 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2954 last ? " last" : "", put ? " put" : "");
2956 if (last && !flushsnaps)
2957 ceph_check_caps(ci, 0, NULL);
2958 else if (flushsnaps)
2959 ceph_flush_snaps(ci, NULL);
2961 wake_up_all(&ci->i_cap_wq);
2967 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2968 * context. Adjust per-snap dirty page accounting as appropriate.
2969 * Once all dirty data for a cap_snap is flushed, flush snapped file
2970 * metadata back to the MDS. If we dropped the last ref, call
2973 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2974 struct ceph_snap_context *snapc)
2976 struct inode *inode = &ci->vfs_inode;
2977 struct ceph_cap_snap *capsnap = NULL;
2981 bool flush_snaps = false;
2982 bool complete_capsnap = false;
2984 spin_lock(&ci->i_ceph_lock);
2985 ci->i_wrbuffer_ref -= nr;
2986 if (ci->i_wrbuffer_ref == 0) {
2991 if (ci->i_head_snapc == snapc) {
2992 ci->i_wrbuffer_ref_head -= nr;
2993 if (ci->i_wrbuffer_ref_head == 0 &&
2994 ci->i_wr_ref == 0 &&
2995 ci->i_dirty_caps == 0 &&
2996 ci->i_flushing_caps == 0) {
2997 BUG_ON(!ci->i_head_snapc);
2998 ceph_put_snap_context(ci->i_head_snapc);
2999 ci->i_head_snapc = NULL;
3001 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3003 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3004 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3005 last ? " LAST" : "");
3007 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3008 if (capsnap->context == snapc) {
3014 capsnap->dirty_pages -= nr;
3015 if (capsnap->dirty_pages == 0) {
3016 complete_capsnap = true;
3017 if (!capsnap->writing) {
3018 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3021 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3026 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3027 " snap %lld %d/%d -> %d/%d %s%s\n",
3028 inode, capsnap, capsnap->context->seq,
3029 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3030 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3031 last ? " (wrbuffer last)" : "",
3032 complete_capsnap ? " (complete capsnap)" : "");
3035 spin_unlock(&ci->i_ceph_lock);
3038 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
3039 } else if (flush_snaps) {
3040 ceph_flush_snaps(ci, NULL);
3042 if (complete_capsnap)
3043 wake_up_all(&ci->i_cap_wq);
3045 /* avoid calling iput_final() in osd dispatch threads */
3046 ceph_async_iput(inode);
3051 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3053 static void invalidate_aliases(struct inode *inode)
3055 struct dentry *dn, *prev = NULL;
3057 dout("invalidate_aliases inode %p\n", inode);
3058 d_prune_aliases(inode);
3060 * For non-directory inode, d_find_alias() only returns
3061 * hashed dentry. After calling d_invalidate(), the
3062 * dentry becomes unhashed.
3064 * For directory inode, d_find_alias() can return
3065 * unhashed dentry. But directory inode should have
3066 * one alias at most.
3068 while ((dn = d_find_alias(inode))) {
3082 struct cap_extra_info {
3083 struct ceph_string *pool_ns;
3093 /* currently issued */
3095 struct timespec64 btime;
3099 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3100 * actually be a revocation if it specifies a smaller cap set.)
3102 * caller holds s_mutex and i_ceph_lock, we drop both.
3104 static void handle_cap_grant(struct inode *inode,
3105 struct ceph_mds_session *session,
3106 struct ceph_cap *cap,
3107 struct ceph_mds_caps *grant,
3108 struct ceph_buffer *xattr_buf,
3109 struct cap_extra_info *extra_info)
3110 __releases(ci->i_ceph_lock)
3111 __releases(session->s_mdsc->snap_rwsem)
3113 struct ceph_inode_info *ci = ceph_inode(inode);
3114 int seq = le32_to_cpu(grant->seq);
3115 int newcaps = le32_to_cpu(grant->caps);
3116 int used, wanted, dirty;
3117 u64 size = le64_to_cpu(grant->size);
3118 u64 max_size = le64_to_cpu(grant->max_size);
3119 unsigned char check_caps = 0;
3120 bool was_stale = cap->cap_gen < session->s_cap_gen;
3122 bool writeback = false;
3123 bool queue_trunc = false;
3124 bool queue_invalidate = false;
3125 bool deleted_inode = false;
3126 bool fill_inline = false;
3128 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3129 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3130 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3135 * If CACHE is being revoked, and we have no dirty buffers,
3136 * try to invalidate (once). (If there are dirty buffers, we
3137 * will invalidate _after_ writeback.)
3139 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
3140 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3141 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3142 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3143 if (try_nonblocking_invalidate(inode)) {
3144 /* there were locked pages.. invalidate later
3145 in a separate thread. */
3146 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3147 queue_invalidate = true;
3148 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3154 cap->issued = cap->implemented = CEPH_CAP_PIN;
3157 * auth mds of the inode changed. we received the cap export message,
3158 * but still haven't received the cap import message. handle_cap_export
3159 * updated the new auth MDS' cap.
3161 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3162 * that was sent before the cap import message. So don't remove caps.
3164 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3165 WARN_ON(cap != ci->i_auth_cap);
3166 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3168 newcaps |= cap->issued;
3171 /* side effects now are allowed */
3172 cap->cap_gen = session->s_cap_gen;
3175 __check_cap_issue(ci, cap, newcaps);
3177 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3179 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3180 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3181 inode->i_mode = le32_to_cpu(grant->mode);
3182 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3183 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3184 ci->i_btime = extra_info->btime;
3185 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3186 from_kuid(&init_user_ns, inode->i_uid),
3187 from_kgid(&init_user_ns, inode->i_gid));
3190 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3191 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3192 set_nlink(inode, le32_to_cpu(grant->nlink));
3193 if (inode->i_nlink == 0 &&
3194 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3195 deleted_inode = true;
3198 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3200 int len = le32_to_cpu(grant->xattr_len);
3201 u64 version = le64_to_cpu(grant->xattr_version);
3203 if (version > ci->i_xattrs.version) {
3204 dout(" got new xattrs v%llu on %p len %d\n",
3205 version, inode, len);
3206 if (ci->i_xattrs.blob)
3207 ceph_buffer_put(ci->i_xattrs.blob);
3208 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3209 ci->i_xattrs.version = version;
3210 ceph_forget_all_cached_acls(inode);
3211 ceph_security_invalidate_secctx(inode);
3215 if (newcaps & CEPH_CAP_ANY_RD) {
3216 struct timespec64 mtime, atime, ctime;
3217 /* ctime/mtime/atime? */
3218 ceph_decode_timespec64(&mtime, &grant->mtime);
3219 ceph_decode_timespec64(&atime, &grant->atime);
3220 ceph_decode_timespec64(&ctime, &grant->ctime);
3221 ceph_fill_file_time(inode, extra_info->issued,
3222 le32_to_cpu(grant->time_warp_seq),
3223 &ctime, &mtime, &atime);
3226 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3227 ci->i_files = extra_info->nfiles;
3228 ci->i_subdirs = extra_info->nsubdirs;
3231 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3232 /* file layout may have changed */
3233 s64 old_pool = ci->i_layout.pool_id;
3234 struct ceph_string *old_ns;
3236 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3237 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3238 lockdep_is_held(&ci->i_ceph_lock));
3239 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3241 if (ci->i_layout.pool_id != old_pool ||
3242 extra_info->pool_ns != old_ns)
3243 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3245 extra_info->pool_ns = old_ns;
3247 /* size/truncate_seq? */
3248 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3249 le32_to_cpu(grant->truncate_seq),
3250 le64_to_cpu(grant->truncate_size),
3254 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3255 if (max_size != ci->i_max_size) {
3256 dout("max_size %lld -> %llu\n",
3257 ci->i_max_size, max_size);
3258 ci->i_max_size = max_size;
3259 if (max_size >= ci->i_wanted_max_size) {
3260 ci->i_wanted_max_size = 0; /* reset */
3261 ci->i_requested_max_size = 0;
3264 } else if (ci->i_wanted_max_size > ci->i_max_size &&
3265 ci->i_wanted_max_size > ci->i_requested_max_size) {
3266 /* CEPH_CAP_OP_IMPORT */
3271 /* check cap bits */
3272 wanted = __ceph_caps_wanted(ci);
3273 used = __ceph_caps_used(ci);
3274 dirty = __ceph_caps_dirty(ci);
3275 dout(" my wanted = %s, used = %s, dirty %s\n",
3276 ceph_cap_string(wanted),
3277 ceph_cap_string(used),
3278 ceph_cap_string(dirty));
3280 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3281 (wanted & ~(cap->mds_wanted | newcaps))) {
3283 * If mds is importing cap, prior cap messages that update
3284 * 'wanted' may get dropped by mds (migrate seq mismatch).
3286 * We don't send cap message to update 'wanted' if what we
3287 * want are already issued. If mds revokes caps, cap message
3288 * that releases caps also tells mds what we want. But if
3289 * caps got revoked by mds forcedly (session stale). We may
3290 * haven't told mds what we want.
3295 /* revocation, grant, or no-op? */
3296 if (cap->issued & ~newcaps) {
3297 int revoking = cap->issued & ~newcaps;
3299 dout("revocation: %s -> %s (revoking %s)\n",
3300 ceph_cap_string(cap->issued),
3301 ceph_cap_string(newcaps),
3302 ceph_cap_string(revoking));
3303 if (revoking & used & CEPH_CAP_FILE_BUFFER)
3304 writeback = true; /* initiate writeback; will delay ack */
3305 else if (revoking == CEPH_CAP_FILE_CACHE &&
3306 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3308 ; /* do nothing yet, invalidation will be queued */
3309 else if (cap == ci->i_auth_cap)
3310 check_caps = 1; /* check auth cap only */
3312 check_caps = 2; /* check all caps */
3313 cap->issued = newcaps;
3314 cap->implemented |= newcaps;
3315 } else if (cap->issued == newcaps) {
3316 dout("caps unchanged: %s -> %s\n",
3317 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3319 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3320 ceph_cap_string(newcaps));
3321 /* non-auth MDS is revoking the newly grant caps ? */
3322 if (cap == ci->i_auth_cap &&
3323 __ceph_caps_revoking_other(ci, cap, newcaps))
3326 cap->issued = newcaps;
3327 cap->implemented |= newcaps; /* add bits only, to
3328 * avoid stepping on a
3329 * pending revocation */
3332 BUG_ON(cap->issued & ~cap->implemented);
3334 if (extra_info->inline_version > 0 &&
3335 extra_info->inline_version >= ci->i_inline_version) {
3336 ci->i_inline_version = extra_info->inline_version;
3337 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3338 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3342 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3343 if (newcaps & ~extra_info->issued)
3345 kick_flushing_inode_caps(session->s_mdsc, session, inode);
3346 up_read(&session->s_mdsc->snap_rwsem);
3348 spin_unlock(&ci->i_ceph_lock);
3352 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3353 extra_info->inline_len);
3356 ceph_queue_vmtruncate(inode);
3360 * queue inode for writeback: we can't actually call
3361 * filemap_write_and_wait, etc. from message handler
3364 ceph_queue_writeback(inode);
3365 if (queue_invalidate)
3366 ceph_queue_invalidate(inode);
3368 invalidate_aliases(inode);
3370 wake_up_all(&ci->i_cap_wq);
3372 if (check_caps == 1)
3373 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
3375 else if (check_caps == 2)
3376 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
3378 mutex_unlock(&session->s_mutex);
3382 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3383 * MDS has been safely committed.
3385 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3386 struct ceph_mds_caps *m,
3387 struct ceph_mds_session *session,
3388 struct ceph_cap *cap)
3389 __releases(ci->i_ceph_lock)
3391 struct ceph_inode_info *ci = ceph_inode(inode);
3392 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3393 struct ceph_cap_flush *cf, *tmp_cf;
3394 LIST_HEAD(to_remove);
3395 unsigned seq = le32_to_cpu(m->seq);
3396 int dirty = le32_to_cpu(m->dirty);
3399 bool wake_ci = false;
3400 bool wake_mdsc = false;
3402 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3403 if (cf->tid == flush_tid)
3405 if (cf->caps == 0) /* capsnap */
3407 if (cf->tid <= flush_tid) {
3408 if (__finish_cap_flush(NULL, ci, cf))
3410 list_add_tail(&cf->i_list, &to_remove);
3412 cleaned &= ~cf->caps;
3418 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3419 " flushing %s -> %s\n",
3420 inode, session->s_mds, seq, ceph_cap_string(dirty),
3421 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3422 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3424 if (list_empty(&to_remove) && !cleaned)
3427 ci->i_flushing_caps &= ~cleaned;
3429 spin_lock(&mdsc->cap_dirty_lock);
3431 list_for_each_entry(cf, &to_remove, i_list) {
3432 if (__finish_cap_flush(mdsc, NULL, cf))
3436 if (ci->i_flushing_caps == 0) {
3437 if (list_empty(&ci->i_cap_flush_list)) {
3438 list_del_init(&ci->i_flushing_item);
3439 if (!list_empty(&session->s_cap_flushing)) {
3440 dout(" mds%d still flushing cap on %p\n",
3442 &list_first_entry(&session->s_cap_flushing,
3443 struct ceph_inode_info,
3444 i_flushing_item)->vfs_inode);
3447 mdsc->num_cap_flushing--;
3448 dout(" inode %p now !flushing\n", inode);
3450 if (ci->i_dirty_caps == 0) {
3451 dout(" inode %p now clean\n", inode);
3452 BUG_ON(!list_empty(&ci->i_dirty_item));
3454 if (ci->i_wr_ref == 0 &&
3455 ci->i_wrbuffer_ref_head == 0) {
3456 BUG_ON(!ci->i_head_snapc);
3457 ceph_put_snap_context(ci->i_head_snapc);
3458 ci->i_head_snapc = NULL;
3461 BUG_ON(list_empty(&ci->i_dirty_item));
3464 spin_unlock(&mdsc->cap_dirty_lock);
3467 spin_unlock(&ci->i_ceph_lock);
3469 while (!list_empty(&to_remove)) {
3470 cf = list_first_entry(&to_remove,
3471 struct ceph_cap_flush, i_list);
3472 list_del(&cf->i_list);
3473 ceph_free_cap_flush(cf);
3477 wake_up_all(&ci->i_cap_wq);
3479 wake_up_all(&mdsc->cap_flushing_wq);
3485 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3486 * throw away our cap_snap.
3488 * Caller hold s_mutex.
3490 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3491 struct ceph_mds_caps *m,
3492 struct ceph_mds_session *session)
3494 struct ceph_inode_info *ci = ceph_inode(inode);
3495 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3496 u64 follows = le64_to_cpu(m->snap_follows);
3497 struct ceph_cap_snap *capsnap;
3498 bool flushed = false;
3499 bool wake_ci = false;
3500 bool wake_mdsc = false;
3502 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3503 inode, ci, session->s_mds, follows);
3505 spin_lock(&ci->i_ceph_lock);
3506 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3507 if (capsnap->follows == follows) {
3508 if (capsnap->cap_flush.tid != flush_tid) {
3509 dout(" cap_snap %p follows %lld tid %lld !="
3510 " %lld\n", capsnap, follows,
3511 flush_tid, capsnap->cap_flush.tid);
3517 dout(" skipping cap_snap %p follows %lld\n",
3518 capsnap, capsnap->follows);
3522 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3523 dout(" removing %p cap_snap %p follows %lld\n",
3524 inode, capsnap, follows);
3525 list_del(&capsnap->ci_item);
3526 if (__finish_cap_flush(NULL, ci, &capsnap->cap_flush))
3529 spin_lock(&mdsc->cap_dirty_lock);
3531 if (list_empty(&ci->i_cap_flush_list))
3532 list_del_init(&ci->i_flushing_item);
3534 if (__finish_cap_flush(mdsc, NULL, &capsnap->cap_flush))
3537 spin_unlock(&mdsc->cap_dirty_lock);
3539 spin_unlock(&ci->i_ceph_lock);
3541 ceph_put_snap_context(capsnap->context);
3542 ceph_put_cap_snap(capsnap);
3544 wake_up_all(&ci->i_cap_wq);
3546 wake_up_all(&mdsc->cap_flushing_wq);
3552 * Handle TRUNC from MDS, indicating file truncation.
3554 * caller hold s_mutex.
3556 static void handle_cap_trunc(struct inode *inode,
3557 struct ceph_mds_caps *trunc,
3558 struct ceph_mds_session *session)
3559 __releases(ci->i_ceph_lock)
3561 struct ceph_inode_info *ci = ceph_inode(inode);
3562 int mds = session->s_mds;
3563 int seq = le32_to_cpu(trunc->seq);
3564 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3565 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3566 u64 size = le64_to_cpu(trunc->size);
3567 int implemented = 0;
3568 int dirty = __ceph_caps_dirty(ci);
3569 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3570 int queue_trunc = 0;
3572 issued |= implemented | dirty;
3574 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3575 inode, mds, seq, truncate_size, truncate_seq);
3576 queue_trunc = ceph_fill_file_size(inode, issued,
3577 truncate_seq, truncate_size, size);
3578 spin_unlock(&ci->i_ceph_lock);
3581 ceph_queue_vmtruncate(inode);
3585 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3586 * different one. If we are the most recent migration we've seen (as
3587 * indicated by mseq), make note of the migrating cap bits for the
3588 * duration (until we see the corresponding IMPORT).
3590 * caller holds s_mutex
3592 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3593 struct ceph_mds_cap_peer *ph,
3594 struct ceph_mds_session *session)
3596 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3597 struct ceph_mds_session *tsession = NULL;
3598 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3599 struct ceph_inode_info *ci = ceph_inode(inode);
3601 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3602 unsigned t_seq, t_mseq;
3604 int mds = session->s_mds;
3607 t_cap_id = le64_to_cpu(ph->cap_id);
3608 t_seq = le32_to_cpu(ph->seq);
3609 t_mseq = le32_to_cpu(ph->mseq);
3610 target = le32_to_cpu(ph->mds);
3612 t_cap_id = t_seq = t_mseq = 0;
3616 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3617 inode, ci, mds, mseq, target);
3619 spin_lock(&ci->i_ceph_lock);
3620 cap = __get_cap_for_mds(ci, mds);
3621 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3625 if (cap->mds_wanted | cap->issued)
3626 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
3627 __ceph_remove_cap(cap, false);
3632 * now we know we haven't received the cap import message yet
3633 * because the exported cap still exist.
3636 issued = cap->issued;
3637 if (issued != cap->implemented)
3638 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3639 "ino (%llx.%llx) mds%d seq %d mseq %d "
3640 "issued %s implemented %s\n",
3641 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3642 ceph_cap_string(issued),
3643 ceph_cap_string(cap->implemented));
3646 tcap = __get_cap_for_mds(ci, target);
3648 /* already have caps from the target */
3649 if (tcap->cap_id == t_cap_id &&
3650 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3651 dout(" updating import cap %p mds%d\n", tcap, target);
3652 tcap->cap_id = t_cap_id;
3653 tcap->seq = t_seq - 1;
3654 tcap->issue_seq = t_seq - 1;
3655 tcap->issued |= issued;
3656 tcap->implemented |= issued;
3657 if (cap == ci->i_auth_cap)
3658 ci->i_auth_cap = tcap;
3660 if (!list_empty(&ci->i_cap_flush_list) &&
3661 ci->i_auth_cap == tcap) {
3662 spin_lock(&mdsc->cap_dirty_lock);
3663 list_move_tail(&ci->i_flushing_item,
3664 &tcap->session->s_cap_flushing);
3665 spin_unlock(&mdsc->cap_dirty_lock);
3668 __ceph_remove_cap(cap, false);
3670 } else if (tsession) {
3671 /* add placeholder for the export tagert */
3672 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3674 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3675 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3677 if (!list_empty(&ci->i_cap_flush_list) &&
3678 ci->i_auth_cap == tcap) {
3679 spin_lock(&mdsc->cap_dirty_lock);
3680 list_move_tail(&ci->i_flushing_item,
3681 &tcap->session->s_cap_flushing);
3682 spin_unlock(&mdsc->cap_dirty_lock);
3685 __ceph_remove_cap(cap, false);
3689 spin_unlock(&ci->i_ceph_lock);
3690 mutex_unlock(&session->s_mutex);
3692 /* open target session */
3693 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3694 if (!IS_ERR(tsession)) {
3696 mutex_lock(&session->s_mutex);
3697 mutex_lock_nested(&tsession->s_mutex,
3698 SINGLE_DEPTH_NESTING);
3700 mutex_lock(&tsession->s_mutex);
3701 mutex_lock_nested(&session->s_mutex,
3702 SINGLE_DEPTH_NESTING);
3704 new_cap = ceph_get_cap(mdsc, NULL);
3713 spin_unlock(&ci->i_ceph_lock);
3714 mutex_unlock(&session->s_mutex);
3716 mutex_unlock(&tsession->s_mutex);
3717 ceph_put_mds_session(tsession);
3720 ceph_put_cap(mdsc, new_cap);
3724 * Handle cap IMPORT.
3726 * caller holds s_mutex. acquires i_ceph_lock
3728 static void handle_cap_import(struct ceph_mds_client *mdsc,
3729 struct inode *inode, struct ceph_mds_caps *im,
3730 struct ceph_mds_cap_peer *ph,
3731 struct ceph_mds_session *session,
3732 struct ceph_cap **target_cap, int *old_issued)
3733 __acquires(ci->i_ceph_lock)
3735 struct ceph_inode_info *ci = ceph_inode(inode);
3736 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3737 int mds = session->s_mds;
3739 unsigned caps = le32_to_cpu(im->caps);
3740 unsigned wanted = le32_to_cpu(im->wanted);
3741 unsigned seq = le32_to_cpu(im->seq);
3742 unsigned mseq = le32_to_cpu(im->migrate_seq);
3743 u64 realmino = le64_to_cpu(im->realm);
3744 u64 cap_id = le64_to_cpu(im->cap_id);
3749 p_cap_id = le64_to_cpu(ph->cap_id);
3750 peer = le32_to_cpu(ph->mds);
3756 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3757 inode, ci, mds, mseq, peer);
3760 spin_lock(&ci->i_ceph_lock);
3761 cap = __get_cap_for_mds(ci, mds);
3764 spin_unlock(&ci->i_ceph_lock);
3765 new_cap = ceph_get_cap(mdsc, NULL);
3771 ceph_put_cap(mdsc, new_cap);
3776 __ceph_caps_issued(ci, &issued);
3777 issued |= __ceph_caps_dirty(ci);
3779 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3780 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3782 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3783 if (ocap && ocap->cap_id == p_cap_id) {
3784 dout(" remove export cap %p mds%d flags %d\n",
3785 ocap, peer, ph->flags);
3786 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3787 (ocap->seq != le32_to_cpu(ph->seq) ||
3788 ocap->mseq != le32_to_cpu(ph->mseq))) {
3789 pr_err_ratelimited("handle_cap_import: "
3790 "mismatched seq/mseq: ino (%llx.%llx) "
3791 "mds%d seq %d mseq %d importer mds%d "
3792 "has peer seq %d mseq %d\n",
3793 ceph_vinop(inode), peer, ocap->seq,
3794 ocap->mseq, mds, le32_to_cpu(ph->seq),
3795 le32_to_cpu(ph->mseq));
3797 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3800 /* make sure we re-request max_size, if necessary */
3801 ci->i_requested_max_size = 0;
3803 *old_issued = issued;
3808 * Handle a caps message from the MDS.
3810 * Identify the appropriate session, inode, and call the right handler
3811 * based on the cap op.
3813 void ceph_handle_caps(struct ceph_mds_session *session,
3814 struct ceph_msg *msg)
3816 struct ceph_mds_client *mdsc = session->s_mdsc;
3817 struct inode *inode;
3818 struct ceph_inode_info *ci;
3819 struct ceph_cap *cap;
3820 struct ceph_mds_caps *h;
3821 struct ceph_mds_cap_peer *peer = NULL;
3822 struct ceph_snap_realm *realm = NULL;
3824 int msg_version = le16_to_cpu(msg->hdr.version);
3826 struct ceph_vino vino;
3828 size_t snaptrace_len;
3830 struct cap_extra_info extra_info = {};
3832 dout("handle_caps from mds%d\n", session->s_mds);
3835 end = msg->front.iov_base + msg->front.iov_len;
3836 if (msg->front.iov_len < sizeof(*h))
3838 h = msg->front.iov_base;
3839 op = le32_to_cpu(h->op);
3840 vino.ino = le64_to_cpu(h->ino);
3841 vino.snap = CEPH_NOSNAP;
3842 seq = le32_to_cpu(h->seq);
3843 mseq = le32_to_cpu(h->migrate_seq);
3846 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3847 p = snaptrace + snaptrace_len;
3849 if (msg_version >= 2) {
3851 ceph_decode_32_safe(&p, end, flock_len, bad);
3852 if (p + flock_len > end)
3857 if (msg_version >= 3) {
3858 if (op == CEPH_CAP_OP_IMPORT) {
3859 if (p + sizeof(*peer) > end)
3863 } else if (op == CEPH_CAP_OP_EXPORT) {
3864 /* recorded in unused fields */
3865 peer = (void *)&h->size;
3869 if (msg_version >= 4) {
3870 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
3871 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
3872 if (p + extra_info.inline_len > end)
3874 extra_info.inline_data = p;
3875 p += extra_info.inline_len;
3878 if (msg_version >= 5) {
3879 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
3882 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
3883 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
3886 if (msg_version >= 8) {
3888 u32 caller_uid, caller_gid;
3892 ceph_decode_64_safe(&p, end, flush_tid, bad);
3894 ceph_decode_32_safe(&p, end, caller_uid, bad);
3895 ceph_decode_32_safe(&p, end, caller_gid, bad);
3897 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
3898 if (pool_ns_len > 0) {
3899 ceph_decode_need(&p, end, pool_ns_len, bad);
3900 extra_info.pool_ns =
3901 ceph_find_or_create_string(p, pool_ns_len);
3906 if (msg_version >= 9) {
3907 struct ceph_timespec *btime;
3909 if (p + sizeof(*btime) > end)
3912 ceph_decode_timespec64(&extra_info.btime, btime);
3913 p += sizeof(*btime);
3914 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
3917 if (msg_version >= 11) {
3920 ceph_decode_32_safe(&p, end, flags, bad);
3922 extra_info.dirstat_valid = true;
3923 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
3924 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
3928 inode = ceph_find_inode(mdsc->fsc->sb, vino);
3929 ci = ceph_inode(inode);
3930 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3933 mutex_lock(&session->s_mutex);
3935 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3939 dout(" i don't have ino %llx\n", vino.ino);
3941 if (op == CEPH_CAP_OP_IMPORT) {
3942 cap = ceph_get_cap(mdsc, NULL);
3943 cap->cap_ino = vino.ino;
3944 cap->queue_release = 1;
3945 cap->cap_id = le64_to_cpu(h->cap_id);
3948 cap->issue_seq = seq;
3949 spin_lock(&session->s_cap_lock);
3950 __ceph_queue_cap_release(session, cap);
3951 spin_unlock(&session->s_cap_lock);
3956 /* these will work even if we don't have a cap yet */
3958 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3959 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
3963 case CEPH_CAP_OP_EXPORT:
3964 handle_cap_export(inode, h, peer, session);
3967 case CEPH_CAP_OP_IMPORT:
3969 if (snaptrace_len) {
3970 down_write(&mdsc->snap_rwsem);
3971 ceph_update_snap_trace(mdsc, snaptrace,
3972 snaptrace + snaptrace_len,
3974 downgrade_write(&mdsc->snap_rwsem);
3976 down_read(&mdsc->snap_rwsem);
3978 handle_cap_import(mdsc, inode, h, peer, session,
3979 &cap, &extra_info.issued);
3980 handle_cap_grant(inode, session, cap,
3981 h, msg->middle, &extra_info);
3983 ceph_put_snap_realm(mdsc, realm);
3987 /* the rest require a cap */
3988 spin_lock(&ci->i_ceph_lock);
3989 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
3991 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3992 inode, ceph_ino(inode), ceph_snap(inode),
3994 spin_unlock(&ci->i_ceph_lock);
3995 goto flush_cap_releases;
3998 /* note that each of these drops i_ceph_lock for us */
4000 case CEPH_CAP_OP_REVOKE:
4001 case CEPH_CAP_OP_GRANT:
4002 __ceph_caps_issued(ci, &extra_info.issued);
4003 extra_info.issued |= __ceph_caps_dirty(ci);
4004 handle_cap_grant(inode, session, cap,
4005 h, msg->middle, &extra_info);
4008 case CEPH_CAP_OP_FLUSH_ACK:
4009 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4013 case CEPH_CAP_OP_TRUNC:
4014 handle_cap_trunc(inode, h, session);
4018 spin_unlock(&ci->i_ceph_lock);
4019 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4020 ceph_cap_op_name(op));
4024 mutex_unlock(&session->s_mutex);
4026 ceph_put_string(extra_info.pool_ns);
4027 /* avoid calling iput_final() in mds dispatch threads */
4028 ceph_async_iput(inode);
4033 * send any cap release message to try to move things
4034 * along for the mds (who clearly thinks we still have this
4037 ceph_flush_cap_releases(mdsc, session);
4041 pr_err("ceph_handle_caps: corrupt message\n");
4047 * Delayed work handler to process end of delayed cap release LRU list.
4049 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4051 struct inode *inode;
4052 struct ceph_inode_info *ci;
4053 int flags = CHECK_CAPS_NODELAY;
4055 dout("check_delayed_caps\n");
4057 spin_lock(&mdsc->cap_delay_lock);
4058 if (list_empty(&mdsc->cap_delay_list))
4060 ci = list_first_entry(&mdsc->cap_delay_list,
4061 struct ceph_inode_info,
4063 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4064 time_before(jiffies, ci->i_hold_caps_max))
4066 list_del_init(&ci->i_cap_delay_list);
4068 inode = igrab(&ci->vfs_inode);
4069 spin_unlock(&mdsc->cap_delay_lock);
4072 dout("check_delayed_caps on %p\n", inode);
4073 ceph_check_caps(ci, flags, NULL);
4074 /* avoid calling iput_final() in tick thread */
4075 ceph_async_iput(inode);
4078 spin_unlock(&mdsc->cap_delay_lock);
4082 * Flush all dirty caps to the mds
4084 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4086 struct ceph_inode_info *ci;
4087 struct inode *inode;
4089 dout("flush_dirty_caps\n");
4090 spin_lock(&mdsc->cap_dirty_lock);
4091 while (!list_empty(&mdsc->cap_dirty)) {
4092 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
4094 inode = &ci->vfs_inode;
4096 dout("flush_dirty_caps %p\n", inode);
4097 spin_unlock(&mdsc->cap_dirty_lock);
4098 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
4100 spin_lock(&mdsc->cap_dirty_lock);
4102 spin_unlock(&mdsc->cap_dirty_lock);
4103 dout("flush_dirty_caps done\n");
4106 void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode)
4109 int bits = (fmode << 1) | 1;
4110 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4111 if (bits & (1 << i))
4112 ci->i_nr_by_mode[i]++;
4117 * Drop open file reference. If we were the last open file,
4118 * we may need to release capabilities to the MDS (or schedule
4119 * their delayed release).
4121 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
4124 int bits = (fmode << 1) | 1;
4125 spin_lock(&ci->i_ceph_lock);
4126 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4127 if (bits & (1 << i)) {
4128 BUG_ON(ci->i_nr_by_mode[i] == 0);
4129 if (--ci->i_nr_by_mode[i] == 0)
4133 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
4134 &ci->vfs_inode, fmode,
4135 ci->i_nr_by_mode[0], ci->i_nr_by_mode[1],
4136 ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]);
4137 spin_unlock(&ci->i_ceph_lock);
4139 if (last && ci->i_vino.snap == CEPH_NOSNAP)
4140 ceph_check_caps(ci, 0, NULL);
4144 * For a soon-to-be unlinked file, drop the LINK caps. If it
4145 * looks like the link count will hit 0, drop any other caps (other
4146 * than PIN) we don't specifically want (due to the file still being
4149 int ceph_drop_caps_for_unlink(struct inode *inode)
4151 struct ceph_inode_info *ci = ceph_inode(inode);
4152 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4154 spin_lock(&ci->i_ceph_lock);
4155 if (inode->i_nlink == 1) {
4156 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4158 ci->i_ceph_flags |= CEPH_I_NODELAY;
4159 if (__ceph_caps_dirty(ci)) {
4160 struct ceph_mds_client *mdsc =
4161 ceph_inode_to_client(inode)->mdsc;
4162 __cap_delay_requeue_front(mdsc, ci);
4165 spin_unlock(&ci->i_ceph_lock);
4170 * Helpers for embedding cap and dentry lease releases into mds
4173 * @force is used by dentry_release (below) to force inclusion of a
4174 * record for the directory inode, even when there aren't any caps to
4177 int ceph_encode_inode_release(void **p, struct inode *inode,
4178 int mds, int drop, int unless, int force)
4180 struct ceph_inode_info *ci = ceph_inode(inode);
4181 struct ceph_cap *cap;
4182 struct ceph_mds_request_release *rel = *p;
4186 spin_lock(&ci->i_ceph_lock);
4187 used = __ceph_caps_used(ci);
4188 dirty = __ceph_caps_dirty(ci);
4190 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4191 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4192 ceph_cap_string(unless));
4194 /* only drop unused, clean caps */
4195 drop &= ~(used | dirty);
4197 cap = __get_cap_for_mds(ci, mds);
4198 if (cap && __cap_is_valid(cap)) {
4199 unless &= cap->issued;
4201 if (unless & CEPH_CAP_AUTH_EXCL)
4202 drop &= ~CEPH_CAP_AUTH_SHARED;
4203 if (unless & CEPH_CAP_LINK_EXCL)
4204 drop &= ~CEPH_CAP_LINK_SHARED;
4205 if (unless & CEPH_CAP_XATTR_EXCL)
4206 drop &= ~CEPH_CAP_XATTR_SHARED;
4207 if (unless & CEPH_CAP_FILE_EXCL)
4208 drop &= ~CEPH_CAP_FILE_SHARED;
4211 if (force || (cap->issued & drop)) {
4212 if (cap->issued & drop) {
4213 int wanted = __ceph_caps_wanted(ci);
4214 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
4215 wanted |= cap->mds_wanted;
4216 dout("encode_inode_release %p cap %p "
4217 "%s -> %s, wanted %s -> %s\n", inode, cap,
4218 ceph_cap_string(cap->issued),
4219 ceph_cap_string(cap->issued & ~drop),
4220 ceph_cap_string(cap->mds_wanted),
4221 ceph_cap_string(wanted));
4223 cap->issued &= ~drop;
4224 cap->implemented &= ~drop;
4225 cap->mds_wanted = wanted;
4227 dout("encode_inode_release %p cap %p %s"
4228 " (force)\n", inode, cap,
4229 ceph_cap_string(cap->issued));
4232 rel->ino = cpu_to_le64(ceph_ino(inode));
4233 rel->cap_id = cpu_to_le64(cap->cap_id);
4234 rel->seq = cpu_to_le32(cap->seq);
4235 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4236 rel->mseq = cpu_to_le32(cap->mseq);
4237 rel->caps = cpu_to_le32(cap->implemented);
4238 rel->wanted = cpu_to_le32(cap->mds_wanted);
4244 dout("encode_inode_release %p cap %p %s (noop)\n",
4245 inode, cap, ceph_cap_string(cap->issued));
4248 spin_unlock(&ci->i_ceph_lock);
4252 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4254 int mds, int drop, int unless)
4256 struct dentry *parent = NULL;
4257 struct ceph_mds_request_release *rel = *p;
4258 struct ceph_dentry_info *di = ceph_dentry(dentry);
4263 * force an record for the directory caps if we have a dentry lease.
4264 * this is racy (can't take i_ceph_lock and d_lock together), but it
4265 * doesn't have to be perfect; the mds will revoke anything we don't
4268 spin_lock(&dentry->d_lock);
4269 if (di->lease_session && di->lease_session->s_mds == mds)
4272 parent = dget(dentry->d_parent);
4273 dir = d_inode(parent);
4275 spin_unlock(&dentry->d_lock);
4277 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4280 spin_lock(&dentry->d_lock);
4281 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4282 dout("encode_dentry_release %p mds%d seq %d\n",
4283 dentry, mds, (int)di->lease_seq);
4284 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4285 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4286 *p += dentry->d_name.len;
4287 rel->dname_seq = cpu_to_le32(di->lease_seq);
4288 __ceph_mdsc_drop_dentry_lease(dentry);
4290 spin_unlock(&dentry->d_lock);