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;
1058 /* 'ci' being NULL means the remove have already occurred */
1060 dout("%s: cap inode is NULL\n", __func__);
1064 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1066 mdsc = ceph_inode_to_client(&ci->vfs_inode)->mdsc;
1068 /* remove from inode's cap rbtree, and clear auth cap */
1069 rb_erase(&cap->ci_node, &ci->i_caps);
1070 if (ci->i_auth_cap == cap)
1071 ci->i_auth_cap = NULL;
1073 /* remove from session list */
1074 spin_lock(&session->s_cap_lock);
1075 if (session->s_cap_iterator == cap) {
1076 /* not yet, we are iterating over this very cap */
1077 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1080 list_del_init(&cap->session_caps);
1081 session->s_nr_caps--;
1082 cap->session = NULL;
1085 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1089 * s_cap_reconnect is protected by s_cap_lock. no one changes
1090 * s_cap_gen while session is in the reconnect state.
1092 if (queue_release &&
1093 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
1094 cap->queue_release = 1;
1096 __ceph_queue_cap_release(session, cap);
1100 cap->queue_release = 0;
1102 cap->cap_ino = ci->i_vino.ino;
1104 spin_unlock(&session->s_cap_lock);
1107 ceph_put_cap(mdsc, cap);
1109 /* when reconnect denied, we remove session caps forcibly,
1110 * i_wr_ref can be non-zero. If there are ongoing write,
1111 * keep i_snap_realm.
1113 if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm)
1114 drop_inode_snap_realm(ci);
1116 if (!__ceph_is_any_real_caps(ci))
1117 __cap_delay_cancel(mdsc, ci);
1120 struct cap_msg_args {
1121 struct ceph_mds_session *session;
1122 u64 ino, cid, follows;
1123 u64 flush_tid, oldest_flush_tid, size, max_size;
1126 struct ceph_buffer *xattr_buf;
1127 struct timespec64 atime, mtime, ctime, btime;
1128 int op, caps, wanted, dirty;
1129 u32 seq, issue_seq, mseq, time_warp_seq;
1138 * Build and send a cap message to the given MDS.
1140 * Caller should be holding s_mutex.
1142 static int send_cap_msg(struct cap_msg_args *arg)
1144 struct ceph_mds_caps *fc;
1145 struct ceph_msg *msg;
1148 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1150 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1151 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1152 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(arg->op),
1153 arg->cid, arg->ino, ceph_cap_string(arg->caps),
1154 ceph_cap_string(arg->wanted), ceph_cap_string(arg->dirty),
1155 arg->seq, arg->issue_seq, arg->flush_tid, arg->oldest_flush_tid,
1156 arg->mseq, arg->follows, arg->size, arg->max_size,
1158 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1160 /* flock buffer size + inline version + inline data size +
1161 * osd_epoch_barrier + oldest_flush_tid */
1162 extra_len = 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4;
1163 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1168 msg->hdr.version = cpu_to_le16(10);
1169 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1171 fc = msg->front.iov_base;
1172 memset(fc, 0, sizeof(*fc));
1174 fc->cap_id = cpu_to_le64(arg->cid);
1175 fc->op = cpu_to_le32(arg->op);
1176 fc->seq = cpu_to_le32(arg->seq);
1177 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1178 fc->migrate_seq = cpu_to_le32(arg->mseq);
1179 fc->caps = cpu_to_le32(arg->caps);
1180 fc->wanted = cpu_to_le32(arg->wanted);
1181 fc->dirty = cpu_to_le32(arg->dirty);
1182 fc->ino = cpu_to_le64(arg->ino);
1183 fc->snap_follows = cpu_to_le64(arg->follows);
1185 fc->size = cpu_to_le64(arg->size);
1186 fc->max_size = cpu_to_le64(arg->max_size);
1187 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1188 ceph_encode_timespec64(&fc->atime, &arg->atime);
1189 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1190 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1192 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1193 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1194 fc->mode = cpu_to_le32(arg->mode);
1196 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1197 if (arg->xattr_buf) {
1198 msg->middle = ceph_buffer_get(arg->xattr_buf);
1199 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1200 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1204 /* flock buffer size (version 2) */
1205 ceph_encode_32(&p, 0);
1206 /* inline version (version 4) */
1207 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1208 /* inline data size */
1209 ceph_encode_32(&p, 0);
1211 * osd_epoch_barrier (version 5)
1212 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1213 * case it was recently changed
1215 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1216 /* oldest_flush_tid (version 6) */
1217 ceph_encode_64(&p, arg->oldest_flush_tid);
1220 * caller_uid/caller_gid (version 7)
1222 * Currently, we don't properly track which caller dirtied the caps
1223 * last, and force a flush of them when there is a conflict. For now,
1224 * just set this to 0:0, to emulate how the MDS has worked up to now.
1226 ceph_encode_32(&p, 0);
1227 ceph_encode_32(&p, 0);
1229 /* pool namespace (version 8) (mds always ignores this) */
1230 ceph_encode_32(&p, 0);
1232 /* btime and change_attr (version 9) */
1233 ceph_encode_timespec64(p, &arg->btime);
1234 p += sizeof(struct ceph_timespec);
1235 ceph_encode_64(&p, arg->change_attr);
1237 /* Advisory flags (version 10) */
1238 ceph_encode_32(&p, arg->flags);
1240 ceph_con_send(&arg->session->s_con, msg);
1245 * Queue cap releases when an inode is dropped from our cache.
1247 void __ceph_remove_caps(struct ceph_inode_info *ci)
1251 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1252 * may call __ceph_caps_issued_mask() on a freeing inode. */
1253 spin_lock(&ci->i_ceph_lock);
1254 p = rb_first(&ci->i_caps);
1256 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1258 __ceph_remove_cap(cap, true);
1260 spin_unlock(&ci->i_ceph_lock);
1264 * Send a cap msg on the given inode. Update our caps state, then
1265 * drop i_ceph_lock and send the message.
1267 * Make note of max_size reported/requested from mds, revoked caps
1268 * that have now been implemented.
1270 * Return non-zero if delayed release, or we experienced an error
1271 * such that the caller should requeue + retry later.
1273 * called with i_ceph_lock, then drops it.
1274 * caller should hold snap_rwsem (read), s_mutex.
1276 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1277 int op, int flags, int used, int want, int retain,
1278 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1279 __releases(cap->ci->i_ceph_lock)
1281 struct ceph_inode_info *ci = cap->ci;
1282 struct inode *inode = &ci->vfs_inode;
1283 struct ceph_buffer *old_blob = NULL;
1284 struct cap_msg_args arg;
1290 held = cap->issued | cap->implemented;
1291 revoking = cap->implemented & ~cap->issued;
1292 retain &= ~revoking;
1294 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1295 inode, cap, cap->session,
1296 ceph_cap_string(held), ceph_cap_string(held & retain),
1297 ceph_cap_string(revoking));
1298 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1300 arg.session = cap->session;
1302 /* don't release wanted unless we've waited a bit. */
1303 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1304 time_before(jiffies, ci->i_hold_caps_min)) {
1305 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1306 ceph_cap_string(cap->issued),
1307 ceph_cap_string(cap->issued & retain),
1308 ceph_cap_string(cap->mds_wanted),
1309 ceph_cap_string(want));
1310 want |= cap->mds_wanted;
1311 retain |= cap->issued;
1314 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1315 if (want & ~cap->mds_wanted) {
1316 /* user space may open/close single file frequently.
1317 * This avoids droping mds_wanted immediately after
1318 * requesting new mds_wanted.
1320 __cap_set_timeouts(mdsc, ci);
1323 cap->issued &= retain; /* drop bits we don't want */
1324 if (cap->implemented & ~cap->issued) {
1326 * Wake up any waiters on wanted -> needed transition.
1327 * This is due to the weird transition from buffered
1328 * to sync IO... we need to flush dirty pages _before_
1329 * allowing sync writes to avoid reordering.
1333 cap->implemented &= cap->issued | used;
1334 cap->mds_wanted = want;
1336 arg.ino = ceph_vino(inode).ino;
1337 arg.cid = cap->cap_id;
1338 arg.follows = flushing ? ci->i_head_snapc->seq : 0;
1339 arg.flush_tid = flush_tid;
1340 arg.oldest_flush_tid = oldest_flush_tid;
1342 arg.size = inode->i_size;
1343 ci->i_reported_size = arg.size;
1344 arg.max_size = ci->i_wanted_max_size;
1345 ci->i_requested_max_size = arg.max_size;
1347 if (flushing & CEPH_CAP_XATTR_EXCL) {
1348 old_blob = __ceph_build_xattrs_blob(ci);
1349 arg.xattr_version = ci->i_xattrs.version;
1350 arg.xattr_buf = ci->i_xattrs.blob;
1352 arg.xattr_buf = NULL;
1355 arg.mtime = inode->i_mtime;
1356 arg.atime = inode->i_atime;
1357 arg.ctime = inode->i_ctime;
1358 arg.btime = ci->i_btime;
1359 arg.change_attr = inode_peek_iversion_raw(inode);
1362 arg.caps = cap->implemented;
1364 arg.dirty = flushing;
1367 arg.issue_seq = cap->issue_seq;
1368 arg.mseq = cap->mseq;
1369 arg.time_warp_seq = ci->i_time_warp_seq;
1371 arg.uid = inode->i_uid;
1372 arg.gid = inode->i_gid;
1373 arg.mode = inode->i_mode;
1375 arg.inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1376 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1377 !list_empty(&ci->i_cap_snaps)) {
1378 struct ceph_cap_snap *capsnap;
1379 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1380 if (capsnap->cap_flush.tid)
1382 if (capsnap->need_flush) {
1383 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1390 spin_unlock(&ci->i_ceph_lock);
1392 ceph_buffer_put(old_blob);
1394 ret = send_cap_msg(&arg);
1396 dout("error sending cap msg, must requeue %p\n", inode);
1401 wake_up_all(&ci->i_cap_wq);
1406 static inline int __send_flush_snap(struct inode *inode,
1407 struct ceph_mds_session *session,
1408 struct ceph_cap_snap *capsnap,
1409 u32 mseq, u64 oldest_flush_tid)
1411 struct cap_msg_args arg;
1413 arg.session = session;
1414 arg.ino = ceph_vino(inode).ino;
1416 arg.follows = capsnap->follows;
1417 arg.flush_tid = capsnap->cap_flush.tid;
1418 arg.oldest_flush_tid = oldest_flush_tid;
1420 arg.size = capsnap->size;
1422 arg.xattr_version = capsnap->xattr_version;
1423 arg.xattr_buf = capsnap->xattr_blob;
1425 arg.atime = capsnap->atime;
1426 arg.mtime = capsnap->mtime;
1427 arg.ctime = capsnap->ctime;
1428 arg.btime = capsnap->btime;
1429 arg.change_attr = capsnap->change_attr;
1431 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1432 arg.caps = capsnap->issued;
1434 arg.dirty = capsnap->dirty;
1439 arg.time_warp_seq = capsnap->time_warp_seq;
1441 arg.uid = capsnap->uid;
1442 arg.gid = capsnap->gid;
1443 arg.mode = capsnap->mode;
1445 arg.inline_data = capsnap->inline_data;
1448 return send_cap_msg(&arg);
1452 * When a snapshot is taken, clients accumulate dirty metadata on
1453 * inodes with capabilities in ceph_cap_snaps to describe the file
1454 * state at the time the snapshot was taken. This must be flushed
1455 * asynchronously back to the MDS once sync writes complete and dirty
1456 * data is written out.
1458 * Called under i_ceph_lock. Takes s_mutex as needed.
1460 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1461 struct ceph_mds_session *session)
1462 __releases(ci->i_ceph_lock)
1463 __acquires(ci->i_ceph_lock)
1465 struct inode *inode = &ci->vfs_inode;
1466 struct ceph_mds_client *mdsc = session->s_mdsc;
1467 struct ceph_cap_snap *capsnap;
1468 u64 oldest_flush_tid = 0;
1469 u64 first_tid = 1, last_tid = 0;
1471 dout("__flush_snaps %p session %p\n", inode, session);
1473 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1475 * we need to wait for sync writes to complete and for dirty
1476 * pages to be written out.
1478 if (capsnap->dirty_pages || capsnap->writing)
1481 /* should be removed by ceph_try_drop_cap_snap() */
1482 BUG_ON(!capsnap->need_flush);
1484 /* only flush each capsnap once */
1485 if (capsnap->cap_flush.tid > 0) {
1486 dout(" already flushed %p, skipping\n", capsnap);
1490 spin_lock(&mdsc->cap_dirty_lock);
1491 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1492 list_add_tail(&capsnap->cap_flush.g_list,
1493 &mdsc->cap_flush_list);
1494 if (oldest_flush_tid == 0)
1495 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1496 if (list_empty(&ci->i_flushing_item)) {
1497 list_add_tail(&ci->i_flushing_item,
1498 &session->s_cap_flushing);
1500 spin_unlock(&mdsc->cap_dirty_lock);
1502 list_add_tail(&capsnap->cap_flush.i_list,
1503 &ci->i_cap_flush_list);
1506 first_tid = capsnap->cap_flush.tid;
1507 last_tid = capsnap->cap_flush.tid;
1510 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1512 while (first_tid <= last_tid) {
1513 struct ceph_cap *cap = ci->i_auth_cap;
1514 struct ceph_cap_flush *cf;
1517 if (!(cap && cap->session == session)) {
1518 dout("__flush_snaps %p auth cap %p not mds%d, "
1519 "stop\n", inode, cap, session->s_mds);
1524 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1525 if (cf->tid >= first_tid) {
1533 first_tid = cf->tid + 1;
1535 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1536 refcount_inc(&capsnap->nref);
1537 spin_unlock(&ci->i_ceph_lock);
1539 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1540 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1542 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1545 pr_err("__flush_snaps: error sending cap flushsnap, "
1546 "ino (%llx.%llx) tid %llu follows %llu\n",
1547 ceph_vinop(inode), cf->tid, capsnap->follows);
1550 ceph_put_cap_snap(capsnap);
1551 spin_lock(&ci->i_ceph_lock);
1555 void ceph_flush_snaps(struct ceph_inode_info *ci,
1556 struct ceph_mds_session **psession)
1558 struct inode *inode = &ci->vfs_inode;
1559 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1560 struct ceph_mds_session *session = NULL;
1561 bool need_put = false;
1564 dout("ceph_flush_snaps %p\n", inode);
1566 session = *psession;
1568 spin_lock(&ci->i_ceph_lock);
1569 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1570 dout(" no capsnap needs flush, doing nothing\n");
1573 if (!ci->i_auth_cap) {
1574 dout(" no auth cap (migrating?), doing nothing\n");
1578 mds = ci->i_auth_cap->session->s_mds;
1579 if (session && session->s_mds != mds) {
1580 dout(" oops, wrong session %p mutex\n", session);
1581 mutex_unlock(&session->s_mutex);
1582 ceph_put_mds_session(session);
1586 spin_unlock(&ci->i_ceph_lock);
1587 mutex_lock(&mdsc->mutex);
1588 session = __ceph_lookup_mds_session(mdsc, mds);
1589 mutex_unlock(&mdsc->mutex);
1591 dout(" inverting session/ino locks on %p\n", session);
1592 mutex_lock(&session->s_mutex);
1597 // make sure flushsnap messages are sent in proper order.
1598 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1599 __kick_flushing_caps(mdsc, session, ci, 0);
1601 __ceph_flush_snaps(ci, session);
1603 spin_unlock(&ci->i_ceph_lock);
1606 *psession = session;
1607 } else if (session) {
1608 mutex_unlock(&session->s_mutex);
1609 ceph_put_mds_session(session);
1611 /* we flushed them all; remove this inode from the queue */
1612 spin_lock(&mdsc->snap_flush_lock);
1613 if (!list_empty(&ci->i_snap_flush_item))
1615 list_del_init(&ci->i_snap_flush_item);
1616 spin_unlock(&mdsc->snap_flush_lock);
1623 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1624 * Caller is then responsible for calling __mark_inode_dirty with the
1625 * returned flags value.
1627 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1628 struct ceph_cap_flush **pcf)
1630 struct ceph_mds_client *mdsc =
1631 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1632 struct inode *inode = &ci->vfs_inode;
1633 int was = ci->i_dirty_caps;
1636 if (!ci->i_auth_cap) {
1637 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1638 "but no auth cap (session was closed?)\n",
1639 inode, ceph_ino(inode), ceph_cap_string(mask));
1643 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1644 ceph_cap_string(mask), ceph_cap_string(was),
1645 ceph_cap_string(was | mask));
1646 ci->i_dirty_caps |= mask;
1648 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1649 swap(ci->i_prealloc_cap_flush, *pcf);
1651 if (!ci->i_head_snapc) {
1652 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1653 ci->i_head_snapc = ceph_get_snap_context(
1654 ci->i_snap_realm->cached_context);
1656 dout(" inode %p now dirty snapc %p auth cap %p\n",
1657 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1658 BUG_ON(!list_empty(&ci->i_dirty_item));
1659 spin_lock(&mdsc->cap_dirty_lock);
1660 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1661 spin_unlock(&mdsc->cap_dirty_lock);
1662 if (ci->i_flushing_caps == 0) {
1664 dirty |= I_DIRTY_SYNC;
1667 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1669 BUG_ON(list_empty(&ci->i_dirty_item));
1670 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1671 (mask & CEPH_CAP_FILE_BUFFER))
1672 dirty |= I_DIRTY_DATASYNC;
1673 __cap_delay_requeue(mdsc, ci, true);
1677 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1679 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1682 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1685 kmem_cache_free(ceph_cap_flush_cachep, cf);
1688 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1690 if (!list_empty(&mdsc->cap_flush_list)) {
1691 struct ceph_cap_flush *cf =
1692 list_first_entry(&mdsc->cap_flush_list,
1693 struct ceph_cap_flush, g_list);
1700 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1701 * Return true if caller needs to wake up flush waiters.
1703 static bool __finish_cap_flush(struct ceph_mds_client *mdsc,
1704 struct ceph_inode_info *ci,
1705 struct ceph_cap_flush *cf)
1707 struct ceph_cap_flush *prev;
1708 bool wake = cf->wake;
1710 /* are there older pending cap flushes? */
1711 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1712 prev = list_prev_entry(cf, g_list);
1716 list_del(&cf->g_list);
1718 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1719 prev = list_prev_entry(cf, i_list);
1723 list_del(&cf->i_list);
1731 * Add dirty inode to the flushing list. Assigned a seq number so we
1732 * can wait for caps to flush without starving.
1734 * Called under i_ceph_lock. Returns the flush tid.
1736 static u64 __mark_caps_flushing(struct inode *inode,
1737 struct ceph_mds_session *session, bool wake,
1738 u64 *oldest_flush_tid)
1740 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1741 struct ceph_inode_info *ci = ceph_inode(inode);
1742 struct ceph_cap_flush *cf = NULL;
1745 BUG_ON(ci->i_dirty_caps == 0);
1746 BUG_ON(list_empty(&ci->i_dirty_item));
1747 BUG_ON(!ci->i_prealloc_cap_flush);
1749 flushing = ci->i_dirty_caps;
1750 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1751 ceph_cap_string(flushing),
1752 ceph_cap_string(ci->i_flushing_caps),
1753 ceph_cap_string(ci->i_flushing_caps | flushing));
1754 ci->i_flushing_caps |= flushing;
1755 ci->i_dirty_caps = 0;
1756 dout(" inode %p now !dirty\n", inode);
1758 swap(cf, ci->i_prealloc_cap_flush);
1759 cf->caps = flushing;
1762 spin_lock(&mdsc->cap_dirty_lock);
1763 list_del_init(&ci->i_dirty_item);
1765 cf->tid = ++mdsc->last_cap_flush_tid;
1766 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1767 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1769 if (list_empty(&ci->i_flushing_item)) {
1770 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1771 mdsc->num_cap_flushing++;
1773 spin_unlock(&mdsc->cap_dirty_lock);
1775 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1781 * try to invalidate mapping pages without blocking.
1783 static int try_nonblocking_invalidate(struct inode *inode)
1784 __releases(ci->i_ceph_lock)
1785 __acquires(ci->i_ceph_lock)
1787 struct ceph_inode_info *ci = ceph_inode(inode);
1788 u32 invalidating_gen = ci->i_rdcache_gen;
1790 spin_unlock(&ci->i_ceph_lock);
1791 ceph_fscache_invalidate(inode);
1792 invalidate_mapping_pages(&inode->i_data, 0, -1);
1793 spin_lock(&ci->i_ceph_lock);
1795 if (inode->i_data.nrpages == 0 &&
1796 invalidating_gen == ci->i_rdcache_gen) {
1798 dout("try_nonblocking_invalidate %p success\n", inode);
1799 /* save any racing async invalidate some trouble */
1800 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1803 dout("try_nonblocking_invalidate %p failed\n", inode);
1807 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1809 loff_t size = ci->vfs_inode.i_size;
1810 /* mds will adjust max size according to the reported size */
1811 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1813 if (size >= ci->i_max_size)
1815 /* half of previous max_size increment has been used */
1816 if (ci->i_max_size > ci->i_reported_size &&
1817 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1823 * Swiss army knife function to examine currently used and wanted
1824 * versus held caps. Release, flush, ack revoked caps to mds as
1827 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1828 * cap release further.
1829 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1830 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1833 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1834 struct ceph_mds_session *session)
1836 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1837 struct ceph_mds_client *mdsc = fsc->mdsc;
1838 struct inode *inode = &ci->vfs_inode;
1839 struct ceph_cap *cap;
1840 u64 flush_tid, oldest_flush_tid;
1841 int file_wanted, used, cap_used;
1842 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1843 int issued, implemented, want, retain, revoking, flushing = 0;
1844 int mds = -1; /* keep track of how far we've gone through i_caps list
1845 to avoid an infinite loop on retry */
1847 int delayed = 0, sent = 0;
1848 bool no_delay = flags & CHECK_CAPS_NODELAY;
1849 bool queue_invalidate = false;
1850 bool tried_invalidate = false;
1852 /* if we are unmounting, flush any unused caps immediately. */
1856 spin_lock(&ci->i_ceph_lock);
1858 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1859 flags |= CHECK_CAPS_FLUSH;
1861 if (!(flags & CHECK_CAPS_AUTHONLY) ||
1862 (ci->i_auth_cap && __ceph_is_single_caps(ci)))
1863 __cap_delay_cancel(mdsc, ci);
1867 spin_lock(&ci->i_ceph_lock);
1869 file_wanted = __ceph_caps_file_wanted(ci);
1870 used = __ceph_caps_used(ci);
1871 issued = __ceph_caps_issued(ci, &implemented);
1872 revoking = implemented & ~issued;
1875 retain = file_wanted | used | CEPH_CAP_PIN;
1876 if (!mdsc->stopping && inode->i_nlink > 0) {
1878 retain |= CEPH_CAP_ANY; /* be greedy */
1879 } else if (S_ISDIR(inode->i_mode) &&
1880 (issued & CEPH_CAP_FILE_SHARED) &&
1881 __ceph_dir_is_complete(ci)) {
1883 * If a directory is complete, we want to keep
1884 * the exclusive cap. So that MDS does not end up
1885 * revoking the shared cap on every create/unlink
1888 if (IS_RDONLY(inode))
1889 want = CEPH_CAP_ANY_SHARED;
1891 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1895 retain |= CEPH_CAP_ANY_SHARED;
1897 * keep RD only if we didn't have the file open RW,
1898 * because then the mds would revoke it anyway to
1899 * journal max_size=0.
1901 if (ci->i_max_size == 0)
1902 retain |= CEPH_CAP_ANY_RD;
1906 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1907 " issued %s revoking %s retain %s %s%s%s\n", inode,
1908 ceph_cap_string(file_wanted),
1909 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1910 ceph_cap_string(ci->i_flushing_caps),
1911 ceph_cap_string(issued), ceph_cap_string(revoking),
1912 ceph_cap_string(retain),
1913 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1914 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1915 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1918 * If we no longer need to hold onto old our caps, and we may
1919 * have cached pages, but don't want them, then try to invalidate.
1920 * If we fail, it's because pages are locked.... try again later.
1922 if ((!no_delay || mdsc->stopping) &&
1923 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */
1924 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1925 inode->i_data.nrpages && /* have cached pages */
1926 (revoking & (CEPH_CAP_FILE_CACHE|
1927 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1928 !tried_invalidate) {
1929 dout("check_caps trying to invalidate on %p\n", inode);
1930 if (try_nonblocking_invalidate(inode) < 0) {
1931 dout("check_caps queuing invalidate\n");
1932 queue_invalidate = true;
1933 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1935 tried_invalidate = true;
1939 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1940 cap = rb_entry(p, struct ceph_cap, ci_node);
1942 /* avoid looping forever */
1943 if (mds >= cap->mds ||
1944 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1947 /* NOTE: no side-effects allowed, until we take s_mutex */
1950 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1951 cap_used &= ~ci->i_auth_cap->issued;
1953 revoking = cap->implemented & ~cap->issued;
1954 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1955 cap->mds, cap, ceph_cap_string(cap_used),
1956 ceph_cap_string(cap->issued),
1957 ceph_cap_string(cap->implemented),
1958 ceph_cap_string(revoking));
1960 if (cap == ci->i_auth_cap &&
1961 (cap->issued & CEPH_CAP_FILE_WR)) {
1962 /* request larger max_size from MDS? */
1963 if (ci->i_wanted_max_size > ci->i_max_size &&
1964 ci->i_wanted_max_size > ci->i_requested_max_size) {
1965 dout("requesting new max_size\n");
1969 /* approaching file_max? */
1970 if (__ceph_should_report_size(ci)) {
1971 dout("i_size approaching max_size\n");
1975 /* flush anything dirty? */
1976 if (cap == ci->i_auth_cap) {
1977 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
1978 dout("flushing dirty caps\n");
1981 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
1982 dout("flushing snap caps\n");
1987 /* completed revocation? going down and there are no caps? */
1988 if (revoking && (revoking & cap_used) == 0) {
1989 dout("completed revocation of %s\n",
1990 ceph_cap_string(cap->implemented & ~cap->issued));
1994 /* want more caps from mds? */
1995 if (want & ~cap->mds_wanted) {
1996 if (want & ~(cap->mds_wanted | cap->issued))
1998 if (!__cap_is_valid(cap))
2002 /* things we might delay */
2003 if ((cap->issued & ~retain) == 0)
2004 continue; /* nope, all good */
2010 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
2011 time_before(jiffies, ci->i_hold_caps_max)) {
2012 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
2013 ceph_cap_string(cap->issued),
2014 ceph_cap_string(cap->issued & retain),
2015 ceph_cap_string(cap->mds_wanted),
2016 ceph_cap_string(want));
2022 if (session && session != cap->session) {
2023 dout("oops, wrong session %p mutex\n", session);
2024 mutex_unlock(&session->s_mutex);
2028 session = cap->session;
2029 if (mutex_trylock(&session->s_mutex) == 0) {
2030 dout("inverting session/ino locks on %p\n",
2032 session = ceph_get_mds_session(session);
2033 spin_unlock(&ci->i_ceph_lock);
2034 if (took_snap_rwsem) {
2035 up_read(&mdsc->snap_rwsem);
2036 took_snap_rwsem = 0;
2039 mutex_lock(&session->s_mutex);
2040 ceph_put_mds_session(session);
2043 * Because we take the reference while
2044 * holding the i_ceph_lock, it should
2045 * never be NULL. Throw a warning if it
2054 /* kick flushing and flush snaps before sending normal
2056 if (cap == ci->i_auth_cap &&
2058 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2059 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2060 __kick_flushing_caps(mdsc, session, ci, 0);
2061 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2062 __ceph_flush_snaps(ci, session);
2067 /* take snap_rwsem after session mutex */
2068 if (!took_snap_rwsem) {
2069 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
2070 dout("inverting snap/in locks on %p\n",
2072 spin_unlock(&ci->i_ceph_lock);
2073 down_read(&mdsc->snap_rwsem);
2074 took_snap_rwsem = 1;
2077 took_snap_rwsem = 1;
2080 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2081 flushing = ci->i_dirty_caps;
2082 flush_tid = __mark_caps_flushing(inode, session, false,
2087 spin_lock(&mdsc->cap_dirty_lock);
2088 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2089 spin_unlock(&mdsc->cap_dirty_lock);
2092 mds = cap->mds; /* remember mds, so we don't repeat */
2095 /* __send_cap drops i_ceph_lock */
2096 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, 0,
2097 cap_used, want, retain, flushing,
2098 flush_tid, oldest_flush_tid);
2099 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2102 /* Reschedule delayed caps release if we delayed anything */
2104 __cap_delay_requeue(mdsc, ci, false);
2106 spin_unlock(&ci->i_ceph_lock);
2108 if (queue_invalidate)
2109 ceph_queue_invalidate(inode);
2112 mutex_unlock(&session->s_mutex);
2113 if (took_snap_rwsem)
2114 up_read(&mdsc->snap_rwsem);
2118 * Try to flush dirty caps back to the auth mds.
2120 static int try_flush_caps(struct inode *inode, u64 *ptid)
2122 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2123 struct ceph_inode_info *ci = ceph_inode(inode);
2124 struct ceph_mds_session *session = NULL;
2126 u64 flush_tid = 0, oldest_flush_tid = 0;
2129 spin_lock(&ci->i_ceph_lock);
2131 if (ci->i_dirty_caps && ci->i_auth_cap) {
2132 struct ceph_cap *cap = ci->i_auth_cap;
2135 if (session != cap->session) {
2136 spin_unlock(&ci->i_ceph_lock);
2138 mutex_unlock(&session->s_mutex);
2139 session = cap->session;
2140 mutex_lock(&session->s_mutex);
2143 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) {
2144 spin_unlock(&ci->i_ceph_lock);
2148 if (ci->i_ceph_flags &
2149 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2150 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2151 __kick_flushing_caps(mdsc, session, ci, 0);
2152 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2153 __ceph_flush_snaps(ci, session);
2157 flushing = ci->i_dirty_caps;
2158 flush_tid = __mark_caps_flushing(inode, session, true,
2161 /* __send_cap drops i_ceph_lock */
2162 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2163 CEPH_CLIENT_CAPS_SYNC,
2164 __ceph_caps_used(ci),
2165 __ceph_caps_wanted(ci),
2166 (cap->issued | cap->implemented),
2167 flushing, flush_tid, oldest_flush_tid);
2170 spin_lock(&ci->i_ceph_lock);
2171 __cap_delay_requeue(mdsc, ci, true);
2172 spin_unlock(&ci->i_ceph_lock);
2175 if (!list_empty(&ci->i_cap_flush_list)) {
2176 struct ceph_cap_flush *cf =
2177 list_last_entry(&ci->i_cap_flush_list,
2178 struct ceph_cap_flush, i_list);
2180 flush_tid = cf->tid;
2182 flushing = ci->i_flushing_caps;
2183 spin_unlock(&ci->i_ceph_lock);
2187 mutex_unlock(&session->s_mutex);
2194 * Return true if we've flushed caps through the given flush_tid.
2196 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2198 struct ceph_inode_info *ci = ceph_inode(inode);
2201 spin_lock(&ci->i_ceph_lock);
2202 if (!list_empty(&ci->i_cap_flush_list)) {
2203 struct ceph_cap_flush * cf =
2204 list_first_entry(&ci->i_cap_flush_list,
2205 struct ceph_cap_flush, i_list);
2206 if (cf->tid <= flush_tid)
2209 spin_unlock(&ci->i_ceph_lock);
2214 * wait for any unsafe requests to complete.
2216 static int unsafe_request_wait(struct inode *inode)
2218 struct ceph_inode_info *ci = ceph_inode(inode);
2219 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2222 spin_lock(&ci->i_unsafe_lock);
2223 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2224 req1 = list_last_entry(&ci->i_unsafe_dirops,
2225 struct ceph_mds_request,
2227 ceph_mdsc_get_request(req1);
2229 if (!list_empty(&ci->i_unsafe_iops)) {
2230 req2 = list_last_entry(&ci->i_unsafe_iops,
2231 struct ceph_mds_request,
2232 r_unsafe_target_item);
2233 ceph_mdsc_get_request(req2);
2235 spin_unlock(&ci->i_unsafe_lock);
2237 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2238 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2240 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2241 ceph_timeout_jiffies(req1->r_timeout));
2244 ceph_mdsc_put_request(req1);
2247 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2248 ceph_timeout_jiffies(req2->r_timeout));
2251 ceph_mdsc_put_request(req2);
2256 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2258 struct inode *inode = file->f_mapping->host;
2259 struct ceph_inode_info *ci = ceph_inode(inode);
2264 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2266 ret = file_write_and_wait_range(file, start, end);
2270 dirty = try_flush_caps(inode, &flush_tid);
2271 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2273 err = unsafe_request_wait(inode);
2276 * only wait on non-file metadata writeback (the mds
2277 * can recover size and mtime, so we don't need to
2280 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2281 err = wait_event_interruptible(ci->i_cap_wq,
2282 caps_are_flushed(inode, flush_tid));
2288 err = file_check_and_advance_wb_err(file);
2292 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2297 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2298 * queue inode for flush but don't do so immediately, because we can
2299 * get by with fewer MDS messages if we wait for data writeback to
2302 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2304 struct ceph_inode_info *ci = ceph_inode(inode);
2308 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2310 dout("write_inode %p wait=%d\n", inode, wait);
2312 dirty = try_flush_caps(inode, &flush_tid);
2314 err = wait_event_interruptible(ci->i_cap_wq,
2315 caps_are_flushed(inode, flush_tid));
2317 struct ceph_mds_client *mdsc =
2318 ceph_sb_to_client(inode->i_sb)->mdsc;
2320 spin_lock(&ci->i_ceph_lock);
2321 if (__ceph_caps_dirty(ci))
2322 __cap_delay_requeue_front(mdsc, ci);
2323 spin_unlock(&ci->i_ceph_lock);
2328 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2329 struct ceph_mds_session *session,
2330 struct ceph_inode_info *ci,
2331 u64 oldest_flush_tid)
2332 __releases(ci->i_ceph_lock)
2333 __acquires(ci->i_ceph_lock)
2335 struct inode *inode = &ci->vfs_inode;
2336 struct ceph_cap *cap;
2337 struct ceph_cap_flush *cf;
2340 u64 last_snap_flush = 0;
2342 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2344 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2346 last_snap_flush = cf->tid;
2351 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2352 if (cf->tid < first_tid)
2355 cap = ci->i_auth_cap;
2356 if (!(cap && cap->session == session)) {
2357 pr_err("%p auth cap %p not mds%d ???\n",
2358 inode, cap, session->s_mds);
2362 first_tid = cf->tid + 1;
2365 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2366 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2367 ci->i_ceph_flags |= CEPH_I_NODELAY;
2369 ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2370 (cf->tid < last_snap_flush ?
2371 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2372 __ceph_caps_used(ci),
2373 __ceph_caps_wanted(ci),
2374 (cap->issued | cap->implemented),
2375 cf->caps, cf->tid, oldest_flush_tid);
2377 pr_err("kick_flushing_caps: error sending "
2378 "cap flush, ino (%llx.%llx) "
2379 "tid %llu flushing %s\n",
2380 ceph_vinop(inode), cf->tid,
2381 ceph_cap_string(cf->caps));
2384 struct ceph_cap_snap *capsnap =
2385 container_of(cf, struct ceph_cap_snap,
2387 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2388 inode, capsnap, cf->tid,
2389 ceph_cap_string(capsnap->dirty));
2391 refcount_inc(&capsnap->nref);
2392 spin_unlock(&ci->i_ceph_lock);
2394 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2397 pr_err("kick_flushing_caps: error sending "
2398 "cap flushsnap, ino (%llx.%llx) "
2399 "tid %llu follows %llu\n",
2400 ceph_vinop(inode), cf->tid,
2404 ceph_put_cap_snap(capsnap);
2407 spin_lock(&ci->i_ceph_lock);
2411 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2412 struct ceph_mds_session *session)
2414 struct ceph_inode_info *ci;
2415 struct ceph_cap *cap;
2416 u64 oldest_flush_tid;
2418 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2420 spin_lock(&mdsc->cap_dirty_lock);
2421 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2422 spin_unlock(&mdsc->cap_dirty_lock);
2424 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2425 spin_lock(&ci->i_ceph_lock);
2426 cap = ci->i_auth_cap;
2427 if (!(cap && cap->session == session)) {
2428 pr_err("%p auth cap %p not mds%d ???\n",
2429 &ci->vfs_inode, cap, session->s_mds);
2430 spin_unlock(&ci->i_ceph_lock);
2436 * if flushing caps were revoked, we re-send the cap flush
2437 * in client reconnect stage. This guarantees MDS * processes
2438 * the cap flush message before issuing the flushing caps to
2441 if ((cap->issued & ci->i_flushing_caps) !=
2442 ci->i_flushing_caps) {
2443 /* encode_caps_cb() also will reset these sequence
2444 * numbers. make sure sequence numbers in cap flush
2445 * message match later reconnect message */
2449 __kick_flushing_caps(mdsc, session, ci,
2452 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2455 spin_unlock(&ci->i_ceph_lock);
2459 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2460 struct ceph_mds_session *session)
2462 struct ceph_inode_info *ci;
2463 struct ceph_cap *cap;
2464 u64 oldest_flush_tid;
2466 dout("kick_flushing_caps mds%d\n", session->s_mds);
2468 spin_lock(&mdsc->cap_dirty_lock);
2469 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2470 spin_unlock(&mdsc->cap_dirty_lock);
2472 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2473 spin_lock(&ci->i_ceph_lock);
2474 cap = ci->i_auth_cap;
2475 if (!(cap && cap->session == session)) {
2476 pr_err("%p auth cap %p not mds%d ???\n",
2477 &ci->vfs_inode, cap, session->s_mds);
2478 spin_unlock(&ci->i_ceph_lock);
2481 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2482 __kick_flushing_caps(mdsc, session, ci,
2485 spin_unlock(&ci->i_ceph_lock);
2489 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2490 struct ceph_mds_session *session,
2491 struct inode *inode)
2492 __releases(ci->i_ceph_lock)
2494 struct ceph_inode_info *ci = ceph_inode(inode);
2495 struct ceph_cap *cap;
2497 cap = ci->i_auth_cap;
2498 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2499 ceph_cap_string(ci->i_flushing_caps));
2501 if (!list_empty(&ci->i_cap_flush_list)) {
2502 u64 oldest_flush_tid;
2503 spin_lock(&mdsc->cap_dirty_lock);
2504 list_move_tail(&ci->i_flushing_item,
2505 &cap->session->s_cap_flushing);
2506 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2507 spin_unlock(&mdsc->cap_dirty_lock);
2509 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2510 spin_unlock(&ci->i_ceph_lock);
2512 spin_unlock(&ci->i_ceph_lock);
2518 * Take references to capabilities we hold, so that we don't release
2519 * them to the MDS prematurely.
2521 * Protected by i_ceph_lock.
2523 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2524 bool snap_rwsem_locked)
2526 if (got & CEPH_CAP_PIN)
2528 if (got & CEPH_CAP_FILE_RD)
2530 if (got & CEPH_CAP_FILE_CACHE)
2531 ci->i_rdcache_ref++;
2532 if (got & CEPH_CAP_FILE_WR) {
2533 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2534 BUG_ON(!snap_rwsem_locked);
2535 ci->i_head_snapc = ceph_get_snap_context(
2536 ci->i_snap_realm->cached_context);
2540 if (got & CEPH_CAP_FILE_BUFFER) {
2541 if (ci->i_wb_ref == 0)
2542 ihold(&ci->vfs_inode);
2544 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2545 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2550 * Try to grab cap references. Specify those refs we @want, and the
2551 * minimal set we @need. Also include the larger offset we are writing
2552 * to (when applicable), and check against max_size here as well.
2553 * Note that caller is responsible for ensuring max_size increases are
2554 * requested from the MDS.
2556 * Returns 0 if caps were not able to be acquired (yet), a 1 if they were,
2557 * or a negative error code.
2559 * FIXME: how does a 0 return differ from -EAGAIN?
2566 static int try_get_cap_refs(struct inode *inode, int need, int want,
2567 loff_t endoff, int flags, int *got)
2569 struct ceph_inode_info *ci = ceph_inode(inode);
2570 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2572 int have, implemented;
2574 bool snap_rwsem_locked = false;
2576 dout("get_cap_refs %p need %s want %s\n", inode,
2577 ceph_cap_string(need), ceph_cap_string(want));
2580 spin_lock(&ci->i_ceph_lock);
2582 if ((flags & CHECK_FILELOCK) &&
2583 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2584 dout("try_get_cap_refs %p error filelock\n", inode);
2589 /* make sure file is actually open */
2590 file_wanted = __ceph_caps_file_wanted(ci);
2591 if ((file_wanted & need) != need) {
2592 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2593 ceph_cap_string(need), ceph_cap_string(file_wanted));
2598 /* finish pending truncate */
2599 while (ci->i_truncate_pending) {
2600 spin_unlock(&ci->i_ceph_lock);
2601 if (snap_rwsem_locked) {
2602 up_read(&mdsc->snap_rwsem);
2603 snap_rwsem_locked = false;
2605 __ceph_do_pending_vmtruncate(inode);
2606 spin_lock(&ci->i_ceph_lock);
2609 have = __ceph_caps_issued(ci, &implemented);
2611 if (have & need & CEPH_CAP_FILE_WR) {
2612 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2613 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2614 inode, endoff, ci->i_max_size);
2615 if (endoff > ci->i_requested_max_size)
2620 * If a sync write is in progress, we must wait, so that we
2621 * can get a final snapshot value for size+mtime.
2623 if (__ceph_have_pending_cap_snap(ci)) {
2624 dout("get_cap_refs %p cap_snap_pending\n", inode);
2629 if ((have & need) == need) {
2631 * Look at (implemented & ~have & not) so that we keep waiting
2632 * on transition from wanted -> needed caps. This is needed
2633 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2634 * going before a prior buffered writeback happens.
2636 int not = want & ~(have & need);
2637 int revoking = implemented & ~have;
2638 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2639 inode, ceph_cap_string(have), ceph_cap_string(not),
2640 ceph_cap_string(revoking));
2641 if ((revoking & not) == 0) {
2642 if (!snap_rwsem_locked &&
2643 !ci->i_head_snapc &&
2644 (need & CEPH_CAP_FILE_WR)) {
2645 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2647 * we can not call down_read() when
2648 * task isn't in TASK_RUNNING state
2650 if (flags & NON_BLOCKING) {
2655 spin_unlock(&ci->i_ceph_lock);
2656 down_read(&mdsc->snap_rwsem);
2657 snap_rwsem_locked = true;
2660 snap_rwsem_locked = true;
2662 *got = need | (have & want);
2663 if ((need & CEPH_CAP_FILE_RD) &&
2664 !(*got & CEPH_CAP_FILE_CACHE))
2665 ceph_disable_fscache_readpage(ci);
2666 __take_cap_refs(ci, *got, true);
2670 int session_readonly = false;
2671 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2672 struct ceph_mds_session *s = ci->i_auth_cap->session;
2673 spin_lock(&s->s_cap_lock);
2674 session_readonly = s->s_readonly;
2675 spin_unlock(&s->s_cap_lock);
2677 if (session_readonly) {
2678 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2679 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2684 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) {
2686 if (READ_ONCE(mdsc->fsc->mount_state) ==
2687 CEPH_MOUNT_SHUTDOWN) {
2688 dout("get_cap_refs %p forced umount\n", inode);
2692 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2693 if (need & ~(mds_wanted & need)) {
2694 dout("get_cap_refs %p caps were dropped"
2695 " (session killed?)\n", inode);
2699 if (!(file_wanted & ~mds_wanted))
2700 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED;
2703 dout("get_cap_refs %p have %s needed %s\n", inode,
2704 ceph_cap_string(have), ceph_cap_string(need));
2707 spin_unlock(&ci->i_ceph_lock);
2708 if (snap_rwsem_locked)
2709 up_read(&mdsc->snap_rwsem);
2711 dout("get_cap_refs %p ret %d got %s\n", inode,
2712 ret, ceph_cap_string(*got));
2717 * Check the offset we are writing up to against our current
2718 * max_size. If necessary, tell the MDS we want to write to
2721 static void check_max_size(struct inode *inode, loff_t endoff)
2723 struct ceph_inode_info *ci = ceph_inode(inode);
2726 /* do we need to explicitly request a larger max_size? */
2727 spin_lock(&ci->i_ceph_lock);
2728 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2729 dout("write %p at large endoff %llu, req max_size\n",
2731 ci->i_wanted_max_size = endoff;
2733 /* duplicate ceph_check_caps()'s logic */
2734 if (ci->i_auth_cap &&
2735 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2736 ci->i_wanted_max_size > ci->i_max_size &&
2737 ci->i_wanted_max_size > ci->i_requested_max_size)
2739 spin_unlock(&ci->i_ceph_lock);
2741 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2744 int ceph_try_get_caps(struct inode *inode, int need, int want,
2745 bool nonblock, int *got)
2749 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2750 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO|CEPH_CAP_FILE_SHARED));
2751 ret = ceph_pool_perm_check(inode, need);
2755 ret = try_get_cap_refs(inode, need, want, 0,
2756 (nonblock ? NON_BLOCKING : 0), got);
2757 return ret == -EAGAIN ? 0 : ret;
2761 * Wait for caps, and take cap references. If we can't get a WR cap
2762 * due to a small max_size, make sure we check_max_size (and possibly
2763 * ask the mds) so we don't get hung up indefinitely.
2765 int ceph_get_caps(struct file *filp, int need, int want,
2766 loff_t endoff, int *got, struct page **pinned_page)
2768 struct ceph_file_info *fi = filp->private_data;
2769 struct inode *inode = file_inode(filp);
2770 struct ceph_inode_info *ci = ceph_inode(inode);
2771 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2772 int ret, _got, flags;
2774 ret = ceph_pool_perm_check(inode, need);
2778 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2779 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2784 check_max_size(inode, endoff);
2786 flags = atomic_read(&fi->num_locks) ? CHECK_FILELOCK : 0;
2788 ret = try_get_cap_refs(inode, need, want, endoff,
2793 struct ceph_mds_client *mdsc = fsc->mdsc;
2795 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2797 cw.ino = inode->i_ino;
2798 cw.tgid = current->tgid;
2802 spin_lock(&mdsc->caps_list_lock);
2803 list_add(&cw.list, &mdsc->cap_wait_list);
2804 spin_unlock(&mdsc->caps_list_lock);
2806 add_wait_queue(&ci->i_cap_wq, &wait);
2808 flags |= NON_BLOCKING;
2809 while (!(ret = try_get_cap_refs(inode, need, want,
2810 endoff, flags, &_got))) {
2811 if (signal_pending(current)) {
2815 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2818 remove_wait_queue(&ci->i_cap_wq, &wait);
2820 spin_lock(&mdsc->caps_list_lock);
2822 spin_unlock(&mdsc->caps_list_lock);
2828 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2829 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2830 if (ret >= 0 && _got)
2831 ceph_put_cap_refs(ci, _got);
2836 if (ret == -ESTALE) {
2837 /* session was killed, try renew caps */
2838 ret = ceph_renew_caps(inode);
2845 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2846 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2847 i_size_read(inode) > 0) {
2849 find_get_page(inode->i_mapping, 0);
2851 if (PageUptodate(page)) {
2852 *pinned_page = page;
2858 * drop cap refs first because getattr while
2859 * holding * caps refs can cause deadlock.
2861 ceph_put_cap_refs(ci, _got);
2865 * getattr request will bring inline data into
2868 ret = __ceph_do_getattr(inode, NULL,
2869 CEPH_STAT_CAP_INLINE_DATA,
2878 if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
2879 ceph_fscache_revalidate_cookie(ci);
2886 * Take cap refs. Caller must already know we hold at least one ref
2887 * on the caps in question or we don't know this is safe.
2889 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2891 spin_lock(&ci->i_ceph_lock);
2892 __take_cap_refs(ci, caps, false);
2893 spin_unlock(&ci->i_ceph_lock);
2898 * drop cap_snap that is not associated with any snapshot.
2899 * we don't need to send FLUSHSNAP message for it.
2901 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2902 struct ceph_cap_snap *capsnap)
2904 if (!capsnap->need_flush &&
2905 !capsnap->writing && !capsnap->dirty_pages) {
2906 dout("dropping cap_snap %p follows %llu\n",
2907 capsnap, capsnap->follows);
2908 BUG_ON(capsnap->cap_flush.tid > 0);
2909 ceph_put_snap_context(capsnap->context);
2910 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2911 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2913 list_del(&capsnap->ci_item);
2914 ceph_put_cap_snap(capsnap);
2923 * If we released the last ref on any given cap, call ceph_check_caps
2924 * to release (or schedule a release).
2926 * If we are releasing a WR cap (from a sync write), finalize any affected
2927 * cap_snap, and wake up any waiters.
2929 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2931 struct inode *inode = &ci->vfs_inode;
2932 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2934 spin_lock(&ci->i_ceph_lock);
2935 if (had & CEPH_CAP_PIN)
2937 if (had & CEPH_CAP_FILE_RD)
2938 if (--ci->i_rd_ref == 0)
2940 if (had & CEPH_CAP_FILE_CACHE)
2941 if (--ci->i_rdcache_ref == 0)
2943 if (had & CEPH_CAP_FILE_BUFFER) {
2944 if (--ci->i_wb_ref == 0) {
2948 dout("put_cap_refs %p wb %d -> %d (?)\n",
2949 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2951 if (had & CEPH_CAP_FILE_WR)
2952 if (--ci->i_wr_ref == 0) {
2954 if (__ceph_have_pending_cap_snap(ci)) {
2955 struct ceph_cap_snap *capsnap =
2956 list_last_entry(&ci->i_cap_snaps,
2957 struct ceph_cap_snap,
2959 capsnap->writing = 0;
2960 if (ceph_try_drop_cap_snap(ci, capsnap))
2962 else if (__ceph_finish_cap_snap(ci, capsnap))
2966 if (ci->i_wrbuffer_ref_head == 0 &&
2967 ci->i_dirty_caps == 0 &&
2968 ci->i_flushing_caps == 0) {
2969 BUG_ON(!ci->i_head_snapc);
2970 ceph_put_snap_context(ci->i_head_snapc);
2971 ci->i_head_snapc = NULL;
2973 /* see comment in __ceph_remove_cap() */
2974 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2975 drop_inode_snap_realm(ci);
2977 spin_unlock(&ci->i_ceph_lock);
2979 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2980 last ? " last" : "", put ? " put" : "");
2982 if (last && !flushsnaps)
2983 ceph_check_caps(ci, 0, NULL);
2984 else if (flushsnaps)
2985 ceph_flush_snaps(ci, NULL);
2987 wake_up_all(&ci->i_cap_wq);
2993 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2994 * context. Adjust per-snap dirty page accounting as appropriate.
2995 * Once all dirty data for a cap_snap is flushed, flush snapped file
2996 * metadata back to the MDS. If we dropped the last ref, call
2999 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
3000 struct ceph_snap_context *snapc)
3002 struct inode *inode = &ci->vfs_inode;
3003 struct ceph_cap_snap *capsnap = NULL;
3007 bool flush_snaps = false;
3008 bool complete_capsnap = false;
3010 spin_lock(&ci->i_ceph_lock);
3011 ci->i_wrbuffer_ref -= nr;
3012 if (ci->i_wrbuffer_ref == 0) {
3017 if (ci->i_head_snapc == snapc) {
3018 ci->i_wrbuffer_ref_head -= nr;
3019 if (ci->i_wrbuffer_ref_head == 0 &&
3020 ci->i_wr_ref == 0 &&
3021 ci->i_dirty_caps == 0 &&
3022 ci->i_flushing_caps == 0) {
3023 BUG_ON(!ci->i_head_snapc);
3024 ceph_put_snap_context(ci->i_head_snapc);
3025 ci->i_head_snapc = NULL;
3027 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3029 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3030 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3031 last ? " LAST" : "");
3033 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3034 if (capsnap->context == snapc) {
3040 capsnap->dirty_pages -= nr;
3041 if (capsnap->dirty_pages == 0) {
3042 complete_capsnap = true;
3043 if (!capsnap->writing) {
3044 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3047 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3052 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3053 " snap %lld %d/%d -> %d/%d %s%s\n",
3054 inode, capsnap, capsnap->context->seq,
3055 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3056 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3057 last ? " (wrbuffer last)" : "",
3058 complete_capsnap ? " (complete capsnap)" : "");
3061 spin_unlock(&ci->i_ceph_lock);
3064 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
3065 } else if (flush_snaps) {
3066 ceph_flush_snaps(ci, NULL);
3068 if (complete_capsnap)
3069 wake_up_all(&ci->i_cap_wq);
3071 /* avoid calling iput_final() in osd dispatch threads */
3072 ceph_async_iput(inode);
3077 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3079 static void invalidate_aliases(struct inode *inode)
3081 struct dentry *dn, *prev = NULL;
3083 dout("invalidate_aliases inode %p\n", inode);
3084 d_prune_aliases(inode);
3086 * For non-directory inode, d_find_alias() only returns
3087 * hashed dentry. After calling d_invalidate(), the
3088 * dentry becomes unhashed.
3090 * For directory inode, d_find_alias() can return
3091 * unhashed dentry. But directory inode should have
3092 * one alias at most.
3094 while ((dn = d_find_alias(inode))) {
3108 struct cap_extra_info {
3109 struct ceph_string *pool_ns;
3119 /* currently issued */
3121 struct timespec64 btime;
3125 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3126 * actually be a revocation if it specifies a smaller cap set.)
3128 * caller holds s_mutex and i_ceph_lock, we drop both.
3130 static void handle_cap_grant(struct inode *inode,
3131 struct ceph_mds_session *session,
3132 struct ceph_cap *cap,
3133 struct ceph_mds_caps *grant,
3134 struct ceph_buffer *xattr_buf,
3135 struct cap_extra_info *extra_info)
3136 __releases(ci->i_ceph_lock)
3137 __releases(session->s_mdsc->snap_rwsem)
3139 struct ceph_inode_info *ci = ceph_inode(inode);
3140 int seq = le32_to_cpu(grant->seq);
3141 int newcaps = le32_to_cpu(grant->caps);
3142 int used, wanted, dirty;
3143 u64 size = le64_to_cpu(grant->size);
3144 u64 max_size = le64_to_cpu(grant->max_size);
3145 unsigned char check_caps = 0;
3146 bool was_stale = cap->cap_gen < session->s_cap_gen;
3148 bool writeback = false;
3149 bool queue_trunc = false;
3150 bool queue_invalidate = false;
3151 bool deleted_inode = false;
3152 bool fill_inline = false;
3154 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3155 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3156 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3161 * If CACHE is being revoked, and we have no dirty buffers,
3162 * try to invalidate (once). (If there are dirty buffers, we
3163 * will invalidate _after_ writeback.)
3165 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
3166 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3167 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3168 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3169 if (try_nonblocking_invalidate(inode)) {
3170 /* there were locked pages.. invalidate later
3171 in a separate thread. */
3172 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3173 queue_invalidate = true;
3174 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3180 cap->issued = cap->implemented = CEPH_CAP_PIN;
3183 * auth mds of the inode changed. we received the cap export message,
3184 * but still haven't received the cap import message. handle_cap_export
3185 * updated the new auth MDS' cap.
3187 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3188 * that was sent before the cap import message. So don't remove caps.
3190 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3191 WARN_ON(cap != ci->i_auth_cap);
3192 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3194 newcaps |= cap->issued;
3197 /* side effects now are allowed */
3198 cap->cap_gen = session->s_cap_gen;
3201 __check_cap_issue(ci, cap, newcaps);
3203 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3205 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3206 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3207 inode->i_mode = le32_to_cpu(grant->mode);
3208 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3209 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3210 ci->i_btime = extra_info->btime;
3211 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3212 from_kuid(&init_user_ns, inode->i_uid),
3213 from_kgid(&init_user_ns, inode->i_gid));
3216 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3217 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3218 set_nlink(inode, le32_to_cpu(grant->nlink));
3219 if (inode->i_nlink == 0 &&
3220 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3221 deleted_inode = true;
3224 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3226 int len = le32_to_cpu(grant->xattr_len);
3227 u64 version = le64_to_cpu(grant->xattr_version);
3229 if (version > ci->i_xattrs.version) {
3230 dout(" got new xattrs v%llu on %p len %d\n",
3231 version, inode, len);
3232 if (ci->i_xattrs.blob)
3233 ceph_buffer_put(ci->i_xattrs.blob);
3234 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3235 ci->i_xattrs.version = version;
3236 ceph_forget_all_cached_acls(inode);
3237 ceph_security_invalidate_secctx(inode);
3241 if (newcaps & CEPH_CAP_ANY_RD) {
3242 struct timespec64 mtime, atime, ctime;
3243 /* ctime/mtime/atime? */
3244 ceph_decode_timespec64(&mtime, &grant->mtime);
3245 ceph_decode_timespec64(&atime, &grant->atime);
3246 ceph_decode_timespec64(&ctime, &grant->ctime);
3247 ceph_fill_file_time(inode, extra_info->issued,
3248 le32_to_cpu(grant->time_warp_seq),
3249 &ctime, &mtime, &atime);
3252 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3253 ci->i_files = extra_info->nfiles;
3254 ci->i_subdirs = extra_info->nsubdirs;
3257 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3258 /* file layout may have changed */
3259 s64 old_pool = ci->i_layout.pool_id;
3260 struct ceph_string *old_ns;
3262 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3263 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3264 lockdep_is_held(&ci->i_ceph_lock));
3265 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3267 if (ci->i_layout.pool_id != old_pool ||
3268 extra_info->pool_ns != old_ns)
3269 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3271 extra_info->pool_ns = old_ns;
3273 /* size/truncate_seq? */
3274 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3275 le32_to_cpu(grant->truncate_seq),
3276 le64_to_cpu(grant->truncate_size),
3280 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3281 if (max_size != ci->i_max_size) {
3282 dout("max_size %lld -> %llu\n",
3283 ci->i_max_size, max_size);
3284 ci->i_max_size = max_size;
3285 if (max_size >= ci->i_wanted_max_size) {
3286 ci->i_wanted_max_size = 0; /* reset */
3287 ci->i_requested_max_size = 0;
3290 } else if (ci->i_wanted_max_size > ci->i_max_size &&
3291 ci->i_wanted_max_size > ci->i_requested_max_size) {
3292 /* CEPH_CAP_OP_IMPORT */
3297 /* check cap bits */
3298 wanted = __ceph_caps_wanted(ci);
3299 used = __ceph_caps_used(ci);
3300 dirty = __ceph_caps_dirty(ci);
3301 dout(" my wanted = %s, used = %s, dirty %s\n",
3302 ceph_cap_string(wanted),
3303 ceph_cap_string(used),
3304 ceph_cap_string(dirty));
3306 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3307 (wanted & ~(cap->mds_wanted | newcaps))) {
3309 * If mds is importing cap, prior cap messages that update
3310 * 'wanted' may get dropped by mds (migrate seq mismatch).
3312 * We don't send cap message to update 'wanted' if what we
3313 * want are already issued. If mds revokes caps, cap message
3314 * that releases caps also tells mds what we want. But if
3315 * caps got revoked by mds forcedly (session stale). We may
3316 * haven't told mds what we want.
3321 /* revocation, grant, or no-op? */
3322 if (cap->issued & ~newcaps) {
3323 int revoking = cap->issued & ~newcaps;
3325 dout("revocation: %s -> %s (revoking %s)\n",
3326 ceph_cap_string(cap->issued),
3327 ceph_cap_string(newcaps),
3328 ceph_cap_string(revoking));
3329 if (revoking & used & CEPH_CAP_FILE_BUFFER)
3330 writeback = true; /* initiate writeback; will delay ack */
3331 else if (revoking == CEPH_CAP_FILE_CACHE &&
3332 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3334 ; /* do nothing yet, invalidation will be queued */
3335 else if (cap == ci->i_auth_cap)
3336 check_caps = 1; /* check auth cap only */
3338 check_caps = 2; /* check all caps */
3339 cap->issued = newcaps;
3340 cap->implemented |= newcaps;
3341 } else if (cap->issued == newcaps) {
3342 dout("caps unchanged: %s -> %s\n",
3343 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3345 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3346 ceph_cap_string(newcaps));
3347 /* non-auth MDS is revoking the newly grant caps ? */
3348 if (cap == ci->i_auth_cap &&
3349 __ceph_caps_revoking_other(ci, cap, newcaps))
3352 cap->issued = newcaps;
3353 cap->implemented |= newcaps; /* add bits only, to
3354 * avoid stepping on a
3355 * pending revocation */
3358 BUG_ON(cap->issued & ~cap->implemented);
3360 /* don't let check_caps skip sending a response to MDS for revoke msgs */
3361 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_REVOKE) {
3362 cap->mds_wanted = 0;
3363 if (cap == ci->i_auth_cap)
3364 check_caps = 1; /* check auth cap only */
3366 check_caps = 2; /* check all caps */
3369 if (extra_info->inline_version > 0 &&
3370 extra_info->inline_version >= ci->i_inline_version) {
3371 ci->i_inline_version = extra_info->inline_version;
3372 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3373 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3377 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3378 if (newcaps & ~extra_info->issued)
3380 kick_flushing_inode_caps(session->s_mdsc, session, inode);
3381 up_read(&session->s_mdsc->snap_rwsem);
3383 spin_unlock(&ci->i_ceph_lock);
3387 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3388 extra_info->inline_len);
3391 ceph_queue_vmtruncate(inode);
3395 * queue inode for writeback: we can't actually call
3396 * filemap_write_and_wait, etc. from message handler
3399 ceph_queue_writeback(inode);
3400 if (queue_invalidate)
3401 ceph_queue_invalidate(inode);
3403 invalidate_aliases(inode);
3405 wake_up_all(&ci->i_cap_wq);
3407 if (check_caps == 1)
3408 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
3410 else if (check_caps == 2)
3411 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
3413 mutex_unlock(&session->s_mutex);
3417 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3418 * MDS has been safely committed.
3420 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3421 struct ceph_mds_caps *m,
3422 struct ceph_mds_session *session,
3423 struct ceph_cap *cap)
3424 __releases(ci->i_ceph_lock)
3426 struct ceph_inode_info *ci = ceph_inode(inode);
3427 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3428 struct ceph_cap_flush *cf, *tmp_cf;
3429 LIST_HEAD(to_remove);
3430 unsigned seq = le32_to_cpu(m->seq);
3431 int dirty = le32_to_cpu(m->dirty);
3434 bool wake_ci = false;
3435 bool wake_mdsc = false;
3437 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3438 if (cf->tid == flush_tid)
3440 if (cf->caps == 0) /* capsnap */
3442 if (cf->tid <= flush_tid) {
3443 if (__finish_cap_flush(NULL, ci, cf))
3445 list_add_tail(&cf->i_list, &to_remove);
3447 cleaned &= ~cf->caps;
3453 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3454 " flushing %s -> %s\n",
3455 inode, session->s_mds, seq, ceph_cap_string(dirty),
3456 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3457 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3459 if (list_empty(&to_remove) && !cleaned)
3462 ci->i_flushing_caps &= ~cleaned;
3464 spin_lock(&mdsc->cap_dirty_lock);
3466 list_for_each_entry(cf, &to_remove, i_list) {
3467 if (__finish_cap_flush(mdsc, NULL, cf))
3471 if (ci->i_flushing_caps == 0) {
3472 if (list_empty(&ci->i_cap_flush_list)) {
3473 list_del_init(&ci->i_flushing_item);
3474 if (!list_empty(&session->s_cap_flushing)) {
3475 dout(" mds%d still flushing cap on %p\n",
3477 &list_first_entry(&session->s_cap_flushing,
3478 struct ceph_inode_info,
3479 i_flushing_item)->vfs_inode);
3482 mdsc->num_cap_flushing--;
3483 dout(" inode %p now !flushing\n", inode);
3485 if (ci->i_dirty_caps == 0) {
3486 dout(" inode %p now clean\n", inode);
3487 BUG_ON(!list_empty(&ci->i_dirty_item));
3489 if (ci->i_wr_ref == 0 &&
3490 ci->i_wrbuffer_ref_head == 0) {
3491 BUG_ON(!ci->i_head_snapc);
3492 ceph_put_snap_context(ci->i_head_snapc);
3493 ci->i_head_snapc = NULL;
3496 BUG_ON(list_empty(&ci->i_dirty_item));
3499 spin_unlock(&mdsc->cap_dirty_lock);
3502 spin_unlock(&ci->i_ceph_lock);
3504 while (!list_empty(&to_remove)) {
3505 cf = list_first_entry(&to_remove,
3506 struct ceph_cap_flush, i_list);
3507 list_del(&cf->i_list);
3508 ceph_free_cap_flush(cf);
3512 wake_up_all(&ci->i_cap_wq);
3514 wake_up_all(&mdsc->cap_flushing_wq);
3520 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3521 * throw away our cap_snap.
3523 * Caller hold s_mutex.
3525 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3526 struct ceph_mds_caps *m,
3527 struct ceph_mds_session *session)
3529 struct ceph_inode_info *ci = ceph_inode(inode);
3530 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3531 u64 follows = le64_to_cpu(m->snap_follows);
3532 struct ceph_cap_snap *capsnap;
3533 bool flushed = false;
3534 bool wake_ci = false;
3535 bool wake_mdsc = false;
3537 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3538 inode, ci, session->s_mds, follows);
3540 spin_lock(&ci->i_ceph_lock);
3541 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3542 if (capsnap->follows == follows) {
3543 if (capsnap->cap_flush.tid != flush_tid) {
3544 dout(" cap_snap %p follows %lld tid %lld !="
3545 " %lld\n", capsnap, follows,
3546 flush_tid, capsnap->cap_flush.tid);
3552 dout(" skipping cap_snap %p follows %lld\n",
3553 capsnap, capsnap->follows);
3557 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3558 dout(" removing %p cap_snap %p follows %lld\n",
3559 inode, capsnap, follows);
3560 list_del(&capsnap->ci_item);
3561 if (__finish_cap_flush(NULL, ci, &capsnap->cap_flush))
3564 spin_lock(&mdsc->cap_dirty_lock);
3566 if (list_empty(&ci->i_cap_flush_list))
3567 list_del_init(&ci->i_flushing_item);
3569 if (__finish_cap_flush(mdsc, NULL, &capsnap->cap_flush))
3572 spin_unlock(&mdsc->cap_dirty_lock);
3574 spin_unlock(&ci->i_ceph_lock);
3576 ceph_put_snap_context(capsnap->context);
3577 ceph_put_cap_snap(capsnap);
3579 wake_up_all(&ci->i_cap_wq);
3581 wake_up_all(&mdsc->cap_flushing_wq);
3587 * Handle TRUNC from MDS, indicating file truncation.
3589 * caller hold s_mutex.
3591 static void handle_cap_trunc(struct inode *inode,
3592 struct ceph_mds_caps *trunc,
3593 struct ceph_mds_session *session)
3594 __releases(ci->i_ceph_lock)
3596 struct ceph_inode_info *ci = ceph_inode(inode);
3597 int mds = session->s_mds;
3598 int seq = le32_to_cpu(trunc->seq);
3599 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3600 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3601 u64 size = le64_to_cpu(trunc->size);
3602 int implemented = 0;
3603 int dirty = __ceph_caps_dirty(ci);
3604 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3605 int queue_trunc = 0;
3607 issued |= implemented | dirty;
3609 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3610 inode, mds, seq, truncate_size, truncate_seq);
3611 queue_trunc = ceph_fill_file_size(inode, issued,
3612 truncate_seq, truncate_size, size);
3613 spin_unlock(&ci->i_ceph_lock);
3616 ceph_queue_vmtruncate(inode);
3620 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3621 * different one. If we are the most recent migration we've seen (as
3622 * indicated by mseq), make note of the migrating cap bits for the
3623 * duration (until we see the corresponding IMPORT).
3625 * caller holds s_mutex
3627 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3628 struct ceph_mds_cap_peer *ph,
3629 struct ceph_mds_session *session)
3631 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3632 struct ceph_mds_session *tsession = NULL;
3633 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3634 struct ceph_inode_info *ci = ceph_inode(inode);
3636 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3637 unsigned t_seq, t_mseq;
3639 int mds = session->s_mds;
3642 t_cap_id = le64_to_cpu(ph->cap_id);
3643 t_seq = le32_to_cpu(ph->seq);
3644 t_mseq = le32_to_cpu(ph->mseq);
3645 target = le32_to_cpu(ph->mds);
3647 t_cap_id = t_seq = t_mseq = 0;
3651 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3652 inode, ci, mds, mseq, target);
3654 spin_lock(&ci->i_ceph_lock);
3655 cap = __get_cap_for_mds(ci, mds);
3656 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3660 if (cap->mds_wanted | cap->issued)
3661 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
3662 __ceph_remove_cap(cap, false);
3667 * now we know we haven't received the cap import message yet
3668 * because the exported cap still exist.
3671 issued = cap->issued;
3672 if (issued != cap->implemented)
3673 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3674 "ino (%llx.%llx) mds%d seq %d mseq %d "
3675 "issued %s implemented %s\n",
3676 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3677 ceph_cap_string(issued),
3678 ceph_cap_string(cap->implemented));
3681 tcap = __get_cap_for_mds(ci, target);
3683 /* already have caps from the target */
3684 if (tcap->cap_id == t_cap_id &&
3685 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3686 dout(" updating import cap %p mds%d\n", tcap, target);
3687 tcap->cap_id = t_cap_id;
3688 tcap->seq = t_seq - 1;
3689 tcap->issue_seq = t_seq - 1;
3690 tcap->issued |= issued;
3691 tcap->implemented |= issued;
3692 if (cap == ci->i_auth_cap)
3693 ci->i_auth_cap = tcap;
3695 if (!list_empty(&ci->i_cap_flush_list) &&
3696 ci->i_auth_cap == tcap) {
3697 spin_lock(&mdsc->cap_dirty_lock);
3698 list_move_tail(&ci->i_flushing_item,
3699 &tcap->session->s_cap_flushing);
3700 spin_unlock(&mdsc->cap_dirty_lock);
3703 __ceph_remove_cap(cap, false);
3705 } else if (tsession) {
3706 /* add placeholder for the export tagert */
3707 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3709 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3710 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3712 if (!list_empty(&ci->i_cap_flush_list) &&
3713 ci->i_auth_cap == tcap) {
3714 spin_lock(&mdsc->cap_dirty_lock);
3715 list_move_tail(&ci->i_flushing_item,
3716 &tcap->session->s_cap_flushing);
3717 spin_unlock(&mdsc->cap_dirty_lock);
3720 __ceph_remove_cap(cap, false);
3724 spin_unlock(&ci->i_ceph_lock);
3725 mutex_unlock(&session->s_mutex);
3727 /* open target session */
3728 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3729 if (!IS_ERR(tsession)) {
3731 mutex_lock(&session->s_mutex);
3732 mutex_lock_nested(&tsession->s_mutex,
3733 SINGLE_DEPTH_NESTING);
3735 mutex_lock(&tsession->s_mutex);
3736 mutex_lock_nested(&session->s_mutex,
3737 SINGLE_DEPTH_NESTING);
3739 new_cap = ceph_get_cap(mdsc, NULL);
3744 mutex_lock(&session->s_mutex);
3749 spin_unlock(&ci->i_ceph_lock);
3750 mutex_unlock(&session->s_mutex);
3752 mutex_unlock(&tsession->s_mutex);
3753 ceph_put_mds_session(tsession);
3756 ceph_put_cap(mdsc, new_cap);
3760 * Handle cap IMPORT.
3762 * caller holds s_mutex. acquires i_ceph_lock
3764 static void handle_cap_import(struct ceph_mds_client *mdsc,
3765 struct inode *inode, struct ceph_mds_caps *im,
3766 struct ceph_mds_cap_peer *ph,
3767 struct ceph_mds_session *session,
3768 struct ceph_cap **target_cap, int *old_issued)
3769 __acquires(ci->i_ceph_lock)
3771 struct ceph_inode_info *ci = ceph_inode(inode);
3772 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3773 int mds = session->s_mds;
3775 unsigned caps = le32_to_cpu(im->caps);
3776 unsigned wanted = le32_to_cpu(im->wanted);
3777 unsigned seq = le32_to_cpu(im->seq);
3778 unsigned mseq = le32_to_cpu(im->migrate_seq);
3779 u64 realmino = le64_to_cpu(im->realm);
3780 u64 cap_id = le64_to_cpu(im->cap_id);
3785 p_cap_id = le64_to_cpu(ph->cap_id);
3786 peer = le32_to_cpu(ph->mds);
3792 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3793 inode, ci, mds, mseq, peer);
3796 spin_lock(&ci->i_ceph_lock);
3797 cap = __get_cap_for_mds(ci, mds);
3800 spin_unlock(&ci->i_ceph_lock);
3801 new_cap = ceph_get_cap(mdsc, NULL);
3807 ceph_put_cap(mdsc, new_cap);
3812 __ceph_caps_issued(ci, &issued);
3813 issued |= __ceph_caps_dirty(ci);
3815 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3816 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3818 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3819 if (ocap && ocap->cap_id == p_cap_id) {
3820 dout(" remove export cap %p mds%d flags %d\n",
3821 ocap, peer, ph->flags);
3822 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3823 (ocap->seq != le32_to_cpu(ph->seq) ||
3824 ocap->mseq != le32_to_cpu(ph->mseq))) {
3825 pr_err_ratelimited("handle_cap_import: "
3826 "mismatched seq/mseq: ino (%llx.%llx) "
3827 "mds%d seq %d mseq %d importer mds%d "
3828 "has peer seq %d mseq %d\n",
3829 ceph_vinop(inode), peer, ocap->seq,
3830 ocap->mseq, mds, le32_to_cpu(ph->seq),
3831 le32_to_cpu(ph->mseq));
3833 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3836 /* make sure we re-request max_size, if necessary */
3837 ci->i_requested_max_size = 0;
3839 *old_issued = issued;
3844 * Handle a caps message from the MDS.
3846 * Identify the appropriate session, inode, and call the right handler
3847 * based on the cap op.
3849 void ceph_handle_caps(struct ceph_mds_session *session,
3850 struct ceph_msg *msg)
3852 struct ceph_mds_client *mdsc = session->s_mdsc;
3853 struct inode *inode;
3854 struct ceph_inode_info *ci;
3855 struct ceph_cap *cap;
3856 struct ceph_mds_caps *h;
3857 struct ceph_mds_cap_peer *peer = NULL;
3858 struct ceph_snap_realm *realm = NULL;
3860 int msg_version = le16_to_cpu(msg->hdr.version);
3862 struct ceph_vino vino;
3864 size_t snaptrace_len;
3866 struct cap_extra_info extra_info = {};
3868 dout("handle_caps from mds%d\n", session->s_mds);
3871 end = msg->front.iov_base + msg->front.iov_len;
3872 if (msg->front.iov_len < sizeof(*h))
3874 h = msg->front.iov_base;
3875 op = le32_to_cpu(h->op);
3876 vino.ino = le64_to_cpu(h->ino);
3877 vino.snap = CEPH_NOSNAP;
3878 seq = le32_to_cpu(h->seq);
3879 mseq = le32_to_cpu(h->migrate_seq);
3882 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3883 p = snaptrace + snaptrace_len;
3885 if (msg_version >= 2) {
3887 ceph_decode_32_safe(&p, end, flock_len, bad);
3888 if (p + flock_len > end)
3893 if (msg_version >= 3) {
3894 if (op == CEPH_CAP_OP_IMPORT) {
3895 if (p + sizeof(*peer) > end)
3899 } else if (op == CEPH_CAP_OP_EXPORT) {
3900 /* recorded in unused fields */
3901 peer = (void *)&h->size;
3905 if (msg_version >= 4) {
3906 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
3907 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
3908 if (p + extra_info.inline_len > end)
3910 extra_info.inline_data = p;
3911 p += extra_info.inline_len;
3914 if (msg_version >= 5) {
3915 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
3918 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
3919 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
3922 if (msg_version >= 8) {
3924 u32 caller_uid, caller_gid;
3928 ceph_decode_64_safe(&p, end, flush_tid, bad);
3930 ceph_decode_32_safe(&p, end, caller_uid, bad);
3931 ceph_decode_32_safe(&p, end, caller_gid, bad);
3933 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
3934 if (pool_ns_len > 0) {
3935 ceph_decode_need(&p, end, pool_ns_len, bad);
3936 extra_info.pool_ns =
3937 ceph_find_or_create_string(p, pool_ns_len);
3942 if (msg_version >= 9) {
3943 struct ceph_timespec *btime;
3945 if (p + sizeof(*btime) > end)
3948 ceph_decode_timespec64(&extra_info.btime, btime);
3949 p += sizeof(*btime);
3950 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
3953 if (msg_version >= 11) {
3956 ceph_decode_32_safe(&p, end, flags, bad);
3958 extra_info.dirstat_valid = true;
3959 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
3960 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
3964 inode = ceph_find_inode(mdsc->fsc->sb, vino);
3965 ci = ceph_inode(inode);
3966 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3969 mutex_lock(&session->s_mutex);
3971 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3975 dout(" i don't have ino %llx\n", vino.ino);
3977 if (op == CEPH_CAP_OP_IMPORT) {
3978 cap = ceph_get_cap(mdsc, NULL);
3979 cap->cap_ino = vino.ino;
3980 cap->queue_release = 1;
3981 cap->cap_id = le64_to_cpu(h->cap_id);
3984 cap->issue_seq = seq;
3985 spin_lock(&session->s_cap_lock);
3986 __ceph_queue_cap_release(session, cap);
3987 spin_unlock(&session->s_cap_lock);
3989 goto flush_cap_releases;
3992 /* these will work even if we don't have a cap yet */
3994 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3995 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
3999 case CEPH_CAP_OP_EXPORT:
4000 handle_cap_export(inode, h, peer, session);
4003 case CEPH_CAP_OP_IMPORT:
4005 if (snaptrace_len) {
4006 down_write(&mdsc->snap_rwsem);
4007 ceph_update_snap_trace(mdsc, snaptrace,
4008 snaptrace + snaptrace_len,
4010 downgrade_write(&mdsc->snap_rwsem);
4012 down_read(&mdsc->snap_rwsem);
4014 handle_cap_import(mdsc, inode, h, peer, session,
4015 &cap, &extra_info.issued);
4016 handle_cap_grant(inode, session, cap,
4017 h, msg->middle, &extra_info);
4019 ceph_put_snap_realm(mdsc, realm);
4023 /* the rest require a cap */
4024 spin_lock(&ci->i_ceph_lock);
4025 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
4027 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4028 inode, ceph_ino(inode), ceph_snap(inode),
4030 spin_unlock(&ci->i_ceph_lock);
4031 goto flush_cap_releases;
4034 /* note that each of these drops i_ceph_lock for us */
4036 case CEPH_CAP_OP_REVOKE:
4037 case CEPH_CAP_OP_GRANT:
4038 __ceph_caps_issued(ci, &extra_info.issued);
4039 extra_info.issued |= __ceph_caps_dirty(ci);
4040 handle_cap_grant(inode, session, cap,
4041 h, msg->middle, &extra_info);
4044 case CEPH_CAP_OP_FLUSH_ACK:
4045 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4049 case CEPH_CAP_OP_TRUNC:
4050 handle_cap_trunc(inode, h, session);
4054 spin_unlock(&ci->i_ceph_lock);
4055 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4056 ceph_cap_op_name(op));
4060 mutex_unlock(&session->s_mutex);
4062 ceph_put_string(extra_info.pool_ns);
4063 /* avoid calling iput_final() in mds dispatch threads */
4064 ceph_async_iput(inode);
4069 * send any cap release message to try to move things
4070 * along for the mds (who clearly thinks we still have this
4073 ceph_flush_cap_releases(mdsc, session);
4077 pr_err("ceph_handle_caps: corrupt message\n");
4083 * Delayed work handler to process end of delayed cap release LRU list.
4085 * If new caps are added to the list while processing it, these won't get
4086 * processed in this run. In this case, the ci->i_hold_caps_max will be
4087 * returned so that the work can be scheduled accordingly.
4089 unsigned long ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4091 struct inode *inode;
4092 struct ceph_inode_info *ci;
4093 int flags = CHECK_CAPS_NODELAY;
4094 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
4095 unsigned long delay_max = opt->caps_wanted_delay_max * HZ;
4096 unsigned long loop_start = jiffies;
4097 unsigned long delay = 0;
4099 dout("check_delayed_caps\n");
4101 spin_lock(&mdsc->cap_delay_lock);
4102 if (list_empty(&mdsc->cap_delay_list))
4104 ci = list_first_entry(&mdsc->cap_delay_list,
4105 struct ceph_inode_info,
4107 if (time_before(loop_start, ci->i_hold_caps_max - delay_max)) {
4108 dout("%s caps added recently. Exiting loop", __func__);
4109 delay = ci->i_hold_caps_max;
4112 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4113 time_before(jiffies, ci->i_hold_caps_max))
4115 list_del_init(&ci->i_cap_delay_list);
4117 inode = igrab(&ci->vfs_inode);
4118 spin_unlock(&mdsc->cap_delay_lock);
4121 dout("check_delayed_caps on %p\n", inode);
4122 ceph_check_caps(ci, flags, NULL);
4123 /* avoid calling iput_final() in tick thread */
4124 ceph_async_iput(inode);
4127 spin_unlock(&mdsc->cap_delay_lock);
4133 * Flush all dirty caps to the mds
4135 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4137 struct ceph_inode_info *ci;
4138 struct inode *inode;
4140 dout("flush_dirty_caps\n");
4141 spin_lock(&mdsc->cap_dirty_lock);
4142 while (!list_empty(&mdsc->cap_dirty)) {
4143 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
4145 inode = &ci->vfs_inode;
4147 dout("flush_dirty_caps %p\n", inode);
4148 spin_unlock(&mdsc->cap_dirty_lock);
4149 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
4151 spin_lock(&mdsc->cap_dirty_lock);
4153 spin_unlock(&mdsc->cap_dirty_lock);
4154 dout("flush_dirty_caps done\n");
4157 void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode)
4160 int bits = (fmode << 1) | 1;
4161 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4162 if (bits & (1 << i))
4163 ci->i_nr_by_mode[i]++;
4168 * Drop open file reference. If we were the last open file,
4169 * we may need to release capabilities to the MDS (or schedule
4170 * their delayed release).
4172 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
4175 int bits = (fmode << 1) | 1;
4176 spin_lock(&ci->i_ceph_lock);
4177 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4178 if (bits & (1 << i)) {
4179 BUG_ON(ci->i_nr_by_mode[i] == 0);
4180 if (--ci->i_nr_by_mode[i] == 0)
4184 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
4185 &ci->vfs_inode, fmode,
4186 ci->i_nr_by_mode[0], ci->i_nr_by_mode[1],
4187 ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]);
4188 spin_unlock(&ci->i_ceph_lock);
4190 if (last && ci->i_vino.snap == CEPH_NOSNAP)
4191 ceph_check_caps(ci, 0, NULL);
4195 * For a soon-to-be unlinked file, drop the LINK caps. If it
4196 * looks like the link count will hit 0, drop any other caps (other
4197 * than PIN) we don't specifically want (due to the file still being
4200 int ceph_drop_caps_for_unlink(struct inode *inode)
4202 struct ceph_inode_info *ci = ceph_inode(inode);
4203 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4205 spin_lock(&ci->i_ceph_lock);
4206 if (inode->i_nlink == 1) {
4207 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4209 ci->i_ceph_flags |= CEPH_I_NODELAY;
4210 if (__ceph_caps_dirty(ci)) {
4211 struct ceph_mds_client *mdsc =
4212 ceph_inode_to_client(inode)->mdsc;
4213 __cap_delay_requeue_front(mdsc, ci);
4216 spin_unlock(&ci->i_ceph_lock);
4221 * Helpers for embedding cap and dentry lease releases into mds
4224 * @force is used by dentry_release (below) to force inclusion of a
4225 * record for the directory inode, even when there aren't any caps to
4228 int ceph_encode_inode_release(void **p, struct inode *inode,
4229 int mds, int drop, int unless, int force)
4231 struct ceph_inode_info *ci = ceph_inode(inode);
4232 struct ceph_cap *cap;
4233 struct ceph_mds_request_release *rel = *p;
4237 spin_lock(&ci->i_ceph_lock);
4238 used = __ceph_caps_used(ci);
4239 dirty = __ceph_caps_dirty(ci);
4241 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4242 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4243 ceph_cap_string(unless));
4245 /* only drop unused, clean caps */
4246 drop &= ~(used | dirty);
4248 cap = __get_cap_for_mds(ci, mds);
4249 if (cap && __cap_is_valid(cap)) {
4250 unless &= cap->issued;
4252 if (unless & CEPH_CAP_AUTH_EXCL)
4253 drop &= ~CEPH_CAP_AUTH_SHARED;
4254 if (unless & CEPH_CAP_LINK_EXCL)
4255 drop &= ~CEPH_CAP_LINK_SHARED;
4256 if (unless & CEPH_CAP_XATTR_EXCL)
4257 drop &= ~CEPH_CAP_XATTR_SHARED;
4258 if (unless & CEPH_CAP_FILE_EXCL)
4259 drop &= ~CEPH_CAP_FILE_SHARED;
4262 if (force || (cap->issued & drop)) {
4263 if (cap->issued & drop) {
4264 int wanted = __ceph_caps_wanted(ci);
4265 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
4266 wanted |= cap->mds_wanted;
4267 dout("encode_inode_release %p cap %p "
4268 "%s -> %s, wanted %s -> %s\n", inode, cap,
4269 ceph_cap_string(cap->issued),
4270 ceph_cap_string(cap->issued & ~drop),
4271 ceph_cap_string(cap->mds_wanted),
4272 ceph_cap_string(wanted));
4274 cap->issued &= ~drop;
4275 cap->implemented &= ~drop;
4276 cap->mds_wanted = wanted;
4278 dout("encode_inode_release %p cap %p %s"
4279 " (force)\n", inode, cap,
4280 ceph_cap_string(cap->issued));
4283 rel->ino = cpu_to_le64(ceph_ino(inode));
4284 rel->cap_id = cpu_to_le64(cap->cap_id);
4285 rel->seq = cpu_to_le32(cap->seq);
4286 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4287 rel->mseq = cpu_to_le32(cap->mseq);
4288 rel->caps = cpu_to_le32(cap->implemented);
4289 rel->wanted = cpu_to_le32(cap->mds_wanted);
4295 dout("encode_inode_release %p cap %p %s (noop)\n",
4296 inode, cap, ceph_cap_string(cap->issued));
4299 spin_unlock(&ci->i_ceph_lock);
4303 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4305 int mds, int drop, int unless)
4307 struct dentry *parent = NULL;
4308 struct ceph_mds_request_release *rel = *p;
4309 struct ceph_dentry_info *di = ceph_dentry(dentry);
4314 * force an record for the directory caps if we have a dentry lease.
4315 * this is racy (can't take i_ceph_lock and d_lock together), but it
4316 * doesn't have to be perfect; the mds will revoke anything we don't
4319 spin_lock(&dentry->d_lock);
4320 if (di->lease_session && di->lease_session->s_mds == mds)
4323 parent = dget(dentry->d_parent);
4324 dir = d_inode(parent);
4326 spin_unlock(&dentry->d_lock);
4328 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4331 spin_lock(&dentry->d_lock);
4332 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4333 dout("encode_dentry_release %p mds%d seq %d\n",
4334 dentry, mds, (int)di->lease_seq);
4335 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4336 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4337 *p += dentry->d_name.len;
4338 rel->dname_seq = cpu_to_le32(di->lease_seq);
4339 __ceph_mdsc_drop_dentry_lease(dentry);
4341 spin_unlock(&dentry->d_lock);