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 int ceph_is_any_caps(struct inode *inode)
1016 struct ceph_inode_info *ci = ceph_inode(inode);
1019 spin_lock(&ci->i_ceph_lock);
1020 ret = __ceph_is_any_real_caps(ci);
1021 spin_unlock(&ci->i_ceph_lock);
1026 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
1028 struct ceph_snap_realm *realm = ci->i_snap_realm;
1029 spin_lock(&realm->inodes_with_caps_lock);
1030 list_del_init(&ci->i_snap_realm_item);
1031 ci->i_snap_realm_counter++;
1032 ci->i_snap_realm = NULL;
1033 if (realm->ino == ci->i_vino.ino)
1034 realm->inode = NULL;
1035 spin_unlock(&realm->inodes_with_caps_lock);
1036 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
1041 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
1043 * caller should hold i_ceph_lock.
1044 * caller will not hold session s_mutex if called from destroy_inode.
1046 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1048 struct ceph_mds_session *session = cap->session;
1049 struct ceph_inode_info *ci = cap->ci;
1050 struct ceph_mds_client *mdsc =
1051 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1054 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1056 /* remove from inode's cap rbtree, and clear auth cap */
1057 rb_erase(&cap->ci_node, &ci->i_caps);
1058 if (ci->i_auth_cap == cap)
1059 ci->i_auth_cap = NULL;
1061 /* remove from session list */
1062 spin_lock(&session->s_cap_lock);
1063 if (session->s_cap_iterator == cap) {
1064 /* not yet, we are iterating over this very cap */
1065 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1068 list_del_init(&cap->session_caps);
1069 session->s_nr_caps--;
1070 cap->session = NULL;
1073 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1077 * s_cap_reconnect is protected by s_cap_lock. no one changes
1078 * s_cap_gen while session is in the reconnect state.
1080 if (queue_release &&
1081 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
1082 cap->queue_release = 1;
1084 __ceph_queue_cap_release(session, cap);
1088 cap->queue_release = 0;
1090 cap->cap_ino = ci->i_vino.ino;
1092 spin_unlock(&session->s_cap_lock);
1095 ceph_put_cap(mdsc, cap);
1097 if (!__ceph_is_any_real_caps(ci)) {
1098 /* when reconnect denied, we remove session caps forcibly,
1099 * i_wr_ref can be non-zero. If there are ongoing write,
1100 * keep i_snap_realm.
1102 if (ci->i_wr_ref == 0 && ci->i_snap_realm)
1103 drop_inode_snap_realm(ci);
1105 __cap_delay_cancel(mdsc, ci);
1109 struct cap_msg_args {
1110 struct ceph_mds_session *session;
1111 u64 ino, cid, follows;
1112 u64 flush_tid, oldest_flush_tid, size, max_size;
1115 struct ceph_buffer *xattr_buf;
1116 struct timespec64 atime, mtime, ctime, btime;
1117 int op, caps, wanted, dirty;
1118 u32 seq, issue_seq, mseq, time_warp_seq;
1127 * Build and send a cap message to the given MDS.
1129 * Caller should be holding s_mutex.
1131 static int send_cap_msg(struct cap_msg_args *arg)
1133 struct ceph_mds_caps *fc;
1134 struct ceph_msg *msg;
1137 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1139 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1140 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1141 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(arg->op),
1142 arg->cid, arg->ino, ceph_cap_string(arg->caps),
1143 ceph_cap_string(arg->wanted), ceph_cap_string(arg->dirty),
1144 arg->seq, arg->issue_seq, arg->flush_tid, arg->oldest_flush_tid,
1145 arg->mseq, arg->follows, arg->size, arg->max_size,
1147 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1149 /* flock buffer size + inline version + inline data size +
1150 * osd_epoch_barrier + oldest_flush_tid */
1151 extra_len = 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4;
1152 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1157 msg->hdr.version = cpu_to_le16(10);
1158 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1160 fc = msg->front.iov_base;
1161 memset(fc, 0, sizeof(*fc));
1163 fc->cap_id = cpu_to_le64(arg->cid);
1164 fc->op = cpu_to_le32(arg->op);
1165 fc->seq = cpu_to_le32(arg->seq);
1166 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1167 fc->migrate_seq = cpu_to_le32(arg->mseq);
1168 fc->caps = cpu_to_le32(arg->caps);
1169 fc->wanted = cpu_to_le32(arg->wanted);
1170 fc->dirty = cpu_to_le32(arg->dirty);
1171 fc->ino = cpu_to_le64(arg->ino);
1172 fc->snap_follows = cpu_to_le64(arg->follows);
1174 fc->size = cpu_to_le64(arg->size);
1175 fc->max_size = cpu_to_le64(arg->max_size);
1176 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1177 ceph_encode_timespec64(&fc->atime, &arg->atime);
1178 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1179 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1181 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1182 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1183 fc->mode = cpu_to_le32(arg->mode);
1185 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1186 if (arg->xattr_buf) {
1187 msg->middle = ceph_buffer_get(arg->xattr_buf);
1188 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1189 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1193 /* flock buffer size (version 2) */
1194 ceph_encode_32(&p, 0);
1195 /* inline version (version 4) */
1196 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1197 /* inline data size */
1198 ceph_encode_32(&p, 0);
1200 * osd_epoch_barrier (version 5)
1201 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1202 * case it was recently changed
1204 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1205 /* oldest_flush_tid (version 6) */
1206 ceph_encode_64(&p, arg->oldest_flush_tid);
1209 * caller_uid/caller_gid (version 7)
1211 * Currently, we don't properly track which caller dirtied the caps
1212 * last, and force a flush of them when there is a conflict. For now,
1213 * just set this to 0:0, to emulate how the MDS has worked up to now.
1215 ceph_encode_32(&p, 0);
1216 ceph_encode_32(&p, 0);
1218 /* pool namespace (version 8) (mds always ignores this) */
1219 ceph_encode_32(&p, 0);
1221 /* btime and change_attr (version 9) */
1222 ceph_encode_timespec64(p, &arg->btime);
1223 p += sizeof(struct ceph_timespec);
1224 ceph_encode_64(&p, arg->change_attr);
1226 /* Advisory flags (version 10) */
1227 ceph_encode_32(&p, arg->flags);
1229 ceph_con_send(&arg->session->s_con, msg);
1234 * Queue cap releases when an inode is dropped from our cache.
1236 void __ceph_remove_caps(struct ceph_inode_info *ci)
1240 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1241 * may call __ceph_caps_issued_mask() on a freeing inode. */
1242 spin_lock(&ci->i_ceph_lock);
1243 p = rb_first(&ci->i_caps);
1245 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1247 __ceph_remove_cap(cap, true);
1249 spin_unlock(&ci->i_ceph_lock);
1253 * Send a cap msg on the given inode. Update our caps state, then
1254 * drop i_ceph_lock and send the message.
1256 * Make note of max_size reported/requested from mds, revoked caps
1257 * that have now been implemented.
1259 * Return non-zero if delayed release, or we experienced an error
1260 * such that the caller should requeue + retry later.
1262 * called with i_ceph_lock, then drops it.
1263 * caller should hold snap_rwsem (read), s_mutex.
1265 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1266 int op, int flags, int used, int want, int retain,
1267 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1268 __releases(cap->ci->i_ceph_lock)
1270 struct ceph_inode_info *ci = cap->ci;
1271 struct inode *inode = &ci->vfs_inode;
1272 struct ceph_buffer *old_blob = NULL;
1273 struct cap_msg_args arg;
1279 held = cap->issued | cap->implemented;
1280 revoking = cap->implemented & ~cap->issued;
1281 retain &= ~revoking;
1283 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1284 inode, cap, cap->session,
1285 ceph_cap_string(held), ceph_cap_string(held & retain),
1286 ceph_cap_string(revoking));
1287 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1289 arg.session = cap->session;
1291 /* don't release wanted unless we've waited a bit. */
1292 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1293 time_before(jiffies, ci->i_hold_caps_min)) {
1294 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1295 ceph_cap_string(cap->issued),
1296 ceph_cap_string(cap->issued & retain),
1297 ceph_cap_string(cap->mds_wanted),
1298 ceph_cap_string(want));
1299 want |= cap->mds_wanted;
1300 retain |= cap->issued;
1303 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1304 if (want & ~cap->mds_wanted) {
1305 /* user space may open/close single file frequently.
1306 * This avoids droping mds_wanted immediately after
1307 * requesting new mds_wanted.
1309 __cap_set_timeouts(mdsc, ci);
1312 cap->issued &= retain; /* drop bits we don't want */
1313 if (cap->implemented & ~cap->issued) {
1315 * Wake up any waiters on wanted -> needed transition.
1316 * This is due to the weird transition from buffered
1317 * to sync IO... we need to flush dirty pages _before_
1318 * allowing sync writes to avoid reordering.
1322 cap->implemented &= cap->issued | used;
1323 cap->mds_wanted = want;
1325 arg.ino = ceph_vino(inode).ino;
1326 arg.cid = cap->cap_id;
1327 arg.follows = flushing ? ci->i_head_snapc->seq : 0;
1328 arg.flush_tid = flush_tid;
1329 arg.oldest_flush_tid = oldest_flush_tid;
1331 arg.size = inode->i_size;
1332 ci->i_reported_size = arg.size;
1333 arg.max_size = ci->i_wanted_max_size;
1334 ci->i_requested_max_size = arg.max_size;
1336 if (flushing & CEPH_CAP_XATTR_EXCL) {
1337 old_blob = __ceph_build_xattrs_blob(ci);
1338 arg.xattr_version = ci->i_xattrs.version;
1339 arg.xattr_buf = ci->i_xattrs.blob;
1341 arg.xattr_buf = NULL;
1344 arg.mtime = inode->i_mtime;
1345 arg.atime = inode->i_atime;
1346 arg.ctime = inode->i_ctime;
1347 arg.btime = ci->i_btime;
1348 arg.change_attr = inode_peek_iversion_raw(inode);
1351 arg.caps = cap->implemented;
1353 arg.dirty = flushing;
1356 arg.issue_seq = cap->issue_seq;
1357 arg.mseq = cap->mseq;
1358 arg.time_warp_seq = ci->i_time_warp_seq;
1360 arg.uid = inode->i_uid;
1361 arg.gid = inode->i_gid;
1362 arg.mode = inode->i_mode;
1364 arg.inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1365 if (!(flags & CEPH_CLIENT_CAPS_PENDING_CAPSNAP) &&
1366 !list_empty(&ci->i_cap_snaps)) {
1367 struct ceph_cap_snap *capsnap;
1368 list_for_each_entry_reverse(capsnap, &ci->i_cap_snaps, ci_item) {
1369 if (capsnap->cap_flush.tid)
1371 if (capsnap->need_flush) {
1372 flags |= CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1379 spin_unlock(&ci->i_ceph_lock);
1381 ceph_buffer_put(old_blob);
1383 ret = send_cap_msg(&arg);
1385 dout("error sending cap msg, must requeue %p\n", inode);
1390 wake_up_all(&ci->i_cap_wq);
1395 static inline int __send_flush_snap(struct inode *inode,
1396 struct ceph_mds_session *session,
1397 struct ceph_cap_snap *capsnap,
1398 u32 mseq, u64 oldest_flush_tid)
1400 struct cap_msg_args arg;
1402 arg.session = session;
1403 arg.ino = ceph_vino(inode).ino;
1405 arg.follows = capsnap->follows;
1406 arg.flush_tid = capsnap->cap_flush.tid;
1407 arg.oldest_flush_tid = oldest_flush_tid;
1409 arg.size = capsnap->size;
1411 arg.xattr_version = capsnap->xattr_version;
1412 arg.xattr_buf = capsnap->xattr_blob;
1414 arg.atime = capsnap->atime;
1415 arg.mtime = capsnap->mtime;
1416 arg.ctime = capsnap->ctime;
1417 arg.btime = capsnap->btime;
1418 arg.change_attr = capsnap->change_attr;
1420 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1421 arg.caps = capsnap->issued;
1423 arg.dirty = capsnap->dirty;
1428 arg.time_warp_seq = capsnap->time_warp_seq;
1430 arg.uid = capsnap->uid;
1431 arg.gid = capsnap->gid;
1432 arg.mode = capsnap->mode;
1434 arg.inline_data = capsnap->inline_data;
1437 return send_cap_msg(&arg);
1441 * When a snapshot is taken, clients accumulate dirty metadata on
1442 * inodes with capabilities in ceph_cap_snaps to describe the file
1443 * state at the time the snapshot was taken. This must be flushed
1444 * asynchronously back to the MDS once sync writes complete and dirty
1445 * data is written out.
1447 * Called under i_ceph_lock. Takes s_mutex as needed.
1449 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1450 struct ceph_mds_session *session)
1451 __releases(ci->i_ceph_lock)
1452 __acquires(ci->i_ceph_lock)
1454 struct inode *inode = &ci->vfs_inode;
1455 struct ceph_mds_client *mdsc = session->s_mdsc;
1456 struct ceph_cap_snap *capsnap;
1457 u64 oldest_flush_tid = 0;
1458 u64 first_tid = 1, last_tid = 0;
1460 dout("__flush_snaps %p session %p\n", inode, session);
1462 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1464 * we need to wait for sync writes to complete and for dirty
1465 * pages to be written out.
1467 if (capsnap->dirty_pages || capsnap->writing)
1470 /* should be removed by ceph_try_drop_cap_snap() */
1471 BUG_ON(!capsnap->need_flush);
1473 /* only flush each capsnap once */
1474 if (capsnap->cap_flush.tid > 0) {
1475 dout(" already flushed %p, skipping\n", capsnap);
1479 spin_lock(&mdsc->cap_dirty_lock);
1480 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1481 list_add_tail(&capsnap->cap_flush.g_list,
1482 &mdsc->cap_flush_list);
1483 if (oldest_flush_tid == 0)
1484 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1485 if (list_empty(&ci->i_flushing_item)) {
1486 list_add_tail(&ci->i_flushing_item,
1487 &session->s_cap_flushing);
1489 spin_unlock(&mdsc->cap_dirty_lock);
1491 list_add_tail(&capsnap->cap_flush.i_list,
1492 &ci->i_cap_flush_list);
1495 first_tid = capsnap->cap_flush.tid;
1496 last_tid = capsnap->cap_flush.tid;
1499 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1501 while (first_tid <= last_tid) {
1502 struct ceph_cap *cap = ci->i_auth_cap;
1503 struct ceph_cap_flush *cf;
1506 if (!(cap && cap->session == session)) {
1507 dout("__flush_snaps %p auth cap %p not mds%d, "
1508 "stop\n", inode, cap, session->s_mds);
1513 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1514 if (cf->tid >= first_tid) {
1522 first_tid = cf->tid + 1;
1524 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1525 refcount_inc(&capsnap->nref);
1526 spin_unlock(&ci->i_ceph_lock);
1528 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1529 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1531 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1534 pr_err("__flush_snaps: error sending cap flushsnap, "
1535 "ino (%llx.%llx) tid %llu follows %llu\n",
1536 ceph_vinop(inode), cf->tid, capsnap->follows);
1539 ceph_put_cap_snap(capsnap);
1540 spin_lock(&ci->i_ceph_lock);
1544 void ceph_flush_snaps(struct ceph_inode_info *ci,
1545 struct ceph_mds_session **psession)
1547 struct inode *inode = &ci->vfs_inode;
1548 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1549 struct ceph_mds_session *session = NULL;
1552 dout("ceph_flush_snaps %p\n", inode);
1554 session = *psession;
1556 spin_lock(&ci->i_ceph_lock);
1557 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1558 dout(" no capsnap needs flush, doing nothing\n");
1561 if (!ci->i_auth_cap) {
1562 dout(" no auth cap (migrating?), doing nothing\n");
1566 mds = ci->i_auth_cap->session->s_mds;
1567 if (session && session->s_mds != mds) {
1568 dout(" oops, wrong session %p mutex\n", session);
1569 mutex_unlock(&session->s_mutex);
1570 ceph_put_mds_session(session);
1574 spin_unlock(&ci->i_ceph_lock);
1575 mutex_lock(&mdsc->mutex);
1576 session = __ceph_lookup_mds_session(mdsc, mds);
1577 mutex_unlock(&mdsc->mutex);
1579 dout(" inverting session/ino locks on %p\n", session);
1580 mutex_lock(&session->s_mutex);
1585 // make sure flushsnap messages are sent in proper order.
1586 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
1587 __kick_flushing_caps(mdsc, session, ci, 0);
1589 __ceph_flush_snaps(ci, session);
1591 spin_unlock(&ci->i_ceph_lock);
1594 *psession = session;
1595 } else if (session) {
1596 mutex_unlock(&session->s_mutex);
1597 ceph_put_mds_session(session);
1599 /* we flushed them all; remove this inode from the queue */
1600 spin_lock(&mdsc->snap_flush_lock);
1601 list_del_init(&ci->i_snap_flush_item);
1602 spin_unlock(&mdsc->snap_flush_lock);
1606 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1607 * Caller is then responsible for calling __mark_inode_dirty with the
1608 * returned flags value.
1610 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1611 struct ceph_cap_flush **pcf)
1613 struct ceph_mds_client *mdsc =
1614 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1615 struct inode *inode = &ci->vfs_inode;
1616 int was = ci->i_dirty_caps;
1619 if (!ci->i_auth_cap) {
1620 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1621 "but no auth cap (session was closed?)\n",
1622 inode, ceph_ino(inode), ceph_cap_string(mask));
1626 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1627 ceph_cap_string(mask), ceph_cap_string(was),
1628 ceph_cap_string(was | mask));
1629 ci->i_dirty_caps |= mask;
1631 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1632 swap(ci->i_prealloc_cap_flush, *pcf);
1634 if (!ci->i_head_snapc) {
1635 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1636 ci->i_head_snapc = ceph_get_snap_context(
1637 ci->i_snap_realm->cached_context);
1639 dout(" inode %p now dirty snapc %p auth cap %p\n",
1640 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1641 BUG_ON(!list_empty(&ci->i_dirty_item));
1642 spin_lock(&mdsc->cap_dirty_lock);
1643 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1644 spin_unlock(&mdsc->cap_dirty_lock);
1645 if (ci->i_flushing_caps == 0) {
1647 dirty |= I_DIRTY_SYNC;
1650 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1652 BUG_ON(list_empty(&ci->i_dirty_item));
1653 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1654 (mask & CEPH_CAP_FILE_BUFFER))
1655 dirty |= I_DIRTY_DATASYNC;
1656 __cap_delay_requeue(mdsc, ci, true);
1660 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1662 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1665 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1668 kmem_cache_free(ceph_cap_flush_cachep, cf);
1671 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1673 if (!list_empty(&mdsc->cap_flush_list)) {
1674 struct ceph_cap_flush *cf =
1675 list_first_entry(&mdsc->cap_flush_list,
1676 struct ceph_cap_flush, g_list);
1683 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1684 * Return true if caller needs to wake up flush waiters.
1686 static bool __finish_cap_flush(struct ceph_mds_client *mdsc,
1687 struct ceph_inode_info *ci,
1688 struct ceph_cap_flush *cf)
1690 struct ceph_cap_flush *prev;
1691 bool wake = cf->wake;
1693 /* are there older pending cap flushes? */
1694 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1695 prev = list_prev_entry(cf, g_list);
1699 list_del(&cf->g_list);
1701 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1702 prev = list_prev_entry(cf, i_list);
1706 list_del(&cf->i_list);
1714 * Add dirty inode to the flushing list. Assigned a seq number so we
1715 * can wait for caps to flush without starving.
1717 * Called under i_ceph_lock. Returns the flush tid.
1719 static u64 __mark_caps_flushing(struct inode *inode,
1720 struct ceph_mds_session *session, bool wake,
1721 u64 *oldest_flush_tid)
1723 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1724 struct ceph_inode_info *ci = ceph_inode(inode);
1725 struct ceph_cap_flush *cf = NULL;
1728 BUG_ON(ci->i_dirty_caps == 0);
1729 BUG_ON(list_empty(&ci->i_dirty_item));
1730 BUG_ON(!ci->i_prealloc_cap_flush);
1732 flushing = ci->i_dirty_caps;
1733 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1734 ceph_cap_string(flushing),
1735 ceph_cap_string(ci->i_flushing_caps),
1736 ceph_cap_string(ci->i_flushing_caps | flushing));
1737 ci->i_flushing_caps |= flushing;
1738 ci->i_dirty_caps = 0;
1739 dout(" inode %p now !dirty\n", inode);
1741 swap(cf, ci->i_prealloc_cap_flush);
1742 cf->caps = flushing;
1745 spin_lock(&mdsc->cap_dirty_lock);
1746 list_del_init(&ci->i_dirty_item);
1748 cf->tid = ++mdsc->last_cap_flush_tid;
1749 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1750 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1752 if (list_empty(&ci->i_flushing_item)) {
1753 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1754 mdsc->num_cap_flushing++;
1756 spin_unlock(&mdsc->cap_dirty_lock);
1758 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1764 * try to invalidate mapping pages without blocking.
1766 static int try_nonblocking_invalidate(struct inode *inode)
1768 struct ceph_inode_info *ci = ceph_inode(inode);
1769 u32 invalidating_gen = ci->i_rdcache_gen;
1771 spin_unlock(&ci->i_ceph_lock);
1772 invalidate_mapping_pages(&inode->i_data, 0, -1);
1773 spin_lock(&ci->i_ceph_lock);
1775 if (inode->i_data.nrpages == 0 &&
1776 invalidating_gen == ci->i_rdcache_gen) {
1778 dout("try_nonblocking_invalidate %p success\n", inode);
1779 /* save any racing async invalidate some trouble */
1780 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1783 dout("try_nonblocking_invalidate %p failed\n", inode);
1787 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1789 loff_t size = ci->vfs_inode.i_size;
1790 /* mds will adjust max size according to the reported size */
1791 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1793 if (size >= ci->i_max_size)
1795 /* half of previous max_size increment has been used */
1796 if (ci->i_max_size > ci->i_reported_size &&
1797 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1803 * Swiss army knife function to examine currently used and wanted
1804 * versus held caps. Release, flush, ack revoked caps to mds as
1807 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1808 * cap release further.
1809 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1810 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1813 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1814 struct ceph_mds_session *session)
1816 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1817 struct ceph_mds_client *mdsc = fsc->mdsc;
1818 struct inode *inode = &ci->vfs_inode;
1819 struct ceph_cap *cap;
1820 u64 flush_tid, oldest_flush_tid;
1821 int file_wanted, used, cap_used;
1822 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1823 int issued, implemented, want, retain, revoking, flushing = 0;
1824 int mds = -1; /* keep track of how far we've gone through i_caps list
1825 to avoid an infinite loop on retry */
1827 int delayed = 0, sent = 0;
1828 bool no_delay = flags & CHECK_CAPS_NODELAY;
1829 bool queue_invalidate = false;
1830 bool tried_invalidate = false;
1832 /* if we are unmounting, flush any unused caps immediately. */
1836 spin_lock(&ci->i_ceph_lock);
1838 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1839 flags |= CHECK_CAPS_FLUSH;
1841 if (!(flags & CHECK_CAPS_AUTHONLY) ||
1842 (ci->i_auth_cap && __ceph_is_single_caps(ci)))
1843 __cap_delay_cancel(mdsc, ci);
1847 spin_lock(&ci->i_ceph_lock);
1849 file_wanted = __ceph_caps_file_wanted(ci);
1850 used = __ceph_caps_used(ci);
1851 issued = __ceph_caps_issued(ci, &implemented);
1852 revoking = implemented & ~issued;
1855 retain = file_wanted | used | CEPH_CAP_PIN;
1856 if (!mdsc->stopping && inode->i_nlink > 0) {
1858 retain |= CEPH_CAP_ANY; /* be greedy */
1859 } else if (S_ISDIR(inode->i_mode) &&
1860 (issued & CEPH_CAP_FILE_SHARED) &&
1861 __ceph_dir_is_complete(ci)) {
1863 * If a directory is complete, we want to keep
1864 * the exclusive cap. So that MDS does not end up
1865 * revoking the shared cap on every create/unlink
1868 if (IS_RDONLY(inode))
1869 want = CEPH_CAP_ANY_SHARED;
1871 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1875 retain |= CEPH_CAP_ANY_SHARED;
1877 * keep RD only if we didn't have the file open RW,
1878 * because then the mds would revoke it anyway to
1879 * journal max_size=0.
1881 if (ci->i_max_size == 0)
1882 retain |= CEPH_CAP_ANY_RD;
1886 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1887 " issued %s revoking %s retain %s %s%s%s\n", inode,
1888 ceph_cap_string(file_wanted),
1889 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1890 ceph_cap_string(ci->i_flushing_caps),
1891 ceph_cap_string(issued), ceph_cap_string(revoking),
1892 ceph_cap_string(retain),
1893 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1894 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1895 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1898 * If we no longer need to hold onto old our caps, and we may
1899 * have cached pages, but don't want them, then try to invalidate.
1900 * If we fail, it's because pages are locked.... try again later.
1902 if ((!no_delay || mdsc->stopping) &&
1903 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */
1904 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1905 inode->i_data.nrpages && /* have cached pages */
1906 (revoking & (CEPH_CAP_FILE_CACHE|
1907 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1908 !tried_invalidate) {
1909 dout("check_caps trying to invalidate on %p\n", inode);
1910 if (try_nonblocking_invalidate(inode) < 0) {
1911 dout("check_caps queuing invalidate\n");
1912 queue_invalidate = true;
1913 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1915 tried_invalidate = true;
1919 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1920 cap = rb_entry(p, struct ceph_cap, ci_node);
1922 /* avoid looping forever */
1923 if (mds >= cap->mds ||
1924 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1927 /* NOTE: no side-effects allowed, until we take s_mutex */
1930 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1931 cap_used &= ~ci->i_auth_cap->issued;
1933 revoking = cap->implemented & ~cap->issued;
1934 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1935 cap->mds, cap, ceph_cap_string(cap_used),
1936 ceph_cap_string(cap->issued),
1937 ceph_cap_string(cap->implemented),
1938 ceph_cap_string(revoking));
1940 if (cap == ci->i_auth_cap &&
1941 (cap->issued & CEPH_CAP_FILE_WR)) {
1942 /* request larger max_size from MDS? */
1943 if (ci->i_wanted_max_size > ci->i_max_size &&
1944 ci->i_wanted_max_size > ci->i_requested_max_size) {
1945 dout("requesting new max_size\n");
1949 /* approaching file_max? */
1950 if (__ceph_should_report_size(ci)) {
1951 dout("i_size approaching max_size\n");
1955 /* flush anything dirty? */
1956 if (cap == ci->i_auth_cap) {
1957 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
1958 dout("flushing dirty caps\n");
1961 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
1962 dout("flushing snap caps\n");
1967 /* completed revocation? going down and there are no caps? */
1968 if (revoking && (revoking & cap_used) == 0) {
1969 dout("completed revocation of %s\n",
1970 ceph_cap_string(cap->implemented & ~cap->issued));
1974 /* want more caps from mds? */
1975 if (want & ~(cap->mds_wanted | cap->issued))
1978 /* things we might delay */
1979 if ((cap->issued & ~retain) == 0)
1980 continue; /* nope, all good */
1986 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1987 time_before(jiffies, ci->i_hold_caps_max)) {
1988 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1989 ceph_cap_string(cap->issued),
1990 ceph_cap_string(cap->issued & retain),
1991 ceph_cap_string(cap->mds_wanted),
1992 ceph_cap_string(want));
1998 if (session && session != cap->session) {
1999 dout("oops, wrong session %p mutex\n", session);
2000 mutex_unlock(&session->s_mutex);
2004 session = cap->session;
2005 if (mutex_trylock(&session->s_mutex) == 0) {
2006 dout("inverting session/ino locks on %p\n",
2008 spin_unlock(&ci->i_ceph_lock);
2009 if (took_snap_rwsem) {
2010 up_read(&mdsc->snap_rwsem);
2011 took_snap_rwsem = 0;
2013 mutex_lock(&session->s_mutex);
2018 /* kick flushing and flush snaps before sending normal
2020 if (cap == ci->i_auth_cap &&
2022 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2023 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2024 __kick_flushing_caps(mdsc, session, ci, 0);
2025 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2026 __ceph_flush_snaps(ci, session);
2031 /* take snap_rwsem after session mutex */
2032 if (!took_snap_rwsem) {
2033 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
2034 dout("inverting snap/in locks on %p\n",
2036 spin_unlock(&ci->i_ceph_lock);
2037 down_read(&mdsc->snap_rwsem);
2038 took_snap_rwsem = 1;
2041 took_snap_rwsem = 1;
2044 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2045 flushing = ci->i_dirty_caps;
2046 flush_tid = __mark_caps_flushing(inode, session, false,
2051 spin_lock(&mdsc->cap_dirty_lock);
2052 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2053 spin_unlock(&mdsc->cap_dirty_lock);
2056 mds = cap->mds; /* remember mds, so we don't repeat */
2059 /* __send_cap drops i_ceph_lock */
2060 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, 0,
2061 cap_used, want, retain, flushing,
2062 flush_tid, oldest_flush_tid);
2063 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2066 /* Reschedule delayed caps release if we delayed anything */
2068 __cap_delay_requeue(mdsc, ci, false);
2070 spin_unlock(&ci->i_ceph_lock);
2072 if (queue_invalidate)
2073 ceph_queue_invalidate(inode);
2076 mutex_unlock(&session->s_mutex);
2077 if (took_snap_rwsem)
2078 up_read(&mdsc->snap_rwsem);
2082 * Try to flush dirty caps back to the auth mds.
2084 static int try_flush_caps(struct inode *inode, u64 *ptid)
2086 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2087 struct ceph_inode_info *ci = ceph_inode(inode);
2088 struct ceph_mds_session *session = NULL;
2090 u64 flush_tid = 0, oldest_flush_tid = 0;
2093 spin_lock(&ci->i_ceph_lock);
2095 if (ci->i_dirty_caps && ci->i_auth_cap) {
2096 struct ceph_cap *cap = ci->i_auth_cap;
2099 if (session != cap->session) {
2100 spin_unlock(&ci->i_ceph_lock);
2102 mutex_unlock(&session->s_mutex);
2103 session = cap->session;
2104 mutex_lock(&session->s_mutex);
2107 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) {
2108 spin_unlock(&ci->i_ceph_lock);
2112 if (ci->i_ceph_flags &
2113 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS)) {
2114 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH)
2115 __kick_flushing_caps(mdsc, session, ci, 0);
2116 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2117 __ceph_flush_snaps(ci, session);
2121 flushing = ci->i_dirty_caps;
2122 flush_tid = __mark_caps_flushing(inode, session, true,
2125 /* __send_cap drops i_ceph_lock */
2126 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2127 CEPH_CLIENT_CAPS_SYNC,
2128 __ceph_caps_used(ci),
2129 __ceph_caps_wanted(ci),
2130 (cap->issued | cap->implemented),
2131 flushing, flush_tid, oldest_flush_tid);
2134 spin_lock(&ci->i_ceph_lock);
2135 __cap_delay_requeue(mdsc, ci, true);
2136 spin_unlock(&ci->i_ceph_lock);
2139 if (!list_empty(&ci->i_cap_flush_list)) {
2140 struct ceph_cap_flush *cf =
2141 list_last_entry(&ci->i_cap_flush_list,
2142 struct ceph_cap_flush, i_list);
2144 flush_tid = cf->tid;
2146 flushing = ci->i_flushing_caps;
2147 spin_unlock(&ci->i_ceph_lock);
2151 mutex_unlock(&session->s_mutex);
2158 * Return true if we've flushed caps through the given flush_tid.
2160 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2162 struct ceph_inode_info *ci = ceph_inode(inode);
2165 spin_lock(&ci->i_ceph_lock);
2166 if (!list_empty(&ci->i_cap_flush_list)) {
2167 struct ceph_cap_flush * cf =
2168 list_first_entry(&ci->i_cap_flush_list,
2169 struct ceph_cap_flush, i_list);
2170 if (cf->tid <= flush_tid)
2173 spin_unlock(&ci->i_ceph_lock);
2178 * wait for any unsafe requests to complete.
2180 static int unsafe_request_wait(struct inode *inode)
2182 struct ceph_inode_info *ci = ceph_inode(inode);
2183 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2186 spin_lock(&ci->i_unsafe_lock);
2187 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2188 req1 = list_last_entry(&ci->i_unsafe_dirops,
2189 struct ceph_mds_request,
2191 ceph_mdsc_get_request(req1);
2193 if (!list_empty(&ci->i_unsafe_iops)) {
2194 req2 = list_last_entry(&ci->i_unsafe_iops,
2195 struct ceph_mds_request,
2196 r_unsafe_target_item);
2197 ceph_mdsc_get_request(req2);
2199 spin_unlock(&ci->i_unsafe_lock);
2201 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2202 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2204 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2205 ceph_timeout_jiffies(req1->r_timeout));
2208 ceph_mdsc_put_request(req1);
2211 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2212 ceph_timeout_jiffies(req2->r_timeout));
2215 ceph_mdsc_put_request(req2);
2220 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2222 struct ceph_file_info *fi = file->private_data;
2223 struct inode *inode = file->f_mapping->host;
2224 struct ceph_inode_info *ci = ceph_inode(inode);
2229 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2231 ret = file_write_and_wait_range(file, start, end);
2235 dirty = try_flush_caps(inode, &flush_tid);
2236 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2238 err = unsafe_request_wait(inode);
2241 * only wait on non-file metadata writeback (the mds
2242 * can recover size and mtime, so we don't need to
2245 if (!err && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2246 err = wait_event_interruptible(ci->i_cap_wq,
2247 caps_are_flushed(inode, flush_tid));
2253 if (errseq_check(&ci->i_meta_err, READ_ONCE(fi->meta_err))) {
2254 spin_lock(&file->f_lock);
2255 err = errseq_check_and_advance(&ci->i_meta_err,
2257 spin_unlock(&file->f_lock);
2262 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2267 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2268 * queue inode for flush but don't do so immediately, because we can
2269 * get by with fewer MDS messages if we wait for data writeback to
2272 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2274 struct ceph_inode_info *ci = ceph_inode(inode);
2278 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2280 dout("write_inode %p wait=%d\n", inode, wait);
2282 dirty = try_flush_caps(inode, &flush_tid);
2284 err = wait_event_interruptible(ci->i_cap_wq,
2285 caps_are_flushed(inode, flush_tid));
2287 struct ceph_mds_client *mdsc =
2288 ceph_sb_to_client(inode->i_sb)->mdsc;
2290 spin_lock(&ci->i_ceph_lock);
2291 if (__ceph_caps_dirty(ci))
2292 __cap_delay_requeue_front(mdsc, ci);
2293 spin_unlock(&ci->i_ceph_lock);
2298 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2299 struct ceph_mds_session *session,
2300 struct ceph_inode_info *ci,
2301 u64 oldest_flush_tid)
2302 __releases(ci->i_ceph_lock)
2303 __acquires(ci->i_ceph_lock)
2305 struct inode *inode = &ci->vfs_inode;
2306 struct ceph_cap *cap;
2307 struct ceph_cap_flush *cf;
2310 u64 last_snap_flush = 0;
2312 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2314 list_for_each_entry_reverse(cf, &ci->i_cap_flush_list, i_list) {
2316 last_snap_flush = cf->tid;
2321 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2322 if (cf->tid < first_tid)
2325 cap = ci->i_auth_cap;
2326 if (!(cap && cap->session == session)) {
2327 pr_err("%p auth cap %p not mds%d ???\n",
2328 inode, cap, session->s_mds);
2332 first_tid = cf->tid + 1;
2335 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2336 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2337 ci->i_ceph_flags |= CEPH_I_NODELAY;
2339 ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2340 (cf->tid < last_snap_flush ?
2341 CEPH_CLIENT_CAPS_PENDING_CAPSNAP : 0),
2342 __ceph_caps_used(ci),
2343 __ceph_caps_wanted(ci),
2344 (cap->issued | cap->implemented),
2345 cf->caps, cf->tid, oldest_flush_tid);
2347 pr_err("kick_flushing_caps: error sending "
2348 "cap flush, ino (%llx.%llx) "
2349 "tid %llu flushing %s\n",
2350 ceph_vinop(inode), cf->tid,
2351 ceph_cap_string(cf->caps));
2354 struct ceph_cap_snap *capsnap =
2355 container_of(cf, struct ceph_cap_snap,
2357 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2358 inode, capsnap, cf->tid,
2359 ceph_cap_string(capsnap->dirty));
2361 refcount_inc(&capsnap->nref);
2362 spin_unlock(&ci->i_ceph_lock);
2364 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2367 pr_err("kick_flushing_caps: error sending "
2368 "cap flushsnap, ino (%llx.%llx) "
2369 "tid %llu follows %llu\n",
2370 ceph_vinop(inode), cf->tid,
2374 ceph_put_cap_snap(capsnap);
2377 spin_lock(&ci->i_ceph_lock);
2381 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2382 struct ceph_mds_session *session)
2384 struct ceph_inode_info *ci;
2385 struct ceph_cap *cap;
2386 u64 oldest_flush_tid;
2388 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2390 spin_lock(&mdsc->cap_dirty_lock);
2391 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2392 spin_unlock(&mdsc->cap_dirty_lock);
2394 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2395 spin_lock(&ci->i_ceph_lock);
2396 cap = ci->i_auth_cap;
2397 if (!(cap && cap->session == session)) {
2398 pr_err("%p auth cap %p not mds%d ???\n",
2399 &ci->vfs_inode, cap, session->s_mds);
2400 spin_unlock(&ci->i_ceph_lock);
2406 * if flushing caps were revoked, we re-send the cap flush
2407 * in client reconnect stage. This guarantees MDS * processes
2408 * the cap flush message before issuing the flushing caps to
2411 if ((cap->issued & ci->i_flushing_caps) !=
2412 ci->i_flushing_caps) {
2413 /* encode_caps_cb() also will reset these sequence
2414 * numbers. make sure sequence numbers in cap flush
2415 * message match later reconnect message */
2419 __kick_flushing_caps(mdsc, session, ci,
2422 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2425 spin_unlock(&ci->i_ceph_lock);
2429 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2430 struct ceph_mds_session *session)
2432 struct ceph_inode_info *ci;
2433 struct ceph_cap *cap;
2434 u64 oldest_flush_tid;
2436 dout("kick_flushing_caps mds%d\n", session->s_mds);
2438 spin_lock(&mdsc->cap_dirty_lock);
2439 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2440 spin_unlock(&mdsc->cap_dirty_lock);
2442 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2443 spin_lock(&ci->i_ceph_lock);
2444 cap = ci->i_auth_cap;
2445 if (!(cap && cap->session == session)) {
2446 pr_err("%p auth cap %p not mds%d ???\n",
2447 &ci->vfs_inode, cap, session->s_mds);
2448 spin_unlock(&ci->i_ceph_lock);
2451 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2452 __kick_flushing_caps(mdsc, session, ci,
2455 spin_unlock(&ci->i_ceph_lock);
2459 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2460 struct ceph_mds_session *session,
2461 struct inode *inode)
2462 __releases(ci->i_ceph_lock)
2464 struct ceph_inode_info *ci = ceph_inode(inode);
2465 struct ceph_cap *cap;
2467 cap = ci->i_auth_cap;
2468 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2469 ceph_cap_string(ci->i_flushing_caps));
2471 if (!list_empty(&ci->i_cap_flush_list)) {
2472 u64 oldest_flush_tid;
2473 spin_lock(&mdsc->cap_dirty_lock);
2474 list_move_tail(&ci->i_flushing_item,
2475 &cap->session->s_cap_flushing);
2476 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2477 spin_unlock(&mdsc->cap_dirty_lock);
2479 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2480 spin_unlock(&ci->i_ceph_lock);
2482 spin_unlock(&ci->i_ceph_lock);
2488 * Take references to capabilities we hold, so that we don't release
2489 * them to the MDS prematurely.
2491 * Protected by i_ceph_lock.
2493 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2494 bool snap_rwsem_locked)
2496 if (got & CEPH_CAP_PIN)
2498 if (got & CEPH_CAP_FILE_RD)
2500 if (got & CEPH_CAP_FILE_CACHE)
2501 ci->i_rdcache_ref++;
2502 if (got & CEPH_CAP_FILE_WR) {
2503 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2504 BUG_ON(!snap_rwsem_locked);
2505 ci->i_head_snapc = ceph_get_snap_context(
2506 ci->i_snap_realm->cached_context);
2510 if (got & CEPH_CAP_FILE_BUFFER) {
2511 if (ci->i_wb_ref == 0)
2512 ihold(&ci->vfs_inode);
2514 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2515 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2520 * Try to grab cap references. Specify those refs we @want, and the
2521 * minimal set we @need. Also include the larger offset we are writing
2522 * to (when applicable), and check against max_size here as well.
2523 * Note that caller is responsible for ensuring max_size increases are
2524 * requested from the MDS.
2526 * Returns 0 if caps were not able to be acquired (yet), a 1 if they were,
2527 * or a negative error code.
2529 * FIXME: how does a 0 return differ from -EAGAIN?
2536 static int try_get_cap_refs(struct inode *inode, int need, int want,
2537 loff_t endoff, int flags, int *got)
2539 struct ceph_inode_info *ci = ceph_inode(inode);
2540 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2542 int have, implemented;
2544 bool snap_rwsem_locked = false;
2546 dout("get_cap_refs %p need %s want %s\n", inode,
2547 ceph_cap_string(need), ceph_cap_string(want));
2550 spin_lock(&ci->i_ceph_lock);
2552 if ((flags & CHECK_FILELOCK) &&
2553 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK)) {
2554 dout("try_get_cap_refs %p error filelock\n", inode);
2559 /* make sure file is actually open */
2560 file_wanted = __ceph_caps_file_wanted(ci);
2561 if ((file_wanted & need) != need) {
2562 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2563 ceph_cap_string(need), ceph_cap_string(file_wanted));
2568 /* finish pending truncate */
2569 while (ci->i_truncate_pending) {
2570 spin_unlock(&ci->i_ceph_lock);
2571 if (snap_rwsem_locked) {
2572 up_read(&mdsc->snap_rwsem);
2573 snap_rwsem_locked = false;
2575 __ceph_do_pending_vmtruncate(inode);
2576 spin_lock(&ci->i_ceph_lock);
2579 have = __ceph_caps_issued(ci, &implemented);
2581 if (have & need & CEPH_CAP_FILE_WR) {
2582 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2583 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2584 inode, endoff, ci->i_max_size);
2585 if (endoff > ci->i_requested_max_size)
2590 * If a sync write is in progress, we must wait, so that we
2591 * can get a final snapshot value for size+mtime.
2593 if (__ceph_have_pending_cap_snap(ci)) {
2594 dout("get_cap_refs %p cap_snap_pending\n", inode);
2599 if ((have & need) == need) {
2601 * Look at (implemented & ~have & not) so that we keep waiting
2602 * on transition from wanted -> needed caps. This is needed
2603 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2604 * going before a prior buffered writeback happens.
2606 int not = want & ~(have & need);
2607 int revoking = implemented & ~have;
2608 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2609 inode, ceph_cap_string(have), ceph_cap_string(not),
2610 ceph_cap_string(revoking));
2611 if ((revoking & not) == 0) {
2612 if (!snap_rwsem_locked &&
2613 !ci->i_head_snapc &&
2614 (need & CEPH_CAP_FILE_WR)) {
2615 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2617 * we can not call down_read() when
2618 * task isn't in TASK_RUNNING state
2620 if (flags & NON_BLOCKING) {
2625 spin_unlock(&ci->i_ceph_lock);
2626 down_read(&mdsc->snap_rwsem);
2627 snap_rwsem_locked = true;
2630 snap_rwsem_locked = true;
2632 *got = need | (have & want);
2633 if ((need & CEPH_CAP_FILE_RD) &&
2634 !(*got & CEPH_CAP_FILE_CACHE))
2635 ceph_disable_fscache_readpage(ci);
2636 __take_cap_refs(ci, *got, true);
2640 int session_readonly = false;
2641 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2642 struct ceph_mds_session *s = ci->i_auth_cap->session;
2643 spin_lock(&s->s_cap_lock);
2644 session_readonly = s->s_readonly;
2645 spin_unlock(&s->s_cap_lock);
2647 if (session_readonly) {
2648 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2649 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2654 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) {
2656 if (READ_ONCE(mdsc->fsc->mount_state) ==
2657 CEPH_MOUNT_SHUTDOWN) {
2658 dout("get_cap_refs %p forced umount\n", inode);
2662 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2663 if (need & ~(mds_wanted & need)) {
2664 dout("get_cap_refs %p caps were dropped"
2665 " (session killed?)\n", inode);
2669 if (!(file_wanted & ~mds_wanted))
2670 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED;
2673 dout("get_cap_refs %p have %s needed %s\n", inode,
2674 ceph_cap_string(have), ceph_cap_string(need));
2677 spin_unlock(&ci->i_ceph_lock);
2678 if (snap_rwsem_locked)
2679 up_read(&mdsc->snap_rwsem);
2681 dout("get_cap_refs %p ret %d got %s\n", inode,
2682 ret, ceph_cap_string(*got));
2687 * Check the offset we are writing up to against our current
2688 * max_size. If necessary, tell the MDS we want to write to
2691 static void check_max_size(struct inode *inode, loff_t endoff)
2693 struct ceph_inode_info *ci = ceph_inode(inode);
2696 /* do we need to explicitly request a larger max_size? */
2697 spin_lock(&ci->i_ceph_lock);
2698 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2699 dout("write %p at large endoff %llu, req max_size\n",
2701 ci->i_wanted_max_size = endoff;
2703 /* duplicate ceph_check_caps()'s logic */
2704 if (ci->i_auth_cap &&
2705 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2706 ci->i_wanted_max_size > ci->i_max_size &&
2707 ci->i_wanted_max_size > ci->i_requested_max_size)
2709 spin_unlock(&ci->i_ceph_lock);
2711 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2714 int ceph_try_get_caps(struct inode *inode, int need, int want,
2715 bool nonblock, int *got)
2719 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2720 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO|CEPH_CAP_FILE_SHARED));
2721 ret = ceph_pool_perm_check(inode, need);
2725 ret = try_get_cap_refs(inode, need, want, 0,
2726 (nonblock ? NON_BLOCKING : 0), got);
2727 return ret == -EAGAIN ? 0 : ret;
2731 * Wait for caps, and take cap references. If we can't get a WR cap
2732 * due to a small max_size, make sure we check_max_size (and possibly
2733 * ask the mds) so we don't get hung up indefinitely.
2735 int ceph_get_caps(struct file *filp, int need, int want,
2736 loff_t endoff, int *got, struct page **pinned_page)
2738 struct ceph_file_info *fi = filp->private_data;
2739 struct inode *inode = file_inode(filp);
2740 struct ceph_inode_info *ci = ceph_inode(inode);
2741 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2742 int ret, _got, flags;
2744 ret = ceph_pool_perm_check(inode, need);
2748 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2749 fi->filp_gen != READ_ONCE(fsc->filp_gen))
2754 check_max_size(inode, endoff);
2756 flags = atomic_read(&fi->num_locks) ? CHECK_FILELOCK : 0;
2758 ret = try_get_cap_refs(inode, need, want, endoff,
2763 struct ceph_mds_client *mdsc = fsc->mdsc;
2765 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2767 cw.ino = inode->i_ino;
2768 cw.tgid = current->tgid;
2772 spin_lock(&mdsc->caps_list_lock);
2773 list_add(&cw.list, &mdsc->cap_wait_list);
2774 spin_unlock(&mdsc->caps_list_lock);
2776 add_wait_queue(&ci->i_cap_wq, &wait);
2778 flags |= NON_BLOCKING;
2779 while (!(ret = try_get_cap_refs(inode, need, want,
2780 endoff, flags, &_got))) {
2781 if (signal_pending(current)) {
2785 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2788 remove_wait_queue(&ci->i_cap_wq, &wait);
2790 spin_lock(&mdsc->caps_list_lock);
2792 spin_unlock(&mdsc->caps_list_lock);
2798 if ((fi->fmode & CEPH_FILE_MODE_WR) &&
2799 fi->filp_gen != READ_ONCE(fsc->filp_gen)) {
2800 if (ret >= 0 && _got)
2801 ceph_put_cap_refs(ci, _got);
2806 if (ret == -ESTALE) {
2807 /* session was killed, try renew caps */
2808 ret = ceph_renew_caps(inode);
2815 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2816 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2817 i_size_read(inode) > 0) {
2819 find_get_page(inode->i_mapping, 0);
2821 if (PageUptodate(page)) {
2822 *pinned_page = page;
2828 * drop cap refs first because getattr while
2829 * holding * caps refs can cause deadlock.
2831 ceph_put_cap_refs(ci, _got);
2835 * getattr request will bring inline data into
2838 ret = __ceph_do_getattr(inode, NULL,
2839 CEPH_STAT_CAP_INLINE_DATA,
2848 if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
2849 ceph_fscache_revalidate_cookie(ci);
2856 * Take cap refs. Caller must already know we hold at least one ref
2857 * on the caps in question or we don't know this is safe.
2859 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2861 spin_lock(&ci->i_ceph_lock);
2862 __take_cap_refs(ci, caps, false);
2863 spin_unlock(&ci->i_ceph_lock);
2868 * drop cap_snap that is not associated with any snapshot.
2869 * we don't need to send FLUSHSNAP message for it.
2871 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2872 struct ceph_cap_snap *capsnap)
2874 if (!capsnap->need_flush &&
2875 !capsnap->writing && !capsnap->dirty_pages) {
2876 dout("dropping cap_snap %p follows %llu\n",
2877 capsnap, capsnap->follows);
2878 BUG_ON(capsnap->cap_flush.tid > 0);
2879 ceph_put_snap_context(capsnap->context);
2880 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2881 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2883 list_del(&capsnap->ci_item);
2884 ceph_put_cap_snap(capsnap);
2893 * If we released the last ref on any given cap, call ceph_check_caps
2894 * to release (or schedule a release).
2896 * If we are releasing a WR cap (from a sync write), finalize any affected
2897 * cap_snap, and wake up any waiters.
2899 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2901 struct inode *inode = &ci->vfs_inode;
2902 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2904 spin_lock(&ci->i_ceph_lock);
2905 if (had & CEPH_CAP_PIN)
2907 if (had & CEPH_CAP_FILE_RD)
2908 if (--ci->i_rd_ref == 0)
2910 if (had & CEPH_CAP_FILE_CACHE)
2911 if (--ci->i_rdcache_ref == 0)
2913 if (had & CEPH_CAP_FILE_BUFFER) {
2914 if (--ci->i_wb_ref == 0) {
2918 dout("put_cap_refs %p wb %d -> %d (?)\n",
2919 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2921 if (had & CEPH_CAP_FILE_WR)
2922 if (--ci->i_wr_ref == 0) {
2924 if (__ceph_have_pending_cap_snap(ci)) {
2925 struct ceph_cap_snap *capsnap =
2926 list_last_entry(&ci->i_cap_snaps,
2927 struct ceph_cap_snap,
2929 capsnap->writing = 0;
2930 if (ceph_try_drop_cap_snap(ci, capsnap))
2932 else if (__ceph_finish_cap_snap(ci, capsnap))
2936 if (ci->i_wrbuffer_ref_head == 0 &&
2937 ci->i_dirty_caps == 0 &&
2938 ci->i_flushing_caps == 0) {
2939 BUG_ON(!ci->i_head_snapc);
2940 ceph_put_snap_context(ci->i_head_snapc);
2941 ci->i_head_snapc = NULL;
2943 /* see comment in __ceph_remove_cap() */
2944 if (!__ceph_is_any_real_caps(ci) && ci->i_snap_realm)
2945 drop_inode_snap_realm(ci);
2947 spin_unlock(&ci->i_ceph_lock);
2949 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2950 last ? " last" : "", put ? " put" : "");
2952 if (last && !flushsnaps)
2953 ceph_check_caps(ci, 0, NULL);
2954 else if (flushsnaps)
2955 ceph_flush_snaps(ci, NULL);
2957 wake_up_all(&ci->i_cap_wq);
2963 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2964 * context. Adjust per-snap dirty page accounting as appropriate.
2965 * Once all dirty data for a cap_snap is flushed, flush snapped file
2966 * metadata back to the MDS. If we dropped the last ref, call
2969 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2970 struct ceph_snap_context *snapc)
2972 struct inode *inode = &ci->vfs_inode;
2973 struct ceph_cap_snap *capsnap = NULL;
2977 bool flush_snaps = false;
2978 bool complete_capsnap = false;
2980 spin_lock(&ci->i_ceph_lock);
2981 ci->i_wrbuffer_ref -= nr;
2982 if (ci->i_wrbuffer_ref == 0) {
2987 if (ci->i_head_snapc == snapc) {
2988 ci->i_wrbuffer_ref_head -= nr;
2989 if (ci->i_wrbuffer_ref_head == 0 &&
2990 ci->i_wr_ref == 0 &&
2991 ci->i_dirty_caps == 0 &&
2992 ci->i_flushing_caps == 0) {
2993 BUG_ON(!ci->i_head_snapc);
2994 ceph_put_snap_context(ci->i_head_snapc);
2995 ci->i_head_snapc = NULL;
2997 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2999 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
3000 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
3001 last ? " LAST" : "");
3003 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3004 if (capsnap->context == snapc) {
3010 capsnap->dirty_pages -= nr;
3011 if (capsnap->dirty_pages == 0) {
3012 complete_capsnap = true;
3013 if (!capsnap->writing) {
3014 if (ceph_try_drop_cap_snap(ci, capsnap)) {
3017 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
3022 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3023 " snap %lld %d/%d -> %d/%d %s%s\n",
3024 inode, capsnap, capsnap->context->seq,
3025 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
3026 ci->i_wrbuffer_ref, capsnap->dirty_pages,
3027 last ? " (wrbuffer last)" : "",
3028 complete_capsnap ? " (complete capsnap)" : "");
3031 spin_unlock(&ci->i_ceph_lock);
3034 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
3035 } else if (flush_snaps) {
3036 ceph_flush_snaps(ci, NULL);
3038 if (complete_capsnap)
3039 wake_up_all(&ci->i_cap_wq);
3041 /* avoid calling iput_final() in osd dispatch threads */
3042 ceph_async_iput(inode);
3047 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3049 static void invalidate_aliases(struct inode *inode)
3051 struct dentry *dn, *prev = NULL;
3053 dout("invalidate_aliases inode %p\n", inode);
3054 d_prune_aliases(inode);
3056 * For non-directory inode, d_find_alias() only returns
3057 * hashed dentry. After calling d_invalidate(), the
3058 * dentry becomes unhashed.
3060 * For directory inode, d_find_alias() can return
3061 * unhashed dentry. But directory inode should have
3062 * one alias at most.
3064 while ((dn = d_find_alias(inode))) {
3078 struct cap_extra_info {
3079 struct ceph_string *pool_ns;
3089 /* currently issued */
3091 struct timespec64 btime;
3095 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3096 * actually be a revocation if it specifies a smaller cap set.)
3098 * caller holds s_mutex and i_ceph_lock, we drop both.
3100 static void handle_cap_grant(struct inode *inode,
3101 struct ceph_mds_session *session,
3102 struct ceph_cap *cap,
3103 struct ceph_mds_caps *grant,
3104 struct ceph_buffer *xattr_buf,
3105 struct cap_extra_info *extra_info)
3106 __releases(ci->i_ceph_lock)
3107 __releases(session->s_mdsc->snap_rwsem)
3109 struct ceph_inode_info *ci = ceph_inode(inode);
3110 int seq = le32_to_cpu(grant->seq);
3111 int newcaps = le32_to_cpu(grant->caps);
3112 int used, wanted, dirty;
3113 u64 size = le64_to_cpu(grant->size);
3114 u64 max_size = le64_to_cpu(grant->max_size);
3115 unsigned char check_caps = 0;
3116 bool was_stale = cap->cap_gen < session->s_cap_gen;
3118 bool writeback = false;
3119 bool queue_trunc = false;
3120 bool queue_invalidate = false;
3121 bool deleted_inode = false;
3122 bool fill_inline = false;
3124 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3125 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3126 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3131 * If CACHE is being revoked, and we have no dirty buffers,
3132 * try to invalidate (once). (If there are dirty buffers, we
3133 * will invalidate _after_ writeback.)
3135 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
3136 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3137 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3138 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3139 if (try_nonblocking_invalidate(inode)) {
3140 /* there were locked pages.. invalidate later
3141 in a separate thread. */
3142 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3143 queue_invalidate = true;
3144 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3150 cap->issued = cap->implemented = CEPH_CAP_PIN;
3153 * auth mds of the inode changed. we received the cap export message,
3154 * but still haven't received the cap import message. handle_cap_export
3155 * updated the new auth MDS' cap.
3157 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3158 * that was sent before the cap import message. So don't remove caps.
3160 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3161 WARN_ON(cap != ci->i_auth_cap);
3162 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3164 newcaps |= cap->issued;
3167 /* side effects now are allowed */
3168 cap->cap_gen = session->s_cap_gen;
3171 __check_cap_issue(ci, cap, newcaps);
3173 inode_set_max_iversion_raw(inode, extra_info->change_attr);
3175 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3176 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3177 inode->i_mode = le32_to_cpu(grant->mode);
3178 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3179 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3180 ci->i_btime = extra_info->btime;
3181 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3182 from_kuid(&init_user_ns, inode->i_uid),
3183 from_kgid(&init_user_ns, inode->i_gid));
3186 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3187 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3188 set_nlink(inode, le32_to_cpu(grant->nlink));
3189 if (inode->i_nlink == 0 &&
3190 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3191 deleted_inode = true;
3194 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3196 int len = le32_to_cpu(grant->xattr_len);
3197 u64 version = le64_to_cpu(grant->xattr_version);
3199 if (version > ci->i_xattrs.version) {
3200 dout(" got new xattrs v%llu on %p len %d\n",
3201 version, inode, len);
3202 if (ci->i_xattrs.blob)
3203 ceph_buffer_put(ci->i_xattrs.blob);
3204 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3205 ci->i_xattrs.version = version;
3206 ceph_forget_all_cached_acls(inode);
3207 ceph_security_invalidate_secctx(inode);
3211 if (newcaps & CEPH_CAP_ANY_RD) {
3212 struct timespec64 mtime, atime, ctime;
3213 /* ctime/mtime/atime? */
3214 ceph_decode_timespec64(&mtime, &grant->mtime);
3215 ceph_decode_timespec64(&atime, &grant->atime);
3216 ceph_decode_timespec64(&ctime, &grant->ctime);
3217 ceph_fill_file_time(inode, extra_info->issued,
3218 le32_to_cpu(grant->time_warp_seq),
3219 &ctime, &mtime, &atime);
3222 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3223 ci->i_files = extra_info->nfiles;
3224 ci->i_subdirs = extra_info->nsubdirs;
3227 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3228 /* file layout may have changed */
3229 s64 old_pool = ci->i_layout.pool_id;
3230 struct ceph_string *old_ns;
3232 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3233 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3234 lockdep_is_held(&ci->i_ceph_lock));
3235 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3237 if (ci->i_layout.pool_id != old_pool ||
3238 extra_info->pool_ns != old_ns)
3239 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3241 extra_info->pool_ns = old_ns;
3243 /* size/truncate_seq? */
3244 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3245 le32_to_cpu(grant->truncate_seq),
3246 le64_to_cpu(grant->truncate_size),
3250 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3251 if (max_size != ci->i_max_size) {
3252 dout("max_size %lld -> %llu\n",
3253 ci->i_max_size, max_size);
3254 ci->i_max_size = max_size;
3255 if (max_size >= ci->i_wanted_max_size) {
3256 ci->i_wanted_max_size = 0; /* reset */
3257 ci->i_requested_max_size = 0;
3260 } else if (ci->i_wanted_max_size > ci->i_max_size &&
3261 ci->i_wanted_max_size > ci->i_requested_max_size) {
3262 /* CEPH_CAP_OP_IMPORT */
3267 /* check cap bits */
3268 wanted = __ceph_caps_wanted(ci);
3269 used = __ceph_caps_used(ci);
3270 dirty = __ceph_caps_dirty(ci);
3271 dout(" my wanted = %s, used = %s, dirty %s\n",
3272 ceph_cap_string(wanted),
3273 ceph_cap_string(used),
3274 ceph_cap_string(dirty));
3276 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3277 (wanted & ~(cap->mds_wanted | newcaps))) {
3279 * If mds is importing cap, prior cap messages that update
3280 * 'wanted' may get dropped by mds (migrate seq mismatch).
3282 * We don't send cap message to update 'wanted' if what we
3283 * want are already issued. If mds revokes caps, cap message
3284 * that releases caps also tells mds what we want. But if
3285 * caps got revoked by mds forcedly (session stale). We may
3286 * haven't told mds what we want.
3291 /* revocation, grant, or no-op? */
3292 if (cap->issued & ~newcaps) {
3293 int revoking = cap->issued & ~newcaps;
3295 dout("revocation: %s -> %s (revoking %s)\n",
3296 ceph_cap_string(cap->issued),
3297 ceph_cap_string(newcaps),
3298 ceph_cap_string(revoking));
3299 if (revoking & used & CEPH_CAP_FILE_BUFFER)
3300 writeback = true; /* initiate writeback; will delay ack */
3301 else if (revoking == CEPH_CAP_FILE_CACHE &&
3302 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3304 ; /* do nothing yet, invalidation will be queued */
3305 else if (cap == ci->i_auth_cap)
3306 check_caps = 1; /* check auth cap only */
3308 check_caps = 2; /* check all caps */
3309 cap->issued = newcaps;
3310 cap->implemented |= newcaps;
3311 } else if (cap->issued == newcaps) {
3312 dout("caps unchanged: %s -> %s\n",
3313 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3315 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3316 ceph_cap_string(newcaps));
3317 /* non-auth MDS is revoking the newly grant caps ? */
3318 if (cap == ci->i_auth_cap &&
3319 __ceph_caps_revoking_other(ci, cap, newcaps))
3322 cap->issued = newcaps;
3323 cap->implemented |= newcaps; /* add bits only, to
3324 * avoid stepping on a
3325 * pending revocation */
3328 BUG_ON(cap->issued & ~cap->implemented);
3330 if (extra_info->inline_version > 0 &&
3331 extra_info->inline_version >= ci->i_inline_version) {
3332 ci->i_inline_version = extra_info->inline_version;
3333 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3334 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3338 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3339 if (newcaps & ~extra_info->issued)
3341 kick_flushing_inode_caps(session->s_mdsc, session, inode);
3342 up_read(&session->s_mdsc->snap_rwsem);
3344 spin_unlock(&ci->i_ceph_lock);
3348 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3349 extra_info->inline_len);
3352 ceph_queue_vmtruncate(inode);
3356 * queue inode for writeback: we can't actually call
3357 * filemap_write_and_wait, etc. from message handler
3360 ceph_queue_writeback(inode);
3361 if (queue_invalidate)
3362 ceph_queue_invalidate(inode);
3364 invalidate_aliases(inode);
3366 wake_up_all(&ci->i_cap_wq);
3368 if (check_caps == 1)
3369 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
3371 else if (check_caps == 2)
3372 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
3374 mutex_unlock(&session->s_mutex);
3378 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3379 * MDS has been safely committed.
3381 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3382 struct ceph_mds_caps *m,
3383 struct ceph_mds_session *session,
3384 struct ceph_cap *cap)
3385 __releases(ci->i_ceph_lock)
3387 struct ceph_inode_info *ci = ceph_inode(inode);
3388 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3389 struct ceph_cap_flush *cf, *tmp_cf;
3390 LIST_HEAD(to_remove);
3391 unsigned seq = le32_to_cpu(m->seq);
3392 int dirty = le32_to_cpu(m->dirty);
3395 bool wake_ci = false;
3396 bool wake_mdsc = false;
3398 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3399 if (cf->tid == flush_tid)
3401 if (cf->caps == 0) /* capsnap */
3403 if (cf->tid <= flush_tid) {
3404 if (__finish_cap_flush(NULL, ci, cf))
3406 list_add_tail(&cf->i_list, &to_remove);
3408 cleaned &= ~cf->caps;
3414 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3415 " flushing %s -> %s\n",
3416 inode, session->s_mds, seq, ceph_cap_string(dirty),
3417 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3418 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3420 if (list_empty(&to_remove) && !cleaned)
3423 ci->i_flushing_caps &= ~cleaned;
3425 spin_lock(&mdsc->cap_dirty_lock);
3427 list_for_each_entry(cf, &to_remove, i_list) {
3428 if (__finish_cap_flush(mdsc, NULL, cf))
3432 if (ci->i_flushing_caps == 0) {
3433 if (list_empty(&ci->i_cap_flush_list)) {
3434 list_del_init(&ci->i_flushing_item);
3435 if (!list_empty(&session->s_cap_flushing)) {
3436 dout(" mds%d still flushing cap on %p\n",
3438 &list_first_entry(&session->s_cap_flushing,
3439 struct ceph_inode_info,
3440 i_flushing_item)->vfs_inode);
3443 mdsc->num_cap_flushing--;
3444 dout(" inode %p now !flushing\n", inode);
3446 if (ci->i_dirty_caps == 0) {
3447 dout(" inode %p now clean\n", inode);
3448 BUG_ON(!list_empty(&ci->i_dirty_item));
3450 if (ci->i_wr_ref == 0 &&
3451 ci->i_wrbuffer_ref_head == 0) {
3452 BUG_ON(!ci->i_head_snapc);
3453 ceph_put_snap_context(ci->i_head_snapc);
3454 ci->i_head_snapc = NULL;
3457 BUG_ON(list_empty(&ci->i_dirty_item));
3460 spin_unlock(&mdsc->cap_dirty_lock);
3463 spin_unlock(&ci->i_ceph_lock);
3465 while (!list_empty(&to_remove)) {
3466 cf = list_first_entry(&to_remove,
3467 struct ceph_cap_flush, i_list);
3468 list_del(&cf->i_list);
3469 ceph_free_cap_flush(cf);
3473 wake_up_all(&ci->i_cap_wq);
3475 wake_up_all(&mdsc->cap_flushing_wq);
3481 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3482 * throw away our cap_snap.
3484 * Caller hold s_mutex.
3486 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3487 struct ceph_mds_caps *m,
3488 struct ceph_mds_session *session)
3490 struct ceph_inode_info *ci = ceph_inode(inode);
3491 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3492 u64 follows = le64_to_cpu(m->snap_follows);
3493 struct ceph_cap_snap *capsnap;
3494 bool flushed = false;
3495 bool wake_ci = false;
3496 bool wake_mdsc = false;
3498 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3499 inode, ci, session->s_mds, follows);
3501 spin_lock(&ci->i_ceph_lock);
3502 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3503 if (capsnap->follows == follows) {
3504 if (capsnap->cap_flush.tid != flush_tid) {
3505 dout(" cap_snap %p follows %lld tid %lld !="
3506 " %lld\n", capsnap, follows,
3507 flush_tid, capsnap->cap_flush.tid);
3513 dout(" skipping cap_snap %p follows %lld\n",
3514 capsnap, capsnap->follows);
3518 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3519 dout(" removing %p cap_snap %p follows %lld\n",
3520 inode, capsnap, follows);
3521 list_del(&capsnap->ci_item);
3522 if (__finish_cap_flush(NULL, ci, &capsnap->cap_flush))
3525 spin_lock(&mdsc->cap_dirty_lock);
3527 if (list_empty(&ci->i_cap_flush_list))
3528 list_del_init(&ci->i_flushing_item);
3530 if (__finish_cap_flush(mdsc, NULL, &capsnap->cap_flush))
3533 spin_unlock(&mdsc->cap_dirty_lock);
3535 spin_unlock(&ci->i_ceph_lock);
3537 ceph_put_snap_context(capsnap->context);
3538 ceph_put_cap_snap(capsnap);
3540 wake_up_all(&ci->i_cap_wq);
3542 wake_up_all(&mdsc->cap_flushing_wq);
3548 * Handle TRUNC from MDS, indicating file truncation.
3550 * caller hold s_mutex.
3552 static void handle_cap_trunc(struct inode *inode,
3553 struct ceph_mds_caps *trunc,
3554 struct ceph_mds_session *session)
3555 __releases(ci->i_ceph_lock)
3557 struct ceph_inode_info *ci = ceph_inode(inode);
3558 int mds = session->s_mds;
3559 int seq = le32_to_cpu(trunc->seq);
3560 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3561 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3562 u64 size = le64_to_cpu(trunc->size);
3563 int implemented = 0;
3564 int dirty = __ceph_caps_dirty(ci);
3565 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3566 int queue_trunc = 0;
3568 issued |= implemented | dirty;
3570 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3571 inode, mds, seq, truncate_size, truncate_seq);
3572 queue_trunc = ceph_fill_file_size(inode, issued,
3573 truncate_seq, truncate_size, size);
3574 spin_unlock(&ci->i_ceph_lock);
3577 ceph_queue_vmtruncate(inode);
3581 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3582 * different one. If we are the most recent migration we've seen (as
3583 * indicated by mseq), make note of the migrating cap bits for the
3584 * duration (until we see the corresponding IMPORT).
3586 * caller holds s_mutex
3588 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3589 struct ceph_mds_cap_peer *ph,
3590 struct ceph_mds_session *session)
3592 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3593 struct ceph_mds_session *tsession = NULL;
3594 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3595 struct ceph_inode_info *ci = ceph_inode(inode);
3597 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3598 unsigned t_seq, t_mseq;
3600 int mds = session->s_mds;
3603 t_cap_id = le64_to_cpu(ph->cap_id);
3604 t_seq = le32_to_cpu(ph->seq);
3605 t_mseq = le32_to_cpu(ph->mseq);
3606 target = le32_to_cpu(ph->mds);
3608 t_cap_id = t_seq = t_mseq = 0;
3612 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3613 inode, ci, mds, mseq, target);
3615 spin_lock(&ci->i_ceph_lock);
3616 cap = __get_cap_for_mds(ci, mds);
3617 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3621 if (cap->mds_wanted | cap->issued)
3622 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
3623 __ceph_remove_cap(cap, false);
3628 * now we know we haven't received the cap import message yet
3629 * because the exported cap still exist.
3632 issued = cap->issued;
3633 if (issued != cap->implemented)
3634 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3635 "ino (%llx.%llx) mds%d seq %d mseq %d "
3636 "issued %s implemented %s\n",
3637 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3638 ceph_cap_string(issued),
3639 ceph_cap_string(cap->implemented));
3642 tcap = __get_cap_for_mds(ci, target);
3644 /* already have caps from the target */
3645 if (tcap->cap_id == t_cap_id &&
3646 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3647 dout(" updating import cap %p mds%d\n", tcap, target);
3648 tcap->cap_id = t_cap_id;
3649 tcap->seq = t_seq - 1;
3650 tcap->issue_seq = t_seq - 1;
3651 tcap->issued |= issued;
3652 tcap->implemented |= issued;
3653 if (cap == ci->i_auth_cap)
3654 ci->i_auth_cap = tcap;
3656 if (!list_empty(&ci->i_cap_flush_list) &&
3657 ci->i_auth_cap == tcap) {
3658 spin_lock(&mdsc->cap_dirty_lock);
3659 list_move_tail(&ci->i_flushing_item,
3660 &tcap->session->s_cap_flushing);
3661 spin_unlock(&mdsc->cap_dirty_lock);
3664 __ceph_remove_cap(cap, false);
3666 } else if (tsession) {
3667 /* add placeholder for the export tagert */
3668 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3670 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3671 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3673 if (!list_empty(&ci->i_cap_flush_list) &&
3674 ci->i_auth_cap == tcap) {
3675 spin_lock(&mdsc->cap_dirty_lock);
3676 list_move_tail(&ci->i_flushing_item,
3677 &tcap->session->s_cap_flushing);
3678 spin_unlock(&mdsc->cap_dirty_lock);
3681 __ceph_remove_cap(cap, false);
3685 spin_unlock(&ci->i_ceph_lock);
3686 mutex_unlock(&session->s_mutex);
3688 /* open target session */
3689 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3690 if (!IS_ERR(tsession)) {
3692 mutex_lock(&session->s_mutex);
3693 mutex_lock_nested(&tsession->s_mutex,
3694 SINGLE_DEPTH_NESTING);
3696 mutex_lock(&tsession->s_mutex);
3697 mutex_lock_nested(&session->s_mutex,
3698 SINGLE_DEPTH_NESTING);
3700 new_cap = ceph_get_cap(mdsc, NULL);
3709 spin_unlock(&ci->i_ceph_lock);
3710 mutex_unlock(&session->s_mutex);
3712 mutex_unlock(&tsession->s_mutex);
3713 ceph_put_mds_session(tsession);
3716 ceph_put_cap(mdsc, new_cap);
3720 * Handle cap IMPORT.
3722 * caller holds s_mutex. acquires i_ceph_lock
3724 static void handle_cap_import(struct ceph_mds_client *mdsc,
3725 struct inode *inode, struct ceph_mds_caps *im,
3726 struct ceph_mds_cap_peer *ph,
3727 struct ceph_mds_session *session,
3728 struct ceph_cap **target_cap, int *old_issued)
3729 __acquires(ci->i_ceph_lock)
3731 struct ceph_inode_info *ci = ceph_inode(inode);
3732 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3733 int mds = session->s_mds;
3735 unsigned caps = le32_to_cpu(im->caps);
3736 unsigned wanted = le32_to_cpu(im->wanted);
3737 unsigned seq = le32_to_cpu(im->seq);
3738 unsigned mseq = le32_to_cpu(im->migrate_seq);
3739 u64 realmino = le64_to_cpu(im->realm);
3740 u64 cap_id = le64_to_cpu(im->cap_id);
3745 p_cap_id = le64_to_cpu(ph->cap_id);
3746 peer = le32_to_cpu(ph->mds);
3752 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3753 inode, ci, mds, mseq, peer);
3756 spin_lock(&ci->i_ceph_lock);
3757 cap = __get_cap_for_mds(ci, mds);
3760 spin_unlock(&ci->i_ceph_lock);
3761 new_cap = ceph_get_cap(mdsc, NULL);
3767 ceph_put_cap(mdsc, new_cap);
3772 __ceph_caps_issued(ci, &issued);
3773 issued |= __ceph_caps_dirty(ci);
3775 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3776 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3778 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3779 if (ocap && ocap->cap_id == p_cap_id) {
3780 dout(" remove export cap %p mds%d flags %d\n",
3781 ocap, peer, ph->flags);
3782 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3783 (ocap->seq != le32_to_cpu(ph->seq) ||
3784 ocap->mseq != le32_to_cpu(ph->mseq))) {
3785 pr_err_ratelimited("handle_cap_import: "
3786 "mismatched seq/mseq: ino (%llx.%llx) "
3787 "mds%d seq %d mseq %d importer mds%d "
3788 "has peer seq %d mseq %d\n",
3789 ceph_vinop(inode), peer, ocap->seq,
3790 ocap->mseq, mds, le32_to_cpu(ph->seq),
3791 le32_to_cpu(ph->mseq));
3793 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3796 /* make sure we re-request max_size, if necessary */
3797 ci->i_requested_max_size = 0;
3799 *old_issued = issued;
3804 * Handle a caps message from the MDS.
3806 * Identify the appropriate session, inode, and call the right handler
3807 * based on the cap op.
3809 void ceph_handle_caps(struct ceph_mds_session *session,
3810 struct ceph_msg *msg)
3812 struct ceph_mds_client *mdsc = session->s_mdsc;
3813 struct inode *inode;
3814 struct ceph_inode_info *ci;
3815 struct ceph_cap *cap;
3816 struct ceph_mds_caps *h;
3817 struct ceph_mds_cap_peer *peer = NULL;
3818 struct ceph_snap_realm *realm = NULL;
3820 int msg_version = le16_to_cpu(msg->hdr.version);
3822 struct ceph_vino vino;
3824 size_t snaptrace_len;
3826 struct cap_extra_info extra_info = {};
3828 dout("handle_caps from mds%d\n", session->s_mds);
3831 end = msg->front.iov_base + msg->front.iov_len;
3832 if (msg->front.iov_len < sizeof(*h))
3834 h = msg->front.iov_base;
3835 op = le32_to_cpu(h->op);
3836 vino.ino = le64_to_cpu(h->ino);
3837 vino.snap = CEPH_NOSNAP;
3838 seq = le32_to_cpu(h->seq);
3839 mseq = le32_to_cpu(h->migrate_seq);
3842 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3843 p = snaptrace + snaptrace_len;
3845 if (msg_version >= 2) {
3847 ceph_decode_32_safe(&p, end, flock_len, bad);
3848 if (p + flock_len > end)
3853 if (msg_version >= 3) {
3854 if (op == CEPH_CAP_OP_IMPORT) {
3855 if (p + sizeof(*peer) > end)
3859 } else if (op == CEPH_CAP_OP_EXPORT) {
3860 /* recorded in unused fields */
3861 peer = (void *)&h->size;
3865 if (msg_version >= 4) {
3866 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
3867 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
3868 if (p + extra_info.inline_len > end)
3870 extra_info.inline_data = p;
3871 p += extra_info.inline_len;
3874 if (msg_version >= 5) {
3875 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
3878 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
3879 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
3882 if (msg_version >= 8) {
3884 u32 caller_uid, caller_gid;
3888 ceph_decode_64_safe(&p, end, flush_tid, bad);
3890 ceph_decode_32_safe(&p, end, caller_uid, bad);
3891 ceph_decode_32_safe(&p, end, caller_gid, bad);
3893 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
3894 if (pool_ns_len > 0) {
3895 ceph_decode_need(&p, end, pool_ns_len, bad);
3896 extra_info.pool_ns =
3897 ceph_find_or_create_string(p, pool_ns_len);
3902 if (msg_version >= 9) {
3903 struct ceph_timespec *btime;
3905 if (p + sizeof(*btime) > end)
3908 ceph_decode_timespec64(&extra_info.btime, btime);
3909 p += sizeof(*btime);
3910 ceph_decode_64_safe(&p, end, extra_info.change_attr, bad);
3913 if (msg_version >= 11) {
3916 ceph_decode_32_safe(&p, end, flags, bad);
3918 extra_info.dirstat_valid = true;
3919 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
3920 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
3924 inode = ceph_find_inode(mdsc->fsc->sb, vino);
3925 ci = ceph_inode(inode);
3926 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3929 mutex_lock(&session->s_mutex);
3931 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3935 dout(" i don't have ino %llx\n", vino.ino);
3937 if (op == CEPH_CAP_OP_IMPORT) {
3938 cap = ceph_get_cap(mdsc, NULL);
3939 cap->cap_ino = vino.ino;
3940 cap->queue_release = 1;
3941 cap->cap_id = le64_to_cpu(h->cap_id);
3944 cap->issue_seq = seq;
3945 spin_lock(&session->s_cap_lock);
3946 __ceph_queue_cap_release(session, cap);
3947 spin_unlock(&session->s_cap_lock);
3952 /* these will work even if we don't have a cap yet */
3954 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3955 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
3959 case CEPH_CAP_OP_EXPORT:
3960 handle_cap_export(inode, h, peer, session);
3963 case CEPH_CAP_OP_IMPORT:
3965 if (snaptrace_len) {
3966 down_write(&mdsc->snap_rwsem);
3967 ceph_update_snap_trace(mdsc, snaptrace,
3968 snaptrace + snaptrace_len,
3970 downgrade_write(&mdsc->snap_rwsem);
3972 down_read(&mdsc->snap_rwsem);
3974 handle_cap_import(mdsc, inode, h, peer, session,
3975 &cap, &extra_info.issued);
3976 handle_cap_grant(inode, session, cap,
3977 h, msg->middle, &extra_info);
3979 ceph_put_snap_realm(mdsc, realm);
3983 /* the rest require a cap */
3984 spin_lock(&ci->i_ceph_lock);
3985 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
3987 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3988 inode, ceph_ino(inode), ceph_snap(inode),
3990 spin_unlock(&ci->i_ceph_lock);
3991 goto flush_cap_releases;
3994 /* note that each of these drops i_ceph_lock for us */
3996 case CEPH_CAP_OP_REVOKE:
3997 case CEPH_CAP_OP_GRANT:
3998 __ceph_caps_issued(ci, &extra_info.issued);
3999 extra_info.issued |= __ceph_caps_dirty(ci);
4000 handle_cap_grant(inode, session, cap,
4001 h, msg->middle, &extra_info);
4004 case CEPH_CAP_OP_FLUSH_ACK:
4005 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
4009 case CEPH_CAP_OP_TRUNC:
4010 handle_cap_trunc(inode, h, session);
4014 spin_unlock(&ci->i_ceph_lock);
4015 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
4016 ceph_cap_op_name(op));
4020 mutex_unlock(&session->s_mutex);
4022 ceph_put_string(extra_info.pool_ns);
4023 /* avoid calling iput_final() in mds dispatch threads */
4024 ceph_async_iput(inode);
4029 * send any cap release message to try to move things
4030 * along for the mds (who clearly thinks we still have this
4033 ceph_flush_cap_releases(mdsc, session);
4037 pr_err("ceph_handle_caps: corrupt message\n");
4043 * Delayed work handler to process end of delayed cap release LRU list.
4045 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
4047 struct inode *inode;
4048 struct ceph_inode_info *ci;
4049 int flags = CHECK_CAPS_NODELAY;
4051 dout("check_delayed_caps\n");
4053 spin_lock(&mdsc->cap_delay_lock);
4054 if (list_empty(&mdsc->cap_delay_list))
4056 ci = list_first_entry(&mdsc->cap_delay_list,
4057 struct ceph_inode_info,
4059 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4060 time_before(jiffies, ci->i_hold_caps_max))
4062 list_del_init(&ci->i_cap_delay_list);
4064 inode = igrab(&ci->vfs_inode);
4065 spin_unlock(&mdsc->cap_delay_lock);
4068 dout("check_delayed_caps on %p\n", inode);
4069 ceph_check_caps(ci, flags, NULL);
4070 /* avoid calling iput_final() in tick thread */
4071 ceph_async_iput(inode);
4074 spin_unlock(&mdsc->cap_delay_lock);
4078 * Flush all dirty caps to the mds
4080 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4082 struct ceph_inode_info *ci;
4083 struct inode *inode;
4085 dout("flush_dirty_caps\n");
4086 spin_lock(&mdsc->cap_dirty_lock);
4087 while (!list_empty(&mdsc->cap_dirty)) {
4088 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
4090 inode = &ci->vfs_inode;
4092 dout("flush_dirty_caps %p\n", inode);
4093 spin_unlock(&mdsc->cap_dirty_lock);
4094 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
4096 spin_lock(&mdsc->cap_dirty_lock);
4098 spin_unlock(&mdsc->cap_dirty_lock);
4099 dout("flush_dirty_caps done\n");
4102 void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode)
4105 int bits = (fmode << 1) | 1;
4106 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4107 if (bits & (1 << i))
4108 ci->i_nr_by_mode[i]++;
4113 * Drop open file reference. If we were the last open file,
4114 * we may need to release capabilities to the MDS (or schedule
4115 * their delayed release).
4117 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
4120 int bits = (fmode << 1) | 1;
4121 spin_lock(&ci->i_ceph_lock);
4122 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4123 if (bits & (1 << i)) {
4124 BUG_ON(ci->i_nr_by_mode[i] == 0);
4125 if (--ci->i_nr_by_mode[i] == 0)
4129 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
4130 &ci->vfs_inode, fmode,
4131 ci->i_nr_by_mode[0], ci->i_nr_by_mode[1],
4132 ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]);
4133 spin_unlock(&ci->i_ceph_lock);
4135 if (last && ci->i_vino.snap == CEPH_NOSNAP)
4136 ceph_check_caps(ci, 0, NULL);
4140 * For a soon-to-be unlinked file, drop the LINK caps. If it
4141 * looks like the link count will hit 0, drop any other caps (other
4142 * than PIN) we don't specifically want (due to the file still being
4145 int ceph_drop_caps_for_unlink(struct inode *inode)
4147 struct ceph_inode_info *ci = ceph_inode(inode);
4148 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4150 spin_lock(&ci->i_ceph_lock);
4151 if (inode->i_nlink == 1) {
4152 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4154 ci->i_ceph_flags |= CEPH_I_NODELAY;
4155 if (__ceph_caps_dirty(ci)) {
4156 struct ceph_mds_client *mdsc =
4157 ceph_inode_to_client(inode)->mdsc;
4158 __cap_delay_requeue_front(mdsc, ci);
4161 spin_unlock(&ci->i_ceph_lock);
4166 * Helpers for embedding cap and dentry lease releases into mds
4169 * @force is used by dentry_release (below) to force inclusion of a
4170 * record for the directory inode, even when there aren't any caps to
4173 int ceph_encode_inode_release(void **p, struct inode *inode,
4174 int mds, int drop, int unless, int force)
4176 struct ceph_inode_info *ci = ceph_inode(inode);
4177 struct ceph_cap *cap;
4178 struct ceph_mds_request_release *rel = *p;
4182 spin_lock(&ci->i_ceph_lock);
4183 used = __ceph_caps_used(ci);
4184 dirty = __ceph_caps_dirty(ci);
4186 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4187 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4188 ceph_cap_string(unless));
4190 /* only drop unused, clean caps */
4191 drop &= ~(used | dirty);
4193 cap = __get_cap_for_mds(ci, mds);
4194 if (cap && __cap_is_valid(cap)) {
4195 unless &= cap->issued;
4197 if (unless & CEPH_CAP_AUTH_EXCL)
4198 drop &= ~CEPH_CAP_AUTH_SHARED;
4199 if (unless & CEPH_CAP_LINK_EXCL)
4200 drop &= ~CEPH_CAP_LINK_SHARED;
4201 if (unless & CEPH_CAP_XATTR_EXCL)
4202 drop &= ~CEPH_CAP_XATTR_SHARED;
4203 if (unless & CEPH_CAP_FILE_EXCL)
4204 drop &= ~CEPH_CAP_FILE_SHARED;
4207 if (force || (cap->issued & drop)) {
4208 if (cap->issued & drop) {
4209 int wanted = __ceph_caps_wanted(ci);
4210 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
4211 wanted |= cap->mds_wanted;
4212 dout("encode_inode_release %p cap %p "
4213 "%s -> %s, wanted %s -> %s\n", inode, cap,
4214 ceph_cap_string(cap->issued),
4215 ceph_cap_string(cap->issued & ~drop),
4216 ceph_cap_string(cap->mds_wanted),
4217 ceph_cap_string(wanted));
4219 cap->issued &= ~drop;
4220 cap->implemented &= ~drop;
4221 cap->mds_wanted = wanted;
4223 dout("encode_inode_release %p cap %p %s"
4224 " (force)\n", inode, cap,
4225 ceph_cap_string(cap->issued));
4228 rel->ino = cpu_to_le64(ceph_ino(inode));
4229 rel->cap_id = cpu_to_le64(cap->cap_id);
4230 rel->seq = cpu_to_le32(cap->seq);
4231 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4232 rel->mseq = cpu_to_le32(cap->mseq);
4233 rel->caps = cpu_to_le32(cap->implemented);
4234 rel->wanted = cpu_to_le32(cap->mds_wanted);
4240 dout("encode_inode_release %p cap %p %s (noop)\n",
4241 inode, cap, ceph_cap_string(cap->issued));
4244 spin_unlock(&ci->i_ceph_lock);
4248 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4250 int mds, int drop, int unless)
4252 struct dentry *parent = NULL;
4253 struct ceph_mds_request_release *rel = *p;
4254 struct ceph_dentry_info *di = ceph_dentry(dentry);
4259 * force an record for the directory caps if we have a dentry lease.
4260 * this is racy (can't take i_ceph_lock and d_lock together), but it
4261 * doesn't have to be perfect; the mds will revoke anything we don't
4264 spin_lock(&dentry->d_lock);
4265 if (di->lease_session && di->lease_session->s_mds == mds)
4268 parent = dget(dentry->d_parent);
4269 dir = d_inode(parent);
4271 spin_unlock(&dentry->d_lock);
4273 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4276 spin_lock(&dentry->d_lock);
4277 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4278 dout("encode_dentry_release %p mds%d seq %d\n",
4279 dentry, mds, (int)di->lease_seq);
4280 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4281 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4282 *p += dentry->d_name.len;
4283 rel->dname_seq = cpu_to_le32(di->lease_seq);
4284 __ceph_mdsc_drop_dentry_lease(dentry);
4286 spin_unlock(&dentry->d_lock);