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>
13 #include "mds_client.h"
15 #include <linux/ceph/decode.h>
16 #include <linux/ceph/messenger.h>
19 * Capability management
21 * The Ceph metadata servers control client access to inode metadata
22 * and file data by issuing capabilities, granting clients permission
23 * to read and/or write both inode field and file data to OSDs
24 * (storage nodes). Each capability consists of a set of bits
25 * indicating which operations are allowed.
27 * If the client holds a *_SHARED cap, the client has a coherent value
28 * that can be safely read from the cached inode.
30 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
31 * client is allowed to change inode attributes (e.g., file size,
32 * mtime), note its dirty state in the ceph_cap, and asynchronously
33 * flush that metadata change to the MDS.
35 * In the event of a conflicting operation (perhaps by another
36 * client), the MDS will revoke the conflicting client capabilities.
38 * In order for a client to cache an inode, it must hold a capability
39 * with at least one MDS server. When inodes are released, release
40 * notifications are batched and periodically sent en masse to the MDS
41 * cluster to release server state.
44 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc);
45 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
46 struct ceph_mds_session *session,
47 struct ceph_inode_info *ci,
48 u64 oldest_flush_tid);
51 * Generate readable cap strings for debugging output.
53 #define MAX_CAP_STR 20
54 static char cap_str[MAX_CAP_STR][40];
55 static DEFINE_SPINLOCK(cap_str_lock);
56 static int last_cap_str;
58 static char *gcap_string(char *s, int c)
60 if (c & CEPH_CAP_GSHARED)
62 if (c & CEPH_CAP_GEXCL)
64 if (c & CEPH_CAP_GCACHE)
70 if (c & CEPH_CAP_GBUFFER)
72 if (c & CEPH_CAP_GWREXTEND)
74 if (c & CEPH_CAP_GLAZYIO)
79 const char *ceph_cap_string(int caps)
85 spin_lock(&cap_str_lock);
87 if (last_cap_str == MAX_CAP_STR)
89 spin_unlock(&cap_str_lock);
93 if (caps & CEPH_CAP_PIN)
96 c = (caps >> CEPH_CAP_SAUTH) & 3;
99 s = gcap_string(s, c);
102 c = (caps >> CEPH_CAP_SLINK) & 3;
105 s = gcap_string(s, c);
108 c = (caps >> CEPH_CAP_SXATTR) & 3;
111 s = gcap_string(s, c);
114 c = caps >> CEPH_CAP_SFILE;
117 s = gcap_string(s, c);
126 void ceph_caps_init(struct ceph_mds_client *mdsc)
128 INIT_LIST_HEAD(&mdsc->caps_list);
129 spin_lock_init(&mdsc->caps_list_lock);
132 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
134 struct ceph_cap *cap;
136 spin_lock(&mdsc->caps_list_lock);
137 while (!list_empty(&mdsc->caps_list)) {
138 cap = list_first_entry(&mdsc->caps_list,
139 struct ceph_cap, caps_item);
140 list_del(&cap->caps_item);
141 kmem_cache_free(ceph_cap_cachep, cap);
143 mdsc->caps_total_count = 0;
144 mdsc->caps_avail_count = 0;
145 mdsc->caps_use_count = 0;
146 mdsc->caps_reserve_count = 0;
147 mdsc->caps_min_count = 0;
148 spin_unlock(&mdsc->caps_list_lock);
151 void ceph_adjust_caps_max_min(struct ceph_mds_client *mdsc,
152 struct ceph_mount_options *fsopt)
154 spin_lock(&mdsc->caps_list_lock);
155 mdsc->caps_min_count = fsopt->max_readdir;
156 if (mdsc->caps_min_count < 1024)
157 mdsc->caps_min_count = 1024;
158 mdsc->caps_use_max = fsopt->caps_max;
159 if (mdsc->caps_use_max > 0 &&
160 mdsc->caps_use_max < mdsc->caps_min_count)
161 mdsc->caps_use_max = mdsc->caps_min_count;
162 spin_unlock(&mdsc->caps_list_lock);
165 static void __ceph_unreserve_caps(struct ceph_mds_client *mdsc, int nr_caps)
167 struct ceph_cap *cap;
171 BUG_ON(mdsc->caps_reserve_count < nr_caps);
172 mdsc->caps_reserve_count -= nr_caps;
173 if (mdsc->caps_avail_count >=
174 mdsc->caps_reserve_count + mdsc->caps_min_count) {
175 mdsc->caps_total_count -= nr_caps;
176 for (i = 0; i < nr_caps; i++) {
177 cap = list_first_entry(&mdsc->caps_list,
178 struct ceph_cap, caps_item);
179 list_del(&cap->caps_item);
180 kmem_cache_free(ceph_cap_cachep, cap);
183 mdsc->caps_avail_count += nr_caps;
186 dout("%s: caps %d = %d used + %d resv + %d avail\n",
188 mdsc->caps_total_count, mdsc->caps_use_count,
189 mdsc->caps_reserve_count, mdsc->caps_avail_count);
190 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
191 mdsc->caps_reserve_count +
192 mdsc->caps_avail_count);
197 * Called under mdsc->mutex.
199 int ceph_reserve_caps(struct ceph_mds_client *mdsc,
200 struct ceph_cap_reservation *ctx, int need)
203 struct ceph_cap *cap;
208 bool trimmed = false;
209 struct ceph_mds_session *s;
212 dout("reserve caps ctx=%p need=%d\n", ctx, need);
214 /* first reserve any caps that are already allocated */
215 spin_lock(&mdsc->caps_list_lock);
216 if (mdsc->caps_avail_count >= need)
219 have = mdsc->caps_avail_count;
220 mdsc->caps_avail_count -= have;
221 mdsc->caps_reserve_count += have;
222 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
223 mdsc->caps_reserve_count +
224 mdsc->caps_avail_count);
225 spin_unlock(&mdsc->caps_list_lock);
227 for (i = have; i < need; ) {
228 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
230 list_add(&cap->caps_item, &newcaps);
237 for (j = 0; j < mdsc->max_sessions; j++) {
238 s = __ceph_lookup_mds_session(mdsc, j);
241 mutex_unlock(&mdsc->mutex);
243 mutex_lock(&s->s_mutex);
244 max_caps = s->s_nr_caps - (need - i);
245 ceph_trim_caps(mdsc, s, max_caps);
246 mutex_unlock(&s->s_mutex);
248 ceph_put_mds_session(s);
249 mutex_lock(&mdsc->mutex);
253 spin_lock(&mdsc->caps_list_lock);
254 if (mdsc->caps_avail_count) {
256 if (mdsc->caps_avail_count >= need - i)
257 more_have = need - i;
259 more_have = mdsc->caps_avail_count;
263 mdsc->caps_avail_count -= more_have;
264 mdsc->caps_reserve_count += more_have;
267 spin_unlock(&mdsc->caps_list_lock);
272 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
273 ctx, need, have + alloc);
279 BUG_ON(have + alloc != need);
284 spin_lock(&mdsc->caps_list_lock);
285 mdsc->caps_total_count += alloc;
286 mdsc->caps_reserve_count += alloc;
287 list_splice(&newcaps, &mdsc->caps_list);
289 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
290 mdsc->caps_reserve_count +
291 mdsc->caps_avail_count);
294 __ceph_unreserve_caps(mdsc, have + alloc);
296 spin_unlock(&mdsc->caps_list_lock);
298 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
299 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
300 mdsc->caps_reserve_count, mdsc->caps_avail_count);
304 void ceph_unreserve_caps(struct ceph_mds_client *mdsc,
305 struct ceph_cap_reservation *ctx)
307 bool reclaim = false;
311 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
312 spin_lock(&mdsc->caps_list_lock);
313 __ceph_unreserve_caps(mdsc, ctx->count);
316 if (mdsc->caps_use_max > 0 &&
317 mdsc->caps_use_count > mdsc->caps_use_max)
319 spin_unlock(&mdsc->caps_list_lock);
322 ceph_reclaim_caps_nr(mdsc, ctx->used);
325 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
326 struct ceph_cap_reservation *ctx)
328 struct ceph_cap *cap = NULL;
330 /* temporary, until we do something about cap import/export */
332 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
334 spin_lock(&mdsc->caps_list_lock);
335 mdsc->caps_use_count++;
336 mdsc->caps_total_count++;
337 spin_unlock(&mdsc->caps_list_lock);
339 spin_lock(&mdsc->caps_list_lock);
340 if (mdsc->caps_avail_count) {
341 BUG_ON(list_empty(&mdsc->caps_list));
343 mdsc->caps_avail_count--;
344 mdsc->caps_use_count++;
345 cap = list_first_entry(&mdsc->caps_list,
346 struct ceph_cap, caps_item);
347 list_del(&cap->caps_item);
349 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
350 mdsc->caps_reserve_count + mdsc->caps_avail_count);
352 spin_unlock(&mdsc->caps_list_lock);
358 spin_lock(&mdsc->caps_list_lock);
359 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
360 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
361 mdsc->caps_reserve_count, mdsc->caps_avail_count);
363 BUG_ON(ctx->count > mdsc->caps_reserve_count);
364 BUG_ON(list_empty(&mdsc->caps_list));
368 mdsc->caps_reserve_count--;
369 mdsc->caps_use_count++;
371 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
372 list_del(&cap->caps_item);
374 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
375 mdsc->caps_reserve_count + mdsc->caps_avail_count);
376 spin_unlock(&mdsc->caps_list_lock);
380 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
382 spin_lock(&mdsc->caps_list_lock);
383 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
384 cap, mdsc->caps_total_count, mdsc->caps_use_count,
385 mdsc->caps_reserve_count, mdsc->caps_avail_count);
386 mdsc->caps_use_count--;
388 * Keep some preallocated caps around (ceph_min_count), to
389 * avoid lots of free/alloc churn.
391 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
392 mdsc->caps_min_count) {
393 mdsc->caps_total_count--;
394 kmem_cache_free(ceph_cap_cachep, cap);
396 mdsc->caps_avail_count++;
397 list_add(&cap->caps_item, &mdsc->caps_list);
400 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
401 mdsc->caps_reserve_count + mdsc->caps_avail_count);
402 spin_unlock(&mdsc->caps_list_lock);
405 void ceph_reservation_status(struct ceph_fs_client *fsc,
406 int *total, int *avail, int *used, int *reserved,
409 struct ceph_mds_client *mdsc = fsc->mdsc;
411 spin_lock(&mdsc->caps_list_lock);
414 *total = mdsc->caps_total_count;
416 *avail = mdsc->caps_avail_count;
418 *used = mdsc->caps_use_count;
420 *reserved = mdsc->caps_reserve_count;
422 *min = mdsc->caps_min_count;
424 spin_unlock(&mdsc->caps_list_lock);
428 * Find ceph_cap for given mds, if any.
430 * Called with i_ceph_lock held.
432 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
434 struct ceph_cap *cap;
435 struct rb_node *n = ci->i_caps.rb_node;
438 cap = rb_entry(n, struct ceph_cap, ci_node);
441 else if (mds > cap->mds)
449 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
451 struct ceph_cap *cap;
453 spin_lock(&ci->i_ceph_lock);
454 cap = __get_cap_for_mds(ci, mds);
455 spin_unlock(&ci->i_ceph_lock);
460 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
462 static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
464 struct ceph_cap *cap;
468 /* prefer mds with WR|BUFFER|EXCL caps */
469 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
470 cap = rb_entry(p, struct ceph_cap, ci_node);
472 if (cap->issued & (CEPH_CAP_FILE_WR |
473 CEPH_CAP_FILE_BUFFER |
480 int ceph_get_cap_mds(struct inode *inode)
482 struct ceph_inode_info *ci = ceph_inode(inode);
484 spin_lock(&ci->i_ceph_lock);
485 mds = __ceph_get_cap_mds(ceph_inode(inode));
486 spin_unlock(&ci->i_ceph_lock);
491 * Called under i_ceph_lock.
493 static void __insert_cap_node(struct ceph_inode_info *ci,
494 struct ceph_cap *new)
496 struct rb_node **p = &ci->i_caps.rb_node;
497 struct rb_node *parent = NULL;
498 struct ceph_cap *cap = NULL;
502 cap = rb_entry(parent, struct ceph_cap, ci_node);
503 if (new->mds < cap->mds)
505 else if (new->mds > cap->mds)
511 rb_link_node(&new->ci_node, parent, p);
512 rb_insert_color(&new->ci_node, &ci->i_caps);
516 * (re)set cap hold timeouts, which control the delayed release
517 * of unused caps back to the MDS. Should be called on cap use.
519 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
520 struct ceph_inode_info *ci)
522 struct ceph_mount_options *opt = mdsc->fsc->mount_options;
524 ci->i_hold_caps_min = round_jiffies(jiffies +
525 opt->caps_wanted_delay_min * HZ);
526 ci->i_hold_caps_max = round_jiffies(jiffies +
527 opt->caps_wanted_delay_max * HZ);
528 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
529 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
533 * (Re)queue cap at the end of the delayed cap release list.
535 * If I_FLUSH is set, leave the inode at the front of the list.
537 * Caller holds i_ceph_lock
538 * -> we take mdsc->cap_delay_lock
540 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
541 struct ceph_inode_info *ci,
544 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
545 ci->i_ceph_flags, ci->i_hold_caps_max);
546 if (!mdsc->stopping) {
547 spin_lock(&mdsc->cap_delay_lock);
548 if (!list_empty(&ci->i_cap_delay_list)) {
549 if (ci->i_ceph_flags & CEPH_I_FLUSH)
551 list_del_init(&ci->i_cap_delay_list);
554 __cap_set_timeouts(mdsc, ci);
555 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
557 spin_unlock(&mdsc->cap_delay_lock);
562 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
563 * indicating we should send a cap message to flush dirty metadata
564 * asap, and move to the front of the delayed cap list.
566 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
567 struct ceph_inode_info *ci)
569 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
570 spin_lock(&mdsc->cap_delay_lock);
571 ci->i_ceph_flags |= CEPH_I_FLUSH;
572 if (!list_empty(&ci->i_cap_delay_list))
573 list_del_init(&ci->i_cap_delay_list);
574 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
575 spin_unlock(&mdsc->cap_delay_lock);
579 * Cancel delayed work on cap.
581 * Caller must hold i_ceph_lock.
583 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
584 struct ceph_inode_info *ci)
586 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
587 if (list_empty(&ci->i_cap_delay_list))
589 spin_lock(&mdsc->cap_delay_lock);
590 list_del_init(&ci->i_cap_delay_list);
591 spin_unlock(&mdsc->cap_delay_lock);
595 * Common issue checks for add_cap, handle_cap_grant.
597 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
600 unsigned had = __ceph_caps_issued(ci, NULL);
603 * Each time we receive FILE_CACHE anew, we increment
606 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
607 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
612 * If FILE_SHARED is newly issued, mark dir not complete. We don't
613 * know what happened to this directory while we didn't have the cap.
614 * If FILE_SHARED is being revoked, also mark dir not complete. It
615 * stops on-going cached readdir.
617 if ((issued & CEPH_CAP_FILE_SHARED) != (had & CEPH_CAP_FILE_SHARED)) {
618 if (issued & CEPH_CAP_FILE_SHARED)
619 atomic_inc(&ci->i_shared_gen);
620 if (S_ISDIR(ci->vfs_inode.i_mode)) {
621 dout(" marking %p NOT complete\n", &ci->vfs_inode);
622 __ceph_dir_clear_complete(ci);
628 * Add a capability under the given MDS session.
630 * Caller should hold session snap_rwsem (read) and s_mutex.
632 * @fmode is the open file mode, if we are opening a file, otherwise
633 * it is < 0. (This is so we can atomically add the cap and add an
634 * open file reference to it.)
636 void ceph_add_cap(struct inode *inode,
637 struct ceph_mds_session *session, u64 cap_id,
638 int fmode, unsigned issued, unsigned wanted,
639 unsigned seq, unsigned mseq, u64 realmino, int flags,
640 struct ceph_cap **new_cap)
642 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
643 struct ceph_inode_info *ci = ceph_inode(inode);
644 struct ceph_cap *cap;
645 int mds = session->s_mds;
648 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
649 session->s_mds, cap_id, ceph_cap_string(issued), seq);
652 * If we are opening the file, include file mode wanted bits
656 wanted |= ceph_caps_for_mode(fmode);
658 cap = __get_cap_for_mds(ci, mds);
664 cap->implemented = 0;
670 __insert_cap_node(ci, cap);
672 /* add to session cap list */
673 cap->session = session;
674 spin_lock(&session->s_cap_lock);
675 list_add_tail(&cap->session_caps, &session->s_caps);
676 session->s_nr_caps++;
677 spin_unlock(&session->s_cap_lock);
679 spin_lock(&session->s_cap_lock);
680 list_move_tail(&cap->session_caps, &session->s_caps);
681 spin_unlock(&session->s_cap_lock);
683 if (cap->cap_gen < session->s_cap_gen)
684 cap->issued = cap->implemented = CEPH_CAP_PIN;
687 * auth mds of the inode changed. we received the cap export
688 * message, but still haven't received the cap import message.
689 * handle_cap_export() updated the new auth MDS' cap.
691 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
692 * a message that was send before the cap import message. So
695 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
696 WARN_ON(cap != ci->i_auth_cap);
697 WARN_ON(cap->cap_id != cap_id);
700 issued |= cap->issued;
701 flags |= CEPH_CAP_FLAG_AUTH;
705 if (!ci->i_snap_realm ||
706 ((flags & CEPH_CAP_FLAG_AUTH) &&
707 realmino != (u64)-1 && ci->i_snap_realm->ino != realmino)) {
709 * add this inode to the appropriate snap realm
711 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
714 struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
716 spin_lock(&oldrealm->inodes_with_caps_lock);
717 list_del_init(&ci->i_snap_realm_item);
718 spin_unlock(&oldrealm->inodes_with_caps_lock);
721 spin_lock(&realm->inodes_with_caps_lock);
722 list_add(&ci->i_snap_realm_item,
723 &realm->inodes_with_caps);
724 ci->i_snap_realm = realm;
725 if (realm->ino == ci->i_vino.ino)
726 realm->inode = inode;
727 spin_unlock(&realm->inodes_with_caps_lock);
730 ceph_put_snap_realm(mdsc, oldrealm);
732 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
738 __check_cap_issue(ci, cap, issued);
741 * If we are issued caps we don't want, or the mds' wanted
742 * value appears to be off, queue a check so we'll release
743 * later and/or update the mds wanted value.
745 actual_wanted = __ceph_caps_wanted(ci);
746 if ((wanted & ~actual_wanted) ||
747 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
748 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
749 ceph_cap_string(issued), ceph_cap_string(wanted),
750 ceph_cap_string(actual_wanted));
751 __cap_delay_requeue(mdsc, ci, true);
754 if (flags & CEPH_CAP_FLAG_AUTH) {
755 if (!ci->i_auth_cap ||
756 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
757 ci->i_auth_cap = cap;
758 cap->mds_wanted = wanted;
761 WARN_ON(ci->i_auth_cap == cap);
764 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
765 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
766 ceph_cap_string(issued|cap->issued), seq, mds);
767 cap->cap_id = cap_id;
768 cap->issued = issued;
769 cap->implemented |= issued;
770 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
771 cap->mds_wanted = wanted;
773 cap->mds_wanted |= wanted;
775 cap->issue_seq = seq;
777 cap->cap_gen = session->s_cap_gen;
780 __ceph_get_fmode(ci, fmode);
784 * Return true if cap has not timed out and belongs to the current
785 * generation of the MDS session (i.e. has not gone 'stale' due to
786 * us losing touch with the mds).
788 static int __cap_is_valid(struct ceph_cap *cap)
793 spin_lock(&cap->session->s_gen_ttl_lock);
794 gen = cap->session->s_cap_gen;
795 ttl = cap->session->s_cap_ttl;
796 spin_unlock(&cap->session->s_gen_ttl_lock);
798 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
799 dout("__cap_is_valid %p cap %p issued %s "
800 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
801 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
809 * Return set of valid cap bits issued to us. Note that caps time
810 * out, and may be invalidated in bulk if the client session times out
811 * and session->s_cap_gen is bumped.
813 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
815 int have = ci->i_snap_caps;
816 struct ceph_cap *cap;
821 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
822 cap = rb_entry(p, struct ceph_cap, ci_node);
823 if (!__cap_is_valid(cap))
825 dout("__ceph_caps_issued %p cap %p issued %s\n",
826 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
829 *implemented |= cap->implemented;
832 * exclude caps issued by non-auth MDS, but are been revoking
833 * by the auth MDS. The non-auth MDS should be revoking/exporting
834 * these caps, but the message is delayed.
836 if (ci->i_auth_cap) {
837 cap = ci->i_auth_cap;
838 have &= ~cap->implemented | cap->issued;
844 * Get cap bits issued by caps other than @ocap
846 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
848 int have = ci->i_snap_caps;
849 struct ceph_cap *cap;
852 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
853 cap = rb_entry(p, struct ceph_cap, ci_node);
856 if (!__cap_is_valid(cap))
864 * Move a cap to the end of the LRU (oldest caps at list head, newest
867 static void __touch_cap(struct ceph_cap *cap)
869 struct ceph_mds_session *s = cap->session;
871 spin_lock(&s->s_cap_lock);
872 if (!s->s_cap_iterator) {
873 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
875 list_move_tail(&cap->session_caps, &s->s_caps);
877 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
878 &cap->ci->vfs_inode, cap, s->s_mds);
880 spin_unlock(&s->s_cap_lock);
884 * Check if we hold the given mask. If so, move the cap(s) to the
885 * front of their respective LRUs. (This is the preferred way for
886 * callers to check for caps they want.)
888 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
890 struct ceph_cap *cap;
892 int have = ci->i_snap_caps;
894 if ((have & mask) == mask) {
895 dout("__ceph_caps_issued_mask ino 0x%lx snap issued %s"
896 " (mask %s)\n", ci->vfs_inode.i_ino,
897 ceph_cap_string(have),
898 ceph_cap_string(mask));
902 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
903 cap = rb_entry(p, struct ceph_cap, ci_node);
904 if (!__cap_is_valid(cap))
906 if ((cap->issued & mask) == mask) {
907 dout("__ceph_caps_issued_mask ino 0x%lx cap %p issued %s"
908 " (mask %s)\n", ci->vfs_inode.i_ino, cap,
909 ceph_cap_string(cap->issued),
910 ceph_cap_string(mask));
916 /* does a combination of caps satisfy mask? */
918 if ((have & mask) == mask) {
919 dout("__ceph_caps_issued_mask ino 0x%lx combo issued %s"
920 " (mask %s)\n", ci->vfs_inode.i_ino,
921 ceph_cap_string(cap->issued),
922 ceph_cap_string(mask));
926 /* touch this + preceding caps */
928 for (q = rb_first(&ci->i_caps); q != p;
930 cap = rb_entry(q, struct ceph_cap,
932 if (!__cap_is_valid(cap))
945 * Return true if mask caps are currently being revoked by an MDS.
947 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
948 struct ceph_cap *ocap, int mask)
950 struct ceph_cap *cap;
953 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
954 cap = rb_entry(p, struct ceph_cap, ci_node);
956 (cap->implemented & ~cap->issued & mask))
962 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
964 struct inode *inode = &ci->vfs_inode;
967 spin_lock(&ci->i_ceph_lock);
968 ret = __ceph_caps_revoking_other(ci, NULL, mask);
969 spin_unlock(&ci->i_ceph_lock);
970 dout("ceph_caps_revoking %p %s = %d\n", inode,
971 ceph_cap_string(mask), ret);
975 int __ceph_caps_used(struct ceph_inode_info *ci)
979 used |= CEPH_CAP_PIN;
981 used |= CEPH_CAP_FILE_RD;
982 if (ci->i_rdcache_ref ||
983 (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */
984 ci->vfs_inode.i_data.nrpages))
985 used |= CEPH_CAP_FILE_CACHE;
987 used |= CEPH_CAP_FILE_WR;
988 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
989 used |= CEPH_CAP_FILE_BUFFER;
994 * wanted, by virtue of open file modes
996 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
999 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
1000 if (ci->i_nr_by_mode[i])
1005 return ceph_caps_for_mode(bits >> 1);
1009 * Return caps we have registered with the MDS(s) as 'wanted'.
1011 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
1013 struct ceph_cap *cap;
1017 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1018 cap = rb_entry(p, struct ceph_cap, ci_node);
1019 if (check && !__cap_is_valid(cap))
1021 if (cap == ci->i_auth_cap)
1022 mds_wanted |= cap->mds_wanted;
1024 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
1030 * called under i_ceph_lock
1032 static int __ceph_is_single_caps(struct ceph_inode_info *ci)
1034 return rb_first(&ci->i_caps) == rb_last(&ci->i_caps);
1037 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
1039 return !RB_EMPTY_ROOT(&ci->i_caps);
1042 int ceph_is_any_caps(struct inode *inode)
1044 struct ceph_inode_info *ci = ceph_inode(inode);
1047 spin_lock(&ci->i_ceph_lock);
1048 ret = __ceph_is_any_caps(ci);
1049 spin_unlock(&ci->i_ceph_lock);
1054 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
1056 struct ceph_snap_realm *realm = ci->i_snap_realm;
1057 spin_lock(&realm->inodes_with_caps_lock);
1058 list_del_init(&ci->i_snap_realm_item);
1059 ci->i_snap_realm_counter++;
1060 ci->i_snap_realm = NULL;
1061 if (realm->ino == ci->i_vino.ino)
1062 realm->inode = NULL;
1063 spin_unlock(&realm->inodes_with_caps_lock);
1064 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
1069 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
1071 * caller should hold i_ceph_lock.
1072 * caller will not hold session s_mutex if called from destroy_inode.
1074 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1076 struct ceph_mds_session *session = cap->session;
1077 struct ceph_inode_info *ci = cap->ci;
1078 struct ceph_mds_client *mdsc =
1079 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1082 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1084 /* remove from session list */
1085 spin_lock(&session->s_cap_lock);
1086 if (session->s_cap_iterator == cap) {
1087 /* not yet, we are iterating over this very cap */
1088 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1091 list_del_init(&cap->session_caps);
1092 session->s_nr_caps--;
1093 cap->session = NULL;
1096 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1100 * s_cap_reconnect is protected by s_cap_lock. no one changes
1101 * s_cap_gen while session is in the reconnect state.
1103 if (queue_release &&
1104 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
1105 cap->queue_release = 1;
1107 __ceph_queue_cap_release(session, cap);
1111 cap->queue_release = 0;
1113 cap->cap_ino = ci->i_vino.ino;
1115 spin_unlock(&session->s_cap_lock);
1117 /* remove from inode list */
1118 rb_erase(&cap->ci_node, &ci->i_caps);
1119 if (ci->i_auth_cap == cap)
1120 ci->i_auth_cap = NULL;
1123 ceph_put_cap(mdsc, cap);
1125 /* when reconnect denied, we remove session caps forcibly,
1126 * i_wr_ref can be non-zero. If there are ongoing write,
1127 * keep i_snap_realm.
1129 if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm)
1130 drop_inode_snap_realm(ci);
1132 if (!__ceph_is_any_real_caps(ci))
1133 __cap_delay_cancel(mdsc, ci);
1136 struct cap_msg_args {
1137 struct ceph_mds_session *session;
1138 u64 ino, cid, follows;
1139 u64 flush_tid, oldest_flush_tid, size, max_size;
1141 struct ceph_buffer *xattr_buf;
1142 struct timespec64 atime, mtime, ctime;
1143 int op, caps, wanted, dirty;
1144 u32 seq, issue_seq, mseq, time_warp_seq;
1153 * Build and send a cap message to the given MDS.
1155 * Caller should be holding s_mutex.
1157 static int send_cap_msg(struct cap_msg_args *arg)
1159 struct ceph_mds_caps *fc;
1160 struct ceph_msg *msg;
1163 struct timespec64 zerotime = {0};
1164 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1166 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1167 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1168 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(arg->op),
1169 arg->cid, arg->ino, ceph_cap_string(arg->caps),
1170 ceph_cap_string(arg->wanted), ceph_cap_string(arg->dirty),
1171 arg->seq, arg->issue_seq, arg->flush_tid, arg->oldest_flush_tid,
1172 arg->mseq, arg->follows, arg->size, arg->max_size,
1174 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1176 /* flock buffer size + inline version + inline data size +
1177 * osd_epoch_barrier + oldest_flush_tid */
1178 extra_len = 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4;
1179 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1184 msg->hdr.version = cpu_to_le16(10);
1185 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1187 fc = msg->front.iov_base;
1188 memset(fc, 0, sizeof(*fc));
1190 fc->cap_id = cpu_to_le64(arg->cid);
1191 fc->op = cpu_to_le32(arg->op);
1192 fc->seq = cpu_to_le32(arg->seq);
1193 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1194 fc->migrate_seq = cpu_to_le32(arg->mseq);
1195 fc->caps = cpu_to_le32(arg->caps);
1196 fc->wanted = cpu_to_le32(arg->wanted);
1197 fc->dirty = cpu_to_le32(arg->dirty);
1198 fc->ino = cpu_to_le64(arg->ino);
1199 fc->snap_follows = cpu_to_le64(arg->follows);
1201 fc->size = cpu_to_le64(arg->size);
1202 fc->max_size = cpu_to_le64(arg->max_size);
1203 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1204 ceph_encode_timespec64(&fc->atime, &arg->atime);
1205 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1206 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1208 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1209 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1210 fc->mode = cpu_to_le32(arg->mode);
1212 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1213 if (arg->xattr_buf) {
1214 msg->middle = ceph_buffer_get(arg->xattr_buf);
1215 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1216 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1220 /* flock buffer size (version 2) */
1221 ceph_encode_32(&p, 0);
1222 /* inline version (version 4) */
1223 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1224 /* inline data size */
1225 ceph_encode_32(&p, 0);
1227 * osd_epoch_barrier (version 5)
1228 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1229 * case it was recently changed
1231 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1232 /* oldest_flush_tid (version 6) */
1233 ceph_encode_64(&p, arg->oldest_flush_tid);
1236 * caller_uid/caller_gid (version 7)
1238 * Currently, we don't properly track which caller dirtied the caps
1239 * last, and force a flush of them when there is a conflict. For now,
1240 * just set this to 0:0, to emulate how the MDS has worked up to now.
1242 ceph_encode_32(&p, 0);
1243 ceph_encode_32(&p, 0);
1245 /* pool namespace (version 8) (mds always ignores this) */
1246 ceph_encode_32(&p, 0);
1249 * btime and change_attr (version 9)
1251 * We just zero these out for now, as the MDS ignores them unless
1252 * the requisite feature flags are set (which we don't do yet).
1254 ceph_encode_timespec64(p, &zerotime);
1255 p += sizeof(struct ceph_timespec);
1256 ceph_encode_64(&p, 0);
1258 /* Advisory flags (version 10) */
1259 ceph_encode_32(&p, arg->flags);
1261 ceph_con_send(&arg->session->s_con, msg);
1266 * Queue cap releases when an inode is dropped from our cache. Since
1267 * inode is about to be destroyed, there is no need for i_ceph_lock.
1269 void __ceph_remove_caps(struct inode *inode)
1271 struct ceph_inode_info *ci = ceph_inode(inode);
1274 p = rb_first(&ci->i_caps);
1276 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1278 __ceph_remove_cap(cap, true);
1283 * Send a cap msg on the given inode. Update our caps state, then
1284 * drop i_ceph_lock and send the message.
1286 * Make note of max_size reported/requested from mds, revoked caps
1287 * that have now been implemented.
1289 * Make half-hearted attempt ot to invalidate page cache if we are
1290 * dropping RDCACHE. Note that this will leave behind locked pages
1291 * that we'll then need to deal with elsewhere.
1293 * Return non-zero if delayed release, or we experienced an error
1294 * such that the caller should requeue + retry later.
1296 * called with i_ceph_lock, then drops it.
1297 * caller should hold snap_rwsem (read), s_mutex.
1299 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1300 int op, bool sync, int used, int want, int retain,
1301 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1302 __releases(cap->ci->i_ceph_lock)
1304 struct ceph_inode_info *ci = cap->ci;
1305 struct inode *inode = &ci->vfs_inode;
1306 struct cap_msg_args arg;
1312 held = cap->issued | cap->implemented;
1313 revoking = cap->implemented & ~cap->issued;
1314 retain &= ~revoking;
1316 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1317 inode, cap, cap->session,
1318 ceph_cap_string(held), ceph_cap_string(held & retain),
1319 ceph_cap_string(revoking));
1320 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1322 arg.session = cap->session;
1324 /* don't release wanted unless we've waited a bit. */
1325 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1326 time_before(jiffies, ci->i_hold_caps_min)) {
1327 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1328 ceph_cap_string(cap->issued),
1329 ceph_cap_string(cap->issued & retain),
1330 ceph_cap_string(cap->mds_wanted),
1331 ceph_cap_string(want));
1332 want |= cap->mds_wanted;
1333 retain |= cap->issued;
1336 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1337 if (want & ~cap->mds_wanted) {
1338 /* user space may open/close single file frequently.
1339 * This avoids droping mds_wanted immediately after
1340 * requesting new mds_wanted.
1342 __cap_set_timeouts(mdsc, ci);
1345 cap->issued &= retain; /* drop bits we don't want */
1346 if (cap->implemented & ~cap->issued) {
1348 * Wake up any waiters on wanted -> needed transition.
1349 * This is due to the weird transition from buffered
1350 * to sync IO... we need to flush dirty pages _before_
1351 * allowing sync writes to avoid reordering.
1355 cap->implemented &= cap->issued | used;
1356 cap->mds_wanted = want;
1358 arg.ino = ceph_vino(inode).ino;
1359 arg.cid = cap->cap_id;
1360 arg.follows = flushing ? ci->i_head_snapc->seq : 0;
1361 arg.flush_tid = flush_tid;
1362 arg.oldest_flush_tid = oldest_flush_tid;
1364 arg.size = inode->i_size;
1365 ci->i_reported_size = arg.size;
1366 arg.max_size = ci->i_wanted_max_size;
1367 ci->i_requested_max_size = arg.max_size;
1369 if (flushing & CEPH_CAP_XATTR_EXCL) {
1370 __ceph_build_xattrs_blob(ci);
1371 arg.xattr_version = ci->i_xattrs.version;
1372 arg.xattr_buf = ci->i_xattrs.blob;
1374 arg.xattr_buf = NULL;
1377 arg.mtime = inode->i_mtime;
1378 arg.atime = inode->i_atime;
1379 arg.ctime = inode->i_ctime;
1382 arg.caps = cap->implemented;
1384 arg.dirty = flushing;
1387 arg.issue_seq = cap->issue_seq;
1388 arg.mseq = cap->mseq;
1389 arg.time_warp_seq = ci->i_time_warp_seq;
1391 arg.uid = inode->i_uid;
1392 arg.gid = inode->i_gid;
1393 arg.mode = inode->i_mode;
1395 arg.inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1396 if (list_empty(&ci->i_cap_snaps))
1397 arg.flags = CEPH_CLIENT_CAPS_NO_CAPSNAP;
1399 arg.flags = CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1401 arg.flags |= CEPH_CLIENT_CAPS_SYNC;
1403 spin_unlock(&ci->i_ceph_lock);
1405 ret = send_cap_msg(&arg);
1407 dout("error sending cap msg, must requeue %p\n", inode);
1412 wake_up_all(&ci->i_cap_wq);
1417 static inline int __send_flush_snap(struct inode *inode,
1418 struct ceph_mds_session *session,
1419 struct ceph_cap_snap *capsnap,
1420 u32 mseq, u64 oldest_flush_tid)
1422 struct cap_msg_args arg;
1424 arg.session = session;
1425 arg.ino = ceph_vino(inode).ino;
1427 arg.follows = capsnap->follows;
1428 arg.flush_tid = capsnap->cap_flush.tid;
1429 arg.oldest_flush_tid = oldest_flush_tid;
1431 arg.size = capsnap->size;
1433 arg.xattr_version = capsnap->xattr_version;
1434 arg.xattr_buf = capsnap->xattr_blob;
1436 arg.atime = capsnap->atime;
1437 arg.mtime = capsnap->mtime;
1438 arg.ctime = capsnap->ctime;
1440 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1441 arg.caps = capsnap->issued;
1443 arg.dirty = capsnap->dirty;
1448 arg.time_warp_seq = capsnap->time_warp_seq;
1450 arg.uid = capsnap->uid;
1451 arg.gid = capsnap->gid;
1452 arg.mode = capsnap->mode;
1454 arg.inline_data = capsnap->inline_data;
1457 return send_cap_msg(&arg);
1461 * When a snapshot is taken, clients accumulate dirty metadata on
1462 * inodes with capabilities in ceph_cap_snaps to describe the file
1463 * state at the time the snapshot was taken. This must be flushed
1464 * asynchronously back to the MDS once sync writes complete and dirty
1465 * data is written out.
1467 * Called under i_ceph_lock. Takes s_mutex as needed.
1469 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1470 struct ceph_mds_session *session)
1471 __releases(ci->i_ceph_lock)
1472 __acquires(ci->i_ceph_lock)
1474 struct inode *inode = &ci->vfs_inode;
1475 struct ceph_mds_client *mdsc = session->s_mdsc;
1476 struct ceph_cap_snap *capsnap;
1477 u64 oldest_flush_tid = 0;
1478 u64 first_tid = 1, last_tid = 0;
1480 dout("__flush_snaps %p session %p\n", inode, session);
1482 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1484 * we need to wait for sync writes to complete and for dirty
1485 * pages to be written out.
1487 if (capsnap->dirty_pages || capsnap->writing)
1490 /* should be removed by ceph_try_drop_cap_snap() */
1491 BUG_ON(!capsnap->need_flush);
1493 /* only flush each capsnap once */
1494 if (capsnap->cap_flush.tid > 0) {
1495 dout(" already flushed %p, skipping\n", capsnap);
1499 spin_lock(&mdsc->cap_dirty_lock);
1500 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1501 list_add_tail(&capsnap->cap_flush.g_list,
1502 &mdsc->cap_flush_list);
1503 if (oldest_flush_tid == 0)
1504 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1505 if (list_empty(&ci->i_flushing_item)) {
1506 list_add_tail(&ci->i_flushing_item,
1507 &session->s_cap_flushing);
1509 spin_unlock(&mdsc->cap_dirty_lock);
1511 list_add_tail(&capsnap->cap_flush.i_list,
1512 &ci->i_cap_flush_list);
1515 first_tid = capsnap->cap_flush.tid;
1516 last_tid = capsnap->cap_flush.tid;
1519 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1521 while (first_tid <= last_tid) {
1522 struct ceph_cap *cap = ci->i_auth_cap;
1523 struct ceph_cap_flush *cf;
1526 if (!(cap && cap->session == session)) {
1527 dout("__flush_snaps %p auth cap %p not mds%d, "
1528 "stop\n", inode, cap, session->s_mds);
1533 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1534 if (cf->tid >= first_tid) {
1542 first_tid = cf->tid + 1;
1544 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1545 refcount_inc(&capsnap->nref);
1546 spin_unlock(&ci->i_ceph_lock);
1548 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1549 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1551 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1554 pr_err("__flush_snaps: error sending cap flushsnap, "
1555 "ino (%llx.%llx) tid %llu follows %llu\n",
1556 ceph_vinop(inode), cf->tid, capsnap->follows);
1559 ceph_put_cap_snap(capsnap);
1560 spin_lock(&ci->i_ceph_lock);
1564 void ceph_flush_snaps(struct ceph_inode_info *ci,
1565 struct ceph_mds_session **psession)
1567 struct inode *inode = &ci->vfs_inode;
1568 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1569 struct ceph_mds_session *session = NULL;
1572 dout("ceph_flush_snaps %p\n", inode);
1574 session = *psession;
1576 spin_lock(&ci->i_ceph_lock);
1577 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1578 dout(" no capsnap needs flush, doing nothing\n");
1581 if (!ci->i_auth_cap) {
1582 dout(" no auth cap (migrating?), doing nothing\n");
1586 mds = ci->i_auth_cap->session->s_mds;
1587 if (session && session->s_mds != mds) {
1588 dout(" oops, wrong session %p mutex\n", session);
1589 mutex_unlock(&session->s_mutex);
1590 ceph_put_mds_session(session);
1594 spin_unlock(&ci->i_ceph_lock);
1595 mutex_lock(&mdsc->mutex);
1596 session = __ceph_lookup_mds_session(mdsc, mds);
1597 mutex_unlock(&mdsc->mutex);
1599 dout(" inverting session/ino locks on %p\n", session);
1600 mutex_lock(&session->s_mutex);
1605 // make sure flushsnap messages are sent in proper order.
1606 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
1607 __kick_flushing_caps(mdsc, session, ci, 0);
1608 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
1611 __ceph_flush_snaps(ci, session);
1613 spin_unlock(&ci->i_ceph_lock);
1616 *psession = session;
1617 } else if (session) {
1618 mutex_unlock(&session->s_mutex);
1619 ceph_put_mds_session(session);
1621 /* we flushed them all; remove this inode from the queue */
1622 spin_lock(&mdsc->snap_flush_lock);
1623 list_del_init(&ci->i_snap_flush_item);
1624 spin_unlock(&mdsc->snap_flush_lock);
1628 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1629 * Caller is then responsible for calling __mark_inode_dirty with the
1630 * returned flags value.
1632 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1633 struct ceph_cap_flush **pcf)
1635 struct ceph_mds_client *mdsc =
1636 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1637 struct inode *inode = &ci->vfs_inode;
1638 int was = ci->i_dirty_caps;
1641 if (!ci->i_auth_cap) {
1642 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1643 "but no auth cap (session was closed?)\n",
1644 inode, ceph_ino(inode), ceph_cap_string(mask));
1648 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1649 ceph_cap_string(mask), ceph_cap_string(was),
1650 ceph_cap_string(was | mask));
1651 ci->i_dirty_caps |= mask;
1653 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1654 swap(ci->i_prealloc_cap_flush, *pcf);
1656 if (!ci->i_head_snapc) {
1657 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1658 ci->i_head_snapc = ceph_get_snap_context(
1659 ci->i_snap_realm->cached_context);
1661 dout(" inode %p now dirty snapc %p auth cap %p\n",
1662 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1663 BUG_ON(!list_empty(&ci->i_dirty_item));
1664 spin_lock(&mdsc->cap_dirty_lock);
1665 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1666 spin_unlock(&mdsc->cap_dirty_lock);
1667 if (ci->i_flushing_caps == 0) {
1669 dirty |= I_DIRTY_SYNC;
1672 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1674 BUG_ON(list_empty(&ci->i_dirty_item));
1675 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1676 (mask & CEPH_CAP_FILE_BUFFER))
1677 dirty |= I_DIRTY_DATASYNC;
1678 __cap_delay_requeue(mdsc, ci, true);
1682 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1684 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1687 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1690 kmem_cache_free(ceph_cap_flush_cachep, cf);
1693 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1695 if (!list_empty(&mdsc->cap_flush_list)) {
1696 struct ceph_cap_flush *cf =
1697 list_first_entry(&mdsc->cap_flush_list,
1698 struct ceph_cap_flush, g_list);
1705 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1706 * Return true if caller needs to wake up flush waiters.
1708 static bool __finish_cap_flush(struct ceph_mds_client *mdsc,
1709 struct ceph_inode_info *ci,
1710 struct ceph_cap_flush *cf)
1712 struct ceph_cap_flush *prev;
1713 bool wake = cf->wake;
1715 /* are there older pending cap flushes? */
1716 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1717 prev = list_prev_entry(cf, g_list);
1721 list_del(&cf->g_list);
1723 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1724 prev = list_prev_entry(cf, i_list);
1728 list_del(&cf->i_list);
1736 * Add dirty inode to the flushing list. Assigned a seq number so we
1737 * can wait for caps to flush without starving.
1739 * Called under i_ceph_lock.
1741 static int __mark_caps_flushing(struct inode *inode,
1742 struct ceph_mds_session *session, bool wake,
1743 u64 *flush_tid, u64 *oldest_flush_tid)
1745 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1746 struct ceph_inode_info *ci = ceph_inode(inode);
1747 struct ceph_cap_flush *cf = NULL;
1750 BUG_ON(ci->i_dirty_caps == 0);
1751 BUG_ON(list_empty(&ci->i_dirty_item));
1752 BUG_ON(!ci->i_prealloc_cap_flush);
1754 flushing = ci->i_dirty_caps;
1755 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1756 ceph_cap_string(flushing),
1757 ceph_cap_string(ci->i_flushing_caps),
1758 ceph_cap_string(ci->i_flushing_caps | flushing));
1759 ci->i_flushing_caps |= flushing;
1760 ci->i_dirty_caps = 0;
1761 dout(" inode %p now !dirty\n", inode);
1763 swap(cf, ci->i_prealloc_cap_flush);
1764 cf->caps = flushing;
1767 spin_lock(&mdsc->cap_dirty_lock);
1768 list_del_init(&ci->i_dirty_item);
1770 cf->tid = ++mdsc->last_cap_flush_tid;
1771 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1772 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1774 if (list_empty(&ci->i_flushing_item)) {
1775 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1776 mdsc->num_cap_flushing++;
1778 spin_unlock(&mdsc->cap_dirty_lock);
1780 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1782 *flush_tid = cf->tid;
1787 * try to invalidate mapping pages without blocking.
1789 static int try_nonblocking_invalidate(struct inode *inode)
1791 struct ceph_inode_info *ci = ceph_inode(inode);
1792 u32 invalidating_gen = ci->i_rdcache_gen;
1794 spin_unlock(&ci->i_ceph_lock);
1795 invalidate_mapping_pages(&inode->i_data, 0, -1);
1796 spin_lock(&ci->i_ceph_lock);
1798 if (inode->i_data.nrpages == 0 &&
1799 invalidating_gen == ci->i_rdcache_gen) {
1801 dout("try_nonblocking_invalidate %p success\n", inode);
1802 /* save any racing async invalidate some trouble */
1803 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1806 dout("try_nonblocking_invalidate %p failed\n", inode);
1810 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1812 loff_t size = ci->vfs_inode.i_size;
1813 /* mds will adjust max size according to the reported size */
1814 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1816 if (size >= ci->i_max_size)
1818 /* half of previous max_size increment has been used */
1819 if (ci->i_max_size > ci->i_reported_size &&
1820 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1826 * Swiss army knife function to examine currently used and wanted
1827 * versus held caps. Release, flush, ack revoked caps to mds as
1830 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1831 * cap release further.
1832 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1833 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1836 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1837 struct ceph_mds_session *session)
1839 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1840 struct ceph_mds_client *mdsc = fsc->mdsc;
1841 struct inode *inode = &ci->vfs_inode;
1842 struct ceph_cap *cap;
1843 u64 flush_tid, oldest_flush_tid;
1844 int file_wanted, used, cap_used;
1845 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1846 int issued, implemented, want, retain, revoking, flushing = 0;
1847 int mds = -1; /* keep track of how far we've gone through i_caps list
1848 to avoid an infinite loop on retry */
1850 int delayed = 0, sent = 0;
1851 bool no_delay = flags & CHECK_CAPS_NODELAY;
1852 bool queue_invalidate = false;
1853 bool tried_invalidate = false;
1855 /* if we are unmounting, flush any unused caps immediately. */
1859 spin_lock(&ci->i_ceph_lock);
1861 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1862 flags |= CHECK_CAPS_FLUSH;
1864 if (!(flags & CHECK_CAPS_AUTHONLY) ||
1865 (ci->i_auth_cap && __ceph_is_single_caps(ci)))
1866 __cap_delay_cancel(mdsc, ci);
1870 spin_lock(&ci->i_ceph_lock);
1872 file_wanted = __ceph_caps_file_wanted(ci);
1873 used = __ceph_caps_used(ci);
1874 issued = __ceph_caps_issued(ci, &implemented);
1875 revoking = implemented & ~issued;
1878 retain = file_wanted | used | CEPH_CAP_PIN;
1879 if (!mdsc->stopping && inode->i_nlink > 0) {
1881 retain |= CEPH_CAP_ANY; /* be greedy */
1882 } else if (S_ISDIR(inode->i_mode) &&
1883 (issued & CEPH_CAP_FILE_SHARED) &&
1884 __ceph_dir_is_complete(ci)) {
1886 * If a directory is complete, we want to keep
1887 * the exclusive cap. So that MDS does not end up
1888 * revoking the shared cap on every create/unlink
1891 if (IS_RDONLY(inode))
1892 want = CEPH_CAP_ANY_SHARED;
1894 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1898 retain |= CEPH_CAP_ANY_SHARED;
1900 * keep RD only if we didn't have the file open RW,
1901 * because then the mds would revoke it anyway to
1902 * journal max_size=0.
1904 if (ci->i_max_size == 0)
1905 retain |= CEPH_CAP_ANY_RD;
1909 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1910 " issued %s revoking %s retain %s %s%s%s\n", inode,
1911 ceph_cap_string(file_wanted),
1912 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1913 ceph_cap_string(ci->i_flushing_caps),
1914 ceph_cap_string(issued), ceph_cap_string(revoking),
1915 ceph_cap_string(retain),
1916 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1917 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1918 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1921 * If we no longer need to hold onto old our caps, and we may
1922 * have cached pages, but don't want them, then try to invalidate.
1923 * If we fail, it's because pages are locked.... try again later.
1925 if ((!no_delay || mdsc->stopping) &&
1926 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */
1927 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1928 inode->i_data.nrpages && /* have cached pages */
1929 (revoking & (CEPH_CAP_FILE_CACHE|
1930 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1931 !tried_invalidate) {
1932 dout("check_caps trying to invalidate on %p\n", inode);
1933 if (try_nonblocking_invalidate(inode) < 0) {
1934 dout("check_caps queuing invalidate\n");
1935 queue_invalidate = true;
1936 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1938 tried_invalidate = true;
1942 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1943 cap = rb_entry(p, struct ceph_cap, ci_node);
1945 /* avoid looping forever */
1946 if (mds >= cap->mds ||
1947 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1950 /* NOTE: no side-effects allowed, until we take s_mutex */
1953 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1954 cap_used &= ~ci->i_auth_cap->issued;
1956 revoking = cap->implemented & ~cap->issued;
1957 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1958 cap->mds, cap, ceph_cap_string(cap_used),
1959 ceph_cap_string(cap->issued),
1960 ceph_cap_string(cap->implemented),
1961 ceph_cap_string(revoking));
1963 if (cap == ci->i_auth_cap &&
1964 (cap->issued & CEPH_CAP_FILE_WR)) {
1965 /* request larger max_size from MDS? */
1966 if (ci->i_wanted_max_size > ci->i_max_size &&
1967 ci->i_wanted_max_size > ci->i_requested_max_size) {
1968 dout("requesting new max_size\n");
1972 /* approaching file_max? */
1973 if (__ceph_should_report_size(ci)) {
1974 dout("i_size approaching max_size\n");
1978 /* flush anything dirty? */
1979 if (cap == ci->i_auth_cap) {
1980 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
1981 dout("flushing dirty caps\n");
1984 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
1985 dout("flushing snap caps\n");
1990 /* completed revocation? going down and there are no caps? */
1991 if (revoking && (revoking & cap_used) == 0) {
1992 dout("completed revocation of %s\n",
1993 ceph_cap_string(cap->implemented & ~cap->issued));
1997 /* want more caps from mds? */
1998 if (want & ~(cap->mds_wanted | cap->issued))
2001 /* things we might delay */
2002 if ((cap->issued & ~retain) == 0)
2003 continue; /* nope, all good */
2009 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
2010 time_before(jiffies, ci->i_hold_caps_max)) {
2011 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
2012 ceph_cap_string(cap->issued),
2013 ceph_cap_string(cap->issued & retain),
2014 ceph_cap_string(cap->mds_wanted),
2015 ceph_cap_string(want));
2021 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
2022 dout(" skipping %p I_NOFLUSH set\n", inode);
2026 if (session && session != cap->session) {
2027 dout("oops, wrong session %p mutex\n", session);
2028 mutex_unlock(&session->s_mutex);
2032 session = cap->session;
2033 if (mutex_trylock(&session->s_mutex) == 0) {
2034 dout("inverting session/ino locks on %p\n",
2036 spin_unlock(&ci->i_ceph_lock);
2037 if (took_snap_rwsem) {
2038 up_read(&mdsc->snap_rwsem);
2039 took_snap_rwsem = 0;
2041 mutex_lock(&session->s_mutex);
2046 /* kick flushing and flush snaps before sending normal
2048 if (cap == ci->i_auth_cap &&
2050 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2051 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2052 __kick_flushing_caps(mdsc, session, ci, 0);
2053 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2055 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2056 __ceph_flush_snaps(ci, session);
2061 /* take snap_rwsem after session mutex */
2062 if (!took_snap_rwsem) {
2063 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
2064 dout("inverting snap/in locks on %p\n",
2066 spin_unlock(&ci->i_ceph_lock);
2067 down_read(&mdsc->snap_rwsem);
2068 took_snap_rwsem = 1;
2071 took_snap_rwsem = 1;
2074 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2075 flushing = __mark_caps_flushing(inode, session, false,
2081 spin_lock(&mdsc->cap_dirty_lock);
2082 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2083 spin_unlock(&mdsc->cap_dirty_lock);
2086 mds = cap->mds; /* remember mds, so we don't repeat */
2089 /* __send_cap drops i_ceph_lock */
2090 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, false,
2091 cap_used, want, retain, flushing,
2092 flush_tid, oldest_flush_tid);
2093 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2096 /* Reschedule delayed caps release if we delayed anything */
2098 __cap_delay_requeue(mdsc, ci, false);
2100 spin_unlock(&ci->i_ceph_lock);
2102 if (queue_invalidate)
2103 ceph_queue_invalidate(inode);
2106 mutex_unlock(&session->s_mutex);
2107 if (took_snap_rwsem)
2108 up_read(&mdsc->snap_rwsem);
2112 * Try to flush dirty caps back to the auth mds.
2114 static int try_flush_caps(struct inode *inode, u64 *ptid)
2116 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2117 struct ceph_inode_info *ci = ceph_inode(inode);
2118 struct ceph_mds_session *session = NULL;
2120 u64 flush_tid = 0, oldest_flush_tid = 0;
2123 spin_lock(&ci->i_ceph_lock);
2124 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
2125 spin_unlock(&ci->i_ceph_lock);
2126 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
2129 if (ci->i_dirty_caps && ci->i_auth_cap) {
2130 struct ceph_cap *cap = ci->i_auth_cap;
2131 int used = __ceph_caps_used(ci);
2132 int want = __ceph_caps_wanted(ci);
2135 if (!session || session != cap->session) {
2136 spin_unlock(&ci->i_ceph_lock);
2138 mutex_unlock(&session->s_mutex);
2139 session = cap->session;
2140 mutex_lock(&session->s_mutex);
2143 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) {
2144 spin_unlock(&ci->i_ceph_lock);
2148 flushing = __mark_caps_flushing(inode, session, true,
2149 &flush_tid, &oldest_flush_tid);
2151 /* __send_cap drops i_ceph_lock */
2152 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, true,
2153 used, want, (cap->issued | cap->implemented),
2154 flushing, flush_tid, oldest_flush_tid);
2157 spin_lock(&ci->i_ceph_lock);
2158 __cap_delay_requeue(mdsc, ci, true);
2159 spin_unlock(&ci->i_ceph_lock);
2162 if (!list_empty(&ci->i_cap_flush_list)) {
2163 struct ceph_cap_flush *cf =
2164 list_last_entry(&ci->i_cap_flush_list,
2165 struct ceph_cap_flush, i_list);
2167 flush_tid = cf->tid;
2169 flushing = ci->i_flushing_caps;
2170 spin_unlock(&ci->i_ceph_lock);
2174 mutex_unlock(&session->s_mutex);
2181 * Return true if we've flushed caps through the given flush_tid.
2183 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2185 struct ceph_inode_info *ci = ceph_inode(inode);
2188 spin_lock(&ci->i_ceph_lock);
2189 if (!list_empty(&ci->i_cap_flush_list)) {
2190 struct ceph_cap_flush * cf =
2191 list_first_entry(&ci->i_cap_flush_list,
2192 struct ceph_cap_flush, i_list);
2193 if (cf->tid <= flush_tid)
2196 spin_unlock(&ci->i_ceph_lock);
2201 * wait for any unsafe requests to complete.
2203 static int unsafe_request_wait(struct inode *inode)
2205 struct ceph_inode_info *ci = ceph_inode(inode);
2206 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2209 spin_lock(&ci->i_unsafe_lock);
2210 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2211 req1 = list_last_entry(&ci->i_unsafe_dirops,
2212 struct ceph_mds_request,
2214 ceph_mdsc_get_request(req1);
2216 if (!list_empty(&ci->i_unsafe_iops)) {
2217 req2 = list_last_entry(&ci->i_unsafe_iops,
2218 struct ceph_mds_request,
2219 r_unsafe_target_item);
2220 ceph_mdsc_get_request(req2);
2222 spin_unlock(&ci->i_unsafe_lock);
2224 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2225 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2227 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2228 ceph_timeout_jiffies(req1->r_timeout));
2231 ceph_mdsc_put_request(req1);
2234 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2235 ceph_timeout_jiffies(req2->r_timeout));
2238 ceph_mdsc_put_request(req2);
2243 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2245 struct inode *inode = file->f_mapping->host;
2246 struct ceph_inode_info *ci = ceph_inode(inode);
2251 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2253 ret = file_write_and_wait_range(file, start, end);
2260 dirty = try_flush_caps(inode, &flush_tid);
2261 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2263 ret = unsafe_request_wait(inode);
2266 * only wait on non-file metadata writeback (the mds
2267 * can recover size and mtime, so we don't need to
2270 if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2271 ret = wait_event_interruptible(ci->i_cap_wq,
2272 caps_are_flushed(inode, flush_tid));
2275 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2280 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2281 * queue inode for flush but don't do so immediately, because we can
2282 * get by with fewer MDS messages if we wait for data writeback to
2285 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2287 struct ceph_inode_info *ci = ceph_inode(inode);
2291 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2293 dout("write_inode %p wait=%d\n", inode, wait);
2295 dirty = try_flush_caps(inode, &flush_tid);
2297 err = wait_event_interruptible(ci->i_cap_wq,
2298 caps_are_flushed(inode, flush_tid));
2300 struct ceph_mds_client *mdsc =
2301 ceph_sb_to_client(inode->i_sb)->mdsc;
2303 spin_lock(&ci->i_ceph_lock);
2304 if (__ceph_caps_dirty(ci))
2305 __cap_delay_requeue_front(mdsc, ci);
2306 spin_unlock(&ci->i_ceph_lock);
2311 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2312 struct ceph_mds_session *session,
2313 struct ceph_inode_info *ci,
2314 u64 oldest_flush_tid)
2315 __releases(ci->i_ceph_lock)
2316 __acquires(ci->i_ceph_lock)
2318 struct inode *inode = &ci->vfs_inode;
2319 struct ceph_cap *cap;
2320 struct ceph_cap_flush *cf;
2324 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2325 if (cf->tid < first_tid)
2328 cap = ci->i_auth_cap;
2329 if (!(cap && cap->session == session)) {
2330 pr_err("%p auth cap %p not mds%d ???\n",
2331 inode, cap, session->s_mds);
2335 first_tid = cf->tid + 1;
2338 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2339 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2340 ci->i_ceph_flags |= CEPH_I_NODELAY;
2341 ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2342 false, __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 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2414 /* encode_caps_cb() also will reset these sequence
2415 * numbers. make sure sequence numbers in cap flush
2416 * message match later reconnect message */
2420 __kick_flushing_caps(mdsc, session, ci,
2423 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2426 spin_unlock(&ci->i_ceph_lock);
2430 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2431 struct ceph_mds_session *session)
2433 struct ceph_inode_info *ci;
2434 struct ceph_cap *cap;
2435 u64 oldest_flush_tid;
2437 dout("kick_flushing_caps mds%d\n", session->s_mds);
2439 spin_lock(&mdsc->cap_dirty_lock);
2440 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2441 spin_unlock(&mdsc->cap_dirty_lock);
2443 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2444 spin_lock(&ci->i_ceph_lock);
2445 cap = ci->i_auth_cap;
2446 if (!(cap && cap->session == session)) {
2447 pr_err("%p auth cap %p not mds%d ???\n",
2448 &ci->vfs_inode, cap, session->s_mds);
2449 spin_unlock(&ci->i_ceph_lock);
2452 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2453 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2454 __kick_flushing_caps(mdsc, session, ci,
2457 spin_unlock(&ci->i_ceph_lock);
2461 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2462 struct ceph_mds_session *session,
2463 struct inode *inode)
2464 __releases(ci->i_ceph_lock)
2466 struct ceph_inode_info *ci = ceph_inode(inode);
2467 struct ceph_cap *cap;
2469 cap = ci->i_auth_cap;
2470 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2471 ceph_cap_string(ci->i_flushing_caps));
2473 if (!list_empty(&ci->i_cap_flush_list)) {
2474 u64 oldest_flush_tid;
2475 spin_lock(&mdsc->cap_dirty_lock);
2476 list_move_tail(&ci->i_flushing_item,
2477 &cap->session->s_cap_flushing);
2478 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2479 spin_unlock(&mdsc->cap_dirty_lock);
2481 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2482 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2483 spin_unlock(&ci->i_ceph_lock);
2485 spin_unlock(&ci->i_ceph_lock);
2491 * Take references to capabilities we hold, so that we don't release
2492 * them to the MDS prematurely.
2494 * Protected by i_ceph_lock.
2496 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2497 bool snap_rwsem_locked)
2499 if (got & CEPH_CAP_PIN)
2501 if (got & CEPH_CAP_FILE_RD)
2503 if (got & CEPH_CAP_FILE_CACHE)
2504 ci->i_rdcache_ref++;
2505 if (got & CEPH_CAP_FILE_WR) {
2506 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2507 BUG_ON(!snap_rwsem_locked);
2508 ci->i_head_snapc = ceph_get_snap_context(
2509 ci->i_snap_realm->cached_context);
2513 if (got & CEPH_CAP_FILE_BUFFER) {
2514 if (ci->i_wb_ref == 0)
2515 ihold(&ci->vfs_inode);
2517 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2518 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2523 * Try to grab cap references. Specify those refs we @want, and the
2524 * minimal set we @need. Also include the larger offset we are writing
2525 * to (when applicable), and check against max_size here as well.
2526 * Note that caller is responsible for ensuring max_size increases are
2527 * requested from the MDS.
2529 * Returns 0 if caps were not able to be acquired (yet), a 1 if they were,
2530 * or a negative error code.
2532 * FIXME: how does a 0 return differ from -EAGAIN?
2534 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2535 loff_t endoff, bool nonblock, int *got)
2537 struct inode *inode = &ci->vfs_inode;
2538 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2540 int have, implemented;
2542 bool snap_rwsem_locked = false;
2544 dout("get_cap_refs %p need %s want %s\n", inode,
2545 ceph_cap_string(need), ceph_cap_string(want));
2548 spin_lock(&ci->i_ceph_lock);
2550 /* make sure file is actually open */
2551 file_wanted = __ceph_caps_file_wanted(ci);
2552 if ((file_wanted & need) != need) {
2553 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2554 ceph_cap_string(need), ceph_cap_string(file_wanted));
2559 /* finish pending truncate */
2560 while (ci->i_truncate_pending) {
2561 spin_unlock(&ci->i_ceph_lock);
2562 if (snap_rwsem_locked) {
2563 up_read(&mdsc->snap_rwsem);
2564 snap_rwsem_locked = false;
2566 __ceph_do_pending_vmtruncate(inode);
2567 spin_lock(&ci->i_ceph_lock);
2570 have = __ceph_caps_issued(ci, &implemented);
2572 if (have & need & CEPH_CAP_FILE_WR) {
2573 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2574 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2575 inode, endoff, ci->i_max_size);
2576 if (endoff > ci->i_requested_max_size)
2581 * If a sync write is in progress, we must wait, so that we
2582 * can get a final snapshot value for size+mtime.
2584 if (__ceph_have_pending_cap_snap(ci)) {
2585 dout("get_cap_refs %p cap_snap_pending\n", inode);
2590 if ((have & need) == need) {
2592 * Look at (implemented & ~have & not) so that we keep waiting
2593 * on transition from wanted -> needed caps. This is needed
2594 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2595 * going before a prior buffered writeback happens.
2597 int not = want & ~(have & need);
2598 int revoking = implemented & ~have;
2599 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2600 inode, ceph_cap_string(have), ceph_cap_string(not),
2601 ceph_cap_string(revoking));
2602 if ((revoking & not) == 0) {
2603 if (!snap_rwsem_locked &&
2604 !ci->i_head_snapc &&
2605 (need & CEPH_CAP_FILE_WR)) {
2606 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2608 * we can not call down_read() when
2609 * task isn't in TASK_RUNNING state
2616 spin_unlock(&ci->i_ceph_lock);
2617 down_read(&mdsc->snap_rwsem);
2618 snap_rwsem_locked = true;
2621 snap_rwsem_locked = true;
2623 *got = need | (have & want);
2624 if ((need & CEPH_CAP_FILE_RD) &&
2625 !(*got & CEPH_CAP_FILE_CACHE))
2626 ceph_disable_fscache_readpage(ci);
2627 __take_cap_refs(ci, *got, true);
2631 int session_readonly = false;
2632 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2633 struct ceph_mds_session *s = ci->i_auth_cap->session;
2634 spin_lock(&s->s_cap_lock);
2635 session_readonly = s->s_readonly;
2636 spin_unlock(&s->s_cap_lock);
2638 if (session_readonly) {
2639 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2640 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2645 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) {
2647 if (READ_ONCE(mdsc->fsc->mount_state) ==
2648 CEPH_MOUNT_SHUTDOWN) {
2649 dout("get_cap_refs %p forced umount\n", inode);
2653 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2654 if (need & ~(mds_wanted & need)) {
2655 dout("get_cap_refs %p caps were dropped"
2656 " (session killed?)\n", inode);
2660 if (!(file_wanted & ~mds_wanted))
2661 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED;
2664 dout("get_cap_refs %p have %s needed %s\n", inode,
2665 ceph_cap_string(have), ceph_cap_string(need));
2668 spin_unlock(&ci->i_ceph_lock);
2669 if (snap_rwsem_locked)
2670 up_read(&mdsc->snap_rwsem);
2672 dout("get_cap_refs %p ret %d got %s\n", inode,
2673 ret, ceph_cap_string(*got));
2678 * Check the offset we are writing up to against our current
2679 * max_size. If necessary, tell the MDS we want to write to
2682 static void check_max_size(struct inode *inode, loff_t endoff)
2684 struct ceph_inode_info *ci = ceph_inode(inode);
2687 /* do we need to explicitly request a larger max_size? */
2688 spin_lock(&ci->i_ceph_lock);
2689 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2690 dout("write %p at large endoff %llu, req max_size\n",
2692 ci->i_wanted_max_size = endoff;
2694 /* duplicate ceph_check_caps()'s logic */
2695 if (ci->i_auth_cap &&
2696 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2697 ci->i_wanted_max_size > ci->i_max_size &&
2698 ci->i_wanted_max_size > ci->i_requested_max_size)
2700 spin_unlock(&ci->i_ceph_lock);
2702 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2705 int ceph_try_get_caps(struct ceph_inode_info *ci, int need, int want,
2706 bool nonblock, int *got)
2710 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2711 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO|CEPH_CAP_FILE_SHARED));
2712 ret = ceph_pool_perm_check(ci, need);
2716 ret = try_get_cap_refs(ci, need, want, 0, nonblock, got);
2717 return ret == -EAGAIN ? 0 : ret;
2721 * Wait for caps, and take cap references. If we can't get a WR cap
2722 * due to a small max_size, make sure we check_max_size (and possibly
2723 * ask the mds) so we don't get hung up indefinitely.
2725 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2726 loff_t endoff, int *got, struct page **pinned_page)
2730 ret = ceph_pool_perm_check(ci, need);
2736 check_max_size(&ci->vfs_inode, endoff);
2739 ret = try_get_cap_refs(ci, need, want, endoff,
2741 if (ret == -EAGAIN) {
2746 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2747 add_wait_queue(&ci->i_cap_wq, &wait);
2749 while (!(err = try_get_cap_refs(ci, need, want, endoff,
2751 if (signal_pending(current)) {
2755 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2758 remove_wait_queue(&ci->i_cap_wq, &wait);
2762 if (ret == -ESTALE) {
2763 /* session was killed, try renew caps */
2764 ret = ceph_renew_caps(&ci->vfs_inode);
2770 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2771 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2772 i_size_read(&ci->vfs_inode) > 0) {
2774 find_get_page(ci->vfs_inode.i_mapping, 0);
2776 if (PageUptodate(page)) {
2777 *pinned_page = page;
2783 * drop cap refs first because getattr while
2784 * holding * caps refs can cause deadlock.
2786 ceph_put_cap_refs(ci, _got);
2790 * getattr request will bring inline data into
2793 ret = __ceph_do_getattr(&ci->vfs_inode, NULL,
2794 CEPH_STAT_CAP_INLINE_DATA,
2803 if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
2804 ceph_fscache_revalidate_cookie(ci);
2811 * Take cap refs. Caller must already know we hold at least one ref
2812 * on the caps in question or we don't know this is safe.
2814 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2816 spin_lock(&ci->i_ceph_lock);
2817 __take_cap_refs(ci, caps, false);
2818 spin_unlock(&ci->i_ceph_lock);
2823 * drop cap_snap that is not associated with any snapshot.
2824 * we don't need to send FLUSHSNAP message for it.
2826 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2827 struct ceph_cap_snap *capsnap)
2829 if (!capsnap->need_flush &&
2830 !capsnap->writing && !capsnap->dirty_pages) {
2831 dout("dropping cap_snap %p follows %llu\n",
2832 capsnap, capsnap->follows);
2833 BUG_ON(capsnap->cap_flush.tid > 0);
2834 ceph_put_snap_context(capsnap->context);
2835 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2836 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2838 list_del(&capsnap->ci_item);
2839 ceph_put_cap_snap(capsnap);
2848 * If we released the last ref on any given cap, call ceph_check_caps
2849 * to release (or schedule a release).
2851 * If we are releasing a WR cap (from a sync write), finalize any affected
2852 * cap_snap, and wake up any waiters.
2854 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2856 struct inode *inode = &ci->vfs_inode;
2857 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2859 spin_lock(&ci->i_ceph_lock);
2860 if (had & CEPH_CAP_PIN)
2862 if (had & CEPH_CAP_FILE_RD)
2863 if (--ci->i_rd_ref == 0)
2865 if (had & CEPH_CAP_FILE_CACHE)
2866 if (--ci->i_rdcache_ref == 0)
2868 if (had & CEPH_CAP_FILE_BUFFER) {
2869 if (--ci->i_wb_ref == 0) {
2873 dout("put_cap_refs %p wb %d -> %d (?)\n",
2874 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2876 if (had & CEPH_CAP_FILE_WR)
2877 if (--ci->i_wr_ref == 0) {
2879 if (__ceph_have_pending_cap_snap(ci)) {
2880 struct ceph_cap_snap *capsnap =
2881 list_last_entry(&ci->i_cap_snaps,
2882 struct ceph_cap_snap,
2884 capsnap->writing = 0;
2885 if (ceph_try_drop_cap_snap(ci, capsnap))
2887 else if (__ceph_finish_cap_snap(ci, capsnap))
2891 if (ci->i_wrbuffer_ref_head == 0 &&
2892 ci->i_dirty_caps == 0 &&
2893 ci->i_flushing_caps == 0) {
2894 BUG_ON(!ci->i_head_snapc);
2895 ceph_put_snap_context(ci->i_head_snapc);
2896 ci->i_head_snapc = NULL;
2898 /* see comment in __ceph_remove_cap() */
2899 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2900 drop_inode_snap_realm(ci);
2902 spin_unlock(&ci->i_ceph_lock);
2904 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2905 last ? " last" : "", put ? " put" : "");
2907 if (last && !flushsnaps)
2908 ceph_check_caps(ci, 0, NULL);
2909 else if (flushsnaps)
2910 ceph_flush_snaps(ci, NULL);
2912 wake_up_all(&ci->i_cap_wq);
2918 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2919 * context. Adjust per-snap dirty page accounting as appropriate.
2920 * Once all dirty data for a cap_snap is flushed, flush snapped file
2921 * metadata back to the MDS. If we dropped the last ref, call
2924 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2925 struct ceph_snap_context *snapc)
2927 struct inode *inode = &ci->vfs_inode;
2928 struct ceph_cap_snap *capsnap = NULL;
2932 bool flush_snaps = false;
2933 bool complete_capsnap = false;
2935 spin_lock(&ci->i_ceph_lock);
2936 ci->i_wrbuffer_ref -= nr;
2937 if (ci->i_wrbuffer_ref == 0) {
2942 if (ci->i_head_snapc == snapc) {
2943 ci->i_wrbuffer_ref_head -= nr;
2944 if (ci->i_wrbuffer_ref_head == 0 &&
2945 ci->i_wr_ref == 0 &&
2946 ci->i_dirty_caps == 0 &&
2947 ci->i_flushing_caps == 0) {
2948 BUG_ON(!ci->i_head_snapc);
2949 ceph_put_snap_context(ci->i_head_snapc);
2950 ci->i_head_snapc = NULL;
2952 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2954 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2955 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2956 last ? " LAST" : "");
2958 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2959 if (capsnap->context == snapc) {
2965 capsnap->dirty_pages -= nr;
2966 if (capsnap->dirty_pages == 0) {
2967 complete_capsnap = true;
2968 if (!capsnap->writing) {
2969 if (ceph_try_drop_cap_snap(ci, capsnap)) {
2972 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2977 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2978 " snap %lld %d/%d -> %d/%d %s%s\n",
2979 inode, capsnap, capsnap->context->seq,
2980 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2981 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2982 last ? " (wrbuffer last)" : "",
2983 complete_capsnap ? " (complete capsnap)" : "");
2986 spin_unlock(&ci->i_ceph_lock);
2989 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2990 } else if (flush_snaps) {
2991 ceph_flush_snaps(ci, NULL);
2993 if (complete_capsnap)
2994 wake_up_all(&ci->i_cap_wq);
2996 /* avoid calling iput_final() in osd dispatch threads */
2997 ceph_async_iput(inode);
3002 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3004 static void invalidate_aliases(struct inode *inode)
3006 struct dentry *dn, *prev = NULL;
3008 dout("invalidate_aliases inode %p\n", inode);
3009 d_prune_aliases(inode);
3011 * For non-directory inode, d_find_alias() only returns
3012 * hashed dentry. After calling d_invalidate(), the
3013 * dentry becomes unhashed.
3015 * For directory inode, d_find_alias() can return
3016 * unhashed dentry. But directory inode should have
3017 * one alias at most.
3019 while ((dn = d_find_alias(inode))) {
3033 struct cap_extra_info {
3034 struct ceph_string *pool_ns;
3043 /* currently issued */
3048 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3049 * actually be a revocation if it specifies a smaller cap set.)
3051 * caller holds s_mutex and i_ceph_lock, we drop both.
3053 static void handle_cap_grant(struct inode *inode,
3054 struct ceph_mds_session *session,
3055 struct ceph_cap *cap,
3056 struct ceph_mds_caps *grant,
3057 struct ceph_buffer *xattr_buf,
3058 struct cap_extra_info *extra_info)
3059 __releases(ci->i_ceph_lock)
3060 __releases(session->s_mdsc->snap_rwsem)
3062 struct ceph_inode_info *ci = ceph_inode(inode);
3063 int seq = le32_to_cpu(grant->seq);
3064 int newcaps = le32_to_cpu(grant->caps);
3065 int used, wanted, dirty;
3066 u64 size = le64_to_cpu(grant->size);
3067 u64 max_size = le64_to_cpu(grant->max_size);
3068 unsigned char check_caps = 0;
3069 bool was_stale = cap->cap_gen < session->s_cap_gen;
3071 bool writeback = false;
3072 bool queue_trunc = false;
3073 bool queue_invalidate = false;
3074 bool deleted_inode = false;
3075 bool fill_inline = false;
3077 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3078 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3079 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3084 * If CACHE is being revoked, and we have no dirty buffers,
3085 * try to invalidate (once). (If there are dirty buffers, we
3086 * will invalidate _after_ writeback.)
3088 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
3089 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3090 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3091 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3092 if (try_nonblocking_invalidate(inode)) {
3093 /* there were locked pages.. invalidate later
3094 in a separate thread. */
3095 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3096 queue_invalidate = true;
3097 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3103 cap->issued = cap->implemented = CEPH_CAP_PIN;
3106 * auth mds of the inode changed. we received the cap export message,
3107 * but still haven't received the cap import message. handle_cap_export
3108 * updated the new auth MDS' cap.
3110 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3111 * that was sent before the cap import message. So don't remove caps.
3113 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3114 WARN_ON(cap != ci->i_auth_cap);
3115 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3117 newcaps |= cap->issued;
3120 /* side effects now are allowed */
3121 cap->cap_gen = session->s_cap_gen;
3124 __check_cap_issue(ci, cap, newcaps);
3126 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3127 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3128 inode->i_mode = le32_to_cpu(grant->mode);
3129 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3130 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3131 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3132 from_kuid(&init_user_ns, inode->i_uid),
3133 from_kgid(&init_user_ns, inode->i_gid));
3136 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3137 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3138 set_nlink(inode, le32_to_cpu(grant->nlink));
3139 if (inode->i_nlink == 0 &&
3140 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3141 deleted_inode = true;
3144 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3146 int len = le32_to_cpu(grant->xattr_len);
3147 u64 version = le64_to_cpu(grant->xattr_version);
3149 if (version > ci->i_xattrs.version) {
3150 dout(" got new xattrs v%llu on %p len %d\n",
3151 version, inode, len);
3152 if (ci->i_xattrs.blob)
3153 ceph_buffer_put(ci->i_xattrs.blob);
3154 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3155 ci->i_xattrs.version = version;
3156 ceph_forget_all_cached_acls(inode);
3160 if (newcaps & CEPH_CAP_ANY_RD) {
3161 struct timespec64 mtime, atime, ctime;
3162 /* ctime/mtime/atime? */
3163 ceph_decode_timespec64(&mtime, &grant->mtime);
3164 ceph_decode_timespec64(&atime, &grant->atime);
3165 ceph_decode_timespec64(&ctime, &grant->ctime);
3166 ceph_fill_file_time(inode, extra_info->issued,
3167 le32_to_cpu(grant->time_warp_seq),
3168 &ctime, &mtime, &atime);
3171 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3172 ci->i_files = extra_info->nfiles;
3173 ci->i_subdirs = extra_info->nsubdirs;
3176 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3177 /* file layout may have changed */
3178 s64 old_pool = ci->i_layout.pool_id;
3179 struct ceph_string *old_ns;
3181 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3182 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3183 lockdep_is_held(&ci->i_ceph_lock));
3184 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3186 if (ci->i_layout.pool_id != old_pool ||
3187 extra_info->pool_ns != old_ns)
3188 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3190 extra_info->pool_ns = old_ns;
3192 /* size/truncate_seq? */
3193 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3194 le32_to_cpu(grant->truncate_seq),
3195 le64_to_cpu(grant->truncate_size),
3199 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3200 if (max_size != ci->i_max_size) {
3201 dout("max_size %lld -> %llu\n",
3202 ci->i_max_size, max_size);
3203 ci->i_max_size = max_size;
3204 if (max_size >= ci->i_wanted_max_size) {
3205 ci->i_wanted_max_size = 0; /* reset */
3206 ci->i_requested_max_size = 0;
3209 } else if (ci->i_wanted_max_size > ci->i_max_size &&
3210 ci->i_wanted_max_size > ci->i_requested_max_size) {
3211 /* CEPH_CAP_OP_IMPORT */
3216 /* check cap bits */
3217 wanted = __ceph_caps_wanted(ci);
3218 used = __ceph_caps_used(ci);
3219 dirty = __ceph_caps_dirty(ci);
3220 dout(" my wanted = %s, used = %s, dirty %s\n",
3221 ceph_cap_string(wanted),
3222 ceph_cap_string(used),
3223 ceph_cap_string(dirty));
3225 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3226 (wanted & ~(cap->mds_wanted | newcaps))) {
3228 * If mds is importing cap, prior cap messages that update
3229 * 'wanted' may get dropped by mds (migrate seq mismatch).
3231 * We don't send cap message to update 'wanted' if what we
3232 * want are already issued. If mds revokes caps, cap message
3233 * that releases caps also tells mds what we want. But if
3234 * caps got revoked by mds forcedly (session stale). We may
3235 * haven't told mds what we want.
3240 /* revocation, grant, or no-op? */
3241 if (cap->issued & ~newcaps) {
3242 int revoking = cap->issued & ~newcaps;
3244 dout("revocation: %s -> %s (revoking %s)\n",
3245 ceph_cap_string(cap->issued),
3246 ceph_cap_string(newcaps),
3247 ceph_cap_string(revoking));
3248 if (revoking & used & CEPH_CAP_FILE_BUFFER)
3249 writeback = true; /* initiate writeback; will delay ack */
3250 else if (revoking == CEPH_CAP_FILE_CACHE &&
3251 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3253 ; /* do nothing yet, invalidation will be queued */
3254 else if (cap == ci->i_auth_cap)
3255 check_caps = 1; /* check auth cap only */
3257 check_caps = 2; /* check all caps */
3258 cap->issued = newcaps;
3259 cap->implemented |= newcaps;
3260 } else if (cap->issued == newcaps) {
3261 dout("caps unchanged: %s -> %s\n",
3262 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3264 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3265 ceph_cap_string(newcaps));
3266 /* non-auth MDS is revoking the newly grant caps ? */
3267 if (cap == ci->i_auth_cap &&
3268 __ceph_caps_revoking_other(ci, cap, newcaps))
3271 cap->issued = newcaps;
3272 cap->implemented |= newcaps; /* add bits only, to
3273 * avoid stepping on a
3274 * pending revocation */
3277 BUG_ON(cap->issued & ~cap->implemented);
3279 if (extra_info->inline_version > 0 &&
3280 extra_info->inline_version >= ci->i_inline_version) {
3281 ci->i_inline_version = extra_info->inline_version;
3282 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3283 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3287 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3288 if (newcaps & ~extra_info->issued)
3290 kick_flushing_inode_caps(session->s_mdsc, session, inode);
3291 up_read(&session->s_mdsc->snap_rwsem);
3293 spin_unlock(&ci->i_ceph_lock);
3297 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3298 extra_info->inline_len);
3301 ceph_queue_vmtruncate(inode);
3305 * queue inode for writeback: we can't actually call
3306 * filemap_write_and_wait, etc. from message handler
3309 ceph_queue_writeback(inode);
3310 if (queue_invalidate)
3311 ceph_queue_invalidate(inode);
3313 invalidate_aliases(inode);
3315 wake_up_all(&ci->i_cap_wq);
3317 if (check_caps == 1)
3318 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
3320 else if (check_caps == 2)
3321 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
3323 mutex_unlock(&session->s_mutex);
3327 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3328 * MDS has been safely committed.
3330 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3331 struct ceph_mds_caps *m,
3332 struct ceph_mds_session *session,
3333 struct ceph_cap *cap)
3334 __releases(ci->i_ceph_lock)
3336 struct ceph_inode_info *ci = ceph_inode(inode);
3337 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3338 struct ceph_cap_flush *cf, *tmp_cf;
3339 LIST_HEAD(to_remove);
3340 unsigned seq = le32_to_cpu(m->seq);
3341 int dirty = le32_to_cpu(m->dirty);
3344 bool wake_ci = false;
3345 bool wake_mdsc = false;
3347 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3348 if (cf->tid == flush_tid)
3350 if (cf->caps == 0) /* capsnap */
3352 if (cf->tid <= flush_tid) {
3353 if (__finish_cap_flush(NULL, ci, cf))
3355 list_add_tail(&cf->i_list, &to_remove);
3357 cleaned &= ~cf->caps;
3363 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3364 " flushing %s -> %s\n",
3365 inode, session->s_mds, seq, ceph_cap_string(dirty),
3366 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3367 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3369 if (list_empty(&to_remove) && !cleaned)
3372 ci->i_flushing_caps &= ~cleaned;
3374 spin_lock(&mdsc->cap_dirty_lock);
3376 list_for_each_entry(cf, &to_remove, i_list) {
3377 if (__finish_cap_flush(mdsc, NULL, cf))
3381 if (ci->i_flushing_caps == 0) {
3382 if (list_empty(&ci->i_cap_flush_list)) {
3383 list_del_init(&ci->i_flushing_item);
3384 if (!list_empty(&session->s_cap_flushing)) {
3385 dout(" mds%d still flushing cap on %p\n",
3387 &list_first_entry(&session->s_cap_flushing,
3388 struct ceph_inode_info,
3389 i_flushing_item)->vfs_inode);
3392 mdsc->num_cap_flushing--;
3393 dout(" inode %p now !flushing\n", inode);
3395 if (ci->i_dirty_caps == 0) {
3396 dout(" inode %p now clean\n", inode);
3397 BUG_ON(!list_empty(&ci->i_dirty_item));
3399 if (ci->i_wr_ref == 0 &&
3400 ci->i_wrbuffer_ref_head == 0) {
3401 BUG_ON(!ci->i_head_snapc);
3402 ceph_put_snap_context(ci->i_head_snapc);
3403 ci->i_head_snapc = NULL;
3406 BUG_ON(list_empty(&ci->i_dirty_item));
3409 spin_unlock(&mdsc->cap_dirty_lock);
3412 spin_unlock(&ci->i_ceph_lock);
3414 while (!list_empty(&to_remove)) {
3415 cf = list_first_entry(&to_remove,
3416 struct ceph_cap_flush, i_list);
3417 list_del(&cf->i_list);
3418 ceph_free_cap_flush(cf);
3422 wake_up_all(&ci->i_cap_wq);
3424 wake_up_all(&mdsc->cap_flushing_wq);
3430 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3431 * throw away our cap_snap.
3433 * Caller hold s_mutex.
3435 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3436 struct ceph_mds_caps *m,
3437 struct ceph_mds_session *session)
3439 struct ceph_inode_info *ci = ceph_inode(inode);
3440 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3441 u64 follows = le64_to_cpu(m->snap_follows);
3442 struct ceph_cap_snap *capsnap;
3443 bool flushed = false;
3444 bool wake_ci = false;
3445 bool wake_mdsc = false;
3447 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3448 inode, ci, session->s_mds, follows);
3450 spin_lock(&ci->i_ceph_lock);
3451 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3452 if (capsnap->follows == follows) {
3453 if (capsnap->cap_flush.tid != flush_tid) {
3454 dout(" cap_snap %p follows %lld tid %lld !="
3455 " %lld\n", capsnap, follows,
3456 flush_tid, capsnap->cap_flush.tid);
3462 dout(" skipping cap_snap %p follows %lld\n",
3463 capsnap, capsnap->follows);
3467 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3468 dout(" removing %p cap_snap %p follows %lld\n",
3469 inode, capsnap, follows);
3470 list_del(&capsnap->ci_item);
3471 if (__finish_cap_flush(NULL, ci, &capsnap->cap_flush))
3474 spin_lock(&mdsc->cap_dirty_lock);
3476 if (list_empty(&ci->i_cap_flush_list))
3477 list_del_init(&ci->i_flushing_item);
3479 if (__finish_cap_flush(mdsc, NULL, &capsnap->cap_flush))
3482 spin_unlock(&mdsc->cap_dirty_lock);
3484 spin_unlock(&ci->i_ceph_lock);
3486 ceph_put_snap_context(capsnap->context);
3487 ceph_put_cap_snap(capsnap);
3489 wake_up_all(&ci->i_cap_wq);
3491 wake_up_all(&mdsc->cap_flushing_wq);
3497 * Handle TRUNC from MDS, indicating file truncation.
3499 * caller hold s_mutex.
3501 static void handle_cap_trunc(struct inode *inode,
3502 struct ceph_mds_caps *trunc,
3503 struct ceph_mds_session *session)
3504 __releases(ci->i_ceph_lock)
3506 struct ceph_inode_info *ci = ceph_inode(inode);
3507 int mds = session->s_mds;
3508 int seq = le32_to_cpu(trunc->seq);
3509 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3510 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3511 u64 size = le64_to_cpu(trunc->size);
3512 int implemented = 0;
3513 int dirty = __ceph_caps_dirty(ci);
3514 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3515 int queue_trunc = 0;
3517 issued |= implemented | dirty;
3519 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3520 inode, mds, seq, truncate_size, truncate_seq);
3521 queue_trunc = ceph_fill_file_size(inode, issued,
3522 truncate_seq, truncate_size, size);
3523 spin_unlock(&ci->i_ceph_lock);
3526 ceph_queue_vmtruncate(inode);
3530 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3531 * different one. If we are the most recent migration we've seen (as
3532 * indicated by mseq), make note of the migrating cap bits for the
3533 * duration (until we see the corresponding IMPORT).
3535 * caller holds s_mutex
3537 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3538 struct ceph_mds_cap_peer *ph,
3539 struct ceph_mds_session *session)
3541 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3542 struct ceph_mds_session *tsession = NULL;
3543 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3544 struct ceph_inode_info *ci = ceph_inode(inode);
3546 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3547 unsigned t_seq, t_mseq;
3549 int mds = session->s_mds;
3552 t_cap_id = le64_to_cpu(ph->cap_id);
3553 t_seq = le32_to_cpu(ph->seq);
3554 t_mseq = le32_to_cpu(ph->mseq);
3555 target = le32_to_cpu(ph->mds);
3557 t_cap_id = t_seq = t_mseq = 0;
3561 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3562 inode, ci, mds, mseq, target);
3564 spin_lock(&ci->i_ceph_lock);
3565 cap = __get_cap_for_mds(ci, mds);
3566 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3570 if (cap->mds_wanted | cap->issued)
3571 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
3572 __ceph_remove_cap(cap, false);
3577 * now we know we haven't received the cap import message yet
3578 * because the exported cap still exist.
3581 issued = cap->issued;
3582 if (issued != cap->implemented)
3583 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3584 "ino (%llx.%llx) mds%d seq %d mseq %d "
3585 "issued %s implemented %s\n",
3586 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3587 ceph_cap_string(issued),
3588 ceph_cap_string(cap->implemented));
3591 tcap = __get_cap_for_mds(ci, target);
3593 /* already have caps from the target */
3594 if (tcap->cap_id == t_cap_id &&
3595 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3596 dout(" updating import cap %p mds%d\n", tcap, target);
3597 tcap->cap_id = t_cap_id;
3598 tcap->seq = t_seq - 1;
3599 tcap->issue_seq = t_seq - 1;
3600 tcap->issued |= issued;
3601 tcap->implemented |= issued;
3602 if (cap == ci->i_auth_cap)
3603 ci->i_auth_cap = tcap;
3605 if (!list_empty(&ci->i_cap_flush_list) &&
3606 ci->i_auth_cap == tcap) {
3607 spin_lock(&mdsc->cap_dirty_lock);
3608 list_move_tail(&ci->i_flushing_item,
3609 &tcap->session->s_cap_flushing);
3610 spin_unlock(&mdsc->cap_dirty_lock);
3613 __ceph_remove_cap(cap, false);
3615 } else if (tsession) {
3616 /* add placeholder for the export tagert */
3617 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3619 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3620 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3622 if (!list_empty(&ci->i_cap_flush_list) &&
3623 ci->i_auth_cap == tcap) {
3624 spin_lock(&mdsc->cap_dirty_lock);
3625 list_move_tail(&ci->i_flushing_item,
3626 &tcap->session->s_cap_flushing);
3627 spin_unlock(&mdsc->cap_dirty_lock);
3630 __ceph_remove_cap(cap, false);
3634 spin_unlock(&ci->i_ceph_lock);
3635 mutex_unlock(&session->s_mutex);
3637 /* open target session */
3638 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3639 if (!IS_ERR(tsession)) {
3641 mutex_lock(&session->s_mutex);
3642 mutex_lock_nested(&tsession->s_mutex,
3643 SINGLE_DEPTH_NESTING);
3645 mutex_lock(&tsession->s_mutex);
3646 mutex_lock_nested(&session->s_mutex,
3647 SINGLE_DEPTH_NESTING);
3649 new_cap = ceph_get_cap(mdsc, NULL);
3658 spin_unlock(&ci->i_ceph_lock);
3659 mutex_unlock(&session->s_mutex);
3661 mutex_unlock(&tsession->s_mutex);
3662 ceph_put_mds_session(tsession);
3665 ceph_put_cap(mdsc, new_cap);
3669 * Handle cap IMPORT.
3671 * caller holds s_mutex. acquires i_ceph_lock
3673 static void handle_cap_import(struct ceph_mds_client *mdsc,
3674 struct inode *inode, struct ceph_mds_caps *im,
3675 struct ceph_mds_cap_peer *ph,
3676 struct ceph_mds_session *session,
3677 struct ceph_cap **target_cap, int *old_issued)
3678 __acquires(ci->i_ceph_lock)
3680 struct ceph_inode_info *ci = ceph_inode(inode);
3681 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3682 int mds = session->s_mds;
3684 unsigned caps = le32_to_cpu(im->caps);
3685 unsigned wanted = le32_to_cpu(im->wanted);
3686 unsigned seq = le32_to_cpu(im->seq);
3687 unsigned mseq = le32_to_cpu(im->migrate_seq);
3688 u64 realmino = le64_to_cpu(im->realm);
3689 u64 cap_id = le64_to_cpu(im->cap_id);
3694 p_cap_id = le64_to_cpu(ph->cap_id);
3695 peer = le32_to_cpu(ph->mds);
3701 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3702 inode, ci, mds, mseq, peer);
3705 spin_lock(&ci->i_ceph_lock);
3706 cap = __get_cap_for_mds(ci, mds);
3709 spin_unlock(&ci->i_ceph_lock);
3710 new_cap = ceph_get_cap(mdsc, NULL);
3716 ceph_put_cap(mdsc, new_cap);
3721 __ceph_caps_issued(ci, &issued);
3722 issued |= __ceph_caps_dirty(ci);
3724 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3725 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3727 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3728 if (ocap && ocap->cap_id == p_cap_id) {
3729 dout(" remove export cap %p mds%d flags %d\n",
3730 ocap, peer, ph->flags);
3731 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3732 (ocap->seq != le32_to_cpu(ph->seq) ||
3733 ocap->mseq != le32_to_cpu(ph->mseq))) {
3734 pr_err_ratelimited("handle_cap_import: "
3735 "mismatched seq/mseq: ino (%llx.%llx) "
3736 "mds%d seq %d mseq %d importer mds%d "
3737 "has peer seq %d mseq %d\n",
3738 ceph_vinop(inode), peer, ocap->seq,
3739 ocap->mseq, mds, le32_to_cpu(ph->seq),
3740 le32_to_cpu(ph->mseq));
3742 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3745 /* make sure we re-request max_size, if necessary */
3746 ci->i_requested_max_size = 0;
3748 *old_issued = issued;
3753 * Handle a caps message from the MDS.
3755 * Identify the appropriate session, inode, and call the right handler
3756 * based on the cap op.
3758 void ceph_handle_caps(struct ceph_mds_session *session,
3759 struct ceph_msg *msg)
3761 struct ceph_mds_client *mdsc = session->s_mdsc;
3762 struct inode *inode;
3763 struct ceph_inode_info *ci;
3764 struct ceph_cap *cap;
3765 struct ceph_mds_caps *h;
3766 struct ceph_mds_cap_peer *peer = NULL;
3767 struct ceph_snap_realm *realm = NULL;
3769 int msg_version = le16_to_cpu(msg->hdr.version);
3771 struct ceph_vino vino;
3773 size_t snaptrace_len;
3775 struct cap_extra_info extra_info = {};
3777 dout("handle_caps from mds%d\n", session->s_mds);
3780 end = msg->front.iov_base + msg->front.iov_len;
3781 if (msg->front.iov_len < sizeof(*h))
3783 h = msg->front.iov_base;
3784 op = le32_to_cpu(h->op);
3785 vino.ino = le64_to_cpu(h->ino);
3786 vino.snap = CEPH_NOSNAP;
3787 seq = le32_to_cpu(h->seq);
3788 mseq = le32_to_cpu(h->migrate_seq);
3791 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3792 p = snaptrace + snaptrace_len;
3794 if (msg_version >= 2) {
3796 ceph_decode_32_safe(&p, end, flock_len, bad);
3797 if (p + flock_len > end)
3802 if (msg_version >= 3) {
3803 if (op == CEPH_CAP_OP_IMPORT) {
3804 if (p + sizeof(*peer) > end)
3808 } else if (op == CEPH_CAP_OP_EXPORT) {
3809 /* recorded in unused fields */
3810 peer = (void *)&h->size;
3814 if (msg_version >= 4) {
3815 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
3816 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
3817 if (p + extra_info.inline_len > end)
3819 extra_info.inline_data = p;
3820 p += extra_info.inline_len;
3823 if (msg_version >= 5) {
3824 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
3827 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
3828 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
3831 if (msg_version >= 8) {
3833 u32 caller_uid, caller_gid;
3837 ceph_decode_64_safe(&p, end, flush_tid, bad);
3839 ceph_decode_32_safe(&p, end, caller_uid, bad);
3840 ceph_decode_32_safe(&p, end, caller_gid, bad);
3842 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
3843 if (pool_ns_len > 0) {
3844 ceph_decode_need(&p, end, pool_ns_len, bad);
3845 extra_info.pool_ns =
3846 ceph_find_or_create_string(p, pool_ns_len);
3851 if (msg_version >= 11) {
3852 struct ceph_timespec *btime;
3857 if (p + sizeof(*btime) > end)
3860 p += sizeof(*btime);
3861 ceph_decode_64_safe(&p, end, change_attr, bad);
3863 ceph_decode_32_safe(&p, end, flags, bad);
3865 extra_info.dirstat_valid = true;
3866 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
3867 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
3871 inode = ceph_find_inode(mdsc->fsc->sb, vino);
3872 ci = ceph_inode(inode);
3873 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3876 mutex_lock(&session->s_mutex);
3878 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3882 dout(" i don't have ino %llx\n", vino.ino);
3884 if (op == CEPH_CAP_OP_IMPORT) {
3885 cap = ceph_get_cap(mdsc, NULL);
3886 cap->cap_ino = vino.ino;
3887 cap->queue_release = 1;
3888 cap->cap_id = le64_to_cpu(h->cap_id);
3891 cap->issue_seq = seq;
3892 spin_lock(&session->s_cap_lock);
3893 __ceph_queue_cap_release(session, cap);
3894 spin_unlock(&session->s_cap_lock);
3899 /* these will work even if we don't have a cap yet */
3901 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3902 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
3906 case CEPH_CAP_OP_EXPORT:
3907 handle_cap_export(inode, h, peer, session);
3910 case CEPH_CAP_OP_IMPORT:
3912 if (snaptrace_len) {
3913 down_write(&mdsc->snap_rwsem);
3914 ceph_update_snap_trace(mdsc, snaptrace,
3915 snaptrace + snaptrace_len,
3917 downgrade_write(&mdsc->snap_rwsem);
3919 down_read(&mdsc->snap_rwsem);
3921 handle_cap_import(mdsc, inode, h, peer, session,
3922 &cap, &extra_info.issued);
3923 handle_cap_grant(inode, session, cap,
3924 h, msg->middle, &extra_info);
3926 ceph_put_snap_realm(mdsc, realm);
3930 /* the rest require a cap */
3931 spin_lock(&ci->i_ceph_lock);
3932 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
3934 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3935 inode, ceph_ino(inode), ceph_snap(inode),
3937 spin_unlock(&ci->i_ceph_lock);
3938 goto flush_cap_releases;
3941 /* note that each of these drops i_ceph_lock for us */
3943 case CEPH_CAP_OP_REVOKE:
3944 case CEPH_CAP_OP_GRANT:
3945 __ceph_caps_issued(ci, &extra_info.issued);
3946 extra_info.issued |= __ceph_caps_dirty(ci);
3947 handle_cap_grant(inode, session, cap,
3948 h, msg->middle, &extra_info);
3951 case CEPH_CAP_OP_FLUSH_ACK:
3952 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
3956 case CEPH_CAP_OP_TRUNC:
3957 handle_cap_trunc(inode, h, session);
3961 spin_unlock(&ci->i_ceph_lock);
3962 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3963 ceph_cap_op_name(op));
3967 mutex_unlock(&session->s_mutex);
3969 ceph_put_string(extra_info.pool_ns);
3970 /* avoid calling iput_final() in mds dispatch threads */
3971 ceph_async_iput(inode);
3976 * send any cap release message to try to move things
3977 * along for the mds (who clearly thinks we still have this
3980 ceph_flush_cap_releases(mdsc, session);
3984 pr_err("ceph_handle_caps: corrupt message\n");
3990 * Delayed work handler to process end of delayed cap release LRU list.
3992 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3994 struct inode *inode;
3995 struct ceph_inode_info *ci;
3996 int flags = CHECK_CAPS_NODELAY;
3998 dout("check_delayed_caps\n");
4000 spin_lock(&mdsc->cap_delay_lock);
4001 if (list_empty(&mdsc->cap_delay_list))
4003 ci = list_first_entry(&mdsc->cap_delay_list,
4004 struct ceph_inode_info,
4006 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
4007 time_before(jiffies, ci->i_hold_caps_max))
4009 list_del_init(&ci->i_cap_delay_list);
4011 inode = igrab(&ci->vfs_inode);
4012 spin_unlock(&mdsc->cap_delay_lock);
4015 dout("check_delayed_caps on %p\n", inode);
4016 ceph_check_caps(ci, flags, NULL);
4017 /* avoid calling iput_final() in tick thread */
4018 ceph_async_iput(inode);
4021 spin_unlock(&mdsc->cap_delay_lock);
4025 * Flush all dirty caps to the mds
4027 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4029 struct ceph_inode_info *ci;
4030 struct inode *inode;
4032 dout("flush_dirty_caps\n");
4033 spin_lock(&mdsc->cap_dirty_lock);
4034 while (!list_empty(&mdsc->cap_dirty)) {
4035 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
4037 inode = &ci->vfs_inode;
4039 dout("flush_dirty_caps %p\n", inode);
4040 spin_unlock(&mdsc->cap_dirty_lock);
4041 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
4043 spin_lock(&mdsc->cap_dirty_lock);
4045 spin_unlock(&mdsc->cap_dirty_lock);
4046 dout("flush_dirty_caps done\n");
4049 void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode)
4052 int bits = (fmode << 1) | 1;
4053 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4054 if (bits & (1 << i))
4055 ci->i_nr_by_mode[i]++;
4060 * Drop open file reference. If we were the last open file,
4061 * we may need to release capabilities to the MDS (or schedule
4062 * their delayed release).
4064 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
4067 int bits = (fmode << 1) | 1;
4068 spin_lock(&ci->i_ceph_lock);
4069 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4070 if (bits & (1 << i)) {
4071 BUG_ON(ci->i_nr_by_mode[i] == 0);
4072 if (--ci->i_nr_by_mode[i] == 0)
4076 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
4077 &ci->vfs_inode, fmode,
4078 ci->i_nr_by_mode[0], ci->i_nr_by_mode[1],
4079 ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]);
4080 spin_unlock(&ci->i_ceph_lock);
4082 if (last && ci->i_vino.snap == CEPH_NOSNAP)
4083 ceph_check_caps(ci, 0, NULL);
4087 * For a soon-to-be unlinked file, drop the LINK caps. If it
4088 * looks like the link count will hit 0, drop any other caps (other
4089 * than PIN) we don't specifically want (due to the file still being
4092 int ceph_drop_caps_for_unlink(struct inode *inode)
4094 struct ceph_inode_info *ci = ceph_inode(inode);
4095 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4097 spin_lock(&ci->i_ceph_lock);
4098 if (inode->i_nlink == 1) {
4099 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4101 ci->i_ceph_flags |= CEPH_I_NODELAY;
4102 if (__ceph_caps_dirty(ci)) {
4103 struct ceph_mds_client *mdsc =
4104 ceph_inode_to_client(inode)->mdsc;
4105 __cap_delay_requeue_front(mdsc, ci);
4108 spin_unlock(&ci->i_ceph_lock);
4113 * Helpers for embedding cap and dentry lease releases into mds
4116 * @force is used by dentry_release (below) to force inclusion of a
4117 * record for the directory inode, even when there aren't any caps to
4120 int ceph_encode_inode_release(void **p, struct inode *inode,
4121 int mds, int drop, int unless, int force)
4123 struct ceph_inode_info *ci = ceph_inode(inode);
4124 struct ceph_cap *cap;
4125 struct ceph_mds_request_release *rel = *p;
4129 spin_lock(&ci->i_ceph_lock);
4130 used = __ceph_caps_used(ci);
4131 dirty = __ceph_caps_dirty(ci);
4133 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4134 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4135 ceph_cap_string(unless));
4137 /* only drop unused, clean caps */
4138 drop &= ~(used | dirty);
4140 cap = __get_cap_for_mds(ci, mds);
4141 if (cap && __cap_is_valid(cap)) {
4142 unless &= cap->issued;
4144 if (unless & CEPH_CAP_AUTH_EXCL)
4145 drop &= ~CEPH_CAP_AUTH_SHARED;
4146 if (unless & CEPH_CAP_LINK_EXCL)
4147 drop &= ~CEPH_CAP_LINK_SHARED;
4148 if (unless & CEPH_CAP_XATTR_EXCL)
4149 drop &= ~CEPH_CAP_XATTR_SHARED;
4150 if (unless & CEPH_CAP_FILE_EXCL)
4151 drop &= ~CEPH_CAP_FILE_SHARED;
4154 if (force || (cap->issued & drop)) {
4155 if (cap->issued & drop) {
4156 int wanted = __ceph_caps_wanted(ci);
4157 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
4158 wanted |= cap->mds_wanted;
4159 dout("encode_inode_release %p cap %p "
4160 "%s -> %s, wanted %s -> %s\n", inode, cap,
4161 ceph_cap_string(cap->issued),
4162 ceph_cap_string(cap->issued & ~drop),
4163 ceph_cap_string(cap->mds_wanted),
4164 ceph_cap_string(wanted));
4166 cap->issued &= ~drop;
4167 cap->implemented &= ~drop;
4168 cap->mds_wanted = wanted;
4170 dout("encode_inode_release %p cap %p %s"
4171 " (force)\n", inode, cap,
4172 ceph_cap_string(cap->issued));
4175 rel->ino = cpu_to_le64(ceph_ino(inode));
4176 rel->cap_id = cpu_to_le64(cap->cap_id);
4177 rel->seq = cpu_to_le32(cap->seq);
4178 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4179 rel->mseq = cpu_to_le32(cap->mseq);
4180 rel->caps = cpu_to_le32(cap->implemented);
4181 rel->wanted = cpu_to_le32(cap->mds_wanted);
4187 dout("encode_inode_release %p cap %p %s (noop)\n",
4188 inode, cap, ceph_cap_string(cap->issued));
4191 spin_unlock(&ci->i_ceph_lock);
4195 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4197 int mds, int drop, int unless)
4199 struct dentry *parent = NULL;
4200 struct ceph_mds_request_release *rel = *p;
4201 struct ceph_dentry_info *di = ceph_dentry(dentry);
4206 * force an record for the directory caps if we have a dentry lease.
4207 * this is racy (can't take i_ceph_lock and d_lock together), but it
4208 * doesn't have to be perfect; the mds will revoke anything we don't
4211 spin_lock(&dentry->d_lock);
4212 if (di->lease_session && di->lease_session->s_mds == mds)
4215 parent = dget(dentry->d_parent);
4216 dir = d_inode(parent);
4218 spin_unlock(&dentry->d_lock);
4220 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4223 spin_lock(&dentry->d_lock);
4224 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4225 dout("encode_dentry_release %p mds%d seq %d\n",
4226 dentry, mds, (int)di->lease_seq);
4227 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4228 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4229 *p += dentry->d_name.len;
4230 rel->dname_seq = cpu_to_le32(di->lease_seq);
4231 __ceph_mdsc_drop_dentry_lease(dentry);
4233 spin_unlock(&dentry->d_lock);