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_min_caps(struct ceph_mds_client *mdsc, int delta)
153 spin_lock(&mdsc->caps_list_lock);
154 mdsc->caps_min_count += delta;
155 BUG_ON(mdsc->caps_min_count < 0);
156 spin_unlock(&mdsc->caps_list_lock);
159 static void __ceph_unreserve_caps(struct ceph_mds_client *mdsc, int nr_caps)
161 struct ceph_cap *cap;
165 BUG_ON(mdsc->caps_reserve_count < nr_caps);
166 mdsc->caps_reserve_count -= nr_caps;
167 if (mdsc->caps_avail_count >=
168 mdsc->caps_reserve_count + mdsc->caps_min_count) {
169 mdsc->caps_total_count -= nr_caps;
170 for (i = 0; i < nr_caps; i++) {
171 cap = list_first_entry(&mdsc->caps_list,
172 struct ceph_cap, caps_item);
173 list_del(&cap->caps_item);
174 kmem_cache_free(ceph_cap_cachep, cap);
177 mdsc->caps_avail_count += nr_caps;
180 dout("%s: caps %d = %d used + %d resv + %d avail\n",
182 mdsc->caps_total_count, mdsc->caps_use_count,
183 mdsc->caps_reserve_count, mdsc->caps_avail_count);
184 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
185 mdsc->caps_reserve_count +
186 mdsc->caps_avail_count);
191 * Called under mdsc->mutex.
193 int ceph_reserve_caps(struct ceph_mds_client *mdsc,
194 struct ceph_cap_reservation *ctx, int need)
197 struct ceph_cap *cap;
202 bool trimmed = false;
203 struct ceph_mds_session *s;
206 dout("reserve caps ctx=%p need=%d\n", ctx, need);
208 /* first reserve any caps that are already allocated */
209 spin_lock(&mdsc->caps_list_lock);
210 if (mdsc->caps_avail_count >= need)
213 have = mdsc->caps_avail_count;
214 mdsc->caps_avail_count -= have;
215 mdsc->caps_reserve_count += have;
216 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
217 mdsc->caps_reserve_count +
218 mdsc->caps_avail_count);
219 spin_unlock(&mdsc->caps_list_lock);
221 for (i = have; i < need; ) {
222 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
224 list_add(&cap->caps_item, &newcaps);
231 for (j = 0; j < mdsc->max_sessions; j++) {
232 s = __ceph_lookup_mds_session(mdsc, j);
235 mutex_unlock(&mdsc->mutex);
237 mutex_lock(&s->s_mutex);
238 max_caps = s->s_nr_caps - (need - i);
239 ceph_trim_caps(mdsc, s, max_caps);
240 mutex_unlock(&s->s_mutex);
242 ceph_put_mds_session(s);
243 mutex_lock(&mdsc->mutex);
247 spin_lock(&mdsc->caps_list_lock);
248 if (mdsc->caps_avail_count) {
250 if (mdsc->caps_avail_count >= need - i)
251 more_have = need - i;
253 more_have = mdsc->caps_avail_count;
257 mdsc->caps_avail_count -= more_have;
258 mdsc->caps_reserve_count += more_have;
261 spin_unlock(&mdsc->caps_list_lock);
266 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
267 ctx, need, have + alloc);
273 BUG_ON(have + alloc != need);
277 spin_lock(&mdsc->caps_list_lock);
278 mdsc->caps_total_count += alloc;
279 mdsc->caps_reserve_count += alloc;
280 list_splice(&newcaps, &mdsc->caps_list);
282 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
283 mdsc->caps_reserve_count +
284 mdsc->caps_avail_count);
287 __ceph_unreserve_caps(mdsc, have + alloc);
289 spin_unlock(&mdsc->caps_list_lock);
291 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
292 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
293 mdsc->caps_reserve_count, mdsc->caps_avail_count);
297 void ceph_unreserve_caps(struct ceph_mds_client *mdsc,
298 struct ceph_cap_reservation *ctx)
300 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
301 spin_lock(&mdsc->caps_list_lock);
302 __ceph_unreserve_caps(mdsc, ctx->count);
304 spin_unlock(&mdsc->caps_list_lock);
307 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
308 struct ceph_cap_reservation *ctx)
310 struct ceph_cap *cap = NULL;
312 /* temporary, until we do something about cap import/export */
314 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
316 spin_lock(&mdsc->caps_list_lock);
317 mdsc->caps_use_count++;
318 mdsc->caps_total_count++;
319 spin_unlock(&mdsc->caps_list_lock);
321 spin_lock(&mdsc->caps_list_lock);
322 if (mdsc->caps_avail_count) {
323 BUG_ON(list_empty(&mdsc->caps_list));
325 mdsc->caps_avail_count--;
326 mdsc->caps_use_count++;
327 cap = list_first_entry(&mdsc->caps_list,
328 struct ceph_cap, caps_item);
329 list_del(&cap->caps_item);
331 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
332 mdsc->caps_reserve_count + mdsc->caps_avail_count);
334 spin_unlock(&mdsc->caps_list_lock);
340 spin_lock(&mdsc->caps_list_lock);
341 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
342 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
343 mdsc->caps_reserve_count, mdsc->caps_avail_count);
345 BUG_ON(ctx->count > mdsc->caps_reserve_count);
346 BUG_ON(list_empty(&mdsc->caps_list));
349 mdsc->caps_reserve_count--;
350 mdsc->caps_use_count++;
352 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
353 list_del(&cap->caps_item);
355 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
356 mdsc->caps_reserve_count + mdsc->caps_avail_count);
357 spin_unlock(&mdsc->caps_list_lock);
361 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
363 spin_lock(&mdsc->caps_list_lock);
364 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
365 cap, mdsc->caps_total_count, mdsc->caps_use_count,
366 mdsc->caps_reserve_count, mdsc->caps_avail_count);
367 mdsc->caps_use_count--;
369 * Keep some preallocated caps around (ceph_min_count), to
370 * avoid lots of free/alloc churn.
372 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
373 mdsc->caps_min_count) {
374 mdsc->caps_total_count--;
375 kmem_cache_free(ceph_cap_cachep, cap);
377 mdsc->caps_avail_count++;
378 list_add(&cap->caps_item, &mdsc->caps_list);
381 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
382 mdsc->caps_reserve_count + mdsc->caps_avail_count);
383 spin_unlock(&mdsc->caps_list_lock);
386 void ceph_reservation_status(struct ceph_fs_client *fsc,
387 int *total, int *avail, int *used, int *reserved,
390 struct ceph_mds_client *mdsc = fsc->mdsc;
392 spin_lock(&mdsc->caps_list_lock);
395 *total = mdsc->caps_total_count;
397 *avail = mdsc->caps_avail_count;
399 *used = mdsc->caps_use_count;
401 *reserved = mdsc->caps_reserve_count;
403 *min = mdsc->caps_min_count;
405 spin_unlock(&mdsc->caps_list_lock);
409 * Find ceph_cap for given mds, if any.
411 * Called with i_ceph_lock held.
413 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
415 struct ceph_cap *cap;
416 struct rb_node *n = ci->i_caps.rb_node;
419 cap = rb_entry(n, struct ceph_cap, ci_node);
422 else if (mds > cap->mds)
430 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
432 struct ceph_cap *cap;
434 spin_lock(&ci->i_ceph_lock);
435 cap = __get_cap_for_mds(ci, mds);
436 spin_unlock(&ci->i_ceph_lock);
441 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
443 static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
445 struct ceph_cap *cap;
449 /* prefer mds with WR|BUFFER|EXCL caps */
450 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
451 cap = rb_entry(p, struct ceph_cap, ci_node);
453 if (cap->issued & (CEPH_CAP_FILE_WR |
454 CEPH_CAP_FILE_BUFFER |
461 int ceph_get_cap_mds(struct inode *inode)
463 struct ceph_inode_info *ci = ceph_inode(inode);
465 spin_lock(&ci->i_ceph_lock);
466 mds = __ceph_get_cap_mds(ceph_inode(inode));
467 spin_unlock(&ci->i_ceph_lock);
472 * Called under i_ceph_lock.
474 static void __insert_cap_node(struct ceph_inode_info *ci,
475 struct ceph_cap *new)
477 struct rb_node **p = &ci->i_caps.rb_node;
478 struct rb_node *parent = NULL;
479 struct ceph_cap *cap = NULL;
483 cap = rb_entry(parent, struct ceph_cap, ci_node);
484 if (new->mds < cap->mds)
486 else if (new->mds > cap->mds)
492 rb_link_node(&new->ci_node, parent, p);
493 rb_insert_color(&new->ci_node, &ci->i_caps);
497 * (re)set cap hold timeouts, which control the delayed release
498 * of unused caps back to the MDS. Should be called on cap use.
500 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
501 struct ceph_inode_info *ci)
503 struct ceph_mount_options *ma = mdsc->fsc->mount_options;
505 ci->i_hold_caps_min = round_jiffies(jiffies +
506 ma->caps_wanted_delay_min * HZ);
507 ci->i_hold_caps_max = round_jiffies(jiffies +
508 ma->caps_wanted_delay_max * HZ);
509 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
510 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
514 * (Re)queue cap at the end of the delayed cap release list.
516 * If I_FLUSH is set, leave the inode at the front of the list.
518 * Caller holds i_ceph_lock
519 * -> we take mdsc->cap_delay_lock
521 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
522 struct ceph_inode_info *ci,
525 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
526 ci->i_ceph_flags, ci->i_hold_caps_max);
527 if (!mdsc->stopping) {
528 spin_lock(&mdsc->cap_delay_lock);
529 if (!list_empty(&ci->i_cap_delay_list)) {
530 if (ci->i_ceph_flags & CEPH_I_FLUSH)
532 list_del_init(&ci->i_cap_delay_list);
535 __cap_set_timeouts(mdsc, ci);
536 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
538 spin_unlock(&mdsc->cap_delay_lock);
543 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
544 * indicating we should send a cap message to flush dirty metadata
545 * asap, and move to the front of the delayed cap list.
547 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
548 struct ceph_inode_info *ci)
550 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
551 spin_lock(&mdsc->cap_delay_lock);
552 ci->i_ceph_flags |= CEPH_I_FLUSH;
553 if (!list_empty(&ci->i_cap_delay_list))
554 list_del_init(&ci->i_cap_delay_list);
555 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
556 spin_unlock(&mdsc->cap_delay_lock);
560 * Cancel delayed work on cap.
562 * Caller must hold i_ceph_lock.
564 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
565 struct ceph_inode_info *ci)
567 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
568 if (list_empty(&ci->i_cap_delay_list))
570 spin_lock(&mdsc->cap_delay_lock);
571 list_del_init(&ci->i_cap_delay_list);
572 spin_unlock(&mdsc->cap_delay_lock);
576 * Common issue checks for add_cap, handle_cap_grant.
578 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
581 unsigned had = __ceph_caps_issued(ci, NULL);
584 * Each time we receive FILE_CACHE anew, we increment
587 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
588 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
593 * If FILE_SHARED is newly issued, mark dir not complete. We don't
594 * know what happened to this directory while we didn't have the cap.
595 * If FILE_SHARED is being revoked, also mark dir not complete. It
596 * stops on-going cached readdir.
598 if ((issued & CEPH_CAP_FILE_SHARED) != (had & CEPH_CAP_FILE_SHARED)) {
599 if (issued & CEPH_CAP_FILE_SHARED)
600 atomic_inc(&ci->i_shared_gen);
601 if (S_ISDIR(ci->vfs_inode.i_mode)) {
602 dout(" marking %p NOT complete\n", &ci->vfs_inode);
603 __ceph_dir_clear_complete(ci);
609 * Add a capability under the given MDS session.
611 * Caller should hold session snap_rwsem (read) and s_mutex.
613 * @fmode is the open file mode, if we are opening a file, otherwise
614 * it is < 0. (This is so we can atomically add the cap and add an
615 * open file reference to it.)
617 void ceph_add_cap(struct inode *inode,
618 struct ceph_mds_session *session, u64 cap_id,
619 int fmode, unsigned issued, unsigned wanted,
620 unsigned seq, unsigned mseq, u64 realmino, int flags,
621 struct ceph_cap **new_cap)
623 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
624 struct ceph_inode_info *ci = ceph_inode(inode);
625 struct ceph_cap *cap;
626 int mds = session->s_mds;
629 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
630 session->s_mds, cap_id, ceph_cap_string(issued), seq);
633 * If we are opening the file, include file mode wanted bits
637 wanted |= ceph_caps_for_mode(fmode);
639 cap = __get_cap_for_mds(ci, mds);
645 cap->implemented = 0;
651 __insert_cap_node(ci, cap);
653 /* add to session cap list */
654 cap->session = session;
655 spin_lock(&session->s_cap_lock);
656 list_add_tail(&cap->session_caps, &session->s_caps);
657 session->s_nr_caps++;
658 spin_unlock(&session->s_cap_lock);
660 if (cap->cap_gen < session->s_cap_gen)
661 cap->issued = cap->implemented = CEPH_CAP_PIN;
664 * auth mds of the inode changed. we received the cap export
665 * message, but still haven't received the cap import message.
666 * handle_cap_export() updated the new auth MDS' cap.
668 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
669 * a message that was send before the cap import message. So
672 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
673 WARN_ON(cap != ci->i_auth_cap);
674 WARN_ON(cap->cap_id != cap_id);
677 issued |= cap->issued;
678 flags |= CEPH_CAP_FLAG_AUTH;
682 if (!ci->i_snap_realm ||
683 ((flags & CEPH_CAP_FLAG_AUTH) &&
684 realmino != (u64)-1 && ci->i_snap_realm->ino != realmino)) {
686 * add this inode to the appropriate snap realm
688 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
691 struct ceph_snap_realm *oldrealm = ci->i_snap_realm;
693 spin_lock(&oldrealm->inodes_with_caps_lock);
694 list_del_init(&ci->i_snap_realm_item);
695 spin_unlock(&oldrealm->inodes_with_caps_lock);
698 spin_lock(&realm->inodes_with_caps_lock);
699 list_add(&ci->i_snap_realm_item,
700 &realm->inodes_with_caps);
701 ci->i_snap_realm = realm;
702 if (realm->ino == ci->i_vino.ino)
703 realm->inode = inode;
704 spin_unlock(&realm->inodes_with_caps_lock);
707 ceph_put_snap_realm(mdsc, oldrealm);
709 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
715 __check_cap_issue(ci, cap, issued);
718 * If we are issued caps we don't want, or the mds' wanted
719 * value appears to be off, queue a check so we'll release
720 * later and/or update the mds wanted value.
722 actual_wanted = __ceph_caps_wanted(ci);
723 if ((wanted & ~actual_wanted) ||
724 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
725 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
726 ceph_cap_string(issued), ceph_cap_string(wanted),
727 ceph_cap_string(actual_wanted));
728 __cap_delay_requeue(mdsc, ci, true);
731 if (flags & CEPH_CAP_FLAG_AUTH) {
732 if (!ci->i_auth_cap ||
733 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
734 ci->i_auth_cap = cap;
735 cap->mds_wanted = wanted;
738 WARN_ON(ci->i_auth_cap == cap);
741 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
742 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
743 ceph_cap_string(issued|cap->issued), seq, mds);
744 cap->cap_id = cap_id;
745 cap->issued = issued;
746 cap->implemented |= issued;
747 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
748 cap->mds_wanted = wanted;
750 cap->mds_wanted |= wanted;
752 cap->issue_seq = seq;
754 cap->cap_gen = session->s_cap_gen;
757 __ceph_get_fmode(ci, fmode);
761 * Return true if cap has not timed out and belongs to the current
762 * generation of the MDS session (i.e. has not gone 'stale' due to
763 * us losing touch with the mds).
765 static int __cap_is_valid(struct ceph_cap *cap)
770 spin_lock(&cap->session->s_gen_ttl_lock);
771 gen = cap->session->s_cap_gen;
772 ttl = cap->session->s_cap_ttl;
773 spin_unlock(&cap->session->s_gen_ttl_lock);
775 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
776 dout("__cap_is_valid %p cap %p issued %s "
777 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
778 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
786 * Return set of valid cap bits issued to us. Note that caps time
787 * out, and may be invalidated in bulk if the client session times out
788 * and session->s_cap_gen is bumped.
790 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
792 int have = ci->i_snap_caps;
793 struct ceph_cap *cap;
798 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
799 cap = rb_entry(p, struct ceph_cap, ci_node);
800 if (!__cap_is_valid(cap))
802 dout("__ceph_caps_issued %p cap %p issued %s\n",
803 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
806 *implemented |= cap->implemented;
809 * exclude caps issued by non-auth MDS, but are been revoking
810 * by the auth MDS. The non-auth MDS should be revoking/exporting
811 * these caps, but the message is delayed.
813 if (ci->i_auth_cap) {
814 cap = ci->i_auth_cap;
815 have &= ~cap->implemented | cap->issued;
821 * Get cap bits issued by caps other than @ocap
823 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
825 int have = ci->i_snap_caps;
826 struct ceph_cap *cap;
829 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
830 cap = rb_entry(p, struct ceph_cap, ci_node);
833 if (!__cap_is_valid(cap))
841 * Move a cap to the end of the LRU (oldest caps at list head, newest
844 static void __touch_cap(struct ceph_cap *cap)
846 struct ceph_mds_session *s = cap->session;
848 spin_lock(&s->s_cap_lock);
849 if (!s->s_cap_iterator) {
850 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
852 list_move_tail(&cap->session_caps, &s->s_caps);
854 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
855 &cap->ci->vfs_inode, cap, s->s_mds);
857 spin_unlock(&s->s_cap_lock);
861 * Check if we hold the given mask. If so, move the cap(s) to the
862 * front of their respective LRUs. (This is the preferred way for
863 * callers to check for caps they want.)
865 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
867 struct ceph_cap *cap;
869 int have = ci->i_snap_caps;
871 if ((have & mask) == mask) {
872 dout("__ceph_caps_issued_mask %p snap issued %s"
873 " (mask %s)\n", &ci->vfs_inode,
874 ceph_cap_string(have),
875 ceph_cap_string(mask));
879 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
880 cap = rb_entry(p, struct ceph_cap, ci_node);
881 if (!__cap_is_valid(cap))
883 if ((cap->issued & mask) == mask) {
884 dout("__ceph_caps_issued_mask %p cap %p issued %s"
885 " (mask %s)\n", &ci->vfs_inode, cap,
886 ceph_cap_string(cap->issued),
887 ceph_cap_string(mask));
893 /* does a combination of caps satisfy mask? */
895 if ((have & mask) == mask) {
896 dout("__ceph_caps_issued_mask %p combo issued %s"
897 " (mask %s)\n", &ci->vfs_inode,
898 ceph_cap_string(cap->issued),
899 ceph_cap_string(mask));
903 /* touch this + preceding caps */
905 for (q = rb_first(&ci->i_caps); q != p;
907 cap = rb_entry(q, struct ceph_cap,
909 if (!__cap_is_valid(cap))
922 * Return true if mask caps are currently being revoked by an MDS.
924 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
925 struct ceph_cap *ocap, int mask)
927 struct ceph_cap *cap;
930 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
931 cap = rb_entry(p, struct ceph_cap, ci_node);
933 (cap->implemented & ~cap->issued & mask))
939 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
941 struct inode *inode = &ci->vfs_inode;
944 spin_lock(&ci->i_ceph_lock);
945 ret = __ceph_caps_revoking_other(ci, NULL, mask);
946 spin_unlock(&ci->i_ceph_lock);
947 dout("ceph_caps_revoking %p %s = %d\n", inode,
948 ceph_cap_string(mask), ret);
952 int __ceph_caps_used(struct ceph_inode_info *ci)
956 used |= CEPH_CAP_PIN;
958 used |= CEPH_CAP_FILE_RD;
959 if (ci->i_rdcache_ref ||
960 (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */
961 ci->vfs_inode.i_data.nrpages))
962 used |= CEPH_CAP_FILE_CACHE;
964 used |= CEPH_CAP_FILE_WR;
965 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
966 used |= CEPH_CAP_FILE_BUFFER;
971 * wanted, by virtue of open file modes
973 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
976 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
977 if (ci->i_nr_by_mode[i])
982 return ceph_caps_for_mode(bits >> 1);
986 * Return caps we have registered with the MDS(s) as 'wanted'.
988 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci, bool check)
990 struct ceph_cap *cap;
994 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
995 cap = rb_entry(p, struct ceph_cap, ci_node);
996 if (check && !__cap_is_valid(cap))
998 if (cap == ci->i_auth_cap)
999 mds_wanted |= cap->mds_wanted;
1001 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
1007 * called under i_ceph_lock
1009 static int __ceph_is_single_caps(struct ceph_inode_info *ci)
1011 return rb_first(&ci->i_caps) == rb_last(&ci->i_caps);
1014 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
1016 return !RB_EMPTY_ROOT(&ci->i_caps);
1019 int ceph_is_any_caps(struct inode *inode)
1021 struct ceph_inode_info *ci = ceph_inode(inode);
1024 spin_lock(&ci->i_ceph_lock);
1025 ret = __ceph_is_any_caps(ci);
1026 spin_unlock(&ci->i_ceph_lock);
1031 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
1033 struct ceph_snap_realm *realm = ci->i_snap_realm;
1034 spin_lock(&realm->inodes_with_caps_lock);
1035 list_del_init(&ci->i_snap_realm_item);
1036 ci->i_snap_realm_counter++;
1037 ci->i_snap_realm = NULL;
1038 if (realm->ino == ci->i_vino.ino)
1039 realm->inode = NULL;
1040 spin_unlock(&realm->inodes_with_caps_lock);
1041 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
1046 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
1048 * caller should hold i_ceph_lock.
1049 * caller will not hold session s_mutex if called from destroy_inode.
1051 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
1053 struct ceph_mds_session *session = cap->session;
1054 struct ceph_inode_info *ci = cap->ci;
1055 struct ceph_mds_client *mdsc =
1056 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1059 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
1061 /* remove from session list */
1062 spin_lock(&session->s_cap_lock);
1063 if (session->s_cap_iterator == cap) {
1064 /* not yet, we are iterating over this very cap */
1065 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1068 list_del_init(&cap->session_caps);
1069 session->s_nr_caps--;
1070 cap->session = NULL;
1073 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1077 * s_cap_reconnect is protected by s_cap_lock. no one changes
1078 * s_cap_gen while session is in the reconnect state.
1080 if (queue_release &&
1081 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
1082 cap->queue_release = 1;
1084 list_add_tail(&cap->session_caps,
1085 &session->s_cap_releases);
1086 session->s_num_cap_releases++;
1090 cap->queue_release = 0;
1092 cap->cap_ino = ci->i_vino.ino;
1094 spin_unlock(&session->s_cap_lock);
1096 /* remove from inode list */
1097 rb_erase(&cap->ci_node, &ci->i_caps);
1098 if (ci->i_auth_cap == cap)
1099 ci->i_auth_cap = NULL;
1102 ceph_put_cap(mdsc, cap);
1104 /* when reconnect denied, we remove session caps forcibly,
1105 * i_wr_ref can be non-zero. If there are ongoing write,
1106 * keep i_snap_realm.
1108 if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm)
1109 drop_inode_snap_realm(ci);
1111 if (!__ceph_is_any_real_caps(ci))
1112 __cap_delay_cancel(mdsc, ci);
1115 struct cap_msg_args {
1116 struct ceph_mds_session *session;
1117 u64 ino, cid, follows;
1118 u64 flush_tid, oldest_flush_tid, size, max_size;
1120 struct ceph_buffer *xattr_buf;
1121 struct timespec64 atime, mtime, ctime;
1122 int op, caps, wanted, dirty;
1123 u32 seq, issue_seq, mseq, time_warp_seq;
1132 * Build and send a cap message to the given MDS.
1134 * Caller should be holding s_mutex.
1136 static int send_cap_msg(struct cap_msg_args *arg)
1138 struct ceph_mds_caps *fc;
1139 struct ceph_msg *msg;
1142 struct timespec64 zerotime = {0};
1143 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1145 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1146 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1147 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(arg->op),
1148 arg->cid, arg->ino, ceph_cap_string(arg->caps),
1149 ceph_cap_string(arg->wanted), ceph_cap_string(arg->dirty),
1150 arg->seq, arg->issue_seq, arg->flush_tid, arg->oldest_flush_tid,
1151 arg->mseq, arg->follows, arg->size, arg->max_size,
1153 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1155 /* flock buffer size + inline version + inline data size +
1156 * osd_epoch_barrier + oldest_flush_tid */
1157 extra_len = 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4;
1158 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1163 msg->hdr.version = cpu_to_le16(10);
1164 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1166 fc = msg->front.iov_base;
1167 memset(fc, 0, sizeof(*fc));
1169 fc->cap_id = cpu_to_le64(arg->cid);
1170 fc->op = cpu_to_le32(arg->op);
1171 fc->seq = cpu_to_le32(arg->seq);
1172 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1173 fc->migrate_seq = cpu_to_le32(arg->mseq);
1174 fc->caps = cpu_to_le32(arg->caps);
1175 fc->wanted = cpu_to_le32(arg->wanted);
1176 fc->dirty = cpu_to_le32(arg->dirty);
1177 fc->ino = cpu_to_le64(arg->ino);
1178 fc->snap_follows = cpu_to_le64(arg->follows);
1180 fc->size = cpu_to_le64(arg->size);
1181 fc->max_size = cpu_to_le64(arg->max_size);
1182 ceph_encode_timespec64(&fc->mtime, &arg->mtime);
1183 ceph_encode_timespec64(&fc->atime, &arg->atime);
1184 ceph_encode_timespec64(&fc->ctime, &arg->ctime);
1185 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1187 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1188 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1189 fc->mode = cpu_to_le32(arg->mode);
1191 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1192 if (arg->xattr_buf) {
1193 msg->middle = ceph_buffer_get(arg->xattr_buf);
1194 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1195 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1199 /* flock buffer size (version 2) */
1200 ceph_encode_32(&p, 0);
1201 /* inline version (version 4) */
1202 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1203 /* inline data size */
1204 ceph_encode_32(&p, 0);
1206 * osd_epoch_barrier (version 5)
1207 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1208 * case it was recently changed
1210 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1211 /* oldest_flush_tid (version 6) */
1212 ceph_encode_64(&p, arg->oldest_flush_tid);
1215 * caller_uid/caller_gid (version 7)
1217 * Currently, we don't properly track which caller dirtied the caps
1218 * last, and force a flush of them when there is a conflict. For now,
1219 * just set this to 0:0, to emulate how the MDS has worked up to now.
1221 ceph_encode_32(&p, 0);
1222 ceph_encode_32(&p, 0);
1224 /* pool namespace (version 8) (mds always ignores this) */
1225 ceph_encode_32(&p, 0);
1228 * btime and change_attr (version 9)
1230 * We just zero these out for now, as the MDS ignores them unless
1231 * the requisite feature flags are set (which we don't do yet).
1233 ceph_encode_timespec64(p, &zerotime);
1234 p += sizeof(struct ceph_timespec);
1235 ceph_encode_64(&p, 0);
1237 /* Advisory flags (version 10) */
1238 ceph_encode_32(&p, arg->flags);
1240 ceph_con_send(&arg->session->s_con, msg);
1245 * Queue cap releases when an inode is dropped from our cache. Since
1246 * inode is about to be destroyed, there is no need for i_ceph_lock.
1248 void ceph_queue_caps_release(struct inode *inode)
1250 struct ceph_inode_info *ci = ceph_inode(inode);
1253 p = rb_first(&ci->i_caps);
1255 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1257 __ceph_remove_cap(cap, true);
1262 * Send a cap msg on the given inode. Update our caps state, then
1263 * drop i_ceph_lock and send the message.
1265 * Make note of max_size reported/requested from mds, revoked caps
1266 * that have now been implemented.
1268 * Make half-hearted attempt ot to invalidate page cache if we are
1269 * dropping RDCACHE. Note that this will leave behind locked pages
1270 * that we'll then need to deal with elsewhere.
1272 * Return non-zero if delayed release, or we experienced an error
1273 * such that the caller should requeue + retry later.
1275 * called with i_ceph_lock, then drops it.
1276 * caller should hold snap_rwsem (read), s_mutex.
1278 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1279 int op, bool sync, int used, int want, int retain,
1280 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1281 __releases(cap->ci->i_ceph_lock)
1283 struct ceph_inode_info *ci = cap->ci;
1284 struct inode *inode = &ci->vfs_inode;
1285 struct cap_msg_args arg;
1291 held = cap->issued | cap->implemented;
1292 revoking = cap->implemented & ~cap->issued;
1293 retain &= ~revoking;
1295 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1296 inode, cap, cap->session,
1297 ceph_cap_string(held), ceph_cap_string(held & retain),
1298 ceph_cap_string(revoking));
1299 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1301 arg.session = cap->session;
1303 /* don't release wanted unless we've waited a bit. */
1304 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1305 time_before(jiffies, ci->i_hold_caps_min)) {
1306 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1307 ceph_cap_string(cap->issued),
1308 ceph_cap_string(cap->issued & retain),
1309 ceph_cap_string(cap->mds_wanted),
1310 ceph_cap_string(want));
1311 want |= cap->mds_wanted;
1312 retain |= cap->issued;
1315 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1316 if (want & ~cap->mds_wanted) {
1317 /* user space may open/close single file frequently.
1318 * This avoids droping mds_wanted immediately after
1319 * requesting new mds_wanted.
1321 __cap_set_timeouts(mdsc, ci);
1324 cap->issued &= retain; /* drop bits we don't want */
1325 if (cap->implemented & ~cap->issued) {
1327 * Wake up any waiters on wanted -> needed transition.
1328 * This is due to the weird transition from buffered
1329 * to sync IO... we need to flush dirty pages _before_
1330 * allowing sync writes to avoid reordering.
1334 cap->implemented &= cap->issued | used;
1335 cap->mds_wanted = want;
1337 arg.ino = ceph_vino(inode).ino;
1338 arg.cid = cap->cap_id;
1339 arg.follows = flushing ? ci->i_head_snapc->seq : 0;
1340 arg.flush_tid = flush_tid;
1341 arg.oldest_flush_tid = oldest_flush_tid;
1343 arg.size = inode->i_size;
1344 ci->i_reported_size = arg.size;
1345 arg.max_size = ci->i_wanted_max_size;
1346 ci->i_requested_max_size = arg.max_size;
1348 if (flushing & CEPH_CAP_XATTR_EXCL) {
1349 __ceph_build_xattrs_blob(ci);
1350 arg.xattr_version = ci->i_xattrs.version;
1351 arg.xattr_buf = ci->i_xattrs.blob;
1353 arg.xattr_buf = NULL;
1356 arg.mtime = inode->i_mtime;
1357 arg.atime = inode->i_atime;
1358 arg.ctime = inode->i_ctime;
1361 arg.caps = cap->implemented;
1363 arg.dirty = flushing;
1366 arg.issue_seq = cap->issue_seq;
1367 arg.mseq = cap->mseq;
1368 arg.time_warp_seq = ci->i_time_warp_seq;
1370 arg.uid = inode->i_uid;
1371 arg.gid = inode->i_gid;
1372 arg.mode = inode->i_mode;
1374 arg.inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1375 if (list_empty(&ci->i_cap_snaps))
1376 arg.flags = CEPH_CLIENT_CAPS_NO_CAPSNAP;
1378 arg.flags = CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1380 arg.flags |= CEPH_CLIENT_CAPS_SYNC;
1382 spin_unlock(&ci->i_ceph_lock);
1384 ret = send_cap_msg(&arg);
1386 dout("error sending cap msg, must requeue %p\n", inode);
1391 wake_up_all(&ci->i_cap_wq);
1396 static inline int __send_flush_snap(struct inode *inode,
1397 struct ceph_mds_session *session,
1398 struct ceph_cap_snap *capsnap,
1399 u32 mseq, u64 oldest_flush_tid)
1401 struct cap_msg_args arg;
1403 arg.session = session;
1404 arg.ino = ceph_vino(inode).ino;
1406 arg.follows = capsnap->follows;
1407 arg.flush_tid = capsnap->cap_flush.tid;
1408 arg.oldest_flush_tid = oldest_flush_tid;
1410 arg.size = capsnap->size;
1412 arg.xattr_version = capsnap->xattr_version;
1413 arg.xattr_buf = capsnap->xattr_blob;
1415 arg.atime = capsnap->atime;
1416 arg.mtime = capsnap->mtime;
1417 arg.ctime = capsnap->ctime;
1419 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1420 arg.caps = capsnap->issued;
1422 arg.dirty = capsnap->dirty;
1427 arg.time_warp_seq = capsnap->time_warp_seq;
1429 arg.uid = capsnap->uid;
1430 arg.gid = capsnap->gid;
1431 arg.mode = capsnap->mode;
1433 arg.inline_data = capsnap->inline_data;
1436 return send_cap_msg(&arg);
1440 * When a snapshot is taken, clients accumulate dirty metadata on
1441 * inodes with capabilities in ceph_cap_snaps to describe the file
1442 * state at the time the snapshot was taken. This must be flushed
1443 * asynchronously back to the MDS once sync writes complete and dirty
1444 * data is written out.
1446 * Called under i_ceph_lock. Takes s_mutex as needed.
1448 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1449 struct ceph_mds_session *session)
1450 __releases(ci->i_ceph_lock)
1451 __acquires(ci->i_ceph_lock)
1453 struct inode *inode = &ci->vfs_inode;
1454 struct ceph_mds_client *mdsc = session->s_mdsc;
1455 struct ceph_cap_snap *capsnap;
1456 u64 oldest_flush_tid = 0;
1457 u64 first_tid = 1, last_tid = 0;
1459 dout("__flush_snaps %p session %p\n", inode, session);
1461 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1463 * we need to wait for sync writes to complete and for dirty
1464 * pages to be written out.
1466 if (capsnap->dirty_pages || capsnap->writing)
1469 /* should be removed by ceph_try_drop_cap_snap() */
1470 BUG_ON(!capsnap->need_flush);
1472 /* only flush each capsnap once */
1473 if (capsnap->cap_flush.tid > 0) {
1474 dout(" already flushed %p, skipping\n", capsnap);
1478 spin_lock(&mdsc->cap_dirty_lock);
1479 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1480 list_add_tail(&capsnap->cap_flush.g_list,
1481 &mdsc->cap_flush_list);
1482 if (oldest_flush_tid == 0)
1483 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1484 if (list_empty(&ci->i_flushing_item)) {
1485 list_add_tail(&ci->i_flushing_item,
1486 &session->s_cap_flushing);
1488 spin_unlock(&mdsc->cap_dirty_lock);
1490 list_add_tail(&capsnap->cap_flush.i_list,
1491 &ci->i_cap_flush_list);
1494 first_tid = capsnap->cap_flush.tid;
1495 last_tid = capsnap->cap_flush.tid;
1498 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1500 while (first_tid <= last_tid) {
1501 struct ceph_cap *cap = ci->i_auth_cap;
1502 struct ceph_cap_flush *cf;
1505 if (!(cap && cap->session == session)) {
1506 dout("__flush_snaps %p auth cap %p not mds%d, "
1507 "stop\n", inode, cap, session->s_mds);
1512 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1513 if (cf->tid >= first_tid) {
1521 first_tid = cf->tid + 1;
1523 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1524 refcount_inc(&capsnap->nref);
1525 spin_unlock(&ci->i_ceph_lock);
1527 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1528 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1530 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1533 pr_err("__flush_snaps: error sending cap flushsnap, "
1534 "ino (%llx.%llx) tid %llu follows %llu\n",
1535 ceph_vinop(inode), cf->tid, capsnap->follows);
1538 ceph_put_cap_snap(capsnap);
1539 spin_lock(&ci->i_ceph_lock);
1543 void ceph_flush_snaps(struct ceph_inode_info *ci,
1544 struct ceph_mds_session **psession)
1546 struct inode *inode = &ci->vfs_inode;
1547 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1548 struct ceph_mds_session *session = NULL;
1551 dout("ceph_flush_snaps %p\n", inode);
1553 session = *psession;
1555 spin_lock(&ci->i_ceph_lock);
1556 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1557 dout(" no capsnap needs flush, doing nothing\n");
1560 if (!ci->i_auth_cap) {
1561 dout(" no auth cap (migrating?), doing nothing\n");
1565 mds = ci->i_auth_cap->session->s_mds;
1566 if (session && session->s_mds != mds) {
1567 dout(" oops, wrong session %p mutex\n", session);
1568 mutex_unlock(&session->s_mutex);
1569 ceph_put_mds_session(session);
1573 spin_unlock(&ci->i_ceph_lock);
1574 mutex_lock(&mdsc->mutex);
1575 session = __ceph_lookup_mds_session(mdsc, mds);
1576 mutex_unlock(&mdsc->mutex);
1578 dout(" inverting session/ino locks on %p\n", session);
1579 mutex_lock(&session->s_mutex);
1584 // make sure flushsnap messages are sent in proper order.
1585 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
1586 __kick_flushing_caps(mdsc, session, ci, 0);
1587 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
1590 __ceph_flush_snaps(ci, session);
1592 spin_unlock(&ci->i_ceph_lock);
1595 *psession = session;
1596 } else if (session) {
1597 mutex_unlock(&session->s_mutex);
1598 ceph_put_mds_session(session);
1600 /* we flushed them all; remove this inode from the queue */
1601 spin_lock(&mdsc->snap_flush_lock);
1602 list_del_init(&ci->i_snap_flush_item);
1603 spin_unlock(&mdsc->snap_flush_lock);
1607 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1608 * Caller is then responsible for calling __mark_inode_dirty with the
1609 * returned flags value.
1611 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1612 struct ceph_cap_flush **pcf)
1614 struct ceph_mds_client *mdsc =
1615 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1616 struct inode *inode = &ci->vfs_inode;
1617 int was = ci->i_dirty_caps;
1620 if (!ci->i_auth_cap) {
1621 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1622 "but no auth cap (session was closed?)\n",
1623 inode, ceph_ino(inode), ceph_cap_string(mask));
1627 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1628 ceph_cap_string(mask), ceph_cap_string(was),
1629 ceph_cap_string(was | mask));
1630 ci->i_dirty_caps |= mask;
1632 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1633 swap(ci->i_prealloc_cap_flush, *pcf);
1635 if (!ci->i_head_snapc) {
1636 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1637 ci->i_head_snapc = ceph_get_snap_context(
1638 ci->i_snap_realm->cached_context);
1640 dout(" inode %p now dirty snapc %p auth cap %p\n",
1641 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1642 BUG_ON(!list_empty(&ci->i_dirty_item));
1643 spin_lock(&mdsc->cap_dirty_lock);
1644 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1645 spin_unlock(&mdsc->cap_dirty_lock);
1646 if (ci->i_flushing_caps == 0) {
1648 dirty |= I_DIRTY_SYNC;
1651 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1653 BUG_ON(list_empty(&ci->i_dirty_item));
1654 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1655 (mask & CEPH_CAP_FILE_BUFFER))
1656 dirty |= I_DIRTY_DATASYNC;
1657 __cap_delay_requeue(mdsc, ci, true);
1661 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1663 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1666 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1669 kmem_cache_free(ceph_cap_flush_cachep, cf);
1672 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1674 if (!list_empty(&mdsc->cap_flush_list)) {
1675 struct ceph_cap_flush *cf =
1676 list_first_entry(&mdsc->cap_flush_list,
1677 struct ceph_cap_flush, g_list);
1684 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1685 * Return true if caller needs to wake up flush waiters.
1687 static bool __finish_cap_flush(struct ceph_mds_client *mdsc,
1688 struct ceph_inode_info *ci,
1689 struct ceph_cap_flush *cf)
1691 struct ceph_cap_flush *prev;
1692 bool wake = cf->wake;
1694 /* are there older pending cap flushes? */
1695 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1696 prev = list_prev_entry(cf, g_list);
1700 list_del(&cf->g_list);
1702 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1703 prev = list_prev_entry(cf, i_list);
1707 list_del(&cf->i_list);
1715 * Add dirty inode to the flushing list. Assigned a seq number so we
1716 * can wait for caps to flush without starving.
1718 * Called under i_ceph_lock.
1720 static int __mark_caps_flushing(struct inode *inode,
1721 struct ceph_mds_session *session, bool wake,
1722 u64 *flush_tid, u64 *oldest_flush_tid)
1724 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1725 struct ceph_inode_info *ci = ceph_inode(inode);
1726 struct ceph_cap_flush *cf = NULL;
1729 BUG_ON(ci->i_dirty_caps == 0);
1730 BUG_ON(list_empty(&ci->i_dirty_item));
1731 BUG_ON(!ci->i_prealloc_cap_flush);
1733 flushing = ci->i_dirty_caps;
1734 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1735 ceph_cap_string(flushing),
1736 ceph_cap_string(ci->i_flushing_caps),
1737 ceph_cap_string(ci->i_flushing_caps | flushing));
1738 ci->i_flushing_caps |= flushing;
1739 ci->i_dirty_caps = 0;
1740 dout(" inode %p now !dirty\n", inode);
1742 swap(cf, ci->i_prealloc_cap_flush);
1743 cf->caps = flushing;
1746 spin_lock(&mdsc->cap_dirty_lock);
1747 list_del_init(&ci->i_dirty_item);
1749 cf->tid = ++mdsc->last_cap_flush_tid;
1750 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1751 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1753 if (list_empty(&ci->i_flushing_item)) {
1754 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1755 mdsc->num_cap_flushing++;
1757 spin_unlock(&mdsc->cap_dirty_lock);
1759 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1761 *flush_tid = cf->tid;
1766 * try to invalidate mapping pages without blocking.
1768 static int try_nonblocking_invalidate(struct inode *inode)
1770 struct ceph_inode_info *ci = ceph_inode(inode);
1771 u32 invalidating_gen = ci->i_rdcache_gen;
1773 spin_unlock(&ci->i_ceph_lock);
1774 invalidate_mapping_pages(&inode->i_data, 0, -1);
1775 spin_lock(&ci->i_ceph_lock);
1777 if (inode->i_data.nrpages == 0 &&
1778 invalidating_gen == ci->i_rdcache_gen) {
1780 dout("try_nonblocking_invalidate %p success\n", inode);
1781 /* save any racing async invalidate some trouble */
1782 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1785 dout("try_nonblocking_invalidate %p failed\n", inode);
1789 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1791 loff_t size = ci->vfs_inode.i_size;
1792 /* mds will adjust max size according to the reported size */
1793 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1795 if (size >= ci->i_max_size)
1797 /* half of previous max_size increment has been used */
1798 if (ci->i_max_size > ci->i_reported_size &&
1799 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1805 * Swiss army knife function to examine currently used and wanted
1806 * versus held caps. Release, flush, ack revoked caps to mds as
1809 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1810 * cap release further.
1811 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1812 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1815 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1816 struct ceph_mds_session *session)
1818 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1819 struct ceph_mds_client *mdsc = fsc->mdsc;
1820 struct inode *inode = &ci->vfs_inode;
1821 struct ceph_cap *cap;
1822 u64 flush_tid, oldest_flush_tid;
1823 int file_wanted, used, cap_used;
1824 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1825 int issued, implemented, want, retain, revoking, flushing = 0;
1826 int mds = -1; /* keep track of how far we've gone through i_caps list
1827 to avoid an infinite loop on retry */
1829 int delayed = 0, sent = 0;
1830 bool no_delay = flags & CHECK_CAPS_NODELAY;
1831 bool queue_invalidate = false;
1832 bool tried_invalidate = false;
1834 /* if we are unmounting, flush any unused caps immediately. */
1838 spin_lock(&ci->i_ceph_lock);
1840 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1841 flags |= CHECK_CAPS_FLUSH;
1843 if (!(flags & CHECK_CAPS_AUTHONLY) ||
1844 (ci->i_auth_cap && __ceph_is_single_caps(ci)))
1845 __cap_delay_cancel(mdsc, ci);
1849 spin_lock(&ci->i_ceph_lock);
1851 file_wanted = __ceph_caps_file_wanted(ci);
1852 used = __ceph_caps_used(ci);
1853 issued = __ceph_caps_issued(ci, &implemented);
1854 revoking = implemented & ~issued;
1857 retain = file_wanted | used | CEPH_CAP_PIN;
1858 if (!mdsc->stopping && inode->i_nlink > 0) {
1860 retain |= CEPH_CAP_ANY; /* be greedy */
1861 } else if (S_ISDIR(inode->i_mode) &&
1862 (issued & CEPH_CAP_FILE_SHARED) &&
1863 __ceph_dir_is_complete(ci)) {
1865 * If a directory is complete, we want to keep
1866 * the exclusive cap. So that MDS does not end up
1867 * revoking the shared cap on every create/unlink
1870 if (IS_RDONLY(inode))
1871 want = CEPH_CAP_ANY_SHARED;
1873 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1877 retain |= CEPH_CAP_ANY_SHARED;
1879 * keep RD only if we didn't have the file open RW,
1880 * because then the mds would revoke it anyway to
1881 * journal max_size=0.
1883 if (ci->i_max_size == 0)
1884 retain |= CEPH_CAP_ANY_RD;
1888 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1889 " issued %s revoking %s retain %s %s%s%s\n", inode,
1890 ceph_cap_string(file_wanted),
1891 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1892 ceph_cap_string(ci->i_flushing_caps),
1893 ceph_cap_string(issued), ceph_cap_string(revoking),
1894 ceph_cap_string(retain),
1895 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1896 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1897 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1900 * If we no longer need to hold onto old our caps, and we may
1901 * have cached pages, but don't want them, then try to invalidate.
1902 * If we fail, it's because pages are locked.... try again later.
1904 if ((!no_delay || mdsc->stopping) &&
1905 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */
1906 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1907 inode->i_data.nrpages && /* have cached pages */
1908 (revoking & (CEPH_CAP_FILE_CACHE|
1909 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1910 !tried_invalidate) {
1911 dout("check_caps trying to invalidate on %p\n", inode);
1912 if (try_nonblocking_invalidate(inode) < 0) {
1913 dout("check_caps queuing invalidate\n");
1914 queue_invalidate = true;
1915 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1917 tried_invalidate = true;
1921 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1922 cap = rb_entry(p, struct ceph_cap, ci_node);
1924 /* avoid looping forever */
1925 if (mds >= cap->mds ||
1926 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1929 /* NOTE: no side-effects allowed, until we take s_mutex */
1932 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1933 cap_used &= ~ci->i_auth_cap->issued;
1935 revoking = cap->implemented & ~cap->issued;
1936 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1937 cap->mds, cap, ceph_cap_string(cap_used),
1938 ceph_cap_string(cap->issued),
1939 ceph_cap_string(cap->implemented),
1940 ceph_cap_string(revoking));
1942 if (cap == ci->i_auth_cap &&
1943 (cap->issued & CEPH_CAP_FILE_WR)) {
1944 /* request larger max_size from MDS? */
1945 if (ci->i_wanted_max_size > ci->i_max_size &&
1946 ci->i_wanted_max_size > ci->i_requested_max_size) {
1947 dout("requesting new max_size\n");
1951 /* approaching file_max? */
1952 if (__ceph_should_report_size(ci)) {
1953 dout("i_size approaching max_size\n");
1957 /* flush anything dirty? */
1958 if (cap == ci->i_auth_cap) {
1959 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
1960 dout("flushing dirty caps\n");
1963 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
1964 dout("flushing snap caps\n");
1969 /* completed revocation? going down and there are no caps? */
1970 if (revoking && (revoking & cap_used) == 0) {
1971 dout("completed revocation of %s\n",
1972 ceph_cap_string(cap->implemented & ~cap->issued));
1976 /* want more caps from mds? */
1977 if (want & ~(cap->mds_wanted | cap->issued))
1980 /* things we might delay */
1981 if ((cap->issued & ~retain) == 0)
1982 continue; /* nope, all good */
1988 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1989 time_before(jiffies, ci->i_hold_caps_max)) {
1990 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1991 ceph_cap_string(cap->issued),
1992 ceph_cap_string(cap->issued & retain),
1993 ceph_cap_string(cap->mds_wanted),
1994 ceph_cap_string(want));
2000 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
2001 dout(" skipping %p I_NOFLUSH set\n", inode);
2005 if (session && session != cap->session) {
2006 dout("oops, wrong session %p mutex\n", session);
2007 mutex_unlock(&session->s_mutex);
2011 session = cap->session;
2012 if (mutex_trylock(&session->s_mutex) == 0) {
2013 dout("inverting session/ino locks on %p\n",
2015 spin_unlock(&ci->i_ceph_lock);
2016 if (took_snap_rwsem) {
2017 up_read(&mdsc->snap_rwsem);
2018 took_snap_rwsem = 0;
2020 mutex_lock(&session->s_mutex);
2025 /* kick flushing and flush snaps before sending normal
2027 if (cap == ci->i_auth_cap &&
2029 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
2030 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2031 __kick_flushing_caps(mdsc, session, ci, 0);
2032 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2034 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
2035 __ceph_flush_snaps(ci, session);
2040 /* take snap_rwsem after session mutex */
2041 if (!took_snap_rwsem) {
2042 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
2043 dout("inverting snap/in locks on %p\n",
2045 spin_unlock(&ci->i_ceph_lock);
2046 down_read(&mdsc->snap_rwsem);
2047 took_snap_rwsem = 1;
2050 took_snap_rwsem = 1;
2053 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
2054 flushing = __mark_caps_flushing(inode, session, false,
2060 spin_lock(&mdsc->cap_dirty_lock);
2061 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2062 spin_unlock(&mdsc->cap_dirty_lock);
2065 mds = cap->mds; /* remember mds, so we don't repeat */
2068 /* __send_cap drops i_ceph_lock */
2069 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, false,
2070 cap_used, want, retain, flushing,
2071 flush_tid, oldest_flush_tid);
2072 goto retry; /* retake i_ceph_lock and restart our cap scan. */
2075 /* Reschedule delayed caps release if we delayed anything */
2077 __cap_delay_requeue(mdsc, ci, false);
2079 spin_unlock(&ci->i_ceph_lock);
2081 if (queue_invalidate)
2082 ceph_queue_invalidate(inode);
2085 mutex_unlock(&session->s_mutex);
2086 if (took_snap_rwsem)
2087 up_read(&mdsc->snap_rwsem);
2091 * Try to flush dirty caps back to the auth mds.
2093 static int try_flush_caps(struct inode *inode, u64 *ptid)
2095 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2096 struct ceph_inode_info *ci = ceph_inode(inode);
2097 struct ceph_mds_session *session = NULL;
2099 u64 flush_tid = 0, oldest_flush_tid = 0;
2102 spin_lock(&ci->i_ceph_lock);
2103 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
2104 spin_unlock(&ci->i_ceph_lock);
2105 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
2108 if (ci->i_dirty_caps && ci->i_auth_cap) {
2109 struct ceph_cap *cap = ci->i_auth_cap;
2110 int used = __ceph_caps_used(ci);
2111 int want = __ceph_caps_wanted(ci);
2114 if (!session || session != cap->session) {
2115 spin_unlock(&ci->i_ceph_lock);
2117 mutex_unlock(&session->s_mutex);
2118 session = cap->session;
2119 mutex_lock(&session->s_mutex);
2122 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) {
2123 spin_unlock(&ci->i_ceph_lock);
2127 flushing = __mark_caps_flushing(inode, session, true,
2128 &flush_tid, &oldest_flush_tid);
2130 /* __send_cap drops i_ceph_lock */
2131 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, true,
2132 used, want, (cap->issued | cap->implemented),
2133 flushing, flush_tid, oldest_flush_tid);
2136 spin_lock(&ci->i_ceph_lock);
2137 __cap_delay_requeue(mdsc, ci, true);
2138 spin_unlock(&ci->i_ceph_lock);
2141 if (!list_empty(&ci->i_cap_flush_list)) {
2142 struct ceph_cap_flush *cf =
2143 list_last_entry(&ci->i_cap_flush_list,
2144 struct ceph_cap_flush, i_list);
2146 flush_tid = cf->tid;
2148 flushing = ci->i_flushing_caps;
2149 spin_unlock(&ci->i_ceph_lock);
2153 mutex_unlock(&session->s_mutex);
2160 * Return true if we've flushed caps through the given flush_tid.
2162 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2164 struct ceph_inode_info *ci = ceph_inode(inode);
2167 spin_lock(&ci->i_ceph_lock);
2168 if (!list_empty(&ci->i_cap_flush_list)) {
2169 struct ceph_cap_flush * cf =
2170 list_first_entry(&ci->i_cap_flush_list,
2171 struct ceph_cap_flush, i_list);
2172 if (cf->tid <= flush_tid)
2175 spin_unlock(&ci->i_ceph_lock);
2180 * wait for any unsafe requests to complete.
2182 static int unsafe_request_wait(struct inode *inode)
2184 struct ceph_inode_info *ci = ceph_inode(inode);
2185 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2188 spin_lock(&ci->i_unsafe_lock);
2189 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2190 req1 = list_last_entry(&ci->i_unsafe_dirops,
2191 struct ceph_mds_request,
2193 ceph_mdsc_get_request(req1);
2195 if (!list_empty(&ci->i_unsafe_iops)) {
2196 req2 = list_last_entry(&ci->i_unsafe_iops,
2197 struct ceph_mds_request,
2198 r_unsafe_target_item);
2199 ceph_mdsc_get_request(req2);
2201 spin_unlock(&ci->i_unsafe_lock);
2203 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2204 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2206 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2207 ceph_timeout_jiffies(req1->r_timeout));
2210 ceph_mdsc_put_request(req1);
2213 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2214 ceph_timeout_jiffies(req2->r_timeout));
2217 ceph_mdsc_put_request(req2);
2222 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2224 struct inode *inode = file->f_mapping->host;
2225 struct ceph_inode_info *ci = ceph_inode(inode);
2230 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2232 ret = file_write_and_wait_range(file, start, end);
2241 dirty = try_flush_caps(inode, &flush_tid);
2242 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2244 ret = unsafe_request_wait(inode);
2247 * only wait on non-file metadata writeback (the mds
2248 * can recover size and mtime, so we don't need to
2251 if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2252 ret = wait_event_interruptible(ci->i_cap_wq,
2253 caps_are_flushed(inode, flush_tid));
2255 inode_unlock(inode);
2257 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2262 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2263 * queue inode for flush but don't do so immediately, because we can
2264 * get by with fewer MDS messages if we wait for data writeback to
2267 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2269 struct ceph_inode_info *ci = ceph_inode(inode);
2273 int wait = (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync);
2275 dout("write_inode %p wait=%d\n", inode, wait);
2277 dirty = try_flush_caps(inode, &flush_tid);
2279 err = wait_event_interruptible(ci->i_cap_wq,
2280 caps_are_flushed(inode, flush_tid));
2282 struct ceph_mds_client *mdsc =
2283 ceph_sb_to_client(inode->i_sb)->mdsc;
2285 spin_lock(&ci->i_ceph_lock);
2286 if (__ceph_caps_dirty(ci))
2287 __cap_delay_requeue_front(mdsc, ci);
2288 spin_unlock(&ci->i_ceph_lock);
2293 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2294 struct ceph_mds_session *session,
2295 struct ceph_inode_info *ci,
2296 u64 oldest_flush_tid)
2297 __releases(ci->i_ceph_lock)
2298 __acquires(ci->i_ceph_lock)
2300 struct inode *inode = &ci->vfs_inode;
2301 struct ceph_cap *cap;
2302 struct ceph_cap_flush *cf;
2306 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2307 if (cf->tid < first_tid)
2310 cap = ci->i_auth_cap;
2311 if (!(cap && cap->session == session)) {
2312 pr_err("%p auth cap %p not mds%d ???\n",
2313 inode, cap, session->s_mds);
2317 first_tid = cf->tid + 1;
2320 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2321 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2322 ci->i_ceph_flags |= CEPH_I_NODELAY;
2323 ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2324 false, __ceph_caps_used(ci),
2325 __ceph_caps_wanted(ci),
2326 cap->issued | cap->implemented,
2327 cf->caps, cf->tid, oldest_flush_tid);
2329 pr_err("kick_flushing_caps: error sending "
2330 "cap flush, ino (%llx.%llx) "
2331 "tid %llu flushing %s\n",
2332 ceph_vinop(inode), cf->tid,
2333 ceph_cap_string(cf->caps));
2336 struct ceph_cap_snap *capsnap =
2337 container_of(cf, struct ceph_cap_snap,
2339 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2340 inode, capsnap, cf->tid,
2341 ceph_cap_string(capsnap->dirty));
2343 refcount_inc(&capsnap->nref);
2344 spin_unlock(&ci->i_ceph_lock);
2346 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2349 pr_err("kick_flushing_caps: error sending "
2350 "cap flushsnap, ino (%llx.%llx) "
2351 "tid %llu follows %llu\n",
2352 ceph_vinop(inode), cf->tid,
2356 ceph_put_cap_snap(capsnap);
2359 spin_lock(&ci->i_ceph_lock);
2363 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2364 struct ceph_mds_session *session)
2366 struct ceph_inode_info *ci;
2367 struct ceph_cap *cap;
2368 u64 oldest_flush_tid;
2370 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2372 spin_lock(&mdsc->cap_dirty_lock);
2373 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2374 spin_unlock(&mdsc->cap_dirty_lock);
2376 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2377 spin_lock(&ci->i_ceph_lock);
2378 cap = ci->i_auth_cap;
2379 if (!(cap && cap->session == session)) {
2380 pr_err("%p auth cap %p not mds%d ???\n",
2381 &ci->vfs_inode, cap, session->s_mds);
2382 spin_unlock(&ci->i_ceph_lock);
2388 * if flushing caps were revoked, we re-send the cap flush
2389 * in client reconnect stage. This guarantees MDS * processes
2390 * the cap flush message before issuing the flushing caps to
2393 if ((cap->issued & ci->i_flushing_caps) !=
2394 ci->i_flushing_caps) {
2395 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2396 __kick_flushing_caps(mdsc, session, ci,
2399 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2402 spin_unlock(&ci->i_ceph_lock);
2406 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2407 struct ceph_mds_session *session)
2409 struct ceph_inode_info *ci;
2410 struct ceph_cap *cap;
2411 u64 oldest_flush_tid;
2413 dout("kick_flushing_caps mds%d\n", session->s_mds);
2415 spin_lock(&mdsc->cap_dirty_lock);
2416 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2417 spin_unlock(&mdsc->cap_dirty_lock);
2419 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2420 spin_lock(&ci->i_ceph_lock);
2421 cap = ci->i_auth_cap;
2422 if (!(cap && cap->session == session)) {
2423 pr_err("%p auth cap %p not mds%d ???\n",
2424 &ci->vfs_inode, cap, session->s_mds);
2425 spin_unlock(&ci->i_ceph_lock);
2428 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2429 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2430 __kick_flushing_caps(mdsc, session, ci,
2433 spin_unlock(&ci->i_ceph_lock);
2437 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2438 struct ceph_mds_session *session,
2439 struct inode *inode)
2440 __releases(ci->i_ceph_lock)
2442 struct ceph_inode_info *ci = ceph_inode(inode);
2443 struct ceph_cap *cap;
2445 cap = ci->i_auth_cap;
2446 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2447 ceph_cap_string(ci->i_flushing_caps));
2449 if (!list_empty(&ci->i_cap_flush_list)) {
2450 u64 oldest_flush_tid;
2451 spin_lock(&mdsc->cap_dirty_lock);
2452 list_move_tail(&ci->i_flushing_item,
2453 &cap->session->s_cap_flushing);
2454 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2455 spin_unlock(&mdsc->cap_dirty_lock);
2457 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2458 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2459 spin_unlock(&ci->i_ceph_lock);
2461 spin_unlock(&ci->i_ceph_lock);
2467 * Take references to capabilities we hold, so that we don't release
2468 * them to the MDS prematurely.
2470 * Protected by i_ceph_lock.
2472 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2473 bool snap_rwsem_locked)
2475 if (got & CEPH_CAP_PIN)
2477 if (got & CEPH_CAP_FILE_RD)
2479 if (got & CEPH_CAP_FILE_CACHE)
2480 ci->i_rdcache_ref++;
2481 if (got & CEPH_CAP_FILE_WR) {
2482 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2483 BUG_ON(!snap_rwsem_locked);
2484 ci->i_head_snapc = ceph_get_snap_context(
2485 ci->i_snap_realm->cached_context);
2489 if (got & CEPH_CAP_FILE_BUFFER) {
2490 if (ci->i_wb_ref == 0)
2491 ihold(&ci->vfs_inode);
2493 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2494 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2499 * Try to grab cap references. Specify those refs we @want, and the
2500 * minimal set we @need. Also include the larger offset we are writing
2501 * to (when applicable), and check against max_size here as well.
2502 * Note that caller is responsible for ensuring max_size increases are
2503 * requested from the MDS.
2505 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2506 loff_t endoff, bool nonblock, int *got, int *err)
2508 struct inode *inode = &ci->vfs_inode;
2509 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2511 int have, implemented;
2513 bool snap_rwsem_locked = false;
2515 dout("get_cap_refs %p need %s want %s\n", inode,
2516 ceph_cap_string(need), ceph_cap_string(want));
2519 spin_lock(&ci->i_ceph_lock);
2521 /* make sure file is actually open */
2522 file_wanted = __ceph_caps_file_wanted(ci);
2523 if ((file_wanted & need) != need) {
2524 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2525 ceph_cap_string(need), ceph_cap_string(file_wanted));
2531 /* finish pending truncate */
2532 while (ci->i_truncate_pending) {
2533 spin_unlock(&ci->i_ceph_lock);
2534 if (snap_rwsem_locked) {
2535 up_read(&mdsc->snap_rwsem);
2536 snap_rwsem_locked = false;
2538 __ceph_do_pending_vmtruncate(inode);
2539 spin_lock(&ci->i_ceph_lock);
2542 have = __ceph_caps_issued(ci, &implemented);
2544 if (have & need & CEPH_CAP_FILE_WR) {
2545 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2546 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2547 inode, endoff, ci->i_max_size);
2548 if (endoff > ci->i_requested_max_size) {
2555 * If a sync write is in progress, we must wait, so that we
2556 * can get a final snapshot value for size+mtime.
2558 if (__ceph_have_pending_cap_snap(ci)) {
2559 dout("get_cap_refs %p cap_snap_pending\n", inode);
2564 if ((have & need) == need) {
2566 * Look at (implemented & ~have & not) so that we keep waiting
2567 * on transition from wanted -> needed caps. This is needed
2568 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2569 * going before a prior buffered writeback happens.
2571 int not = want & ~(have & need);
2572 int revoking = implemented & ~have;
2573 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2574 inode, ceph_cap_string(have), ceph_cap_string(not),
2575 ceph_cap_string(revoking));
2576 if ((revoking & not) == 0) {
2577 if (!snap_rwsem_locked &&
2578 !ci->i_head_snapc &&
2579 (need & CEPH_CAP_FILE_WR)) {
2580 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2582 * we can not call down_read() when
2583 * task isn't in TASK_RUNNING state
2591 spin_unlock(&ci->i_ceph_lock);
2592 down_read(&mdsc->snap_rwsem);
2593 snap_rwsem_locked = true;
2596 snap_rwsem_locked = true;
2598 *got = need | (have & want);
2599 if ((need & CEPH_CAP_FILE_RD) &&
2600 !(*got & CEPH_CAP_FILE_CACHE))
2601 ceph_disable_fscache_readpage(ci);
2602 __take_cap_refs(ci, *got, true);
2606 int session_readonly = false;
2607 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2608 struct ceph_mds_session *s = ci->i_auth_cap->session;
2609 spin_lock(&s->s_cap_lock);
2610 session_readonly = s->s_readonly;
2611 spin_unlock(&s->s_cap_lock);
2613 if (session_readonly) {
2614 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2615 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2621 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) {
2623 if (READ_ONCE(mdsc->fsc->mount_state) ==
2624 CEPH_MOUNT_SHUTDOWN) {
2625 dout("get_cap_refs %p forced umount\n", inode);
2630 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2631 if (need & ~(mds_wanted & need)) {
2632 dout("get_cap_refs %p caps were dropped"
2633 " (session killed?)\n", inode);
2638 if (!(file_wanted & ~mds_wanted))
2639 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED;
2642 dout("get_cap_refs %p have %s needed %s\n", inode,
2643 ceph_cap_string(have), ceph_cap_string(need));
2646 spin_unlock(&ci->i_ceph_lock);
2647 if (snap_rwsem_locked)
2648 up_read(&mdsc->snap_rwsem);
2650 dout("get_cap_refs %p ret %d got %s\n", inode,
2651 ret, ceph_cap_string(*got));
2656 * Check the offset we are writing up to against our current
2657 * max_size. If necessary, tell the MDS we want to write to
2660 static void check_max_size(struct inode *inode, loff_t endoff)
2662 struct ceph_inode_info *ci = ceph_inode(inode);
2665 /* do we need to explicitly request a larger max_size? */
2666 spin_lock(&ci->i_ceph_lock);
2667 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2668 dout("write %p at large endoff %llu, req max_size\n",
2670 ci->i_wanted_max_size = endoff;
2672 /* duplicate ceph_check_caps()'s logic */
2673 if (ci->i_auth_cap &&
2674 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2675 ci->i_wanted_max_size > ci->i_max_size &&
2676 ci->i_wanted_max_size > ci->i_requested_max_size)
2678 spin_unlock(&ci->i_ceph_lock);
2680 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2683 int ceph_try_get_caps(struct ceph_inode_info *ci, int need, int want,
2684 bool nonblock, int *got)
2688 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2689 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO|CEPH_CAP_FILE_SHARED));
2690 ret = ceph_pool_perm_check(ci, need);
2694 ret = try_get_cap_refs(ci, need, want, 0, nonblock, got, &err);
2696 if (err == -EAGAIN) {
2698 } else if (err < 0) {
2706 * Wait for caps, and take cap references. If we can't get a WR cap
2707 * due to a small max_size, make sure we check_max_size (and possibly
2708 * ask the mds) so we don't get hung up indefinitely.
2710 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2711 loff_t endoff, int *got, struct page **pinned_page)
2713 int _got, ret, err = 0;
2715 ret = ceph_pool_perm_check(ci, need);
2721 check_max_size(&ci->vfs_inode, endoff);
2725 ret = try_get_cap_refs(ci, need, want, endoff,
2726 false, &_got, &err);
2733 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2734 add_wait_queue(&ci->i_cap_wq, &wait);
2736 while (!try_get_cap_refs(ci, need, want, endoff,
2737 true, &_got, &err)) {
2738 if (signal_pending(current)) {
2742 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2745 remove_wait_queue(&ci->i_cap_wq, &wait);
2753 if (err == -ESTALE) {
2754 /* session was killed, try renew caps */
2755 ret = ceph_renew_caps(&ci->vfs_inode);
2762 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2763 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2764 i_size_read(&ci->vfs_inode) > 0) {
2766 find_get_page(ci->vfs_inode.i_mapping, 0);
2768 if (PageUptodate(page)) {
2769 *pinned_page = page;
2775 * drop cap refs first because getattr while
2776 * holding * caps refs can cause deadlock.
2778 ceph_put_cap_refs(ci, _got);
2782 * getattr request will bring inline data into
2785 ret = __ceph_do_getattr(&ci->vfs_inode, NULL,
2786 CEPH_STAT_CAP_INLINE_DATA,
2795 if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
2796 ceph_fscache_revalidate_cookie(ci);
2803 * Take cap refs. Caller must already know we hold at least one ref
2804 * on the caps in question or we don't know this is safe.
2806 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2808 spin_lock(&ci->i_ceph_lock);
2809 __take_cap_refs(ci, caps, false);
2810 spin_unlock(&ci->i_ceph_lock);
2815 * drop cap_snap that is not associated with any snapshot.
2816 * we don't need to send FLUSHSNAP message for it.
2818 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2819 struct ceph_cap_snap *capsnap)
2821 if (!capsnap->need_flush &&
2822 !capsnap->writing && !capsnap->dirty_pages) {
2823 dout("dropping cap_snap %p follows %llu\n",
2824 capsnap, capsnap->follows);
2825 BUG_ON(capsnap->cap_flush.tid > 0);
2826 ceph_put_snap_context(capsnap->context);
2827 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2828 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2830 list_del(&capsnap->ci_item);
2831 ceph_put_cap_snap(capsnap);
2840 * If we released the last ref on any given cap, call ceph_check_caps
2841 * to release (or schedule a release).
2843 * If we are releasing a WR cap (from a sync write), finalize any affected
2844 * cap_snap, and wake up any waiters.
2846 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2848 struct inode *inode = &ci->vfs_inode;
2849 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2851 spin_lock(&ci->i_ceph_lock);
2852 if (had & CEPH_CAP_PIN)
2854 if (had & CEPH_CAP_FILE_RD)
2855 if (--ci->i_rd_ref == 0)
2857 if (had & CEPH_CAP_FILE_CACHE)
2858 if (--ci->i_rdcache_ref == 0)
2860 if (had & CEPH_CAP_FILE_BUFFER) {
2861 if (--ci->i_wb_ref == 0) {
2865 dout("put_cap_refs %p wb %d -> %d (?)\n",
2866 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2868 if (had & CEPH_CAP_FILE_WR)
2869 if (--ci->i_wr_ref == 0) {
2871 if (__ceph_have_pending_cap_snap(ci)) {
2872 struct ceph_cap_snap *capsnap =
2873 list_last_entry(&ci->i_cap_snaps,
2874 struct ceph_cap_snap,
2876 capsnap->writing = 0;
2877 if (ceph_try_drop_cap_snap(ci, capsnap))
2879 else if (__ceph_finish_cap_snap(ci, capsnap))
2883 if (ci->i_wrbuffer_ref_head == 0 &&
2884 ci->i_dirty_caps == 0 &&
2885 ci->i_flushing_caps == 0) {
2886 BUG_ON(!ci->i_head_snapc);
2887 ceph_put_snap_context(ci->i_head_snapc);
2888 ci->i_head_snapc = NULL;
2890 /* see comment in __ceph_remove_cap() */
2891 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2892 drop_inode_snap_realm(ci);
2894 spin_unlock(&ci->i_ceph_lock);
2896 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2897 last ? " last" : "", put ? " put" : "");
2899 if (last && !flushsnaps)
2900 ceph_check_caps(ci, 0, NULL);
2901 else if (flushsnaps)
2902 ceph_flush_snaps(ci, NULL);
2904 wake_up_all(&ci->i_cap_wq);
2910 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2911 * context. Adjust per-snap dirty page accounting as appropriate.
2912 * Once all dirty data for a cap_snap is flushed, flush snapped file
2913 * metadata back to the MDS. If we dropped the last ref, call
2916 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2917 struct ceph_snap_context *snapc)
2919 struct inode *inode = &ci->vfs_inode;
2920 struct ceph_cap_snap *capsnap = NULL;
2924 bool flush_snaps = false;
2925 bool complete_capsnap = false;
2927 spin_lock(&ci->i_ceph_lock);
2928 ci->i_wrbuffer_ref -= nr;
2929 if (ci->i_wrbuffer_ref == 0) {
2934 if (ci->i_head_snapc == snapc) {
2935 ci->i_wrbuffer_ref_head -= nr;
2936 if (ci->i_wrbuffer_ref_head == 0 &&
2937 ci->i_wr_ref == 0 &&
2938 ci->i_dirty_caps == 0 &&
2939 ci->i_flushing_caps == 0) {
2940 BUG_ON(!ci->i_head_snapc);
2941 ceph_put_snap_context(ci->i_head_snapc);
2942 ci->i_head_snapc = NULL;
2944 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2946 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2947 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2948 last ? " LAST" : "");
2950 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2951 if (capsnap->context == snapc) {
2957 capsnap->dirty_pages -= nr;
2958 if (capsnap->dirty_pages == 0) {
2959 complete_capsnap = true;
2960 if (!capsnap->writing) {
2961 if (ceph_try_drop_cap_snap(ci, capsnap)) {
2964 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2969 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2970 " snap %lld %d/%d -> %d/%d %s%s\n",
2971 inode, capsnap, capsnap->context->seq,
2972 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2973 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2974 last ? " (wrbuffer last)" : "",
2975 complete_capsnap ? " (complete capsnap)" : "");
2978 spin_unlock(&ci->i_ceph_lock);
2981 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2982 } else if (flush_snaps) {
2983 ceph_flush_snaps(ci, NULL);
2985 if (complete_capsnap)
2986 wake_up_all(&ci->i_cap_wq);
2992 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2994 static void invalidate_aliases(struct inode *inode)
2996 struct dentry *dn, *prev = NULL;
2998 dout("invalidate_aliases inode %p\n", inode);
2999 d_prune_aliases(inode);
3001 * For non-directory inode, d_find_alias() only returns
3002 * hashed dentry. After calling d_invalidate(), the
3003 * dentry becomes unhashed.
3005 * For directory inode, d_find_alias() can return
3006 * unhashed dentry. But directory inode should have
3007 * one alias at most.
3009 while ((dn = d_find_alias(inode))) {
3023 struct cap_extra_info {
3024 struct ceph_string *pool_ns;
3033 /* currently issued */
3038 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3039 * actually be a revocation if it specifies a smaller cap set.)
3041 * caller holds s_mutex and i_ceph_lock, we drop both.
3043 static void handle_cap_grant(struct inode *inode,
3044 struct ceph_mds_session *session,
3045 struct ceph_cap *cap,
3046 struct ceph_mds_caps *grant,
3047 struct ceph_buffer *xattr_buf,
3048 struct cap_extra_info *extra_info)
3049 __releases(ci->i_ceph_lock)
3050 __releases(session->s_mdsc->snap_rwsem)
3052 struct ceph_inode_info *ci = ceph_inode(inode);
3053 int seq = le32_to_cpu(grant->seq);
3054 int newcaps = le32_to_cpu(grant->caps);
3055 int used, wanted, dirty;
3056 u64 size = le64_to_cpu(grant->size);
3057 u64 max_size = le64_to_cpu(grant->max_size);
3058 unsigned char check_caps = 0;
3059 bool was_stale = cap->cap_gen < session->s_cap_gen;
3061 bool writeback = false;
3062 bool queue_trunc = false;
3063 bool queue_invalidate = false;
3064 bool deleted_inode = false;
3065 bool fill_inline = false;
3067 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3068 inode, cap, session->s_mds, seq, ceph_cap_string(newcaps));
3069 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
3074 * If CACHE is being revoked, and we have no dirty buffers,
3075 * try to invalidate (once). (If there are dirty buffers, we
3076 * will invalidate _after_ writeback.)
3078 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
3079 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
3080 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3081 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
3082 if (try_nonblocking_invalidate(inode)) {
3083 /* there were locked pages.. invalidate later
3084 in a separate thread. */
3085 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
3086 queue_invalidate = true;
3087 ci->i_rdcache_revoking = ci->i_rdcache_gen;
3093 cap->issued = cap->implemented = CEPH_CAP_PIN;
3096 * auth mds of the inode changed. we received the cap export message,
3097 * but still haven't received the cap import message. handle_cap_export
3098 * updated the new auth MDS' cap.
3100 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3101 * that was sent before the cap import message. So don't remove caps.
3103 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
3104 WARN_ON(cap != ci->i_auth_cap);
3105 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
3107 newcaps |= cap->issued;
3110 /* side effects now are allowed */
3111 cap->cap_gen = session->s_cap_gen;
3114 __check_cap_issue(ci, cap, newcaps);
3116 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
3117 (extra_info->issued & CEPH_CAP_AUTH_EXCL) == 0) {
3118 inode->i_mode = le32_to_cpu(grant->mode);
3119 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
3120 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
3121 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
3122 from_kuid(&init_user_ns, inode->i_uid),
3123 from_kgid(&init_user_ns, inode->i_gid));
3126 if ((newcaps & CEPH_CAP_LINK_SHARED) &&
3127 (extra_info->issued & CEPH_CAP_LINK_EXCL) == 0) {
3128 set_nlink(inode, le32_to_cpu(grant->nlink));
3129 if (inode->i_nlink == 0 &&
3130 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3131 deleted_inode = true;
3134 if ((extra_info->issued & CEPH_CAP_XATTR_EXCL) == 0 &&
3136 int len = le32_to_cpu(grant->xattr_len);
3137 u64 version = le64_to_cpu(grant->xattr_version);
3139 if (version > ci->i_xattrs.version) {
3140 dout(" got new xattrs v%llu on %p len %d\n",
3141 version, inode, len);
3142 if (ci->i_xattrs.blob)
3143 ceph_buffer_put(ci->i_xattrs.blob);
3144 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3145 ci->i_xattrs.version = version;
3146 ceph_forget_all_cached_acls(inode);
3150 if (newcaps & CEPH_CAP_ANY_RD) {
3151 struct timespec64 mtime, atime, ctime;
3152 /* ctime/mtime/atime? */
3153 ceph_decode_timespec64(&mtime, &grant->mtime);
3154 ceph_decode_timespec64(&atime, &grant->atime);
3155 ceph_decode_timespec64(&ctime, &grant->ctime);
3156 ceph_fill_file_time(inode, extra_info->issued,
3157 le32_to_cpu(grant->time_warp_seq),
3158 &ctime, &mtime, &atime);
3161 if ((newcaps & CEPH_CAP_FILE_SHARED) && extra_info->dirstat_valid) {
3162 ci->i_files = extra_info->nfiles;
3163 ci->i_subdirs = extra_info->nsubdirs;
3166 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3167 /* file layout may have changed */
3168 s64 old_pool = ci->i_layout.pool_id;
3169 struct ceph_string *old_ns;
3171 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3172 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3173 lockdep_is_held(&ci->i_ceph_lock));
3174 rcu_assign_pointer(ci->i_layout.pool_ns, extra_info->pool_ns);
3176 if (ci->i_layout.pool_id != old_pool ||
3177 extra_info->pool_ns != old_ns)
3178 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3180 extra_info->pool_ns = old_ns;
3182 /* size/truncate_seq? */
3183 queue_trunc = ceph_fill_file_size(inode, extra_info->issued,
3184 le32_to_cpu(grant->truncate_seq),
3185 le64_to_cpu(grant->truncate_size),
3189 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3190 if (max_size != ci->i_max_size) {
3191 dout("max_size %lld -> %llu\n",
3192 ci->i_max_size, max_size);
3193 ci->i_max_size = max_size;
3194 if (max_size >= ci->i_wanted_max_size) {
3195 ci->i_wanted_max_size = 0; /* reset */
3196 ci->i_requested_max_size = 0;
3199 } else if (ci->i_wanted_max_size > ci->i_max_size &&
3200 ci->i_wanted_max_size > ci->i_requested_max_size) {
3201 /* CEPH_CAP_OP_IMPORT */
3206 /* check cap bits */
3207 wanted = __ceph_caps_wanted(ci);
3208 used = __ceph_caps_used(ci);
3209 dirty = __ceph_caps_dirty(ci);
3210 dout(" my wanted = %s, used = %s, dirty %s\n",
3211 ceph_cap_string(wanted),
3212 ceph_cap_string(used),
3213 ceph_cap_string(dirty));
3215 if ((was_stale || le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) &&
3216 (wanted & ~(cap->mds_wanted | newcaps))) {
3218 * If mds is importing cap, prior cap messages that update
3219 * 'wanted' may get dropped by mds (migrate seq mismatch).
3221 * We don't send cap message to update 'wanted' if what we
3222 * want are already issued. If mds revokes caps, cap message
3223 * that releases caps also tells mds what we want. But if
3224 * caps got revoked by mds forcedly (session stale). We may
3225 * haven't told mds what we want.
3230 /* revocation, grant, or no-op? */
3231 if (cap->issued & ~newcaps) {
3232 int revoking = cap->issued & ~newcaps;
3234 dout("revocation: %s -> %s (revoking %s)\n",
3235 ceph_cap_string(cap->issued),
3236 ceph_cap_string(newcaps),
3237 ceph_cap_string(revoking));
3238 if (revoking & used & CEPH_CAP_FILE_BUFFER)
3239 writeback = true; /* initiate writeback; will delay ack */
3240 else if (revoking == CEPH_CAP_FILE_CACHE &&
3241 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3243 ; /* do nothing yet, invalidation will be queued */
3244 else if (cap == ci->i_auth_cap)
3245 check_caps = 1; /* check auth cap only */
3247 check_caps = 2; /* check all caps */
3248 cap->issued = newcaps;
3249 cap->implemented |= newcaps;
3250 } else if (cap->issued == newcaps) {
3251 dout("caps unchanged: %s -> %s\n",
3252 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3254 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3255 ceph_cap_string(newcaps));
3256 /* non-auth MDS is revoking the newly grant caps ? */
3257 if (cap == ci->i_auth_cap &&
3258 __ceph_caps_revoking_other(ci, cap, newcaps))
3261 cap->issued = newcaps;
3262 cap->implemented |= newcaps; /* add bits only, to
3263 * avoid stepping on a
3264 * pending revocation */
3267 BUG_ON(cap->issued & ~cap->implemented);
3269 if (extra_info->inline_version > 0 &&
3270 extra_info->inline_version >= ci->i_inline_version) {
3271 ci->i_inline_version = extra_info->inline_version;
3272 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3273 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3277 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3278 if (newcaps & ~extra_info->issued)
3280 kick_flushing_inode_caps(session->s_mdsc, session, inode);
3281 up_read(&session->s_mdsc->snap_rwsem);
3283 spin_unlock(&ci->i_ceph_lock);
3287 ceph_fill_inline_data(inode, NULL, extra_info->inline_data,
3288 extra_info->inline_len);
3291 ceph_queue_vmtruncate(inode);
3295 * queue inode for writeback: we can't actually call
3296 * filemap_write_and_wait, etc. from message handler
3299 ceph_queue_writeback(inode);
3300 if (queue_invalidate)
3301 ceph_queue_invalidate(inode);
3303 invalidate_aliases(inode);
3305 wake_up_all(&ci->i_cap_wq);
3307 if (check_caps == 1)
3308 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
3310 else if (check_caps == 2)
3311 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
3313 mutex_unlock(&session->s_mutex);
3317 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3318 * MDS has been safely committed.
3320 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3321 struct ceph_mds_caps *m,
3322 struct ceph_mds_session *session,
3323 struct ceph_cap *cap)
3324 __releases(ci->i_ceph_lock)
3326 struct ceph_inode_info *ci = ceph_inode(inode);
3327 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3328 struct ceph_cap_flush *cf, *tmp_cf;
3329 LIST_HEAD(to_remove);
3330 unsigned seq = le32_to_cpu(m->seq);
3331 int dirty = le32_to_cpu(m->dirty);
3334 bool wake_ci = false;
3335 bool wake_mdsc = false;
3337 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3338 if (cf->tid == flush_tid)
3340 if (cf->caps == 0) /* capsnap */
3342 if (cf->tid <= flush_tid) {
3343 if (__finish_cap_flush(NULL, ci, cf))
3345 list_add_tail(&cf->i_list, &to_remove);
3347 cleaned &= ~cf->caps;
3353 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3354 " flushing %s -> %s\n",
3355 inode, session->s_mds, seq, ceph_cap_string(dirty),
3356 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3357 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3359 if (list_empty(&to_remove) && !cleaned)
3362 ci->i_flushing_caps &= ~cleaned;
3364 spin_lock(&mdsc->cap_dirty_lock);
3366 list_for_each_entry(cf, &to_remove, i_list) {
3367 if (__finish_cap_flush(mdsc, NULL, cf))
3371 if (ci->i_flushing_caps == 0) {
3372 if (list_empty(&ci->i_cap_flush_list)) {
3373 list_del_init(&ci->i_flushing_item);
3374 if (!list_empty(&session->s_cap_flushing)) {
3375 dout(" mds%d still flushing cap on %p\n",
3377 &list_first_entry(&session->s_cap_flushing,
3378 struct ceph_inode_info,
3379 i_flushing_item)->vfs_inode);
3382 mdsc->num_cap_flushing--;
3383 dout(" inode %p now !flushing\n", inode);
3385 if (ci->i_dirty_caps == 0) {
3386 dout(" inode %p now clean\n", inode);
3387 BUG_ON(!list_empty(&ci->i_dirty_item));
3389 if (ci->i_wr_ref == 0 &&
3390 ci->i_wrbuffer_ref_head == 0) {
3391 BUG_ON(!ci->i_head_snapc);
3392 ceph_put_snap_context(ci->i_head_snapc);
3393 ci->i_head_snapc = NULL;
3396 BUG_ON(list_empty(&ci->i_dirty_item));
3399 spin_unlock(&mdsc->cap_dirty_lock);
3402 spin_unlock(&ci->i_ceph_lock);
3404 while (!list_empty(&to_remove)) {
3405 cf = list_first_entry(&to_remove,
3406 struct ceph_cap_flush, i_list);
3407 list_del(&cf->i_list);
3408 ceph_free_cap_flush(cf);
3412 wake_up_all(&ci->i_cap_wq);
3414 wake_up_all(&mdsc->cap_flushing_wq);
3420 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3421 * throw away our cap_snap.
3423 * Caller hold s_mutex.
3425 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3426 struct ceph_mds_caps *m,
3427 struct ceph_mds_session *session)
3429 struct ceph_inode_info *ci = ceph_inode(inode);
3430 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3431 u64 follows = le64_to_cpu(m->snap_follows);
3432 struct ceph_cap_snap *capsnap;
3433 bool flushed = false;
3434 bool wake_ci = false;
3435 bool wake_mdsc = false;
3437 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3438 inode, ci, session->s_mds, follows);
3440 spin_lock(&ci->i_ceph_lock);
3441 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3442 if (capsnap->follows == follows) {
3443 if (capsnap->cap_flush.tid != flush_tid) {
3444 dout(" cap_snap %p follows %lld tid %lld !="
3445 " %lld\n", capsnap, follows,
3446 flush_tid, capsnap->cap_flush.tid);
3452 dout(" skipping cap_snap %p follows %lld\n",
3453 capsnap, capsnap->follows);
3457 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3458 dout(" removing %p cap_snap %p follows %lld\n",
3459 inode, capsnap, follows);
3460 list_del(&capsnap->ci_item);
3461 if (__finish_cap_flush(NULL, ci, &capsnap->cap_flush))
3464 spin_lock(&mdsc->cap_dirty_lock);
3466 if (list_empty(&ci->i_cap_flush_list))
3467 list_del_init(&ci->i_flushing_item);
3469 if (__finish_cap_flush(mdsc, NULL, &capsnap->cap_flush))
3472 spin_unlock(&mdsc->cap_dirty_lock);
3474 spin_unlock(&ci->i_ceph_lock);
3476 ceph_put_snap_context(capsnap->context);
3477 ceph_put_cap_snap(capsnap);
3479 wake_up_all(&ci->i_cap_wq);
3481 wake_up_all(&mdsc->cap_flushing_wq);
3487 * Handle TRUNC from MDS, indicating file truncation.
3489 * caller hold s_mutex.
3491 static void handle_cap_trunc(struct inode *inode,
3492 struct ceph_mds_caps *trunc,
3493 struct ceph_mds_session *session)
3494 __releases(ci->i_ceph_lock)
3496 struct ceph_inode_info *ci = ceph_inode(inode);
3497 int mds = session->s_mds;
3498 int seq = le32_to_cpu(trunc->seq);
3499 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3500 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3501 u64 size = le64_to_cpu(trunc->size);
3502 int implemented = 0;
3503 int dirty = __ceph_caps_dirty(ci);
3504 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3505 int queue_trunc = 0;
3507 issued |= implemented | dirty;
3509 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3510 inode, mds, seq, truncate_size, truncate_seq);
3511 queue_trunc = ceph_fill_file_size(inode, issued,
3512 truncate_seq, truncate_size, size);
3513 spin_unlock(&ci->i_ceph_lock);
3516 ceph_queue_vmtruncate(inode);
3520 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3521 * different one. If we are the most recent migration we've seen (as
3522 * indicated by mseq), make note of the migrating cap bits for the
3523 * duration (until we see the corresponding IMPORT).
3525 * caller holds s_mutex
3527 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3528 struct ceph_mds_cap_peer *ph,
3529 struct ceph_mds_session *session)
3531 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3532 struct ceph_mds_session *tsession = NULL;
3533 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3534 struct ceph_inode_info *ci = ceph_inode(inode);
3536 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3537 unsigned t_seq, t_mseq;
3539 int mds = session->s_mds;
3542 t_cap_id = le64_to_cpu(ph->cap_id);
3543 t_seq = le32_to_cpu(ph->seq);
3544 t_mseq = le32_to_cpu(ph->mseq);
3545 target = le32_to_cpu(ph->mds);
3547 t_cap_id = t_seq = t_mseq = 0;
3551 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3552 inode, ci, mds, mseq, target);
3554 spin_lock(&ci->i_ceph_lock);
3555 cap = __get_cap_for_mds(ci, mds);
3556 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3560 if (cap->mds_wanted | cap->issued)
3561 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
3562 __ceph_remove_cap(cap, false);
3567 * now we know we haven't received the cap import message yet
3568 * because the exported cap still exist.
3571 issued = cap->issued;
3572 if (issued != cap->implemented)
3573 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3574 "ino (%llx.%llx) mds%d seq %d mseq %d "
3575 "issued %s implemented %s\n",
3576 ceph_vinop(inode), mds, cap->seq, cap->mseq,
3577 ceph_cap_string(issued),
3578 ceph_cap_string(cap->implemented));
3581 tcap = __get_cap_for_mds(ci, target);
3583 /* already have caps from the target */
3584 if (tcap->cap_id == t_cap_id &&
3585 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3586 dout(" updating import cap %p mds%d\n", tcap, target);
3587 tcap->cap_id = t_cap_id;
3588 tcap->seq = t_seq - 1;
3589 tcap->issue_seq = t_seq - 1;
3590 tcap->issued |= issued;
3591 tcap->implemented |= issued;
3592 if (cap == ci->i_auth_cap)
3593 ci->i_auth_cap = tcap;
3595 if (!list_empty(&ci->i_cap_flush_list) &&
3596 ci->i_auth_cap == tcap) {
3597 spin_lock(&mdsc->cap_dirty_lock);
3598 list_move_tail(&ci->i_flushing_item,
3599 &tcap->session->s_cap_flushing);
3600 spin_unlock(&mdsc->cap_dirty_lock);
3603 __ceph_remove_cap(cap, false);
3605 } else if (tsession) {
3606 /* add placeholder for the export tagert */
3607 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3609 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3610 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3612 if (!list_empty(&ci->i_cap_flush_list) &&
3613 ci->i_auth_cap == tcap) {
3614 spin_lock(&mdsc->cap_dirty_lock);
3615 list_move_tail(&ci->i_flushing_item,
3616 &tcap->session->s_cap_flushing);
3617 spin_unlock(&mdsc->cap_dirty_lock);
3620 __ceph_remove_cap(cap, false);
3624 spin_unlock(&ci->i_ceph_lock);
3625 mutex_unlock(&session->s_mutex);
3627 /* open target session */
3628 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3629 if (!IS_ERR(tsession)) {
3631 mutex_lock(&session->s_mutex);
3632 mutex_lock_nested(&tsession->s_mutex,
3633 SINGLE_DEPTH_NESTING);
3635 mutex_lock(&tsession->s_mutex);
3636 mutex_lock_nested(&session->s_mutex,
3637 SINGLE_DEPTH_NESTING);
3639 new_cap = ceph_get_cap(mdsc, NULL);
3648 spin_unlock(&ci->i_ceph_lock);
3649 mutex_unlock(&session->s_mutex);
3651 mutex_unlock(&tsession->s_mutex);
3652 ceph_put_mds_session(tsession);
3655 ceph_put_cap(mdsc, new_cap);
3659 * Handle cap IMPORT.
3661 * caller holds s_mutex. acquires i_ceph_lock
3663 static void handle_cap_import(struct ceph_mds_client *mdsc,
3664 struct inode *inode, struct ceph_mds_caps *im,
3665 struct ceph_mds_cap_peer *ph,
3666 struct ceph_mds_session *session,
3667 struct ceph_cap **target_cap, int *old_issued)
3668 __acquires(ci->i_ceph_lock)
3670 struct ceph_inode_info *ci = ceph_inode(inode);
3671 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3672 int mds = session->s_mds;
3674 unsigned caps = le32_to_cpu(im->caps);
3675 unsigned wanted = le32_to_cpu(im->wanted);
3676 unsigned seq = le32_to_cpu(im->seq);
3677 unsigned mseq = le32_to_cpu(im->migrate_seq);
3678 u64 realmino = le64_to_cpu(im->realm);
3679 u64 cap_id = le64_to_cpu(im->cap_id);
3684 p_cap_id = le64_to_cpu(ph->cap_id);
3685 peer = le32_to_cpu(ph->mds);
3691 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3692 inode, ci, mds, mseq, peer);
3695 spin_lock(&ci->i_ceph_lock);
3696 cap = __get_cap_for_mds(ci, mds);
3699 spin_unlock(&ci->i_ceph_lock);
3700 new_cap = ceph_get_cap(mdsc, NULL);
3706 ceph_put_cap(mdsc, new_cap);
3711 __ceph_caps_issued(ci, &issued);
3712 issued |= __ceph_caps_dirty(ci);
3714 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3715 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3717 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3718 if (ocap && ocap->cap_id == p_cap_id) {
3719 dout(" remove export cap %p mds%d flags %d\n",
3720 ocap, peer, ph->flags);
3721 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3722 (ocap->seq != le32_to_cpu(ph->seq) ||
3723 ocap->mseq != le32_to_cpu(ph->mseq))) {
3724 pr_err_ratelimited("handle_cap_import: "
3725 "mismatched seq/mseq: ino (%llx.%llx) "
3726 "mds%d seq %d mseq %d importer mds%d "
3727 "has peer seq %d mseq %d\n",
3728 ceph_vinop(inode), peer, ocap->seq,
3729 ocap->mseq, mds, le32_to_cpu(ph->seq),
3730 le32_to_cpu(ph->mseq));
3732 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3735 /* make sure we re-request max_size, if necessary */
3736 ci->i_requested_max_size = 0;
3738 *old_issued = issued;
3743 * Handle a caps message from the MDS.
3745 * Identify the appropriate session, inode, and call the right handler
3746 * based on the cap op.
3748 void ceph_handle_caps(struct ceph_mds_session *session,
3749 struct ceph_msg *msg)
3751 struct ceph_mds_client *mdsc = session->s_mdsc;
3752 struct inode *inode;
3753 struct ceph_inode_info *ci;
3754 struct ceph_cap *cap;
3755 struct ceph_mds_caps *h;
3756 struct ceph_mds_cap_peer *peer = NULL;
3757 struct ceph_snap_realm *realm = NULL;
3759 int msg_version = le16_to_cpu(msg->hdr.version);
3761 struct ceph_vino vino;
3763 size_t snaptrace_len;
3765 struct cap_extra_info extra_info = {};
3767 dout("handle_caps from mds%d\n", session->s_mds);
3770 end = msg->front.iov_base + msg->front.iov_len;
3771 if (msg->front.iov_len < sizeof(*h))
3773 h = msg->front.iov_base;
3774 op = le32_to_cpu(h->op);
3775 vino.ino = le64_to_cpu(h->ino);
3776 vino.snap = CEPH_NOSNAP;
3777 seq = le32_to_cpu(h->seq);
3778 mseq = le32_to_cpu(h->migrate_seq);
3781 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3782 p = snaptrace + snaptrace_len;
3784 if (msg_version >= 2) {
3786 ceph_decode_32_safe(&p, end, flock_len, bad);
3787 if (p + flock_len > end)
3792 if (msg_version >= 3) {
3793 if (op == CEPH_CAP_OP_IMPORT) {
3794 if (p + sizeof(*peer) > end)
3798 } else if (op == CEPH_CAP_OP_EXPORT) {
3799 /* recorded in unused fields */
3800 peer = (void *)&h->size;
3804 if (msg_version >= 4) {
3805 ceph_decode_64_safe(&p, end, extra_info.inline_version, bad);
3806 ceph_decode_32_safe(&p, end, extra_info.inline_len, bad);
3807 if (p + extra_info.inline_len > end)
3809 extra_info.inline_data = p;
3810 p += extra_info.inline_len;
3813 if (msg_version >= 5) {
3814 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
3817 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
3818 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
3821 if (msg_version >= 8) {
3823 u32 caller_uid, caller_gid;
3827 ceph_decode_64_safe(&p, end, flush_tid, bad);
3829 ceph_decode_32_safe(&p, end, caller_uid, bad);
3830 ceph_decode_32_safe(&p, end, caller_gid, bad);
3832 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
3833 if (pool_ns_len > 0) {
3834 ceph_decode_need(&p, end, pool_ns_len, bad);
3835 extra_info.pool_ns =
3836 ceph_find_or_create_string(p, pool_ns_len);
3841 if (msg_version >= 11) {
3842 struct ceph_timespec *btime;
3847 if (p + sizeof(*btime) > end)
3850 p += sizeof(*btime);
3851 ceph_decode_64_safe(&p, end, change_attr, bad);
3853 ceph_decode_32_safe(&p, end, flags, bad);
3855 extra_info.dirstat_valid = true;
3856 ceph_decode_64_safe(&p, end, extra_info.nfiles, bad);
3857 ceph_decode_64_safe(&p, end, extra_info.nsubdirs, bad);
3861 inode = ceph_find_inode(mdsc->fsc->sb, vino);
3862 ci = ceph_inode(inode);
3863 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3866 mutex_lock(&session->s_mutex);
3868 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3872 dout(" i don't have ino %llx\n", vino.ino);
3874 if (op == CEPH_CAP_OP_IMPORT) {
3875 cap = ceph_get_cap(mdsc, NULL);
3876 cap->cap_ino = vino.ino;
3877 cap->queue_release = 1;
3878 cap->cap_id = le64_to_cpu(h->cap_id);
3881 cap->issue_seq = seq;
3882 spin_lock(&session->s_cap_lock);
3883 list_add_tail(&cap->session_caps,
3884 &session->s_cap_releases);
3885 session->s_num_cap_releases++;
3886 spin_unlock(&session->s_cap_lock);
3888 goto flush_cap_releases;
3891 /* these will work even if we don't have a cap yet */
3893 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3894 handle_cap_flushsnap_ack(inode, le64_to_cpu(msg->hdr.tid),
3898 case CEPH_CAP_OP_EXPORT:
3899 handle_cap_export(inode, h, peer, session);
3902 case CEPH_CAP_OP_IMPORT:
3904 if (snaptrace_len) {
3905 down_write(&mdsc->snap_rwsem);
3906 ceph_update_snap_trace(mdsc, snaptrace,
3907 snaptrace + snaptrace_len,
3909 downgrade_write(&mdsc->snap_rwsem);
3911 down_read(&mdsc->snap_rwsem);
3913 handle_cap_import(mdsc, inode, h, peer, session,
3914 &cap, &extra_info.issued);
3915 handle_cap_grant(inode, session, cap,
3916 h, msg->middle, &extra_info);
3918 ceph_put_snap_realm(mdsc, realm);
3922 /* the rest require a cap */
3923 spin_lock(&ci->i_ceph_lock);
3924 cap = __get_cap_for_mds(ceph_inode(inode), session->s_mds);
3926 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3927 inode, ceph_ino(inode), ceph_snap(inode),
3929 spin_unlock(&ci->i_ceph_lock);
3930 goto flush_cap_releases;
3933 /* note that each of these drops i_ceph_lock for us */
3935 case CEPH_CAP_OP_REVOKE:
3936 case CEPH_CAP_OP_GRANT:
3937 __ceph_caps_issued(ci, &extra_info.issued);
3938 extra_info.issued |= __ceph_caps_dirty(ci);
3939 handle_cap_grant(inode, session, cap,
3940 h, msg->middle, &extra_info);
3943 case CEPH_CAP_OP_FLUSH_ACK:
3944 handle_cap_flush_ack(inode, le64_to_cpu(msg->hdr.tid),
3948 case CEPH_CAP_OP_TRUNC:
3949 handle_cap_trunc(inode, h, session);
3953 spin_unlock(&ci->i_ceph_lock);
3954 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3955 ceph_cap_op_name(op));
3962 * send any cap release message to try to move things
3963 * along for the mds (who clearly thinks we still have this
3966 ceph_send_cap_releases(mdsc, session);
3969 mutex_unlock(&session->s_mutex);
3972 ceph_put_string(extra_info.pool_ns);
3976 pr_err("ceph_handle_caps: corrupt message\n");
3982 * Delayed work handler to process end of delayed cap release LRU list.
3984 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3986 struct inode *inode;
3987 struct ceph_inode_info *ci;
3988 int flags = CHECK_CAPS_NODELAY;
3990 dout("check_delayed_caps\n");
3992 spin_lock(&mdsc->cap_delay_lock);
3993 if (list_empty(&mdsc->cap_delay_list))
3995 ci = list_first_entry(&mdsc->cap_delay_list,
3996 struct ceph_inode_info,
3998 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3999 time_before(jiffies, ci->i_hold_caps_max))
4001 list_del_init(&ci->i_cap_delay_list);
4003 inode = igrab(&ci->vfs_inode);
4004 spin_unlock(&mdsc->cap_delay_lock);
4007 dout("check_delayed_caps on %p\n", inode);
4008 ceph_check_caps(ci, flags, NULL);
4012 spin_unlock(&mdsc->cap_delay_lock);
4016 * Flush all dirty caps to the mds
4018 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
4020 struct ceph_inode_info *ci;
4021 struct inode *inode;
4023 dout("flush_dirty_caps\n");
4024 spin_lock(&mdsc->cap_dirty_lock);
4025 while (!list_empty(&mdsc->cap_dirty)) {
4026 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
4028 inode = &ci->vfs_inode;
4030 dout("flush_dirty_caps %p\n", inode);
4031 spin_unlock(&mdsc->cap_dirty_lock);
4032 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
4034 spin_lock(&mdsc->cap_dirty_lock);
4036 spin_unlock(&mdsc->cap_dirty_lock);
4037 dout("flush_dirty_caps done\n");
4040 void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode)
4043 int bits = (fmode << 1) | 1;
4044 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4045 if (bits & (1 << i))
4046 ci->i_nr_by_mode[i]++;
4051 * Drop open file reference. If we were the last open file,
4052 * we may need to release capabilities to the MDS (or schedule
4053 * their delayed release).
4055 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
4058 int bits = (fmode << 1) | 1;
4059 spin_lock(&ci->i_ceph_lock);
4060 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
4061 if (bits & (1 << i)) {
4062 BUG_ON(ci->i_nr_by_mode[i] == 0);
4063 if (--ci->i_nr_by_mode[i] == 0)
4067 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
4068 &ci->vfs_inode, fmode,
4069 ci->i_nr_by_mode[0], ci->i_nr_by_mode[1],
4070 ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]);
4071 spin_unlock(&ci->i_ceph_lock);
4073 if (last && ci->i_vino.snap == CEPH_NOSNAP)
4074 ceph_check_caps(ci, 0, NULL);
4078 * For a soon-to-be unlinked file, drop the AUTH_RDCACHE caps. If it
4079 * looks like the link count will hit 0, drop any other caps (other
4080 * than PIN) we don't specifically want (due to the file still being
4083 int ceph_drop_caps_for_unlink(struct inode *inode)
4085 struct ceph_inode_info *ci = ceph_inode(inode);
4086 int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
4088 spin_lock(&ci->i_ceph_lock);
4089 if (inode->i_nlink == 1) {
4090 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
4092 ci->i_ceph_flags |= CEPH_I_NODELAY;
4093 if (__ceph_caps_dirty(ci)) {
4094 struct ceph_mds_client *mdsc =
4095 ceph_inode_to_client(inode)->mdsc;
4096 __cap_delay_requeue_front(mdsc, ci);
4099 spin_unlock(&ci->i_ceph_lock);
4104 * Helpers for embedding cap and dentry lease releases into mds
4107 * @force is used by dentry_release (below) to force inclusion of a
4108 * record for the directory inode, even when there aren't any caps to
4111 int ceph_encode_inode_release(void **p, struct inode *inode,
4112 int mds, int drop, int unless, int force)
4114 struct ceph_inode_info *ci = ceph_inode(inode);
4115 struct ceph_cap *cap;
4116 struct ceph_mds_request_release *rel = *p;
4120 spin_lock(&ci->i_ceph_lock);
4121 used = __ceph_caps_used(ci);
4122 dirty = __ceph_caps_dirty(ci);
4124 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4125 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
4126 ceph_cap_string(unless));
4128 /* only drop unused, clean caps */
4129 drop &= ~(used | dirty);
4131 cap = __get_cap_for_mds(ci, mds);
4132 if (cap && __cap_is_valid(cap)) {
4133 unless &= cap->issued;
4135 if (unless & CEPH_CAP_AUTH_EXCL)
4136 drop &= ~CEPH_CAP_AUTH_SHARED;
4137 if (unless & CEPH_CAP_LINK_EXCL)
4138 drop &= ~CEPH_CAP_LINK_SHARED;
4139 if (unless & CEPH_CAP_XATTR_EXCL)
4140 drop &= ~CEPH_CAP_XATTR_SHARED;
4141 if (unless & CEPH_CAP_FILE_EXCL)
4142 drop &= ~CEPH_CAP_FILE_SHARED;
4145 if (force || (cap->issued & drop)) {
4146 if (cap->issued & drop) {
4147 int wanted = __ceph_caps_wanted(ci);
4148 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
4149 wanted |= cap->mds_wanted;
4150 dout("encode_inode_release %p cap %p "
4151 "%s -> %s, wanted %s -> %s\n", inode, cap,
4152 ceph_cap_string(cap->issued),
4153 ceph_cap_string(cap->issued & ~drop),
4154 ceph_cap_string(cap->mds_wanted),
4155 ceph_cap_string(wanted));
4157 cap->issued &= ~drop;
4158 cap->implemented &= ~drop;
4159 cap->mds_wanted = wanted;
4161 dout("encode_inode_release %p cap %p %s"
4162 " (force)\n", inode, cap,
4163 ceph_cap_string(cap->issued));
4166 rel->ino = cpu_to_le64(ceph_ino(inode));
4167 rel->cap_id = cpu_to_le64(cap->cap_id);
4168 rel->seq = cpu_to_le32(cap->seq);
4169 rel->issue_seq = cpu_to_le32(cap->issue_seq);
4170 rel->mseq = cpu_to_le32(cap->mseq);
4171 rel->caps = cpu_to_le32(cap->implemented);
4172 rel->wanted = cpu_to_le32(cap->mds_wanted);
4178 dout("encode_inode_release %p cap %p %s (noop)\n",
4179 inode, cap, ceph_cap_string(cap->issued));
4182 spin_unlock(&ci->i_ceph_lock);
4186 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
4188 int mds, int drop, int unless)
4190 struct dentry *parent = NULL;
4191 struct ceph_mds_request_release *rel = *p;
4192 struct ceph_dentry_info *di = ceph_dentry(dentry);
4197 * force an record for the directory caps if we have a dentry lease.
4198 * this is racy (can't take i_ceph_lock and d_lock together), but it
4199 * doesn't have to be perfect; the mds will revoke anything we don't
4202 spin_lock(&dentry->d_lock);
4203 if (di->lease_session && di->lease_session->s_mds == mds)
4206 parent = dget(dentry->d_parent);
4207 dir = d_inode(parent);
4209 spin_unlock(&dentry->d_lock);
4211 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4214 spin_lock(&dentry->d_lock);
4215 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4216 dout("encode_dentry_release %p mds%d seq %d\n",
4217 dentry, mds, (int)di->lease_seq);
4218 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4219 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4220 *p += dentry->d_name.len;
4221 rel->dname_seq = cpu_to_le32(di->lease_seq);
4222 __ceph_mdsc_drop_dentry_lease(dentry);
4224 spin_unlock(&dentry->d_lock);