1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/kernel.h>
5 #include <linux/sched.h>
6 #include <linux/slab.h>
7 #include <linux/vmalloc.h>
8 #include <linux/wait.h>
9 #include <linux/writeback.h>
12 #include "mds_client.h"
14 #include <linux/ceph/decode.h>
15 #include <linux/ceph/messenger.h>
18 * Capability management
20 * The Ceph metadata servers control client access to inode metadata
21 * and file data by issuing capabilities, granting clients permission
22 * to read and/or write both inode field and file data to OSDs
23 * (storage nodes). Each capability consists of a set of bits
24 * indicating which operations are allowed.
26 * If the client holds a *_SHARED cap, the client has a coherent value
27 * that can be safely read from the cached inode.
29 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
30 * client is allowed to change inode attributes (e.g., file size,
31 * mtime), note its dirty state in the ceph_cap, and asynchronously
32 * flush that metadata change to the MDS.
34 * In the event of a conflicting operation (perhaps by another
35 * client), the MDS will revoke the conflicting client capabilities.
37 * In order for a client to cache an inode, it must hold a capability
38 * with at least one MDS server. When inodes are released, release
39 * notifications are batched and periodically sent en masse to the MDS
40 * cluster to release server state.
45 * Generate readable cap strings for debugging output.
47 #define MAX_CAP_STR 20
48 static char cap_str[MAX_CAP_STR][40];
49 static DEFINE_SPINLOCK(cap_str_lock);
50 static int last_cap_str;
52 static char *gcap_string(char *s, int c)
54 if (c & CEPH_CAP_GSHARED)
56 if (c & CEPH_CAP_GEXCL)
58 if (c & CEPH_CAP_GCACHE)
64 if (c & CEPH_CAP_GBUFFER)
66 if (c & CEPH_CAP_GLAZYIO)
71 const char *ceph_cap_string(int caps)
77 spin_lock(&cap_str_lock);
79 if (last_cap_str == MAX_CAP_STR)
81 spin_unlock(&cap_str_lock);
85 if (caps & CEPH_CAP_PIN)
88 c = (caps >> CEPH_CAP_SAUTH) & 3;
91 s = gcap_string(s, c);
94 c = (caps >> CEPH_CAP_SLINK) & 3;
97 s = gcap_string(s, c);
100 c = (caps >> CEPH_CAP_SXATTR) & 3;
103 s = gcap_string(s, c);
106 c = caps >> CEPH_CAP_SFILE;
109 s = gcap_string(s, c);
118 void ceph_caps_init(struct ceph_mds_client *mdsc)
120 INIT_LIST_HEAD(&mdsc->caps_list);
121 spin_lock_init(&mdsc->caps_list_lock);
124 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
126 struct ceph_cap *cap;
128 spin_lock(&mdsc->caps_list_lock);
129 while (!list_empty(&mdsc->caps_list)) {
130 cap = list_first_entry(&mdsc->caps_list,
131 struct ceph_cap, caps_item);
132 list_del(&cap->caps_item);
133 kmem_cache_free(ceph_cap_cachep, cap);
135 mdsc->caps_total_count = 0;
136 mdsc->caps_avail_count = 0;
137 mdsc->caps_use_count = 0;
138 mdsc->caps_reserve_count = 0;
139 mdsc->caps_min_count = 0;
140 spin_unlock(&mdsc->caps_list_lock);
143 void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta)
145 spin_lock(&mdsc->caps_list_lock);
146 mdsc->caps_min_count += delta;
147 BUG_ON(mdsc->caps_min_count < 0);
148 spin_unlock(&mdsc->caps_list_lock);
151 void ceph_reserve_caps(struct ceph_mds_client *mdsc,
152 struct ceph_cap_reservation *ctx, int need)
155 struct ceph_cap *cap;
160 dout("reserve caps ctx=%p need=%d\n", ctx, need);
162 /* first reserve any caps that are already allocated */
163 spin_lock(&mdsc->caps_list_lock);
164 if (mdsc->caps_avail_count >= need)
167 have = mdsc->caps_avail_count;
168 mdsc->caps_avail_count -= have;
169 mdsc->caps_reserve_count += have;
170 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
171 mdsc->caps_reserve_count +
172 mdsc->caps_avail_count);
173 spin_unlock(&mdsc->caps_list_lock);
175 for (i = have; i < need; i++) {
176 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
179 list_add(&cap->caps_item, &newcaps);
182 /* we didn't manage to reserve as much as we needed */
183 if (have + alloc != need)
184 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
185 ctx, need, have + alloc);
187 spin_lock(&mdsc->caps_list_lock);
188 mdsc->caps_total_count += alloc;
189 mdsc->caps_reserve_count += alloc;
190 list_splice(&newcaps, &mdsc->caps_list);
192 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
193 mdsc->caps_reserve_count +
194 mdsc->caps_avail_count);
195 spin_unlock(&mdsc->caps_list_lock);
198 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
199 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
200 mdsc->caps_reserve_count, mdsc->caps_avail_count);
203 int ceph_unreserve_caps(struct ceph_mds_client *mdsc,
204 struct ceph_cap_reservation *ctx)
206 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
208 spin_lock(&mdsc->caps_list_lock);
209 BUG_ON(mdsc->caps_reserve_count < ctx->count);
210 mdsc->caps_reserve_count -= ctx->count;
211 mdsc->caps_avail_count += ctx->count;
213 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
214 mdsc->caps_total_count, mdsc->caps_use_count,
215 mdsc->caps_reserve_count, mdsc->caps_avail_count);
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);
224 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
225 struct ceph_cap_reservation *ctx)
227 struct ceph_cap *cap = NULL;
229 /* temporary, until we do something about cap import/export */
231 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
233 spin_lock(&mdsc->caps_list_lock);
234 mdsc->caps_use_count++;
235 mdsc->caps_total_count++;
236 spin_unlock(&mdsc->caps_list_lock);
241 spin_lock(&mdsc->caps_list_lock);
242 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
243 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
244 mdsc->caps_reserve_count, mdsc->caps_avail_count);
246 BUG_ON(ctx->count > mdsc->caps_reserve_count);
247 BUG_ON(list_empty(&mdsc->caps_list));
250 mdsc->caps_reserve_count--;
251 mdsc->caps_use_count++;
253 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
254 list_del(&cap->caps_item);
256 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
257 mdsc->caps_reserve_count + mdsc->caps_avail_count);
258 spin_unlock(&mdsc->caps_list_lock);
262 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
264 spin_lock(&mdsc->caps_list_lock);
265 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
266 cap, mdsc->caps_total_count, mdsc->caps_use_count,
267 mdsc->caps_reserve_count, mdsc->caps_avail_count);
268 mdsc->caps_use_count--;
270 * Keep some preallocated caps around (ceph_min_count), to
271 * avoid lots of free/alloc churn.
273 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
274 mdsc->caps_min_count) {
275 mdsc->caps_total_count--;
276 kmem_cache_free(ceph_cap_cachep, cap);
278 mdsc->caps_avail_count++;
279 list_add(&cap->caps_item, &mdsc->caps_list);
282 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
283 mdsc->caps_reserve_count + mdsc->caps_avail_count);
284 spin_unlock(&mdsc->caps_list_lock);
287 void ceph_reservation_status(struct ceph_fs_client *fsc,
288 int *total, int *avail, int *used, int *reserved,
291 struct ceph_mds_client *mdsc = fsc->mdsc;
294 *total = mdsc->caps_total_count;
296 *avail = mdsc->caps_avail_count;
298 *used = mdsc->caps_use_count;
300 *reserved = mdsc->caps_reserve_count;
302 *min = mdsc->caps_min_count;
306 * Find ceph_cap for given mds, if any.
308 * Called with i_ceph_lock held.
310 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
312 struct ceph_cap *cap;
313 struct rb_node *n = ci->i_caps.rb_node;
316 cap = rb_entry(n, struct ceph_cap, ci_node);
319 else if (mds > cap->mds)
327 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
329 struct ceph_cap *cap;
331 spin_lock(&ci->i_ceph_lock);
332 cap = __get_cap_for_mds(ci, mds);
333 spin_unlock(&ci->i_ceph_lock);
338 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
340 static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
342 struct ceph_cap *cap;
346 /* prefer mds with WR|BUFFER|EXCL caps */
347 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
348 cap = rb_entry(p, struct ceph_cap, ci_node);
350 if (cap->issued & (CEPH_CAP_FILE_WR |
351 CEPH_CAP_FILE_BUFFER |
358 int ceph_get_cap_mds(struct inode *inode)
360 struct ceph_inode_info *ci = ceph_inode(inode);
362 spin_lock(&ci->i_ceph_lock);
363 mds = __ceph_get_cap_mds(ceph_inode(inode));
364 spin_unlock(&ci->i_ceph_lock);
369 * Called under i_ceph_lock.
371 static void __insert_cap_node(struct ceph_inode_info *ci,
372 struct ceph_cap *new)
374 struct rb_node **p = &ci->i_caps.rb_node;
375 struct rb_node *parent = NULL;
376 struct ceph_cap *cap = NULL;
380 cap = rb_entry(parent, struct ceph_cap, ci_node);
381 if (new->mds < cap->mds)
383 else if (new->mds > cap->mds)
389 rb_link_node(&new->ci_node, parent, p);
390 rb_insert_color(&new->ci_node, &ci->i_caps);
394 * (re)set cap hold timeouts, which control the delayed release
395 * of unused caps back to the MDS. Should be called on cap use.
397 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
398 struct ceph_inode_info *ci)
400 struct ceph_mount_options *ma = mdsc->fsc->mount_options;
402 ci->i_hold_caps_min = round_jiffies(jiffies +
403 ma->caps_wanted_delay_min * HZ);
404 ci->i_hold_caps_max = round_jiffies(jiffies +
405 ma->caps_wanted_delay_max * HZ);
406 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
407 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
411 * (Re)queue cap at the end of the delayed cap release list.
413 * If I_FLUSH is set, leave the inode at the front of the list.
415 * Caller holds i_ceph_lock
416 * -> we take mdsc->cap_delay_lock
418 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
419 struct ceph_inode_info *ci)
421 __cap_set_timeouts(mdsc, ci);
422 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
423 ci->i_ceph_flags, ci->i_hold_caps_max);
424 if (!mdsc->stopping) {
425 spin_lock(&mdsc->cap_delay_lock);
426 if (!list_empty(&ci->i_cap_delay_list)) {
427 if (ci->i_ceph_flags & CEPH_I_FLUSH)
429 list_del_init(&ci->i_cap_delay_list);
431 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
433 spin_unlock(&mdsc->cap_delay_lock);
438 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
439 * indicating we should send a cap message to flush dirty metadata
440 * asap, and move to the front of the delayed cap list.
442 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
443 struct ceph_inode_info *ci)
445 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
446 spin_lock(&mdsc->cap_delay_lock);
447 ci->i_ceph_flags |= CEPH_I_FLUSH;
448 if (!list_empty(&ci->i_cap_delay_list))
449 list_del_init(&ci->i_cap_delay_list);
450 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
451 spin_unlock(&mdsc->cap_delay_lock);
455 * Cancel delayed work on cap.
457 * Caller must hold i_ceph_lock.
459 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
460 struct ceph_inode_info *ci)
462 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
463 if (list_empty(&ci->i_cap_delay_list))
465 spin_lock(&mdsc->cap_delay_lock);
466 list_del_init(&ci->i_cap_delay_list);
467 spin_unlock(&mdsc->cap_delay_lock);
471 * Common issue checks for add_cap, handle_cap_grant.
473 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
476 unsigned had = __ceph_caps_issued(ci, NULL);
479 * Each time we receive FILE_CACHE anew, we increment
482 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
483 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
488 * if we are newly issued FILE_SHARED, mark dir not complete; we
489 * don't know what happened to this directory while we didn't
492 if ((issued & CEPH_CAP_FILE_SHARED) &&
493 (had & CEPH_CAP_FILE_SHARED) == 0) {
495 if (S_ISDIR(ci->vfs_inode.i_mode)) {
496 dout(" marking %p NOT complete\n", &ci->vfs_inode);
497 __ceph_dir_clear_complete(ci);
503 * Add a capability under the given MDS session.
505 * Caller should hold session snap_rwsem (read) and s_mutex.
507 * @fmode is the open file mode, if we are opening a file, otherwise
508 * it is < 0. (This is so we can atomically add the cap and add an
509 * open file reference to it.)
511 void ceph_add_cap(struct inode *inode,
512 struct ceph_mds_session *session, u64 cap_id,
513 int fmode, unsigned issued, unsigned wanted,
514 unsigned seq, unsigned mseq, u64 realmino, int flags,
515 struct ceph_cap **new_cap)
517 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
518 struct ceph_inode_info *ci = ceph_inode(inode);
519 struct ceph_cap *cap;
520 int mds = session->s_mds;
523 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
524 session->s_mds, cap_id, ceph_cap_string(issued), seq);
527 * If we are opening the file, include file mode wanted bits
531 wanted |= ceph_caps_for_mode(fmode);
533 cap = __get_cap_for_mds(ci, mds);
539 cap->implemented = 0;
545 __insert_cap_node(ci, cap);
547 /* add to session cap list */
548 cap->session = session;
549 spin_lock(&session->s_cap_lock);
550 list_add_tail(&cap->session_caps, &session->s_caps);
551 session->s_nr_caps++;
552 spin_unlock(&session->s_cap_lock);
555 * auth mds of the inode changed. we received the cap export
556 * message, but still haven't received the cap import message.
557 * handle_cap_export() updated the new auth MDS' cap.
559 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
560 * a message that was send before the cap import message. So
563 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
564 WARN_ON(cap != ci->i_auth_cap);
565 WARN_ON(cap->cap_id != cap_id);
568 issued |= cap->issued;
569 flags |= CEPH_CAP_FLAG_AUTH;
573 if (!ci->i_snap_realm) {
575 * add this inode to the appropriate snap realm
577 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
580 spin_lock(&realm->inodes_with_caps_lock);
581 ci->i_snap_realm = realm;
582 list_add(&ci->i_snap_realm_item,
583 &realm->inodes_with_caps);
584 spin_unlock(&realm->inodes_with_caps_lock);
586 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
592 __check_cap_issue(ci, cap, issued);
595 * If we are issued caps we don't want, or the mds' wanted
596 * value appears to be off, queue a check so we'll release
597 * later and/or update the mds wanted value.
599 actual_wanted = __ceph_caps_wanted(ci);
600 if ((wanted & ~actual_wanted) ||
601 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
602 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
603 ceph_cap_string(issued), ceph_cap_string(wanted),
604 ceph_cap_string(actual_wanted));
605 __cap_delay_requeue(mdsc, ci);
608 if (flags & CEPH_CAP_FLAG_AUTH) {
609 if (ci->i_auth_cap == NULL ||
610 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
611 ci->i_auth_cap = cap;
612 cap->mds_wanted = wanted;
615 WARN_ON(ci->i_auth_cap == cap);
618 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
619 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
620 ceph_cap_string(issued|cap->issued), seq, mds);
621 cap->cap_id = cap_id;
622 cap->issued = issued;
623 cap->implemented |= issued;
624 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
625 cap->mds_wanted = wanted;
627 cap->mds_wanted |= wanted;
629 cap->issue_seq = seq;
631 cap->cap_gen = session->s_cap_gen;
634 __ceph_get_fmode(ci, fmode);
638 * Return true if cap has not timed out and belongs to the current
639 * generation of the MDS session (i.e. has not gone 'stale' due to
640 * us losing touch with the mds).
642 static int __cap_is_valid(struct ceph_cap *cap)
647 spin_lock(&cap->session->s_gen_ttl_lock);
648 gen = cap->session->s_cap_gen;
649 ttl = cap->session->s_cap_ttl;
650 spin_unlock(&cap->session->s_gen_ttl_lock);
652 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
653 dout("__cap_is_valid %p cap %p issued %s "
654 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
655 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
663 * Return set of valid cap bits issued to us. Note that caps time
664 * out, and may be invalidated in bulk if the client session times out
665 * and session->s_cap_gen is bumped.
667 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
669 int have = ci->i_snap_caps;
670 struct ceph_cap *cap;
675 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
676 cap = rb_entry(p, struct ceph_cap, ci_node);
677 if (!__cap_is_valid(cap))
679 dout("__ceph_caps_issued %p cap %p issued %s\n",
680 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
683 *implemented |= cap->implemented;
686 * exclude caps issued by non-auth MDS, but are been revoking
687 * by the auth MDS. The non-auth MDS should be revoking/exporting
688 * these caps, but the message is delayed.
690 if (ci->i_auth_cap) {
691 cap = ci->i_auth_cap;
692 have &= ~cap->implemented | cap->issued;
698 * Get cap bits issued by caps other than @ocap
700 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
702 int have = ci->i_snap_caps;
703 struct ceph_cap *cap;
706 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
707 cap = rb_entry(p, struct ceph_cap, ci_node);
710 if (!__cap_is_valid(cap))
718 * Move a cap to the end of the LRU (oldest caps at list head, newest
721 static void __touch_cap(struct ceph_cap *cap)
723 struct ceph_mds_session *s = cap->session;
725 spin_lock(&s->s_cap_lock);
726 if (s->s_cap_iterator == NULL) {
727 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
729 list_move_tail(&cap->session_caps, &s->s_caps);
731 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
732 &cap->ci->vfs_inode, cap, s->s_mds);
734 spin_unlock(&s->s_cap_lock);
738 * Check if we hold the given mask. If so, move the cap(s) to the
739 * front of their respective LRUs. (This is the preferred way for
740 * callers to check for caps they want.)
742 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
744 struct ceph_cap *cap;
746 int have = ci->i_snap_caps;
748 if ((have & mask) == mask) {
749 dout("__ceph_caps_issued_mask %p snap issued %s"
750 " (mask %s)\n", &ci->vfs_inode,
751 ceph_cap_string(have),
752 ceph_cap_string(mask));
756 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
757 cap = rb_entry(p, struct ceph_cap, ci_node);
758 if (!__cap_is_valid(cap))
760 if ((cap->issued & mask) == mask) {
761 dout("__ceph_caps_issued_mask %p cap %p issued %s"
762 " (mask %s)\n", &ci->vfs_inode, cap,
763 ceph_cap_string(cap->issued),
764 ceph_cap_string(mask));
770 /* does a combination of caps satisfy mask? */
772 if ((have & mask) == mask) {
773 dout("__ceph_caps_issued_mask %p combo issued %s"
774 " (mask %s)\n", &ci->vfs_inode,
775 ceph_cap_string(cap->issued),
776 ceph_cap_string(mask));
780 /* touch this + preceding caps */
782 for (q = rb_first(&ci->i_caps); q != p;
784 cap = rb_entry(q, struct ceph_cap,
786 if (!__cap_is_valid(cap))
799 * Return true if mask caps are currently being revoked by an MDS.
801 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
802 struct ceph_cap *ocap, int mask)
804 struct ceph_cap *cap;
807 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
808 cap = rb_entry(p, struct ceph_cap, ci_node);
810 (cap->implemented & ~cap->issued & mask))
816 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
818 struct inode *inode = &ci->vfs_inode;
821 spin_lock(&ci->i_ceph_lock);
822 ret = __ceph_caps_revoking_other(ci, NULL, mask);
823 spin_unlock(&ci->i_ceph_lock);
824 dout("ceph_caps_revoking %p %s = %d\n", inode,
825 ceph_cap_string(mask), ret);
829 int __ceph_caps_used(struct ceph_inode_info *ci)
833 used |= CEPH_CAP_PIN;
835 used |= CEPH_CAP_FILE_RD;
836 if (ci->i_rdcache_ref || ci->vfs_inode.i_data.nrpages)
837 used |= CEPH_CAP_FILE_CACHE;
839 used |= CEPH_CAP_FILE_WR;
840 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
841 used |= CEPH_CAP_FILE_BUFFER;
846 * wanted, by virtue of open file modes
848 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
852 for (mode = 0; mode < CEPH_FILE_MODE_NUM; mode++)
853 if (ci->i_nr_by_mode[mode])
854 want |= ceph_caps_for_mode(mode);
859 * Return caps we have registered with the MDS(s) as 'wanted'.
861 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
863 struct ceph_cap *cap;
867 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
868 cap = rb_entry(p, struct ceph_cap, ci_node);
869 if (!__cap_is_valid(cap))
871 if (cap == ci->i_auth_cap)
872 mds_wanted |= cap->mds_wanted;
874 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
880 * called under i_ceph_lock
882 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
884 return !RB_EMPTY_ROOT(&ci->i_caps);
887 int ceph_is_any_caps(struct inode *inode)
889 struct ceph_inode_info *ci = ceph_inode(inode);
892 spin_lock(&ci->i_ceph_lock);
893 ret = __ceph_is_any_caps(ci);
894 spin_unlock(&ci->i_ceph_lock);
899 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
901 struct ceph_snap_realm *realm = ci->i_snap_realm;
902 spin_lock(&realm->inodes_with_caps_lock);
903 list_del_init(&ci->i_snap_realm_item);
904 ci->i_snap_realm_counter++;
905 ci->i_snap_realm = NULL;
906 spin_unlock(&realm->inodes_with_caps_lock);
907 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
912 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
914 * caller should hold i_ceph_lock.
915 * caller will not hold session s_mutex if called from destroy_inode.
917 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
919 struct ceph_mds_session *session = cap->session;
920 struct ceph_inode_info *ci = cap->ci;
921 struct ceph_mds_client *mdsc =
922 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
925 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
927 /* remove from session list */
928 spin_lock(&session->s_cap_lock);
929 if (session->s_cap_iterator == cap) {
930 /* not yet, we are iterating over this very cap */
931 dout("__ceph_remove_cap delaying %p removal from session %p\n",
934 list_del_init(&cap->session_caps);
935 session->s_nr_caps--;
939 /* protect backpointer with s_cap_lock: see iterate_session_caps */
943 * s_cap_reconnect is protected by s_cap_lock. no one changes
944 * s_cap_gen while session is in the reconnect state.
947 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
948 cap->queue_release = 1;
950 list_add_tail(&cap->session_caps,
951 &session->s_cap_releases);
952 session->s_num_cap_releases++;
956 cap->queue_release = 0;
958 cap->cap_ino = ci->i_vino.ino;
960 spin_unlock(&session->s_cap_lock);
962 /* remove from inode list */
963 rb_erase(&cap->ci_node, &ci->i_caps);
964 if (ci->i_auth_cap == cap)
965 ci->i_auth_cap = NULL;
968 ceph_put_cap(mdsc, cap);
970 /* when reconnect denied, we remove session caps forcibly,
971 * i_wr_ref can be non-zero. If there are ongoing write,
974 if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm)
975 drop_inode_snap_realm(ci);
977 if (!__ceph_is_any_real_caps(ci))
978 __cap_delay_cancel(mdsc, ci);
982 * Build and send a cap message to the given MDS.
984 * Caller should be holding s_mutex.
986 static int send_cap_msg(struct ceph_mds_session *session,
987 u64 ino, u64 cid, int op,
988 int caps, int wanted, int dirty,
989 u32 seq, u64 flush_tid, u64 oldest_flush_tid,
990 u32 issue_seq, u32 mseq, u64 size, u64 max_size,
991 struct timespec *mtime, struct timespec *atime,
993 kuid_t uid, kgid_t gid, umode_t mode,
995 struct ceph_buffer *xattrs_buf,
996 u64 follows, bool inline_data)
998 struct ceph_mds_caps *fc;
999 struct ceph_msg *msg;
1003 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1004 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1005 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
1006 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
1007 ceph_cap_string(dirty),
1008 seq, issue_seq, flush_tid, oldest_flush_tid,
1009 mseq, follows, size, max_size,
1010 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
1012 /* flock buffer size + inline version + inline data size +
1013 * osd_epoch_barrier + oldest_flush_tid */
1014 extra_len = 4 + 8 + 4 + 4 + 8;
1015 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1020 msg->hdr.version = cpu_to_le16(6);
1021 msg->hdr.tid = cpu_to_le64(flush_tid);
1023 fc = msg->front.iov_base;
1024 memset(fc, 0, sizeof(*fc));
1026 fc->cap_id = cpu_to_le64(cid);
1027 fc->op = cpu_to_le32(op);
1028 fc->seq = cpu_to_le32(seq);
1029 fc->issue_seq = cpu_to_le32(issue_seq);
1030 fc->migrate_seq = cpu_to_le32(mseq);
1031 fc->caps = cpu_to_le32(caps);
1032 fc->wanted = cpu_to_le32(wanted);
1033 fc->dirty = cpu_to_le32(dirty);
1034 fc->ino = cpu_to_le64(ino);
1035 fc->snap_follows = cpu_to_le64(follows);
1037 fc->size = cpu_to_le64(size);
1038 fc->max_size = cpu_to_le64(max_size);
1040 ceph_encode_timespec(&fc->mtime, mtime);
1042 ceph_encode_timespec(&fc->atime, atime);
1043 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
1045 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, uid));
1046 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, gid));
1047 fc->mode = cpu_to_le32(mode);
1050 /* flock buffer size */
1051 ceph_encode_32(&p, 0);
1052 /* inline version */
1053 ceph_encode_64(&p, inline_data ? 0 : CEPH_INLINE_NONE);
1054 /* inline data size */
1055 ceph_encode_32(&p, 0);
1056 /* osd_epoch_barrier */
1057 ceph_encode_32(&p, 0);
1058 /* oldest_flush_tid */
1059 ceph_encode_64(&p, oldest_flush_tid);
1061 fc->xattr_version = cpu_to_le64(xattr_version);
1063 msg->middle = ceph_buffer_get(xattrs_buf);
1064 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1065 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1068 ceph_con_send(&session->s_con, msg);
1073 * Queue cap releases when an inode is dropped from our cache. Since
1074 * inode is about to be destroyed, there is no need for i_ceph_lock.
1076 void ceph_queue_caps_release(struct inode *inode)
1078 struct ceph_inode_info *ci = ceph_inode(inode);
1081 p = rb_first(&ci->i_caps);
1083 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1085 __ceph_remove_cap(cap, true);
1090 * Send a cap msg on the given inode. Update our caps state, then
1091 * drop i_ceph_lock and send the message.
1093 * Make note of max_size reported/requested from mds, revoked caps
1094 * that have now been implemented.
1096 * Make half-hearted attempt ot to invalidate page cache if we are
1097 * dropping RDCACHE. Note that this will leave behind locked pages
1098 * that we'll then need to deal with elsewhere.
1100 * Return non-zero if delayed release, or we experienced an error
1101 * such that the caller should requeue + retry later.
1103 * called with i_ceph_lock, then drops it.
1104 * caller should hold snap_rwsem (read), s_mutex.
1106 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1107 int op, int used, int want, int retain, int flushing,
1108 u64 flush_tid, u64 oldest_flush_tid)
1109 __releases(cap->ci->i_ceph_lock)
1111 struct ceph_inode_info *ci = cap->ci;
1112 struct inode *inode = &ci->vfs_inode;
1113 u64 cap_id = cap->cap_id;
1114 int held, revoking, dropping, keep;
1115 u64 seq, issue_seq, mseq, time_warp_seq, follows;
1117 struct timespec mtime, atime;
1122 struct ceph_mds_session *session;
1123 u64 xattr_version = 0;
1124 struct ceph_buffer *xattr_blob = NULL;
1129 held = cap->issued | cap->implemented;
1130 revoking = cap->implemented & ~cap->issued;
1131 retain &= ~revoking;
1132 dropping = cap->issued & ~retain;
1134 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1135 inode, cap, cap->session,
1136 ceph_cap_string(held), ceph_cap_string(held & retain),
1137 ceph_cap_string(revoking));
1138 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1140 session = cap->session;
1142 /* don't release wanted unless we've waited a bit. */
1143 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1144 time_before(jiffies, ci->i_hold_caps_min)) {
1145 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1146 ceph_cap_string(cap->issued),
1147 ceph_cap_string(cap->issued & retain),
1148 ceph_cap_string(cap->mds_wanted),
1149 ceph_cap_string(want));
1150 want |= cap->mds_wanted;
1151 retain |= cap->issued;
1154 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1156 cap->issued &= retain; /* drop bits we don't want */
1157 if (cap->implemented & ~cap->issued) {
1159 * Wake up any waiters on wanted -> needed transition.
1160 * This is due to the weird transition from buffered
1161 * to sync IO... we need to flush dirty pages _before_
1162 * allowing sync writes to avoid reordering.
1166 cap->implemented &= cap->issued | used;
1167 cap->mds_wanted = want;
1169 follows = flushing ? ci->i_head_snapc->seq : 0;
1171 keep = cap->implemented;
1173 issue_seq = cap->issue_seq;
1175 size = inode->i_size;
1176 ci->i_reported_size = size;
1177 max_size = ci->i_wanted_max_size;
1178 ci->i_requested_max_size = max_size;
1179 mtime = inode->i_mtime;
1180 atime = inode->i_atime;
1181 time_warp_seq = ci->i_time_warp_seq;
1184 mode = inode->i_mode;
1186 if (flushing & CEPH_CAP_XATTR_EXCL) {
1187 __ceph_build_xattrs_blob(ci);
1188 xattr_blob = ci->i_xattrs.blob;
1189 xattr_version = ci->i_xattrs.version;
1192 inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1194 spin_unlock(&ci->i_ceph_lock);
1196 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1197 op, keep, want, flushing, seq,
1198 flush_tid, oldest_flush_tid, issue_seq, mseq,
1199 size, max_size, &mtime, &atime, time_warp_seq,
1200 uid, gid, mode, xattr_version, xattr_blob,
1201 follows, inline_data);
1203 dout("error sending cap msg, must requeue %p\n", inode);
1208 wake_up_all(&ci->i_cap_wq);
1214 * When a snapshot is taken, clients accumulate dirty metadata on
1215 * inodes with capabilities in ceph_cap_snaps to describe the file
1216 * state at the time the snapshot was taken. This must be flushed
1217 * asynchronously back to the MDS once sync writes complete and dirty
1218 * data is written out.
1220 * Unless @kick is true, skip cap_snaps that were already sent to
1221 * the MDS (i.e., during this session).
1223 * Called under i_ceph_lock. Takes s_mutex as needed.
1225 void __ceph_flush_snaps(struct ceph_inode_info *ci,
1226 struct ceph_mds_session **psession,
1228 __releases(ci->i_ceph_lock)
1229 __acquires(ci->i_ceph_lock)
1231 struct inode *inode = &ci->vfs_inode;
1233 struct ceph_cap_snap *capsnap;
1235 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1236 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1238 u64 next_follows = 0; /* keep track of how far we've gotten through the
1239 i_cap_snaps list, and skip these entries next time
1240 around to avoid an infinite loop */
1243 session = *psession;
1245 dout("__flush_snaps %p\n", inode);
1247 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1248 /* avoid an infiniute loop after retry */
1249 if (capsnap->follows < next_follows)
1252 * we need to wait for sync writes to complete and for dirty
1253 * pages to be written out.
1255 if (capsnap->dirty_pages || capsnap->writing)
1258 /* should be removed by ceph_try_drop_cap_snap() */
1259 BUG_ON(!capsnap->need_flush);
1261 /* pick mds, take s_mutex */
1262 if (ci->i_auth_cap == NULL) {
1263 dout("no auth cap (migrating?), doing nothing\n");
1267 /* only flush each capsnap once */
1268 if (!kick && !list_empty(&capsnap->flushing_item)) {
1269 dout("already flushed %p, skipping\n", capsnap);
1273 mds = ci->i_auth_cap->session->s_mds;
1274 mseq = ci->i_auth_cap->mseq;
1276 if (session && session->s_mds != mds) {
1277 dout("oops, wrong session %p mutex\n", session);
1281 mutex_unlock(&session->s_mutex);
1282 ceph_put_mds_session(session);
1286 spin_unlock(&ci->i_ceph_lock);
1287 mutex_lock(&mdsc->mutex);
1288 session = __ceph_lookup_mds_session(mdsc, mds);
1289 mutex_unlock(&mdsc->mutex);
1291 dout("inverting session/ino locks on %p\n",
1293 mutex_lock(&session->s_mutex);
1296 * if session == NULL, we raced against a cap
1297 * deletion or migration. retry, and we'll
1298 * get a better @mds value next time.
1300 spin_lock(&ci->i_ceph_lock);
1304 spin_lock(&mdsc->cap_dirty_lock);
1305 capsnap->flush_tid = ++mdsc->last_cap_flush_tid;
1306 spin_unlock(&mdsc->cap_dirty_lock);
1308 atomic_inc(&capsnap->nref);
1309 if (list_empty(&capsnap->flushing_item))
1310 list_add_tail(&capsnap->flushing_item,
1311 &session->s_cap_snaps_flushing);
1312 spin_unlock(&ci->i_ceph_lock);
1314 dout("flush_snaps %p cap_snap %p follows %lld tid %llu\n",
1315 inode, capsnap, capsnap->follows, capsnap->flush_tid);
1316 send_cap_msg(session, ceph_vino(inode).ino, 0,
1317 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1318 capsnap->dirty, 0, capsnap->flush_tid, 0,
1319 0, mseq, capsnap->size, 0,
1320 &capsnap->mtime, &capsnap->atime,
1321 capsnap->time_warp_seq,
1322 capsnap->uid, capsnap->gid, capsnap->mode,
1323 capsnap->xattr_version, capsnap->xattr_blob,
1324 capsnap->follows, capsnap->inline_data);
1326 next_follows = capsnap->follows + 1;
1327 ceph_put_cap_snap(capsnap);
1329 spin_lock(&ci->i_ceph_lock);
1333 /* we flushed them all; remove this inode from the queue */
1334 spin_lock(&mdsc->snap_flush_lock);
1335 list_del_init(&ci->i_snap_flush_item);
1336 spin_unlock(&mdsc->snap_flush_lock);
1340 *psession = session;
1342 mutex_unlock(&session->s_mutex);
1343 ceph_put_mds_session(session);
1347 static void ceph_flush_snaps(struct ceph_inode_info *ci)
1349 spin_lock(&ci->i_ceph_lock);
1350 __ceph_flush_snaps(ci, NULL, 0);
1351 spin_unlock(&ci->i_ceph_lock);
1355 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1356 * Caller is then responsible for calling __mark_inode_dirty with the
1357 * returned flags value.
1359 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1360 struct ceph_cap_flush **pcf)
1362 struct ceph_mds_client *mdsc =
1363 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1364 struct inode *inode = &ci->vfs_inode;
1365 int was = ci->i_dirty_caps;
1368 if (!ci->i_auth_cap) {
1369 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1370 "but no auth cap (session was closed?)\n",
1371 inode, ceph_ino(inode), ceph_cap_string(mask));
1375 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1376 ceph_cap_string(mask), ceph_cap_string(was),
1377 ceph_cap_string(was | mask));
1378 ci->i_dirty_caps |= mask;
1380 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1381 swap(ci->i_prealloc_cap_flush, *pcf);
1383 if (!ci->i_head_snapc) {
1384 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1385 ci->i_head_snapc = ceph_get_snap_context(
1386 ci->i_snap_realm->cached_context);
1388 dout(" inode %p now dirty snapc %p auth cap %p\n",
1389 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1390 BUG_ON(!list_empty(&ci->i_dirty_item));
1391 spin_lock(&mdsc->cap_dirty_lock);
1392 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1393 spin_unlock(&mdsc->cap_dirty_lock);
1394 if (ci->i_flushing_caps == 0) {
1396 dirty |= I_DIRTY_SYNC;
1399 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1401 BUG_ON(list_empty(&ci->i_dirty_item));
1402 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1403 (mask & CEPH_CAP_FILE_BUFFER))
1404 dirty |= I_DIRTY_DATASYNC;
1405 __cap_delay_requeue(mdsc, ci);
1409 static void __add_cap_flushing_to_inode(struct ceph_inode_info *ci,
1410 struct ceph_cap_flush *cf)
1412 struct rb_node **p = &ci->i_cap_flush_tree.rb_node;
1413 struct rb_node *parent = NULL;
1414 struct ceph_cap_flush *other = NULL;
1418 other = rb_entry(parent, struct ceph_cap_flush, i_node);
1420 if (cf->tid < other->tid)
1422 else if (cf->tid > other->tid)
1423 p = &(*p)->rb_right;
1428 rb_link_node(&cf->i_node, parent, p);
1429 rb_insert_color(&cf->i_node, &ci->i_cap_flush_tree);
1432 static void __add_cap_flushing_to_mdsc(struct ceph_mds_client *mdsc,
1433 struct ceph_cap_flush *cf)
1435 struct rb_node **p = &mdsc->cap_flush_tree.rb_node;
1436 struct rb_node *parent = NULL;
1437 struct ceph_cap_flush *other = NULL;
1441 other = rb_entry(parent, struct ceph_cap_flush, g_node);
1443 if (cf->tid < other->tid)
1445 else if (cf->tid > other->tid)
1446 p = &(*p)->rb_right;
1451 rb_link_node(&cf->g_node, parent, p);
1452 rb_insert_color(&cf->g_node, &mdsc->cap_flush_tree);
1455 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1457 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1460 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1463 kmem_cache_free(ceph_cap_flush_cachep, cf);
1466 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1468 struct rb_node *n = rb_first(&mdsc->cap_flush_tree);
1470 struct ceph_cap_flush *cf =
1471 rb_entry(n, struct ceph_cap_flush, g_node);
1478 * Add dirty inode to the flushing list. Assigned a seq number so we
1479 * can wait for caps to flush without starving.
1481 * Called under i_ceph_lock.
1483 static int __mark_caps_flushing(struct inode *inode,
1484 struct ceph_mds_session *session,
1485 u64 *flush_tid, u64 *oldest_flush_tid)
1487 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1488 struct ceph_inode_info *ci = ceph_inode(inode);
1489 struct ceph_cap_flush *cf = NULL;
1492 BUG_ON(ci->i_dirty_caps == 0);
1493 BUG_ON(list_empty(&ci->i_dirty_item));
1494 BUG_ON(!ci->i_prealloc_cap_flush);
1496 flushing = ci->i_dirty_caps;
1497 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1498 ceph_cap_string(flushing),
1499 ceph_cap_string(ci->i_flushing_caps),
1500 ceph_cap_string(ci->i_flushing_caps | flushing));
1501 ci->i_flushing_caps |= flushing;
1502 ci->i_dirty_caps = 0;
1503 dout(" inode %p now !dirty\n", inode);
1505 swap(cf, ci->i_prealloc_cap_flush);
1506 cf->caps = flushing;
1509 spin_lock(&mdsc->cap_dirty_lock);
1510 list_del_init(&ci->i_dirty_item);
1512 cf->tid = ++mdsc->last_cap_flush_tid;
1513 __add_cap_flushing_to_mdsc(mdsc, cf);
1514 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1516 if (list_empty(&ci->i_flushing_item)) {
1517 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1518 mdsc->num_cap_flushing++;
1519 dout(" inode %p now flushing tid %llu\n", inode, cf->tid);
1521 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1522 dout(" inode %p now flushing (more) tid %llu\n",
1525 spin_unlock(&mdsc->cap_dirty_lock);
1527 __add_cap_flushing_to_inode(ci, cf);
1529 *flush_tid = cf->tid;
1534 * try to invalidate mapping pages without blocking.
1536 static int try_nonblocking_invalidate(struct inode *inode)
1538 struct ceph_inode_info *ci = ceph_inode(inode);
1539 u32 invalidating_gen = ci->i_rdcache_gen;
1541 spin_unlock(&ci->i_ceph_lock);
1542 invalidate_mapping_pages(&inode->i_data, 0, -1);
1543 spin_lock(&ci->i_ceph_lock);
1545 if (inode->i_data.nrpages == 0 &&
1546 invalidating_gen == ci->i_rdcache_gen) {
1548 dout("try_nonblocking_invalidate %p success\n", inode);
1549 /* save any racing async invalidate some trouble */
1550 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1553 dout("try_nonblocking_invalidate %p failed\n", inode);
1558 * Swiss army knife function to examine currently used and wanted
1559 * versus held caps. Release, flush, ack revoked caps to mds as
1562 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1563 * cap release further.
1564 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1565 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1568 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1569 struct ceph_mds_session *session)
1571 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1572 struct ceph_mds_client *mdsc = fsc->mdsc;
1573 struct inode *inode = &ci->vfs_inode;
1574 struct ceph_cap *cap;
1575 u64 flush_tid, oldest_flush_tid;
1576 int file_wanted, used, cap_used;
1577 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1578 int issued, implemented, want, retain, revoking, flushing = 0;
1579 int mds = -1; /* keep track of how far we've gone through i_caps list
1580 to avoid an infinite loop on retry */
1582 int tried_invalidate = 0;
1583 int delayed = 0, sent = 0, force_requeue = 0, num;
1584 int queue_invalidate = 0;
1585 int is_delayed = flags & CHECK_CAPS_NODELAY;
1587 /* if we are unmounting, flush any unused caps immediately. */
1591 spin_lock(&ci->i_ceph_lock);
1593 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1594 flags |= CHECK_CAPS_FLUSH;
1596 /* flush snaps first time around only */
1597 if (!list_empty(&ci->i_cap_snaps))
1598 __ceph_flush_snaps(ci, &session, 0);
1601 spin_lock(&ci->i_ceph_lock);
1603 file_wanted = __ceph_caps_file_wanted(ci);
1604 used = __ceph_caps_used(ci);
1605 issued = __ceph_caps_issued(ci, &implemented);
1606 revoking = implemented & ~issued;
1609 retain = file_wanted | used | CEPH_CAP_PIN;
1610 if (!mdsc->stopping && inode->i_nlink > 0) {
1612 retain |= CEPH_CAP_ANY; /* be greedy */
1613 } else if (S_ISDIR(inode->i_mode) &&
1614 (issued & CEPH_CAP_FILE_SHARED) &&
1615 __ceph_dir_is_complete(ci)) {
1617 * If a directory is complete, we want to keep
1618 * the exclusive cap. So that MDS does not end up
1619 * revoking the shared cap on every create/unlink
1622 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1626 retain |= CEPH_CAP_ANY_SHARED;
1628 * keep RD only if we didn't have the file open RW,
1629 * because then the mds would revoke it anyway to
1630 * journal max_size=0.
1632 if (ci->i_max_size == 0)
1633 retain |= CEPH_CAP_ANY_RD;
1637 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1638 " issued %s revoking %s retain %s %s%s%s\n", inode,
1639 ceph_cap_string(file_wanted),
1640 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1641 ceph_cap_string(ci->i_flushing_caps),
1642 ceph_cap_string(issued), ceph_cap_string(revoking),
1643 ceph_cap_string(retain),
1644 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1645 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1646 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1649 * If we no longer need to hold onto old our caps, and we may
1650 * have cached pages, but don't want them, then try to invalidate.
1651 * If we fail, it's because pages are locked.... try again later.
1653 if ((!is_delayed || mdsc->stopping) &&
1654 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */
1655 inode->i_data.nrpages && /* have cached pages */
1656 (file_wanted == 0 || /* no open files */
1657 (revoking & (CEPH_CAP_FILE_CACHE|
1658 CEPH_CAP_FILE_LAZYIO))) && /* or revoking cache */
1659 !tried_invalidate) {
1660 dout("check_caps trying to invalidate on %p\n", inode);
1661 if (try_nonblocking_invalidate(inode) < 0) {
1662 if (revoking & (CEPH_CAP_FILE_CACHE|
1663 CEPH_CAP_FILE_LAZYIO)) {
1664 dout("check_caps queuing invalidate\n");
1665 queue_invalidate = 1;
1666 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1668 dout("check_caps failed to invalidate pages\n");
1669 /* we failed to invalidate pages. check these
1670 caps again later. */
1672 __cap_set_timeouts(mdsc, ci);
1675 tried_invalidate = 1;
1680 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1681 cap = rb_entry(p, struct ceph_cap, ci_node);
1684 /* avoid looping forever */
1685 if (mds >= cap->mds ||
1686 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1689 /* NOTE: no side-effects allowed, until we take s_mutex */
1692 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1693 cap_used &= ~ci->i_auth_cap->issued;
1695 revoking = cap->implemented & ~cap->issued;
1696 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1697 cap->mds, cap, ceph_cap_string(cap->issued),
1698 ceph_cap_string(cap_used),
1699 ceph_cap_string(cap->implemented),
1700 ceph_cap_string(revoking));
1702 if (cap == ci->i_auth_cap &&
1703 (cap->issued & CEPH_CAP_FILE_WR)) {
1704 /* request larger max_size from MDS? */
1705 if (ci->i_wanted_max_size > ci->i_max_size &&
1706 ci->i_wanted_max_size > ci->i_requested_max_size) {
1707 dout("requesting new max_size\n");
1711 /* approaching file_max? */
1712 if ((inode->i_size << 1) >= ci->i_max_size &&
1713 (ci->i_reported_size << 1) < ci->i_max_size) {
1714 dout("i_size approaching max_size\n");
1718 /* flush anything dirty? */
1719 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1721 dout("flushing dirty caps\n");
1725 /* completed revocation? going down and there are no caps? */
1726 if (revoking && (revoking & cap_used) == 0) {
1727 dout("completed revocation of %s\n",
1728 ceph_cap_string(cap->implemented & ~cap->issued));
1732 /* want more caps from mds? */
1733 if (want & ~(cap->mds_wanted | cap->issued))
1736 /* things we might delay */
1737 if ((cap->issued & ~retain) == 0 &&
1738 cap->mds_wanted == want)
1739 continue; /* nope, all good */
1745 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1746 time_before(jiffies, ci->i_hold_caps_max)) {
1747 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1748 ceph_cap_string(cap->issued),
1749 ceph_cap_string(cap->issued & retain),
1750 ceph_cap_string(cap->mds_wanted),
1751 ceph_cap_string(want));
1757 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1758 dout(" skipping %p I_NOFLUSH set\n", inode);
1762 if (session && session != cap->session) {
1763 dout("oops, wrong session %p mutex\n", session);
1764 mutex_unlock(&session->s_mutex);
1768 session = cap->session;
1769 if (mutex_trylock(&session->s_mutex) == 0) {
1770 dout("inverting session/ino locks on %p\n",
1772 spin_unlock(&ci->i_ceph_lock);
1773 if (took_snap_rwsem) {
1774 up_read(&mdsc->snap_rwsem);
1775 took_snap_rwsem = 0;
1777 mutex_lock(&session->s_mutex);
1781 /* take snap_rwsem after session mutex */
1782 if (!took_snap_rwsem) {
1783 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1784 dout("inverting snap/in locks on %p\n",
1786 spin_unlock(&ci->i_ceph_lock);
1787 down_read(&mdsc->snap_rwsem);
1788 took_snap_rwsem = 1;
1791 took_snap_rwsem = 1;
1794 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
1795 flushing = __mark_caps_flushing(inode, session,
1801 spin_lock(&mdsc->cap_dirty_lock);
1802 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1803 spin_unlock(&mdsc->cap_dirty_lock);
1806 mds = cap->mds; /* remember mds, so we don't repeat */
1809 /* __send_cap drops i_ceph_lock */
1810 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used,
1811 want, retain, flushing,
1812 flush_tid, oldest_flush_tid);
1813 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1817 * Reschedule delayed caps release if we delayed anything,
1820 if (delayed && is_delayed)
1821 force_requeue = 1; /* __send_cap delayed release; requeue */
1822 if (!delayed && !is_delayed)
1823 __cap_delay_cancel(mdsc, ci);
1824 else if (!is_delayed || force_requeue)
1825 __cap_delay_requeue(mdsc, ci);
1827 spin_unlock(&ci->i_ceph_lock);
1829 if (queue_invalidate)
1830 ceph_queue_invalidate(inode);
1833 mutex_unlock(&session->s_mutex);
1834 if (took_snap_rwsem)
1835 up_read(&mdsc->snap_rwsem);
1839 * Try to flush dirty caps back to the auth mds.
1841 static int try_flush_caps(struct inode *inode, u64 *ptid)
1843 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1844 struct ceph_inode_info *ci = ceph_inode(inode);
1845 struct ceph_mds_session *session = NULL;
1847 u64 flush_tid = 0, oldest_flush_tid = 0;
1850 spin_lock(&ci->i_ceph_lock);
1851 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1852 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1855 if (ci->i_dirty_caps && ci->i_auth_cap) {
1856 struct ceph_cap *cap = ci->i_auth_cap;
1857 int used = __ceph_caps_used(ci);
1858 int want = __ceph_caps_wanted(ci);
1861 if (!session || session != cap->session) {
1862 spin_unlock(&ci->i_ceph_lock);
1864 mutex_unlock(&session->s_mutex);
1865 session = cap->session;
1866 mutex_lock(&session->s_mutex);
1869 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1872 flushing = __mark_caps_flushing(inode, session, &flush_tid,
1875 /* __send_cap drops i_ceph_lock */
1876 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1877 (cap->issued | cap->implemented),
1878 flushing, flush_tid, oldest_flush_tid);
1881 spin_lock(&ci->i_ceph_lock);
1882 __cap_delay_requeue(mdsc, ci);
1883 spin_unlock(&ci->i_ceph_lock);
1886 struct rb_node *n = rb_last(&ci->i_cap_flush_tree);
1888 struct ceph_cap_flush *cf =
1889 rb_entry(n, struct ceph_cap_flush, i_node);
1890 flush_tid = cf->tid;
1892 flushing = ci->i_flushing_caps;
1893 spin_unlock(&ci->i_ceph_lock);
1897 mutex_unlock(&session->s_mutex);
1904 * Return true if we've flushed caps through the given flush_tid.
1906 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
1908 struct ceph_inode_info *ci = ceph_inode(inode);
1909 struct ceph_cap_flush *cf;
1913 spin_lock(&ci->i_ceph_lock);
1914 n = rb_first(&ci->i_cap_flush_tree);
1916 cf = rb_entry(n, struct ceph_cap_flush, i_node);
1917 if (cf->tid <= flush_tid)
1920 spin_unlock(&ci->i_ceph_lock);
1925 * Wait on any unsafe replies for the given inode. First wait on the
1926 * newest request, and make that the upper bound. Then, if there are
1927 * more requests, keep waiting on the oldest as long as it is still older
1928 * than the original request.
1930 static void sync_write_wait(struct inode *inode)
1932 struct ceph_inode_info *ci = ceph_inode(inode);
1933 struct list_head *head = &ci->i_unsafe_writes;
1934 struct ceph_osd_request *req;
1937 if (!S_ISREG(inode->i_mode))
1940 spin_lock(&ci->i_unsafe_lock);
1941 if (list_empty(head))
1944 /* set upper bound as _last_ entry in chain */
1945 req = list_last_entry(head, struct ceph_osd_request,
1947 last_tid = req->r_tid;
1950 ceph_osdc_get_request(req);
1951 spin_unlock(&ci->i_unsafe_lock);
1952 dout("sync_write_wait on tid %llu (until %llu)\n",
1953 req->r_tid, last_tid);
1954 wait_for_completion(&req->r_safe_completion);
1955 spin_lock(&ci->i_unsafe_lock);
1956 ceph_osdc_put_request(req);
1959 * from here on look at first entry in chain, since we
1960 * only want to wait for anything older than last_tid
1962 if (list_empty(head))
1964 req = list_first_entry(head, struct ceph_osd_request,
1966 } while (req->r_tid < last_tid);
1968 spin_unlock(&ci->i_unsafe_lock);
1972 * wait for any uncommitted directory operations to commit.
1974 static int unsafe_dirop_wait(struct inode *inode)
1976 struct ceph_inode_info *ci = ceph_inode(inode);
1977 struct list_head *head = &ci->i_unsafe_dirops;
1978 struct ceph_mds_request *req;
1982 if (!S_ISDIR(inode->i_mode))
1985 spin_lock(&ci->i_unsafe_lock);
1986 if (list_empty(head))
1989 req = list_last_entry(head, struct ceph_mds_request,
1991 last_tid = req->r_tid;
1994 ceph_mdsc_get_request(req);
1995 spin_unlock(&ci->i_unsafe_lock);
1997 dout("unsafe_dirop_wait %p wait on tid %llu (until %llu)\n",
1998 inode, req->r_tid, last_tid);
1999 ret = !wait_for_completion_timeout(&req->r_safe_completion,
2000 ceph_timeout_jiffies(req->r_timeout));
2002 ret = -EIO; /* timed out */
2004 ceph_mdsc_put_request(req);
2006 spin_lock(&ci->i_unsafe_lock);
2007 if (ret || list_empty(head))
2009 req = list_first_entry(head, struct ceph_mds_request,
2011 } while (req->r_tid < last_tid);
2013 spin_unlock(&ci->i_unsafe_lock);
2017 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2019 struct inode *inode = file->f_mapping->host;
2020 struct ceph_inode_info *ci = ceph_inode(inode);
2025 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2026 sync_write_wait(inode);
2028 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
2035 mutex_lock(&inode->i_mutex);
2037 dirty = try_flush_caps(inode, &flush_tid);
2038 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2040 ret = unsafe_dirop_wait(inode);
2043 * only wait on non-file metadata writeback (the mds
2044 * can recover size and mtime, so we don't need to
2047 if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2048 ret = wait_event_interruptible(ci->i_cap_wq,
2049 caps_are_flushed(inode, flush_tid));
2051 mutex_unlock(&inode->i_mutex);
2053 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2058 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2059 * queue inode for flush but don't do so immediately, because we can
2060 * get by with fewer MDS messages if we wait for data writeback to
2063 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2065 struct ceph_inode_info *ci = ceph_inode(inode);
2069 int wait = wbc->sync_mode == WB_SYNC_ALL;
2071 dout("write_inode %p wait=%d\n", inode, wait);
2073 dirty = try_flush_caps(inode, &flush_tid);
2075 err = wait_event_interruptible(ci->i_cap_wq,
2076 caps_are_flushed(inode, flush_tid));
2078 struct ceph_mds_client *mdsc =
2079 ceph_sb_to_client(inode->i_sb)->mdsc;
2081 spin_lock(&ci->i_ceph_lock);
2082 if (__ceph_caps_dirty(ci))
2083 __cap_delay_requeue_front(mdsc, ci);
2084 spin_unlock(&ci->i_ceph_lock);
2090 * After a recovering MDS goes active, we need to resend any caps
2093 * Caller holds session->s_mutex.
2095 static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
2096 struct ceph_mds_session *session)
2098 struct ceph_cap_snap *capsnap;
2100 dout("kick_flushing_capsnaps mds%d\n", session->s_mds);
2101 list_for_each_entry(capsnap, &session->s_cap_snaps_flushing,
2103 struct ceph_inode_info *ci = capsnap->ci;
2104 struct inode *inode = &ci->vfs_inode;
2105 struct ceph_cap *cap;
2107 spin_lock(&ci->i_ceph_lock);
2108 cap = ci->i_auth_cap;
2109 if (cap && cap->session == session) {
2110 dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
2112 __ceph_flush_snaps(ci, &session, 1);
2114 pr_err("%p auth cap %p not mds%d ???\n", inode,
2115 cap, session->s_mds);
2117 spin_unlock(&ci->i_ceph_lock);
2121 static int __kick_flushing_caps(struct ceph_mds_client *mdsc,
2122 struct ceph_mds_session *session,
2123 struct ceph_inode_info *ci,
2126 struct inode *inode = &ci->vfs_inode;
2127 struct ceph_cap *cap;
2128 struct ceph_cap_flush *cf;
2132 u64 oldest_flush_tid;
2134 spin_lock(&mdsc->cap_dirty_lock);
2135 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2136 spin_unlock(&mdsc->cap_dirty_lock);
2139 spin_lock(&ci->i_ceph_lock);
2140 cap = ci->i_auth_cap;
2141 if (!(cap && cap->session == session)) {
2142 pr_err("%p auth cap %p not mds%d ???\n", inode,
2143 cap, session->s_mds);
2144 spin_unlock(&ci->i_ceph_lock);
2148 for (n = rb_first(&ci->i_cap_flush_tree); n; n = rb_next(n)) {
2149 cf = rb_entry(n, struct ceph_cap_flush, i_node);
2150 if (cf->tid < first_tid)
2152 if (kick_all || cf->kick)
2156 spin_unlock(&ci->i_ceph_lock);
2160 cf = rb_entry(n, struct ceph_cap_flush, i_node);
2163 first_tid = cf->tid + 1;
2165 dout("kick_flushing_caps %p cap %p tid %llu %s\n", inode,
2166 cap, cf->tid, ceph_cap_string(cf->caps));
2167 delayed |= __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2168 __ceph_caps_used(ci),
2169 __ceph_caps_wanted(ci),
2170 cap->issued | cap->implemented,
2171 cf->caps, cf->tid, oldest_flush_tid);
2176 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2177 struct ceph_mds_session *session)
2179 struct ceph_inode_info *ci;
2180 struct ceph_cap *cap;
2181 struct ceph_cap_flush *cf;
2184 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2185 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2186 spin_lock(&ci->i_ceph_lock);
2187 cap = ci->i_auth_cap;
2188 if (!(cap && cap->session == session)) {
2189 pr_err("%p auth cap %p not mds%d ???\n",
2190 &ci->vfs_inode, cap, session->s_mds);
2191 spin_unlock(&ci->i_ceph_lock);
2197 * if flushing caps were revoked, we re-send the cap flush
2198 * in client reconnect stage. This guarantees MDS * processes
2199 * the cap flush message before issuing the flushing caps to
2202 if ((cap->issued & ci->i_flushing_caps) !=
2203 ci->i_flushing_caps) {
2204 spin_unlock(&ci->i_ceph_lock);
2205 if (!__kick_flushing_caps(mdsc, session, ci, true))
2207 spin_lock(&ci->i_ceph_lock);
2210 for (n = rb_first(&ci->i_cap_flush_tree); n; n = rb_next(n)) {
2211 cf = rb_entry(n, struct ceph_cap_flush, i_node);
2215 spin_unlock(&ci->i_ceph_lock);
2219 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2220 struct ceph_mds_session *session)
2222 struct ceph_inode_info *ci;
2224 kick_flushing_capsnaps(mdsc, session);
2226 dout("kick_flushing_caps mds%d\n", session->s_mds);
2227 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2228 int delayed = __kick_flushing_caps(mdsc, session, ci, false);
2230 spin_lock(&ci->i_ceph_lock);
2231 __cap_delay_requeue(mdsc, ci);
2232 spin_unlock(&ci->i_ceph_lock);
2237 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2238 struct ceph_mds_session *session,
2239 struct inode *inode)
2241 struct ceph_inode_info *ci = ceph_inode(inode);
2242 struct ceph_cap *cap;
2244 spin_lock(&ci->i_ceph_lock);
2245 cap = ci->i_auth_cap;
2246 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2247 ceph_cap_string(ci->i_flushing_caps));
2249 __ceph_flush_snaps(ci, &session, 1);
2251 if (ci->i_flushing_caps) {
2254 spin_lock(&mdsc->cap_dirty_lock);
2255 list_move_tail(&ci->i_flushing_item,
2256 &cap->session->s_cap_flushing);
2257 spin_unlock(&mdsc->cap_dirty_lock);
2259 spin_unlock(&ci->i_ceph_lock);
2261 delayed = __kick_flushing_caps(mdsc, session, ci, true);
2263 spin_lock(&ci->i_ceph_lock);
2264 __cap_delay_requeue(mdsc, ci);
2265 spin_unlock(&ci->i_ceph_lock);
2268 spin_unlock(&ci->i_ceph_lock);
2274 * Take references to capabilities we hold, so that we don't release
2275 * them to the MDS prematurely.
2277 * Protected by i_ceph_lock.
2279 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2280 bool snap_rwsem_locked)
2282 if (got & CEPH_CAP_PIN)
2284 if (got & CEPH_CAP_FILE_RD)
2286 if (got & CEPH_CAP_FILE_CACHE)
2287 ci->i_rdcache_ref++;
2288 if (got & CEPH_CAP_FILE_WR) {
2289 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2290 BUG_ON(!snap_rwsem_locked);
2291 ci->i_head_snapc = ceph_get_snap_context(
2292 ci->i_snap_realm->cached_context);
2296 if (got & CEPH_CAP_FILE_BUFFER) {
2297 if (ci->i_wb_ref == 0)
2298 ihold(&ci->vfs_inode);
2300 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2301 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2306 * Try to grab cap references. Specify those refs we @want, and the
2307 * minimal set we @need. Also include the larger offset we are writing
2308 * to (when applicable), and check against max_size here as well.
2309 * Note that caller is responsible for ensuring max_size increases are
2310 * requested from the MDS.
2312 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2313 loff_t endoff, bool nonblock, int *got, int *err)
2315 struct inode *inode = &ci->vfs_inode;
2316 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2318 int have, implemented;
2320 bool snap_rwsem_locked = false;
2322 dout("get_cap_refs %p need %s want %s\n", inode,
2323 ceph_cap_string(need), ceph_cap_string(want));
2326 spin_lock(&ci->i_ceph_lock);
2328 /* make sure file is actually open */
2329 file_wanted = __ceph_caps_file_wanted(ci);
2330 if ((file_wanted & need) == 0) {
2331 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2332 ceph_cap_string(need), ceph_cap_string(file_wanted));
2338 /* finish pending truncate */
2339 while (ci->i_truncate_pending) {
2340 spin_unlock(&ci->i_ceph_lock);
2341 if (snap_rwsem_locked) {
2342 up_read(&mdsc->snap_rwsem);
2343 snap_rwsem_locked = false;
2345 __ceph_do_pending_vmtruncate(inode);
2346 spin_lock(&ci->i_ceph_lock);
2349 have = __ceph_caps_issued(ci, &implemented);
2351 if (have & need & CEPH_CAP_FILE_WR) {
2352 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2353 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2354 inode, endoff, ci->i_max_size);
2355 if (endoff > ci->i_requested_max_size) {
2362 * If a sync write is in progress, we must wait, so that we
2363 * can get a final snapshot value for size+mtime.
2365 if (__ceph_have_pending_cap_snap(ci)) {
2366 dout("get_cap_refs %p cap_snap_pending\n", inode);
2371 if ((have & need) == need) {
2373 * Look at (implemented & ~have & not) so that we keep waiting
2374 * on transition from wanted -> needed caps. This is needed
2375 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2376 * going before a prior buffered writeback happens.
2378 int not = want & ~(have & need);
2379 int revoking = implemented & ~have;
2380 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2381 inode, ceph_cap_string(have), ceph_cap_string(not),
2382 ceph_cap_string(revoking));
2383 if ((revoking & not) == 0) {
2384 if (!snap_rwsem_locked &&
2385 !ci->i_head_snapc &&
2386 (need & CEPH_CAP_FILE_WR)) {
2387 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2389 * we can not call down_read() when
2390 * task isn't in TASK_RUNNING state
2398 spin_unlock(&ci->i_ceph_lock);
2399 down_read(&mdsc->snap_rwsem);
2400 snap_rwsem_locked = true;
2403 snap_rwsem_locked = true;
2405 *got = need | (have & want);
2406 __take_cap_refs(ci, *got, true);
2410 int session_readonly = false;
2411 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2412 struct ceph_mds_session *s = ci->i_auth_cap->session;
2413 spin_lock(&s->s_cap_lock);
2414 session_readonly = s->s_readonly;
2415 spin_unlock(&s->s_cap_lock);
2417 if (session_readonly) {
2418 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2419 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2425 dout("get_cap_refs %p have %s needed %s\n", inode,
2426 ceph_cap_string(have), ceph_cap_string(need));
2429 spin_unlock(&ci->i_ceph_lock);
2430 if (snap_rwsem_locked)
2431 up_read(&mdsc->snap_rwsem);
2433 dout("get_cap_refs %p ret %d got %s\n", inode,
2434 ret, ceph_cap_string(*got));
2439 * Check the offset we are writing up to against our current
2440 * max_size. If necessary, tell the MDS we want to write to
2443 static void check_max_size(struct inode *inode, loff_t endoff)
2445 struct ceph_inode_info *ci = ceph_inode(inode);
2448 /* do we need to explicitly request a larger max_size? */
2449 spin_lock(&ci->i_ceph_lock);
2450 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2451 dout("write %p at large endoff %llu, req max_size\n",
2453 ci->i_wanted_max_size = endoff;
2455 /* duplicate ceph_check_caps()'s logic */
2456 if (ci->i_auth_cap &&
2457 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2458 ci->i_wanted_max_size > ci->i_max_size &&
2459 ci->i_wanted_max_size > ci->i_requested_max_size)
2461 spin_unlock(&ci->i_ceph_lock);
2463 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2467 * Wait for caps, and take cap references. If we can't get a WR cap
2468 * due to a small max_size, make sure we check_max_size (and possibly
2469 * ask the mds) so we don't get hung up indefinitely.
2471 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2472 loff_t endoff, int *got, struct page **pinned_page)
2474 int _got, ret, err = 0;
2476 ret = ceph_pool_perm_check(ci, need);
2482 check_max_size(&ci->vfs_inode, endoff);
2486 ret = try_get_cap_refs(ci, need, want, endoff,
2487 false, &_got, &err);
2494 ret = wait_event_interruptible(ci->i_cap_wq,
2495 try_get_cap_refs(ci, need, want, endoff,
2496 true, &_got, &err));
2505 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2506 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2507 i_size_read(&ci->vfs_inode) > 0) {
2509 find_get_page(ci->vfs_inode.i_mapping, 0);
2511 if (PageUptodate(page)) {
2512 *pinned_page = page;
2515 page_cache_release(page);
2518 * drop cap refs first because getattr while
2519 * holding * caps refs can cause deadlock.
2521 ceph_put_cap_refs(ci, _got);
2525 * getattr request will bring inline data into
2528 ret = __ceph_do_getattr(&ci->vfs_inode, NULL,
2529 CEPH_STAT_CAP_INLINE_DATA,
2543 * Take cap refs. Caller must already know we hold at least one ref
2544 * on the caps in question or we don't know this is safe.
2546 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2548 spin_lock(&ci->i_ceph_lock);
2549 __take_cap_refs(ci, caps, false);
2550 spin_unlock(&ci->i_ceph_lock);
2555 * drop cap_snap that is not associated with any snapshot.
2556 * we don't need to send FLUSHSNAP message for it.
2558 static int ceph_try_drop_cap_snap(struct ceph_cap_snap *capsnap)
2560 if (!capsnap->need_flush &&
2561 !capsnap->writing && !capsnap->dirty_pages) {
2563 dout("dropping cap_snap %p follows %llu\n",
2564 capsnap, capsnap->follows);
2565 ceph_put_snap_context(capsnap->context);
2566 list_del(&capsnap->ci_item);
2567 list_del(&capsnap->flushing_item);
2568 ceph_put_cap_snap(capsnap);
2577 * If we released the last ref on any given cap, call ceph_check_caps
2578 * to release (or schedule a release).
2580 * If we are releasing a WR cap (from a sync write), finalize any affected
2581 * cap_snap, and wake up any waiters.
2583 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2585 struct inode *inode = &ci->vfs_inode;
2586 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2588 spin_lock(&ci->i_ceph_lock);
2589 if (had & CEPH_CAP_PIN)
2591 if (had & CEPH_CAP_FILE_RD)
2592 if (--ci->i_rd_ref == 0)
2594 if (had & CEPH_CAP_FILE_CACHE)
2595 if (--ci->i_rdcache_ref == 0)
2597 if (had & CEPH_CAP_FILE_BUFFER) {
2598 if (--ci->i_wb_ref == 0) {
2602 dout("put_cap_refs %p wb %d -> %d (?)\n",
2603 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2605 if (had & CEPH_CAP_FILE_WR)
2606 if (--ci->i_wr_ref == 0) {
2608 if (__ceph_have_pending_cap_snap(ci)) {
2609 struct ceph_cap_snap *capsnap =
2610 list_last_entry(&ci->i_cap_snaps,
2611 struct ceph_cap_snap,
2613 capsnap->writing = 0;
2614 if (ceph_try_drop_cap_snap(capsnap))
2616 else if (__ceph_finish_cap_snap(ci, capsnap))
2620 if (ci->i_wrbuffer_ref_head == 0 &&
2621 ci->i_dirty_caps == 0 &&
2622 ci->i_flushing_caps == 0) {
2623 BUG_ON(!ci->i_head_snapc);
2624 ceph_put_snap_context(ci->i_head_snapc);
2625 ci->i_head_snapc = NULL;
2627 /* see comment in __ceph_remove_cap() */
2628 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2629 drop_inode_snap_realm(ci);
2631 spin_unlock(&ci->i_ceph_lock);
2633 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2634 last ? " last" : "", put ? " put" : "");
2636 if (last && !flushsnaps)
2637 ceph_check_caps(ci, 0, NULL);
2638 else if (flushsnaps)
2639 ceph_flush_snaps(ci);
2641 wake_up_all(&ci->i_cap_wq);
2647 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2648 * context. Adjust per-snap dirty page accounting as appropriate.
2649 * Once all dirty data for a cap_snap is flushed, flush snapped file
2650 * metadata back to the MDS. If we dropped the last ref, call
2653 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2654 struct ceph_snap_context *snapc)
2656 struct inode *inode = &ci->vfs_inode;
2658 int complete_capsnap = 0;
2659 int drop_capsnap = 0;
2661 struct ceph_cap_snap *capsnap = NULL;
2663 spin_lock(&ci->i_ceph_lock);
2664 ci->i_wrbuffer_ref -= nr;
2665 last = !ci->i_wrbuffer_ref;
2667 if (ci->i_head_snapc == snapc) {
2668 ci->i_wrbuffer_ref_head -= nr;
2669 if (ci->i_wrbuffer_ref_head == 0 &&
2670 ci->i_wr_ref == 0 &&
2671 ci->i_dirty_caps == 0 &&
2672 ci->i_flushing_caps == 0) {
2673 BUG_ON(!ci->i_head_snapc);
2674 ceph_put_snap_context(ci->i_head_snapc);
2675 ci->i_head_snapc = NULL;
2677 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2679 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2680 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2681 last ? " LAST" : "");
2683 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2684 if (capsnap->context == snapc) {
2690 capsnap->dirty_pages -= nr;
2691 if (capsnap->dirty_pages == 0) {
2692 complete_capsnap = 1;
2693 drop_capsnap = ceph_try_drop_cap_snap(capsnap);
2695 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2696 " snap %lld %d/%d -> %d/%d %s%s\n",
2697 inode, capsnap, capsnap->context->seq,
2698 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2699 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2700 last ? " (wrbuffer last)" : "",
2701 complete_capsnap ? " (complete capsnap)" : "");
2704 spin_unlock(&ci->i_ceph_lock);
2707 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2709 } else if (complete_capsnap) {
2710 ceph_flush_snaps(ci);
2711 wake_up_all(&ci->i_cap_wq);
2718 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2720 static void invalidate_aliases(struct inode *inode)
2722 struct dentry *dn, *prev = NULL;
2724 dout("invalidate_aliases inode %p\n", inode);
2725 d_prune_aliases(inode);
2727 * For non-directory inode, d_find_alias() only returns
2728 * hashed dentry. After calling d_invalidate(), the
2729 * dentry becomes unhashed.
2731 * For directory inode, d_find_alias() can return
2732 * unhashed dentry. But directory inode should have
2733 * one alias at most.
2735 while ((dn = d_find_alias(inode))) {
2750 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2751 * actually be a revocation if it specifies a smaller cap set.)
2753 * caller holds s_mutex and i_ceph_lock, we drop both.
2755 static void handle_cap_grant(struct ceph_mds_client *mdsc,
2756 struct inode *inode, struct ceph_mds_caps *grant,
2758 void *inline_data, int inline_len,
2759 struct ceph_buffer *xattr_buf,
2760 struct ceph_mds_session *session,
2761 struct ceph_cap *cap, int issued)
2762 __releases(ci->i_ceph_lock)
2763 __releases(mdsc->snap_rwsem)
2765 struct ceph_inode_info *ci = ceph_inode(inode);
2766 int mds = session->s_mds;
2767 int seq = le32_to_cpu(grant->seq);
2768 int newcaps = le32_to_cpu(grant->caps);
2769 int used, wanted, dirty;
2770 u64 size = le64_to_cpu(grant->size);
2771 u64 max_size = le64_to_cpu(grant->max_size);
2772 struct timespec mtime, atime, ctime;
2775 bool writeback = false;
2776 bool queue_trunc = false;
2777 bool queue_invalidate = false;
2778 bool queue_revalidate = false;
2779 bool deleted_inode = false;
2780 bool fill_inline = false;
2782 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2783 inode, cap, mds, seq, ceph_cap_string(newcaps));
2784 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2789 * auth mds of the inode changed. we received the cap export message,
2790 * but still haven't received the cap import message. handle_cap_export
2791 * updated the new auth MDS' cap.
2793 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2794 * that was sent before the cap import message. So don't remove caps.
2796 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
2797 WARN_ON(cap != ci->i_auth_cap);
2798 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
2800 newcaps |= cap->issued;
2804 * If CACHE is being revoked, and we have no dirty buffers,
2805 * try to invalidate (once). (If there are dirty buffers, we
2806 * will invalidate _after_ writeback.)
2808 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2809 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2810 !ci->i_wrbuffer_ref) {
2811 if (try_nonblocking_invalidate(inode)) {
2812 /* there were locked pages.. invalidate later
2813 in a separate thread. */
2814 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2815 queue_invalidate = true;
2816 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2820 ceph_fscache_invalidate(inode);
2823 /* side effects now are allowed */
2824 cap->cap_gen = session->s_cap_gen;
2827 __check_cap_issue(ci, cap, newcaps);
2829 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2830 (issued & CEPH_CAP_AUTH_EXCL) == 0) {
2831 inode->i_mode = le32_to_cpu(grant->mode);
2832 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
2833 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
2834 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2835 from_kuid(&init_user_ns, inode->i_uid),
2836 from_kgid(&init_user_ns, inode->i_gid));
2839 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2840 (issued & CEPH_CAP_LINK_EXCL) == 0) {
2841 set_nlink(inode, le32_to_cpu(grant->nlink));
2842 if (inode->i_nlink == 0 &&
2843 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
2844 deleted_inode = true;
2847 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2848 int len = le32_to_cpu(grant->xattr_len);
2849 u64 version = le64_to_cpu(grant->xattr_version);
2851 if (version > ci->i_xattrs.version) {
2852 dout(" got new xattrs v%llu on %p len %d\n",
2853 version, inode, len);
2854 if (ci->i_xattrs.blob)
2855 ceph_buffer_put(ci->i_xattrs.blob);
2856 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2857 ci->i_xattrs.version = version;
2858 ceph_forget_all_cached_acls(inode);
2862 /* Do we need to revalidate our fscache cookie. Don't bother on the
2863 * first cache cap as we already validate at cookie creation time. */
2864 if ((issued & CEPH_CAP_FILE_CACHE) && ci->i_rdcache_gen > 1)
2865 queue_revalidate = true;
2867 if (newcaps & CEPH_CAP_ANY_RD) {
2868 /* ctime/mtime/atime? */
2869 ceph_decode_timespec(&mtime, &grant->mtime);
2870 ceph_decode_timespec(&atime, &grant->atime);
2871 ceph_decode_timespec(&ctime, &grant->ctime);
2872 ceph_fill_file_time(inode, issued,
2873 le32_to_cpu(grant->time_warp_seq),
2874 &ctime, &mtime, &atime);
2877 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
2878 /* file layout may have changed */
2879 ci->i_layout = grant->layout;
2880 /* size/truncate_seq? */
2881 queue_trunc = ceph_fill_file_size(inode, issued,
2882 le32_to_cpu(grant->truncate_seq),
2883 le64_to_cpu(grant->truncate_size),
2885 /* max size increase? */
2886 if (ci->i_auth_cap == cap && max_size != ci->i_max_size) {
2887 dout("max_size %lld -> %llu\n",
2888 ci->i_max_size, max_size);
2889 ci->i_max_size = max_size;
2890 if (max_size >= ci->i_wanted_max_size) {
2891 ci->i_wanted_max_size = 0; /* reset */
2892 ci->i_requested_max_size = 0;
2898 /* check cap bits */
2899 wanted = __ceph_caps_wanted(ci);
2900 used = __ceph_caps_used(ci);
2901 dirty = __ceph_caps_dirty(ci);
2902 dout(" my wanted = %s, used = %s, dirty %s\n",
2903 ceph_cap_string(wanted),
2904 ceph_cap_string(used),
2905 ceph_cap_string(dirty));
2906 if (wanted != le32_to_cpu(grant->wanted)) {
2907 dout("mds wanted %s -> %s\n",
2908 ceph_cap_string(le32_to_cpu(grant->wanted)),
2909 ceph_cap_string(wanted));
2910 /* imported cap may not have correct mds_wanted */
2911 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
2915 /* revocation, grant, or no-op? */
2916 if (cap->issued & ~newcaps) {
2917 int revoking = cap->issued & ~newcaps;
2919 dout("revocation: %s -> %s (revoking %s)\n",
2920 ceph_cap_string(cap->issued),
2921 ceph_cap_string(newcaps),
2922 ceph_cap_string(revoking));
2923 if (revoking & used & CEPH_CAP_FILE_BUFFER)
2924 writeback = true; /* initiate writeback; will delay ack */
2925 else if (revoking == CEPH_CAP_FILE_CACHE &&
2926 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2928 ; /* do nothing yet, invalidation will be queued */
2929 else if (cap == ci->i_auth_cap)
2930 check_caps = 1; /* check auth cap only */
2932 check_caps = 2; /* check all caps */
2933 cap->issued = newcaps;
2934 cap->implemented |= newcaps;
2935 } else if (cap->issued == newcaps) {
2936 dout("caps unchanged: %s -> %s\n",
2937 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2939 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2940 ceph_cap_string(newcaps));
2941 /* non-auth MDS is revoking the newly grant caps ? */
2942 if (cap == ci->i_auth_cap &&
2943 __ceph_caps_revoking_other(ci, cap, newcaps))
2946 cap->issued = newcaps;
2947 cap->implemented |= newcaps; /* add bits only, to
2948 * avoid stepping on a
2949 * pending revocation */
2952 BUG_ON(cap->issued & ~cap->implemented);
2954 if (inline_version > 0 && inline_version >= ci->i_inline_version) {
2955 ci->i_inline_version = inline_version;
2956 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2957 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
2961 spin_unlock(&ci->i_ceph_lock);
2963 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
2964 kick_flushing_inode_caps(mdsc, session, inode);
2965 up_read(&mdsc->snap_rwsem);
2966 if (newcaps & ~issued)
2971 ceph_fill_inline_data(inode, NULL, inline_data, inline_len);
2974 ceph_queue_vmtruncate(inode);
2975 ceph_queue_revalidate(inode);
2976 } else if (queue_revalidate)
2977 ceph_queue_revalidate(inode);
2981 * queue inode for writeback: we can't actually call
2982 * filemap_write_and_wait, etc. from message handler
2985 ceph_queue_writeback(inode);
2986 if (queue_invalidate)
2987 ceph_queue_invalidate(inode);
2989 invalidate_aliases(inode);
2991 wake_up_all(&ci->i_cap_wq);
2993 if (check_caps == 1)
2994 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2996 else if (check_caps == 2)
2997 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
2999 mutex_unlock(&session->s_mutex);
3003 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3004 * MDS has been safely committed.
3006 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3007 struct ceph_mds_caps *m,
3008 struct ceph_mds_session *session,
3009 struct ceph_cap *cap)
3010 __releases(ci->i_ceph_lock)
3012 struct ceph_inode_info *ci = ceph_inode(inode);
3013 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3014 struct ceph_cap_flush *cf;
3016 LIST_HEAD(to_remove);
3017 unsigned seq = le32_to_cpu(m->seq);
3018 int dirty = le32_to_cpu(m->dirty);
3022 n = rb_first(&ci->i_cap_flush_tree);
3024 cf = rb_entry(n, struct ceph_cap_flush, i_node);
3025 n = rb_next(&cf->i_node);
3026 if (cf->tid == flush_tid)
3028 if (cf->tid <= flush_tid) {
3029 rb_erase(&cf->i_node, &ci->i_cap_flush_tree);
3030 list_add_tail(&cf->list, &to_remove);
3032 cleaned &= ~cf->caps;
3038 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3039 " flushing %s -> %s\n",
3040 inode, session->s_mds, seq, ceph_cap_string(dirty),
3041 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3042 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3044 if (list_empty(&to_remove) && !cleaned)
3047 ci->i_flushing_caps &= ~cleaned;
3049 spin_lock(&mdsc->cap_dirty_lock);
3051 if (!list_empty(&to_remove)) {
3052 list_for_each_entry(cf, &to_remove, list)
3053 rb_erase(&cf->g_node, &mdsc->cap_flush_tree);
3055 n = rb_first(&mdsc->cap_flush_tree);
3056 cf = n ? rb_entry(n, struct ceph_cap_flush, g_node) : NULL;
3057 if (!cf || cf->tid > flush_tid)
3058 wake_up_all(&mdsc->cap_flushing_wq);
3061 if (ci->i_flushing_caps == 0) {
3062 list_del_init(&ci->i_flushing_item);
3063 if (!list_empty(&session->s_cap_flushing))
3064 dout(" mds%d still flushing cap on %p\n",
3066 &list_entry(session->s_cap_flushing.next,
3067 struct ceph_inode_info,
3068 i_flushing_item)->vfs_inode);
3069 mdsc->num_cap_flushing--;
3070 dout(" inode %p now !flushing\n", inode);
3072 if (ci->i_dirty_caps == 0) {
3073 dout(" inode %p now clean\n", inode);
3074 BUG_ON(!list_empty(&ci->i_dirty_item));
3076 if (ci->i_wr_ref == 0 &&
3077 ci->i_wrbuffer_ref_head == 0) {
3078 BUG_ON(!ci->i_head_snapc);
3079 ceph_put_snap_context(ci->i_head_snapc);
3080 ci->i_head_snapc = NULL;
3083 BUG_ON(list_empty(&ci->i_dirty_item));
3086 spin_unlock(&mdsc->cap_dirty_lock);
3087 wake_up_all(&ci->i_cap_wq);
3090 spin_unlock(&ci->i_ceph_lock);
3092 while (!list_empty(&to_remove)) {
3093 cf = list_first_entry(&to_remove,
3094 struct ceph_cap_flush, list);
3095 list_del(&cf->list);
3096 ceph_free_cap_flush(cf);
3103 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3104 * throw away our cap_snap.
3106 * Caller hold s_mutex.
3108 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3109 struct ceph_mds_caps *m,
3110 struct ceph_mds_session *session)
3112 struct ceph_inode_info *ci = ceph_inode(inode);
3113 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3114 u64 follows = le64_to_cpu(m->snap_follows);
3115 struct ceph_cap_snap *capsnap;
3118 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3119 inode, ci, session->s_mds, follows);
3121 spin_lock(&ci->i_ceph_lock);
3122 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3123 if (capsnap->follows == follows) {
3124 if (capsnap->flush_tid != flush_tid) {
3125 dout(" cap_snap %p follows %lld tid %lld !="
3126 " %lld\n", capsnap, follows,
3127 flush_tid, capsnap->flush_tid);
3130 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3131 dout(" removing %p cap_snap %p follows %lld\n",
3132 inode, capsnap, follows);
3133 ceph_put_snap_context(capsnap->context);
3134 list_del(&capsnap->ci_item);
3135 list_del(&capsnap->flushing_item);
3136 ceph_put_cap_snap(capsnap);
3137 wake_up_all(&mdsc->cap_flushing_wq);
3141 dout(" skipping cap_snap %p follows %lld\n",
3142 capsnap, capsnap->follows);
3145 spin_unlock(&ci->i_ceph_lock);
3151 * Handle TRUNC from MDS, indicating file truncation.
3153 * caller hold s_mutex.
3155 static void handle_cap_trunc(struct inode *inode,
3156 struct ceph_mds_caps *trunc,
3157 struct ceph_mds_session *session)
3158 __releases(ci->i_ceph_lock)
3160 struct ceph_inode_info *ci = ceph_inode(inode);
3161 int mds = session->s_mds;
3162 int seq = le32_to_cpu(trunc->seq);
3163 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3164 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3165 u64 size = le64_to_cpu(trunc->size);
3166 int implemented = 0;
3167 int dirty = __ceph_caps_dirty(ci);
3168 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3169 int queue_trunc = 0;
3171 issued |= implemented | dirty;
3173 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3174 inode, mds, seq, truncate_size, truncate_seq);
3175 queue_trunc = ceph_fill_file_size(inode, issued,
3176 truncate_seq, truncate_size, size);
3177 spin_unlock(&ci->i_ceph_lock);
3180 ceph_queue_vmtruncate(inode);
3181 ceph_fscache_invalidate(inode);
3186 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3187 * different one. If we are the most recent migration we've seen (as
3188 * indicated by mseq), make note of the migrating cap bits for the
3189 * duration (until we see the corresponding IMPORT).
3191 * caller holds s_mutex
3193 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3194 struct ceph_mds_cap_peer *ph,
3195 struct ceph_mds_session *session)
3197 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3198 struct ceph_mds_session *tsession = NULL;
3199 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3200 struct ceph_inode_info *ci = ceph_inode(inode);
3202 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3203 unsigned t_seq, t_mseq;
3205 int mds = session->s_mds;
3208 t_cap_id = le64_to_cpu(ph->cap_id);
3209 t_seq = le32_to_cpu(ph->seq);
3210 t_mseq = le32_to_cpu(ph->mseq);
3211 target = le32_to_cpu(ph->mds);
3213 t_cap_id = t_seq = t_mseq = 0;
3217 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3218 inode, ci, mds, mseq, target);
3220 spin_lock(&ci->i_ceph_lock);
3221 cap = __get_cap_for_mds(ci, mds);
3222 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3226 __ceph_remove_cap(cap, false);
3231 * now we know we haven't received the cap import message yet
3232 * because the exported cap still exist.
3235 issued = cap->issued;
3236 WARN_ON(issued != cap->implemented);
3238 tcap = __get_cap_for_mds(ci, target);
3240 /* already have caps from the target */
3241 if (tcap->cap_id != t_cap_id ||
3242 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3243 dout(" updating import cap %p mds%d\n", tcap, target);
3244 tcap->cap_id = t_cap_id;
3245 tcap->seq = t_seq - 1;
3246 tcap->issue_seq = t_seq - 1;
3247 tcap->mseq = t_mseq;
3248 tcap->issued |= issued;
3249 tcap->implemented |= issued;
3250 if (cap == ci->i_auth_cap)
3251 ci->i_auth_cap = tcap;
3252 if (ci->i_flushing_caps && ci->i_auth_cap == tcap) {
3253 spin_lock(&mdsc->cap_dirty_lock);
3254 list_move_tail(&ci->i_flushing_item,
3255 &tcap->session->s_cap_flushing);
3256 spin_unlock(&mdsc->cap_dirty_lock);
3259 __ceph_remove_cap(cap, false);
3261 } else if (tsession) {
3262 /* add placeholder for the export tagert */
3263 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3264 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3265 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3267 __ceph_remove_cap(cap, false);
3271 spin_unlock(&ci->i_ceph_lock);
3272 mutex_unlock(&session->s_mutex);
3274 /* open target session */
3275 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3276 if (!IS_ERR(tsession)) {
3278 mutex_lock(&session->s_mutex);
3279 mutex_lock_nested(&tsession->s_mutex,
3280 SINGLE_DEPTH_NESTING);
3282 mutex_lock(&tsession->s_mutex);
3283 mutex_lock_nested(&session->s_mutex,
3284 SINGLE_DEPTH_NESTING);
3286 new_cap = ceph_get_cap(mdsc, NULL);
3295 spin_unlock(&ci->i_ceph_lock);
3296 mutex_unlock(&session->s_mutex);
3298 mutex_unlock(&tsession->s_mutex);
3299 ceph_put_mds_session(tsession);
3302 ceph_put_cap(mdsc, new_cap);
3306 * Handle cap IMPORT.
3308 * caller holds s_mutex. acquires i_ceph_lock
3310 static void handle_cap_import(struct ceph_mds_client *mdsc,
3311 struct inode *inode, struct ceph_mds_caps *im,
3312 struct ceph_mds_cap_peer *ph,
3313 struct ceph_mds_session *session,
3314 struct ceph_cap **target_cap, int *old_issued)
3315 __acquires(ci->i_ceph_lock)
3317 struct ceph_inode_info *ci = ceph_inode(inode);
3318 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3319 int mds = session->s_mds;
3321 unsigned caps = le32_to_cpu(im->caps);
3322 unsigned wanted = le32_to_cpu(im->wanted);
3323 unsigned seq = le32_to_cpu(im->seq);
3324 unsigned mseq = le32_to_cpu(im->migrate_seq);
3325 u64 realmino = le64_to_cpu(im->realm);
3326 u64 cap_id = le64_to_cpu(im->cap_id);
3331 p_cap_id = le64_to_cpu(ph->cap_id);
3332 peer = le32_to_cpu(ph->mds);
3338 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3339 inode, ci, mds, mseq, peer);
3342 spin_lock(&ci->i_ceph_lock);
3343 cap = __get_cap_for_mds(ci, mds);
3346 spin_unlock(&ci->i_ceph_lock);
3347 new_cap = ceph_get_cap(mdsc, NULL);
3353 ceph_put_cap(mdsc, new_cap);
3358 __ceph_caps_issued(ci, &issued);
3359 issued |= __ceph_caps_dirty(ci);
3361 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3362 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3364 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3365 if (ocap && ocap->cap_id == p_cap_id) {
3366 dout(" remove export cap %p mds%d flags %d\n",
3367 ocap, peer, ph->flags);
3368 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3369 (ocap->seq != le32_to_cpu(ph->seq) ||
3370 ocap->mseq != le32_to_cpu(ph->mseq))) {
3371 pr_err("handle_cap_import: mismatched seq/mseq: "
3372 "ino (%llx.%llx) mds%d seq %d mseq %d "
3373 "importer mds%d has peer seq %d mseq %d\n",
3374 ceph_vinop(inode), peer, ocap->seq,
3375 ocap->mseq, mds, le32_to_cpu(ph->seq),
3376 le32_to_cpu(ph->mseq));
3378 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3381 /* make sure we re-request max_size, if necessary */
3382 ci->i_wanted_max_size = 0;
3383 ci->i_requested_max_size = 0;
3385 *old_issued = issued;
3390 * Handle a caps message from the MDS.
3392 * Identify the appropriate session, inode, and call the right handler
3393 * based on the cap op.
3395 void ceph_handle_caps(struct ceph_mds_session *session,
3396 struct ceph_msg *msg)
3398 struct ceph_mds_client *mdsc = session->s_mdsc;
3399 struct super_block *sb = mdsc->fsc->sb;
3400 struct inode *inode;
3401 struct ceph_inode_info *ci;
3402 struct ceph_cap *cap;
3403 struct ceph_mds_caps *h;
3404 struct ceph_mds_cap_peer *peer = NULL;
3405 struct ceph_snap_realm *realm;
3406 int mds = session->s_mds;
3409 struct ceph_vino vino;
3413 u64 inline_version = 0;
3414 void *inline_data = NULL;
3417 size_t snaptrace_len;
3420 dout("handle_caps from mds%d\n", mds);
3423 end = msg->front.iov_base + msg->front.iov_len;
3424 tid = le64_to_cpu(msg->hdr.tid);
3425 if (msg->front.iov_len < sizeof(*h))
3427 h = msg->front.iov_base;
3428 op = le32_to_cpu(h->op);
3429 vino.ino = le64_to_cpu(h->ino);
3430 vino.snap = CEPH_NOSNAP;
3431 cap_id = le64_to_cpu(h->cap_id);
3432 seq = le32_to_cpu(h->seq);
3433 mseq = le32_to_cpu(h->migrate_seq);
3434 size = le64_to_cpu(h->size);
3435 max_size = le64_to_cpu(h->max_size);
3438 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3439 p = snaptrace + snaptrace_len;
3441 if (le16_to_cpu(msg->hdr.version) >= 2) {
3443 ceph_decode_32_safe(&p, end, flock_len, bad);
3444 if (p + flock_len > end)
3449 if (le16_to_cpu(msg->hdr.version) >= 3) {
3450 if (op == CEPH_CAP_OP_IMPORT) {
3451 if (p + sizeof(*peer) > end)
3455 } else if (op == CEPH_CAP_OP_EXPORT) {
3456 /* recorded in unused fields */
3457 peer = (void *)&h->size;
3461 if (le16_to_cpu(msg->hdr.version) >= 4) {
3462 ceph_decode_64_safe(&p, end, inline_version, bad);
3463 ceph_decode_32_safe(&p, end, inline_len, bad);
3464 if (p + inline_len > end)
3471 inode = ceph_find_inode(sb, vino);
3472 ci = ceph_inode(inode);
3473 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3476 mutex_lock(&session->s_mutex);
3478 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3482 dout(" i don't have ino %llx\n", vino.ino);
3484 if (op == CEPH_CAP_OP_IMPORT) {
3485 cap = ceph_get_cap(mdsc, NULL);
3486 cap->cap_ino = vino.ino;
3487 cap->queue_release = 1;
3488 cap->cap_id = cap_id;
3491 spin_lock(&session->s_cap_lock);
3492 list_add_tail(&cap->session_caps,
3493 &session->s_cap_releases);
3494 session->s_num_cap_releases++;
3495 spin_unlock(&session->s_cap_lock);
3497 goto flush_cap_releases;
3500 /* these will work even if we don't have a cap yet */
3502 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3503 handle_cap_flushsnap_ack(inode, tid, h, session);
3506 case CEPH_CAP_OP_EXPORT:
3507 handle_cap_export(inode, h, peer, session);
3510 case CEPH_CAP_OP_IMPORT:
3512 if (snaptrace_len) {
3513 down_write(&mdsc->snap_rwsem);
3514 ceph_update_snap_trace(mdsc, snaptrace,
3515 snaptrace + snaptrace_len,
3517 downgrade_write(&mdsc->snap_rwsem);
3519 down_read(&mdsc->snap_rwsem);
3521 handle_cap_import(mdsc, inode, h, peer, session,
3523 handle_cap_grant(mdsc, inode, h,
3524 inline_version, inline_data, inline_len,
3525 msg->middle, session, cap, issued);
3527 ceph_put_snap_realm(mdsc, realm);
3531 /* the rest require a cap */
3532 spin_lock(&ci->i_ceph_lock);
3533 cap = __get_cap_for_mds(ceph_inode(inode), mds);
3535 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3536 inode, ceph_ino(inode), ceph_snap(inode), mds);
3537 spin_unlock(&ci->i_ceph_lock);
3538 goto flush_cap_releases;
3541 /* note that each of these drops i_ceph_lock for us */
3543 case CEPH_CAP_OP_REVOKE:
3544 case CEPH_CAP_OP_GRANT:
3545 __ceph_caps_issued(ci, &issued);
3546 issued |= __ceph_caps_dirty(ci);
3547 handle_cap_grant(mdsc, inode, h,
3548 inline_version, inline_data, inline_len,
3549 msg->middle, session, cap, issued);
3552 case CEPH_CAP_OP_FLUSH_ACK:
3553 handle_cap_flush_ack(inode, tid, h, session, cap);
3556 case CEPH_CAP_OP_TRUNC:
3557 handle_cap_trunc(inode, h, session);
3561 spin_unlock(&ci->i_ceph_lock);
3562 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3563 ceph_cap_op_name(op));
3570 * send any cap release message to try to move things
3571 * along for the mds (who clearly thinks we still have this
3574 ceph_send_cap_releases(mdsc, session);
3577 mutex_unlock(&session->s_mutex);
3583 pr_err("ceph_handle_caps: corrupt message\n");
3589 * Delayed work handler to process end of delayed cap release LRU list.
3591 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3593 struct ceph_inode_info *ci;
3594 int flags = CHECK_CAPS_NODELAY;
3596 dout("check_delayed_caps\n");
3598 spin_lock(&mdsc->cap_delay_lock);
3599 if (list_empty(&mdsc->cap_delay_list))
3601 ci = list_first_entry(&mdsc->cap_delay_list,
3602 struct ceph_inode_info,
3604 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3605 time_before(jiffies, ci->i_hold_caps_max))
3607 list_del_init(&ci->i_cap_delay_list);
3608 spin_unlock(&mdsc->cap_delay_lock);
3609 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
3610 ceph_check_caps(ci, flags, NULL);
3612 spin_unlock(&mdsc->cap_delay_lock);
3616 * Flush all dirty caps to the mds
3618 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
3620 struct ceph_inode_info *ci;
3621 struct inode *inode;
3623 dout("flush_dirty_caps\n");
3624 spin_lock(&mdsc->cap_dirty_lock);
3625 while (!list_empty(&mdsc->cap_dirty)) {
3626 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
3628 inode = &ci->vfs_inode;
3630 dout("flush_dirty_caps %p\n", inode);
3631 spin_unlock(&mdsc->cap_dirty_lock);
3632 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
3634 spin_lock(&mdsc->cap_dirty_lock);
3636 spin_unlock(&mdsc->cap_dirty_lock);
3637 dout("flush_dirty_caps done\n");
3641 * Drop open file reference. If we were the last open file,
3642 * we may need to release capabilities to the MDS (or schedule
3643 * their delayed release).
3645 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
3647 struct inode *inode = &ci->vfs_inode;
3650 spin_lock(&ci->i_ceph_lock);
3651 dout("put_fmode %p fmode %d %d -> %d\n", inode, fmode,
3652 ci->i_nr_by_mode[fmode], ci->i_nr_by_mode[fmode]-1);
3653 BUG_ON(ci->i_nr_by_mode[fmode] == 0);
3654 if (--ci->i_nr_by_mode[fmode] == 0)
3656 spin_unlock(&ci->i_ceph_lock);
3658 if (last && ci->i_vino.snap == CEPH_NOSNAP)
3659 ceph_check_caps(ci, 0, NULL);
3663 * Helpers for embedding cap and dentry lease releases into mds
3666 * @force is used by dentry_release (below) to force inclusion of a
3667 * record for the directory inode, even when there aren't any caps to
3670 int ceph_encode_inode_release(void **p, struct inode *inode,
3671 int mds, int drop, int unless, int force)
3673 struct ceph_inode_info *ci = ceph_inode(inode);
3674 struct ceph_cap *cap;
3675 struct ceph_mds_request_release *rel = *p;
3679 spin_lock(&ci->i_ceph_lock);
3680 used = __ceph_caps_used(ci);
3681 dirty = __ceph_caps_dirty(ci);
3683 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3684 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3685 ceph_cap_string(unless));
3687 /* only drop unused, clean caps */
3688 drop &= ~(used | dirty);
3690 cap = __get_cap_for_mds(ci, mds);
3691 if (cap && __cap_is_valid(cap)) {
3693 ((cap->issued & drop) &&
3694 (cap->issued & unless) == 0)) {
3695 if ((cap->issued & drop) &&
3696 (cap->issued & unless) == 0) {
3697 int wanted = __ceph_caps_wanted(ci);
3698 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
3699 wanted |= cap->mds_wanted;
3700 dout("encode_inode_release %p cap %p "
3701 "%s -> %s, wanted %s -> %s\n", inode, cap,
3702 ceph_cap_string(cap->issued),
3703 ceph_cap_string(cap->issued & ~drop),
3704 ceph_cap_string(cap->mds_wanted),
3705 ceph_cap_string(wanted));
3707 cap->issued &= ~drop;
3708 cap->implemented &= ~drop;
3709 cap->mds_wanted = wanted;
3711 dout("encode_inode_release %p cap %p %s"
3712 " (force)\n", inode, cap,
3713 ceph_cap_string(cap->issued));
3716 rel->ino = cpu_to_le64(ceph_ino(inode));
3717 rel->cap_id = cpu_to_le64(cap->cap_id);
3718 rel->seq = cpu_to_le32(cap->seq);
3719 rel->issue_seq = cpu_to_le32(cap->issue_seq);
3720 rel->mseq = cpu_to_le32(cap->mseq);
3721 rel->caps = cpu_to_le32(cap->implemented);
3722 rel->wanted = cpu_to_le32(cap->mds_wanted);
3728 dout("encode_inode_release %p cap %p %s\n",
3729 inode, cap, ceph_cap_string(cap->issued));
3732 spin_unlock(&ci->i_ceph_lock);
3736 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3737 int mds, int drop, int unless)
3739 struct inode *dir = d_inode(dentry->d_parent);
3740 struct ceph_mds_request_release *rel = *p;
3741 struct ceph_dentry_info *di = ceph_dentry(dentry);
3746 * force an record for the directory caps if we have a dentry lease.
3747 * this is racy (can't take i_ceph_lock and d_lock together), but it
3748 * doesn't have to be perfect; the mds will revoke anything we don't
3751 spin_lock(&dentry->d_lock);
3752 if (di->lease_session && di->lease_session->s_mds == mds)
3754 spin_unlock(&dentry->d_lock);
3756 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3758 spin_lock(&dentry->d_lock);
3759 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3760 dout("encode_dentry_release %p mds%d seq %d\n",
3761 dentry, mds, (int)di->lease_seq);
3762 rel->dname_len = cpu_to_le32(dentry->d_name.len);
3763 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3764 *p += dentry->d_name.len;
3765 rel->dname_seq = cpu_to_le32(di->lease_seq);
3766 __ceph_mdsc_drop_dentry_lease(dentry);
3768 spin_unlock(&dentry->d_lock);