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.
43 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc);
46 * Generate readable cap strings for debugging output.
48 #define MAX_CAP_STR 20
49 static char cap_str[MAX_CAP_STR][40];
50 static DEFINE_SPINLOCK(cap_str_lock);
51 static int last_cap_str;
53 static char *gcap_string(char *s, int c)
55 if (c & CEPH_CAP_GSHARED)
57 if (c & CEPH_CAP_GEXCL)
59 if (c & CEPH_CAP_GCACHE)
65 if (c & CEPH_CAP_GBUFFER)
67 if (c & CEPH_CAP_GLAZYIO)
72 const char *ceph_cap_string(int caps)
78 spin_lock(&cap_str_lock);
80 if (last_cap_str == MAX_CAP_STR)
82 spin_unlock(&cap_str_lock);
86 if (caps & CEPH_CAP_PIN)
89 c = (caps >> CEPH_CAP_SAUTH) & 3;
92 s = gcap_string(s, c);
95 c = (caps >> CEPH_CAP_SLINK) & 3;
98 s = gcap_string(s, c);
101 c = (caps >> CEPH_CAP_SXATTR) & 3;
104 s = gcap_string(s, c);
107 c = caps >> CEPH_CAP_SFILE;
110 s = gcap_string(s, c);
119 void ceph_caps_init(struct ceph_mds_client *mdsc)
121 INIT_LIST_HEAD(&mdsc->caps_list);
122 spin_lock_init(&mdsc->caps_list_lock);
125 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
127 struct ceph_cap *cap;
129 spin_lock(&mdsc->caps_list_lock);
130 while (!list_empty(&mdsc->caps_list)) {
131 cap = list_first_entry(&mdsc->caps_list,
132 struct ceph_cap, caps_item);
133 list_del(&cap->caps_item);
134 kmem_cache_free(ceph_cap_cachep, cap);
136 mdsc->caps_total_count = 0;
137 mdsc->caps_avail_count = 0;
138 mdsc->caps_use_count = 0;
139 mdsc->caps_reserve_count = 0;
140 mdsc->caps_min_count = 0;
141 spin_unlock(&mdsc->caps_list_lock);
144 void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta)
146 spin_lock(&mdsc->caps_list_lock);
147 mdsc->caps_min_count += delta;
148 BUG_ON(mdsc->caps_min_count < 0);
149 spin_unlock(&mdsc->caps_list_lock);
152 void ceph_reserve_caps(struct ceph_mds_client *mdsc,
153 struct ceph_cap_reservation *ctx, int need)
156 struct ceph_cap *cap;
161 dout("reserve caps ctx=%p need=%d\n", ctx, need);
163 /* first reserve any caps that are already allocated */
164 spin_lock(&mdsc->caps_list_lock);
165 if (mdsc->caps_avail_count >= need)
168 have = mdsc->caps_avail_count;
169 mdsc->caps_avail_count -= have;
170 mdsc->caps_reserve_count += have;
171 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
172 mdsc->caps_reserve_count +
173 mdsc->caps_avail_count);
174 spin_unlock(&mdsc->caps_list_lock);
176 for (i = have; i < need; i++) {
177 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
180 list_add(&cap->caps_item, &newcaps);
183 /* we didn't manage to reserve as much as we needed */
184 if (have + alloc != need)
185 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
186 ctx, need, have + alloc);
188 spin_lock(&mdsc->caps_list_lock);
189 mdsc->caps_total_count += alloc;
190 mdsc->caps_reserve_count += alloc;
191 list_splice(&newcaps, &mdsc->caps_list);
193 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
194 mdsc->caps_reserve_count +
195 mdsc->caps_avail_count);
196 spin_unlock(&mdsc->caps_list_lock);
199 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
200 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
201 mdsc->caps_reserve_count, mdsc->caps_avail_count);
204 int ceph_unreserve_caps(struct ceph_mds_client *mdsc,
205 struct ceph_cap_reservation *ctx)
207 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
209 spin_lock(&mdsc->caps_list_lock);
210 BUG_ON(mdsc->caps_reserve_count < ctx->count);
211 mdsc->caps_reserve_count -= ctx->count;
212 mdsc->caps_avail_count += ctx->count;
214 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
215 mdsc->caps_total_count, mdsc->caps_use_count,
216 mdsc->caps_reserve_count, mdsc->caps_avail_count);
217 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
218 mdsc->caps_reserve_count +
219 mdsc->caps_avail_count);
220 spin_unlock(&mdsc->caps_list_lock);
225 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
226 struct ceph_cap_reservation *ctx)
228 struct ceph_cap *cap = NULL;
230 /* temporary, until we do something about cap import/export */
232 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
234 spin_lock(&mdsc->caps_list_lock);
235 mdsc->caps_use_count++;
236 mdsc->caps_total_count++;
237 spin_unlock(&mdsc->caps_list_lock);
242 spin_lock(&mdsc->caps_list_lock);
243 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
244 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
245 mdsc->caps_reserve_count, mdsc->caps_avail_count);
247 BUG_ON(ctx->count > mdsc->caps_reserve_count);
248 BUG_ON(list_empty(&mdsc->caps_list));
251 mdsc->caps_reserve_count--;
252 mdsc->caps_use_count++;
254 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
255 list_del(&cap->caps_item);
257 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
258 mdsc->caps_reserve_count + mdsc->caps_avail_count);
259 spin_unlock(&mdsc->caps_list_lock);
263 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
265 spin_lock(&mdsc->caps_list_lock);
266 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
267 cap, mdsc->caps_total_count, mdsc->caps_use_count,
268 mdsc->caps_reserve_count, mdsc->caps_avail_count);
269 mdsc->caps_use_count--;
271 * Keep some preallocated caps around (ceph_min_count), to
272 * avoid lots of free/alloc churn.
274 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
275 mdsc->caps_min_count) {
276 mdsc->caps_total_count--;
277 kmem_cache_free(ceph_cap_cachep, cap);
279 mdsc->caps_avail_count++;
280 list_add(&cap->caps_item, &mdsc->caps_list);
283 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
284 mdsc->caps_reserve_count + mdsc->caps_avail_count);
285 spin_unlock(&mdsc->caps_list_lock);
288 void ceph_reservation_status(struct ceph_fs_client *fsc,
289 int *total, int *avail, int *used, int *reserved,
292 struct ceph_mds_client *mdsc = fsc->mdsc;
295 *total = mdsc->caps_total_count;
297 *avail = mdsc->caps_avail_count;
299 *used = mdsc->caps_use_count;
301 *reserved = mdsc->caps_reserve_count;
303 *min = mdsc->caps_min_count;
307 * Find ceph_cap for given mds, if any.
309 * Called with i_ceph_lock held.
311 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
313 struct ceph_cap *cap;
314 struct rb_node *n = ci->i_caps.rb_node;
317 cap = rb_entry(n, struct ceph_cap, ci_node);
320 else if (mds > cap->mds)
328 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
330 struct ceph_cap *cap;
332 spin_lock(&ci->i_ceph_lock);
333 cap = __get_cap_for_mds(ci, mds);
334 spin_unlock(&ci->i_ceph_lock);
339 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
341 static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
343 struct ceph_cap *cap;
347 /* prefer mds with WR|BUFFER|EXCL caps */
348 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
349 cap = rb_entry(p, struct ceph_cap, ci_node);
351 if (cap->issued & (CEPH_CAP_FILE_WR |
352 CEPH_CAP_FILE_BUFFER |
359 int ceph_get_cap_mds(struct inode *inode)
361 struct ceph_inode_info *ci = ceph_inode(inode);
363 spin_lock(&ci->i_ceph_lock);
364 mds = __ceph_get_cap_mds(ceph_inode(inode));
365 spin_unlock(&ci->i_ceph_lock);
370 * Called under i_ceph_lock.
372 static void __insert_cap_node(struct ceph_inode_info *ci,
373 struct ceph_cap *new)
375 struct rb_node **p = &ci->i_caps.rb_node;
376 struct rb_node *parent = NULL;
377 struct ceph_cap *cap = NULL;
381 cap = rb_entry(parent, struct ceph_cap, ci_node);
382 if (new->mds < cap->mds)
384 else if (new->mds > cap->mds)
390 rb_link_node(&new->ci_node, parent, p);
391 rb_insert_color(&new->ci_node, &ci->i_caps);
395 * (re)set cap hold timeouts, which control the delayed release
396 * of unused caps back to the MDS. Should be called on cap use.
398 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
399 struct ceph_inode_info *ci)
401 struct ceph_mount_options *ma = mdsc->fsc->mount_options;
403 ci->i_hold_caps_min = round_jiffies(jiffies +
404 ma->caps_wanted_delay_min * HZ);
405 ci->i_hold_caps_max = round_jiffies(jiffies +
406 ma->caps_wanted_delay_max * HZ);
407 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
408 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
412 * (Re)queue cap at the end of the delayed cap release list.
414 * If I_FLUSH is set, leave the inode at the front of the list.
416 * Caller holds i_ceph_lock
417 * -> we take mdsc->cap_delay_lock
419 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
420 struct ceph_inode_info *ci)
422 __cap_set_timeouts(mdsc, ci);
423 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
424 ci->i_ceph_flags, ci->i_hold_caps_max);
425 if (!mdsc->stopping) {
426 spin_lock(&mdsc->cap_delay_lock);
427 if (!list_empty(&ci->i_cap_delay_list)) {
428 if (ci->i_ceph_flags & CEPH_I_FLUSH)
430 list_del_init(&ci->i_cap_delay_list);
432 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
434 spin_unlock(&mdsc->cap_delay_lock);
439 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
440 * indicating we should send a cap message to flush dirty metadata
441 * asap, and move to the front of the delayed cap list.
443 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
444 struct ceph_inode_info *ci)
446 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
447 spin_lock(&mdsc->cap_delay_lock);
448 ci->i_ceph_flags |= CEPH_I_FLUSH;
449 if (!list_empty(&ci->i_cap_delay_list))
450 list_del_init(&ci->i_cap_delay_list);
451 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
452 spin_unlock(&mdsc->cap_delay_lock);
456 * Cancel delayed work on cap.
458 * Caller must hold i_ceph_lock.
460 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
461 struct ceph_inode_info *ci)
463 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
464 if (list_empty(&ci->i_cap_delay_list))
466 spin_lock(&mdsc->cap_delay_lock);
467 list_del_init(&ci->i_cap_delay_list);
468 spin_unlock(&mdsc->cap_delay_lock);
472 * Common issue checks for add_cap, handle_cap_grant.
474 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
477 unsigned had = __ceph_caps_issued(ci, NULL);
480 * Each time we receive FILE_CACHE anew, we increment
483 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
484 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
489 * if we are newly issued FILE_SHARED, mark dir not complete; we
490 * don't know what happened to this directory while we didn't
493 if ((issued & CEPH_CAP_FILE_SHARED) &&
494 (had & CEPH_CAP_FILE_SHARED) == 0) {
496 if (S_ISDIR(ci->vfs_inode.i_mode)) {
497 dout(" marking %p NOT complete\n", &ci->vfs_inode);
498 __ceph_dir_clear_complete(ci);
504 * Add a capability under the given MDS session.
506 * Caller should hold session snap_rwsem (read) and s_mutex.
508 * @fmode is the open file mode, if we are opening a file, otherwise
509 * it is < 0. (This is so we can atomically add the cap and add an
510 * open file reference to it.)
512 void ceph_add_cap(struct inode *inode,
513 struct ceph_mds_session *session, u64 cap_id,
514 int fmode, unsigned issued, unsigned wanted,
515 unsigned seq, unsigned mseq, u64 realmino, int flags,
516 struct ceph_cap **new_cap)
518 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
519 struct ceph_inode_info *ci = ceph_inode(inode);
520 struct ceph_cap *cap;
521 int mds = session->s_mds;
524 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
525 session->s_mds, cap_id, ceph_cap_string(issued), seq);
528 * If we are opening the file, include file mode wanted bits
532 wanted |= ceph_caps_for_mode(fmode);
534 cap = __get_cap_for_mds(ci, mds);
540 cap->implemented = 0;
546 __insert_cap_node(ci, cap);
548 /* add to session cap list */
549 cap->session = session;
550 spin_lock(&session->s_cap_lock);
551 list_add_tail(&cap->session_caps, &session->s_caps);
552 session->s_nr_caps++;
553 spin_unlock(&session->s_cap_lock);
556 * auth mds of the inode changed. we received the cap export
557 * message, but still haven't received the cap import message.
558 * handle_cap_export() updated the new auth MDS' cap.
560 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
561 * a message that was send before the cap import message. So
564 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
565 WARN_ON(cap != ci->i_auth_cap);
566 WARN_ON(cap->cap_id != cap_id);
569 issued |= cap->issued;
570 flags |= CEPH_CAP_FLAG_AUTH;
574 if (!ci->i_snap_realm) {
576 * add this inode to the appropriate snap realm
578 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
581 spin_lock(&realm->inodes_with_caps_lock);
582 ci->i_snap_realm = realm;
583 list_add(&ci->i_snap_realm_item,
584 &realm->inodes_with_caps);
585 spin_unlock(&realm->inodes_with_caps_lock);
587 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
593 __check_cap_issue(ci, cap, issued);
596 * If we are issued caps we don't want, or the mds' wanted
597 * value appears to be off, queue a check so we'll release
598 * later and/or update the mds wanted value.
600 actual_wanted = __ceph_caps_wanted(ci);
601 if ((wanted & ~actual_wanted) ||
602 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
603 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
604 ceph_cap_string(issued), ceph_cap_string(wanted),
605 ceph_cap_string(actual_wanted));
606 __cap_delay_requeue(mdsc, ci);
609 if (flags & CEPH_CAP_FLAG_AUTH) {
610 if (ci->i_auth_cap == NULL ||
611 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
612 ci->i_auth_cap = cap;
613 cap->mds_wanted = wanted;
616 WARN_ON(ci->i_auth_cap == cap);
619 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
620 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
621 ceph_cap_string(issued|cap->issued), seq, mds);
622 cap->cap_id = cap_id;
623 cap->issued = issued;
624 cap->implemented |= issued;
625 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
626 cap->mds_wanted = wanted;
628 cap->mds_wanted |= wanted;
630 cap->issue_seq = seq;
632 cap->cap_gen = session->s_cap_gen;
635 __ceph_get_fmode(ci, fmode);
639 * Return true if cap has not timed out and belongs to the current
640 * generation of the MDS session (i.e. has not gone 'stale' due to
641 * us losing touch with the mds).
643 static int __cap_is_valid(struct ceph_cap *cap)
648 spin_lock(&cap->session->s_gen_ttl_lock);
649 gen = cap->session->s_cap_gen;
650 ttl = cap->session->s_cap_ttl;
651 spin_unlock(&cap->session->s_gen_ttl_lock);
653 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
654 dout("__cap_is_valid %p cap %p issued %s "
655 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
656 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
664 * Return set of valid cap bits issued to us. Note that caps time
665 * out, and may be invalidated in bulk if the client session times out
666 * and session->s_cap_gen is bumped.
668 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
670 int have = ci->i_snap_caps;
671 struct ceph_cap *cap;
676 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
677 cap = rb_entry(p, struct ceph_cap, ci_node);
678 if (!__cap_is_valid(cap))
680 dout("__ceph_caps_issued %p cap %p issued %s\n",
681 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
684 *implemented |= cap->implemented;
687 * exclude caps issued by non-auth MDS, but are been revoking
688 * by the auth MDS. The non-auth MDS should be revoking/exporting
689 * these caps, but the message is delayed.
691 if (ci->i_auth_cap) {
692 cap = ci->i_auth_cap;
693 have &= ~cap->implemented | cap->issued;
699 * Get cap bits issued by caps other than @ocap
701 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
703 int have = ci->i_snap_caps;
704 struct ceph_cap *cap;
707 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
708 cap = rb_entry(p, struct ceph_cap, ci_node);
711 if (!__cap_is_valid(cap))
719 * Move a cap to the end of the LRU (oldest caps at list head, newest
722 static void __touch_cap(struct ceph_cap *cap)
724 struct ceph_mds_session *s = cap->session;
726 spin_lock(&s->s_cap_lock);
727 if (s->s_cap_iterator == NULL) {
728 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
730 list_move_tail(&cap->session_caps, &s->s_caps);
732 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
733 &cap->ci->vfs_inode, cap, s->s_mds);
735 spin_unlock(&s->s_cap_lock);
739 * Check if we hold the given mask. If so, move the cap(s) to the
740 * front of their respective LRUs. (This is the preferred way for
741 * callers to check for caps they want.)
743 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
745 struct ceph_cap *cap;
747 int have = ci->i_snap_caps;
749 if ((have & mask) == mask) {
750 dout("__ceph_caps_issued_mask %p snap issued %s"
751 " (mask %s)\n", &ci->vfs_inode,
752 ceph_cap_string(have),
753 ceph_cap_string(mask));
757 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
758 cap = rb_entry(p, struct ceph_cap, ci_node);
759 if (!__cap_is_valid(cap))
761 if ((cap->issued & mask) == mask) {
762 dout("__ceph_caps_issued_mask %p cap %p issued %s"
763 " (mask %s)\n", &ci->vfs_inode, cap,
764 ceph_cap_string(cap->issued),
765 ceph_cap_string(mask));
771 /* does a combination of caps satisfy mask? */
773 if ((have & mask) == mask) {
774 dout("__ceph_caps_issued_mask %p combo issued %s"
775 " (mask %s)\n", &ci->vfs_inode,
776 ceph_cap_string(cap->issued),
777 ceph_cap_string(mask));
781 /* touch this + preceding caps */
783 for (q = rb_first(&ci->i_caps); q != p;
785 cap = rb_entry(q, struct ceph_cap,
787 if (!__cap_is_valid(cap))
800 * Return true if mask caps are currently being revoked by an MDS.
802 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
803 struct ceph_cap *ocap, int mask)
805 struct ceph_cap *cap;
808 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
809 cap = rb_entry(p, struct ceph_cap, ci_node);
811 (cap->implemented & ~cap->issued & mask))
817 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
819 struct inode *inode = &ci->vfs_inode;
822 spin_lock(&ci->i_ceph_lock);
823 ret = __ceph_caps_revoking_other(ci, NULL, mask);
824 spin_unlock(&ci->i_ceph_lock);
825 dout("ceph_caps_revoking %p %s = %d\n", inode,
826 ceph_cap_string(mask), ret);
830 int __ceph_caps_used(struct ceph_inode_info *ci)
834 used |= CEPH_CAP_PIN;
836 used |= CEPH_CAP_FILE_RD;
837 if (ci->i_rdcache_ref ||
838 (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */
839 ci->vfs_inode.i_data.nrpages))
840 used |= CEPH_CAP_FILE_CACHE;
842 used |= CEPH_CAP_FILE_WR;
843 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
844 used |= CEPH_CAP_FILE_BUFFER;
849 * wanted, by virtue of open file modes
851 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
854 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
855 if (ci->i_nr_by_mode[i])
860 return ceph_caps_for_mode(bits >> 1);
864 * Return caps we have registered with the MDS(s) as 'wanted'.
866 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
868 struct ceph_cap *cap;
872 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
873 cap = rb_entry(p, struct ceph_cap, ci_node);
874 if (!__cap_is_valid(cap))
876 if (cap == ci->i_auth_cap)
877 mds_wanted |= cap->mds_wanted;
879 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
885 * called under i_ceph_lock
887 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
889 return !RB_EMPTY_ROOT(&ci->i_caps);
892 int ceph_is_any_caps(struct inode *inode)
894 struct ceph_inode_info *ci = ceph_inode(inode);
897 spin_lock(&ci->i_ceph_lock);
898 ret = __ceph_is_any_caps(ci);
899 spin_unlock(&ci->i_ceph_lock);
904 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
906 struct ceph_snap_realm *realm = ci->i_snap_realm;
907 spin_lock(&realm->inodes_with_caps_lock);
908 list_del_init(&ci->i_snap_realm_item);
909 ci->i_snap_realm_counter++;
910 ci->i_snap_realm = NULL;
911 spin_unlock(&realm->inodes_with_caps_lock);
912 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
917 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
919 * caller should hold i_ceph_lock.
920 * caller will not hold session s_mutex if called from destroy_inode.
922 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
924 struct ceph_mds_session *session = cap->session;
925 struct ceph_inode_info *ci = cap->ci;
926 struct ceph_mds_client *mdsc =
927 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
930 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
932 /* remove from session list */
933 spin_lock(&session->s_cap_lock);
934 if (session->s_cap_iterator == cap) {
935 /* not yet, we are iterating over this very cap */
936 dout("__ceph_remove_cap delaying %p removal from session %p\n",
939 list_del_init(&cap->session_caps);
940 session->s_nr_caps--;
944 /* protect backpointer with s_cap_lock: see iterate_session_caps */
948 * s_cap_reconnect is protected by s_cap_lock. no one changes
949 * s_cap_gen while session is in the reconnect state.
952 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
953 cap->queue_release = 1;
955 list_add_tail(&cap->session_caps,
956 &session->s_cap_releases);
957 session->s_num_cap_releases++;
961 cap->queue_release = 0;
963 cap->cap_ino = ci->i_vino.ino;
965 spin_unlock(&session->s_cap_lock);
967 /* remove from inode list */
968 rb_erase(&cap->ci_node, &ci->i_caps);
969 if (ci->i_auth_cap == cap)
970 ci->i_auth_cap = NULL;
973 ceph_put_cap(mdsc, cap);
975 /* when reconnect denied, we remove session caps forcibly,
976 * i_wr_ref can be non-zero. If there are ongoing write,
979 if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm)
980 drop_inode_snap_realm(ci);
982 if (!__ceph_is_any_real_caps(ci))
983 __cap_delay_cancel(mdsc, ci);
987 * Build and send a cap message to the given MDS.
989 * Caller should be holding s_mutex.
991 static int send_cap_msg(struct ceph_mds_session *session,
992 u64 ino, u64 cid, int op,
993 int caps, int wanted, int dirty,
994 u32 seq, u64 flush_tid, u64 oldest_flush_tid,
995 u32 issue_seq, u32 mseq, u64 size, u64 max_size,
996 struct timespec *mtime, struct timespec *atime,
997 struct timespec *ctime, u32 time_warp_seq,
998 kuid_t uid, kgid_t gid, umode_t mode,
1000 struct ceph_buffer *xattrs_buf,
1001 u64 follows, bool inline_data)
1003 struct ceph_mds_caps *fc;
1004 struct ceph_msg *msg;
1008 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1009 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1010 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
1011 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
1012 ceph_cap_string(dirty),
1013 seq, issue_seq, flush_tid, oldest_flush_tid,
1014 mseq, follows, size, max_size,
1015 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
1017 /* flock buffer size + inline version + inline data size +
1018 * osd_epoch_barrier + oldest_flush_tid */
1019 extra_len = 4 + 8 + 4 + 4 + 8;
1020 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1025 msg->hdr.version = cpu_to_le16(6);
1026 msg->hdr.tid = cpu_to_le64(flush_tid);
1028 fc = msg->front.iov_base;
1029 memset(fc, 0, sizeof(*fc));
1031 fc->cap_id = cpu_to_le64(cid);
1032 fc->op = cpu_to_le32(op);
1033 fc->seq = cpu_to_le32(seq);
1034 fc->issue_seq = cpu_to_le32(issue_seq);
1035 fc->migrate_seq = cpu_to_le32(mseq);
1036 fc->caps = cpu_to_le32(caps);
1037 fc->wanted = cpu_to_le32(wanted);
1038 fc->dirty = cpu_to_le32(dirty);
1039 fc->ino = cpu_to_le64(ino);
1040 fc->snap_follows = cpu_to_le64(follows);
1042 fc->size = cpu_to_le64(size);
1043 fc->max_size = cpu_to_le64(max_size);
1045 ceph_encode_timespec(&fc->mtime, mtime);
1047 ceph_encode_timespec(&fc->atime, atime);
1049 ceph_encode_timespec(&fc->ctime, ctime);
1050 fc->time_warp_seq = cpu_to_le32(time_warp_seq);
1052 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, uid));
1053 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, gid));
1054 fc->mode = cpu_to_le32(mode);
1057 /* flock buffer size */
1058 ceph_encode_32(&p, 0);
1059 /* inline version */
1060 ceph_encode_64(&p, inline_data ? 0 : CEPH_INLINE_NONE);
1061 /* inline data size */
1062 ceph_encode_32(&p, 0);
1063 /* osd_epoch_barrier */
1064 ceph_encode_32(&p, 0);
1065 /* oldest_flush_tid */
1066 ceph_encode_64(&p, oldest_flush_tid);
1068 fc->xattr_version = cpu_to_le64(xattr_version);
1070 msg->middle = ceph_buffer_get(xattrs_buf);
1071 fc->xattr_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1072 msg->hdr.middle_len = cpu_to_le32(xattrs_buf->vec.iov_len);
1075 ceph_con_send(&session->s_con, msg);
1080 * Queue cap releases when an inode is dropped from our cache. Since
1081 * inode is about to be destroyed, there is no need for i_ceph_lock.
1083 void ceph_queue_caps_release(struct inode *inode)
1085 struct ceph_inode_info *ci = ceph_inode(inode);
1088 p = rb_first(&ci->i_caps);
1090 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1092 __ceph_remove_cap(cap, true);
1097 * Send a cap msg on the given inode. Update our caps state, then
1098 * drop i_ceph_lock and send the message.
1100 * Make note of max_size reported/requested from mds, revoked caps
1101 * that have now been implemented.
1103 * Make half-hearted attempt ot to invalidate page cache if we are
1104 * dropping RDCACHE. Note that this will leave behind locked pages
1105 * that we'll then need to deal with elsewhere.
1107 * Return non-zero if delayed release, or we experienced an error
1108 * such that the caller should requeue + retry later.
1110 * called with i_ceph_lock, then drops it.
1111 * caller should hold snap_rwsem (read), s_mutex.
1113 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1114 int op, int used, int want, int retain, int flushing,
1115 u64 flush_tid, u64 oldest_flush_tid)
1116 __releases(cap->ci->i_ceph_lock)
1118 struct ceph_inode_info *ci = cap->ci;
1119 struct inode *inode = &ci->vfs_inode;
1120 u64 cap_id = cap->cap_id;
1121 int held, revoking, dropping, keep;
1122 u64 follows, size, max_size;
1123 u32 seq, issue_seq, mseq, time_warp_seq;
1124 struct timespec mtime, atime, ctime;
1129 struct ceph_mds_session *session;
1130 u64 xattr_version = 0;
1131 struct ceph_buffer *xattr_blob = NULL;
1136 held = cap->issued | cap->implemented;
1137 revoking = cap->implemented & ~cap->issued;
1138 retain &= ~revoking;
1139 dropping = cap->issued & ~retain;
1141 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1142 inode, cap, cap->session,
1143 ceph_cap_string(held), ceph_cap_string(held & retain),
1144 ceph_cap_string(revoking));
1145 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1147 session = cap->session;
1149 /* don't release wanted unless we've waited a bit. */
1150 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1151 time_before(jiffies, ci->i_hold_caps_min)) {
1152 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1153 ceph_cap_string(cap->issued),
1154 ceph_cap_string(cap->issued & retain),
1155 ceph_cap_string(cap->mds_wanted),
1156 ceph_cap_string(want));
1157 want |= cap->mds_wanted;
1158 retain |= cap->issued;
1161 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1163 cap->issued &= retain; /* drop bits we don't want */
1164 if (cap->implemented & ~cap->issued) {
1166 * Wake up any waiters on wanted -> needed transition.
1167 * This is due to the weird transition from buffered
1168 * to sync IO... we need to flush dirty pages _before_
1169 * allowing sync writes to avoid reordering.
1173 cap->implemented &= cap->issued | used;
1174 cap->mds_wanted = want;
1176 follows = flushing ? ci->i_head_snapc->seq : 0;
1178 keep = cap->implemented;
1180 issue_seq = cap->issue_seq;
1182 size = inode->i_size;
1183 ci->i_reported_size = size;
1184 max_size = ci->i_wanted_max_size;
1185 ci->i_requested_max_size = max_size;
1186 mtime = inode->i_mtime;
1187 atime = inode->i_atime;
1188 ctime = inode->i_ctime;
1189 time_warp_seq = ci->i_time_warp_seq;
1192 mode = inode->i_mode;
1194 if (flushing & CEPH_CAP_XATTR_EXCL) {
1195 __ceph_build_xattrs_blob(ci);
1196 xattr_blob = ci->i_xattrs.blob;
1197 xattr_version = ci->i_xattrs.version;
1200 inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1202 spin_unlock(&ci->i_ceph_lock);
1204 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1205 op, keep, want, flushing, seq,
1206 flush_tid, oldest_flush_tid, issue_seq, mseq,
1207 size, max_size, &mtime, &atime, &ctime, time_warp_seq,
1208 uid, gid, mode, xattr_version, xattr_blob,
1209 follows, inline_data);
1211 dout("error sending cap msg, must requeue %p\n", inode);
1216 wake_up_all(&ci->i_cap_wq);
1221 static inline int __send_flush_snap(struct inode *inode,
1222 struct ceph_mds_session *session,
1223 struct ceph_cap_snap *capsnap,
1224 u32 mseq, u64 oldest_flush_tid)
1226 return send_cap_msg(session, ceph_vino(inode).ino, 0,
1227 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1228 capsnap->dirty, 0, capsnap->cap_flush.tid,
1229 oldest_flush_tid, 0, mseq, capsnap->size, 0,
1230 &capsnap->mtime, &capsnap->atime,
1231 &capsnap->ctime, capsnap->time_warp_seq,
1232 capsnap->uid, capsnap->gid, capsnap->mode,
1233 capsnap->xattr_version, capsnap->xattr_blob,
1234 capsnap->follows, capsnap->inline_data);
1238 * When a snapshot is taken, clients accumulate dirty metadata on
1239 * inodes with capabilities in ceph_cap_snaps to describe the file
1240 * state at the time the snapshot was taken. This must be flushed
1241 * asynchronously back to the MDS once sync writes complete and dirty
1242 * data is written out.
1244 * Called under i_ceph_lock. Takes s_mutex as needed.
1246 void __ceph_flush_snaps(struct ceph_inode_info *ci,
1247 struct ceph_mds_session **psession)
1248 __releases(ci->i_ceph_lock)
1249 __acquires(ci->i_ceph_lock)
1251 struct inode *inode = &ci->vfs_inode;
1253 struct ceph_cap_snap *capsnap;
1255 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1256 struct ceph_mds_session *session = NULL; /* if session != NULL, we hold
1258 u64 oldest_flush_tid;
1259 u64 next_follows = 0; /* keep track of how far we've gotten through the
1260 i_cap_snaps list, and skip these entries next time
1261 around to avoid an infinite loop */
1264 session = *psession;
1266 dout("__flush_snaps %p\n", inode);
1268 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1269 /* avoid an infiniute loop after retry */
1270 if (capsnap->follows < next_follows)
1273 * we need to wait for sync writes to complete and for dirty
1274 * pages to be written out.
1276 if (capsnap->dirty_pages || capsnap->writing)
1279 /* should be removed by ceph_try_drop_cap_snap() */
1280 BUG_ON(!capsnap->need_flush);
1282 /* pick mds, take s_mutex */
1283 if (ci->i_auth_cap == NULL) {
1284 dout("no auth cap (migrating?), doing nothing\n");
1288 /* only flush each capsnap once */
1289 if (capsnap->cap_flush.tid > 0) {
1290 dout("already flushed %p, skipping\n", capsnap);
1294 mds = ci->i_auth_cap->session->s_mds;
1295 mseq = ci->i_auth_cap->mseq;
1297 if (session && session->s_mds != mds) {
1298 dout("oops, wrong session %p mutex\n", session);
1300 mutex_unlock(&session->s_mutex);
1301 ceph_put_mds_session(session);
1305 spin_unlock(&ci->i_ceph_lock);
1306 mutex_lock(&mdsc->mutex);
1307 session = __ceph_lookup_mds_session(mdsc, mds);
1308 mutex_unlock(&mdsc->mutex);
1310 dout("inverting session/ino locks on %p\n",
1312 mutex_lock(&session->s_mutex);
1315 * if session == NULL, we raced against a cap
1316 * deletion or migration. retry, and we'll
1317 * get a better @mds value next time.
1319 spin_lock(&ci->i_ceph_lock);
1323 spin_lock(&mdsc->cap_dirty_lock);
1324 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1325 list_add_tail(&capsnap->cap_flush.g_list,
1326 &mdsc->cap_flush_list);
1327 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1329 if (list_empty(&ci->i_flushing_item)) {
1330 list_add_tail(&ci->i_flushing_item,
1331 &session->s_cap_flushing);
1333 spin_unlock(&mdsc->cap_dirty_lock);
1335 list_add_tail(&capsnap->cap_flush.i_list,
1336 &ci->i_cap_flush_list);
1338 atomic_inc(&capsnap->nref);
1339 spin_unlock(&ci->i_ceph_lock);
1341 dout("flush_snaps %p cap_snap %p follows %lld tid %llu\n",
1342 inode, capsnap, capsnap->follows, capsnap->cap_flush.tid);
1343 __send_flush_snap(inode, session, capsnap, mseq,
1346 next_follows = capsnap->follows + 1;
1347 ceph_put_cap_snap(capsnap);
1349 spin_lock(&ci->i_ceph_lock);
1353 /* we flushed them all; remove this inode from the queue */
1354 spin_lock(&mdsc->snap_flush_lock);
1355 list_del_init(&ci->i_snap_flush_item);
1356 spin_unlock(&mdsc->snap_flush_lock);
1360 *psession = session;
1362 mutex_unlock(&session->s_mutex);
1363 ceph_put_mds_session(session);
1367 static void ceph_flush_snaps(struct ceph_inode_info *ci)
1369 spin_lock(&ci->i_ceph_lock);
1370 __ceph_flush_snaps(ci, NULL);
1371 spin_unlock(&ci->i_ceph_lock);
1375 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1376 * Caller is then responsible for calling __mark_inode_dirty with the
1377 * returned flags value.
1379 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1380 struct ceph_cap_flush **pcf)
1382 struct ceph_mds_client *mdsc =
1383 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1384 struct inode *inode = &ci->vfs_inode;
1385 int was = ci->i_dirty_caps;
1388 if (!ci->i_auth_cap) {
1389 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1390 "but no auth cap (session was closed?)\n",
1391 inode, ceph_ino(inode), ceph_cap_string(mask));
1395 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1396 ceph_cap_string(mask), ceph_cap_string(was),
1397 ceph_cap_string(was | mask));
1398 ci->i_dirty_caps |= mask;
1400 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1401 swap(ci->i_prealloc_cap_flush, *pcf);
1403 if (!ci->i_head_snapc) {
1404 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1405 ci->i_head_snapc = ceph_get_snap_context(
1406 ci->i_snap_realm->cached_context);
1408 dout(" inode %p now dirty snapc %p auth cap %p\n",
1409 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1410 BUG_ON(!list_empty(&ci->i_dirty_item));
1411 spin_lock(&mdsc->cap_dirty_lock);
1412 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1413 spin_unlock(&mdsc->cap_dirty_lock);
1414 if (ci->i_flushing_caps == 0) {
1416 dirty |= I_DIRTY_SYNC;
1419 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1421 BUG_ON(list_empty(&ci->i_dirty_item));
1422 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1423 (mask & CEPH_CAP_FILE_BUFFER))
1424 dirty |= I_DIRTY_DATASYNC;
1425 __cap_delay_requeue(mdsc, ci);
1429 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1431 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1434 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1437 kmem_cache_free(ceph_cap_flush_cachep, cf);
1440 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1442 if (!list_empty(&mdsc->cap_flush_list)) {
1443 struct ceph_cap_flush *cf =
1444 list_first_entry(&mdsc->cap_flush_list,
1445 struct ceph_cap_flush, g_list);
1452 * Add dirty inode to the flushing list. Assigned a seq number so we
1453 * can wait for caps to flush without starving.
1455 * Called under i_ceph_lock.
1457 static int __mark_caps_flushing(struct inode *inode,
1458 struct ceph_mds_session *session,
1459 u64 *flush_tid, u64 *oldest_flush_tid)
1461 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1462 struct ceph_inode_info *ci = ceph_inode(inode);
1463 struct ceph_cap_flush *cf = NULL;
1466 BUG_ON(ci->i_dirty_caps == 0);
1467 BUG_ON(list_empty(&ci->i_dirty_item));
1468 BUG_ON(!ci->i_prealloc_cap_flush);
1470 flushing = ci->i_dirty_caps;
1471 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1472 ceph_cap_string(flushing),
1473 ceph_cap_string(ci->i_flushing_caps),
1474 ceph_cap_string(ci->i_flushing_caps | flushing));
1475 ci->i_flushing_caps |= flushing;
1476 ci->i_dirty_caps = 0;
1477 dout(" inode %p now !dirty\n", inode);
1479 swap(cf, ci->i_prealloc_cap_flush);
1480 cf->caps = flushing;
1482 spin_lock(&mdsc->cap_dirty_lock);
1483 list_del_init(&ci->i_dirty_item);
1485 cf->tid = ++mdsc->last_cap_flush_tid;
1486 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1487 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1489 if (list_empty(&ci->i_flushing_item)) {
1490 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1491 mdsc->num_cap_flushing++;
1493 spin_unlock(&mdsc->cap_dirty_lock);
1495 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1497 *flush_tid = cf->tid;
1502 * try to invalidate mapping pages without blocking.
1504 static int try_nonblocking_invalidate(struct inode *inode)
1506 struct ceph_inode_info *ci = ceph_inode(inode);
1507 u32 invalidating_gen = ci->i_rdcache_gen;
1509 spin_unlock(&ci->i_ceph_lock);
1510 invalidate_mapping_pages(&inode->i_data, 0, -1);
1511 spin_lock(&ci->i_ceph_lock);
1513 if (inode->i_data.nrpages == 0 &&
1514 invalidating_gen == ci->i_rdcache_gen) {
1516 dout("try_nonblocking_invalidate %p success\n", inode);
1517 /* save any racing async invalidate some trouble */
1518 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1521 dout("try_nonblocking_invalidate %p failed\n", inode);
1526 * Swiss army knife function to examine currently used and wanted
1527 * versus held caps. Release, flush, ack revoked caps to mds as
1530 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1531 * cap release further.
1532 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1533 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1536 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1537 struct ceph_mds_session *session)
1539 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1540 struct ceph_mds_client *mdsc = fsc->mdsc;
1541 struct inode *inode = &ci->vfs_inode;
1542 struct ceph_cap *cap;
1543 u64 flush_tid, oldest_flush_tid;
1544 int file_wanted, used, cap_used;
1545 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1546 int issued, implemented, want, retain, revoking, flushing = 0;
1547 int mds = -1; /* keep track of how far we've gone through i_caps list
1548 to avoid an infinite loop on retry */
1550 int delayed = 0, sent = 0, num;
1551 bool is_delayed = flags & CHECK_CAPS_NODELAY;
1552 bool queue_invalidate = false;
1553 bool force_requeue = false;
1554 bool tried_invalidate = false;
1556 /* if we are unmounting, flush any unused caps immediately. */
1560 spin_lock(&ci->i_ceph_lock);
1562 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1563 flags |= CHECK_CAPS_FLUSH;
1565 /* flush snaps first time around only */
1566 if (!list_empty(&ci->i_cap_snaps))
1567 __ceph_flush_snaps(ci, &session);
1570 spin_lock(&ci->i_ceph_lock);
1572 file_wanted = __ceph_caps_file_wanted(ci);
1573 used = __ceph_caps_used(ci);
1574 issued = __ceph_caps_issued(ci, &implemented);
1575 revoking = implemented & ~issued;
1578 retain = file_wanted | used | CEPH_CAP_PIN;
1579 if (!mdsc->stopping && inode->i_nlink > 0) {
1581 retain |= CEPH_CAP_ANY; /* be greedy */
1582 } else if (S_ISDIR(inode->i_mode) &&
1583 (issued & CEPH_CAP_FILE_SHARED) &&
1584 __ceph_dir_is_complete(ci)) {
1586 * If a directory is complete, we want to keep
1587 * the exclusive cap. So that MDS does not end up
1588 * revoking the shared cap on every create/unlink
1591 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1595 retain |= CEPH_CAP_ANY_SHARED;
1597 * keep RD only if we didn't have the file open RW,
1598 * because then the mds would revoke it anyway to
1599 * journal max_size=0.
1601 if (ci->i_max_size == 0)
1602 retain |= CEPH_CAP_ANY_RD;
1606 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1607 " issued %s revoking %s retain %s %s%s%s\n", inode,
1608 ceph_cap_string(file_wanted),
1609 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1610 ceph_cap_string(ci->i_flushing_caps),
1611 ceph_cap_string(issued), ceph_cap_string(revoking),
1612 ceph_cap_string(retain),
1613 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1614 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1615 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1618 * If we no longer need to hold onto old our caps, and we may
1619 * have cached pages, but don't want them, then try to invalidate.
1620 * If we fail, it's because pages are locked.... try again later.
1622 if ((!is_delayed || mdsc->stopping) &&
1623 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */
1624 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1625 inode->i_data.nrpages && /* have cached pages */
1626 (revoking & (CEPH_CAP_FILE_CACHE|
1627 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1628 !tried_invalidate) {
1629 dout("check_caps trying to invalidate on %p\n", inode);
1630 if (try_nonblocking_invalidate(inode) < 0) {
1631 if (revoking & (CEPH_CAP_FILE_CACHE|
1632 CEPH_CAP_FILE_LAZYIO)) {
1633 dout("check_caps queuing invalidate\n");
1634 queue_invalidate = true;
1635 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1637 dout("check_caps failed to invalidate pages\n");
1638 /* we failed to invalidate pages. check these
1639 caps again later. */
1640 force_requeue = true;
1641 __cap_set_timeouts(mdsc, ci);
1644 tried_invalidate = true;
1649 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1650 cap = rb_entry(p, struct ceph_cap, ci_node);
1653 /* avoid looping forever */
1654 if (mds >= cap->mds ||
1655 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1658 /* NOTE: no side-effects allowed, until we take s_mutex */
1661 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1662 cap_used &= ~ci->i_auth_cap->issued;
1664 revoking = cap->implemented & ~cap->issued;
1665 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1666 cap->mds, cap, ceph_cap_string(cap_used),
1667 ceph_cap_string(cap->issued),
1668 ceph_cap_string(cap->implemented),
1669 ceph_cap_string(revoking));
1671 if (cap == ci->i_auth_cap &&
1672 (cap->issued & CEPH_CAP_FILE_WR)) {
1673 /* request larger max_size from MDS? */
1674 if (ci->i_wanted_max_size > ci->i_max_size &&
1675 ci->i_wanted_max_size > ci->i_requested_max_size) {
1676 dout("requesting new max_size\n");
1680 /* approaching file_max? */
1681 if ((inode->i_size << 1) >= ci->i_max_size &&
1682 (ci->i_reported_size << 1) < ci->i_max_size) {
1683 dout("i_size approaching max_size\n");
1687 /* flush anything dirty? */
1688 if (cap == ci->i_auth_cap && (flags & CHECK_CAPS_FLUSH) &&
1690 dout("flushing dirty caps\n");
1694 /* completed revocation? going down and there are no caps? */
1695 if (revoking && (revoking & cap_used) == 0) {
1696 dout("completed revocation of %s\n",
1697 ceph_cap_string(cap->implemented & ~cap->issued));
1701 /* want more caps from mds? */
1702 if (want & ~(cap->mds_wanted | cap->issued))
1705 /* things we might delay */
1706 if ((cap->issued & ~retain) == 0 &&
1707 cap->mds_wanted == want)
1708 continue; /* nope, all good */
1714 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1715 time_before(jiffies, ci->i_hold_caps_max)) {
1716 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1717 ceph_cap_string(cap->issued),
1718 ceph_cap_string(cap->issued & retain),
1719 ceph_cap_string(cap->mds_wanted),
1720 ceph_cap_string(want));
1726 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1727 dout(" skipping %p I_NOFLUSH set\n", inode);
1731 if (session && session != cap->session) {
1732 dout("oops, wrong session %p mutex\n", session);
1733 mutex_unlock(&session->s_mutex);
1737 session = cap->session;
1738 if (mutex_trylock(&session->s_mutex) == 0) {
1739 dout("inverting session/ino locks on %p\n",
1741 spin_unlock(&ci->i_ceph_lock);
1742 if (took_snap_rwsem) {
1743 up_read(&mdsc->snap_rwsem);
1744 took_snap_rwsem = 0;
1746 mutex_lock(&session->s_mutex);
1750 /* take snap_rwsem after session mutex */
1751 if (!took_snap_rwsem) {
1752 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1753 dout("inverting snap/in locks on %p\n",
1755 spin_unlock(&ci->i_ceph_lock);
1756 down_read(&mdsc->snap_rwsem);
1757 took_snap_rwsem = 1;
1760 took_snap_rwsem = 1;
1763 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
1764 flushing = __mark_caps_flushing(inode, session,
1770 spin_lock(&mdsc->cap_dirty_lock);
1771 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1772 spin_unlock(&mdsc->cap_dirty_lock);
1775 mds = cap->mds; /* remember mds, so we don't repeat */
1778 /* __send_cap drops i_ceph_lock */
1779 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used,
1780 want, retain, flushing,
1781 flush_tid, oldest_flush_tid);
1782 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1786 * Reschedule delayed caps release if we delayed anything,
1789 if (delayed && is_delayed)
1790 force_requeue = true; /* __send_cap delayed release; requeue */
1791 if (!delayed && !is_delayed)
1792 __cap_delay_cancel(mdsc, ci);
1793 else if (!is_delayed || force_requeue)
1794 __cap_delay_requeue(mdsc, ci);
1796 spin_unlock(&ci->i_ceph_lock);
1798 if (queue_invalidate)
1799 ceph_queue_invalidate(inode);
1802 mutex_unlock(&session->s_mutex);
1803 if (took_snap_rwsem)
1804 up_read(&mdsc->snap_rwsem);
1808 * Try to flush dirty caps back to the auth mds.
1810 static int try_flush_caps(struct inode *inode, u64 *ptid)
1812 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1813 struct ceph_inode_info *ci = ceph_inode(inode);
1814 struct ceph_mds_session *session = NULL;
1816 u64 flush_tid = 0, oldest_flush_tid = 0;
1819 spin_lock(&ci->i_ceph_lock);
1820 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1821 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1824 if (ci->i_dirty_caps && ci->i_auth_cap) {
1825 struct ceph_cap *cap = ci->i_auth_cap;
1826 int used = __ceph_caps_used(ci);
1827 int want = __ceph_caps_wanted(ci);
1830 if (!session || session != cap->session) {
1831 spin_unlock(&ci->i_ceph_lock);
1833 mutex_unlock(&session->s_mutex);
1834 session = cap->session;
1835 mutex_lock(&session->s_mutex);
1838 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1841 flushing = __mark_caps_flushing(inode, session, &flush_tid,
1844 /* __send_cap drops i_ceph_lock */
1845 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1846 (cap->issued | cap->implemented),
1847 flushing, flush_tid, oldest_flush_tid);
1850 spin_lock(&ci->i_ceph_lock);
1851 __cap_delay_requeue(mdsc, ci);
1852 spin_unlock(&ci->i_ceph_lock);
1855 if (!list_empty(&ci->i_cap_flush_list)) {
1856 struct ceph_cap_flush *cf =
1857 list_last_entry(&ci->i_cap_flush_list,
1858 struct ceph_cap_flush, i_list);
1859 flush_tid = cf->tid;
1861 flushing = ci->i_flushing_caps;
1862 spin_unlock(&ci->i_ceph_lock);
1866 mutex_unlock(&session->s_mutex);
1873 * Return true if we've flushed caps through the given flush_tid.
1875 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
1877 struct ceph_inode_info *ci = ceph_inode(inode);
1880 spin_lock(&ci->i_ceph_lock);
1881 if (!list_empty(&ci->i_cap_flush_list)) {
1882 struct ceph_cap_flush * cf =
1883 list_first_entry(&ci->i_cap_flush_list,
1884 struct ceph_cap_flush, i_list);
1885 if (cf->tid <= flush_tid)
1888 spin_unlock(&ci->i_ceph_lock);
1893 * wait for any unsafe requests to complete.
1895 static int unsafe_request_wait(struct inode *inode)
1897 struct ceph_inode_info *ci = ceph_inode(inode);
1898 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
1901 spin_lock(&ci->i_unsafe_lock);
1902 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
1903 req1 = list_last_entry(&ci->i_unsafe_dirops,
1904 struct ceph_mds_request,
1906 ceph_mdsc_get_request(req1);
1908 if (!list_empty(&ci->i_unsafe_iops)) {
1909 req2 = list_last_entry(&ci->i_unsafe_iops,
1910 struct ceph_mds_request,
1911 r_unsafe_target_item);
1912 ceph_mdsc_get_request(req2);
1914 spin_unlock(&ci->i_unsafe_lock);
1916 dout("unsafe_requeset_wait %p wait on tid %llu %llu\n",
1917 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
1919 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
1920 ceph_timeout_jiffies(req1->r_timeout));
1923 ceph_mdsc_put_request(req1);
1926 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
1927 ceph_timeout_jiffies(req2->r_timeout));
1930 ceph_mdsc_put_request(req2);
1935 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1937 struct inode *inode = file->f_mapping->host;
1938 struct ceph_inode_info *ci = ceph_inode(inode);
1943 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
1945 ceph_sync_write_wait(inode);
1947 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1956 dirty = try_flush_caps(inode, &flush_tid);
1957 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1959 ret = unsafe_request_wait(inode);
1962 * only wait on non-file metadata writeback (the mds
1963 * can recover size and mtime, so we don't need to
1966 if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1967 ret = wait_event_interruptible(ci->i_cap_wq,
1968 caps_are_flushed(inode, flush_tid));
1970 inode_unlock(inode);
1972 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
1977 * Flush any dirty caps back to the mds. If we aren't asked to wait,
1978 * queue inode for flush but don't do so immediately, because we can
1979 * get by with fewer MDS messages if we wait for data writeback to
1982 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
1984 struct ceph_inode_info *ci = ceph_inode(inode);
1988 int wait = wbc->sync_mode == WB_SYNC_ALL;
1990 dout("write_inode %p wait=%d\n", inode, wait);
1992 dirty = try_flush_caps(inode, &flush_tid);
1994 err = wait_event_interruptible(ci->i_cap_wq,
1995 caps_are_flushed(inode, flush_tid));
1997 struct ceph_mds_client *mdsc =
1998 ceph_sb_to_client(inode->i_sb)->mdsc;
2000 spin_lock(&ci->i_ceph_lock);
2001 if (__ceph_caps_dirty(ci))
2002 __cap_delay_requeue_front(mdsc, ci);
2003 spin_unlock(&ci->i_ceph_lock);
2008 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2009 struct ceph_mds_session *session,
2010 struct ceph_inode_info *ci,
2011 u64 oldest_flush_tid)
2012 __releases(ci->i_ceph_lock)
2013 __acquires(ci->i_ceph_lock)
2015 struct inode *inode = &ci->vfs_inode;
2016 struct ceph_cap *cap;
2017 struct ceph_cap_flush *cf;
2021 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2022 if (cf->tid < first_tid)
2025 cap = ci->i_auth_cap;
2026 if (!(cap && cap->session == session)) {
2027 pr_err("%p auth cap %p not mds%d ???\n",
2028 inode, cap, session->s_mds);
2032 first_tid = cf->tid + 1;
2035 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2036 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2037 ci->i_ceph_flags |= CEPH_I_NODELAY;
2038 ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2039 __ceph_caps_used(ci),
2040 __ceph_caps_wanted(ci),
2041 cap->issued | cap->implemented,
2042 cf->caps, cf->tid, oldest_flush_tid);
2044 pr_err("kick_flushing_caps: error sending "
2045 "cap flush, ino (%llx.%llx) "
2046 "tid %llu flushing %s\n",
2047 ceph_vinop(inode), cf->tid,
2048 ceph_cap_string(cf->caps));
2051 struct ceph_cap_snap *capsnap =
2052 container_of(cf, struct ceph_cap_snap,
2054 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2055 inode, capsnap, cf->tid,
2056 ceph_cap_string(capsnap->dirty));
2058 atomic_inc(&capsnap->nref);
2059 spin_unlock(&ci->i_ceph_lock);
2061 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2064 pr_err("kick_flushing_caps: error sending "
2065 "cap flushsnap, ino (%llx.%llx) "
2066 "tid %llu follows %llu\n",
2067 ceph_vinop(inode), cf->tid,
2071 ceph_put_cap_snap(capsnap);
2074 spin_lock(&ci->i_ceph_lock);
2078 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2079 struct ceph_mds_session *session)
2081 struct ceph_inode_info *ci;
2082 struct ceph_cap *cap;
2083 u64 oldest_flush_tid;
2085 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2087 spin_lock(&mdsc->cap_dirty_lock);
2088 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2089 spin_unlock(&mdsc->cap_dirty_lock);
2091 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2092 spin_lock(&ci->i_ceph_lock);
2093 cap = ci->i_auth_cap;
2094 if (!(cap && cap->session == session)) {
2095 pr_err("%p auth cap %p not mds%d ???\n",
2096 &ci->vfs_inode, cap, session->s_mds);
2097 spin_unlock(&ci->i_ceph_lock);
2103 * if flushing caps were revoked, we re-send the cap flush
2104 * in client reconnect stage. This guarantees MDS * processes
2105 * the cap flush message before issuing the flushing caps to
2108 if ((cap->issued & ci->i_flushing_caps) !=
2109 ci->i_flushing_caps) {
2110 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2111 __kick_flushing_caps(mdsc, session, ci,
2114 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2117 spin_unlock(&ci->i_ceph_lock);
2121 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2122 struct ceph_mds_session *session)
2124 struct ceph_inode_info *ci;
2125 struct ceph_cap *cap;
2126 u64 oldest_flush_tid;
2128 dout("kick_flushing_caps mds%d\n", session->s_mds);
2130 spin_lock(&mdsc->cap_dirty_lock);
2131 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2132 spin_unlock(&mdsc->cap_dirty_lock);
2134 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2135 spin_lock(&ci->i_ceph_lock);
2136 cap = ci->i_auth_cap;
2137 if (!(cap && cap->session == session)) {
2138 pr_err("%p auth cap %p not mds%d ???\n",
2139 &ci->vfs_inode, cap, session->s_mds);
2140 spin_unlock(&ci->i_ceph_lock);
2143 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2144 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2145 __kick_flushing_caps(mdsc, session, ci,
2148 spin_unlock(&ci->i_ceph_lock);
2152 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2153 struct ceph_mds_session *session,
2154 struct inode *inode)
2155 __releases(ci->i_ceph_lock)
2157 struct ceph_inode_info *ci = ceph_inode(inode);
2158 struct ceph_cap *cap;
2160 cap = ci->i_auth_cap;
2161 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2162 ceph_cap_string(ci->i_flushing_caps));
2164 if (!list_empty(&ci->i_cap_flush_list)) {
2165 u64 oldest_flush_tid;
2166 spin_lock(&mdsc->cap_dirty_lock);
2167 list_move_tail(&ci->i_flushing_item,
2168 &cap->session->s_cap_flushing);
2169 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2170 spin_unlock(&mdsc->cap_dirty_lock);
2172 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2173 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2174 spin_unlock(&ci->i_ceph_lock);
2176 spin_unlock(&ci->i_ceph_lock);
2182 * Take references to capabilities we hold, so that we don't release
2183 * them to the MDS prematurely.
2185 * Protected by i_ceph_lock.
2187 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2188 bool snap_rwsem_locked)
2190 if (got & CEPH_CAP_PIN)
2192 if (got & CEPH_CAP_FILE_RD)
2194 if (got & CEPH_CAP_FILE_CACHE)
2195 ci->i_rdcache_ref++;
2196 if (got & CEPH_CAP_FILE_WR) {
2197 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2198 BUG_ON(!snap_rwsem_locked);
2199 ci->i_head_snapc = ceph_get_snap_context(
2200 ci->i_snap_realm->cached_context);
2204 if (got & CEPH_CAP_FILE_BUFFER) {
2205 if (ci->i_wb_ref == 0)
2206 ihold(&ci->vfs_inode);
2208 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2209 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2214 * Try to grab cap references. Specify those refs we @want, and the
2215 * minimal set we @need. Also include the larger offset we are writing
2216 * to (when applicable), and check against max_size here as well.
2217 * Note that caller is responsible for ensuring max_size increases are
2218 * requested from the MDS.
2220 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2221 loff_t endoff, bool nonblock, int *got, int *err)
2223 struct inode *inode = &ci->vfs_inode;
2224 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2226 int have, implemented;
2228 bool snap_rwsem_locked = false;
2230 dout("get_cap_refs %p need %s want %s\n", inode,
2231 ceph_cap_string(need), ceph_cap_string(want));
2234 spin_lock(&ci->i_ceph_lock);
2236 /* make sure file is actually open */
2237 file_wanted = __ceph_caps_file_wanted(ci);
2238 if ((file_wanted & need) != need) {
2239 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2240 ceph_cap_string(need), ceph_cap_string(file_wanted));
2246 /* finish pending truncate */
2247 while (ci->i_truncate_pending) {
2248 spin_unlock(&ci->i_ceph_lock);
2249 if (snap_rwsem_locked) {
2250 up_read(&mdsc->snap_rwsem);
2251 snap_rwsem_locked = false;
2253 __ceph_do_pending_vmtruncate(inode);
2254 spin_lock(&ci->i_ceph_lock);
2257 have = __ceph_caps_issued(ci, &implemented);
2259 if (have & need & CEPH_CAP_FILE_WR) {
2260 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2261 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2262 inode, endoff, ci->i_max_size);
2263 if (endoff > ci->i_requested_max_size) {
2270 * If a sync write is in progress, we must wait, so that we
2271 * can get a final snapshot value for size+mtime.
2273 if (__ceph_have_pending_cap_snap(ci)) {
2274 dout("get_cap_refs %p cap_snap_pending\n", inode);
2279 if ((have & need) == need) {
2281 * Look at (implemented & ~have & not) so that we keep waiting
2282 * on transition from wanted -> needed caps. This is needed
2283 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2284 * going before a prior buffered writeback happens.
2286 int not = want & ~(have & need);
2287 int revoking = implemented & ~have;
2288 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2289 inode, ceph_cap_string(have), ceph_cap_string(not),
2290 ceph_cap_string(revoking));
2291 if ((revoking & not) == 0) {
2292 if (!snap_rwsem_locked &&
2293 !ci->i_head_snapc &&
2294 (need & CEPH_CAP_FILE_WR)) {
2295 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2297 * we can not call down_read() when
2298 * task isn't in TASK_RUNNING state
2306 spin_unlock(&ci->i_ceph_lock);
2307 down_read(&mdsc->snap_rwsem);
2308 snap_rwsem_locked = true;
2311 snap_rwsem_locked = true;
2313 *got = need | (have & want);
2314 if ((need & CEPH_CAP_FILE_RD) &&
2315 !(*got & CEPH_CAP_FILE_CACHE))
2316 ceph_disable_fscache_readpage(ci);
2317 __take_cap_refs(ci, *got, true);
2321 int session_readonly = false;
2322 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2323 struct ceph_mds_session *s = ci->i_auth_cap->session;
2324 spin_lock(&s->s_cap_lock);
2325 session_readonly = s->s_readonly;
2326 spin_unlock(&s->s_cap_lock);
2328 if (session_readonly) {
2329 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2330 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2336 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) {
2338 if (ACCESS_ONCE(mdsc->fsc->mount_state) ==
2339 CEPH_MOUNT_SHUTDOWN) {
2340 dout("get_cap_refs %p forced umount\n", inode);
2345 mds_wanted = __ceph_caps_mds_wanted(ci);
2346 if ((mds_wanted & need) != need) {
2347 dout("get_cap_refs %p caps were dropped"
2348 " (session killed?)\n", inode);
2353 if ((mds_wanted & file_wanted) ==
2354 (file_wanted & (CEPH_CAP_FILE_RD|CEPH_CAP_FILE_WR)))
2355 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED;
2358 dout("get_cap_refs %p have %s needed %s\n", inode,
2359 ceph_cap_string(have), ceph_cap_string(need));
2362 spin_unlock(&ci->i_ceph_lock);
2363 if (snap_rwsem_locked)
2364 up_read(&mdsc->snap_rwsem);
2366 dout("get_cap_refs %p ret %d got %s\n", inode,
2367 ret, ceph_cap_string(*got));
2372 * Check the offset we are writing up to against our current
2373 * max_size. If necessary, tell the MDS we want to write to
2376 static void check_max_size(struct inode *inode, loff_t endoff)
2378 struct ceph_inode_info *ci = ceph_inode(inode);
2381 /* do we need to explicitly request a larger max_size? */
2382 spin_lock(&ci->i_ceph_lock);
2383 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2384 dout("write %p at large endoff %llu, req max_size\n",
2386 ci->i_wanted_max_size = endoff;
2388 /* duplicate ceph_check_caps()'s logic */
2389 if (ci->i_auth_cap &&
2390 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2391 ci->i_wanted_max_size > ci->i_max_size &&
2392 ci->i_wanted_max_size > ci->i_requested_max_size)
2394 spin_unlock(&ci->i_ceph_lock);
2396 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2400 * Wait for caps, and take cap references. If we can't get a WR cap
2401 * due to a small max_size, make sure we check_max_size (and possibly
2402 * ask the mds) so we don't get hung up indefinitely.
2404 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2405 loff_t endoff, int *got, struct page **pinned_page)
2407 int _got, ret, err = 0;
2409 ret = ceph_pool_perm_check(ci, need);
2415 check_max_size(&ci->vfs_inode, endoff);
2419 ret = try_get_cap_refs(ci, need, want, endoff,
2420 false, &_got, &err);
2427 ret = wait_event_interruptible(ci->i_cap_wq,
2428 try_get_cap_refs(ci, need, want, endoff,
2429 true, &_got, &err));
2436 if (err == -ESTALE) {
2437 /* session was killed, try renew caps */
2438 ret = ceph_renew_caps(&ci->vfs_inode);
2445 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2446 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2447 i_size_read(&ci->vfs_inode) > 0) {
2449 find_get_page(ci->vfs_inode.i_mapping, 0);
2451 if (PageUptodate(page)) {
2452 *pinned_page = page;
2458 * drop cap refs first because getattr while
2459 * holding * caps refs can cause deadlock.
2461 ceph_put_cap_refs(ci, _got);
2465 * getattr request will bring inline data into
2468 ret = __ceph_do_getattr(&ci->vfs_inode, NULL,
2469 CEPH_STAT_CAP_INLINE_DATA,
2478 if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
2479 ceph_fscache_revalidate_cookie(ci);
2486 * Take cap refs. Caller must already know we hold at least one ref
2487 * on the caps in question or we don't know this is safe.
2489 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2491 spin_lock(&ci->i_ceph_lock);
2492 __take_cap_refs(ci, caps, false);
2493 spin_unlock(&ci->i_ceph_lock);
2498 * drop cap_snap that is not associated with any snapshot.
2499 * we don't need to send FLUSHSNAP message for it.
2501 static int ceph_try_drop_cap_snap(struct ceph_cap_snap *capsnap)
2503 if (!capsnap->need_flush &&
2504 !capsnap->writing && !capsnap->dirty_pages) {
2505 dout("dropping cap_snap %p follows %llu\n",
2506 capsnap, capsnap->follows);
2507 BUG_ON(capsnap->cap_flush.tid > 0);
2508 ceph_put_snap_context(capsnap->context);
2509 list_del(&capsnap->ci_item);
2510 ceph_put_cap_snap(capsnap);
2519 * If we released the last ref on any given cap, call ceph_check_caps
2520 * to release (or schedule a release).
2522 * If we are releasing a WR cap (from a sync write), finalize any affected
2523 * cap_snap, and wake up any waiters.
2525 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2527 struct inode *inode = &ci->vfs_inode;
2528 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2530 spin_lock(&ci->i_ceph_lock);
2531 if (had & CEPH_CAP_PIN)
2533 if (had & CEPH_CAP_FILE_RD)
2534 if (--ci->i_rd_ref == 0)
2536 if (had & CEPH_CAP_FILE_CACHE)
2537 if (--ci->i_rdcache_ref == 0)
2539 if (had & CEPH_CAP_FILE_BUFFER) {
2540 if (--ci->i_wb_ref == 0) {
2544 dout("put_cap_refs %p wb %d -> %d (?)\n",
2545 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2547 if (had & CEPH_CAP_FILE_WR)
2548 if (--ci->i_wr_ref == 0) {
2550 if (__ceph_have_pending_cap_snap(ci)) {
2551 struct ceph_cap_snap *capsnap =
2552 list_last_entry(&ci->i_cap_snaps,
2553 struct ceph_cap_snap,
2555 capsnap->writing = 0;
2556 if (ceph_try_drop_cap_snap(capsnap))
2558 else if (__ceph_finish_cap_snap(ci, capsnap))
2562 if (ci->i_wrbuffer_ref_head == 0 &&
2563 ci->i_dirty_caps == 0 &&
2564 ci->i_flushing_caps == 0) {
2565 BUG_ON(!ci->i_head_snapc);
2566 ceph_put_snap_context(ci->i_head_snapc);
2567 ci->i_head_snapc = NULL;
2569 /* see comment in __ceph_remove_cap() */
2570 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2571 drop_inode_snap_realm(ci);
2573 spin_unlock(&ci->i_ceph_lock);
2575 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2576 last ? " last" : "", put ? " put" : "");
2578 if (last && !flushsnaps)
2579 ceph_check_caps(ci, 0, NULL);
2580 else if (flushsnaps)
2581 ceph_flush_snaps(ci);
2583 wake_up_all(&ci->i_cap_wq);
2589 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2590 * context. Adjust per-snap dirty page accounting as appropriate.
2591 * Once all dirty data for a cap_snap is flushed, flush snapped file
2592 * metadata back to the MDS. If we dropped the last ref, call
2595 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2596 struct ceph_snap_context *snapc)
2598 struct inode *inode = &ci->vfs_inode;
2600 int complete_capsnap = 0;
2601 int drop_capsnap = 0;
2603 struct ceph_cap_snap *capsnap = NULL;
2605 spin_lock(&ci->i_ceph_lock);
2606 ci->i_wrbuffer_ref -= nr;
2607 last = !ci->i_wrbuffer_ref;
2609 if (ci->i_head_snapc == snapc) {
2610 ci->i_wrbuffer_ref_head -= nr;
2611 if (ci->i_wrbuffer_ref_head == 0 &&
2612 ci->i_wr_ref == 0 &&
2613 ci->i_dirty_caps == 0 &&
2614 ci->i_flushing_caps == 0) {
2615 BUG_ON(!ci->i_head_snapc);
2616 ceph_put_snap_context(ci->i_head_snapc);
2617 ci->i_head_snapc = NULL;
2619 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2621 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2622 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2623 last ? " LAST" : "");
2625 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2626 if (capsnap->context == snapc) {
2632 capsnap->dirty_pages -= nr;
2633 if (capsnap->dirty_pages == 0) {
2634 complete_capsnap = 1;
2635 drop_capsnap = ceph_try_drop_cap_snap(capsnap);
2637 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2638 " snap %lld %d/%d -> %d/%d %s%s\n",
2639 inode, capsnap, capsnap->context->seq,
2640 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2641 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2642 last ? " (wrbuffer last)" : "",
2643 complete_capsnap ? " (complete capsnap)" : "");
2646 spin_unlock(&ci->i_ceph_lock);
2649 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2651 } else if (complete_capsnap) {
2652 ceph_flush_snaps(ci);
2653 wake_up_all(&ci->i_cap_wq);
2660 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2662 static void invalidate_aliases(struct inode *inode)
2664 struct dentry *dn, *prev = NULL;
2666 dout("invalidate_aliases inode %p\n", inode);
2667 d_prune_aliases(inode);
2669 * For non-directory inode, d_find_alias() only returns
2670 * hashed dentry. After calling d_invalidate(), the
2671 * dentry becomes unhashed.
2673 * For directory inode, d_find_alias() can return
2674 * unhashed dentry. But directory inode should have
2675 * one alias at most.
2677 while ((dn = d_find_alias(inode))) {
2692 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2693 * actually be a revocation if it specifies a smaller cap set.)
2695 * caller holds s_mutex and i_ceph_lock, we drop both.
2697 static void handle_cap_grant(struct ceph_mds_client *mdsc,
2698 struct inode *inode, struct ceph_mds_caps *grant,
2699 struct ceph_string **pns, u64 inline_version,
2700 void *inline_data, u32 inline_len,
2701 struct ceph_buffer *xattr_buf,
2702 struct ceph_mds_session *session,
2703 struct ceph_cap *cap, int issued)
2704 __releases(ci->i_ceph_lock)
2705 __releases(mdsc->snap_rwsem)
2707 struct ceph_inode_info *ci = ceph_inode(inode);
2708 int mds = session->s_mds;
2709 int seq = le32_to_cpu(grant->seq);
2710 int newcaps = le32_to_cpu(grant->caps);
2711 int used, wanted, dirty;
2712 u64 size = le64_to_cpu(grant->size);
2713 u64 max_size = le64_to_cpu(grant->max_size);
2714 struct timespec mtime, atime, ctime;
2717 bool writeback = false;
2718 bool queue_trunc = false;
2719 bool queue_invalidate = false;
2720 bool deleted_inode = false;
2721 bool fill_inline = false;
2723 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2724 inode, cap, mds, seq, ceph_cap_string(newcaps));
2725 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2730 * auth mds of the inode changed. we received the cap export message,
2731 * but still haven't received the cap import message. handle_cap_export
2732 * updated the new auth MDS' cap.
2734 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2735 * that was sent before the cap import message. So don't remove caps.
2737 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
2738 WARN_ON(cap != ci->i_auth_cap);
2739 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
2741 newcaps |= cap->issued;
2745 * If CACHE is being revoked, and we have no dirty buffers,
2746 * try to invalidate (once). (If there are dirty buffers, we
2747 * will invalidate _after_ writeback.)
2749 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
2750 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2751 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2752 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
2753 if (try_nonblocking_invalidate(inode)) {
2754 /* there were locked pages.. invalidate later
2755 in a separate thread. */
2756 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2757 queue_invalidate = true;
2758 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2763 /* side effects now are allowed */
2764 cap->cap_gen = session->s_cap_gen;
2767 __check_cap_issue(ci, cap, newcaps);
2769 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2770 (issued & CEPH_CAP_AUTH_EXCL) == 0) {
2771 inode->i_mode = le32_to_cpu(grant->mode);
2772 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
2773 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
2774 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2775 from_kuid(&init_user_ns, inode->i_uid),
2776 from_kgid(&init_user_ns, inode->i_gid));
2779 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2780 (issued & CEPH_CAP_LINK_EXCL) == 0) {
2781 set_nlink(inode, le32_to_cpu(grant->nlink));
2782 if (inode->i_nlink == 0 &&
2783 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
2784 deleted_inode = true;
2787 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2788 int len = le32_to_cpu(grant->xattr_len);
2789 u64 version = le64_to_cpu(grant->xattr_version);
2791 if (version > ci->i_xattrs.version) {
2792 dout(" got new xattrs v%llu on %p len %d\n",
2793 version, inode, len);
2794 if (ci->i_xattrs.blob)
2795 ceph_buffer_put(ci->i_xattrs.blob);
2796 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2797 ci->i_xattrs.version = version;
2798 ceph_forget_all_cached_acls(inode);
2802 if (newcaps & CEPH_CAP_ANY_RD) {
2803 /* ctime/mtime/atime? */
2804 ceph_decode_timespec(&mtime, &grant->mtime);
2805 ceph_decode_timespec(&atime, &grant->atime);
2806 ceph_decode_timespec(&ctime, &grant->ctime);
2807 ceph_fill_file_time(inode, issued,
2808 le32_to_cpu(grant->time_warp_seq),
2809 &ctime, &mtime, &atime);
2812 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
2813 /* file layout may have changed */
2814 s64 old_pool = ci->i_layout.pool_id;
2815 struct ceph_string *old_ns;
2817 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
2818 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
2819 lockdep_is_held(&ci->i_ceph_lock));
2820 rcu_assign_pointer(ci->i_layout.pool_ns, *pns);
2822 if (ci->i_layout.pool_id != old_pool || *pns != old_ns)
2823 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
2827 /* size/truncate_seq? */
2828 queue_trunc = ceph_fill_file_size(inode, issued,
2829 le32_to_cpu(grant->truncate_seq),
2830 le64_to_cpu(grant->truncate_size),
2832 /* max size increase? */
2833 if (ci->i_auth_cap == cap && max_size != ci->i_max_size) {
2834 dout("max_size %lld -> %llu\n",
2835 ci->i_max_size, max_size);
2836 ci->i_max_size = max_size;
2837 if (max_size >= ci->i_wanted_max_size) {
2838 ci->i_wanted_max_size = 0; /* reset */
2839 ci->i_requested_max_size = 0;
2845 /* check cap bits */
2846 wanted = __ceph_caps_wanted(ci);
2847 used = __ceph_caps_used(ci);
2848 dirty = __ceph_caps_dirty(ci);
2849 dout(" my wanted = %s, used = %s, dirty %s\n",
2850 ceph_cap_string(wanted),
2851 ceph_cap_string(used),
2852 ceph_cap_string(dirty));
2853 if (wanted != le32_to_cpu(grant->wanted)) {
2854 dout("mds wanted %s -> %s\n",
2855 ceph_cap_string(le32_to_cpu(grant->wanted)),
2856 ceph_cap_string(wanted));
2857 /* imported cap may not have correct mds_wanted */
2858 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
2862 /* revocation, grant, or no-op? */
2863 if (cap->issued & ~newcaps) {
2864 int revoking = cap->issued & ~newcaps;
2866 dout("revocation: %s -> %s (revoking %s)\n",
2867 ceph_cap_string(cap->issued),
2868 ceph_cap_string(newcaps),
2869 ceph_cap_string(revoking));
2870 if (revoking & used & CEPH_CAP_FILE_BUFFER)
2871 writeback = true; /* initiate writeback; will delay ack */
2872 else if (revoking == CEPH_CAP_FILE_CACHE &&
2873 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2875 ; /* do nothing yet, invalidation will be queued */
2876 else if (cap == ci->i_auth_cap)
2877 check_caps = 1; /* check auth cap only */
2879 check_caps = 2; /* check all caps */
2880 cap->issued = newcaps;
2881 cap->implemented |= newcaps;
2882 } else if (cap->issued == newcaps) {
2883 dout("caps unchanged: %s -> %s\n",
2884 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
2886 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
2887 ceph_cap_string(newcaps));
2888 /* non-auth MDS is revoking the newly grant caps ? */
2889 if (cap == ci->i_auth_cap &&
2890 __ceph_caps_revoking_other(ci, cap, newcaps))
2893 cap->issued = newcaps;
2894 cap->implemented |= newcaps; /* add bits only, to
2895 * avoid stepping on a
2896 * pending revocation */
2899 BUG_ON(cap->issued & ~cap->implemented);
2901 if (inline_version > 0 && inline_version >= ci->i_inline_version) {
2902 ci->i_inline_version = inline_version;
2903 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2904 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
2908 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
2909 if (newcaps & ~issued)
2911 kick_flushing_inode_caps(mdsc, session, inode);
2912 up_read(&mdsc->snap_rwsem);
2914 spin_unlock(&ci->i_ceph_lock);
2918 ceph_fill_inline_data(inode, NULL, inline_data, inline_len);
2921 ceph_queue_vmtruncate(inode);
2925 * queue inode for writeback: we can't actually call
2926 * filemap_write_and_wait, etc. from message handler
2929 ceph_queue_writeback(inode);
2930 if (queue_invalidate)
2931 ceph_queue_invalidate(inode);
2933 invalidate_aliases(inode);
2935 wake_up_all(&ci->i_cap_wq);
2937 if (check_caps == 1)
2938 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
2940 else if (check_caps == 2)
2941 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
2943 mutex_unlock(&session->s_mutex);
2947 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
2948 * MDS has been safely committed.
2950 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2951 struct ceph_mds_caps *m,
2952 struct ceph_mds_session *session,
2953 struct ceph_cap *cap)
2954 __releases(ci->i_ceph_lock)
2956 struct ceph_inode_info *ci = ceph_inode(inode);
2957 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2958 struct ceph_cap_flush *cf, *tmp_cf;
2959 LIST_HEAD(to_remove);
2960 unsigned seq = le32_to_cpu(m->seq);
2961 int dirty = le32_to_cpu(m->dirty);
2965 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
2966 if (cf->tid == flush_tid)
2968 if (cf->caps == 0) /* capsnap */
2970 if (cf->tid <= flush_tid) {
2971 list_del(&cf->i_list);
2972 list_add_tail(&cf->i_list, &to_remove);
2974 cleaned &= ~cf->caps;
2980 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2981 " flushing %s -> %s\n",
2982 inode, session->s_mds, seq, ceph_cap_string(dirty),
2983 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2984 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2986 if (list_empty(&to_remove) && !cleaned)
2989 ci->i_flushing_caps &= ~cleaned;
2991 spin_lock(&mdsc->cap_dirty_lock);
2993 if (!list_empty(&to_remove)) {
2994 u64 oldest_flush_tid;
2995 list_for_each_entry(cf, &to_remove, i_list)
2996 list_del(&cf->g_list);
2998 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2999 if (oldest_flush_tid == 0 || oldest_flush_tid > flush_tid)
3000 wake_up_all(&mdsc->cap_flushing_wq);
3003 if (ci->i_flushing_caps == 0) {
3004 if (list_empty(&ci->i_cap_flush_list)) {
3005 list_del_init(&ci->i_flushing_item);
3006 if (!list_empty(&session->s_cap_flushing)) {
3007 dout(" mds%d still flushing cap on %p\n",
3009 &list_first_entry(&session->s_cap_flushing,
3010 struct ceph_inode_info,
3011 i_flushing_item)->vfs_inode);
3014 mdsc->num_cap_flushing--;
3015 dout(" inode %p now !flushing\n", inode);
3017 if (ci->i_dirty_caps == 0) {
3018 dout(" inode %p now clean\n", inode);
3019 BUG_ON(!list_empty(&ci->i_dirty_item));
3021 if (ci->i_wr_ref == 0 &&
3022 ci->i_wrbuffer_ref_head == 0) {
3023 BUG_ON(!ci->i_head_snapc);
3024 ceph_put_snap_context(ci->i_head_snapc);
3025 ci->i_head_snapc = NULL;
3028 BUG_ON(list_empty(&ci->i_dirty_item));
3031 spin_unlock(&mdsc->cap_dirty_lock);
3032 wake_up_all(&ci->i_cap_wq);
3035 spin_unlock(&ci->i_ceph_lock);
3037 while (!list_empty(&to_remove)) {
3038 cf = list_first_entry(&to_remove,
3039 struct ceph_cap_flush, i_list);
3040 list_del(&cf->i_list);
3041 ceph_free_cap_flush(cf);
3048 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3049 * throw away our cap_snap.
3051 * Caller hold s_mutex.
3053 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3054 struct ceph_mds_caps *m,
3055 struct ceph_mds_session *session)
3057 struct ceph_inode_info *ci = ceph_inode(inode);
3058 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3059 u64 follows = le64_to_cpu(m->snap_follows);
3060 struct ceph_cap_snap *capsnap;
3063 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3064 inode, ci, session->s_mds, follows);
3066 spin_lock(&ci->i_ceph_lock);
3067 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3068 if (capsnap->follows == follows) {
3069 if (capsnap->cap_flush.tid != flush_tid) {
3070 dout(" cap_snap %p follows %lld tid %lld !="
3071 " %lld\n", capsnap, follows,
3072 flush_tid, capsnap->cap_flush.tid);
3078 dout(" skipping cap_snap %p follows %lld\n",
3079 capsnap, capsnap->follows);
3083 u64 oldest_flush_tid;
3084 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3085 dout(" removing %p cap_snap %p follows %lld\n",
3086 inode, capsnap, follows);
3087 list_del(&capsnap->ci_item);
3088 list_del(&capsnap->cap_flush.i_list);
3090 spin_lock(&mdsc->cap_dirty_lock);
3092 if (list_empty(&ci->i_cap_flush_list))
3093 list_del_init(&ci->i_flushing_item);
3095 list_del(&capsnap->cap_flush.g_list);
3097 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
3098 if (oldest_flush_tid == 0 || oldest_flush_tid > flush_tid)
3099 wake_up_all(&mdsc->cap_flushing_wq);
3101 spin_unlock(&mdsc->cap_dirty_lock);
3102 wake_up_all(&ci->i_cap_wq);
3104 spin_unlock(&ci->i_ceph_lock);
3106 ceph_put_snap_context(capsnap->context);
3107 ceph_put_cap_snap(capsnap);
3113 * Handle TRUNC from MDS, indicating file truncation.
3115 * caller hold s_mutex.
3117 static void handle_cap_trunc(struct inode *inode,
3118 struct ceph_mds_caps *trunc,
3119 struct ceph_mds_session *session)
3120 __releases(ci->i_ceph_lock)
3122 struct ceph_inode_info *ci = ceph_inode(inode);
3123 int mds = session->s_mds;
3124 int seq = le32_to_cpu(trunc->seq);
3125 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3126 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3127 u64 size = le64_to_cpu(trunc->size);
3128 int implemented = 0;
3129 int dirty = __ceph_caps_dirty(ci);
3130 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3131 int queue_trunc = 0;
3133 issued |= implemented | dirty;
3135 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3136 inode, mds, seq, truncate_size, truncate_seq);
3137 queue_trunc = ceph_fill_file_size(inode, issued,
3138 truncate_seq, truncate_size, size);
3139 spin_unlock(&ci->i_ceph_lock);
3142 ceph_queue_vmtruncate(inode);
3146 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3147 * different one. If we are the most recent migration we've seen (as
3148 * indicated by mseq), make note of the migrating cap bits for the
3149 * duration (until we see the corresponding IMPORT).
3151 * caller holds s_mutex
3153 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3154 struct ceph_mds_cap_peer *ph,
3155 struct ceph_mds_session *session)
3157 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3158 struct ceph_mds_session *tsession = NULL;
3159 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3160 struct ceph_inode_info *ci = ceph_inode(inode);
3162 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3163 unsigned t_seq, t_mseq;
3165 int mds = session->s_mds;
3168 t_cap_id = le64_to_cpu(ph->cap_id);
3169 t_seq = le32_to_cpu(ph->seq);
3170 t_mseq = le32_to_cpu(ph->mseq);
3171 target = le32_to_cpu(ph->mds);
3173 t_cap_id = t_seq = t_mseq = 0;
3177 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3178 inode, ci, mds, mseq, target);
3180 spin_lock(&ci->i_ceph_lock);
3181 cap = __get_cap_for_mds(ci, mds);
3182 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3186 __ceph_remove_cap(cap, false);
3187 if (!ci->i_auth_cap)
3188 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
3193 * now we know we haven't received the cap import message yet
3194 * because the exported cap still exist.
3197 issued = cap->issued;
3198 WARN_ON(issued != cap->implemented);
3200 tcap = __get_cap_for_mds(ci, target);
3202 /* already have caps from the target */
3203 if (tcap->cap_id != t_cap_id ||
3204 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3205 dout(" updating import cap %p mds%d\n", tcap, target);
3206 tcap->cap_id = t_cap_id;
3207 tcap->seq = t_seq - 1;
3208 tcap->issue_seq = t_seq - 1;
3209 tcap->mseq = t_mseq;
3210 tcap->issued |= issued;
3211 tcap->implemented |= issued;
3212 if (cap == ci->i_auth_cap)
3213 ci->i_auth_cap = tcap;
3214 if (!list_empty(&ci->i_cap_flush_list) &&
3215 ci->i_auth_cap == tcap) {
3216 spin_lock(&mdsc->cap_dirty_lock);
3217 list_move_tail(&ci->i_flushing_item,
3218 &tcap->session->s_cap_flushing);
3219 spin_unlock(&mdsc->cap_dirty_lock);
3222 __ceph_remove_cap(cap, false);
3224 } else if (tsession) {
3225 /* add placeholder for the export tagert */
3226 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3227 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3228 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3230 __ceph_remove_cap(cap, false);
3234 spin_unlock(&ci->i_ceph_lock);
3235 mutex_unlock(&session->s_mutex);
3237 /* open target session */
3238 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3239 if (!IS_ERR(tsession)) {
3241 mutex_lock(&session->s_mutex);
3242 mutex_lock_nested(&tsession->s_mutex,
3243 SINGLE_DEPTH_NESTING);
3245 mutex_lock(&tsession->s_mutex);
3246 mutex_lock_nested(&session->s_mutex,
3247 SINGLE_DEPTH_NESTING);
3249 new_cap = ceph_get_cap(mdsc, NULL);
3258 spin_unlock(&ci->i_ceph_lock);
3259 mutex_unlock(&session->s_mutex);
3261 mutex_unlock(&tsession->s_mutex);
3262 ceph_put_mds_session(tsession);
3265 ceph_put_cap(mdsc, new_cap);
3269 * Handle cap IMPORT.
3271 * caller holds s_mutex. acquires i_ceph_lock
3273 static void handle_cap_import(struct ceph_mds_client *mdsc,
3274 struct inode *inode, struct ceph_mds_caps *im,
3275 struct ceph_mds_cap_peer *ph,
3276 struct ceph_mds_session *session,
3277 struct ceph_cap **target_cap, int *old_issued)
3278 __acquires(ci->i_ceph_lock)
3280 struct ceph_inode_info *ci = ceph_inode(inode);
3281 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3282 int mds = session->s_mds;
3284 unsigned caps = le32_to_cpu(im->caps);
3285 unsigned wanted = le32_to_cpu(im->wanted);
3286 unsigned seq = le32_to_cpu(im->seq);
3287 unsigned mseq = le32_to_cpu(im->migrate_seq);
3288 u64 realmino = le64_to_cpu(im->realm);
3289 u64 cap_id = le64_to_cpu(im->cap_id);
3294 p_cap_id = le64_to_cpu(ph->cap_id);
3295 peer = le32_to_cpu(ph->mds);
3301 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3302 inode, ci, mds, mseq, peer);
3305 spin_lock(&ci->i_ceph_lock);
3306 cap = __get_cap_for_mds(ci, mds);
3309 spin_unlock(&ci->i_ceph_lock);
3310 new_cap = ceph_get_cap(mdsc, NULL);
3316 ceph_put_cap(mdsc, new_cap);
3321 __ceph_caps_issued(ci, &issued);
3322 issued |= __ceph_caps_dirty(ci);
3324 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3325 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3327 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3328 if (ocap && ocap->cap_id == p_cap_id) {
3329 dout(" remove export cap %p mds%d flags %d\n",
3330 ocap, peer, ph->flags);
3331 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3332 (ocap->seq != le32_to_cpu(ph->seq) ||
3333 ocap->mseq != le32_to_cpu(ph->mseq))) {
3334 pr_err("handle_cap_import: mismatched seq/mseq: "
3335 "ino (%llx.%llx) mds%d seq %d mseq %d "
3336 "importer mds%d has peer seq %d mseq %d\n",
3337 ceph_vinop(inode), peer, ocap->seq,
3338 ocap->mseq, mds, le32_to_cpu(ph->seq),
3339 le32_to_cpu(ph->mseq));
3341 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3344 /* make sure we re-request max_size, if necessary */
3345 ci->i_wanted_max_size = 0;
3346 ci->i_requested_max_size = 0;
3348 *old_issued = issued;
3353 * Handle a caps message from the MDS.
3355 * Identify the appropriate session, inode, and call the right handler
3356 * based on the cap op.
3358 void ceph_handle_caps(struct ceph_mds_session *session,
3359 struct ceph_msg *msg)
3361 struct ceph_mds_client *mdsc = session->s_mdsc;
3362 struct super_block *sb = mdsc->fsc->sb;
3363 struct inode *inode;
3364 struct ceph_inode_info *ci;
3365 struct ceph_cap *cap;
3366 struct ceph_mds_caps *h;
3367 struct ceph_mds_cap_peer *peer = NULL;
3368 struct ceph_snap_realm *realm = NULL;
3369 struct ceph_string *pool_ns = NULL;
3370 int mds = session->s_mds;
3373 struct ceph_vino vino;
3375 u64 inline_version = 0;
3376 void *inline_data = NULL;
3379 size_t snaptrace_len;
3382 dout("handle_caps from mds%d\n", mds);
3385 end = msg->front.iov_base + msg->front.iov_len;
3386 tid = le64_to_cpu(msg->hdr.tid);
3387 if (msg->front.iov_len < sizeof(*h))
3389 h = msg->front.iov_base;
3390 op = le32_to_cpu(h->op);
3391 vino.ino = le64_to_cpu(h->ino);
3392 vino.snap = CEPH_NOSNAP;
3393 seq = le32_to_cpu(h->seq);
3394 mseq = le32_to_cpu(h->migrate_seq);
3397 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3398 p = snaptrace + snaptrace_len;
3400 if (le16_to_cpu(msg->hdr.version) >= 2) {
3402 ceph_decode_32_safe(&p, end, flock_len, bad);
3403 if (p + flock_len > end)
3408 if (le16_to_cpu(msg->hdr.version) >= 3) {
3409 if (op == CEPH_CAP_OP_IMPORT) {
3410 if (p + sizeof(*peer) > end)
3414 } else if (op == CEPH_CAP_OP_EXPORT) {
3415 /* recorded in unused fields */
3416 peer = (void *)&h->size;
3420 if (le16_to_cpu(msg->hdr.version) >= 4) {
3421 ceph_decode_64_safe(&p, end, inline_version, bad);
3422 ceph_decode_32_safe(&p, end, inline_len, bad);
3423 if (p + inline_len > end)
3429 if (le16_to_cpu(msg->hdr.version) >= 8) {
3431 u32 caller_uid, caller_gid;
3432 u32 osd_epoch_barrier;
3435 ceph_decode_32_safe(&p, end, osd_epoch_barrier, bad);
3437 ceph_decode_64_safe(&p, end, flush_tid, bad);
3439 ceph_decode_32_safe(&p, end, caller_uid, bad);
3440 ceph_decode_32_safe(&p, end, caller_gid, bad);
3442 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
3443 if (pool_ns_len > 0) {
3444 ceph_decode_need(&p, end, pool_ns_len, bad);
3445 pool_ns = ceph_find_or_create_string(p, pool_ns_len);
3451 inode = ceph_find_inode(sb, vino);
3452 ci = ceph_inode(inode);
3453 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3456 mutex_lock(&session->s_mutex);
3458 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3462 dout(" i don't have ino %llx\n", vino.ino);
3464 if (op == CEPH_CAP_OP_IMPORT) {
3465 cap = ceph_get_cap(mdsc, NULL);
3466 cap->cap_ino = vino.ino;
3467 cap->queue_release = 1;
3468 cap->cap_id = le64_to_cpu(h->cap_id);
3471 spin_lock(&session->s_cap_lock);
3472 list_add_tail(&cap->session_caps,
3473 &session->s_cap_releases);
3474 session->s_num_cap_releases++;
3475 spin_unlock(&session->s_cap_lock);
3477 goto flush_cap_releases;
3480 /* these will work even if we don't have a cap yet */
3482 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3483 handle_cap_flushsnap_ack(inode, tid, h, session);
3486 case CEPH_CAP_OP_EXPORT:
3487 handle_cap_export(inode, h, peer, session);
3490 case CEPH_CAP_OP_IMPORT:
3492 if (snaptrace_len) {
3493 down_write(&mdsc->snap_rwsem);
3494 ceph_update_snap_trace(mdsc, snaptrace,
3495 snaptrace + snaptrace_len,
3497 downgrade_write(&mdsc->snap_rwsem);
3499 down_read(&mdsc->snap_rwsem);
3501 handle_cap_import(mdsc, inode, h, peer, session,
3503 handle_cap_grant(mdsc, inode, h, &pool_ns,
3504 inline_version, inline_data, inline_len,
3505 msg->middle, session, cap, issued);
3507 ceph_put_snap_realm(mdsc, realm);
3511 /* the rest require a cap */
3512 spin_lock(&ci->i_ceph_lock);
3513 cap = __get_cap_for_mds(ceph_inode(inode), mds);
3515 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3516 inode, ceph_ino(inode), ceph_snap(inode), mds);
3517 spin_unlock(&ci->i_ceph_lock);
3518 goto flush_cap_releases;
3521 /* note that each of these drops i_ceph_lock for us */
3523 case CEPH_CAP_OP_REVOKE:
3524 case CEPH_CAP_OP_GRANT:
3525 __ceph_caps_issued(ci, &issued);
3526 issued |= __ceph_caps_dirty(ci);
3527 handle_cap_grant(mdsc, inode, h, &pool_ns,
3528 inline_version, inline_data, inline_len,
3529 msg->middle, session, cap, issued);
3532 case CEPH_CAP_OP_FLUSH_ACK:
3533 handle_cap_flush_ack(inode, tid, h, session, cap);
3536 case CEPH_CAP_OP_TRUNC:
3537 handle_cap_trunc(inode, h, session);
3541 spin_unlock(&ci->i_ceph_lock);
3542 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3543 ceph_cap_op_name(op));
3550 * send any cap release message to try to move things
3551 * along for the mds (who clearly thinks we still have this
3554 ceph_send_cap_releases(mdsc, session);
3557 mutex_unlock(&session->s_mutex);
3560 ceph_put_string(pool_ns);
3564 pr_err("ceph_handle_caps: corrupt message\n");
3570 * Delayed work handler to process end of delayed cap release LRU list.
3572 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3574 struct ceph_inode_info *ci;
3575 int flags = CHECK_CAPS_NODELAY;
3577 dout("check_delayed_caps\n");
3579 spin_lock(&mdsc->cap_delay_lock);
3580 if (list_empty(&mdsc->cap_delay_list))
3582 ci = list_first_entry(&mdsc->cap_delay_list,
3583 struct ceph_inode_info,
3585 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3586 time_before(jiffies, ci->i_hold_caps_max))
3588 list_del_init(&ci->i_cap_delay_list);
3589 spin_unlock(&mdsc->cap_delay_lock);
3590 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
3591 ceph_check_caps(ci, flags, NULL);
3593 spin_unlock(&mdsc->cap_delay_lock);
3597 * Flush all dirty caps to the mds
3599 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
3601 struct ceph_inode_info *ci;
3602 struct inode *inode;
3604 dout("flush_dirty_caps\n");
3605 spin_lock(&mdsc->cap_dirty_lock);
3606 while (!list_empty(&mdsc->cap_dirty)) {
3607 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
3609 inode = &ci->vfs_inode;
3611 dout("flush_dirty_caps %p\n", inode);
3612 spin_unlock(&mdsc->cap_dirty_lock);
3613 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
3615 spin_lock(&mdsc->cap_dirty_lock);
3617 spin_unlock(&mdsc->cap_dirty_lock);
3618 dout("flush_dirty_caps done\n");
3621 void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode)
3624 int bits = (fmode << 1) | 1;
3625 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
3626 if (bits & (1 << i))
3627 ci->i_nr_by_mode[i]++;
3632 * Drop open file reference. If we were the last open file,
3633 * we may need to release capabilities to the MDS (or schedule
3634 * their delayed release).
3636 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
3639 int bits = (fmode << 1) | 1;
3640 spin_lock(&ci->i_ceph_lock);
3641 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
3642 if (bits & (1 << i)) {
3643 BUG_ON(ci->i_nr_by_mode[i] == 0);
3644 if (--ci->i_nr_by_mode[i] == 0)
3648 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
3649 &ci->vfs_inode, fmode,
3650 ci->i_nr_by_mode[0], ci->i_nr_by_mode[1],
3651 ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]);
3652 spin_unlock(&ci->i_ceph_lock);
3654 if (last && ci->i_vino.snap == CEPH_NOSNAP)
3655 ceph_check_caps(ci, 0, NULL);
3659 * Helpers for embedding cap and dentry lease releases into mds
3662 * @force is used by dentry_release (below) to force inclusion of a
3663 * record for the directory inode, even when there aren't any caps to
3666 int ceph_encode_inode_release(void **p, struct inode *inode,
3667 int mds, int drop, int unless, int force)
3669 struct ceph_inode_info *ci = ceph_inode(inode);
3670 struct ceph_cap *cap;
3671 struct ceph_mds_request_release *rel = *p;
3675 spin_lock(&ci->i_ceph_lock);
3676 used = __ceph_caps_used(ci);
3677 dirty = __ceph_caps_dirty(ci);
3679 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3680 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3681 ceph_cap_string(unless));
3683 /* only drop unused, clean caps */
3684 drop &= ~(used | dirty);
3686 cap = __get_cap_for_mds(ci, mds);
3687 if (cap && __cap_is_valid(cap)) {
3689 ((cap->issued & drop) &&
3690 (cap->issued & unless) == 0)) {
3691 if ((cap->issued & drop) &&
3692 (cap->issued & unless) == 0) {
3693 int wanted = __ceph_caps_wanted(ci);
3694 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
3695 wanted |= cap->mds_wanted;
3696 dout("encode_inode_release %p cap %p "
3697 "%s -> %s, wanted %s -> %s\n", inode, cap,
3698 ceph_cap_string(cap->issued),
3699 ceph_cap_string(cap->issued & ~drop),
3700 ceph_cap_string(cap->mds_wanted),
3701 ceph_cap_string(wanted));
3703 cap->issued &= ~drop;
3704 cap->implemented &= ~drop;
3705 cap->mds_wanted = wanted;
3707 dout("encode_inode_release %p cap %p %s"
3708 " (force)\n", inode, cap,
3709 ceph_cap_string(cap->issued));
3712 rel->ino = cpu_to_le64(ceph_ino(inode));
3713 rel->cap_id = cpu_to_le64(cap->cap_id);
3714 rel->seq = cpu_to_le32(cap->seq);
3715 rel->issue_seq = cpu_to_le32(cap->issue_seq);
3716 rel->mseq = cpu_to_le32(cap->mseq);
3717 rel->caps = cpu_to_le32(cap->implemented);
3718 rel->wanted = cpu_to_le32(cap->mds_wanted);
3724 dout("encode_inode_release %p cap %p %s\n",
3725 inode, cap, ceph_cap_string(cap->issued));
3728 spin_unlock(&ci->i_ceph_lock);
3732 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3733 int mds, int drop, int unless)
3735 struct inode *dir = d_inode(dentry->d_parent);
3736 struct ceph_mds_request_release *rel = *p;
3737 struct ceph_dentry_info *di = ceph_dentry(dentry);
3742 * force an record for the directory caps if we have a dentry lease.
3743 * this is racy (can't take i_ceph_lock and d_lock together), but it
3744 * doesn't have to be perfect; the mds will revoke anything we don't
3747 spin_lock(&dentry->d_lock);
3748 if (di->lease_session && di->lease_session->s_mds == mds)
3750 spin_unlock(&dentry->d_lock);
3752 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3754 spin_lock(&dentry->d_lock);
3755 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3756 dout("encode_dentry_release %p mds%d seq %d\n",
3757 dentry, mds, (int)di->lease_seq);
3758 rel->dname_len = cpu_to_le32(dentry->d_name.len);
3759 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3760 *p += dentry->d_name.len;
3761 rel->dname_seq = cpu_to_le32(di->lease_seq);
3762 __ceph_mdsc_drop_dentry_lease(dentry);
3764 spin_unlock(&dentry->d_lock);