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);
44 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
45 struct ceph_mds_session *session,
46 struct ceph_inode_info *ci,
47 u64 oldest_flush_tid);
50 * Generate readable cap strings for debugging output.
52 #define MAX_CAP_STR 20
53 static char cap_str[MAX_CAP_STR][40];
54 static DEFINE_SPINLOCK(cap_str_lock);
55 static int last_cap_str;
57 static char *gcap_string(char *s, int c)
59 if (c & CEPH_CAP_GSHARED)
61 if (c & CEPH_CAP_GEXCL)
63 if (c & CEPH_CAP_GCACHE)
69 if (c & CEPH_CAP_GBUFFER)
71 if (c & CEPH_CAP_GLAZYIO)
76 const char *ceph_cap_string(int caps)
82 spin_lock(&cap_str_lock);
84 if (last_cap_str == MAX_CAP_STR)
86 spin_unlock(&cap_str_lock);
90 if (caps & CEPH_CAP_PIN)
93 c = (caps >> CEPH_CAP_SAUTH) & 3;
96 s = gcap_string(s, c);
99 c = (caps >> CEPH_CAP_SLINK) & 3;
102 s = gcap_string(s, c);
105 c = (caps >> CEPH_CAP_SXATTR) & 3;
108 s = gcap_string(s, c);
111 c = caps >> CEPH_CAP_SFILE;
114 s = gcap_string(s, c);
123 void ceph_caps_init(struct ceph_mds_client *mdsc)
125 INIT_LIST_HEAD(&mdsc->caps_list);
126 spin_lock_init(&mdsc->caps_list_lock);
129 void ceph_caps_finalize(struct ceph_mds_client *mdsc)
131 struct ceph_cap *cap;
133 spin_lock(&mdsc->caps_list_lock);
134 while (!list_empty(&mdsc->caps_list)) {
135 cap = list_first_entry(&mdsc->caps_list,
136 struct ceph_cap, caps_item);
137 list_del(&cap->caps_item);
138 kmem_cache_free(ceph_cap_cachep, cap);
140 mdsc->caps_total_count = 0;
141 mdsc->caps_avail_count = 0;
142 mdsc->caps_use_count = 0;
143 mdsc->caps_reserve_count = 0;
144 mdsc->caps_min_count = 0;
145 spin_unlock(&mdsc->caps_list_lock);
148 void ceph_adjust_min_caps(struct ceph_mds_client *mdsc, int delta)
150 spin_lock(&mdsc->caps_list_lock);
151 mdsc->caps_min_count += delta;
152 BUG_ON(mdsc->caps_min_count < 0);
153 spin_unlock(&mdsc->caps_list_lock);
156 void ceph_reserve_caps(struct ceph_mds_client *mdsc,
157 struct ceph_cap_reservation *ctx, int need)
160 struct ceph_cap *cap;
165 dout("reserve caps ctx=%p need=%d\n", ctx, need);
167 /* first reserve any caps that are already allocated */
168 spin_lock(&mdsc->caps_list_lock);
169 if (mdsc->caps_avail_count >= need)
172 have = mdsc->caps_avail_count;
173 mdsc->caps_avail_count -= have;
174 mdsc->caps_reserve_count += have;
175 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
176 mdsc->caps_reserve_count +
177 mdsc->caps_avail_count);
178 spin_unlock(&mdsc->caps_list_lock);
180 for (i = have; i < need; i++) {
181 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
184 list_add(&cap->caps_item, &newcaps);
187 /* we didn't manage to reserve as much as we needed */
188 if (have + alloc != need)
189 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
190 ctx, need, have + alloc);
192 spin_lock(&mdsc->caps_list_lock);
193 mdsc->caps_total_count += alloc;
194 mdsc->caps_reserve_count += alloc;
195 list_splice(&newcaps, &mdsc->caps_list);
197 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
198 mdsc->caps_reserve_count +
199 mdsc->caps_avail_count);
200 spin_unlock(&mdsc->caps_list_lock);
203 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
204 ctx, mdsc->caps_total_count, mdsc->caps_use_count,
205 mdsc->caps_reserve_count, mdsc->caps_avail_count);
208 int ceph_unreserve_caps(struct ceph_mds_client *mdsc,
209 struct ceph_cap_reservation *ctx)
211 dout("unreserve caps ctx=%p count=%d\n", ctx, ctx->count);
213 spin_lock(&mdsc->caps_list_lock);
214 BUG_ON(mdsc->caps_reserve_count < ctx->count);
215 mdsc->caps_reserve_count -= ctx->count;
216 mdsc->caps_avail_count += ctx->count;
218 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
219 mdsc->caps_total_count, mdsc->caps_use_count,
220 mdsc->caps_reserve_count, mdsc->caps_avail_count);
221 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
222 mdsc->caps_reserve_count +
223 mdsc->caps_avail_count);
224 spin_unlock(&mdsc->caps_list_lock);
229 struct ceph_cap *ceph_get_cap(struct ceph_mds_client *mdsc,
230 struct ceph_cap_reservation *ctx)
232 struct ceph_cap *cap = NULL;
234 /* temporary, until we do something about cap import/export */
236 cap = kmem_cache_alloc(ceph_cap_cachep, GFP_NOFS);
238 spin_lock(&mdsc->caps_list_lock);
239 mdsc->caps_use_count++;
240 mdsc->caps_total_count++;
241 spin_unlock(&mdsc->caps_list_lock);
246 spin_lock(&mdsc->caps_list_lock);
247 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
248 ctx, ctx->count, mdsc->caps_total_count, mdsc->caps_use_count,
249 mdsc->caps_reserve_count, mdsc->caps_avail_count);
251 BUG_ON(ctx->count > mdsc->caps_reserve_count);
252 BUG_ON(list_empty(&mdsc->caps_list));
255 mdsc->caps_reserve_count--;
256 mdsc->caps_use_count++;
258 cap = list_first_entry(&mdsc->caps_list, struct ceph_cap, caps_item);
259 list_del(&cap->caps_item);
261 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
262 mdsc->caps_reserve_count + mdsc->caps_avail_count);
263 spin_unlock(&mdsc->caps_list_lock);
267 void ceph_put_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap)
269 spin_lock(&mdsc->caps_list_lock);
270 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
271 cap, mdsc->caps_total_count, mdsc->caps_use_count,
272 mdsc->caps_reserve_count, mdsc->caps_avail_count);
273 mdsc->caps_use_count--;
275 * Keep some preallocated caps around (ceph_min_count), to
276 * avoid lots of free/alloc churn.
278 if (mdsc->caps_avail_count >= mdsc->caps_reserve_count +
279 mdsc->caps_min_count) {
280 mdsc->caps_total_count--;
281 kmem_cache_free(ceph_cap_cachep, cap);
283 mdsc->caps_avail_count++;
284 list_add(&cap->caps_item, &mdsc->caps_list);
287 BUG_ON(mdsc->caps_total_count != mdsc->caps_use_count +
288 mdsc->caps_reserve_count + mdsc->caps_avail_count);
289 spin_unlock(&mdsc->caps_list_lock);
292 void ceph_reservation_status(struct ceph_fs_client *fsc,
293 int *total, int *avail, int *used, int *reserved,
296 struct ceph_mds_client *mdsc = fsc->mdsc;
299 *total = mdsc->caps_total_count;
301 *avail = mdsc->caps_avail_count;
303 *used = mdsc->caps_use_count;
305 *reserved = mdsc->caps_reserve_count;
307 *min = mdsc->caps_min_count;
311 * Find ceph_cap for given mds, if any.
313 * Called with i_ceph_lock held.
315 static struct ceph_cap *__get_cap_for_mds(struct ceph_inode_info *ci, int mds)
317 struct ceph_cap *cap;
318 struct rb_node *n = ci->i_caps.rb_node;
321 cap = rb_entry(n, struct ceph_cap, ci_node);
324 else if (mds > cap->mds)
332 struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, int mds)
334 struct ceph_cap *cap;
336 spin_lock(&ci->i_ceph_lock);
337 cap = __get_cap_for_mds(ci, mds);
338 spin_unlock(&ci->i_ceph_lock);
343 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
345 static int __ceph_get_cap_mds(struct ceph_inode_info *ci)
347 struct ceph_cap *cap;
351 /* prefer mds with WR|BUFFER|EXCL caps */
352 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
353 cap = rb_entry(p, struct ceph_cap, ci_node);
355 if (cap->issued & (CEPH_CAP_FILE_WR |
356 CEPH_CAP_FILE_BUFFER |
363 int ceph_get_cap_mds(struct inode *inode)
365 struct ceph_inode_info *ci = ceph_inode(inode);
367 spin_lock(&ci->i_ceph_lock);
368 mds = __ceph_get_cap_mds(ceph_inode(inode));
369 spin_unlock(&ci->i_ceph_lock);
374 * Called under i_ceph_lock.
376 static void __insert_cap_node(struct ceph_inode_info *ci,
377 struct ceph_cap *new)
379 struct rb_node **p = &ci->i_caps.rb_node;
380 struct rb_node *parent = NULL;
381 struct ceph_cap *cap = NULL;
385 cap = rb_entry(parent, struct ceph_cap, ci_node);
386 if (new->mds < cap->mds)
388 else if (new->mds > cap->mds)
394 rb_link_node(&new->ci_node, parent, p);
395 rb_insert_color(&new->ci_node, &ci->i_caps);
399 * (re)set cap hold timeouts, which control the delayed release
400 * of unused caps back to the MDS. Should be called on cap use.
402 static void __cap_set_timeouts(struct ceph_mds_client *mdsc,
403 struct ceph_inode_info *ci)
405 struct ceph_mount_options *ma = mdsc->fsc->mount_options;
407 ci->i_hold_caps_min = round_jiffies(jiffies +
408 ma->caps_wanted_delay_min * HZ);
409 ci->i_hold_caps_max = round_jiffies(jiffies +
410 ma->caps_wanted_delay_max * HZ);
411 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci->vfs_inode,
412 ci->i_hold_caps_min - jiffies, ci->i_hold_caps_max - jiffies);
416 * (Re)queue cap at the end of the delayed cap release list.
418 * If I_FLUSH is set, leave the inode at the front of the list.
420 * Caller holds i_ceph_lock
421 * -> we take mdsc->cap_delay_lock
423 static void __cap_delay_requeue(struct ceph_mds_client *mdsc,
424 struct ceph_inode_info *ci)
426 __cap_set_timeouts(mdsc, ci);
427 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci->vfs_inode,
428 ci->i_ceph_flags, ci->i_hold_caps_max);
429 if (!mdsc->stopping) {
430 spin_lock(&mdsc->cap_delay_lock);
431 if (!list_empty(&ci->i_cap_delay_list)) {
432 if (ci->i_ceph_flags & CEPH_I_FLUSH)
434 list_del_init(&ci->i_cap_delay_list);
436 list_add_tail(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
438 spin_unlock(&mdsc->cap_delay_lock);
443 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
444 * indicating we should send a cap message to flush dirty metadata
445 * asap, and move to the front of the delayed cap list.
447 static void __cap_delay_requeue_front(struct ceph_mds_client *mdsc,
448 struct ceph_inode_info *ci)
450 dout("__cap_delay_requeue_front %p\n", &ci->vfs_inode);
451 spin_lock(&mdsc->cap_delay_lock);
452 ci->i_ceph_flags |= CEPH_I_FLUSH;
453 if (!list_empty(&ci->i_cap_delay_list))
454 list_del_init(&ci->i_cap_delay_list);
455 list_add(&ci->i_cap_delay_list, &mdsc->cap_delay_list);
456 spin_unlock(&mdsc->cap_delay_lock);
460 * Cancel delayed work on cap.
462 * Caller must hold i_ceph_lock.
464 static void __cap_delay_cancel(struct ceph_mds_client *mdsc,
465 struct ceph_inode_info *ci)
467 dout("__cap_delay_cancel %p\n", &ci->vfs_inode);
468 if (list_empty(&ci->i_cap_delay_list))
470 spin_lock(&mdsc->cap_delay_lock);
471 list_del_init(&ci->i_cap_delay_list);
472 spin_unlock(&mdsc->cap_delay_lock);
476 * Common issue checks for add_cap, handle_cap_grant.
478 static void __check_cap_issue(struct ceph_inode_info *ci, struct ceph_cap *cap,
481 unsigned had = __ceph_caps_issued(ci, NULL);
484 * Each time we receive FILE_CACHE anew, we increment
487 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
488 (had & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0) {
493 * if we are newly issued FILE_SHARED, mark dir not complete; we
494 * don't know what happened to this directory while we didn't
497 if ((issued & CEPH_CAP_FILE_SHARED) &&
498 (had & CEPH_CAP_FILE_SHARED) == 0) {
500 if (S_ISDIR(ci->vfs_inode.i_mode)) {
501 dout(" marking %p NOT complete\n", &ci->vfs_inode);
502 __ceph_dir_clear_complete(ci);
508 * Add a capability under the given MDS session.
510 * Caller should hold session snap_rwsem (read) and s_mutex.
512 * @fmode is the open file mode, if we are opening a file, otherwise
513 * it is < 0. (This is so we can atomically add the cap and add an
514 * open file reference to it.)
516 void ceph_add_cap(struct inode *inode,
517 struct ceph_mds_session *session, u64 cap_id,
518 int fmode, unsigned issued, unsigned wanted,
519 unsigned seq, unsigned mseq, u64 realmino, int flags,
520 struct ceph_cap **new_cap)
522 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
523 struct ceph_inode_info *ci = ceph_inode(inode);
524 struct ceph_cap *cap;
525 int mds = session->s_mds;
528 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode,
529 session->s_mds, cap_id, ceph_cap_string(issued), seq);
532 * If we are opening the file, include file mode wanted bits
536 wanted |= ceph_caps_for_mode(fmode);
538 cap = __get_cap_for_mds(ci, mds);
544 cap->implemented = 0;
550 __insert_cap_node(ci, cap);
552 /* add to session cap list */
553 cap->session = session;
554 spin_lock(&session->s_cap_lock);
555 list_add_tail(&cap->session_caps, &session->s_caps);
556 session->s_nr_caps++;
557 spin_unlock(&session->s_cap_lock);
560 * auth mds of the inode changed. we received the cap export
561 * message, but still haven't received the cap import message.
562 * handle_cap_export() updated the new auth MDS' cap.
564 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
565 * a message that was send before the cap import message. So
568 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
569 WARN_ON(cap != ci->i_auth_cap);
570 WARN_ON(cap->cap_id != cap_id);
573 issued |= cap->issued;
574 flags |= CEPH_CAP_FLAG_AUTH;
578 if (!ci->i_snap_realm) {
580 * add this inode to the appropriate snap realm
582 struct ceph_snap_realm *realm = ceph_lookup_snap_realm(mdsc,
585 spin_lock(&realm->inodes_with_caps_lock);
586 ci->i_snap_realm = realm;
587 list_add(&ci->i_snap_realm_item,
588 &realm->inodes_with_caps);
589 spin_unlock(&realm->inodes_with_caps_lock);
591 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
597 __check_cap_issue(ci, cap, issued);
600 * If we are issued caps we don't want, or the mds' wanted
601 * value appears to be off, queue a check so we'll release
602 * later and/or update the mds wanted value.
604 actual_wanted = __ceph_caps_wanted(ci);
605 if ((wanted & ~actual_wanted) ||
606 (issued & ~actual_wanted & CEPH_CAP_ANY_WR)) {
607 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
608 ceph_cap_string(issued), ceph_cap_string(wanted),
609 ceph_cap_string(actual_wanted));
610 __cap_delay_requeue(mdsc, ci);
613 if (flags & CEPH_CAP_FLAG_AUTH) {
614 if (ci->i_auth_cap == NULL ||
615 ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
616 ci->i_auth_cap = cap;
617 cap->mds_wanted = wanted;
620 WARN_ON(ci->i_auth_cap == cap);
623 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
624 inode, ceph_vinop(inode), cap, ceph_cap_string(issued),
625 ceph_cap_string(issued|cap->issued), seq, mds);
626 cap->cap_id = cap_id;
627 cap->issued = issued;
628 cap->implemented |= issued;
629 if (ceph_seq_cmp(mseq, cap->mseq) > 0)
630 cap->mds_wanted = wanted;
632 cap->mds_wanted |= wanted;
634 cap->issue_seq = seq;
636 cap->cap_gen = session->s_cap_gen;
639 __ceph_get_fmode(ci, fmode);
643 * Return true if cap has not timed out and belongs to the current
644 * generation of the MDS session (i.e. has not gone 'stale' due to
645 * us losing touch with the mds).
647 static int __cap_is_valid(struct ceph_cap *cap)
652 spin_lock(&cap->session->s_gen_ttl_lock);
653 gen = cap->session->s_cap_gen;
654 ttl = cap->session->s_cap_ttl;
655 spin_unlock(&cap->session->s_gen_ttl_lock);
657 if (cap->cap_gen < gen || time_after_eq(jiffies, ttl)) {
658 dout("__cap_is_valid %p cap %p issued %s "
659 "but STALE (gen %u vs %u)\n", &cap->ci->vfs_inode,
660 cap, ceph_cap_string(cap->issued), cap->cap_gen, gen);
668 * Return set of valid cap bits issued to us. Note that caps time
669 * out, and may be invalidated in bulk if the client session times out
670 * and session->s_cap_gen is bumped.
672 int __ceph_caps_issued(struct ceph_inode_info *ci, int *implemented)
674 int have = ci->i_snap_caps;
675 struct ceph_cap *cap;
680 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
681 cap = rb_entry(p, struct ceph_cap, ci_node);
682 if (!__cap_is_valid(cap))
684 dout("__ceph_caps_issued %p cap %p issued %s\n",
685 &ci->vfs_inode, cap, ceph_cap_string(cap->issued));
688 *implemented |= cap->implemented;
691 * exclude caps issued by non-auth MDS, but are been revoking
692 * by the auth MDS. The non-auth MDS should be revoking/exporting
693 * these caps, but the message is delayed.
695 if (ci->i_auth_cap) {
696 cap = ci->i_auth_cap;
697 have &= ~cap->implemented | cap->issued;
703 * Get cap bits issued by caps other than @ocap
705 int __ceph_caps_issued_other(struct ceph_inode_info *ci, struct ceph_cap *ocap)
707 int have = ci->i_snap_caps;
708 struct ceph_cap *cap;
711 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
712 cap = rb_entry(p, struct ceph_cap, ci_node);
715 if (!__cap_is_valid(cap))
723 * Move a cap to the end of the LRU (oldest caps at list head, newest
726 static void __touch_cap(struct ceph_cap *cap)
728 struct ceph_mds_session *s = cap->session;
730 spin_lock(&s->s_cap_lock);
731 if (s->s_cap_iterator == NULL) {
732 dout("__touch_cap %p cap %p mds%d\n", &cap->ci->vfs_inode, cap,
734 list_move_tail(&cap->session_caps, &s->s_caps);
736 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
737 &cap->ci->vfs_inode, cap, s->s_mds);
739 spin_unlock(&s->s_cap_lock);
743 * Check if we hold the given mask. If so, move the cap(s) to the
744 * front of their respective LRUs. (This is the preferred way for
745 * callers to check for caps they want.)
747 int __ceph_caps_issued_mask(struct ceph_inode_info *ci, int mask, int touch)
749 struct ceph_cap *cap;
751 int have = ci->i_snap_caps;
753 if ((have & mask) == mask) {
754 dout("__ceph_caps_issued_mask %p snap issued %s"
755 " (mask %s)\n", &ci->vfs_inode,
756 ceph_cap_string(have),
757 ceph_cap_string(mask));
761 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
762 cap = rb_entry(p, struct ceph_cap, ci_node);
763 if (!__cap_is_valid(cap))
765 if ((cap->issued & mask) == mask) {
766 dout("__ceph_caps_issued_mask %p cap %p issued %s"
767 " (mask %s)\n", &ci->vfs_inode, cap,
768 ceph_cap_string(cap->issued),
769 ceph_cap_string(mask));
775 /* does a combination of caps satisfy mask? */
777 if ((have & mask) == mask) {
778 dout("__ceph_caps_issued_mask %p combo issued %s"
779 " (mask %s)\n", &ci->vfs_inode,
780 ceph_cap_string(cap->issued),
781 ceph_cap_string(mask));
785 /* touch this + preceding caps */
787 for (q = rb_first(&ci->i_caps); q != p;
789 cap = rb_entry(q, struct ceph_cap,
791 if (!__cap_is_valid(cap))
804 * Return true if mask caps are currently being revoked by an MDS.
806 int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
807 struct ceph_cap *ocap, int mask)
809 struct ceph_cap *cap;
812 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
813 cap = rb_entry(p, struct ceph_cap, ci_node);
815 (cap->implemented & ~cap->issued & mask))
821 int ceph_caps_revoking(struct ceph_inode_info *ci, int mask)
823 struct inode *inode = &ci->vfs_inode;
826 spin_lock(&ci->i_ceph_lock);
827 ret = __ceph_caps_revoking_other(ci, NULL, mask);
828 spin_unlock(&ci->i_ceph_lock);
829 dout("ceph_caps_revoking %p %s = %d\n", inode,
830 ceph_cap_string(mask), ret);
834 int __ceph_caps_used(struct ceph_inode_info *ci)
838 used |= CEPH_CAP_PIN;
840 used |= CEPH_CAP_FILE_RD;
841 if (ci->i_rdcache_ref ||
842 (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */
843 ci->vfs_inode.i_data.nrpages))
844 used |= CEPH_CAP_FILE_CACHE;
846 used |= CEPH_CAP_FILE_WR;
847 if (ci->i_wb_ref || ci->i_wrbuffer_ref)
848 used |= CEPH_CAP_FILE_BUFFER;
853 * wanted, by virtue of open file modes
855 int __ceph_caps_file_wanted(struct ceph_inode_info *ci)
858 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
859 if (ci->i_nr_by_mode[i])
864 return ceph_caps_for_mode(bits >> 1);
868 * Return caps we have registered with the MDS(s) as 'wanted'.
870 int __ceph_caps_mds_wanted(struct ceph_inode_info *ci)
872 struct ceph_cap *cap;
876 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
877 cap = rb_entry(p, struct ceph_cap, ci_node);
878 if (!__cap_is_valid(cap))
880 if (cap == ci->i_auth_cap)
881 mds_wanted |= cap->mds_wanted;
883 mds_wanted |= (cap->mds_wanted & ~CEPH_CAP_ANY_FILE_WR);
889 * called under i_ceph_lock
891 static int __ceph_is_any_caps(struct ceph_inode_info *ci)
893 return !RB_EMPTY_ROOT(&ci->i_caps);
896 int ceph_is_any_caps(struct inode *inode)
898 struct ceph_inode_info *ci = ceph_inode(inode);
901 spin_lock(&ci->i_ceph_lock);
902 ret = __ceph_is_any_caps(ci);
903 spin_unlock(&ci->i_ceph_lock);
908 static void drop_inode_snap_realm(struct ceph_inode_info *ci)
910 struct ceph_snap_realm *realm = ci->i_snap_realm;
911 spin_lock(&realm->inodes_with_caps_lock);
912 list_del_init(&ci->i_snap_realm_item);
913 ci->i_snap_realm_counter++;
914 ci->i_snap_realm = NULL;
915 spin_unlock(&realm->inodes_with_caps_lock);
916 ceph_put_snap_realm(ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc,
921 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
923 * caller should hold i_ceph_lock.
924 * caller will not hold session s_mutex if called from destroy_inode.
926 void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
928 struct ceph_mds_session *session = cap->session;
929 struct ceph_inode_info *ci = cap->ci;
930 struct ceph_mds_client *mdsc =
931 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
934 dout("__ceph_remove_cap %p from %p\n", cap, &ci->vfs_inode);
936 /* remove from session list */
937 spin_lock(&session->s_cap_lock);
938 if (session->s_cap_iterator == cap) {
939 /* not yet, we are iterating over this very cap */
940 dout("__ceph_remove_cap delaying %p removal from session %p\n",
943 list_del_init(&cap->session_caps);
944 session->s_nr_caps--;
948 /* protect backpointer with s_cap_lock: see iterate_session_caps */
952 * s_cap_reconnect is protected by s_cap_lock. no one changes
953 * s_cap_gen while session is in the reconnect state.
956 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
957 cap->queue_release = 1;
959 list_add_tail(&cap->session_caps,
960 &session->s_cap_releases);
961 session->s_num_cap_releases++;
965 cap->queue_release = 0;
967 cap->cap_ino = ci->i_vino.ino;
969 spin_unlock(&session->s_cap_lock);
971 /* remove from inode list */
972 rb_erase(&cap->ci_node, &ci->i_caps);
973 if (ci->i_auth_cap == cap)
974 ci->i_auth_cap = NULL;
977 ceph_put_cap(mdsc, cap);
979 /* when reconnect denied, we remove session caps forcibly,
980 * i_wr_ref can be non-zero. If there are ongoing write,
983 if (!__ceph_is_any_caps(ci) && ci->i_wr_ref == 0 && ci->i_snap_realm)
984 drop_inode_snap_realm(ci);
986 if (!__ceph_is_any_real_caps(ci))
987 __cap_delay_cancel(mdsc, ci);
990 struct cap_msg_args {
991 struct ceph_mds_session *session;
992 u64 ino, cid, follows;
993 u64 flush_tid, oldest_flush_tid, size, max_size;
995 struct ceph_buffer *xattr_buf;
996 struct timespec atime, mtime, ctime;
997 int op, caps, wanted, dirty;
998 u32 seq, issue_seq, mseq, time_warp_seq;
1006 * Build and send a cap message to the given MDS.
1008 * Caller should be holding s_mutex.
1010 static int send_cap_msg(struct cap_msg_args *arg)
1012 struct ceph_mds_caps *fc;
1013 struct ceph_msg *msg;
1016 struct timespec zerotime = {0};
1018 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1019 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1020 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(arg->op),
1021 arg->cid, arg->ino, ceph_cap_string(arg->caps),
1022 ceph_cap_string(arg->wanted), ceph_cap_string(arg->dirty),
1023 arg->seq, arg->issue_seq, arg->flush_tid, arg->oldest_flush_tid,
1024 arg->mseq, arg->follows, arg->size, arg->max_size,
1026 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1028 /* flock buffer size + inline version + inline data size +
1029 * osd_epoch_barrier + oldest_flush_tid */
1030 extra_len = 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4;
1031 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1036 msg->hdr.version = cpu_to_le16(10);
1037 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1039 fc = msg->front.iov_base;
1040 memset(fc, 0, sizeof(*fc));
1042 fc->cap_id = cpu_to_le64(arg->cid);
1043 fc->op = cpu_to_le32(arg->op);
1044 fc->seq = cpu_to_le32(arg->seq);
1045 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1046 fc->migrate_seq = cpu_to_le32(arg->mseq);
1047 fc->caps = cpu_to_le32(arg->caps);
1048 fc->wanted = cpu_to_le32(arg->wanted);
1049 fc->dirty = cpu_to_le32(arg->dirty);
1050 fc->ino = cpu_to_le64(arg->ino);
1051 fc->snap_follows = cpu_to_le64(arg->follows);
1053 fc->size = cpu_to_le64(arg->size);
1054 fc->max_size = cpu_to_le64(arg->max_size);
1055 ceph_encode_timespec(&fc->mtime, &arg->mtime);
1056 ceph_encode_timespec(&fc->atime, &arg->atime);
1057 ceph_encode_timespec(&fc->ctime, &arg->ctime);
1058 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1060 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1061 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1062 fc->mode = cpu_to_le32(arg->mode);
1064 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1065 if (arg->xattr_buf) {
1066 msg->middle = ceph_buffer_get(arg->xattr_buf);
1067 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1068 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1072 /* flock buffer size (version 2) */
1073 ceph_encode_32(&p, 0);
1074 /* inline version (version 4) */
1075 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1076 /* inline data size */
1077 ceph_encode_32(&p, 0);
1078 /* osd_epoch_barrier (version 5) */
1079 ceph_encode_32(&p, 0);
1080 /* oldest_flush_tid (version 6) */
1081 ceph_encode_64(&p, arg->oldest_flush_tid);
1084 * caller_uid/caller_gid (version 7)
1086 * Currently, we don't properly track which caller dirtied the caps
1087 * last, and force a flush of them when there is a conflict. For now,
1088 * just set this to 0:0, to emulate how the MDS has worked up to now.
1090 ceph_encode_32(&p, 0);
1091 ceph_encode_32(&p, 0);
1093 /* pool namespace (version 8) (mds always ignores this) */
1094 ceph_encode_32(&p, 0);
1097 * btime and change_attr (version 9)
1099 * We just zero these out for now, as the MDS ignores them unless
1100 * the requisite feature flags are set (which we don't do yet).
1102 ceph_encode_timespec(p, &zerotime);
1103 p += sizeof(struct ceph_timespec);
1104 ceph_encode_64(&p, 0);
1106 /* Advisory flags (version 10) */
1107 ceph_encode_32(&p, 0);
1109 ceph_con_send(&arg->session->s_con, msg);
1114 * Queue cap releases when an inode is dropped from our cache. Since
1115 * inode is about to be destroyed, there is no need for i_ceph_lock.
1117 void ceph_queue_caps_release(struct inode *inode)
1119 struct ceph_inode_info *ci = ceph_inode(inode);
1122 p = rb_first(&ci->i_caps);
1124 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1126 __ceph_remove_cap(cap, true);
1131 * Send a cap msg on the given inode. Update our caps state, then
1132 * drop i_ceph_lock and send the message.
1134 * Make note of max_size reported/requested from mds, revoked caps
1135 * that have now been implemented.
1137 * Make half-hearted attempt ot to invalidate page cache if we are
1138 * dropping RDCACHE. Note that this will leave behind locked pages
1139 * that we'll then need to deal with elsewhere.
1141 * Return non-zero if delayed release, or we experienced an error
1142 * such that the caller should requeue + retry later.
1144 * called with i_ceph_lock, then drops it.
1145 * caller should hold snap_rwsem (read), s_mutex.
1147 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1148 int op, int used, int want, int retain, int flushing,
1149 u64 flush_tid, u64 oldest_flush_tid)
1150 __releases(cap->ci->i_ceph_lock)
1152 struct ceph_inode_info *ci = cap->ci;
1153 struct inode *inode = &ci->vfs_inode;
1154 struct cap_msg_args arg;
1155 int held, revoking, dropping;
1160 held = cap->issued | cap->implemented;
1161 revoking = cap->implemented & ~cap->issued;
1162 retain &= ~revoking;
1163 dropping = cap->issued & ~retain;
1165 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1166 inode, cap, cap->session,
1167 ceph_cap_string(held), ceph_cap_string(held & retain),
1168 ceph_cap_string(revoking));
1169 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1171 arg.session = cap->session;
1173 /* don't release wanted unless we've waited a bit. */
1174 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1175 time_before(jiffies, ci->i_hold_caps_min)) {
1176 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1177 ceph_cap_string(cap->issued),
1178 ceph_cap_string(cap->issued & retain),
1179 ceph_cap_string(cap->mds_wanted),
1180 ceph_cap_string(want));
1181 want |= cap->mds_wanted;
1182 retain |= cap->issued;
1185 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1187 cap->issued &= retain; /* drop bits we don't want */
1188 if (cap->implemented & ~cap->issued) {
1190 * Wake up any waiters on wanted -> needed transition.
1191 * This is due to the weird transition from buffered
1192 * to sync IO... we need to flush dirty pages _before_
1193 * allowing sync writes to avoid reordering.
1197 cap->implemented &= cap->issued | used;
1198 cap->mds_wanted = want;
1200 arg.ino = ceph_vino(inode).ino;
1201 arg.cid = cap->cap_id;
1202 arg.follows = flushing ? ci->i_head_snapc->seq : 0;
1203 arg.flush_tid = flush_tid;
1204 arg.oldest_flush_tid = oldest_flush_tid;
1206 arg.size = inode->i_size;
1207 ci->i_reported_size = arg.size;
1208 arg.max_size = ci->i_wanted_max_size;
1209 ci->i_requested_max_size = arg.max_size;
1211 if (flushing & CEPH_CAP_XATTR_EXCL) {
1212 __ceph_build_xattrs_blob(ci);
1213 arg.xattr_version = ci->i_xattrs.version;
1214 arg.xattr_buf = ci->i_xattrs.blob;
1216 arg.xattr_buf = NULL;
1219 arg.mtime = inode->i_mtime;
1220 arg.atime = inode->i_atime;
1221 arg.ctime = inode->i_ctime;
1224 arg.caps = cap->implemented;
1226 arg.dirty = flushing;
1229 arg.issue_seq = cap->issue_seq;
1230 arg.mseq = cap->mseq;
1231 arg.time_warp_seq = ci->i_time_warp_seq;
1233 arg.uid = inode->i_uid;
1234 arg.gid = inode->i_gid;
1235 arg.mode = inode->i_mode;
1237 arg.inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1239 spin_unlock(&ci->i_ceph_lock);
1241 ret = send_cap_msg(&arg);
1243 dout("error sending cap msg, must requeue %p\n", inode);
1248 wake_up_all(&ci->i_cap_wq);
1253 static inline int __send_flush_snap(struct inode *inode,
1254 struct ceph_mds_session *session,
1255 struct ceph_cap_snap *capsnap,
1256 u32 mseq, u64 oldest_flush_tid)
1258 struct cap_msg_args arg;
1260 arg.session = session;
1261 arg.ino = ceph_vino(inode).ino;
1263 arg.follows = capsnap->follows;
1264 arg.flush_tid = capsnap->cap_flush.tid;
1265 arg.oldest_flush_tid = oldest_flush_tid;
1267 arg.size = capsnap->size;
1269 arg.xattr_version = capsnap->xattr_version;
1270 arg.xattr_buf = capsnap->xattr_blob;
1272 arg.atime = capsnap->atime;
1273 arg.mtime = capsnap->mtime;
1274 arg.ctime = capsnap->ctime;
1276 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1277 arg.caps = capsnap->issued;
1279 arg.dirty = capsnap->dirty;
1284 arg.time_warp_seq = capsnap->time_warp_seq;
1286 arg.uid = capsnap->uid;
1287 arg.gid = capsnap->gid;
1288 arg.mode = capsnap->mode;
1290 arg.inline_data = capsnap->inline_data;
1292 return send_cap_msg(&arg);
1296 * When a snapshot is taken, clients accumulate dirty metadata on
1297 * inodes with capabilities in ceph_cap_snaps to describe the file
1298 * state at the time the snapshot was taken. This must be flushed
1299 * asynchronously back to the MDS once sync writes complete and dirty
1300 * data is written out.
1302 * Called under i_ceph_lock. Takes s_mutex as needed.
1304 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1305 struct ceph_mds_session *session)
1306 __releases(ci->i_ceph_lock)
1307 __acquires(ci->i_ceph_lock)
1309 struct inode *inode = &ci->vfs_inode;
1310 struct ceph_mds_client *mdsc = session->s_mdsc;
1311 struct ceph_cap_snap *capsnap;
1312 u64 oldest_flush_tid = 0;
1313 u64 first_tid = 1, last_tid = 0;
1315 dout("__flush_snaps %p session %p\n", inode, session);
1317 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1319 * we need to wait for sync writes to complete and for dirty
1320 * pages to be written out.
1322 if (capsnap->dirty_pages || capsnap->writing)
1325 /* should be removed by ceph_try_drop_cap_snap() */
1326 BUG_ON(!capsnap->need_flush);
1328 /* only flush each capsnap once */
1329 if (capsnap->cap_flush.tid > 0) {
1330 dout(" already flushed %p, skipping\n", capsnap);
1334 spin_lock(&mdsc->cap_dirty_lock);
1335 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1336 list_add_tail(&capsnap->cap_flush.g_list,
1337 &mdsc->cap_flush_list);
1338 if (oldest_flush_tid == 0)
1339 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1340 if (list_empty(&ci->i_flushing_item)) {
1341 list_add_tail(&ci->i_flushing_item,
1342 &session->s_cap_flushing);
1344 spin_unlock(&mdsc->cap_dirty_lock);
1346 list_add_tail(&capsnap->cap_flush.i_list,
1347 &ci->i_cap_flush_list);
1350 first_tid = capsnap->cap_flush.tid;
1351 last_tid = capsnap->cap_flush.tid;
1354 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1356 while (first_tid <= last_tid) {
1357 struct ceph_cap *cap = ci->i_auth_cap;
1358 struct ceph_cap_flush *cf;
1361 if (!(cap && cap->session == session)) {
1362 dout("__flush_snaps %p auth cap %p not mds%d, "
1363 "stop\n", inode, cap, session->s_mds);
1368 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1369 if (cf->tid >= first_tid) {
1377 first_tid = cf->tid + 1;
1379 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1380 atomic_inc(&capsnap->nref);
1381 spin_unlock(&ci->i_ceph_lock);
1383 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1384 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1386 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1389 pr_err("__flush_snaps: error sending cap flushsnap, "
1390 "ino (%llx.%llx) tid %llu follows %llu\n",
1391 ceph_vinop(inode), cf->tid, capsnap->follows);
1394 ceph_put_cap_snap(capsnap);
1395 spin_lock(&ci->i_ceph_lock);
1399 void ceph_flush_snaps(struct ceph_inode_info *ci,
1400 struct ceph_mds_session **psession)
1402 struct inode *inode = &ci->vfs_inode;
1403 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1404 struct ceph_mds_session *session = NULL;
1407 dout("ceph_flush_snaps %p\n", inode);
1409 session = *psession;
1411 spin_lock(&ci->i_ceph_lock);
1412 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1413 dout(" no capsnap needs flush, doing nothing\n");
1416 if (!ci->i_auth_cap) {
1417 dout(" no auth cap (migrating?), doing nothing\n");
1421 mds = ci->i_auth_cap->session->s_mds;
1422 if (session && session->s_mds != mds) {
1423 dout(" oops, wrong session %p mutex\n", session);
1424 mutex_unlock(&session->s_mutex);
1425 ceph_put_mds_session(session);
1429 spin_unlock(&ci->i_ceph_lock);
1430 mutex_lock(&mdsc->mutex);
1431 session = __ceph_lookup_mds_session(mdsc, mds);
1432 mutex_unlock(&mdsc->mutex);
1434 dout(" inverting session/ino locks on %p\n", session);
1435 mutex_lock(&session->s_mutex);
1440 __ceph_flush_snaps(ci, session);
1442 spin_unlock(&ci->i_ceph_lock);
1445 *psession = session;
1447 mutex_unlock(&session->s_mutex);
1448 ceph_put_mds_session(session);
1450 /* we flushed them all; remove this inode from the queue */
1451 spin_lock(&mdsc->snap_flush_lock);
1452 list_del_init(&ci->i_snap_flush_item);
1453 spin_unlock(&mdsc->snap_flush_lock);
1457 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1458 * Caller is then responsible for calling __mark_inode_dirty with the
1459 * returned flags value.
1461 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1462 struct ceph_cap_flush **pcf)
1464 struct ceph_mds_client *mdsc =
1465 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1466 struct inode *inode = &ci->vfs_inode;
1467 int was = ci->i_dirty_caps;
1470 if (!ci->i_auth_cap) {
1471 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1472 "but no auth cap (session was closed?)\n",
1473 inode, ceph_ino(inode), ceph_cap_string(mask));
1477 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1478 ceph_cap_string(mask), ceph_cap_string(was),
1479 ceph_cap_string(was | mask));
1480 ci->i_dirty_caps |= mask;
1482 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1483 swap(ci->i_prealloc_cap_flush, *pcf);
1485 if (!ci->i_head_snapc) {
1486 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1487 ci->i_head_snapc = ceph_get_snap_context(
1488 ci->i_snap_realm->cached_context);
1490 dout(" inode %p now dirty snapc %p auth cap %p\n",
1491 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1492 BUG_ON(!list_empty(&ci->i_dirty_item));
1493 spin_lock(&mdsc->cap_dirty_lock);
1494 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1495 spin_unlock(&mdsc->cap_dirty_lock);
1496 if (ci->i_flushing_caps == 0) {
1498 dirty |= I_DIRTY_SYNC;
1501 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1503 BUG_ON(list_empty(&ci->i_dirty_item));
1504 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1505 (mask & CEPH_CAP_FILE_BUFFER))
1506 dirty |= I_DIRTY_DATASYNC;
1507 __cap_delay_requeue(mdsc, ci);
1511 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1513 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1516 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1519 kmem_cache_free(ceph_cap_flush_cachep, cf);
1522 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1524 if (!list_empty(&mdsc->cap_flush_list)) {
1525 struct ceph_cap_flush *cf =
1526 list_first_entry(&mdsc->cap_flush_list,
1527 struct ceph_cap_flush, g_list);
1534 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1535 * Return true if caller needs to wake up flush waiters.
1537 static bool __finish_cap_flush(struct ceph_mds_client *mdsc,
1538 struct ceph_inode_info *ci,
1539 struct ceph_cap_flush *cf)
1541 struct ceph_cap_flush *prev;
1542 bool wake = cf->wake;
1544 /* are there older pending cap flushes? */
1545 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1546 prev = list_prev_entry(cf, g_list);
1550 list_del(&cf->g_list);
1552 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1553 prev = list_prev_entry(cf, i_list);
1557 list_del(&cf->i_list);
1565 * Add dirty inode to the flushing list. Assigned a seq number so we
1566 * can wait for caps to flush without starving.
1568 * Called under i_ceph_lock.
1570 static int __mark_caps_flushing(struct inode *inode,
1571 struct ceph_mds_session *session, bool wake,
1572 u64 *flush_tid, u64 *oldest_flush_tid)
1574 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1575 struct ceph_inode_info *ci = ceph_inode(inode);
1576 struct ceph_cap_flush *cf = NULL;
1579 BUG_ON(ci->i_dirty_caps == 0);
1580 BUG_ON(list_empty(&ci->i_dirty_item));
1581 BUG_ON(!ci->i_prealloc_cap_flush);
1583 flushing = ci->i_dirty_caps;
1584 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1585 ceph_cap_string(flushing),
1586 ceph_cap_string(ci->i_flushing_caps),
1587 ceph_cap_string(ci->i_flushing_caps | flushing));
1588 ci->i_flushing_caps |= flushing;
1589 ci->i_dirty_caps = 0;
1590 dout(" inode %p now !dirty\n", inode);
1592 swap(cf, ci->i_prealloc_cap_flush);
1593 cf->caps = flushing;
1596 spin_lock(&mdsc->cap_dirty_lock);
1597 list_del_init(&ci->i_dirty_item);
1599 cf->tid = ++mdsc->last_cap_flush_tid;
1600 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1601 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1603 if (list_empty(&ci->i_flushing_item)) {
1604 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1605 mdsc->num_cap_flushing++;
1607 spin_unlock(&mdsc->cap_dirty_lock);
1609 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1611 *flush_tid = cf->tid;
1616 * try to invalidate mapping pages without blocking.
1618 static int try_nonblocking_invalidate(struct inode *inode)
1620 struct ceph_inode_info *ci = ceph_inode(inode);
1621 u32 invalidating_gen = ci->i_rdcache_gen;
1623 spin_unlock(&ci->i_ceph_lock);
1624 invalidate_mapping_pages(&inode->i_data, 0, -1);
1625 spin_lock(&ci->i_ceph_lock);
1627 if (inode->i_data.nrpages == 0 &&
1628 invalidating_gen == ci->i_rdcache_gen) {
1630 dout("try_nonblocking_invalidate %p success\n", inode);
1631 /* save any racing async invalidate some trouble */
1632 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1635 dout("try_nonblocking_invalidate %p failed\n", inode);
1640 * Swiss army knife function to examine currently used and wanted
1641 * versus held caps. Release, flush, ack revoked caps to mds as
1644 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1645 * cap release further.
1646 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1647 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1650 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1651 struct ceph_mds_session *session)
1653 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1654 struct ceph_mds_client *mdsc = fsc->mdsc;
1655 struct inode *inode = &ci->vfs_inode;
1656 struct ceph_cap *cap;
1657 u64 flush_tid, oldest_flush_tid;
1658 int file_wanted, used, cap_used;
1659 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1660 int issued, implemented, want, retain, revoking, flushing = 0;
1661 int mds = -1; /* keep track of how far we've gone through i_caps list
1662 to avoid an infinite loop on retry */
1664 int delayed = 0, sent = 0, num;
1665 bool is_delayed = flags & CHECK_CAPS_NODELAY;
1666 bool queue_invalidate = false;
1667 bool force_requeue = false;
1668 bool tried_invalidate = false;
1670 /* if we are unmounting, flush any unused caps immediately. */
1674 spin_lock(&ci->i_ceph_lock);
1676 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1677 flags |= CHECK_CAPS_FLUSH;
1681 spin_lock(&ci->i_ceph_lock);
1683 file_wanted = __ceph_caps_file_wanted(ci);
1684 used = __ceph_caps_used(ci);
1685 issued = __ceph_caps_issued(ci, &implemented);
1686 revoking = implemented & ~issued;
1689 retain = file_wanted | used | CEPH_CAP_PIN;
1690 if (!mdsc->stopping && inode->i_nlink > 0) {
1692 retain |= CEPH_CAP_ANY; /* be greedy */
1693 } else if (S_ISDIR(inode->i_mode) &&
1694 (issued & CEPH_CAP_FILE_SHARED) &&
1695 __ceph_dir_is_complete(ci)) {
1697 * If a directory is complete, we want to keep
1698 * the exclusive cap. So that MDS does not end up
1699 * revoking the shared cap on every create/unlink
1702 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1706 retain |= CEPH_CAP_ANY_SHARED;
1708 * keep RD only if we didn't have the file open RW,
1709 * because then the mds would revoke it anyway to
1710 * journal max_size=0.
1712 if (ci->i_max_size == 0)
1713 retain |= CEPH_CAP_ANY_RD;
1717 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1718 " issued %s revoking %s retain %s %s%s%s\n", inode,
1719 ceph_cap_string(file_wanted),
1720 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1721 ceph_cap_string(ci->i_flushing_caps),
1722 ceph_cap_string(issued), ceph_cap_string(revoking),
1723 ceph_cap_string(retain),
1724 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1725 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1726 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1729 * If we no longer need to hold onto old our caps, and we may
1730 * have cached pages, but don't want them, then try to invalidate.
1731 * If we fail, it's because pages are locked.... try again later.
1733 if ((!is_delayed || mdsc->stopping) &&
1734 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */
1735 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1736 inode->i_data.nrpages && /* have cached pages */
1737 (revoking & (CEPH_CAP_FILE_CACHE|
1738 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1739 !tried_invalidate) {
1740 dout("check_caps trying to invalidate on %p\n", inode);
1741 if (try_nonblocking_invalidate(inode) < 0) {
1742 if (revoking & (CEPH_CAP_FILE_CACHE|
1743 CEPH_CAP_FILE_LAZYIO)) {
1744 dout("check_caps queuing invalidate\n");
1745 queue_invalidate = true;
1746 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1748 dout("check_caps failed to invalidate pages\n");
1749 /* we failed to invalidate pages. check these
1750 caps again later. */
1751 force_requeue = true;
1752 __cap_set_timeouts(mdsc, ci);
1755 tried_invalidate = true;
1760 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1761 cap = rb_entry(p, struct ceph_cap, ci_node);
1764 /* avoid looping forever */
1765 if (mds >= cap->mds ||
1766 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1769 /* NOTE: no side-effects allowed, until we take s_mutex */
1772 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1773 cap_used &= ~ci->i_auth_cap->issued;
1775 revoking = cap->implemented & ~cap->issued;
1776 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1777 cap->mds, cap, ceph_cap_string(cap_used),
1778 ceph_cap_string(cap->issued),
1779 ceph_cap_string(cap->implemented),
1780 ceph_cap_string(revoking));
1782 if (cap == ci->i_auth_cap &&
1783 (cap->issued & CEPH_CAP_FILE_WR)) {
1784 /* request larger max_size from MDS? */
1785 if (ci->i_wanted_max_size > ci->i_max_size &&
1786 ci->i_wanted_max_size > ci->i_requested_max_size) {
1787 dout("requesting new max_size\n");
1791 /* approaching file_max? */
1792 if ((inode->i_size << 1) >= ci->i_max_size &&
1793 (ci->i_reported_size << 1) < ci->i_max_size) {
1794 dout("i_size approaching max_size\n");
1798 /* flush anything dirty? */
1799 if (cap == ci->i_auth_cap) {
1800 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
1801 dout("flushing dirty caps\n");
1804 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
1805 dout("flushing snap caps\n");
1810 /* completed revocation? going down and there are no caps? */
1811 if (revoking && (revoking & cap_used) == 0) {
1812 dout("completed revocation of %s\n",
1813 ceph_cap_string(cap->implemented & ~cap->issued));
1817 /* want more caps from mds? */
1818 if (want & ~(cap->mds_wanted | cap->issued))
1821 /* things we might delay */
1822 if ((cap->issued & ~retain) == 0 &&
1823 cap->mds_wanted == want)
1824 continue; /* nope, all good */
1830 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1831 time_before(jiffies, ci->i_hold_caps_max)) {
1832 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1833 ceph_cap_string(cap->issued),
1834 ceph_cap_string(cap->issued & retain),
1835 ceph_cap_string(cap->mds_wanted),
1836 ceph_cap_string(want));
1842 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1843 dout(" skipping %p I_NOFLUSH set\n", inode);
1847 if (session && session != cap->session) {
1848 dout("oops, wrong session %p mutex\n", session);
1849 mutex_unlock(&session->s_mutex);
1853 session = cap->session;
1854 if (mutex_trylock(&session->s_mutex) == 0) {
1855 dout("inverting session/ino locks on %p\n",
1857 spin_unlock(&ci->i_ceph_lock);
1858 if (took_snap_rwsem) {
1859 up_read(&mdsc->snap_rwsem);
1860 took_snap_rwsem = 0;
1862 mutex_lock(&session->s_mutex);
1867 /* kick flushing and flush snaps before sending normal
1869 if (cap == ci->i_auth_cap &&
1871 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
1872 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
1873 spin_lock(&mdsc->cap_dirty_lock);
1874 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1875 spin_unlock(&mdsc->cap_dirty_lock);
1876 __kick_flushing_caps(mdsc, session, ci,
1878 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
1880 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
1881 __ceph_flush_snaps(ci, session);
1886 /* take snap_rwsem after session mutex */
1887 if (!took_snap_rwsem) {
1888 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1889 dout("inverting snap/in locks on %p\n",
1891 spin_unlock(&ci->i_ceph_lock);
1892 down_read(&mdsc->snap_rwsem);
1893 took_snap_rwsem = 1;
1896 took_snap_rwsem = 1;
1899 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
1900 flushing = __mark_caps_flushing(inode, session, false,
1906 spin_lock(&mdsc->cap_dirty_lock);
1907 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1908 spin_unlock(&mdsc->cap_dirty_lock);
1911 mds = cap->mds; /* remember mds, so we don't repeat */
1914 /* __send_cap drops i_ceph_lock */
1915 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used,
1916 want, retain, flushing,
1917 flush_tid, oldest_flush_tid);
1918 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1922 * Reschedule delayed caps release if we delayed anything,
1925 if (delayed && is_delayed)
1926 force_requeue = true; /* __send_cap delayed release; requeue */
1927 if (!delayed && !is_delayed)
1928 __cap_delay_cancel(mdsc, ci);
1929 else if (!is_delayed || force_requeue)
1930 __cap_delay_requeue(mdsc, ci);
1932 spin_unlock(&ci->i_ceph_lock);
1934 if (queue_invalidate)
1935 ceph_queue_invalidate(inode);
1938 mutex_unlock(&session->s_mutex);
1939 if (took_snap_rwsem)
1940 up_read(&mdsc->snap_rwsem);
1944 * Try to flush dirty caps back to the auth mds.
1946 static int try_flush_caps(struct inode *inode, u64 *ptid)
1948 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1949 struct ceph_inode_info *ci = ceph_inode(inode);
1950 struct ceph_mds_session *session = NULL;
1952 u64 flush_tid = 0, oldest_flush_tid = 0;
1955 spin_lock(&ci->i_ceph_lock);
1956 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1957 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1960 if (ci->i_dirty_caps && ci->i_auth_cap) {
1961 struct ceph_cap *cap = ci->i_auth_cap;
1962 int used = __ceph_caps_used(ci);
1963 int want = __ceph_caps_wanted(ci);
1966 if (!session || session != cap->session) {
1967 spin_unlock(&ci->i_ceph_lock);
1969 mutex_unlock(&session->s_mutex);
1970 session = cap->session;
1971 mutex_lock(&session->s_mutex);
1974 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1977 flushing = __mark_caps_flushing(inode, session, true,
1978 &flush_tid, &oldest_flush_tid);
1980 /* __send_cap drops i_ceph_lock */
1981 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1982 (cap->issued | cap->implemented),
1983 flushing, flush_tid, oldest_flush_tid);
1986 spin_lock(&ci->i_ceph_lock);
1987 __cap_delay_requeue(mdsc, ci);
1988 spin_unlock(&ci->i_ceph_lock);
1991 if (!list_empty(&ci->i_cap_flush_list)) {
1992 struct ceph_cap_flush *cf =
1993 list_last_entry(&ci->i_cap_flush_list,
1994 struct ceph_cap_flush, i_list);
1996 flush_tid = cf->tid;
1998 flushing = ci->i_flushing_caps;
1999 spin_unlock(&ci->i_ceph_lock);
2003 mutex_unlock(&session->s_mutex);
2010 * Return true if we've flushed caps through the given flush_tid.
2012 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2014 struct ceph_inode_info *ci = ceph_inode(inode);
2017 spin_lock(&ci->i_ceph_lock);
2018 if (!list_empty(&ci->i_cap_flush_list)) {
2019 struct ceph_cap_flush * cf =
2020 list_first_entry(&ci->i_cap_flush_list,
2021 struct ceph_cap_flush, i_list);
2022 if (cf->tid <= flush_tid)
2025 spin_unlock(&ci->i_ceph_lock);
2030 * wait for any unsafe requests to complete.
2032 static int unsafe_request_wait(struct inode *inode)
2034 struct ceph_inode_info *ci = ceph_inode(inode);
2035 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2038 spin_lock(&ci->i_unsafe_lock);
2039 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2040 req1 = list_last_entry(&ci->i_unsafe_dirops,
2041 struct ceph_mds_request,
2043 ceph_mdsc_get_request(req1);
2045 if (!list_empty(&ci->i_unsafe_iops)) {
2046 req2 = list_last_entry(&ci->i_unsafe_iops,
2047 struct ceph_mds_request,
2048 r_unsafe_target_item);
2049 ceph_mdsc_get_request(req2);
2051 spin_unlock(&ci->i_unsafe_lock);
2053 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2054 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2056 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2057 ceph_timeout_jiffies(req1->r_timeout));
2060 ceph_mdsc_put_request(req1);
2063 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2064 ceph_timeout_jiffies(req2->r_timeout));
2067 ceph_mdsc_put_request(req2);
2072 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2074 struct inode *inode = file->f_mapping->host;
2075 struct ceph_inode_info *ci = ceph_inode(inode);
2080 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2082 ceph_sync_write_wait(inode);
2084 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
2093 dirty = try_flush_caps(inode, &flush_tid);
2094 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2096 ret = unsafe_request_wait(inode);
2099 * only wait on non-file metadata writeback (the mds
2100 * can recover size and mtime, so we don't need to
2103 if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2104 ret = wait_event_interruptible(ci->i_cap_wq,
2105 caps_are_flushed(inode, flush_tid));
2107 inode_unlock(inode);
2109 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2114 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2115 * queue inode for flush but don't do so immediately, because we can
2116 * get by with fewer MDS messages if we wait for data writeback to
2119 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2121 struct ceph_inode_info *ci = ceph_inode(inode);
2125 int wait = wbc->sync_mode == WB_SYNC_ALL;
2127 dout("write_inode %p wait=%d\n", inode, wait);
2129 dirty = try_flush_caps(inode, &flush_tid);
2131 err = wait_event_interruptible(ci->i_cap_wq,
2132 caps_are_flushed(inode, flush_tid));
2134 struct ceph_mds_client *mdsc =
2135 ceph_sb_to_client(inode->i_sb)->mdsc;
2137 spin_lock(&ci->i_ceph_lock);
2138 if (__ceph_caps_dirty(ci))
2139 __cap_delay_requeue_front(mdsc, ci);
2140 spin_unlock(&ci->i_ceph_lock);
2145 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2146 struct ceph_mds_session *session,
2147 struct ceph_inode_info *ci,
2148 u64 oldest_flush_tid)
2149 __releases(ci->i_ceph_lock)
2150 __acquires(ci->i_ceph_lock)
2152 struct inode *inode = &ci->vfs_inode;
2153 struct ceph_cap *cap;
2154 struct ceph_cap_flush *cf;
2158 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2159 if (cf->tid < first_tid)
2162 cap = ci->i_auth_cap;
2163 if (!(cap && cap->session == session)) {
2164 pr_err("%p auth cap %p not mds%d ???\n",
2165 inode, cap, session->s_mds);
2169 first_tid = cf->tid + 1;
2172 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2173 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2174 ci->i_ceph_flags |= CEPH_I_NODELAY;
2175 ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2176 __ceph_caps_used(ci),
2177 __ceph_caps_wanted(ci),
2178 cap->issued | cap->implemented,
2179 cf->caps, cf->tid, oldest_flush_tid);
2181 pr_err("kick_flushing_caps: error sending "
2182 "cap flush, ino (%llx.%llx) "
2183 "tid %llu flushing %s\n",
2184 ceph_vinop(inode), cf->tid,
2185 ceph_cap_string(cf->caps));
2188 struct ceph_cap_snap *capsnap =
2189 container_of(cf, struct ceph_cap_snap,
2191 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2192 inode, capsnap, cf->tid,
2193 ceph_cap_string(capsnap->dirty));
2195 atomic_inc(&capsnap->nref);
2196 spin_unlock(&ci->i_ceph_lock);
2198 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2201 pr_err("kick_flushing_caps: error sending "
2202 "cap flushsnap, ino (%llx.%llx) "
2203 "tid %llu follows %llu\n",
2204 ceph_vinop(inode), cf->tid,
2208 ceph_put_cap_snap(capsnap);
2211 spin_lock(&ci->i_ceph_lock);
2215 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2216 struct ceph_mds_session *session)
2218 struct ceph_inode_info *ci;
2219 struct ceph_cap *cap;
2220 u64 oldest_flush_tid;
2222 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2224 spin_lock(&mdsc->cap_dirty_lock);
2225 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2226 spin_unlock(&mdsc->cap_dirty_lock);
2228 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2229 spin_lock(&ci->i_ceph_lock);
2230 cap = ci->i_auth_cap;
2231 if (!(cap && cap->session == session)) {
2232 pr_err("%p auth cap %p not mds%d ???\n",
2233 &ci->vfs_inode, cap, session->s_mds);
2234 spin_unlock(&ci->i_ceph_lock);
2240 * if flushing caps were revoked, we re-send the cap flush
2241 * in client reconnect stage. This guarantees MDS * processes
2242 * the cap flush message before issuing the flushing caps to
2245 if ((cap->issued & ci->i_flushing_caps) !=
2246 ci->i_flushing_caps) {
2247 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2248 __kick_flushing_caps(mdsc, session, ci,
2251 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2254 spin_unlock(&ci->i_ceph_lock);
2258 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2259 struct ceph_mds_session *session)
2261 struct ceph_inode_info *ci;
2262 struct ceph_cap *cap;
2263 u64 oldest_flush_tid;
2265 dout("kick_flushing_caps mds%d\n", session->s_mds);
2267 spin_lock(&mdsc->cap_dirty_lock);
2268 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2269 spin_unlock(&mdsc->cap_dirty_lock);
2271 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2272 spin_lock(&ci->i_ceph_lock);
2273 cap = ci->i_auth_cap;
2274 if (!(cap && cap->session == session)) {
2275 pr_err("%p auth cap %p not mds%d ???\n",
2276 &ci->vfs_inode, cap, session->s_mds);
2277 spin_unlock(&ci->i_ceph_lock);
2280 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2281 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2282 __kick_flushing_caps(mdsc, session, ci,
2285 spin_unlock(&ci->i_ceph_lock);
2289 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2290 struct ceph_mds_session *session,
2291 struct inode *inode)
2292 __releases(ci->i_ceph_lock)
2294 struct ceph_inode_info *ci = ceph_inode(inode);
2295 struct ceph_cap *cap;
2297 cap = ci->i_auth_cap;
2298 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2299 ceph_cap_string(ci->i_flushing_caps));
2301 if (!list_empty(&ci->i_cap_flush_list)) {
2302 u64 oldest_flush_tid;
2303 spin_lock(&mdsc->cap_dirty_lock);
2304 list_move_tail(&ci->i_flushing_item,
2305 &cap->session->s_cap_flushing);
2306 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2307 spin_unlock(&mdsc->cap_dirty_lock);
2309 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2310 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2311 spin_unlock(&ci->i_ceph_lock);
2313 spin_unlock(&ci->i_ceph_lock);
2319 * Take references to capabilities we hold, so that we don't release
2320 * them to the MDS prematurely.
2322 * Protected by i_ceph_lock.
2324 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2325 bool snap_rwsem_locked)
2327 if (got & CEPH_CAP_PIN)
2329 if (got & CEPH_CAP_FILE_RD)
2331 if (got & CEPH_CAP_FILE_CACHE)
2332 ci->i_rdcache_ref++;
2333 if (got & CEPH_CAP_FILE_WR) {
2334 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2335 BUG_ON(!snap_rwsem_locked);
2336 ci->i_head_snapc = ceph_get_snap_context(
2337 ci->i_snap_realm->cached_context);
2341 if (got & CEPH_CAP_FILE_BUFFER) {
2342 if (ci->i_wb_ref == 0)
2343 ihold(&ci->vfs_inode);
2345 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2346 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2351 * Try to grab cap references. Specify those refs we @want, and the
2352 * minimal set we @need. Also include the larger offset we are writing
2353 * to (when applicable), and check against max_size here as well.
2354 * Note that caller is responsible for ensuring max_size increases are
2355 * requested from the MDS.
2357 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2358 loff_t endoff, bool nonblock, int *got, int *err)
2360 struct inode *inode = &ci->vfs_inode;
2361 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2363 int have, implemented;
2365 bool snap_rwsem_locked = false;
2367 dout("get_cap_refs %p need %s want %s\n", inode,
2368 ceph_cap_string(need), ceph_cap_string(want));
2371 spin_lock(&ci->i_ceph_lock);
2373 /* make sure file is actually open */
2374 file_wanted = __ceph_caps_file_wanted(ci);
2375 if ((file_wanted & need) != need) {
2376 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2377 ceph_cap_string(need), ceph_cap_string(file_wanted));
2383 /* finish pending truncate */
2384 while (ci->i_truncate_pending) {
2385 spin_unlock(&ci->i_ceph_lock);
2386 if (snap_rwsem_locked) {
2387 up_read(&mdsc->snap_rwsem);
2388 snap_rwsem_locked = false;
2390 __ceph_do_pending_vmtruncate(inode);
2391 spin_lock(&ci->i_ceph_lock);
2394 have = __ceph_caps_issued(ci, &implemented);
2396 if (have & need & CEPH_CAP_FILE_WR) {
2397 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2398 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2399 inode, endoff, ci->i_max_size);
2400 if (endoff > ci->i_requested_max_size) {
2407 * If a sync write is in progress, we must wait, so that we
2408 * can get a final snapshot value for size+mtime.
2410 if (__ceph_have_pending_cap_snap(ci)) {
2411 dout("get_cap_refs %p cap_snap_pending\n", inode);
2416 if ((have & need) == need) {
2418 * Look at (implemented & ~have & not) so that we keep waiting
2419 * on transition from wanted -> needed caps. This is needed
2420 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2421 * going before a prior buffered writeback happens.
2423 int not = want & ~(have & need);
2424 int revoking = implemented & ~have;
2425 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2426 inode, ceph_cap_string(have), ceph_cap_string(not),
2427 ceph_cap_string(revoking));
2428 if ((revoking & not) == 0) {
2429 if (!snap_rwsem_locked &&
2430 !ci->i_head_snapc &&
2431 (need & CEPH_CAP_FILE_WR)) {
2432 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2434 * we can not call down_read() when
2435 * task isn't in TASK_RUNNING state
2443 spin_unlock(&ci->i_ceph_lock);
2444 down_read(&mdsc->snap_rwsem);
2445 snap_rwsem_locked = true;
2448 snap_rwsem_locked = true;
2450 *got = need | (have & want);
2451 if ((need & CEPH_CAP_FILE_RD) &&
2452 !(*got & CEPH_CAP_FILE_CACHE))
2453 ceph_disable_fscache_readpage(ci);
2454 __take_cap_refs(ci, *got, true);
2458 int session_readonly = false;
2459 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2460 struct ceph_mds_session *s = ci->i_auth_cap->session;
2461 spin_lock(&s->s_cap_lock);
2462 session_readonly = s->s_readonly;
2463 spin_unlock(&s->s_cap_lock);
2465 if (session_readonly) {
2466 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2467 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2473 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) {
2475 if (ACCESS_ONCE(mdsc->fsc->mount_state) ==
2476 CEPH_MOUNT_SHUTDOWN) {
2477 dout("get_cap_refs %p forced umount\n", inode);
2482 mds_wanted = __ceph_caps_mds_wanted(ci);
2483 if ((mds_wanted & need) != need) {
2484 dout("get_cap_refs %p caps were dropped"
2485 " (session killed?)\n", inode);
2490 if ((mds_wanted & file_wanted) ==
2491 (file_wanted & (CEPH_CAP_FILE_RD|CEPH_CAP_FILE_WR)))
2492 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED;
2495 dout("get_cap_refs %p have %s needed %s\n", inode,
2496 ceph_cap_string(have), ceph_cap_string(need));
2499 spin_unlock(&ci->i_ceph_lock);
2500 if (snap_rwsem_locked)
2501 up_read(&mdsc->snap_rwsem);
2503 dout("get_cap_refs %p ret %d got %s\n", inode,
2504 ret, ceph_cap_string(*got));
2509 * Check the offset we are writing up to against our current
2510 * max_size. If necessary, tell the MDS we want to write to
2513 static void check_max_size(struct inode *inode, loff_t endoff)
2515 struct ceph_inode_info *ci = ceph_inode(inode);
2518 /* do we need to explicitly request a larger max_size? */
2519 spin_lock(&ci->i_ceph_lock);
2520 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2521 dout("write %p at large endoff %llu, req max_size\n",
2523 ci->i_wanted_max_size = endoff;
2525 /* duplicate ceph_check_caps()'s logic */
2526 if (ci->i_auth_cap &&
2527 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2528 ci->i_wanted_max_size > ci->i_max_size &&
2529 ci->i_wanted_max_size > ci->i_requested_max_size)
2531 spin_unlock(&ci->i_ceph_lock);
2533 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2536 int ceph_try_get_caps(struct ceph_inode_info *ci, int need, int want, int *got)
2540 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2541 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
2542 ret = ceph_pool_perm_check(ci, need);
2546 ret = try_get_cap_refs(ci, need, want, 0, true, got, &err);
2548 if (err == -EAGAIN) {
2550 } else if (err < 0) {
2558 * Wait for caps, and take cap references. If we can't get a WR cap
2559 * due to a small max_size, make sure we check_max_size (and possibly
2560 * ask the mds) so we don't get hung up indefinitely.
2562 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2563 loff_t endoff, int *got, struct page **pinned_page)
2565 int _got, ret, err = 0;
2567 ret = ceph_pool_perm_check(ci, need);
2573 check_max_size(&ci->vfs_inode, endoff);
2577 ret = try_get_cap_refs(ci, need, want, endoff,
2578 false, &_got, &err);
2585 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2586 add_wait_queue(&ci->i_cap_wq, &wait);
2588 while (!try_get_cap_refs(ci, need, want, endoff,
2590 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2592 remove_wait_queue(&ci->i_cap_wq, &wait);
2600 if (err == -ESTALE) {
2601 /* session was killed, try renew caps */
2602 ret = ceph_renew_caps(&ci->vfs_inode);
2609 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2610 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2611 i_size_read(&ci->vfs_inode) > 0) {
2613 find_get_page(ci->vfs_inode.i_mapping, 0);
2615 if (PageUptodate(page)) {
2616 *pinned_page = page;
2622 * drop cap refs first because getattr while
2623 * holding * caps refs can cause deadlock.
2625 ceph_put_cap_refs(ci, _got);
2629 * getattr request will bring inline data into
2632 ret = __ceph_do_getattr(&ci->vfs_inode, NULL,
2633 CEPH_STAT_CAP_INLINE_DATA,
2642 if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
2643 ceph_fscache_revalidate_cookie(ci);
2650 * Take cap refs. Caller must already know we hold at least one ref
2651 * on the caps in question or we don't know this is safe.
2653 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2655 spin_lock(&ci->i_ceph_lock);
2656 __take_cap_refs(ci, caps, false);
2657 spin_unlock(&ci->i_ceph_lock);
2662 * drop cap_snap that is not associated with any snapshot.
2663 * we don't need to send FLUSHSNAP message for it.
2665 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2666 struct ceph_cap_snap *capsnap)
2668 if (!capsnap->need_flush &&
2669 !capsnap->writing && !capsnap->dirty_pages) {
2670 dout("dropping cap_snap %p follows %llu\n",
2671 capsnap, capsnap->follows);
2672 BUG_ON(capsnap->cap_flush.tid > 0);
2673 ceph_put_snap_context(capsnap->context);
2674 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2675 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2677 list_del(&capsnap->ci_item);
2678 ceph_put_cap_snap(capsnap);
2687 * If we released the last ref on any given cap, call ceph_check_caps
2688 * to release (or schedule a release).
2690 * If we are releasing a WR cap (from a sync write), finalize any affected
2691 * cap_snap, and wake up any waiters.
2693 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2695 struct inode *inode = &ci->vfs_inode;
2696 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2698 spin_lock(&ci->i_ceph_lock);
2699 if (had & CEPH_CAP_PIN)
2701 if (had & CEPH_CAP_FILE_RD)
2702 if (--ci->i_rd_ref == 0)
2704 if (had & CEPH_CAP_FILE_CACHE)
2705 if (--ci->i_rdcache_ref == 0)
2707 if (had & CEPH_CAP_FILE_BUFFER) {
2708 if (--ci->i_wb_ref == 0) {
2712 dout("put_cap_refs %p wb %d -> %d (?)\n",
2713 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2715 if (had & CEPH_CAP_FILE_WR)
2716 if (--ci->i_wr_ref == 0) {
2718 if (__ceph_have_pending_cap_snap(ci)) {
2719 struct ceph_cap_snap *capsnap =
2720 list_last_entry(&ci->i_cap_snaps,
2721 struct ceph_cap_snap,
2723 capsnap->writing = 0;
2724 if (ceph_try_drop_cap_snap(ci, capsnap))
2726 else if (__ceph_finish_cap_snap(ci, capsnap))
2730 if (ci->i_wrbuffer_ref_head == 0 &&
2731 ci->i_dirty_caps == 0 &&
2732 ci->i_flushing_caps == 0) {
2733 BUG_ON(!ci->i_head_snapc);
2734 ceph_put_snap_context(ci->i_head_snapc);
2735 ci->i_head_snapc = NULL;
2737 /* see comment in __ceph_remove_cap() */
2738 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2739 drop_inode_snap_realm(ci);
2741 spin_unlock(&ci->i_ceph_lock);
2743 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2744 last ? " last" : "", put ? " put" : "");
2746 if (last && !flushsnaps)
2747 ceph_check_caps(ci, 0, NULL);
2748 else if (flushsnaps)
2749 ceph_flush_snaps(ci, NULL);
2751 wake_up_all(&ci->i_cap_wq);
2757 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2758 * context. Adjust per-snap dirty page accounting as appropriate.
2759 * Once all dirty data for a cap_snap is flushed, flush snapped file
2760 * metadata back to the MDS. If we dropped the last ref, call
2763 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2764 struct ceph_snap_context *snapc)
2766 struct inode *inode = &ci->vfs_inode;
2767 struct ceph_cap_snap *capsnap = NULL;
2771 bool flush_snaps = false;
2772 bool complete_capsnap = false;
2774 spin_lock(&ci->i_ceph_lock);
2775 ci->i_wrbuffer_ref -= nr;
2776 if (ci->i_wrbuffer_ref == 0) {
2781 if (ci->i_head_snapc == snapc) {
2782 ci->i_wrbuffer_ref_head -= nr;
2783 if (ci->i_wrbuffer_ref_head == 0 &&
2784 ci->i_wr_ref == 0 &&
2785 ci->i_dirty_caps == 0 &&
2786 ci->i_flushing_caps == 0) {
2787 BUG_ON(!ci->i_head_snapc);
2788 ceph_put_snap_context(ci->i_head_snapc);
2789 ci->i_head_snapc = NULL;
2791 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2793 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2794 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2795 last ? " LAST" : "");
2797 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2798 if (capsnap->context == snapc) {
2804 capsnap->dirty_pages -= nr;
2805 if (capsnap->dirty_pages == 0) {
2806 complete_capsnap = true;
2807 if (!capsnap->writing) {
2808 if (ceph_try_drop_cap_snap(ci, capsnap)) {
2811 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2816 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2817 " snap %lld %d/%d -> %d/%d %s%s\n",
2818 inode, capsnap, capsnap->context->seq,
2819 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2820 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2821 last ? " (wrbuffer last)" : "",
2822 complete_capsnap ? " (complete capsnap)" : "");
2825 spin_unlock(&ci->i_ceph_lock);
2828 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2829 } else if (flush_snaps) {
2830 ceph_flush_snaps(ci, NULL);
2832 if (complete_capsnap)
2833 wake_up_all(&ci->i_cap_wq);
2839 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2841 static void invalidate_aliases(struct inode *inode)
2843 struct dentry *dn, *prev = NULL;
2845 dout("invalidate_aliases inode %p\n", inode);
2846 d_prune_aliases(inode);
2848 * For non-directory inode, d_find_alias() only returns
2849 * hashed dentry. After calling d_invalidate(), the
2850 * dentry becomes unhashed.
2852 * For directory inode, d_find_alias() can return
2853 * unhashed dentry. But directory inode should have
2854 * one alias at most.
2856 while ((dn = d_find_alias(inode))) {
2871 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2872 * actually be a revocation if it specifies a smaller cap set.)
2874 * caller holds s_mutex and i_ceph_lock, we drop both.
2876 static void handle_cap_grant(struct ceph_mds_client *mdsc,
2877 struct inode *inode, struct ceph_mds_caps *grant,
2878 struct ceph_string **pns, u64 inline_version,
2879 void *inline_data, u32 inline_len,
2880 struct ceph_buffer *xattr_buf,
2881 struct ceph_mds_session *session,
2882 struct ceph_cap *cap, int issued)
2883 __releases(ci->i_ceph_lock)
2884 __releases(mdsc->snap_rwsem)
2886 struct ceph_inode_info *ci = ceph_inode(inode);
2887 int mds = session->s_mds;
2888 int seq = le32_to_cpu(grant->seq);
2889 int newcaps = le32_to_cpu(grant->caps);
2890 int used, wanted, dirty;
2891 u64 size = le64_to_cpu(grant->size);
2892 u64 max_size = le64_to_cpu(grant->max_size);
2893 struct timespec mtime, atime, ctime;
2896 bool writeback = false;
2897 bool queue_trunc = false;
2898 bool queue_invalidate = false;
2899 bool deleted_inode = false;
2900 bool fill_inline = false;
2902 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2903 inode, cap, mds, seq, ceph_cap_string(newcaps));
2904 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2909 * auth mds of the inode changed. we received the cap export message,
2910 * but still haven't received the cap import message. handle_cap_export
2911 * updated the new auth MDS' cap.
2913 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2914 * that was sent before the cap import message. So don't remove caps.
2916 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
2917 WARN_ON(cap != ci->i_auth_cap);
2918 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
2920 newcaps |= cap->issued;
2924 * If CACHE is being revoked, and we have no dirty buffers,
2925 * try to invalidate (once). (If there are dirty buffers, we
2926 * will invalidate _after_ writeback.)
2928 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
2929 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2930 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2931 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
2932 if (try_nonblocking_invalidate(inode)) {
2933 /* there were locked pages.. invalidate later
2934 in a separate thread. */
2935 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2936 queue_invalidate = true;
2937 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2942 /* side effects now are allowed */
2943 cap->cap_gen = session->s_cap_gen;
2946 __check_cap_issue(ci, cap, newcaps);
2948 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2949 (issued & CEPH_CAP_AUTH_EXCL) == 0) {
2950 inode->i_mode = le32_to_cpu(grant->mode);
2951 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
2952 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
2953 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2954 from_kuid(&init_user_ns, inode->i_uid),
2955 from_kgid(&init_user_ns, inode->i_gid));
2958 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2959 (issued & CEPH_CAP_LINK_EXCL) == 0) {
2960 set_nlink(inode, le32_to_cpu(grant->nlink));
2961 if (inode->i_nlink == 0 &&
2962 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
2963 deleted_inode = true;
2966 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
2967 int len = le32_to_cpu(grant->xattr_len);
2968 u64 version = le64_to_cpu(grant->xattr_version);
2970 if (version > ci->i_xattrs.version) {
2971 dout(" got new xattrs v%llu on %p len %d\n",
2972 version, inode, len);
2973 if (ci->i_xattrs.blob)
2974 ceph_buffer_put(ci->i_xattrs.blob);
2975 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
2976 ci->i_xattrs.version = version;
2977 ceph_forget_all_cached_acls(inode);
2981 if (newcaps & CEPH_CAP_ANY_RD) {
2982 /* ctime/mtime/atime? */
2983 ceph_decode_timespec(&mtime, &grant->mtime);
2984 ceph_decode_timespec(&atime, &grant->atime);
2985 ceph_decode_timespec(&ctime, &grant->ctime);
2986 ceph_fill_file_time(inode, issued,
2987 le32_to_cpu(grant->time_warp_seq),
2988 &ctime, &mtime, &atime);
2991 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
2992 /* file layout may have changed */
2993 s64 old_pool = ci->i_layout.pool_id;
2994 struct ceph_string *old_ns;
2996 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
2997 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
2998 lockdep_is_held(&ci->i_ceph_lock));
2999 rcu_assign_pointer(ci->i_layout.pool_ns, *pns);
3001 if (ci->i_layout.pool_id != old_pool || *pns != old_ns)
3002 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3006 /* size/truncate_seq? */
3007 queue_trunc = ceph_fill_file_size(inode, issued,
3008 le32_to_cpu(grant->truncate_seq),
3009 le64_to_cpu(grant->truncate_size),
3011 /* max size increase? */
3012 if (ci->i_auth_cap == cap && max_size != ci->i_max_size) {
3013 dout("max_size %lld -> %llu\n",
3014 ci->i_max_size, max_size);
3015 ci->i_max_size = max_size;
3016 if (max_size >= ci->i_wanted_max_size) {
3017 ci->i_wanted_max_size = 0; /* reset */
3018 ci->i_requested_max_size = 0;
3024 /* check cap bits */
3025 wanted = __ceph_caps_wanted(ci);
3026 used = __ceph_caps_used(ci);
3027 dirty = __ceph_caps_dirty(ci);
3028 dout(" my wanted = %s, used = %s, dirty %s\n",
3029 ceph_cap_string(wanted),
3030 ceph_cap_string(used),
3031 ceph_cap_string(dirty));
3032 if (wanted != le32_to_cpu(grant->wanted)) {
3033 dout("mds wanted %s -> %s\n",
3034 ceph_cap_string(le32_to_cpu(grant->wanted)),
3035 ceph_cap_string(wanted));
3036 /* imported cap may not have correct mds_wanted */
3037 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
3041 /* revocation, grant, or no-op? */
3042 if (cap->issued & ~newcaps) {
3043 int revoking = cap->issued & ~newcaps;
3045 dout("revocation: %s -> %s (revoking %s)\n",
3046 ceph_cap_string(cap->issued),
3047 ceph_cap_string(newcaps),
3048 ceph_cap_string(revoking));
3049 if (revoking & used & CEPH_CAP_FILE_BUFFER)
3050 writeback = true; /* initiate writeback; will delay ack */
3051 else if (revoking == CEPH_CAP_FILE_CACHE &&
3052 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3054 ; /* do nothing yet, invalidation will be queued */
3055 else if (cap == ci->i_auth_cap)
3056 check_caps = 1; /* check auth cap only */
3058 check_caps = 2; /* check all caps */
3059 cap->issued = newcaps;
3060 cap->implemented |= newcaps;
3061 } else if (cap->issued == newcaps) {
3062 dout("caps unchanged: %s -> %s\n",
3063 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3065 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3066 ceph_cap_string(newcaps));
3067 /* non-auth MDS is revoking the newly grant caps ? */
3068 if (cap == ci->i_auth_cap &&
3069 __ceph_caps_revoking_other(ci, cap, newcaps))
3072 cap->issued = newcaps;
3073 cap->implemented |= newcaps; /* add bits only, to
3074 * avoid stepping on a
3075 * pending revocation */
3078 BUG_ON(cap->issued & ~cap->implemented);
3080 if (inline_version > 0 && inline_version >= ci->i_inline_version) {
3081 ci->i_inline_version = inline_version;
3082 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3083 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3087 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3088 if (newcaps & ~issued)
3090 kick_flushing_inode_caps(mdsc, session, inode);
3091 up_read(&mdsc->snap_rwsem);
3093 spin_unlock(&ci->i_ceph_lock);
3097 ceph_fill_inline_data(inode, NULL, inline_data, inline_len);
3100 ceph_queue_vmtruncate(inode);
3104 * queue inode for writeback: we can't actually call
3105 * filemap_write_and_wait, etc. from message handler
3108 ceph_queue_writeback(inode);
3109 if (queue_invalidate)
3110 ceph_queue_invalidate(inode);
3112 invalidate_aliases(inode);
3114 wake_up_all(&ci->i_cap_wq);
3116 if (check_caps == 1)
3117 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
3119 else if (check_caps == 2)
3120 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
3122 mutex_unlock(&session->s_mutex);
3126 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3127 * MDS has been safely committed.
3129 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3130 struct ceph_mds_caps *m,
3131 struct ceph_mds_session *session,
3132 struct ceph_cap *cap)
3133 __releases(ci->i_ceph_lock)
3135 struct ceph_inode_info *ci = ceph_inode(inode);
3136 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3137 struct ceph_cap_flush *cf, *tmp_cf;
3138 LIST_HEAD(to_remove);
3139 unsigned seq = le32_to_cpu(m->seq);
3140 int dirty = le32_to_cpu(m->dirty);
3146 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3147 if (cf->tid == flush_tid)
3149 if (cf->caps == 0) /* capsnap */
3151 if (cf->tid <= flush_tid) {
3152 if (__finish_cap_flush(NULL, ci, cf))
3154 list_add_tail(&cf->i_list, &to_remove);
3156 cleaned &= ~cf->caps;
3162 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3163 " flushing %s -> %s\n",
3164 inode, session->s_mds, seq, ceph_cap_string(dirty),
3165 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3166 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3168 if (list_empty(&to_remove) && !cleaned)
3171 ci->i_flushing_caps &= ~cleaned;
3173 spin_lock(&mdsc->cap_dirty_lock);
3175 list_for_each_entry(cf, &to_remove, i_list) {
3176 if (__finish_cap_flush(mdsc, NULL, cf))
3180 if (ci->i_flushing_caps == 0) {
3181 if (list_empty(&ci->i_cap_flush_list)) {
3182 list_del_init(&ci->i_flushing_item);
3183 if (!list_empty(&session->s_cap_flushing)) {
3184 dout(" mds%d still flushing cap on %p\n",
3186 &list_first_entry(&session->s_cap_flushing,
3187 struct ceph_inode_info,
3188 i_flushing_item)->vfs_inode);
3191 mdsc->num_cap_flushing--;
3192 dout(" inode %p now !flushing\n", inode);
3194 if (ci->i_dirty_caps == 0) {
3195 dout(" inode %p now clean\n", inode);
3196 BUG_ON(!list_empty(&ci->i_dirty_item));
3198 if (ci->i_wr_ref == 0 &&
3199 ci->i_wrbuffer_ref_head == 0) {
3200 BUG_ON(!ci->i_head_snapc);
3201 ceph_put_snap_context(ci->i_head_snapc);
3202 ci->i_head_snapc = NULL;
3205 BUG_ON(list_empty(&ci->i_dirty_item));
3208 spin_unlock(&mdsc->cap_dirty_lock);
3211 spin_unlock(&ci->i_ceph_lock);
3213 while (!list_empty(&to_remove)) {
3214 cf = list_first_entry(&to_remove,
3215 struct ceph_cap_flush, i_list);
3216 list_del(&cf->i_list);
3217 ceph_free_cap_flush(cf);
3221 wake_up_all(&ci->i_cap_wq);
3223 wake_up_all(&mdsc->cap_flushing_wq);
3229 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3230 * throw away our cap_snap.
3232 * Caller hold s_mutex.
3234 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3235 struct ceph_mds_caps *m,
3236 struct ceph_mds_session *session)
3238 struct ceph_inode_info *ci = ceph_inode(inode);
3239 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3240 u64 follows = le64_to_cpu(m->snap_follows);
3241 struct ceph_cap_snap *capsnap;
3242 bool flushed = false;
3243 bool wake_ci = false;
3244 bool wake_mdsc = false;
3246 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3247 inode, ci, session->s_mds, follows);
3249 spin_lock(&ci->i_ceph_lock);
3250 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3251 if (capsnap->follows == follows) {
3252 if (capsnap->cap_flush.tid != flush_tid) {
3253 dout(" cap_snap %p follows %lld tid %lld !="
3254 " %lld\n", capsnap, follows,
3255 flush_tid, capsnap->cap_flush.tid);
3261 dout(" skipping cap_snap %p follows %lld\n",
3262 capsnap, capsnap->follows);
3266 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3267 dout(" removing %p cap_snap %p follows %lld\n",
3268 inode, capsnap, follows);
3269 list_del(&capsnap->ci_item);
3270 if (__finish_cap_flush(NULL, ci, &capsnap->cap_flush))
3273 spin_lock(&mdsc->cap_dirty_lock);
3275 if (list_empty(&ci->i_cap_flush_list))
3276 list_del_init(&ci->i_flushing_item);
3278 if (__finish_cap_flush(mdsc, NULL, &capsnap->cap_flush))
3281 spin_unlock(&mdsc->cap_dirty_lock);
3283 spin_unlock(&ci->i_ceph_lock);
3285 ceph_put_snap_context(capsnap->context);
3286 ceph_put_cap_snap(capsnap);
3288 wake_up_all(&ci->i_cap_wq);
3290 wake_up_all(&mdsc->cap_flushing_wq);
3296 * Handle TRUNC from MDS, indicating file truncation.
3298 * caller hold s_mutex.
3300 static void handle_cap_trunc(struct inode *inode,
3301 struct ceph_mds_caps *trunc,
3302 struct ceph_mds_session *session)
3303 __releases(ci->i_ceph_lock)
3305 struct ceph_inode_info *ci = ceph_inode(inode);
3306 int mds = session->s_mds;
3307 int seq = le32_to_cpu(trunc->seq);
3308 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3309 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3310 u64 size = le64_to_cpu(trunc->size);
3311 int implemented = 0;
3312 int dirty = __ceph_caps_dirty(ci);
3313 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3314 int queue_trunc = 0;
3316 issued |= implemented | dirty;
3318 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3319 inode, mds, seq, truncate_size, truncate_seq);
3320 queue_trunc = ceph_fill_file_size(inode, issued,
3321 truncate_seq, truncate_size, size);
3322 spin_unlock(&ci->i_ceph_lock);
3325 ceph_queue_vmtruncate(inode);
3329 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3330 * different one. If we are the most recent migration we've seen (as
3331 * indicated by mseq), make note of the migrating cap bits for the
3332 * duration (until we see the corresponding IMPORT).
3334 * caller holds s_mutex
3336 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3337 struct ceph_mds_cap_peer *ph,
3338 struct ceph_mds_session *session)
3340 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3341 struct ceph_mds_session *tsession = NULL;
3342 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3343 struct ceph_inode_info *ci = ceph_inode(inode);
3345 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3346 unsigned t_seq, t_mseq;
3348 int mds = session->s_mds;
3351 t_cap_id = le64_to_cpu(ph->cap_id);
3352 t_seq = le32_to_cpu(ph->seq);
3353 t_mseq = le32_to_cpu(ph->mseq);
3354 target = le32_to_cpu(ph->mds);
3356 t_cap_id = t_seq = t_mseq = 0;
3360 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3361 inode, ci, mds, mseq, target);
3363 spin_lock(&ci->i_ceph_lock);
3364 cap = __get_cap_for_mds(ci, mds);
3365 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3369 __ceph_remove_cap(cap, false);
3370 if (!ci->i_auth_cap)
3371 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
3376 * now we know we haven't received the cap import message yet
3377 * because the exported cap still exist.
3380 issued = cap->issued;
3381 WARN_ON(issued != cap->implemented);
3383 tcap = __get_cap_for_mds(ci, target);
3385 /* already have caps from the target */
3386 if (tcap->cap_id != t_cap_id ||
3387 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3388 dout(" updating import cap %p mds%d\n", tcap, target);
3389 tcap->cap_id = t_cap_id;
3390 tcap->seq = t_seq - 1;
3391 tcap->issue_seq = t_seq - 1;
3392 tcap->mseq = t_mseq;
3393 tcap->issued |= issued;
3394 tcap->implemented |= issued;
3395 if (cap == ci->i_auth_cap)
3396 ci->i_auth_cap = tcap;
3397 if (!list_empty(&ci->i_cap_flush_list) &&
3398 ci->i_auth_cap == tcap) {
3399 spin_lock(&mdsc->cap_dirty_lock);
3400 list_move_tail(&ci->i_flushing_item,
3401 &tcap->session->s_cap_flushing);
3402 spin_unlock(&mdsc->cap_dirty_lock);
3405 __ceph_remove_cap(cap, false);
3407 } else if (tsession) {
3408 /* add placeholder for the export tagert */
3409 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3410 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3411 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3413 __ceph_remove_cap(cap, false);
3417 spin_unlock(&ci->i_ceph_lock);
3418 mutex_unlock(&session->s_mutex);
3420 /* open target session */
3421 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3422 if (!IS_ERR(tsession)) {
3424 mutex_lock(&session->s_mutex);
3425 mutex_lock_nested(&tsession->s_mutex,
3426 SINGLE_DEPTH_NESTING);
3428 mutex_lock(&tsession->s_mutex);
3429 mutex_lock_nested(&session->s_mutex,
3430 SINGLE_DEPTH_NESTING);
3432 new_cap = ceph_get_cap(mdsc, NULL);
3441 spin_unlock(&ci->i_ceph_lock);
3442 mutex_unlock(&session->s_mutex);
3444 mutex_unlock(&tsession->s_mutex);
3445 ceph_put_mds_session(tsession);
3448 ceph_put_cap(mdsc, new_cap);
3452 * Handle cap IMPORT.
3454 * caller holds s_mutex. acquires i_ceph_lock
3456 static void handle_cap_import(struct ceph_mds_client *mdsc,
3457 struct inode *inode, struct ceph_mds_caps *im,
3458 struct ceph_mds_cap_peer *ph,
3459 struct ceph_mds_session *session,
3460 struct ceph_cap **target_cap, int *old_issued)
3461 __acquires(ci->i_ceph_lock)
3463 struct ceph_inode_info *ci = ceph_inode(inode);
3464 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3465 int mds = session->s_mds;
3467 unsigned caps = le32_to_cpu(im->caps);
3468 unsigned wanted = le32_to_cpu(im->wanted);
3469 unsigned seq = le32_to_cpu(im->seq);
3470 unsigned mseq = le32_to_cpu(im->migrate_seq);
3471 u64 realmino = le64_to_cpu(im->realm);
3472 u64 cap_id = le64_to_cpu(im->cap_id);
3477 p_cap_id = le64_to_cpu(ph->cap_id);
3478 peer = le32_to_cpu(ph->mds);
3484 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3485 inode, ci, mds, mseq, peer);
3488 spin_lock(&ci->i_ceph_lock);
3489 cap = __get_cap_for_mds(ci, mds);
3492 spin_unlock(&ci->i_ceph_lock);
3493 new_cap = ceph_get_cap(mdsc, NULL);
3499 ceph_put_cap(mdsc, new_cap);
3504 __ceph_caps_issued(ci, &issued);
3505 issued |= __ceph_caps_dirty(ci);
3507 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3508 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3510 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3511 if (ocap && ocap->cap_id == p_cap_id) {
3512 dout(" remove export cap %p mds%d flags %d\n",
3513 ocap, peer, ph->flags);
3514 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3515 (ocap->seq != le32_to_cpu(ph->seq) ||
3516 ocap->mseq != le32_to_cpu(ph->mseq))) {
3517 pr_err("handle_cap_import: mismatched seq/mseq: "
3518 "ino (%llx.%llx) mds%d seq %d mseq %d "
3519 "importer mds%d has peer seq %d mseq %d\n",
3520 ceph_vinop(inode), peer, ocap->seq,
3521 ocap->mseq, mds, le32_to_cpu(ph->seq),
3522 le32_to_cpu(ph->mseq));
3524 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3527 /* make sure we re-request max_size, if necessary */
3528 ci->i_wanted_max_size = 0;
3529 ci->i_requested_max_size = 0;
3531 *old_issued = issued;
3536 * Handle a caps message from the MDS.
3538 * Identify the appropriate session, inode, and call the right handler
3539 * based on the cap op.
3541 void ceph_handle_caps(struct ceph_mds_session *session,
3542 struct ceph_msg *msg)
3544 struct ceph_mds_client *mdsc = session->s_mdsc;
3545 struct super_block *sb = mdsc->fsc->sb;
3546 struct inode *inode;
3547 struct ceph_inode_info *ci;
3548 struct ceph_cap *cap;
3549 struct ceph_mds_caps *h;
3550 struct ceph_mds_cap_peer *peer = NULL;
3551 struct ceph_snap_realm *realm = NULL;
3552 struct ceph_string *pool_ns = NULL;
3553 int mds = session->s_mds;
3556 struct ceph_vino vino;
3558 u64 inline_version = 0;
3559 void *inline_data = NULL;
3562 size_t snaptrace_len;
3565 dout("handle_caps from mds%d\n", mds);
3568 end = msg->front.iov_base + msg->front.iov_len;
3569 tid = le64_to_cpu(msg->hdr.tid);
3570 if (msg->front.iov_len < sizeof(*h))
3572 h = msg->front.iov_base;
3573 op = le32_to_cpu(h->op);
3574 vino.ino = le64_to_cpu(h->ino);
3575 vino.snap = CEPH_NOSNAP;
3576 seq = le32_to_cpu(h->seq);
3577 mseq = le32_to_cpu(h->migrate_seq);
3580 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3581 p = snaptrace + snaptrace_len;
3583 if (le16_to_cpu(msg->hdr.version) >= 2) {
3585 ceph_decode_32_safe(&p, end, flock_len, bad);
3586 if (p + flock_len > end)
3591 if (le16_to_cpu(msg->hdr.version) >= 3) {
3592 if (op == CEPH_CAP_OP_IMPORT) {
3593 if (p + sizeof(*peer) > end)
3597 } else if (op == CEPH_CAP_OP_EXPORT) {
3598 /* recorded in unused fields */
3599 peer = (void *)&h->size;
3603 if (le16_to_cpu(msg->hdr.version) >= 4) {
3604 ceph_decode_64_safe(&p, end, inline_version, bad);
3605 ceph_decode_32_safe(&p, end, inline_len, bad);
3606 if (p + inline_len > end)
3612 if (le16_to_cpu(msg->hdr.version) >= 8) {
3614 u32 caller_uid, caller_gid;
3615 u32 osd_epoch_barrier;
3618 ceph_decode_32_safe(&p, end, osd_epoch_barrier, bad);
3620 ceph_decode_64_safe(&p, end, flush_tid, bad);
3622 ceph_decode_32_safe(&p, end, caller_uid, bad);
3623 ceph_decode_32_safe(&p, end, caller_gid, bad);
3625 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
3626 if (pool_ns_len > 0) {
3627 ceph_decode_need(&p, end, pool_ns_len, bad);
3628 pool_ns = ceph_find_or_create_string(p, pool_ns_len);
3634 inode = ceph_find_inode(sb, vino);
3635 ci = ceph_inode(inode);
3636 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3639 mutex_lock(&session->s_mutex);
3641 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3645 dout(" i don't have ino %llx\n", vino.ino);
3647 if (op == CEPH_CAP_OP_IMPORT) {
3648 cap = ceph_get_cap(mdsc, NULL);
3649 cap->cap_ino = vino.ino;
3650 cap->queue_release = 1;
3651 cap->cap_id = le64_to_cpu(h->cap_id);
3654 spin_lock(&session->s_cap_lock);
3655 list_add_tail(&cap->session_caps,
3656 &session->s_cap_releases);
3657 session->s_num_cap_releases++;
3658 spin_unlock(&session->s_cap_lock);
3660 goto flush_cap_releases;
3663 /* these will work even if we don't have a cap yet */
3665 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3666 handle_cap_flushsnap_ack(inode, tid, h, session);
3669 case CEPH_CAP_OP_EXPORT:
3670 handle_cap_export(inode, h, peer, session);
3673 case CEPH_CAP_OP_IMPORT:
3675 if (snaptrace_len) {
3676 down_write(&mdsc->snap_rwsem);
3677 ceph_update_snap_trace(mdsc, snaptrace,
3678 snaptrace + snaptrace_len,
3680 downgrade_write(&mdsc->snap_rwsem);
3682 down_read(&mdsc->snap_rwsem);
3684 handle_cap_import(mdsc, inode, h, peer, session,
3686 handle_cap_grant(mdsc, inode, h, &pool_ns,
3687 inline_version, inline_data, inline_len,
3688 msg->middle, session, cap, issued);
3690 ceph_put_snap_realm(mdsc, realm);
3694 /* the rest require a cap */
3695 spin_lock(&ci->i_ceph_lock);
3696 cap = __get_cap_for_mds(ceph_inode(inode), mds);
3698 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3699 inode, ceph_ino(inode), ceph_snap(inode), mds);
3700 spin_unlock(&ci->i_ceph_lock);
3701 goto flush_cap_releases;
3704 /* note that each of these drops i_ceph_lock for us */
3706 case CEPH_CAP_OP_REVOKE:
3707 case CEPH_CAP_OP_GRANT:
3708 __ceph_caps_issued(ci, &issued);
3709 issued |= __ceph_caps_dirty(ci);
3710 handle_cap_grant(mdsc, inode, h, &pool_ns,
3711 inline_version, inline_data, inline_len,
3712 msg->middle, session, cap, issued);
3715 case CEPH_CAP_OP_FLUSH_ACK:
3716 handle_cap_flush_ack(inode, tid, h, session, cap);
3719 case CEPH_CAP_OP_TRUNC:
3720 handle_cap_trunc(inode, h, session);
3724 spin_unlock(&ci->i_ceph_lock);
3725 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3726 ceph_cap_op_name(op));
3733 * send any cap release message to try to move things
3734 * along for the mds (who clearly thinks we still have this
3737 ceph_send_cap_releases(mdsc, session);
3740 mutex_unlock(&session->s_mutex);
3743 ceph_put_string(pool_ns);
3747 pr_err("ceph_handle_caps: corrupt message\n");
3753 * Delayed work handler to process end of delayed cap release LRU list.
3755 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3757 struct ceph_inode_info *ci;
3758 int flags = CHECK_CAPS_NODELAY;
3760 dout("check_delayed_caps\n");
3762 spin_lock(&mdsc->cap_delay_lock);
3763 if (list_empty(&mdsc->cap_delay_list))
3765 ci = list_first_entry(&mdsc->cap_delay_list,
3766 struct ceph_inode_info,
3768 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3769 time_before(jiffies, ci->i_hold_caps_max))
3771 list_del_init(&ci->i_cap_delay_list);
3772 spin_unlock(&mdsc->cap_delay_lock);
3773 dout("check_delayed_caps on %p\n", &ci->vfs_inode);
3774 ceph_check_caps(ci, flags, NULL);
3776 spin_unlock(&mdsc->cap_delay_lock);
3780 * Flush all dirty caps to the mds
3782 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
3784 struct ceph_inode_info *ci;
3785 struct inode *inode;
3787 dout("flush_dirty_caps\n");
3788 spin_lock(&mdsc->cap_dirty_lock);
3789 while (!list_empty(&mdsc->cap_dirty)) {
3790 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
3792 inode = &ci->vfs_inode;
3794 dout("flush_dirty_caps %p\n", inode);
3795 spin_unlock(&mdsc->cap_dirty_lock);
3796 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
3798 spin_lock(&mdsc->cap_dirty_lock);
3800 spin_unlock(&mdsc->cap_dirty_lock);
3801 dout("flush_dirty_caps done\n");
3804 void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode)
3807 int bits = (fmode << 1) | 1;
3808 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
3809 if (bits & (1 << i))
3810 ci->i_nr_by_mode[i]++;
3815 * Drop open file reference. If we were the last open file,
3816 * we may need to release capabilities to the MDS (or schedule
3817 * their delayed release).
3819 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
3822 int bits = (fmode << 1) | 1;
3823 spin_lock(&ci->i_ceph_lock);
3824 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
3825 if (bits & (1 << i)) {
3826 BUG_ON(ci->i_nr_by_mode[i] == 0);
3827 if (--ci->i_nr_by_mode[i] == 0)
3831 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
3832 &ci->vfs_inode, fmode,
3833 ci->i_nr_by_mode[0], ci->i_nr_by_mode[1],
3834 ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]);
3835 spin_unlock(&ci->i_ceph_lock);
3837 if (last && ci->i_vino.snap == CEPH_NOSNAP)
3838 ceph_check_caps(ci, 0, NULL);
3842 * Helpers for embedding cap and dentry lease releases into mds
3845 * @force is used by dentry_release (below) to force inclusion of a
3846 * record for the directory inode, even when there aren't any caps to
3849 int ceph_encode_inode_release(void **p, struct inode *inode,
3850 int mds, int drop, int unless, int force)
3852 struct ceph_inode_info *ci = ceph_inode(inode);
3853 struct ceph_cap *cap;
3854 struct ceph_mds_request_release *rel = *p;
3858 spin_lock(&ci->i_ceph_lock);
3859 used = __ceph_caps_used(ci);
3860 dirty = __ceph_caps_dirty(ci);
3862 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3863 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3864 ceph_cap_string(unless));
3866 /* only drop unused, clean caps */
3867 drop &= ~(used | dirty);
3869 cap = __get_cap_for_mds(ci, mds);
3870 if (cap && __cap_is_valid(cap)) {
3872 ((cap->issued & drop) &&
3873 (cap->issued & unless) == 0)) {
3874 if ((cap->issued & drop) &&
3875 (cap->issued & unless) == 0) {
3876 int wanted = __ceph_caps_wanted(ci);
3877 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
3878 wanted |= cap->mds_wanted;
3879 dout("encode_inode_release %p cap %p "
3880 "%s -> %s, wanted %s -> %s\n", inode, cap,
3881 ceph_cap_string(cap->issued),
3882 ceph_cap_string(cap->issued & ~drop),
3883 ceph_cap_string(cap->mds_wanted),
3884 ceph_cap_string(wanted));
3886 cap->issued &= ~drop;
3887 cap->implemented &= ~drop;
3888 cap->mds_wanted = wanted;
3890 dout("encode_inode_release %p cap %p %s"
3891 " (force)\n", inode, cap,
3892 ceph_cap_string(cap->issued));
3895 rel->ino = cpu_to_le64(ceph_ino(inode));
3896 rel->cap_id = cpu_to_le64(cap->cap_id);
3897 rel->seq = cpu_to_le32(cap->seq);
3898 rel->issue_seq = cpu_to_le32(cap->issue_seq);
3899 rel->mseq = cpu_to_le32(cap->mseq);
3900 rel->caps = cpu_to_le32(cap->implemented);
3901 rel->wanted = cpu_to_le32(cap->mds_wanted);
3907 dout("encode_inode_release %p cap %p %s\n",
3908 inode, cap, ceph_cap_string(cap->issued));
3911 spin_unlock(&ci->i_ceph_lock);
3915 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3916 int mds, int drop, int unless)
3918 struct inode *dir = d_inode(dentry->d_parent);
3919 struct ceph_mds_request_release *rel = *p;
3920 struct ceph_dentry_info *di = ceph_dentry(dentry);
3925 * force an record for the directory caps if we have a dentry lease.
3926 * this is racy (can't take i_ceph_lock and d_lock together), but it
3927 * doesn't have to be perfect; the mds will revoke anything we don't
3930 spin_lock(&dentry->d_lock);
3931 if (di->lease_session && di->lease_session->s_mds == mds)
3933 spin_unlock(&dentry->d_lock);
3935 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
3937 spin_lock(&dentry->d_lock);
3938 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
3939 dout("encode_dentry_release %p mds%d seq %d\n",
3940 dentry, mds, (int)di->lease_seq);
3941 rel->dname_len = cpu_to_le32(dentry->d_name.len);
3942 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
3943 *p += dentry->d_name.len;
3944 rel->dname_seq = cpu_to_le32(di->lease_seq);
3945 __ceph_mdsc_drop_dentry_lease(dentry);
3947 spin_unlock(&dentry->d_lock);