1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/kernel.h>
5 #include <linux/sched/signal.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 ||
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) {
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, bool check)
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 (check && !__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;
1007 * Build and send a cap message to the given MDS.
1009 * Caller should be holding s_mutex.
1011 static int send_cap_msg(struct cap_msg_args *arg)
1013 struct ceph_mds_caps *fc;
1014 struct ceph_msg *msg;
1017 struct timespec zerotime = {0};
1018 struct ceph_osd_client *osdc = &arg->session->s_mdsc->fsc->client->osdc;
1020 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1021 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1022 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(arg->op),
1023 arg->cid, arg->ino, ceph_cap_string(arg->caps),
1024 ceph_cap_string(arg->wanted), ceph_cap_string(arg->dirty),
1025 arg->seq, arg->issue_seq, arg->flush_tid, arg->oldest_flush_tid,
1026 arg->mseq, arg->follows, arg->size, arg->max_size,
1028 arg->xattr_buf ? (int)arg->xattr_buf->vec.iov_len : 0);
1030 /* flock buffer size + inline version + inline data size +
1031 * osd_epoch_barrier + oldest_flush_tid */
1032 extra_len = 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4;
1033 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1038 msg->hdr.version = cpu_to_le16(10);
1039 msg->hdr.tid = cpu_to_le64(arg->flush_tid);
1041 fc = msg->front.iov_base;
1042 memset(fc, 0, sizeof(*fc));
1044 fc->cap_id = cpu_to_le64(arg->cid);
1045 fc->op = cpu_to_le32(arg->op);
1046 fc->seq = cpu_to_le32(arg->seq);
1047 fc->issue_seq = cpu_to_le32(arg->issue_seq);
1048 fc->migrate_seq = cpu_to_le32(arg->mseq);
1049 fc->caps = cpu_to_le32(arg->caps);
1050 fc->wanted = cpu_to_le32(arg->wanted);
1051 fc->dirty = cpu_to_le32(arg->dirty);
1052 fc->ino = cpu_to_le64(arg->ino);
1053 fc->snap_follows = cpu_to_le64(arg->follows);
1055 fc->size = cpu_to_le64(arg->size);
1056 fc->max_size = cpu_to_le64(arg->max_size);
1057 ceph_encode_timespec(&fc->mtime, &arg->mtime);
1058 ceph_encode_timespec(&fc->atime, &arg->atime);
1059 ceph_encode_timespec(&fc->ctime, &arg->ctime);
1060 fc->time_warp_seq = cpu_to_le32(arg->time_warp_seq);
1062 fc->uid = cpu_to_le32(from_kuid(&init_user_ns, arg->uid));
1063 fc->gid = cpu_to_le32(from_kgid(&init_user_ns, arg->gid));
1064 fc->mode = cpu_to_le32(arg->mode);
1066 fc->xattr_version = cpu_to_le64(arg->xattr_version);
1067 if (arg->xattr_buf) {
1068 msg->middle = ceph_buffer_get(arg->xattr_buf);
1069 fc->xattr_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1070 msg->hdr.middle_len = cpu_to_le32(arg->xattr_buf->vec.iov_len);
1074 /* flock buffer size (version 2) */
1075 ceph_encode_32(&p, 0);
1076 /* inline version (version 4) */
1077 ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
1078 /* inline data size */
1079 ceph_encode_32(&p, 0);
1081 * osd_epoch_barrier (version 5)
1082 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1083 * case it was recently changed
1085 ceph_encode_32(&p, READ_ONCE(osdc->epoch_barrier));
1086 /* oldest_flush_tid (version 6) */
1087 ceph_encode_64(&p, arg->oldest_flush_tid);
1090 * caller_uid/caller_gid (version 7)
1092 * Currently, we don't properly track which caller dirtied the caps
1093 * last, and force a flush of them when there is a conflict. For now,
1094 * just set this to 0:0, to emulate how the MDS has worked up to now.
1096 ceph_encode_32(&p, 0);
1097 ceph_encode_32(&p, 0);
1099 /* pool namespace (version 8) (mds always ignores this) */
1100 ceph_encode_32(&p, 0);
1103 * btime and change_attr (version 9)
1105 * We just zero these out for now, as the MDS ignores them unless
1106 * the requisite feature flags are set (which we don't do yet).
1108 ceph_encode_timespec(p, &zerotime);
1109 p += sizeof(struct ceph_timespec);
1110 ceph_encode_64(&p, 0);
1112 /* Advisory flags (version 10) */
1113 ceph_encode_32(&p, arg->flags);
1115 ceph_con_send(&arg->session->s_con, msg);
1120 * Queue cap releases when an inode is dropped from our cache. Since
1121 * inode is about to be destroyed, there is no need for i_ceph_lock.
1123 void ceph_queue_caps_release(struct inode *inode)
1125 struct ceph_inode_info *ci = ceph_inode(inode);
1128 p = rb_first(&ci->i_caps);
1130 struct ceph_cap *cap = rb_entry(p, struct ceph_cap, ci_node);
1132 __ceph_remove_cap(cap, true);
1137 * Send a cap msg on the given inode. Update our caps state, then
1138 * drop i_ceph_lock and send the message.
1140 * Make note of max_size reported/requested from mds, revoked caps
1141 * that have now been implemented.
1143 * Make half-hearted attempt ot to invalidate page cache if we are
1144 * dropping RDCACHE. Note that this will leave behind locked pages
1145 * that we'll then need to deal with elsewhere.
1147 * Return non-zero if delayed release, or we experienced an error
1148 * such that the caller should requeue + retry later.
1150 * called with i_ceph_lock, then drops it.
1151 * caller should hold snap_rwsem (read), s_mutex.
1153 static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1154 int op, bool sync, int used, int want, int retain,
1155 int flushing, u64 flush_tid, u64 oldest_flush_tid)
1156 __releases(cap->ci->i_ceph_lock)
1158 struct ceph_inode_info *ci = cap->ci;
1159 struct inode *inode = &ci->vfs_inode;
1160 struct cap_msg_args arg;
1161 int held, revoking, dropping;
1166 held = cap->issued | cap->implemented;
1167 revoking = cap->implemented & ~cap->issued;
1168 retain &= ~revoking;
1169 dropping = cap->issued & ~retain;
1171 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1172 inode, cap, cap->session,
1173 ceph_cap_string(held), ceph_cap_string(held & retain),
1174 ceph_cap_string(revoking));
1175 BUG_ON((retain & CEPH_CAP_PIN) == 0);
1177 arg.session = cap->session;
1179 /* don't release wanted unless we've waited a bit. */
1180 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1181 time_before(jiffies, ci->i_hold_caps_min)) {
1182 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1183 ceph_cap_string(cap->issued),
1184 ceph_cap_string(cap->issued & retain),
1185 ceph_cap_string(cap->mds_wanted),
1186 ceph_cap_string(want));
1187 want |= cap->mds_wanted;
1188 retain |= cap->issued;
1191 ci->i_ceph_flags &= ~(CEPH_I_NODELAY | CEPH_I_FLUSH);
1192 if (want & ~cap->mds_wanted) {
1193 /* user space may open/close single file frequently.
1194 * This avoids droping mds_wanted immediately after
1195 * requesting new mds_wanted.
1197 __cap_set_timeouts(mdsc, ci);
1200 cap->issued &= retain; /* drop bits we don't want */
1201 if (cap->implemented & ~cap->issued) {
1203 * Wake up any waiters on wanted -> needed transition.
1204 * This is due to the weird transition from buffered
1205 * to sync IO... we need to flush dirty pages _before_
1206 * allowing sync writes to avoid reordering.
1210 cap->implemented &= cap->issued | used;
1211 cap->mds_wanted = want;
1213 arg.ino = ceph_vino(inode).ino;
1214 arg.cid = cap->cap_id;
1215 arg.follows = flushing ? ci->i_head_snapc->seq : 0;
1216 arg.flush_tid = flush_tid;
1217 arg.oldest_flush_tid = oldest_flush_tid;
1219 arg.size = inode->i_size;
1220 ci->i_reported_size = arg.size;
1221 arg.max_size = ci->i_wanted_max_size;
1222 ci->i_requested_max_size = arg.max_size;
1224 if (flushing & CEPH_CAP_XATTR_EXCL) {
1225 __ceph_build_xattrs_blob(ci);
1226 arg.xattr_version = ci->i_xattrs.version;
1227 arg.xattr_buf = ci->i_xattrs.blob;
1229 arg.xattr_buf = NULL;
1232 arg.mtime = inode->i_mtime;
1233 arg.atime = inode->i_atime;
1234 arg.ctime = inode->i_ctime;
1237 arg.caps = cap->implemented;
1239 arg.dirty = flushing;
1242 arg.issue_seq = cap->issue_seq;
1243 arg.mseq = cap->mseq;
1244 arg.time_warp_seq = ci->i_time_warp_seq;
1246 arg.uid = inode->i_uid;
1247 arg.gid = inode->i_gid;
1248 arg.mode = inode->i_mode;
1250 arg.inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
1251 if (list_empty(&ci->i_cap_snaps))
1252 arg.flags = CEPH_CLIENT_CAPS_NO_CAPSNAP;
1254 arg.flags = CEPH_CLIENT_CAPS_PENDING_CAPSNAP;
1256 arg.flags |= CEPH_CLIENT_CAPS_SYNC;
1258 spin_unlock(&ci->i_ceph_lock);
1260 ret = send_cap_msg(&arg);
1262 dout("error sending cap msg, must requeue %p\n", inode);
1267 wake_up_all(&ci->i_cap_wq);
1272 static inline int __send_flush_snap(struct inode *inode,
1273 struct ceph_mds_session *session,
1274 struct ceph_cap_snap *capsnap,
1275 u32 mseq, u64 oldest_flush_tid)
1277 struct cap_msg_args arg;
1279 arg.session = session;
1280 arg.ino = ceph_vino(inode).ino;
1282 arg.follows = capsnap->follows;
1283 arg.flush_tid = capsnap->cap_flush.tid;
1284 arg.oldest_flush_tid = oldest_flush_tid;
1286 arg.size = capsnap->size;
1288 arg.xattr_version = capsnap->xattr_version;
1289 arg.xattr_buf = capsnap->xattr_blob;
1291 arg.atime = capsnap->atime;
1292 arg.mtime = capsnap->mtime;
1293 arg.ctime = capsnap->ctime;
1295 arg.op = CEPH_CAP_OP_FLUSHSNAP;
1296 arg.caps = capsnap->issued;
1298 arg.dirty = capsnap->dirty;
1303 arg.time_warp_seq = capsnap->time_warp_seq;
1305 arg.uid = capsnap->uid;
1306 arg.gid = capsnap->gid;
1307 arg.mode = capsnap->mode;
1309 arg.inline_data = capsnap->inline_data;
1312 return send_cap_msg(&arg);
1316 * When a snapshot is taken, clients accumulate dirty metadata on
1317 * inodes with capabilities in ceph_cap_snaps to describe the file
1318 * state at the time the snapshot was taken. This must be flushed
1319 * asynchronously back to the MDS once sync writes complete and dirty
1320 * data is written out.
1322 * Called under i_ceph_lock. Takes s_mutex as needed.
1324 static void __ceph_flush_snaps(struct ceph_inode_info *ci,
1325 struct ceph_mds_session *session)
1326 __releases(ci->i_ceph_lock)
1327 __acquires(ci->i_ceph_lock)
1329 struct inode *inode = &ci->vfs_inode;
1330 struct ceph_mds_client *mdsc = session->s_mdsc;
1331 struct ceph_cap_snap *capsnap;
1332 u64 oldest_flush_tid = 0;
1333 u64 first_tid = 1, last_tid = 0;
1335 dout("__flush_snaps %p session %p\n", inode, session);
1337 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
1339 * we need to wait for sync writes to complete and for dirty
1340 * pages to be written out.
1342 if (capsnap->dirty_pages || capsnap->writing)
1345 /* should be removed by ceph_try_drop_cap_snap() */
1346 BUG_ON(!capsnap->need_flush);
1348 /* only flush each capsnap once */
1349 if (capsnap->cap_flush.tid > 0) {
1350 dout(" already flushed %p, skipping\n", capsnap);
1354 spin_lock(&mdsc->cap_dirty_lock);
1355 capsnap->cap_flush.tid = ++mdsc->last_cap_flush_tid;
1356 list_add_tail(&capsnap->cap_flush.g_list,
1357 &mdsc->cap_flush_list);
1358 if (oldest_flush_tid == 0)
1359 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1360 if (list_empty(&ci->i_flushing_item)) {
1361 list_add_tail(&ci->i_flushing_item,
1362 &session->s_cap_flushing);
1364 spin_unlock(&mdsc->cap_dirty_lock);
1366 list_add_tail(&capsnap->cap_flush.i_list,
1367 &ci->i_cap_flush_list);
1370 first_tid = capsnap->cap_flush.tid;
1371 last_tid = capsnap->cap_flush.tid;
1374 ci->i_ceph_flags &= ~CEPH_I_FLUSH_SNAPS;
1376 while (first_tid <= last_tid) {
1377 struct ceph_cap *cap = ci->i_auth_cap;
1378 struct ceph_cap_flush *cf;
1381 if (!(cap && cap->session == session)) {
1382 dout("__flush_snaps %p auth cap %p not mds%d, "
1383 "stop\n", inode, cap, session->s_mds);
1388 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
1389 if (cf->tid >= first_tid) {
1397 first_tid = cf->tid + 1;
1399 capsnap = container_of(cf, struct ceph_cap_snap, cap_flush);
1400 refcount_inc(&capsnap->nref);
1401 spin_unlock(&ci->i_ceph_lock);
1403 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1404 inode, capsnap, cf->tid, ceph_cap_string(capsnap->dirty));
1406 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
1409 pr_err("__flush_snaps: error sending cap flushsnap, "
1410 "ino (%llx.%llx) tid %llu follows %llu\n",
1411 ceph_vinop(inode), cf->tid, capsnap->follows);
1414 ceph_put_cap_snap(capsnap);
1415 spin_lock(&ci->i_ceph_lock);
1419 void ceph_flush_snaps(struct ceph_inode_info *ci,
1420 struct ceph_mds_session **psession)
1422 struct inode *inode = &ci->vfs_inode;
1423 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1424 struct ceph_mds_session *session = NULL;
1427 dout("ceph_flush_snaps %p\n", inode);
1429 session = *psession;
1431 spin_lock(&ci->i_ceph_lock);
1432 if (!(ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)) {
1433 dout(" no capsnap needs flush, doing nothing\n");
1436 if (!ci->i_auth_cap) {
1437 dout(" no auth cap (migrating?), doing nothing\n");
1441 mds = ci->i_auth_cap->session->s_mds;
1442 if (session && session->s_mds != mds) {
1443 dout(" oops, wrong session %p mutex\n", session);
1444 mutex_unlock(&session->s_mutex);
1445 ceph_put_mds_session(session);
1449 spin_unlock(&ci->i_ceph_lock);
1450 mutex_lock(&mdsc->mutex);
1451 session = __ceph_lookup_mds_session(mdsc, mds);
1452 mutex_unlock(&mdsc->mutex);
1454 dout(" inverting session/ino locks on %p\n", session);
1455 mutex_lock(&session->s_mutex);
1460 // make sure flushsnap messages are sent in proper order.
1461 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
1462 __kick_flushing_caps(mdsc, session, ci, 0);
1463 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
1466 __ceph_flush_snaps(ci, session);
1468 spin_unlock(&ci->i_ceph_lock);
1471 *psession = session;
1473 mutex_unlock(&session->s_mutex);
1474 ceph_put_mds_session(session);
1476 /* we flushed them all; remove this inode from the queue */
1477 spin_lock(&mdsc->snap_flush_lock);
1478 list_del_init(&ci->i_snap_flush_item);
1479 spin_unlock(&mdsc->snap_flush_lock);
1483 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1484 * Caller is then responsible for calling __mark_inode_dirty with the
1485 * returned flags value.
1487 int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1488 struct ceph_cap_flush **pcf)
1490 struct ceph_mds_client *mdsc =
1491 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
1492 struct inode *inode = &ci->vfs_inode;
1493 int was = ci->i_dirty_caps;
1496 if (!ci->i_auth_cap) {
1497 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1498 "but no auth cap (session was closed?)\n",
1499 inode, ceph_ino(inode), ceph_cap_string(mask));
1503 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci->vfs_inode,
1504 ceph_cap_string(mask), ceph_cap_string(was),
1505 ceph_cap_string(was | mask));
1506 ci->i_dirty_caps |= mask;
1508 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1509 swap(ci->i_prealloc_cap_flush, *pcf);
1511 if (!ci->i_head_snapc) {
1512 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1513 ci->i_head_snapc = ceph_get_snap_context(
1514 ci->i_snap_realm->cached_context);
1516 dout(" inode %p now dirty snapc %p auth cap %p\n",
1517 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1518 BUG_ON(!list_empty(&ci->i_dirty_item));
1519 spin_lock(&mdsc->cap_dirty_lock);
1520 list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
1521 spin_unlock(&mdsc->cap_dirty_lock);
1522 if (ci->i_flushing_caps == 0) {
1524 dirty |= I_DIRTY_SYNC;
1527 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1529 BUG_ON(list_empty(&ci->i_dirty_item));
1530 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
1531 (mask & CEPH_CAP_FILE_BUFFER))
1532 dirty |= I_DIRTY_DATASYNC;
1533 __cap_delay_requeue(mdsc, ci);
1537 struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1539 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1542 void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1545 kmem_cache_free(ceph_cap_flush_cachep, cf);
1548 static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1550 if (!list_empty(&mdsc->cap_flush_list)) {
1551 struct ceph_cap_flush *cf =
1552 list_first_entry(&mdsc->cap_flush_list,
1553 struct ceph_cap_flush, g_list);
1560 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1561 * Return true if caller needs to wake up flush waiters.
1563 static bool __finish_cap_flush(struct ceph_mds_client *mdsc,
1564 struct ceph_inode_info *ci,
1565 struct ceph_cap_flush *cf)
1567 struct ceph_cap_flush *prev;
1568 bool wake = cf->wake;
1570 /* are there older pending cap flushes? */
1571 if (wake && cf->g_list.prev != &mdsc->cap_flush_list) {
1572 prev = list_prev_entry(cf, g_list);
1576 list_del(&cf->g_list);
1578 if (wake && cf->i_list.prev != &ci->i_cap_flush_list) {
1579 prev = list_prev_entry(cf, i_list);
1583 list_del(&cf->i_list);
1591 * Add dirty inode to the flushing list. Assigned a seq number so we
1592 * can wait for caps to flush without starving.
1594 * Called under i_ceph_lock.
1596 static int __mark_caps_flushing(struct inode *inode,
1597 struct ceph_mds_session *session, bool wake,
1598 u64 *flush_tid, u64 *oldest_flush_tid)
1600 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1601 struct ceph_inode_info *ci = ceph_inode(inode);
1602 struct ceph_cap_flush *cf = NULL;
1605 BUG_ON(ci->i_dirty_caps == 0);
1606 BUG_ON(list_empty(&ci->i_dirty_item));
1607 BUG_ON(!ci->i_prealloc_cap_flush);
1609 flushing = ci->i_dirty_caps;
1610 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1611 ceph_cap_string(flushing),
1612 ceph_cap_string(ci->i_flushing_caps),
1613 ceph_cap_string(ci->i_flushing_caps | flushing));
1614 ci->i_flushing_caps |= flushing;
1615 ci->i_dirty_caps = 0;
1616 dout(" inode %p now !dirty\n", inode);
1618 swap(cf, ci->i_prealloc_cap_flush);
1619 cf->caps = flushing;
1622 spin_lock(&mdsc->cap_dirty_lock);
1623 list_del_init(&ci->i_dirty_item);
1625 cf->tid = ++mdsc->last_cap_flush_tid;
1626 list_add_tail(&cf->g_list, &mdsc->cap_flush_list);
1627 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1629 if (list_empty(&ci->i_flushing_item)) {
1630 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1631 mdsc->num_cap_flushing++;
1633 spin_unlock(&mdsc->cap_dirty_lock);
1635 list_add_tail(&cf->i_list, &ci->i_cap_flush_list);
1637 *flush_tid = cf->tid;
1642 * try to invalidate mapping pages without blocking.
1644 static int try_nonblocking_invalidate(struct inode *inode)
1646 struct ceph_inode_info *ci = ceph_inode(inode);
1647 u32 invalidating_gen = ci->i_rdcache_gen;
1649 spin_unlock(&ci->i_ceph_lock);
1650 invalidate_mapping_pages(&inode->i_data, 0, -1);
1651 spin_lock(&ci->i_ceph_lock);
1653 if (inode->i_data.nrpages == 0 &&
1654 invalidating_gen == ci->i_rdcache_gen) {
1656 dout("try_nonblocking_invalidate %p success\n", inode);
1657 /* save any racing async invalidate some trouble */
1658 ci->i_rdcache_revoking = ci->i_rdcache_gen - 1;
1661 dout("try_nonblocking_invalidate %p failed\n", inode);
1665 bool __ceph_should_report_size(struct ceph_inode_info *ci)
1667 loff_t size = ci->vfs_inode.i_size;
1668 /* mds will adjust max size according to the reported size */
1669 if (ci->i_flushing_caps & CEPH_CAP_FILE_WR)
1671 if (size >= ci->i_max_size)
1673 /* half of previous max_size increment has been used */
1674 if (ci->i_max_size > ci->i_reported_size &&
1675 (size << 1) >= ci->i_max_size + ci->i_reported_size)
1681 * Swiss army knife function to examine currently used and wanted
1682 * versus held caps. Release, flush, ack revoked caps to mds as
1685 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1686 * cap release further.
1687 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1688 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1691 void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1692 struct ceph_mds_session *session)
1694 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1695 struct ceph_mds_client *mdsc = fsc->mdsc;
1696 struct inode *inode = &ci->vfs_inode;
1697 struct ceph_cap *cap;
1698 u64 flush_tid, oldest_flush_tid;
1699 int file_wanted, used, cap_used;
1700 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1701 int issued, implemented, want, retain, revoking, flushing = 0;
1702 int mds = -1; /* keep track of how far we've gone through i_caps list
1703 to avoid an infinite loop on retry */
1705 int delayed = 0, sent = 0, num;
1706 bool is_delayed = flags & CHECK_CAPS_NODELAY;
1707 bool queue_invalidate = false;
1708 bool force_requeue = false;
1709 bool tried_invalidate = false;
1711 /* if we are unmounting, flush any unused caps immediately. */
1715 spin_lock(&ci->i_ceph_lock);
1717 if (ci->i_ceph_flags & CEPH_I_FLUSH)
1718 flags |= CHECK_CAPS_FLUSH;
1722 spin_lock(&ci->i_ceph_lock);
1724 file_wanted = __ceph_caps_file_wanted(ci);
1725 used = __ceph_caps_used(ci);
1726 issued = __ceph_caps_issued(ci, &implemented);
1727 revoking = implemented & ~issued;
1730 retain = file_wanted | used | CEPH_CAP_PIN;
1731 if (!mdsc->stopping && inode->i_nlink > 0) {
1733 retain |= CEPH_CAP_ANY; /* be greedy */
1734 } else if (S_ISDIR(inode->i_mode) &&
1735 (issued & CEPH_CAP_FILE_SHARED) &&
1736 __ceph_dir_is_complete(ci)) {
1738 * If a directory is complete, we want to keep
1739 * the exclusive cap. So that MDS does not end up
1740 * revoking the shared cap on every create/unlink
1743 want = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
1747 retain |= CEPH_CAP_ANY_SHARED;
1749 * keep RD only if we didn't have the file open RW,
1750 * because then the mds would revoke it anyway to
1751 * journal max_size=0.
1753 if (ci->i_max_size == 0)
1754 retain |= CEPH_CAP_ANY_RD;
1758 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1759 " issued %s revoking %s retain %s %s%s%s\n", inode,
1760 ceph_cap_string(file_wanted),
1761 ceph_cap_string(used), ceph_cap_string(ci->i_dirty_caps),
1762 ceph_cap_string(ci->i_flushing_caps),
1763 ceph_cap_string(issued), ceph_cap_string(revoking),
1764 ceph_cap_string(retain),
1765 (flags & CHECK_CAPS_AUTHONLY) ? " AUTHONLY" : "",
1766 (flags & CHECK_CAPS_NODELAY) ? " NODELAY" : "",
1767 (flags & CHECK_CAPS_FLUSH) ? " FLUSH" : "");
1770 * If we no longer need to hold onto old our caps, and we may
1771 * have cached pages, but don't want them, then try to invalidate.
1772 * If we fail, it's because pages are locked.... try again later.
1774 if ((!is_delayed || mdsc->stopping) &&
1775 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */
1776 !(ci->i_wb_ref || ci->i_wrbuffer_ref) && /* no dirty pages... */
1777 inode->i_data.nrpages && /* have cached pages */
1778 (revoking & (CEPH_CAP_FILE_CACHE|
1779 CEPH_CAP_FILE_LAZYIO)) && /* or revoking cache */
1780 !tried_invalidate) {
1781 dout("check_caps trying to invalidate on %p\n", inode);
1782 if (try_nonblocking_invalidate(inode) < 0) {
1783 if (revoking & (CEPH_CAP_FILE_CACHE|
1784 CEPH_CAP_FILE_LAZYIO)) {
1785 dout("check_caps queuing invalidate\n");
1786 queue_invalidate = true;
1787 ci->i_rdcache_revoking = ci->i_rdcache_gen;
1789 dout("check_caps failed to invalidate pages\n");
1790 /* we failed to invalidate pages. check these
1791 caps again later. */
1792 force_requeue = true;
1793 __cap_set_timeouts(mdsc, ci);
1796 tried_invalidate = true;
1801 for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
1802 cap = rb_entry(p, struct ceph_cap, ci_node);
1805 /* avoid looping forever */
1806 if (mds >= cap->mds ||
1807 ((flags & CHECK_CAPS_AUTHONLY) && cap != ci->i_auth_cap))
1810 /* NOTE: no side-effects allowed, until we take s_mutex */
1813 if (ci->i_auth_cap && cap != ci->i_auth_cap)
1814 cap_used &= ~ci->i_auth_cap->issued;
1816 revoking = cap->implemented & ~cap->issued;
1817 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1818 cap->mds, cap, ceph_cap_string(cap_used),
1819 ceph_cap_string(cap->issued),
1820 ceph_cap_string(cap->implemented),
1821 ceph_cap_string(revoking));
1823 if (cap == ci->i_auth_cap &&
1824 (cap->issued & CEPH_CAP_FILE_WR)) {
1825 /* request larger max_size from MDS? */
1826 if (ci->i_wanted_max_size > ci->i_max_size &&
1827 ci->i_wanted_max_size > ci->i_requested_max_size) {
1828 dout("requesting new max_size\n");
1832 /* approaching file_max? */
1833 if (__ceph_should_report_size(ci)) {
1834 dout("i_size approaching max_size\n");
1838 /* flush anything dirty? */
1839 if (cap == ci->i_auth_cap) {
1840 if ((flags & CHECK_CAPS_FLUSH) && ci->i_dirty_caps) {
1841 dout("flushing dirty caps\n");
1844 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS) {
1845 dout("flushing snap caps\n");
1850 /* completed revocation? going down and there are no caps? */
1851 if (revoking && (revoking & cap_used) == 0) {
1852 dout("completed revocation of %s\n",
1853 ceph_cap_string(cap->implemented & ~cap->issued));
1857 /* want more caps from mds? */
1858 if (want & ~(cap->mds_wanted | cap->issued))
1861 /* things we might delay */
1862 if ((cap->issued & ~retain) == 0 &&
1863 cap->mds_wanted == want)
1864 continue; /* nope, all good */
1870 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0 &&
1871 time_before(jiffies, ci->i_hold_caps_max)) {
1872 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1873 ceph_cap_string(cap->issued),
1874 ceph_cap_string(cap->issued & retain),
1875 ceph_cap_string(cap->mds_wanted),
1876 ceph_cap_string(want));
1882 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1883 dout(" skipping %p I_NOFLUSH set\n", inode);
1887 if (session && session != cap->session) {
1888 dout("oops, wrong session %p mutex\n", session);
1889 mutex_unlock(&session->s_mutex);
1893 session = cap->session;
1894 if (mutex_trylock(&session->s_mutex) == 0) {
1895 dout("inverting session/ino locks on %p\n",
1897 spin_unlock(&ci->i_ceph_lock);
1898 if (took_snap_rwsem) {
1899 up_read(&mdsc->snap_rwsem);
1900 took_snap_rwsem = 0;
1902 mutex_lock(&session->s_mutex);
1907 /* kick flushing and flush snaps before sending normal
1909 if (cap == ci->i_auth_cap &&
1911 (CEPH_I_KICK_FLUSH | CEPH_I_FLUSH_SNAPS))) {
1912 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
1913 __kick_flushing_caps(mdsc, session, ci, 0);
1914 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
1916 if (ci->i_ceph_flags & CEPH_I_FLUSH_SNAPS)
1917 __ceph_flush_snaps(ci, session);
1922 /* take snap_rwsem after session mutex */
1923 if (!took_snap_rwsem) {
1924 if (down_read_trylock(&mdsc->snap_rwsem) == 0) {
1925 dout("inverting snap/in locks on %p\n",
1927 spin_unlock(&ci->i_ceph_lock);
1928 down_read(&mdsc->snap_rwsem);
1929 took_snap_rwsem = 1;
1932 took_snap_rwsem = 1;
1935 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
1936 flushing = __mark_caps_flushing(inode, session, false,
1942 spin_lock(&mdsc->cap_dirty_lock);
1943 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1944 spin_unlock(&mdsc->cap_dirty_lock);
1947 mds = cap->mds; /* remember mds, so we don't repeat */
1950 /* __send_cap drops i_ceph_lock */
1951 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, false,
1952 cap_used, want, retain, flushing,
1953 flush_tid, oldest_flush_tid);
1954 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1958 * Reschedule delayed caps release if we delayed anything,
1961 if (delayed && is_delayed)
1962 force_requeue = true; /* __send_cap delayed release; requeue */
1963 if (!delayed && !is_delayed)
1964 __cap_delay_cancel(mdsc, ci);
1965 else if (!is_delayed || force_requeue)
1966 __cap_delay_requeue(mdsc, ci);
1968 spin_unlock(&ci->i_ceph_lock);
1970 if (queue_invalidate)
1971 ceph_queue_invalidate(inode);
1974 mutex_unlock(&session->s_mutex);
1975 if (took_snap_rwsem)
1976 up_read(&mdsc->snap_rwsem);
1980 * Try to flush dirty caps back to the auth mds.
1982 static int try_flush_caps(struct inode *inode, u64 *ptid)
1984 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1985 struct ceph_inode_info *ci = ceph_inode(inode);
1986 struct ceph_mds_session *session = NULL;
1988 u64 flush_tid = 0, oldest_flush_tid = 0;
1991 spin_lock(&ci->i_ceph_lock);
1992 if (ci->i_ceph_flags & CEPH_I_NOFLUSH) {
1993 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode);
1996 if (ci->i_dirty_caps && ci->i_auth_cap) {
1997 struct ceph_cap *cap = ci->i_auth_cap;
1998 int used = __ceph_caps_used(ci);
1999 int want = __ceph_caps_wanted(ci);
2002 if (!session || session != cap->session) {
2003 spin_unlock(&ci->i_ceph_lock);
2005 mutex_unlock(&session->s_mutex);
2006 session = cap->session;
2007 mutex_lock(&session->s_mutex);
2010 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
2013 flushing = __mark_caps_flushing(inode, session, true,
2014 &flush_tid, &oldest_flush_tid);
2016 /* __send_cap drops i_ceph_lock */
2017 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, true,
2018 used, want, (cap->issued | cap->implemented),
2019 flushing, flush_tid, oldest_flush_tid);
2022 spin_lock(&ci->i_ceph_lock);
2023 __cap_delay_requeue(mdsc, ci);
2024 spin_unlock(&ci->i_ceph_lock);
2027 if (!list_empty(&ci->i_cap_flush_list)) {
2028 struct ceph_cap_flush *cf =
2029 list_last_entry(&ci->i_cap_flush_list,
2030 struct ceph_cap_flush, i_list);
2032 flush_tid = cf->tid;
2034 flushing = ci->i_flushing_caps;
2035 spin_unlock(&ci->i_ceph_lock);
2039 mutex_unlock(&session->s_mutex);
2046 * Return true if we've flushed caps through the given flush_tid.
2048 static int caps_are_flushed(struct inode *inode, u64 flush_tid)
2050 struct ceph_inode_info *ci = ceph_inode(inode);
2053 spin_lock(&ci->i_ceph_lock);
2054 if (!list_empty(&ci->i_cap_flush_list)) {
2055 struct ceph_cap_flush * cf =
2056 list_first_entry(&ci->i_cap_flush_list,
2057 struct ceph_cap_flush, i_list);
2058 if (cf->tid <= flush_tid)
2061 spin_unlock(&ci->i_ceph_lock);
2066 * wait for any unsafe requests to complete.
2068 static int unsafe_request_wait(struct inode *inode)
2070 struct ceph_inode_info *ci = ceph_inode(inode);
2071 struct ceph_mds_request *req1 = NULL, *req2 = NULL;
2074 spin_lock(&ci->i_unsafe_lock);
2075 if (S_ISDIR(inode->i_mode) && !list_empty(&ci->i_unsafe_dirops)) {
2076 req1 = list_last_entry(&ci->i_unsafe_dirops,
2077 struct ceph_mds_request,
2079 ceph_mdsc_get_request(req1);
2081 if (!list_empty(&ci->i_unsafe_iops)) {
2082 req2 = list_last_entry(&ci->i_unsafe_iops,
2083 struct ceph_mds_request,
2084 r_unsafe_target_item);
2085 ceph_mdsc_get_request(req2);
2087 spin_unlock(&ci->i_unsafe_lock);
2089 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2090 inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
2092 ret = !wait_for_completion_timeout(&req1->r_safe_completion,
2093 ceph_timeout_jiffies(req1->r_timeout));
2096 ceph_mdsc_put_request(req1);
2099 ret = !wait_for_completion_timeout(&req2->r_safe_completion,
2100 ceph_timeout_jiffies(req2->r_timeout));
2103 ceph_mdsc_put_request(req2);
2108 int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2110 struct inode *inode = file->f_mapping->host;
2111 struct ceph_inode_info *ci = ceph_inode(inode);
2116 dout("fsync %p%s\n", inode, datasync ? " datasync" : "");
2118 ret = file_write_and_wait_range(file, start, end);
2127 dirty = try_flush_caps(inode, &flush_tid);
2128 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
2130 ret = unsafe_request_wait(inode);
2133 * only wait on non-file metadata writeback (the mds
2134 * can recover size and mtime, so we don't need to
2137 if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
2138 ret = wait_event_interruptible(ci->i_cap_wq,
2139 caps_are_flushed(inode, flush_tid));
2141 inode_unlock(inode);
2143 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
2148 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2149 * queue inode for flush but don't do so immediately, because we can
2150 * get by with fewer MDS messages if we wait for data writeback to
2153 int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
2155 struct ceph_inode_info *ci = ceph_inode(inode);
2159 int wait = wbc->sync_mode == WB_SYNC_ALL;
2161 dout("write_inode %p wait=%d\n", inode, wait);
2163 dirty = try_flush_caps(inode, &flush_tid);
2165 err = wait_event_interruptible(ci->i_cap_wq,
2166 caps_are_flushed(inode, flush_tid));
2168 struct ceph_mds_client *mdsc =
2169 ceph_sb_to_client(inode->i_sb)->mdsc;
2171 spin_lock(&ci->i_ceph_lock);
2172 if (__ceph_caps_dirty(ci))
2173 __cap_delay_requeue_front(mdsc, ci);
2174 spin_unlock(&ci->i_ceph_lock);
2179 static void __kick_flushing_caps(struct ceph_mds_client *mdsc,
2180 struct ceph_mds_session *session,
2181 struct ceph_inode_info *ci,
2182 u64 oldest_flush_tid)
2183 __releases(ci->i_ceph_lock)
2184 __acquires(ci->i_ceph_lock)
2186 struct inode *inode = &ci->vfs_inode;
2187 struct ceph_cap *cap;
2188 struct ceph_cap_flush *cf;
2192 list_for_each_entry(cf, &ci->i_cap_flush_list, i_list) {
2193 if (cf->tid < first_tid)
2196 cap = ci->i_auth_cap;
2197 if (!(cap && cap->session == session)) {
2198 pr_err("%p auth cap %p not mds%d ???\n",
2199 inode, cap, session->s_mds);
2203 first_tid = cf->tid + 1;
2206 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2207 inode, cap, cf->tid, ceph_cap_string(cf->caps));
2208 ci->i_ceph_flags |= CEPH_I_NODELAY;
2209 ret = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2210 false, __ceph_caps_used(ci),
2211 __ceph_caps_wanted(ci),
2212 cap->issued | cap->implemented,
2213 cf->caps, cf->tid, oldest_flush_tid);
2215 pr_err("kick_flushing_caps: error sending "
2216 "cap flush, ino (%llx.%llx) "
2217 "tid %llu flushing %s\n",
2218 ceph_vinop(inode), cf->tid,
2219 ceph_cap_string(cf->caps));
2222 struct ceph_cap_snap *capsnap =
2223 container_of(cf, struct ceph_cap_snap,
2225 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2226 inode, capsnap, cf->tid,
2227 ceph_cap_string(capsnap->dirty));
2229 refcount_inc(&capsnap->nref);
2230 spin_unlock(&ci->i_ceph_lock);
2232 ret = __send_flush_snap(inode, session, capsnap, cap->mseq,
2235 pr_err("kick_flushing_caps: error sending "
2236 "cap flushsnap, ino (%llx.%llx) "
2237 "tid %llu follows %llu\n",
2238 ceph_vinop(inode), cf->tid,
2242 ceph_put_cap_snap(capsnap);
2245 spin_lock(&ci->i_ceph_lock);
2249 void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2250 struct ceph_mds_session *session)
2252 struct ceph_inode_info *ci;
2253 struct ceph_cap *cap;
2254 u64 oldest_flush_tid;
2256 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2258 spin_lock(&mdsc->cap_dirty_lock);
2259 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2260 spin_unlock(&mdsc->cap_dirty_lock);
2262 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2263 spin_lock(&ci->i_ceph_lock);
2264 cap = ci->i_auth_cap;
2265 if (!(cap && cap->session == session)) {
2266 pr_err("%p auth cap %p not mds%d ???\n",
2267 &ci->vfs_inode, cap, session->s_mds);
2268 spin_unlock(&ci->i_ceph_lock);
2274 * if flushing caps were revoked, we re-send the cap flush
2275 * in client reconnect stage. This guarantees MDS * processes
2276 * the cap flush message before issuing the flushing caps to
2279 if ((cap->issued & ci->i_flushing_caps) !=
2280 ci->i_flushing_caps) {
2281 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2282 __kick_flushing_caps(mdsc, session, ci,
2285 ci->i_ceph_flags |= CEPH_I_KICK_FLUSH;
2288 spin_unlock(&ci->i_ceph_lock);
2292 void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2293 struct ceph_mds_session *session)
2295 struct ceph_inode_info *ci;
2296 struct ceph_cap *cap;
2297 u64 oldest_flush_tid;
2299 dout("kick_flushing_caps mds%d\n", session->s_mds);
2301 spin_lock(&mdsc->cap_dirty_lock);
2302 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2303 spin_unlock(&mdsc->cap_dirty_lock);
2305 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2306 spin_lock(&ci->i_ceph_lock);
2307 cap = ci->i_auth_cap;
2308 if (!(cap && cap->session == session)) {
2309 pr_err("%p auth cap %p not mds%d ???\n",
2310 &ci->vfs_inode, cap, session->s_mds);
2311 spin_unlock(&ci->i_ceph_lock);
2314 if (ci->i_ceph_flags & CEPH_I_KICK_FLUSH) {
2315 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2316 __kick_flushing_caps(mdsc, session, ci,
2319 spin_unlock(&ci->i_ceph_lock);
2323 static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2324 struct ceph_mds_session *session,
2325 struct inode *inode)
2326 __releases(ci->i_ceph_lock)
2328 struct ceph_inode_info *ci = ceph_inode(inode);
2329 struct ceph_cap *cap;
2331 cap = ci->i_auth_cap;
2332 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2333 ceph_cap_string(ci->i_flushing_caps));
2335 if (!list_empty(&ci->i_cap_flush_list)) {
2336 u64 oldest_flush_tid;
2337 spin_lock(&mdsc->cap_dirty_lock);
2338 list_move_tail(&ci->i_flushing_item,
2339 &cap->session->s_cap_flushing);
2340 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2341 spin_unlock(&mdsc->cap_dirty_lock);
2343 ci->i_ceph_flags &= ~CEPH_I_KICK_FLUSH;
2344 __kick_flushing_caps(mdsc, session, ci, oldest_flush_tid);
2345 spin_unlock(&ci->i_ceph_lock);
2347 spin_unlock(&ci->i_ceph_lock);
2353 * Take references to capabilities we hold, so that we don't release
2354 * them to the MDS prematurely.
2356 * Protected by i_ceph_lock.
2358 static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2359 bool snap_rwsem_locked)
2361 if (got & CEPH_CAP_PIN)
2363 if (got & CEPH_CAP_FILE_RD)
2365 if (got & CEPH_CAP_FILE_CACHE)
2366 ci->i_rdcache_ref++;
2367 if (got & CEPH_CAP_FILE_WR) {
2368 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2369 BUG_ON(!snap_rwsem_locked);
2370 ci->i_head_snapc = ceph_get_snap_context(
2371 ci->i_snap_realm->cached_context);
2375 if (got & CEPH_CAP_FILE_BUFFER) {
2376 if (ci->i_wb_ref == 0)
2377 ihold(&ci->vfs_inode);
2379 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2380 &ci->vfs_inode, ci->i_wb_ref-1, ci->i_wb_ref);
2385 * Try to grab cap references. Specify those refs we @want, and the
2386 * minimal set we @need. Also include the larger offset we are writing
2387 * to (when applicable), and check against max_size here as well.
2388 * Note that caller is responsible for ensuring max_size increases are
2389 * requested from the MDS.
2391 static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2392 loff_t endoff, bool nonblock, int *got, int *err)
2394 struct inode *inode = &ci->vfs_inode;
2395 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2397 int have, implemented;
2399 bool snap_rwsem_locked = false;
2401 dout("get_cap_refs %p need %s want %s\n", inode,
2402 ceph_cap_string(need), ceph_cap_string(want));
2405 spin_lock(&ci->i_ceph_lock);
2407 /* make sure file is actually open */
2408 file_wanted = __ceph_caps_file_wanted(ci);
2409 if ((file_wanted & need) != need) {
2410 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2411 ceph_cap_string(need), ceph_cap_string(file_wanted));
2417 /* finish pending truncate */
2418 while (ci->i_truncate_pending) {
2419 spin_unlock(&ci->i_ceph_lock);
2420 if (snap_rwsem_locked) {
2421 up_read(&mdsc->snap_rwsem);
2422 snap_rwsem_locked = false;
2424 __ceph_do_pending_vmtruncate(inode);
2425 spin_lock(&ci->i_ceph_lock);
2428 have = __ceph_caps_issued(ci, &implemented);
2430 if (have & need & CEPH_CAP_FILE_WR) {
2431 if (endoff >= 0 && endoff > (loff_t)ci->i_max_size) {
2432 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2433 inode, endoff, ci->i_max_size);
2434 if (endoff > ci->i_requested_max_size) {
2441 * If a sync write is in progress, we must wait, so that we
2442 * can get a final snapshot value for size+mtime.
2444 if (__ceph_have_pending_cap_snap(ci)) {
2445 dout("get_cap_refs %p cap_snap_pending\n", inode);
2450 if ((have & need) == need) {
2452 * Look at (implemented & ~have & not) so that we keep waiting
2453 * on transition from wanted -> needed caps. This is needed
2454 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2455 * going before a prior buffered writeback happens.
2457 int not = want & ~(have & need);
2458 int revoking = implemented & ~have;
2459 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2460 inode, ceph_cap_string(have), ceph_cap_string(not),
2461 ceph_cap_string(revoking));
2462 if ((revoking & not) == 0) {
2463 if (!snap_rwsem_locked &&
2464 !ci->i_head_snapc &&
2465 (need & CEPH_CAP_FILE_WR)) {
2466 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2468 * we can not call down_read() when
2469 * task isn't in TASK_RUNNING state
2477 spin_unlock(&ci->i_ceph_lock);
2478 down_read(&mdsc->snap_rwsem);
2479 snap_rwsem_locked = true;
2482 snap_rwsem_locked = true;
2484 *got = need | (have & want);
2485 if ((need & CEPH_CAP_FILE_RD) &&
2486 !(*got & CEPH_CAP_FILE_CACHE))
2487 ceph_disable_fscache_readpage(ci);
2488 __take_cap_refs(ci, *got, true);
2492 int session_readonly = false;
2493 if ((need & CEPH_CAP_FILE_WR) && ci->i_auth_cap) {
2494 struct ceph_mds_session *s = ci->i_auth_cap->session;
2495 spin_lock(&s->s_cap_lock);
2496 session_readonly = s->s_readonly;
2497 spin_unlock(&s->s_cap_lock);
2499 if (session_readonly) {
2500 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2501 inode, ceph_cap_string(need), ci->i_auth_cap->mds);
2507 if (ci->i_ceph_flags & CEPH_I_CAP_DROPPED) {
2509 if (READ_ONCE(mdsc->fsc->mount_state) ==
2510 CEPH_MOUNT_SHUTDOWN) {
2511 dout("get_cap_refs %p forced umount\n", inode);
2516 mds_wanted = __ceph_caps_mds_wanted(ci, false);
2517 if (need & ~(mds_wanted & need)) {
2518 dout("get_cap_refs %p caps were dropped"
2519 " (session killed?)\n", inode);
2524 if (!(file_wanted & ~mds_wanted))
2525 ci->i_ceph_flags &= ~CEPH_I_CAP_DROPPED;
2528 dout("get_cap_refs %p have %s needed %s\n", inode,
2529 ceph_cap_string(have), ceph_cap_string(need));
2532 spin_unlock(&ci->i_ceph_lock);
2533 if (snap_rwsem_locked)
2534 up_read(&mdsc->snap_rwsem);
2536 dout("get_cap_refs %p ret %d got %s\n", inode,
2537 ret, ceph_cap_string(*got));
2542 * Check the offset we are writing up to against our current
2543 * max_size. If necessary, tell the MDS we want to write to
2546 static void check_max_size(struct inode *inode, loff_t endoff)
2548 struct ceph_inode_info *ci = ceph_inode(inode);
2551 /* do we need to explicitly request a larger max_size? */
2552 spin_lock(&ci->i_ceph_lock);
2553 if (endoff >= ci->i_max_size && endoff > ci->i_wanted_max_size) {
2554 dout("write %p at large endoff %llu, req max_size\n",
2556 ci->i_wanted_max_size = endoff;
2558 /* duplicate ceph_check_caps()'s logic */
2559 if (ci->i_auth_cap &&
2560 (ci->i_auth_cap->issued & CEPH_CAP_FILE_WR) &&
2561 ci->i_wanted_max_size > ci->i_max_size &&
2562 ci->i_wanted_max_size > ci->i_requested_max_size)
2564 spin_unlock(&ci->i_ceph_lock);
2566 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2569 int ceph_try_get_caps(struct ceph_inode_info *ci, int need, int want, int *got)
2573 BUG_ON(need & ~CEPH_CAP_FILE_RD);
2574 BUG_ON(want & ~(CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
2575 ret = ceph_pool_perm_check(ci, need);
2579 ret = try_get_cap_refs(ci, need, want, 0, true, got, &err);
2581 if (err == -EAGAIN) {
2583 } else if (err < 0) {
2591 * Wait for caps, and take cap references. If we can't get a WR cap
2592 * due to a small max_size, make sure we check_max_size (and possibly
2593 * ask the mds) so we don't get hung up indefinitely.
2595 int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2596 loff_t endoff, int *got, struct page **pinned_page)
2598 int _got, ret, err = 0;
2600 ret = ceph_pool_perm_check(ci, need);
2606 check_max_size(&ci->vfs_inode, endoff);
2610 ret = try_get_cap_refs(ci, need, want, endoff,
2611 false, &_got, &err);
2618 DEFINE_WAIT_FUNC(wait, woken_wake_function);
2619 add_wait_queue(&ci->i_cap_wq, &wait);
2621 while (!try_get_cap_refs(ci, need, want, endoff,
2622 true, &_got, &err)) {
2623 if (signal_pending(current)) {
2627 wait_woken(&wait, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2630 remove_wait_queue(&ci->i_cap_wq, &wait);
2638 if (err == -ESTALE) {
2639 /* session was killed, try renew caps */
2640 ret = ceph_renew_caps(&ci->vfs_inode);
2647 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2648 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2649 i_size_read(&ci->vfs_inode) > 0) {
2651 find_get_page(ci->vfs_inode.i_mapping, 0);
2653 if (PageUptodate(page)) {
2654 *pinned_page = page;
2660 * drop cap refs first because getattr while
2661 * holding * caps refs can cause deadlock.
2663 ceph_put_cap_refs(ci, _got);
2667 * getattr request will bring inline data into
2670 ret = __ceph_do_getattr(&ci->vfs_inode, NULL,
2671 CEPH_STAT_CAP_INLINE_DATA,
2680 if ((_got & CEPH_CAP_FILE_RD) && (_got & CEPH_CAP_FILE_CACHE))
2681 ceph_fscache_revalidate_cookie(ci);
2688 * Take cap refs. Caller must already know we hold at least one ref
2689 * on the caps in question or we don't know this is safe.
2691 void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2693 spin_lock(&ci->i_ceph_lock);
2694 __take_cap_refs(ci, caps, false);
2695 spin_unlock(&ci->i_ceph_lock);
2700 * drop cap_snap that is not associated with any snapshot.
2701 * we don't need to send FLUSHSNAP message for it.
2703 static int ceph_try_drop_cap_snap(struct ceph_inode_info *ci,
2704 struct ceph_cap_snap *capsnap)
2706 if (!capsnap->need_flush &&
2707 !capsnap->writing && !capsnap->dirty_pages) {
2708 dout("dropping cap_snap %p follows %llu\n",
2709 capsnap, capsnap->follows);
2710 BUG_ON(capsnap->cap_flush.tid > 0);
2711 ceph_put_snap_context(capsnap->context);
2712 if (!list_is_last(&capsnap->ci_item, &ci->i_cap_snaps))
2713 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2715 list_del(&capsnap->ci_item);
2716 ceph_put_cap_snap(capsnap);
2725 * If we released the last ref on any given cap, call ceph_check_caps
2726 * to release (or schedule a release).
2728 * If we are releasing a WR cap (from a sync write), finalize any affected
2729 * cap_snap, and wake up any waiters.
2731 void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2733 struct inode *inode = &ci->vfs_inode;
2734 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2736 spin_lock(&ci->i_ceph_lock);
2737 if (had & CEPH_CAP_PIN)
2739 if (had & CEPH_CAP_FILE_RD)
2740 if (--ci->i_rd_ref == 0)
2742 if (had & CEPH_CAP_FILE_CACHE)
2743 if (--ci->i_rdcache_ref == 0)
2745 if (had & CEPH_CAP_FILE_BUFFER) {
2746 if (--ci->i_wb_ref == 0) {
2750 dout("put_cap_refs %p wb %d -> %d (?)\n",
2751 inode, ci->i_wb_ref+1, ci->i_wb_ref);
2753 if (had & CEPH_CAP_FILE_WR)
2754 if (--ci->i_wr_ref == 0) {
2756 if (__ceph_have_pending_cap_snap(ci)) {
2757 struct ceph_cap_snap *capsnap =
2758 list_last_entry(&ci->i_cap_snaps,
2759 struct ceph_cap_snap,
2761 capsnap->writing = 0;
2762 if (ceph_try_drop_cap_snap(ci, capsnap))
2764 else if (__ceph_finish_cap_snap(ci, capsnap))
2768 if (ci->i_wrbuffer_ref_head == 0 &&
2769 ci->i_dirty_caps == 0 &&
2770 ci->i_flushing_caps == 0) {
2771 BUG_ON(!ci->i_head_snapc);
2772 ceph_put_snap_context(ci->i_head_snapc);
2773 ci->i_head_snapc = NULL;
2775 /* see comment in __ceph_remove_cap() */
2776 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
2777 drop_inode_snap_realm(ci);
2779 spin_unlock(&ci->i_ceph_lock);
2781 dout("put_cap_refs %p had %s%s%s\n", inode, ceph_cap_string(had),
2782 last ? " last" : "", put ? " put" : "");
2784 if (last && !flushsnaps)
2785 ceph_check_caps(ci, 0, NULL);
2786 else if (flushsnaps)
2787 ceph_flush_snaps(ci, NULL);
2789 wake_up_all(&ci->i_cap_wq);
2795 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2796 * context. Adjust per-snap dirty page accounting as appropriate.
2797 * Once all dirty data for a cap_snap is flushed, flush snapped file
2798 * metadata back to the MDS. If we dropped the last ref, call
2801 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2802 struct ceph_snap_context *snapc)
2804 struct inode *inode = &ci->vfs_inode;
2805 struct ceph_cap_snap *capsnap = NULL;
2809 bool flush_snaps = false;
2810 bool complete_capsnap = false;
2812 spin_lock(&ci->i_ceph_lock);
2813 ci->i_wrbuffer_ref -= nr;
2814 if (ci->i_wrbuffer_ref == 0) {
2819 if (ci->i_head_snapc == snapc) {
2820 ci->i_wrbuffer_ref_head -= nr;
2821 if (ci->i_wrbuffer_ref_head == 0 &&
2822 ci->i_wr_ref == 0 &&
2823 ci->i_dirty_caps == 0 &&
2824 ci->i_flushing_caps == 0) {
2825 BUG_ON(!ci->i_head_snapc);
2826 ceph_put_snap_context(ci->i_head_snapc);
2827 ci->i_head_snapc = NULL;
2829 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2831 ci->i_wrbuffer_ref+nr, ci->i_wrbuffer_ref_head+nr,
2832 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
2833 last ? " LAST" : "");
2835 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
2836 if (capsnap->context == snapc) {
2842 capsnap->dirty_pages -= nr;
2843 if (capsnap->dirty_pages == 0) {
2844 complete_capsnap = true;
2845 if (!capsnap->writing) {
2846 if (ceph_try_drop_cap_snap(ci, capsnap)) {
2849 ci->i_ceph_flags |= CEPH_I_FLUSH_SNAPS;
2854 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2855 " snap %lld %d/%d -> %d/%d %s%s\n",
2856 inode, capsnap, capsnap->context->seq,
2857 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2858 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2859 last ? " (wrbuffer last)" : "",
2860 complete_capsnap ? " (complete capsnap)" : "");
2863 spin_unlock(&ci->i_ceph_lock);
2866 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
2867 } else if (flush_snaps) {
2868 ceph_flush_snaps(ci, NULL);
2870 if (complete_capsnap)
2871 wake_up_all(&ci->i_cap_wq);
2877 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2879 static void invalidate_aliases(struct inode *inode)
2881 struct dentry *dn, *prev = NULL;
2883 dout("invalidate_aliases inode %p\n", inode);
2884 d_prune_aliases(inode);
2886 * For non-directory inode, d_find_alias() only returns
2887 * hashed dentry. After calling d_invalidate(), the
2888 * dentry becomes unhashed.
2890 * For directory inode, d_find_alias() can return
2891 * unhashed dentry. But directory inode should have
2892 * one alias at most.
2894 while ((dn = d_find_alias(inode))) {
2909 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2910 * actually be a revocation if it specifies a smaller cap set.)
2912 * caller holds s_mutex and i_ceph_lock, we drop both.
2914 static void handle_cap_grant(struct ceph_mds_client *mdsc,
2915 struct inode *inode, struct ceph_mds_caps *grant,
2916 struct ceph_string **pns, u64 inline_version,
2917 void *inline_data, u32 inline_len,
2918 struct ceph_buffer *xattr_buf,
2919 struct ceph_mds_session *session,
2920 struct ceph_cap *cap, int issued)
2921 __releases(ci->i_ceph_lock)
2922 __releases(mdsc->snap_rwsem)
2924 struct ceph_inode_info *ci = ceph_inode(inode);
2925 int mds = session->s_mds;
2926 int seq = le32_to_cpu(grant->seq);
2927 int newcaps = le32_to_cpu(grant->caps);
2928 int used, wanted, dirty;
2929 u64 size = le64_to_cpu(grant->size);
2930 u64 max_size = le64_to_cpu(grant->max_size);
2931 struct timespec mtime, atime, ctime;
2934 bool writeback = false;
2935 bool queue_trunc = false;
2936 bool queue_invalidate = false;
2937 bool deleted_inode = false;
2938 bool fill_inline = false;
2940 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2941 inode, cap, mds, seq, ceph_cap_string(newcaps));
2942 dout(" size %llu max_size %llu, i_size %llu\n", size, max_size,
2947 * auth mds of the inode changed. we received the cap export message,
2948 * but still haven't received the cap import message. handle_cap_export
2949 * updated the new auth MDS' cap.
2951 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2952 * that was sent before the cap import message. So don't remove caps.
2954 if (ceph_seq_cmp(seq, cap->seq) <= 0) {
2955 WARN_ON(cap != ci->i_auth_cap);
2956 WARN_ON(cap->cap_id != le64_to_cpu(grant->cap_id));
2958 newcaps |= cap->issued;
2962 * If CACHE is being revoked, and we have no dirty buffers,
2963 * try to invalidate (once). (If there are dirty buffers, we
2964 * will invalidate _after_ writeback.)
2966 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
2967 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2968 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2969 !(ci->i_wrbuffer_ref || ci->i_wb_ref)) {
2970 if (try_nonblocking_invalidate(inode)) {
2971 /* there were locked pages.. invalidate later
2972 in a separate thread. */
2973 if (ci->i_rdcache_revoking != ci->i_rdcache_gen) {
2974 queue_invalidate = true;
2975 ci->i_rdcache_revoking = ci->i_rdcache_gen;
2980 /* side effects now are allowed */
2981 cap->cap_gen = session->s_cap_gen;
2984 __check_cap_issue(ci, cap, newcaps);
2986 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2987 (issued & CEPH_CAP_AUTH_EXCL) == 0) {
2988 inode->i_mode = le32_to_cpu(grant->mode);
2989 inode->i_uid = make_kuid(&init_user_ns, le32_to_cpu(grant->uid));
2990 inode->i_gid = make_kgid(&init_user_ns, le32_to_cpu(grant->gid));
2991 dout("%p mode 0%o uid.gid %d.%d\n", inode, inode->i_mode,
2992 from_kuid(&init_user_ns, inode->i_uid),
2993 from_kgid(&init_user_ns, inode->i_gid));
2996 if ((newcaps & CEPH_CAP_AUTH_SHARED) &&
2997 (issued & CEPH_CAP_LINK_EXCL) == 0) {
2998 set_nlink(inode, le32_to_cpu(grant->nlink));
2999 if (inode->i_nlink == 0 &&
3000 (newcaps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
3001 deleted_inode = true;
3004 if ((issued & CEPH_CAP_XATTR_EXCL) == 0 && grant->xattr_len) {
3005 int len = le32_to_cpu(grant->xattr_len);
3006 u64 version = le64_to_cpu(grant->xattr_version);
3008 if (version > ci->i_xattrs.version) {
3009 dout(" got new xattrs v%llu on %p len %d\n",
3010 version, inode, len);
3011 if (ci->i_xattrs.blob)
3012 ceph_buffer_put(ci->i_xattrs.blob);
3013 ci->i_xattrs.blob = ceph_buffer_get(xattr_buf);
3014 ci->i_xattrs.version = version;
3015 ceph_forget_all_cached_acls(inode);
3019 if (newcaps & CEPH_CAP_ANY_RD) {
3020 /* ctime/mtime/atime? */
3021 ceph_decode_timespec(&mtime, &grant->mtime);
3022 ceph_decode_timespec(&atime, &grant->atime);
3023 ceph_decode_timespec(&ctime, &grant->ctime);
3024 ceph_fill_file_time(inode, issued,
3025 le32_to_cpu(grant->time_warp_seq),
3026 &ctime, &mtime, &atime);
3029 if (newcaps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
3030 /* file layout may have changed */
3031 s64 old_pool = ci->i_layout.pool_id;
3032 struct ceph_string *old_ns;
3034 ceph_file_layout_from_legacy(&ci->i_layout, &grant->layout);
3035 old_ns = rcu_dereference_protected(ci->i_layout.pool_ns,
3036 lockdep_is_held(&ci->i_ceph_lock));
3037 rcu_assign_pointer(ci->i_layout.pool_ns, *pns);
3039 if (ci->i_layout.pool_id != old_pool || *pns != old_ns)
3040 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
3044 /* size/truncate_seq? */
3045 queue_trunc = ceph_fill_file_size(inode, issued,
3046 le32_to_cpu(grant->truncate_seq),
3047 le64_to_cpu(grant->truncate_size),
3051 if (ci->i_auth_cap == cap && (newcaps & CEPH_CAP_ANY_FILE_WR)) {
3052 if (max_size != ci->i_max_size) {
3053 dout("max_size %lld -> %llu\n",
3054 ci->i_max_size, max_size);
3055 ci->i_max_size = max_size;
3056 if (max_size >= ci->i_wanted_max_size) {
3057 ci->i_wanted_max_size = 0; /* reset */
3058 ci->i_requested_max_size = 0;
3061 } else if (ci->i_wanted_max_size > ci->i_max_size &&
3062 ci->i_wanted_max_size > ci->i_requested_max_size) {
3063 /* CEPH_CAP_OP_IMPORT */
3068 /* check cap bits */
3069 wanted = __ceph_caps_wanted(ci);
3070 used = __ceph_caps_used(ci);
3071 dirty = __ceph_caps_dirty(ci);
3072 dout(" my wanted = %s, used = %s, dirty %s\n",
3073 ceph_cap_string(wanted),
3074 ceph_cap_string(used),
3075 ceph_cap_string(dirty));
3076 if (wanted != le32_to_cpu(grant->wanted)) {
3077 dout("mds wanted %s -> %s\n",
3078 ceph_cap_string(le32_to_cpu(grant->wanted)),
3079 ceph_cap_string(wanted));
3080 /* imported cap may not have correct mds_wanted */
3081 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT)
3085 /* revocation, grant, or no-op? */
3086 if (cap->issued & ~newcaps) {
3087 int revoking = cap->issued & ~newcaps;
3089 dout("revocation: %s -> %s (revoking %s)\n",
3090 ceph_cap_string(cap->issued),
3091 ceph_cap_string(newcaps),
3092 ceph_cap_string(revoking));
3093 if (revoking & used & CEPH_CAP_FILE_BUFFER)
3094 writeback = true; /* initiate writeback; will delay ack */
3095 else if (revoking == CEPH_CAP_FILE_CACHE &&
3096 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
3098 ; /* do nothing yet, invalidation will be queued */
3099 else if (cap == ci->i_auth_cap)
3100 check_caps = 1; /* check auth cap only */
3102 check_caps = 2; /* check all caps */
3103 cap->issued = newcaps;
3104 cap->implemented |= newcaps;
3105 } else if (cap->issued == newcaps) {
3106 dout("caps unchanged: %s -> %s\n",
3107 ceph_cap_string(cap->issued), ceph_cap_string(newcaps));
3109 dout("grant: %s -> %s\n", ceph_cap_string(cap->issued),
3110 ceph_cap_string(newcaps));
3111 /* non-auth MDS is revoking the newly grant caps ? */
3112 if (cap == ci->i_auth_cap &&
3113 __ceph_caps_revoking_other(ci, cap, newcaps))
3116 cap->issued = newcaps;
3117 cap->implemented |= newcaps; /* add bits only, to
3118 * avoid stepping on a
3119 * pending revocation */
3122 BUG_ON(cap->issued & ~cap->implemented);
3124 if (inline_version > 0 && inline_version >= ci->i_inline_version) {
3125 ci->i_inline_version = inline_version;
3126 if (ci->i_inline_version != CEPH_INLINE_NONE &&
3127 (newcaps & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)))
3131 if (le32_to_cpu(grant->op) == CEPH_CAP_OP_IMPORT) {
3132 if (newcaps & ~issued)
3134 kick_flushing_inode_caps(mdsc, session, inode);
3135 up_read(&mdsc->snap_rwsem);
3137 spin_unlock(&ci->i_ceph_lock);
3141 ceph_fill_inline_data(inode, NULL, inline_data, inline_len);
3144 ceph_queue_vmtruncate(inode);
3148 * queue inode for writeback: we can't actually call
3149 * filemap_write_and_wait, etc. from message handler
3152 ceph_queue_writeback(inode);
3153 if (queue_invalidate)
3154 ceph_queue_invalidate(inode);
3156 invalidate_aliases(inode);
3158 wake_up_all(&ci->i_cap_wq);
3160 if (check_caps == 1)
3161 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_AUTHONLY,
3163 else if (check_caps == 2)
3164 ceph_check_caps(ci, CHECK_CAPS_NODELAY, session);
3166 mutex_unlock(&session->s_mutex);
3170 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3171 * MDS has been safely committed.
3173 static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
3174 struct ceph_mds_caps *m,
3175 struct ceph_mds_session *session,
3176 struct ceph_cap *cap)
3177 __releases(ci->i_ceph_lock)
3179 struct ceph_inode_info *ci = ceph_inode(inode);
3180 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3181 struct ceph_cap_flush *cf, *tmp_cf;
3182 LIST_HEAD(to_remove);
3183 unsigned seq = le32_to_cpu(m->seq);
3184 int dirty = le32_to_cpu(m->dirty);
3190 list_for_each_entry_safe(cf, tmp_cf, &ci->i_cap_flush_list, i_list) {
3191 if (cf->tid == flush_tid)
3193 if (cf->caps == 0) /* capsnap */
3195 if (cf->tid <= flush_tid) {
3196 if (__finish_cap_flush(NULL, ci, cf))
3198 list_add_tail(&cf->i_list, &to_remove);
3200 cleaned &= ~cf->caps;
3206 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3207 " flushing %s -> %s\n",
3208 inode, session->s_mds, seq, ceph_cap_string(dirty),
3209 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
3210 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
3212 if (list_empty(&to_remove) && !cleaned)
3215 ci->i_flushing_caps &= ~cleaned;
3217 spin_lock(&mdsc->cap_dirty_lock);
3219 list_for_each_entry(cf, &to_remove, i_list) {
3220 if (__finish_cap_flush(mdsc, NULL, cf))
3224 if (ci->i_flushing_caps == 0) {
3225 if (list_empty(&ci->i_cap_flush_list)) {
3226 list_del_init(&ci->i_flushing_item);
3227 if (!list_empty(&session->s_cap_flushing)) {
3228 dout(" mds%d still flushing cap on %p\n",
3230 &list_first_entry(&session->s_cap_flushing,
3231 struct ceph_inode_info,
3232 i_flushing_item)->vfs_inode);
3235 mdsc->num_cap_flushing--;
3236 dout(" inode %p now !flushing\n", inode);
3238 if (ci->i_dirty_caps == 0) {
3239 dout(" inode %p now clean\n", inode);
3240 BUG_ON(!list_empty(&ci->i_dirty_item));
3242 if (ci->i_wr_ref == 0 &&
3243 ci->i_wrbuffer_ref_head == 0) {
3244 BUG_ON(!ci->i_head_snapc);
3245 ceph_put_snap_context(ci->i_head_snapc);
3246 ci->i_head_snapc = NULL;
3249 BUG_ON(list_empty(&ci->i_dirty_item));
3252 spin_unlock(&mdsc->cap_dirty_lock);
3255 spin_unlock(&ci->i_ceph_lock);
3257 while (!list_empty(&to_remove)) {
3258 cf = list_first_entry(&to_remove,
3259 struct ceph_cap_flush, i_list);
3260 list_del(&cf->i_list);
3261 ceph_free_cap_flush(cf);
3265 wake_up_all(&ci->i_cap_wq);
3267 wake_up_all(&mdsc->cap_flushing_wq);
3273 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3274 * throw away our cap_snap.
3276 * Caller hold s_mutex.
3278 static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
3279 struct ceph_mds_caps *m,
3280 struct ceph_mds_session *session)
3282 struct ceph_inode_info *ci = ceph_inode(inode);
3283 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3284 u64 follows = le64_to_cpu(m->snap_follows);
3285 struct ceph_cap_snap *capsnap;
3286 bool flushed = false;
3287 bool wake_ci = false;
3288 bool wake_mdsc = false;
3290 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3291 inode, ci, session->s_mds, follows);
3293 spin_lock(&ci->i_ceph_lock);
3294 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
3295 if (capsnap->follows == follows) {
3296 if (capsnap->cap_flush.tid != flush_tid) {
3297 dout(" cap_snap %p follows %lld tid %lld !="
3298 " %lld\n", capsnap, follows,
3299 flush_tid, capsnap->cap_flush.tid);
3305 dout(" skipping cap_snap %p follows %lld\n",
3306 capsnap, capsnap->follows);
3310 WARN_ON(capsnap->dirty_pages || capsnap->writing);
3311 dout(" removing %p cap_snap %p follows %lld\n",
3312 inode, capsnap, follows);
3313 list_del(&capsnap->ci_item);
3314 if (__finish_cap_flush(NULL, ci, &capsnap->cap_flush))
3317 spin_lock(&mdsc->cap_dirty_lock);
3319 if (list_empty(&ci->i_cap_flush_list))
3320 list_del_init(&ci->i_flushing_item);
3322 if (__finish_cap_flush(mdsc, NULL, &capsnap->cap_flush))
3325 spin_unlock(&mdsc->cap_dirty_lock);
3327 spin_unlock(&ci->i_ceph_lock);
3329 ceph_put_snap_context(capsnap->context);
3330 ceph_put_cap_snap(capsnap);
3332 wake_up_all(&ci->i_cap_wq);
3334 wake_up_all(&mdsc->cap_flushing_wq);
3340 * Handle TRUNC from MDS, indicating file truncation.
3342 * caller hold s_mutex.
3344 static void handle_cap_trunc(struct inode *inode,
3345 struct ceph_mds_caps *trunc,
3346 struct ceph_mds_session *session)
3347 __releases(ci->i_ceph_lock)
3349 struct ceph_inode_info *ci = ceph_inode(inode);
3350 int mds = session->s_mds;
3351 int seq = le32_to_cpu(trunc->seq);
3352 u32 truncate_seq = le32_to_cpu(trunc->truncate_seq);
3353 u64 truncate_size = le64_to_cpu(trunc->truncate_size);
3354 u64 size = le64_to_cpu(trunc->size);
3355 int implemented = 0;
3356 int dirty = __ceph_caps_dirty(ci);
3357 int issued = __ceph_caps_issued(ceph_inode(inode), &implemented);
3358 int queue_trunc = 0;
3360 issued |= implemented | dirty;
3362 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3363 inode, mds, seq, truncate_size, truncate_seq);
3364 queue_trunc = ceph_fill_file_size(inode, issued,
3365 truncate_seq, truncate_size, size);
3366 spin_unlock(&ci->i_ceph_lock);
3369 ceph_queue_vmtruncate(inode);
3373 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3374 * different one. If we are the most recent migration we've seen (as
3375 * indicated by mseq), make note of the migrating cap bits for the
3376 * duration (until we see the corresponding IMPORT).
3378 * caller holds s_mutex
3380 static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
3381 struct ceph_mds_cap_peer *ph,
3382 struct ceph_mds_session *session)
3384 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
3385 struct ceph_mds_session *tsession = NULL;
3386 struct ceph_cap *cap, *tcap, *new_cap = NULL;
3387 struct ceph_inode_info *ci = ceph_inode(inode);
3389 unsigned mseq = le32_to_cpu(ex->migrate_seq);
3390 unsigned t_seq, t_mseq;
3392 int mds = session->s_mds;
3395 t_cap_id = le64_to_cpu(ph->cap_id);
3396 t_seq = le32_to_cpu(ph->seq);
3397 t_mseq = le32_to_cpu(ph->mseq);
3398 target = le32_to_cpu(ph->mds);
3400 t_cap_id = t_seq = t_mseq = 0;
3404 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3405 inode, ci, mds, mseq, target);
3407 spin_lock(&ci->i_ceph_lock);
3408 cap = __get_cap_for_mds(ci, mds);
3409 if (!cap || cap->cap_id != le64_to_cpu(ex->cap_id))
3413 __ceph_remove_cap(cap, false);
3414 if (!ci->i_auth_cap)
3415 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
3420 * now we know we haven't received the cap import message yet
3421 * because the exported cap still exist.
3424 issued = cap->issued;
3425 WARN_ON(issued != cap->implemented);
3427 tcap = __get_cap_for_mds(ci, target);
3429 /* already have caps from the target */
3430 if (tcap->cap_id != t_cap_id ||
3431 ceph_seq_cmp(tcap->seq, t_seq) < 0) {
3432 dout(" updating import cap %p mds%d\n", tcap, target);
3433 tcap->cap_id = t_cap_id;
3434 tcap->seq = t_seq - 1;
3435 tcap->issue_seq = t_seq - 1;
3436 tcap->mseq = t_mseq;
3437 tcap->issued |= issued;
3438 tcap->implemented |= issued;
3439 if (cap == ci->i_auth_cap)
3440 ci->i_auth_cap = tcap;
3442 if (!list_empty(&ci->i_cap_flush_list) &&
3443 ci->i_auth_cap == tcap) {
3444 spin_lock(&mdsc->cap_dirty_lock);
3445 list_move_tail(&ci->i_flushing_item,
3446 &tcap->session->s_cap_flushing);
3447 spin_unlock(&mdsc->cap_dirty_lock);
3450 __ceph_remove_cap(cap, false);
3452 } else if (tsession) {
3453 /* add placeholder for the export tagert */
3454 int flag = (cap == ci->i_auth_cap) ? CEPH_CAP_FLAG_AUTH : 0;
3456 ceph_add_cap(inode, tsession, t_cap_id, -1, issued, 0,
3457 t_seq - 1, t_mseq, (u64)-1, flag, &new_cap);
3459 if (!list_empty(&ci->i_cap_flush_list) &&
3460 ci->i_auth_cap == tcap) {
3461 spin_lock(&mdsc->cap_dirty_lock);
3462 list_move_tail(&ci->i_flushing_item,
3463 &tcap->session->s_cap_flushing);
3464 spin_unlock(&mdsc->cap_dirty_lock);
3467 __ceph_remove_cap(cap, false);
3471 spin_unlock(&ci->i_ceph_lock);
3472 mutex_unlock(&session->s_mutex);
3474 /* open target session */
3475 tsession = ceph_mdsc_open_export_target_session(mdsc, target);
3476 if (!IS_ERR(tsession)) {
3478 mutex_lock(&session->s_mutex);
3479 mutex_lock_nested(&tsession->s_mutex,
3480 SINGLE_DEPTH_NESTING);
3482 mutex_lock(&tsession->s_mutex);
3483 mutex_lock_nested(&session->s_mutex,
3484 SINGLE_DEPTH_NESTING);
3486 new_cap = ceph_get_cap(mdsc, NULL);
3495 spin_unlock(&ci->i_ceph_lock);
3496 mutex_unlock(&session->s_mutex);
3498 mutex_unlock(&tsession->s_mutex);
3499 ceph_put_mds_session(tsession);
3502 ceph_put_cap(mdsc, new_cap);
3506 * Handle cap IMPORT.
3508 * caller holds s_mutex. acquires i_ceph_lock
3510 static void handle_cap_import(struct ceph_mds_client *mdsc,
3511 struct inode *inode, struct ceph_mds_caps *im,
3512 struct ceph_mds_cap_peer *ph,
3513 struct ceph_mds_session *session,
3514 struct ceph_cap **target_cap, int *old_issued)
3515 __acquires(ci->i_ceph_lock)
3517 struct ceph_inode_info *ci = ceph_inode(inode);
3518 struct ceph_cap *cap, *ocap, *new_cap = NULL;
3519 int mds = session->s_mds;
3521 unsigned caps = le32_to_cpu(im->caps);
3522 unsigned wanted = le32_to_cpu(im->wanted);
3523 unsigned seq = le32_to_cpu(im->seq);
3524 unsigned mseq = le32_to_cpu(im->migrate_seq);
3525 u64 realmino = le64_to_cpu(im->realm);
3526 u64 cap_id = le64_to_cpu(im->cap_id);
3531 p_cap_id = le64_to_cpu(ph->cap_id);
3532 peer = le32_to_cpu(ph->mds);
3538 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3539 inode, ci, mds, mseq, peer);
3542 spin_lock(&ci->i_ceph_lock);
3543 cap = __get_cap_for_mds(ci, mds);
3546 spin_unlock(&ci->i_ceph_lock);
3547 new_cap = ceph_get_cap(mdsc, NULL);
3553 ceph_put_cap(mdsc, new_cap);
3558 __ceph_caps_issued(ci, &issued);
3559 issued |= __ceph_caps_dirty(ci);
3561 ceph_add_cap(inode, session, cap_id, -1, caps, wanted, seq, mseq,
3562 realmino, CEPH_CAP_FLAG_AUTH, &new_cap);
3564 ocap = peer >= 0 ? __get_cap_for_mds(ci, peer) : NULL;
3565 if (ocap && ocap->cap_id == p_cap_id) {
3566 dout(" remove export cap %p mds%d flags %d\n",
3567 ocap, peer, ph->flags);
3568 if ((ph->flags & CEPH_CAP_FLAG_AUTH) &&
3569 (ocap->seq != le32_to_cpu(ph->seq) ||
3570 ocap->mseq != le32_to_cpu(ph->mseq))) {
3571 pr_err("handle_cap_import: mismatched seq/mseq: "
3572 "ino (%llx.%llx) mds%d seq %d mseq %d "
3573 "importer mds%d has peer seq %d mseq %d\n",
3574 ceph_vinop(inode), peer, ocap->seq,
3575 ocap->mseq, mds, le32_to_cpu(ph->seq),
3576 le32_to_cpu(ph->mseq));
3578 __ceph_remove_cap(ocap, (ph->flags & CEPH_CAP_FLAG_RELEASE));
3581 /* make sure we re-request max_size, if necessary */
3582 ci->i_requested_max_size = 0;
3584 *old_issued = issued;
3589 * Handle a caps message from the MDS.
3591 * Identify the appropriate session, inode, and call the right handler
3592 * based on the cap op.
3594 void ceph_handle_caps(struct ceph_mds_session *session,
3595 struct ceph_msg *msg)
3597 struct ceph_mds_client *mdsc = session->s_mdsc;
3598 struct super_block *sb = mdsc->fsc->sb;
3599 struct inode *inode;
3600 struct ceph_inode_info *ci;
3601 struct ceph_cap *cap;
3602 struct ceph_mds_caps *h;
3603 struct ceph_mds_cap_peer *peer = NULL;
3604 struct ceph_snap_realm *realm = NULL;
3605 struct ceph_string *pool_ns = NULL;
3606 int mds = session->s_mds;
3609 struct ceph_vino vino;
3611 u64 inline_version = 0;
3612 void *inline_data = NULL;
3615 size_t snaptrace_len;
3618 dout("handle_caps from mds%d\n", mds);
3621 end = msg->front.iov_base + msg->front.iov_len;
3622 tid = le64_to_cpu(msg->hdr.tid);
3623 if (msg->front.iov_len < sizeof(*h))
3625 h = msg->front.iov_base;
3626 op = le32_to_cpu(h->op);
3627 vino.ino = le64_to_cpu(h->ino);
3628 vino.snap = CEPH_NOSNAP;
3629 seq = le32_to_cpu(h->seq);
3630 mseq = le32_to_cpu(h->migrate_seq);
3633 snaptrace_len = le32_to_cpu(h->snap_trace_len);
3634 p = snaptrace + snaptrace_len;
3636 if (le16_to_cpu(msg->hdr.version) >= 2) {
3638 ceph_decode_32_safe(&p, end, flock_len, bad);
3639 if (p + flock_len > end)
3644 if (le16_to_cpu(msg->hdr.version) >= 3) {
3645 if (op == CEPH_CAP_OP_IMPORT) {
3646 if (p + sizeof(*peer) > end)
3650 } else if (op == CEPH_CAP_OP_EXPORT) {
3651 /* recorded in unused fields */
3652 peer = (void *)&h->size;
3656 if (le16_to_cpu(msg->hdr.version) >= 4) {
3657 ceph_decode_64_safe(&p, end, inline_version, bad);
3658 ceph_decode_32_safe(&p, end, inline_len, bad);
3659 if (p + inline_len > end)
3665 if (le16_to_cpu(msg->hdr.version) >= 5) {
3666 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
3669 ceph_decode_32_safe(&p, end, epoch_barrier, bad);
3670 ceph_osdc_update_epoch_barrier(osdc, epoch_barrier);
3673 if (le16_to_cpu(msg->hdr.version) >= 8) {
3675 u32 caller_uid, caller_gid;
3679 ceph_decode_64_safe(&p, end, flush_tid, bad);
3681 ceph_decode_32_safe(&p, end, caller_uid, bad);
3682 ceph_decode_32_safe(&p, end, caller_gid, bad);
3684 ceph_decode_32_safe(&p, end, pool_ns_len, bad);
3685 if (pool_ns_len > 0) {
3686 ceph_decode_need(&p, end, pool_ns_len, bad);
3687 pool_ns = ceph_find_or_create_string(p, pool_ns_len);
3693 inode = ceph_find_inode(sb, vino);
3694 ci = ceph_inode(inode);
3695 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op), vino.ino,
3698 mutex_lock(&session->s_mutex);
3700 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3704 dout(" i don't have ino %llx\n", vino.ino);
3706 if (op == CEPH_CAP_OP_IMPORT) {
3707 cap = ceph_get_cap(mdsc, NULL);
3708 cap->cap_ino = vino.ino;
3709 cap->queue_release = 1;
3710 cap->cap_id = le64_to_cpu(h->cap_id);
3713 cap->issue_seq = seq;
3714 spin_lock(&session->s_cap_lock);
3715 list_add_tail(&cap->session_caps,
3716 &session->s_cap_releases);
3717 session->s_num_cap_releases++;
3718 spin_unlock(&session->s_cap_lock);
3720 goto flush_cap_releases;
3723 /* these will work even if we don't have a cap yet */
3725 case CEPH_CAP_OP_FLUSHSNAP_ACK:
3726 handle_cap_flushsnap_ack(inode, tid, h, session);
3729 case CEPH_CAP_OP_EXPORT:
3730 handle_cap_export(inode, h, peer, session);
3733 case CEPH_CAP_OP_IMPORT:
3735 if (snaptrace_len) {
3736 down_write(&mdsc->snap_rwsem);
3737 ceph_update_snap_trace(mdsc, snaptrace,
3738 snaptrace + snaptrace_len,
3740 downgrade_write(&mdsc->snap_rwsem);
3742 down_read(&mdsc->snap_rwsem);
3744 handle_cap_import(mdsc, inode, h, peer, session,
3746 handle_cap_grant(mdsc, inode, h, &pool_ns,
3747 inline_version, inline_data, inline_len,
3748 msg->middle, session, cap, issued);
3750 ceph_put_snap_realm(mdsc, realm);
3754 /* the rest require a cap */
3755 spin_lock(&ci->i_ceph_lock);
3756 cap = __get_cap_for_mds(ceph_inode(inode), mds);
3758 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3759 inode, ceph_ino(inode), ceph_snap(inode), mds);
3760 spin_unlock(&ci->i_ceph_lock);
3761 goto flush_cap_releases;
3764 /* note that each of these drops i_ceph_lock for us */
3766 case CEPH_CAP_OP_REVOKE:
3767 case CEPH_CAP_OP_GRANT:
3768 __ceph_caps_issued(ci, &issued);
3769 issued |= __ceph_caps_dirty(ci);
3770 handle_cap_grant(mdsc, inode, h, &pool_ns,
3771 inline_version, inline_data, inline_len,
3772 msg->middle, session, cap, issued);
3775 case CEPH_CAP_OP_FLUSH_ACK:
3776 handle_cap_flush_ack(inode, tid, h, session, cap);
3779 case CEPH_CAP_OP_TRUNC:
3780 handle_cap_trunc(inode, h, session);
3784 spin_unlock(&ci->i_ceph_lock);
3785 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op,
3786 ceph_cap_op_name(op));
3793 * send any cap release message to try to move things
3794 * along for the mds (who clearly thinks we still have this
3797 ceph_send_cap_releases(mdsc, session);
3800 mutex_unlock(&session->s_mutex);
3803 ceph_put_string(pool_ns);
3807 pr_err("ceph_handle_caps: corrupt message\n");
3813 * Delayed work handler to process end of delayed cap release LRU list.
3815 void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
3817 struct inode *inode;
3818 struct ceph_inode_info *ci;
3819 int flags = CHECK_CAPS_NODELAY;
3821 dout("check_delayed_caps\n");
3823 spin_lock(&mdsc->cap_delay_lock);
3824 if (list_empty(&mdsc->cap_delay_list))
3826 ci = list_first_entry(&mdsc->cap_delay_list,
3827 struct ceph_inode_info,
3829 if ((ci->i_ceph_flags & CEPH_I_FLUSH) == 0 &&
3830 time_before(jiffies, ci->i_hold_caps_max))
3832 list_del_init(&ci->i_cap_delay_list);
3834 inode = igrab(&ci->vfs_inode);
3835 spin_unlock(&mdsc->cap_delay_lock);
3838 dout("check_delayed_caps on %p\n", inode);
3839 ceph_check_caps(ci, flags, NULL);
3843 spin_unlock(&mdsc->cap_delay_lock);
3847 * Flush all dirty caps to the mds
3849 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
3851 struct ceph_inode_info *ci;
3852 struct inode *inode;
3854 dout("flush_dirty_caps\n");
3855 spin_lock(&mdsc->cap_dirty_lock);
3856 while (!list_empty(&mdsc->cap_dirty)) {
3857 ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
3859 inode = &ci->vfs_inode;
3861 dout("flush_dirty_caps %p\n", inode);
3862 spin_unlock(&mdsc->cap_dirty_lock);
3863 ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH, NULL);
3865 spin_lock(&mdsc->cap_dirty_lock);
3867 spin_unlock(&mdsc->cap_dirty_lock);
3868 dout("flush_dirty_caps done\n");
3871 void __ceph_get_fmode(struct ceph_inode_info *ci, int fmode)
3874 int bits = (fmode << 1) | 1;
3875 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
3876 if (bits & (1 << i))
3877 ci->i_nr_by_mode[i]++;
3882 * Drop open file reference. If we were the last open file,
3883 * we may need to release capabilities to the MDS (or schedule
3884 * their delayed release).
3886 void ceph_put_fmode(struct ceph_inode_info *ci, int fmode)
3889 int bits = (fmode << 1) | 1;
3890 spin_lock(&ci->i_ceph_lock);
3891 for (i = 0; i < CEPH_FILE_MODE_BITS; i++) {
3892 if (bits & (1 << i)) {
3893 BUG_ON(ci->i_nr_by_mode[i] == 0);
3894 if (--ci->i_nr_by_mode[i] == 0)
3898 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
3899 &ci->vfs_inode, fmode,
3900 ci->i_nr_by_mode[0], ci->i_nr_by_mode[1],
3901 ci->i_nr_by_mode[2], ci->i_nr_by_mode[3]);
3902 spin_unlock(&ci->i_ceph_lock);
3904 if (last && ci->i_vino.snap == CEPH_NOSNAP)
3905 ceph_check_caps(ci, 0, NULL);
3909 * Helpers for embedding cap and dentry lease releases into mds
3912 * @force is used by dentry_release (below) to force inclusion of a
3913 * record for the directory inode, even when there aren't any caps to
3916 int ceph_encode_inode_release(void **p, struct inode *inode,
3917 int mds, int drop, int unless, int force)
3919 struct ceph_inode_info *ci = ceph_inode(inode);
3920 struct ceph_cap *cap;
3921 struct ceph_mds_request_release *rel = *p;
3925 spin_lock(&ci->i_ceph_lock);
3926 used = __ceph_caps_used(ci);
3927 dirty = __ceph_caps_dirty(ci);
3929 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3930 inode, mds, ceph_cap_string(used|dirty), ceph_cap_string(drop),
3931 ceph_cap_string(unless));
3933 /* only drop unused, clean caps */
3934 drop &= ~(used | dirty);
3936 cap = __get_cap_for_mds(ci, mds);
3937 if (cap && __cap_is_valid(cap)) {
3939 ((cap->issued & drop) &&
3940 (cap->issued & unless) == 0)) {
3941 if ((cap->issued & drop) &&
3942 (cap->issued & unless) == 0) {
3943 int wanted = __ceph_caps_wanted(ci);
3944 if ((ci->i_ceph_flags & CEPH_I_NODELAY) == 0)
3945 wanted |= cap->mds_wanted;
3946 dout("encode_inode_release %p cap %p "
3947 "%s -> %s, wanted %s -> %s\n", inode, cap,
3948 ceph_cap_string(cap->issued),
3949 ceph_cap_string(cap->issued & ~drop),
3950 ceph_cap_string(cap->mds_wanted),
3951 ceph_cap_string(wanted));
3953 cap->issued &= ~drop;
3954 cap->implemented &= ~drop;
3955 cap->mds_wanted = wanted;
3957 dout("encode_inode_release %p cap %p %s"
3958 " (force)\n", inode, cap,
3959 ceph_cap_string(cap->issued));
3962 rel->ino = cpu_to_le64(ceph_ino(inode));
3963 rel->cap_id = cpu_to_le64(cap->cap_id);
3964 rel->seq = cpu_to_le32(cap->seq);
3965 rel->issue_seq = cpu_to_le32(cap->issue_seq);
3966 rel->mseq = cpu_to_le32(cap->mseq);
3967 rel->caps = cpu_to_le32(cap->implemented);
3968 rel->wanted = cpu_to_le32(cap->mds_wanted);
3974 dout("encode_inode_release %p cap %p %s\n",
3975 inode, cap, ceph_cap_string(cap->issued));
3978 spin_unlock(&ci->i_ceph_lock);
3982 int ceph_encode_dentry_release(void **p, struct dentry *dentry,
3984 int mds, int drop, int unless)
3986 struct dentry *parent = NULL;
3987 struct ceph_mds_request_release *rel = *p;
3988 struct ceph_dentry_info *di = ceph_dentry(dentry);
3993 * force an record for the directory caps if we have a dentry lease.
3994 * this is racy (can't take i_ceph_lock and d_lock together), but it
3995 * doesn't have to be perfect; the mds will revoke anything we don't
3998 spin_lock(&dentry->d_lock);
3999 if (di->lease_session && di->lease_session->s_mds == mds)
4002 parent = dget(dentry->d_parent);
4003 dir = d_inode(parent);
4005 spin_unlock(&dentry->d_lock);
4007 ret = ceph_encode_inode_release(p, dir, mds, drop, unless, force);
4010 spin_lock(&dentry->d_lock);
4011 if (ret && di->lease_session && di->lease_session->s_mds == mds) {
4012 dout("encode_dentry_release %p mds%d seq %d\n",
4013 dentry, mds, (int)di->lease_seq);
4014 rel->dname_len = cpu_to_le32(dentry->d_name.len);
4015 memcpy(*p, dentry->d_name.name, dentry->d_name.len);
4016 *p += dentry->d_name.len;
4017 rel->dname_seq = cpu_to_le32(di->lease_seq);
4018 __ceph_mdsc_drop_dentry_lease(dentry);
4020 spin_unlock(&dentry->d_lock);