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
== NULL
||
615 ceph_seq_cmp(ci
->i_auth_cap
->mseq
, mseq
) < 0) {
616 ci
->i_auth_cap
= cap
;
617 cap
->mds_wanted
= wanted
;
620 WARN_ON(ci
->i_auth_cap
== cap
);
623 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
624 inode
, ceph_vinop(inode
), cap
, ceph_cap_string(issued
),
625 ceph_cap_string(issued
|cap
->issued
), seq
, mds
);
626 cap
->cap_id
= cap_id
;
627 cap
->issued
= issued
;
628 cap
->implemented
|= issued
;
629 if (ceph_seq_cmp(mseq
, cap
->mseq
) > 0)
630 cap
->mds_wanted
= wanted
;
632 cap
->mds_wanted
|= wanted
;
634 cap
->issue_seq
= seq
;
636 cap
->cap_gen
= session
->s_cap_gen
;
639 __ceph_get_fmode(ci
, fmode
);
643 * Return true if cap has not timed out and belongs to the current
644 * generation of the MDS session (i.e. has not gone 'stale' due to
645 * us losing touch with the mds).
647 static int __cap_is_valid(struct ceph_cap
*cap
)
652 spin_lock(&cap
->session
->s_gen_ttl_lock
);
653 gen
= cap
->session
->s_cap_gen
;
654 ttl
= cap
->session
->s_cap_ttl
;
655 spin_unlock(&cap
->session
->s_gen_ttl_lock
);
657 if (cap
->cap_gen
< gen
|| time_after_eq(jiffies
, ttl
)) {
658 dout("__cap_is_valid %p cap %p issued %s "
659 "but STALE (gen %u vs %u)\n", &cap
->ci
->vfs_inode
,
660 cap
, ceph_cap_string(cap
->issued
), cap
->cap_gen
, gen
);
668 * Return set of valid cap bits issued to us. Note that caps time
669 * out, and may be invalidated in bulk if the client session times out
670 * and session->s_cap_gen is bumped.
672 int __ceph_caps_issued(struct ceph_inode_info
*ci
, int *implemented
)
674 int have
= ci
->i_snap_caps
;
675 struct ceph_cap
*cap
;
680 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
681 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
682 if (!__cap_is_valid(cap
))
684 dout("__ceph_caps_issued %p cap %p issued %s\n",
685 &ci
->vfs_inode
, cap
, ceph_cap_string(cap
->issued
));
688 *implemented
|= cap
->implemented
;
691 * exclude caps issued by non-auth MDS, but are been revoking
692 * by the auth MDS. The non-auth MDS should be revoking/exporting
693 * these caps, but the message is delayed.
695 if (ci
->i_auth_cap
) {
696 cap
= ci
->i_auth_cap
;
697 have
&= ~cap
->implemented
| cap
->issued
;
703 * Get cap bits issued by caps other than @ocap
705 int __ceph_caps_issued_other(struct ceph_inode_info
*ci
, struct ceph_cap
*ocap
)
707 int have
= ci
->i_snap_caps
;
708 struct ceph_cap
*cap
;
711 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
712 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
715 if (!__cap_is_valid(cap
))
723 * Move a cap to the end of the LRU (oldest caps at list head, newest
726 static void __touch_cap(struct ceph_cap
*cap
)
728 struct ceph_mds_session
*s
= cap
->session
;
730 spin_lock(&s
->s_cap_lock
);
731 if (s
->s_cap_iterator
== NULL
) {
732 dout("__touch_cap %p cap %p mds%d\n", &cap
->ci
->vfs_inode
, cap
,
734 list_move_tail(&cap
->session_caps
, &s
->s_caps
);
736 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
737 &cap
->ci
->vfs_inode
, cap
, s
->s_mds
);
739 spin_unlock(&s
->s_cap_lock
);
743 * Check if we hold the given mask. If so, move the cap(s) to the
744 * front of their respective LRUs. (This is the preferred way for
745 * callers to check for caps they want.)
747 int __ceph_caps_issued_mask(struct ceph_inode_info
*ci
, int mask
, int touch
)
749 struct ceph_cap
*cap
;
751 int have
= ci
->i_snap_caps
;
753 if ((have
& mask
) == mask
) {
754 dout("__ceph_caps_issued_mask %p snap issued %s"
755 " (mask %s)\n", &ci
->vfs_inode
,
756 ceph_cap_string(have
),
757 ceph_cap_string(mask
));
761 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
762 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
763 if (!__cap_is_valid(cap
))
765 if ((cap
->issued
& mask
) == mask
) {
766 dout("__ceph_caps_issued_mask %p cap %p issued %s"
767 " (mask %s)\n", &ci
->vfs_inode
, cap
,
768 ceph_cap_string(cap
->issued
),
769 ceph_cap_string(mask
));
775 /* does a combination of caps satisfy mask? */
777 if ((have
& mask
) == mask
) {
778 dout("__ceph_caps_issued_mask %p combo issued %s"
779 " (mask %s)\n", &ci
->vfs_inode
,
780 ceph_cap_string(cap
->issued
),
781 ceph_cap_string(mask
));
785 /* touch this + preceding caps */
787 for (q
= rb_first(&ci
->i_caps
); q
!= p
;
789 cap
= rb_entry(q
, struct ceph_cap
,
791 if (!__cap_is_valid(cap
))
804 * Return true if mask caps are currently being revoked by an MDS.
806 int __ceph_caps_revoking_other(struct ceph_inode_info
*ci
,
807 struct ceph_cap
*ocap
, int mask
)
809 struct ceph_cap
*cap
;
812 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
813 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
815 (cap
->implemented
& ~cap
->issued
& mask
))
821 int ceph_caps_revoking(struct ceph_inode_info
*ci
, int mask
)
823 struct inode
*inode
= &ci
->vfs_inode
;
826 spin_lock(&ci
->i_ceph_lock
);
827 ret
= __ceph_caps_revoking_other(ci
, NULL
, mask
);
828 spin_unlock(&ci
->i_ceph_lock
);
829 dout("ceph_caps_revoking %p %s = %d\n", inode
,
830 ceph_cap_string(mask
), ret
);
834 int __ceph_caps_used(struct ceph_inode_info
*ci
)
838 used
|= CEPH_CAP_PIN
;
840 used
|= CEPH_CAP_FILE_RD
;
841 if (ci
->i_rdcache_ref
||
842 (!S_ISDIR(ci
->vfs_inode
.i_mode
) && /* ignore readdir cache */
843 ci
->vfs_inode
.i_data
.nrpages
))
844 used
|= CEPH_CAP_FILE_CACHE
;
846 used
|= CEPH_CAP_FILE_WR
;
847 if (ci
->i_wb_ref
|| ci
->i_wrbuffer_ref
)
848 used
|= CEPH_CAP_FILE_BUFFER
;
853 * wanted, by virtue of open file modes
855 int __ceph_caps_file_wanted(struct ceph_inode_info
*ci
)
858 for (i
= 0; i
< CEPH_FILE_MODE_BITS
; i
++) {
859 if (ci
->i_nr_by_mode
[i
])
864 return ceph_caps_for_mode(bits
>> 1);
868 * Return caps we have registered with the MDS(s) as 'wanted'.
870 int __ceph_caps_mds_wanted(struct ceph_inode_info
*ci
, 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};
1019 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1020 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1021 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(arg
->op
),
1022 arg
->cid
, arg
->ino
, ceph_cap_string(arg
->caps
),
1023 ceph_cap_string(arg
->wanted
), ceph_cap_string(arg
->dirty
),
1024 arg
->seq
, arg
->issue_seq
, arg
->flush_tid
, arg
->oldest_flush_tid
,
1025 arg
->mseq
, arg
->follows
, arg
->size
, arg
->max_size
,
1027 arg
->xattr_buf
? (int)arg
->xattr_buf
->vec
.iov_len
: 0);
1029 /* flock buffer size + inline version + inline data size +
1030 * osd_epoch_barrier + oldest_flush_tid */
1031 extra_len
= 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4;
1032 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPS
, sizeof(*fc
) + extra_len
,
1037 msg
->hdr
.version
= cpu_to_le16(10);
1038 msg
->hdr
.tid
= cpu_to_le64(arg
->flush_tid
);
1040 fc
= msg
->front
.iov_base
;
1041 memset(fc
, 0, sizeof(*fc
));
1043 fc
->cap_id
= cpu_to_le64(arg
->cid
);
1044 fc
->op
= cpu_to_le32(arg
->op
);
1045 fc
->seq
= cpu_to_le32(arg
->seq
);
1046 fc
->issue_seq
= cpu_to_le32(arg
->issue_seq
);
1047 fc
->migrate_seq
= cpu_to_le32(arg
->mseq
);
1048 fc
->caps
= cpu_to_le32(arg
->caps
);
1049 fc
->wanted
= cpu_to_le32(arg
->wanted
);
1050 fc
->dirty
= cpu_to_le32(arg
->dirty
);
1051 fc
->ino
= cpu_to_le64(arg
->ino
);
1052 fc
->snap_follows
= cpu_to_le64(arg
->follows
);
1054 fc
->size
= cpu_to_le64(arg
->size
);
1055 fc
->max_size
= cpu_to_le64(arg
->max_size
);
1056 ceph_encode_timespec(&fc
->mtime
, &arg
->mtime
);
1057 ceph_encode_timespec(&fc
->atime
, &arg
->atime
);
1058 ceph_encode_timespec(&fc
->ctime
, &arg
->ctime
);
1059 fc
->time_warp_seq
= cpu_to_le32(arg
->time_warp_seq
);
1061 fc
->uid
= cpu_to_le32(from_kuid(&init_user_ns
, arg
->uid
));
1062 fc
->gid
= cpu_to_le32(from_kgid(&init_user_ns
, arg
->gid
));
1063 fc
->mode
= cpu_to_le32(arg
->mode
);
1065 fc
->xattr_version
= cpu_to_le64(arg
->xattr_version
);
1066 if (arg
->xattr_buf
) {
1067 msg
->middle
= ceph_buffer_get(arg
->xattr_buf
);
1068 fc
->xattr_len
= cpu_to_le32(arg
->xattr_buf
->vec
.iov_len
);
1069 msg
->hdr
.middle_len
= cpu_to_le32(arg
->xattr_buf
->vec
.iov_len
);
1073 /* flock buffer size (version 2) */
1074 ceph_encode_32(&p
, 0);
1075 /* inline version (version 4) */
1076 ceph_encode_64(&p
, arg
->inline_data
? 0 : CEPH_INLINE_NONE
);
1077 /* inline data size */
1078 ceph_encode_32(&p
, 0);
1079 /* osd_epoch_barrier (version 5) */
1080 ceph_encode_32(&p
, 0);
1081 /* oldest_flush_tid (version 6) */
1082 ceph_encode_64(&p
, arg
->oldest_flush_tid
);
1085 * caller_uid/caller_gid (version 7)
1087 * Currently, we don't properly track which caller dirtied the caps
1088 * last, and force a flush of them when there is a conflict. For now,
1089 * just set this to 0:0, to emulate how the MDS has worked up to now.
1091 ceph_encode_32(&p
, 0);
1092 ceph_encode_32(&p
, 0);
1094 /* pool namespace (version 8) (mds always ignores this) */
1095 ceph_encode_32(&p
, 0);
1098 * btime and change_attr (version 9)
1100 * We just zero these out for now, as the MDS ignores them unless
1101 * the requisite feature flags are set (which we don't do yet).
1103 ceph_encode_timespec(p
, &zerotime
);
1104 p
+= sizeof(struct ceph_timespec
);
1105 ceph_encode_64(&p
, 0);
1107 /* Advisory flags (version 10) */
1108 ceph_encode_32(&p
, arg
->flags
);
1110 ceph_con_send(&arg
->session
->s_con
, msg
);
1115 * Queue cap releases when an inode is dropped from our cache. Since
1116 * inode is about to be destroyed, there is no need for i_ceph_lock.
1118 void ceph_queue_caps_release(struct inode
*inode
)
1120 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1123 p
= rb_first(&ci
->i_caps
);
1125 struct ceph_cap
*cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
1127 __ceph_remove_cap(cap
, true);
1132 * Send a cap msg on the given inode. Update our caps state, then
1133 * drop i_ceph_lock and send the message.
1135 * Make note of max_size reported/requested from mds, revoked caps
1136 * that have now been implemented.
1138 * Make half-hearted attempt ot to invalidate page cache if we are
1139 * dropping RDCACHE. Note that this will leave behind locked pages
1140 * that we'll then need to deal with elsewhere.
1142 * Return non-zero if delayed release, or we experienced an error
1143 * such that the caller should requeue + retry later.
1145 * called with i_ceph_lock, then drops it.
1146 * caller should hold snap_rwsem (read), s_mutex.
1148 static int __send_cap(struct ceph_mds_client
*mdsc
, struct ceph_cap
*cap
,
1149 int op
, bool sync
, int used
, int want
, int retain
,
1150 int flushing
, u64 flush_tid
, u64 oldest_flush_tid
)
1151 __releases(cap
->ci
->i_ceph_lock
)
1153 struct ceph_inode_info
*ci
= cap
->ci
;
1154 struct inode
*inode
= &ci
->vfs_inode
;
1155 struct cap_msg_args arg
;
1156 int held
, revoking
, dropping
;
1161 held
= cap
->issued
| cap
->implemented
;
1162 revoking
= cap
->implemented
& ~cap
->issued
;
1163 retain
&= ~revoking
;
1164 dropping
= cap
->issued
& ~retain
;
1166 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1167 inode
, cap
, cap
->session
,
1168 ceph_cap_string(held
), ceph_cap_string(held
& retain
),
1169 ceph_cap_string(revoking
));
1170 BUG_ON((retain
& CEPH_CAP_PIN
) == 0);
1172 arg
.session
= cap
->session
;
1174 /* don't release wanted unless we've waited a bit. */
1175 if ((ci
->i_ceph_flags
& CEPH_I_NODELAY
) == 0 &&
1176 time_before(jiffies
, ci
->i_hold_caps_min
)) {
1177 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1178 ceph_cap_string(cap
->issued
),
1179 ceph_cap_string(cap
->issued
& retain
),
1180 ceph_cap_string(cap
->mds_wanted
),
1181 ceph_cap_string(want
));
1182 want
|= cap
->mds_wanted
;
1183 retain
|= cap
->issued
;
1186 ci
->i_ceph_flags
&= ~(CEPH_I_NODELAY
| CEPH_I_FLUSH
);
1187 if (want
& ~cap
->mds_wanted
) {
1188 /* user space may open/close single file frequently.
1189 * This avoids droping mds_wanted immediately after
1190 * requesting new mds_wanted.
1192 __cap_set_timeouts(mdsc
, ci
);
1195 cap
->issued
&= retain
; /* drop bits we don't want */
1196 if (cap
->implemented
& ~cap
->issued
) {
1198 * Wake up any waiters on wanted -> needed transition.
1199 * This is due to the weird transition from buffered
1200 * to sync IO... we need to flush dirty pages _before_
1201 * allowing sync writes to avoid reordering.
1205 cap
->implemented
&= cap
->issued
| used
;
1206 cap
->mds_wanted
= want
;
1208 arg
.ino
= ceph_vino(inode
).ino
;
1209 arg
.cid
= cap
->cap_id
;
1210 arg
.follows
= flushing
? ci
->i_head_snapc
->seq
: 0;
1211 arg
.flush_tid
= flush_tid
;
1212 arg
.oldest_flush_tid
= oldest_flush_tid
;
1214 arg
.size
= inode
->i_size
;
1215 ci
->i_reported_size
= arg
.size
;
1216 arg
.max_size
= ci
->i_wanted_max_size
;
1217 ci
->i_requested_max_size
= arg
.max_size
;
1219 if (flushing
& CEPH_CAP_XATTR_EXCL
) {
1220 __ceph_build_xattrs_blob(ci
);
1221 arg
.xattr_version
= ci
->i_xattrs
.version
;
1222 arg
.xattr_buf
= ci
->i_xattrs
.blob
;
1224 arg
.xattr_buf
= NULL
;
1227 arg
.mtime
= inode
->i_mtime
;
1228 arg
.atime
= inode
->i_atime
;
1229 arg
.ctime
= inode
->i_ctime
;
1232 arg
.caps
= cap
->implemented
;
1234 arg
.dirty
= flushing
;
1237 arg
.issue_seq
= cap
->issue_seq
;
1238 arg
.mseq
= cap
->mseq
;
1239 arg
.time_warp_seq
= ci
->i_time_warp_seq
;
1241 arg
.uid
= inode
->i_uid
;
1242 arg
.gid
= inode
->i_gid
;
1243 arg
.mode
= inode
->i_mode
;
1245 arg
.inline_data
= ci
->i_inline_version
!= CEPH_INLINE_NONE
;
1248 arg
.flags
|= CEPH_CLIENT_CAPS_SYNC
;
1250 spin_unlock(&ci
->i_ceph_lock
);
1252 ret
= send_cap_msg(&arg
);
1254 dout("error sending cap msg, must requeue %p\n", inode
);
1259 wake_up_all(&ci
->i_cap_wq
);
1264 static inline int __send_flush_snap(struct inode
*inode
,
1265 struct ceph_mds_session
*session
,
1266 struct ceph_cap_snap
*capsnap
,
1267 u32 mseq
, u64 oldest_flush_tid
)
1269 struct cap_msg_args arg
;
1271 arg
.session
= session
;
1272 arg
.ino
= ceph_vino(inode
).ino
;
1274 arg
.follows
= capsnap
->follows
;
1275 arg
.flush_tid
= capsnap
->cap_flush
.tid
;
1276 arg
.oldest_flush_tid
= oldest_flush_tid
;
1278 arg
.size
= capsnap
->size
;
1280 arg
.xattr_version
= capsnap
->xattr_version
;
1281 arg
.xattr_buf
= capsnap
->xattr_blob
;
1283 arg
.atime
= capsnap
->atime
;
1284 arg
.mtime
= capsnap
->mtime
;
1285 arg
.ctime
= capsnap
->ctime
;
1287 arg
.op
= CEPH_CAP_OP_FLUSHSNAP
;
1288 arg
.caps
= capsnap
->issued
;
1290 arg
.dirty
= capsnap
->dirty
;
1295 arg
.time_warp_seq
= capsnap
->time_warp_seq
;
1297 arg
.uid
= capsnap
->uid
;
1298 arg
.gid
= capsnap
->gid
;
1299 arg
.mode
= capsnap
->mode
;
1301 arg
.inline_data
= capsnap
->inline_data
;
1304 return send_cap_msg(&arg
);
1308 * When a snapshot is taken, clients accumulate dirty metadata on
1309 * inodes with capabilities in ceph_cap_snaps to describe the file
1310 * state at the time the snapshot was taken. This must be flushed
1311 * asynchronously back to the MDS once sync writes complete and dirty
1312 * data is written out.
1314 * Called under i_ceph_lock. Takes s_mutex as needed.
1316 static void __ceph_flush_snaps(struct ceph_inode_info
*ci
,
1317 struct ceph_mds_session
*session
)
1318 __releases(ci
->i_ceph_lock
)
1319 __acquires(ci
->i_ceph_lock
)
1321 struct inode
*inode
= &ci
->vfs_inode
;
1322 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
1323 struct ceph_cap_snap
*capsnap
;
1324 u64 oldest_flush_tid
= 0;
1325 u64 first_tid
= 1, last_tid
= 0;
1327 dout("__flush_snaps %p session %p\n", inode
, session
);
1329 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
1331 * we need to wait for sync writes to complete and for dirty
1332 * pages to be written out.
1334 if (capsnap
->dirty_pages
|| capsnap
->writing
)
1337 /* should be removed by ceph_try_drop_cap_snap() */
1338 BUG_ON(!capsnap
->need_flush
);
1340 /* only flush each capsnap once */
1341 if (capsnap
->cap_flush
.tid
> 0) {
1342 dout(" already flushed %p, skipping\n", capsnap
);
1346 spin_lock(&mdsc
->cap_dirty_lock
);
1347 capsnap
->cap_flush
.tid
= ++mdsc
->last_cap_flush_tid
;
1348 list_add_tail(&capsnap
->cap_flush
.g_list
,
1349 &mdsc
->cap_flush_list
);
1350 if (oldest_flush_tid
== 0)
1351 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1352 if (list_empty(&ci
->i_flushing_item
)) {
1353 list_add_tail(&ci
->i_flushing_item
,
1354 &session
->s_cap_flushing
);
1356 spin_unlock(&mdsc
->cap_dirty_lock
);
1358 list_add_tail(&capsnap
->cap_flush
.i_list
,
1359 &ci
->i_cap_flush_list
);
1362 first_tid
= capsnap
->cap_flush
.tid
;
1363 last_tid
= capsnap
->cap_flush
.tid
;
1366 ci
->i_ceph_flags
&= ~CEPH_I_FLUSH_SNAPS
;
1368 while (first_tid
<= last_tid
) {
1369 struct ceph_cap
*cap
= ci
->i_auth_cap
;
1370 struct ceph_cap_flush
*cf
;
1373 if (!(cap
&& cap
->session
== session
)) {
1374 dout("__flush_snaps %p auth cap %p not mds%d, "
1375 "stop\n", inode
, cap
, session
->s_mds
);
1380 list_for_each_entry(cf
, &ci
->i_cap_flush_list
, i_list
) {
1381 if (cf
->tid
>= first_tid
) {
1389 first_tid
= cf
->tid
+ 1;
1391 capsnap
= container_of(cf
, struct ceph_cap_snap
, cap_flush
);
1392 atomic_inc(&capsnap
->nref
);
1393 spin_unlock(&ci
->i_ceph_lock
);
1395 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1396 inode
, capsnap
, cf
->tid
, ceph_cap_string(capsnap
->dirty
));
1398 ret
= __send_flush_snap(inode
, session
, capsnap
, cap
->mseq
,
1401 pr_err("__flush_snaps: error sending cap flushsnap, "
1402 "ino (%llx.%llx) tid %llu follows %llu\n",
1403 ceph_vinop(inode
), cf
->tid
, capsnap
->follows
);
1406 ceph_put_cap_snap(capsnap
);
1407 spin_lock(&ci
->i_ceph_lock
);
1411 void ceph_flush_snaps(struct ceph_inode_info
*ci
,
1412 struct ceph_mds_session
**psession
)
1414 struct inode
*inode
= &ci
->vfs_inode
;
1415 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
1416 struct ceph_mds_session
*session
= NULL
;
1419 dout("ceph_flush_snaps %p\n", inode
);
1421 session
= *psession
;
1423 spin_lock(&ci
->i_ceph_lock
);
1424 if (!(ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
)) {
1425 dout(" no capsnap needs flush, doing nothing\n");
1428 if (!ci
->i_auth_cap
) {
1429 dout(" no auth cap (migrating?), doing nothing\n");
1433 mds
= ci
->i_auth_cap
->session
->s_mds
;
1434 if (session
&& session
->s_mds
!= mds
) {
1435 dout(" oops, wrong session %p mutex\n", session
);
1436 mutex_unlock(&session
->s_mutex
);
1437 ceph_put_mds_session(session
);
1441 spin_unlock(&ci
->i_ceph_lock
);
1442 mutex_lock(&mdsc
->mutex
);
1443 session
= __ceph_lookup_mds_session(mdsc
, mds
);
1444 mutex_unlock(&mdsc
->mutex
);
1446 dout(" inverting session/ino locks on %p\n", session
);
1447 mutex_lock(&session
->s_mutex
);
1452 __ceph_flush_snaps(ci
, session
);
1454 spin_unlock(&ci
->i_ceph_lock
);
1457 *psession
= session
;
1459 mutex_unlock(&session
->s_mutex
);
1460 ceph_put_mds_session(session
);
1462 /* we flushed them all; remove this inode from the queue */
1463 spin_lock(&mdsc
->snap_flush_lock
);
1464 list_del_init(&ci
->i_snap_flush_item
);
1465 spin_unlock(&mdsc
->snap_flush_lock
);
1469 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1470 * Caller is then responsible for calling __mark_inode_dirty with the
1471 * returned flags value.
1473 int __ceph_mark_dirty_caps(struct ceph_inode_info
*ci
, int mask
,
1474 struct ceph_cap_flush
**pcf
)
1476 struct ceph_mds_client
*mdsc
=
1477 ceph_sb_to_client(ci
->vfs_inode
.i_sb
)->mdsc
;
1478 struct inode
*inode
= &ci
->vfs_inode
;
1479 int was
= ci
->i_dirty_caps
;
1482 if (!ci
->i_auth_cap
) {
1483 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1484 "but no auth cap (session was closed?)\n",
1485 inode
, ceph_ino(inode
), ceph_cap_string(mask
));
1489 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci
->vfs_inode
,
1490 ceph_cap_string(mask
), ceph_cap_string(was
),
1491 ceph_cap_string(was
| mask
));
1492 ci
->i_dirty_caps
|= mask
;
1494 WARN_ON_ONCE(ci
->i_prealloc_cap_flush
);
1495 swap(ci
->i_prealloc_cap_flush
, *pcf
);
1497 if (!ci
->i_head_snapc
) {
1498 WARN_ON_ONCE(!rwsem_is_locked(&mdsc
->snap_rwsem
));
1499 ci
->i_head_snapc
= ceph_get_snap_context(
1500 ci
->i_snap_realm
->cached_context
);
1502 dout(" inode %p now dirty snapc %p auth cap %p\n",
1503 &ci
->vfs_inode
, ci
->i_head_snapc
, ci
->i_auth_cap
);
1504 BUG_ON(!list_empty(&ci
->i_dirty_item
));
1505 spin_lock(&mdsc
->cap_dirty_lock
);
1506 list_add(&ci
->i_dirty_item
, &mdsc
->cap_dirty
);
1507 spin_unlock(&mdsc
->cap_dirty_lock
);
1508 if (ci
->i_flushing_caps
== 0) {
1510 dirty
|= I_DIRTY_SYNC
;
1513 WARN_ON_ONCE(!ci
->i_prealloc_cap_flush
);
1515 BUG_ON(list_empty(&ci
->i_dirty_item
));
1516 if (((was
| ci
->i_flushing_caps
) & CEPH_CAP_FILE_BUFFER
) &&
1517 (mask
& CEPH_CAP_FILE_BUFFER
))
1518 dirty
|= I_DIRTY_DATASYNC
;
1519 __cap_delay_requeue(mdsc
, ci
);
1523 struct ceph_cap_flush
*ceph_alloc_cap_flush(void)
1525 return kmem_cache_alloc(ceph_cap_flush_cachep
, GFP_KERNEL
);
1528 void ceph_free_cap_flush(struct ceph_cap_flush
*cf
)
1531 kmem_cache_free(ceph_cap_flush_cachep
, cf
);
1534 static u64
__get_oldest_flush_tid(struct ceph_mds_client
*mdsc
)
1536 if (!list_empty(&mdsc
->cap_flush_list
)) {
1537 struct ceph_cap_flush
*cf
=
1538 list_first_entry(&mdsc
->cap_flush_list
,
1539 struct ceph_cap_flush
, g_list
);
1546 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1547 * Return true if caller needs to wake up flush waiters.
1549 static bool __finish_cap_flush(struct ceph_mds_client
*mdsc
,
1550 struct ceph_inode_info
*ci
,
1551 struct ceph_cap_flush
*cf
)
1553 struct ceph_cap_flush
*prev
;
1554 bool wake
= cf
->wake
;
1556 /* are there older pending cap flushes? */
1557 if (wake
&& cf
->g_list
.prev
!= &mdsc
->cap_flush_list
) {
1558 prev
= list_prev_entry(cf
, g_list
);
1562 list_del(&cf
->g_list
);
1564 if (wake
&& cf
->i_list
.prev
!= &ci
->i_cap_flush_list
) {
1565 prev
= list_prev_entry(cf
, i_list
);
1569 list_del(&cf
->i_list
);
1577 * Add dirty inode to the flushing list. Assigned a seq number so we
1578 * can wait for caps to flush without starving.
1580 * Called under i_ceph_lock.
1582 static int __mark_caps_flushing(struct inode
*inode
,
1583 struct ceph_mds_session
*session
, bool wake
,
1584 u64
*flush_tid
, u64
*oldest_flush_tid
)
1586 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
1587 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1588 struct ceph_cap_flush
*cf
= NULL
;
1591 BUG_ON(ci
->i_dirty_caps
== 0);
1592 BUG_ON(list_empty(&ci
->i_dirty_item
));
1593 BUG_ON(!ci
->i_prealloc_cap_flush
);
1595 flushing
= ci
->i_dirty_caps
;
1596 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1597 ceph_cap_string(flushing
),
1598 ceph_cap_string(ci
->i_flushing_caps
),
1599 ceph_cap_string(ci
->i_flushing_caps
| flushing
));
1600 ci
->i_flushing_caps
|= flushing
;
1601 ci
->i_dirty_caps
= 0;
1602 dout(" inode %p now !dirty\n", inode
);
1604 swap(cf
, ci
->i_prealloc_cap_flush
);
1605 cf
->caps
= flushing
;
1608 spin_lock(&mdsc
->cap_dirty_lock
);
1609 list_del_init(&ci
->i_dirty_item
);
1611 cf
->tid
= ++mdsc
->last_cap_flush_tid
;
1612 list_add_tail(&cf
->g_list
, &mdsc
->cap_flush_list
);
1613 *oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1615 if (list_empty(&ci
->i_flushing_item
)) {
1616 list_add_tail(&ci
->i_flushing_item
, &session
->s_cap_flushing
);
1617 mdsc
->num_cap_flushing
++;
1619 spin_unlock(&mdsc
->cap_dirty_lock
);
1621 list_add_tail(&cf
->i_list
, &ci
->i_cap_flush_list
);
1623 *flush_tid
= cf
->tid
;
1628 * try to invalidate mapping pages without blocking.
1630 static int try_nonblocking_invalidate(struct inode
*inode
)
1632 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1633 u32 invalidating_gen
= ci
->i_rdcache_gen
;
1635 spin_unlock(&ci
->i_ceph_lock
);
1636 invalidate_mapping_pages(&inode
->i_data
, 0, -1);
1637 spin_lock(&ci
->i_ceph_lock
);
1639 if (inode
->i_data
.nrpages
== 0 &&
1640 invalidating_gen
== ci
->i_rdcache_gen
) {
1642 dout("try_nonblocking_invalidate %p success\n", inode
);
1643 /* save any racing async invalidate some trouble */
1644 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
- 1;
1647 dout("try_nonblocking_invalidate %p failed\n", inode
);
1652 * Swiss army knife function to examine currently used and wanted
1653 * versus held caps. Release, flush, ack revoked caps to mds as
1656 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1657 * cap release further.
1658 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1659 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1662 void ceph_check_caps(struct ceph_inode_info
*ci
, int flags
,
1663 struct ceph_mds_session
*session
)
1665 struct ceph_fs_client
*fsc
= ceph_inode_to_client(&ci
->vfs_inode
);
1666 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
1667 struct inode
*inode
= &ci
->vfs_inode
;
1668 struct ceph_cap
*cap
;
1669 u64 flush_tid
, oldest_flush_tid
;
1670 int file_wanted
, used
, cap_used
;
1671 int took_snap_rwsem
= 0; /* true if mdsc->snap_rwsem held */
1672 int issued
, implemented
, want
, retain
, revoking
, flushing
= 0;
1673 int mds
= -1; /* keep track of how far we've gone through i_caps list
1674 to avoid an infinite loop on retry */
1676 int delayed
= 0, sent
= 0, num
;
1677 bool is_delayed
= flags
& CHECK_CAPS_NODELAY
;
1678 bool queue_invalidate
= false;
1679 bool force_requeue
= false;
1680 bool tried_invalidate
= false;
1682 /* if we are unmounting, flush any unused caps immediately. */
1686 spin_lock(&ci
->i_ceph_lock
);
1688 if (ci
->i_ceph_flags
& CEPH_I_FLUSH
)
1689 flags
|= CHECK_CAPS_FLUSH
;
1693 spin_lock(&ci
->i_ceph_lock
);
1695 file_wanted
= __ceph_caps_file_wanted(ci
);
1696 used
= __ceph_caps_used(ci
);
1697 issued
= __ceph_caps_issued(ci
, &implemented
);
1698 revoking
= implemented
& ~issued
;
1701 retain
= file_wanted
| used
| CEPH_CAP_PIN
;
1702 if (!mdsc
->stopping
&& inode
->i_nlink
> 0) {
1704 retain
|= CEPH_CAP_ANY
; /* be greedy */
1705 } else if (S_ISDIR(inode
->i_mode
) &&
1706 (issued
& CEPH_CAP_FILE_SHARED
) &&
1707 __ceph_dir_is_complete(ci
)) {
1709 * If a directory is complete, we want to keep
1710 * the exclusive cap. So that MDS does not end up
1711 * revoking the shared cap on every create/unlink
1714 want
= CEPH_CAP_ANY_SHARED
| CEPH_CAP_FILE_EXCL
;
1718 retain
|= CEPH_CAP_ANY_SHARED
;
1720 * keep RD only if we didn't have the file open RW,
1721 * because then the mds would revoke it anyway to
1722 * journal max_size=0.
1724 if (ci
->i_max_size
== 0)
1725 retain
|= CEPH_CAP_ANY_RD
;
1729 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1730 " issued %s revoking %s retain %s %s%s%s\n", inode
,
1731 ceph_cap_string(file_wanted
),
1732 ceph_cap_string(used
), ceph_cap_string(ci
->i_dirty_caps
),
1733 ceph_cap_string(ci
->i_flushing_caps
),
1734 ceph_cap_string(issued
), ceph_cap_string(revoking
),
1735 ceph_cap_string(retain
),
1736 (flags
& CHECK_CAPS_AUTHONLY
) ? " AUTHONLY" : "",
1737 (flags
& CHECK_CAPS_NODELAY
) ? " NODELAY" : "",
1738 (flags
& CHECK_CAPS_FLUSH
) ? " FLUSH" : "");
1741 * If we no longer need to hold onto old our caps, and we may
1742 * have cached pages, but don't want them, then try to invalidate.
1743 * If we fail, it's because pages are locked.... try again later.
1745 if ((!is_delayed
|| mdsc
->stopping
) &&
1746 !S_ISDIR(inode
->i_mode
) && /* ignore readdir cache */
1747 !(ci
->i_wb_ref
|| ci
->i_wrbuffer_ref
) && /* no dirty pages... */
1748 inode
->i_data
.nrpages
&& /* have cached pages */
1749 (revoking
& (CEPH_CAP_FILE_CACHE
|
1750 CEPH_CAP_FILE_LAZYIO
)) && /* or revoking cache */
1751 !tried_invalidate
) {
1752 dout("check_caps trying to invalidate on %p\n", inode
);
1753 if (try_nonblocking_invalidate(inode
) < 0) {
1754 if (revoking
& (CEPH_CAP_FILE_CACHE
|
1755 CEPH_CAP_FILE_LAZYIO
)) {
1756 dout("check_caps queuing invalidate\n");
1757 queue_invalidate
= true;
1758 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
;
1760 dout("check_caps failed to invalidate pages\n");
1761 /* we failed to invalidate pages. check these
1762 caps again later. */
1763 force_requeue
= true;
1764 __cap_set_timeouts(mdsc
, ci
);
1767 tried_invalidate
= true;
1772 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
1773 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
1776 /* avoid looping forever */
1777 if (mds
>= cap
->mds
||
1778 ((flags
& CHECK_CAPS_AUTHONLY
) && cap
!= ci
->i_auth_cap
))
1781 /* NOTE: no side-effects allowed, until we take s_mutex */
1784 if (ci
->i_auth_cap
&& cap
!= ci
->i_auth_cap
)
1785 cap_used
&= ~ci
->i_auth_cap
->issued
;
1787 revoking
= cap
->implemented
& ~cap
->issued
;
1788 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1789 cap
->mds
, cap
, ceph_cap_string(cap_used
),
1790 ceph_cap_string(cap
->issued
),
1791 ceph_cap_string(cap
->implemented
),
1792 ceph_cap_string(revoking
));
1794 if (cap
== ci
->i_auth_cap
&&
1795 (cap
->issued
& CEPH_CAP_FILE_WR
)) {
1796 /* request larger max_size from MDS? */
1797 if (ci
->i_wanted_max_size
> ci
->i_max_size
&&
1798 ci
->i_wanted_max_size
> ci
->i_requested_max_size
) {
1799 dout("requesting new max_size\n");
1803 /* approaching file_max? */
1804 if ((inode
->i_size
<< 1) >= ci
->i_max_size
&&
1805 (ci
->i_reported_size
<< 1) < ci
->i_max_size
) {
1806 dout("i_size approaching max_size\n");
1810 /* flush anything dirty? */
1811 if (cap
== ci
->i_auth_cap
) {
1812 if ((flags
& CHECK_CAPS_FLUSH
) && ci
->i_dirty_caps
) {
1813 dout("flushing dirty caps\n");
1816 if (ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
) {
1817 dout("flushing snap caps\n");
1822 /* completed revocation? going down and there are no caps? */
1823 if (revoking
&& (revoking
& cap_used
) == 0) {
1824 dout("completed revocation of %s\n",
1825 ceph_cap_string(cap
->implemented
& ~cap
->issued
));
1829 /* want more caps from mds? */
1830 if (want
& ~(cap
->mds_wanted
| cap
->issued
))
1833 /* things we might delay */
1834 if ((cap
->issued
& ~retain
) == 0 &&
1835 cap
->mds_wanted
== want
)
1836 continue; /* nope, all good */
1842 if ((ci
->i_ceph_flags
& CEPH_I_NODELAY
) == 0 &&
1843 time_before(jiffies
, ci
->i_hold_caps_max
)) {
1844 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1845 ceph_cap_string(cap
->issued
),
1846 ceph_cap_string(cap
->issued
& retain
),
1847 ceph_cap_string(cap
->mds_wanted
),
1848 ceph_cap_string(want
));
1854 if (ci
->i_ceph_flags
& CEPH_I_NOFLUSH
) {
1855 dout(" skipping %p I_NOFLUSH set\n", inode
);
1859 if (session
&& session
!= cap
->session
) {
1860 dout("oops, wrong session %p mutex\n", session
);
1861 mutex_unlock(&session
->s_mutex
);
1865 session
= cap
->session
;
1866 if (mutex_trylock(&session
->s_mutex
) == 0) {
1867 dout("inverting session/ino locks on %p\n",
1869 spin_unlock(&ci
->i_ceph_lock
);
1870 if (took_snap_rwsem
) {
1871 up_read(&mdsc
->snap_rwsem
);
1872 took_snap_rwsem
= 0;
1874 mutex_lock(&session
->s_mutex
);
1879 /* kick flushing and flush snaps before sending normal
1881 if (cap
== ci
->i_auth_cap
&&
1883 (CEPH_I_KICK_FLUSH
| CEPH_I_FLUSH_SNAPS
))) {
1884 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
) {
1885 spin_lock(&mdsc
->cap_dirty_lock
);
1886 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1887 spin_unlock(&mdsc
->cap_dirty_lock
);
1888 __kick_flushing_caps(mdsc
, session
, ci
,
1890 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
1892 if (ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
)
1893 __ceph_flush_snaps(ci
, session
);
1898 /* take snap_rwsem after session mutex */
1899 if (!took_snap_rwsem
) {
1900 if (down_read_trylock(&mdsc
->snap_rwsem
) == 0) {
1901 dout("inverting snap/in locks on %p\n",
1903 spin_unlock(&ci
->i_ceph_lock
);
1904 down_read(&mdsc
->snap_rwsem
);
1905 took_snap_rwsem
= 1;
1908 took_snap_rwsem
= 1;
1911 if (cap
== ci
->i_auth_cap
&& ci
->i_dirty_caps
) {
1912 flushing
= __mark_caps_flushing(inode
, session
, false,
1918 spin_lock(&mdsc
->cap_dirty_lock
);
1919 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1920 spin_unlock(&mdsc
->cap_dirty_lock
);
1923 mds
= cap
->mds
; /* remember mds, so we don't repeat */
1926 /* __send_cap drops i_ceph_lock */
1927 delayed
+= __send_cap(mdsc
, cap
, CEPH_CAP_OP_UPDATE
, false,
1928 cap_used
, want
, retain
, flushing
,
1929 flush_tid
, oldest_flush_tid
);
1930 goto retry
; /* retake i_ceph_lock and restart our cap scan. */
1934 * Reschedule delayed caps release if we delayed anything,
1937 if (delayed
&& is_delayed
)
1938 force_requeue
= true; /* __send_cap delayed release; requeue */
1939 if (!delayed
&& !is_delayed
)
1940 __cap_delay_cancel(mdsc
, ci
);
1941 else if (!is_delayed
|| force_requeue
)
1942 __cap_delay_requeue(mdsc
, ci
);
1944 spin_unlock(&ci
->i_ceph_lock
);
1946 if (queue_invalidate
)
1947 ceph_queue_invalidate(inode
);
1950 mutex_unlock(&session
->s_mutex
);
1951 if (took_snap_rwsem
)
1952 up_read(&mdsc
->snap_rwsem
);
1956 * Try to flush dirty caps back to the auth mds.
1958 static int try_flush_caps(struct inode
*inode
, u64
*ptid
)
1960 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
1961 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1962 struct ceph_mds_session
*session
= NULL
;
1964 u64 flush_tid
= 0, oldest_flush_tid
= 0;
1967 spin_lock(&ci
->i_ceph_lock
);
1968 if (ci
->i_ceph_flags
& CEPH_I_NOFLUSH
) {
1969 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode
);
1972 if (ci
->i_dirty_caps
&& ci
->i_auth_cap
) {
1973 struct ceph_cap
*cap
= ci
->i_auth_cap
;
1974 int used
= __ceph_caps_used(ci
);
1975 int want
= __ceph_caps_wanted(ci
);
1978 if (!session
|| session
!= cap
->session
) {
1979 spin_unlock(&ci
->i_ceph_lock
);
1981 mutex_unlock(&session
->s_mutex
);
1982 session
= cap
->session
;
1983 mutex_lock(&session
->s_mutex
);
1986 if (cap
->session
->s_state
< CEPH_MDS_SESSION_OPEN
)
1989 flushing
= __mark_caps_flushing(inode
, session
, true,
1990 &flush_tid
, &oldest_flush_tid
);
1992 /* __send_cap drops i_ceph_lock */
1993 delayed
= __send_cap(mdsc
, cap
, CEPH_CAP_OP_FLUSH
, true,
1994 used
, want
, (cap
->issued
| cap
->implemented
),
1995 flushing
, flush_tid
, oldest_flush_tid
);
1998 spin_lock(&ci
->i_ceph_lock
);
1999 __cap_delay_requeue(mdsc
, ci
);
2000 spin_unlock(&ci
->i_ceph_lock
);
2003 if (!list_empty(&ci
->i_cap_flush_list
)) {
2004 struct ceph_cap_flush
*cf
=
2005 list_last_entry(&ci
->i_cap_flush_list
,
2006 struct ceph_cap_flush
, i_list
);
2008 flush_tid
= cf
->tid
;
2010 flushing
= ci
->i_flushing_caps
;
2011 spin_unlock(&ci
->i_ceph_lock
);
2015 mutex_unlock(&session
->s_mutex
);
2022 * Return true if we've flushed caps through the given flush_tid.
2024 static int caps_are_flushed(struct inode
*inode
, u64 flush_tid
)
2026 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2029 spin_lock(&ci
->i_ceph_lock
);
2030 if (!list_empty(&ci
->i_cap_flush_list
)) {
2031 struct ceph_cap_flush
* cf
=
2032 list_first_entry(&ci
->i_cap_flush_list
,
2033 struct ceph_cap_flush
, i_list
);
2034 if (cf
->tid
<= flush_tid
)
2037 spin_unlock(&ci
->i_ceph_lock
);
2042 * wait for any unsafe requests to complete.
2044 static int unsafe_request_wait(struct inode
*inode
)
2046 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2047 struct ceph_mds_request
*req1
= NULL
, *req2
= NULL
;
2050 spin_lock(&ci
->i_unsafe_lock
);
2051 if (S_ISDIR(inode
->i_mode
) && !list_empty(&ci
->i_unsafe_dirops
)) {
2052 req1
= list_last_entry(&ci
->i_unsafe_dirops
,
2053 struct ceph_mds_request
,
2055 ceph_mdsc_get_request(req1
);
2057 if (!list_empty(&ci
->i_unsafe_iops
)) {
2058 req2
= list_last_entry(&ci
->i_unsafe_iops
,
2059 struct ceph_mds_request
,
2060 r_unsafe_target_item
);
2061 ceph_mdsc_get_request(req2
);
2063 spin_unlock(&ci
->i_unsafe_lock
);
2065 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2066 inode
, req1
? req1
->r_tid
: 0ULL, req2
? req2
->r_tid
: 0ULL);
2068 ret
= !wait_for_completion_timeout(&req1
->r_safe_completion
,
2069 ceph_timeout_jiffies(req1
->r_timeout
));
2072 ceph_mdsc_put_request(req1
);
2075 ret
= !wait_for_completion_timeout(&req2
->r_safe_completion
,
2076 ceph_timeout_jiffies(req2
->r_timeout
));
2079 ceph_mdsc_put_request(req2
);
2084 int ceph_fsync(struct file
*file
, loff_t start
, loff_t end
, int datasync
)
2086 struct inode
*inode
= file
->f_mapping
->host
;
2087 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2092 dout("fsync %p%s\n", inode
, datasync
? " datasync" : "");
2094 ret
= filemap_write_and_wait_range(inode
->i_mapping
, start
, end
);
2103 dirty
= try_flush_caps(inode
, &flush_tid
);
2104 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty
));
2106 ret
= unsafe_request_wait(inode
);
2109 * only wait on non-file metadata writeback (the mds
2110 * can recover size and mtime, so we don't need to
2113 if (!ret
&& (dirty
& ~CEPH_CAP_ANY_FILE_WR
)) {
2114 ret
= wait_event_interruptible(ci
->i_cap_wq
,
2115 caps_are_flushed(inode
, flush_tid
));
2117 inode_unlock(inode
);
2119 dout("fsync %p%s result=%d\n", inode
, datasync
? " datasync" : "", ret
);
2124 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2125 * queue inode for flush but don't do so immediately, because we can
2126 * get by with fewer MDS messages if we wait for data writeback to
2129 int ceph_write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
2131 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2135 int wait
= wbc
->sync_mode
== WB_SYNC_ALL
;
2137 dout("write_inode %p wait=%d\n", inode
, wait
);
2139 dirty
= try_flush_caps(inode
, &flush_tid
);
2141 err
= wait_event_interruptible(ci
->i_cap_wq
,
2142 caps_are_flushed(inode
, flush_tid
));
2144 struct ceph_mds_client
*mdsc
=
2145 ceph_sb_to_client(inode
->i_sb
)->mdsc
;
2147 spin_lock(&ci
->i_ceph_lock
);
2148 if (__ceph_caps_dirty(ci
))
2149 __cap_delay_requeue_front(mdsc
, ci
);
2150 spin_unlock(&ci
->i_ceph_lock
);
2155 static void __kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2156 struct ceph_mds_session
*session
,
2157 struct ceph_inode_info
*ci
,
2158 u64 oldest_flush_tid
)
2159 __releases(ci
->i_ceph_lock
)
2160 __acquires(ci
->i_ceph_lock
)
2162 struct inode
*inode
= &ci
->vfs_inode
;
2163 struct ceph_cap
*cap
;
2164 struct ceph_cap_flush
*cf
;
2168 list_for_each_entry(cf
, &ci
->i_cap_flush_list
, i_list
) {
2169 if (cf
->tid
< first_tid
)
2172 cap
= ci
->i_auth_cap
;
2173 if (!(cap
&& cap
->session
== session
)) {
2174 pr_err("%p auth cap %p not mds%d ???\n",
2175 inode
, cap
, session
->s_mds
);
2179 first_tid
= cf
->tid
+ 1;
2182 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2183 inode
, cap
, cf
->tid
, ceph_cap_string(cf
->caps
));
2184 ci
->i_ceph_flags
|= CEPH_I_NODELAY
;
2185 ret
= __send_cap(mdsc
, cap
, CEPH_CAP_OP_FLUSH
,
2186 false, __ceph_caps_used(ci
),
2187 __ceph_caps_wanted(ci
),
2188 cap
->issued
| cap
->implemented
,
2189 cf
->caps
, cf
->tid
, oldest_flush_tid
);
2191 pr_err("kick_flushing_caps: error sending "
2192 "cap flush, ino (%llx.%llx) "
2193 "tid %llu flushing %s\n",
2194 ceph_vinop(inode
), cf
->tid
,
2195 ceph_cap_string(cf
->caps
));
2198 struct ceph_cap_snap
*capsnap
=
2199 container_of(cf
, struct ceph_cap_snap
,
2201 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2202 inode
, capsnap
, cf
->tid
,
2203 ceph_cap_string(capsnap
->dirty
));
2205 atomic_inc(&capsnap
->nref
);
2206 spin_unlock(&ci
->i_ceph_lock
);
2208 ret
= __send_flush_snap(inode
, session
, capsnap
, cap
->mseq
,
2211 pr_err("kick_flushing_caps: error sending "
2212 "cap flushsnap, ino (%llx.%llx) "
2213 "tid %llu follows %llu\n",
2214 ceph_vinop(inode
), cf
->tid
,
2218 ceph_put_cap_snap(capsnap
);
2221 spin_lock(&ci
->i_ceph_lock
);
2225 void ceph_early_kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2226 struct ceph_mds_session
*session
)
2228 struct ceph_inode_info
*ci
;
2229 struct ceph_cap
*cap
;
2230 u64 oldest_flush_tid
;
2232 dout("early_kick_flushing_caps mds%d\n", session
->s_mds
);
2234 spin_lock(&mdsc
->cap_dirty_lock
);
2235 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2236 spin_unlock(&mdsc
->cap_dirty_lock
);
2238 list_for_each_entry(ci
, &session
->s_cap_flushing
, i_flushing_item
) {
2239 spin_lock(&ci
->i_ceph_lock
);
2240 cap
= ci
->i_auth_cap
;
2241 if (!(cap
&& cap
->session
== session
)) {
2242 pr_err("%p auth cap %p not mds%d ???\n",
2243 &ci
->vfs_inode
, cap
, session
->s_mds
);
2244 spin_unlock(&ci
->i_ceph_lock
);
2250 * if flushing caps were revoked, we re-send the cap flush
2251 * in client reconnect stage. This guarantees MDS * processes
2252 * the cap flush message before issuing the flushing caps to
2255 if ((cap
->issued
& ci
->i_flushing_caps
) !=
2256 ci
->i_flushing_caps
) {
2257 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
2258 __kick_flushing_caps(mdsc
, session
, ci
,
2261 ci
->i_ceph_flags
|= CEPH_I_KICK_FLUSH
;
2264 spin_unlock(&ci
->i_ceph_lock
);
2268 void ceph_kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2269 struct ceph_mds_session
*session
)
2271 struct ceph_inode_info
*ci
;
2272 struct ceph_cap
*cap
;
2273 u64 oldest_flush_tid
;
2275 dout("kick_flushing_caps mds%d\n", session
->s_mds
);
2277 spin_lock(&mdsc
->cap_dirty_lock
);
2278 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2279 spin_unlock(&mdsc
->cap_dirty_lock
);
2281 list_for_each_entry(ci
, &session
->s_cap_flushing
, i_flushing_item
) {
2282 spin_lock(&ci
->i_ceph_lock
);
2283 cap
= ci
->i_auth_cap
;
2284 if (!(cap
&& cap
->session
== session
)) {
2285 pr_err("%p auth cap %p not mds%d ???\n",
2286 &ci
->vfs_inode
, cap
, session
->s_mds
);
2287 spin_unlock(&ci
->i_ceph_lock
);
2290 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
) {
2291 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
2292 __kick_flushing_caps(mdsc
, session
, ci
,
2295 spin_unlock(&ci
->i_ceph_lock
);
2299 static void kick_flushing_inode_caps(struct ceph_mds_client
*mdsc
,
2300 struct ceph_mds_session
*session
,
2301 struct inode
*inode
)
2302 __releases(ci
->i_ceph_lock
)
2304 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2305 struct ceph_cap
*cap
;
2307 cap
= ci
->i_auth_cap
;
2308 dout("kick_flushing_inode_caps %p flushing %s\n", inode
,
2309 ceph_cap_string(ci
->i_flushing_caps
));
2311 if (!list_empty(&ci
->i_cap_flush_list
)) {
2312 u64 oldest_flush_tid
;
2313 spin_lock(&mdsc
->cap_dirty_lock
);
2314 list_move_tail(&ci
->i_flushing_item
,
2315 &cap
->session
->s_cap_flushing
);
2316 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2317 spin_unlock(&mdsc
->cap_dirty_lock
);
2319 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
2320 __kick_flushing_caps(mdsc
, session
, ci
, oldest_flush_tid
);
2321 spin_unlock(&ci
->i_ceph_lock
);
2323 spin_unlock(&ci
->i_ceph_lock
);
2329 * Take references to capabilities we hold, so that we don't release
2330 * them to the MDS prematurely.
2332 * Protected by i_ceph_lock.
2334 static void __take_cap_refs(struct ceph_inode_info
*ci
, int got
,
2335 bool snap_rwsem_locked
)
2337 if (got
& CEPH_CAP_PIN
)
2339 if (got
& CEPH_CAP_FILE_RD
)
2341 if (got
& CEPH_CAP_FILE_CACHE
)
2342 ci
->i_rdcache_ref
++;
2343 if (got
& CEPH_CAP_FILE_WR
) {
2344 if (ci
->i_wr_ref
== 0 && !ci
->i_head_snapc
) {
2345 BUG_ON(!snap_rwsem_locked
);
2346 ci
->i_head_snapc
= ceph_get_snap_context(
2347 ci
->i_snap_realm
->cached_context
);
2351 if (got
& CEPH_CAP_FILE_BUFFER
) {
2352 if (ci
->i_wb_ref
== 0)
2353 ihold(&ci
->vfs_inode
);
2355 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2356 &ci
->vfs_inode
, ci
->i_wb_ref
-1, ci
->i_wb_ref
);
2361 * Try to grab cap references. Specify those refs we @want, and the
2362 * minimal set we @need. Also include the larger offset we are writing
2363 * to (when applicable), and check against max_size here as well.
2364 * Note that caller is responsible for ensuring max_size increases are
2365 * requested from the MDS.
2367 static int try_get_cap_refs(struct ceph_inode_info
*ci
, int need
, int want
,
2368 loff_t endoff
, bool nonblock
, int *got
, int *err
)
2370 struct inode
*inode
= &ci
->vfs_inode
;
2371 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
2373 int have
, implemented
;
2375 bool snap_rwsem_locked
= false;
2377 dout("get_cap_refs %p need %s want %s\n", inode
,
2378 ceph_cap_string(need
), ceph_cap_string(want
));
2381 spin_lock(&ci
->i_ceph_lock
);
2383 /* make sure file is actually open */
2384 file_wanted
= __ceph_caps_file_wanted(ci
);
2385 if ((file_wanted
& need
) != need
) {
2386 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2387 ceph_cap_string(need
), ceph_cap_string(file_wanted
));
2393 /* finish pending truncate */
2394 while (ci
->i_truncate_pending
) {
2395 spin_unlock(&ci
->i_ceph_lock
);
2396 if (snap_rwsem_locked
) {
2397 up_read(&mdsc
->snap_rwsem
);
2398 snap_rwsem_locked
= false;
2400 __ceph_do_pending_vmtruncate(inode
);
2401 spin_lock(&ci
->i_ceph_lock
);
2404 have
= __ceph_caps_issued(ci
, &implemented
);
2406 if (have
& need
& CEPH_CAP_FILE_WR
) {
2407 if (endoff
>= 0 && endoff
> (loff_t
)ci
->i_max_size
) {
2408 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2409 inode
, endoff
, ci
->i_max_size
);
2410 if (endoff
> ci
->i_requested_max_size
) {
2417 * If a sync write is in progress, we must wait, so that we
2418 * can get a final snapshot value for size+mtime.
2420 if (__ceph_have_pending_cap_snap(ci
)) {
2421 dout("get_cap_refs %p cap_snap_pending\n", inode
);
2426 if ((have
& need
) == need
) {
2428 * Look at (implemented & ~have & not) so that we keep waiting
2429 * on transition from wanted -> needed caps. This is needed
2430 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2431 * going before a prior buffered writeback happens.
2433 int not = want
& ~(have
& need
);
2434 int revoking
= implemented
& ~have
;
2435 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2436 inode
, ceph_cap_string(have
), ceph_cap_string(not),
2437 ceph_cap_string(revoking
));
2438 if ((revoking
& not) == 0) {
2439 if (!snap_rwsem_locked
&&
2440 !ci
->i_head_snapc
&&
2441 (need
& CEPH_CAP_FILE_WR
)) {
2442 if (!down_read_trylock(&mdsc
->snap_rwsem
)) {
2444 * we can not call down_read() when
2445 * task isn't in TASK_RUNNING state
2453 spin_unlock(&ci
->i_ceph_lock
);
2454 down_read(&mdsc
->snap_rwsem
);
2455 snap_rwsem_locked
= true;
2458 snap_rwsem_locked
= true;
2460 *got
= need
| (have
& want
);
2461 if ((need
& CEPH_CAP_FILE_RD
) &&
2462 !(*got
& CEPH_CAP_FILE_CACHE
))
2463 ceph_disable_fscache_readpage(ci
);
2464 __take_cap_refs(ci
, *got
, true);
2468 int session_readonly
= false;
2469 if ((need
& CEPH_CAP_FILE_WR
) && ci
->i_auth_cap
) {
2470 struct ceph_mds_session
*s
= ci
->i_auth_cap
->session
;
2471 spin_lock(&s
->s_cap_lock
);
2472 session_readonly
= s
->s_readonly
;
2473 spin_unlock(&s
->s_cap_lock
);
2475 if (session_readonly
) {
2476 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2477 inode
, ceph_cap_string(need
), ci
->i_auth_cap
->mds
);
2483 if (ci
->i_ceph_flags
& CEPH_I_CAP_DROPPED
) {
2485 if (READ_ONCE(mdsc
->fsc
->mount_state
) ==
2486 CEPH_MOUNT_SHUTDOWN
) {
2487 dout("get_cap_refs %p forced umount\n", inode
);
2492 mds_wanted
= __ceph_caps_mds_wanted(ci
, false);
2493 if (need
& ~(mds_wanted
& need
)) {
2494 dout("get_cap_refs %p caps were dropped"
2495 " (session killed?)\n", inode
);
2500 if (!(file_wanted
& ~mds_wanted
))
2501 ci
->i_ceph_flags
&= ~CEPH_I_CAP_DROPPED
;
2504 dout("get_cap_refs %p have %s needed %s\n", inode
,
2505 ceph_cap_string(have
), ceph_cap_string(need
));
2508 spin_unlock(&ci
->i_ceph_lock
);
2509 if (snap_rwsem_locked
)
2510 up_read(&mdsc
->snap_rwsem
);
2512 dout("get_cap_refs %p ret %d got %s\n", inode
,
2513 ret
, ceph_cap_string(*got
));
2518 * Check the offset we are writing up to against our current
2519 * max_size. If necessary, tell the MDS we want to write to
2522 static void check_max_size(struct inode
*inode
, loff_t endoff
)
2524 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2527 /* do we need to explicitly request a larger max_size? */
2528 spin_lock(&ci
->i_ceph_lock
);
2529 if (endoff
>= ci
->i_max_size
&& endoff
> ci
->i_wanted_max_size
) {
2530 dout("write %p at large endoff %llu, req max_size\n",
2532 ci
->i_wanted_max_size
= endoff
;
2534 /* duplicate ceph_check_caps()'s logic */
2535 if (ci
->i_auth_cap
&&
2536 (ci
->i_auth_cap
->issued
& CEPH_CAP_FILE_WR
) &&
2537 ci
->i_wanted_max_size
> ci
->i_max_size
&&
2538 ci
->i_wanted_max_size
> ci
->i_requested_max_size
)
2540 spin_unlock(&ci
->i_ceph_lock
);
2542 ceph_check_caps(ci
, CHECK_CAPS_AUTHONLY
, NULL
);
2545 int ceph_try_get_caps(struct ceph_inode_info
*ci
, int need
, int want
, int *got
)
2549 BUG_ON(need
& ~CEPH_CAP_FILE_RD
);
2550 BUG_ON(want
& ~(CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
));
2551 ret
= ceph_pool_perm_check(ci
, need
);
2555 ret
= try_get_cap_refs(ci
, need
, want
, 0, true, got
, &err
);
2557 if (err
== -EAGAIN
) {
2559 } else if (err
< 0) {
2567 * Wait for caps, and take cap references. If we can't get a WR cap
2568 * due to a small max_size, make sure we check_max_size (and possibly
2569 * ask the mds) so we don't get hung up indefinitely.
2571 int ceph_get_caps(struct ceph_inode_info
*ci
, int need
, int want
,
2572 loff_t endoff
, int *got
, struct page
**pinned_page
)
2574 int _got
, ret
, err
= 0;
2576 ret
= ceph_pool_perm_check(ci
, need
);
2582 check_max_size(&ci
->vfs_inode
, endoff
);
2586 ret
= try_get_cap_refs(ci
, need
, want
, endoff
,
2587 false, &_got
, &err
);
2594 DEFINE_WAIT_FUNC(wait
, woken_wake_function
);
2595 add_wait_queue(&ci
->i_cap_wq
, &wait
);
2597 while (!try_get_cap_refs(ci
, need
, want
, endoff
,
2598 true, &_got
, &err
)) {
2599 if (signal_pending(current
)) {
2603 wait_woken(&wait
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
2606 remove_wait_queue(&ci
->i_cap_wq
, &wait
);
2614 if (err
== -ESTALE
) {
2615 /* session was killed, try renew caps */
2616 ret
= ceph_renew_caps(&ci
->vfs_inode
);
2623 if (ci
->i_inline_version
!= CEPH_INLINE_NONE
&&
2624 (_got
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)) &&
2625 i_size_read(&ci
->vfs_inode
) > 0) {
2627 find_get_page(ci
->vfs_inode
.i_mapping
, 0);
2629 if (PageUptodate(page
)) {
2630 *pinned_page
= page
;
2636 * drop cap refs first because getattr while
2637 * holding * caps refs can cause deadlock.
2639 ceph_put_cap_refs(ci
, _got
);
2643 * getattr request will bring inline data into
2646 ret
= __ceph_do_getattr(&ci
->vfs_inode
, NULL
,
2647 CEPH_STAT_CAP_INLINE_DATA
,
2656 if ((_got
& CEPH_CAP_FILE_RD
) && (_got
& CEPH_CAP_FILE_CACHE
))
2657 ceph_fscache_revalidate_cookie(ci
);
2664 * Take cap refs. Caller must already know we hold at least one ref
2665 * on the caps in question or we don't know this is safe.
2667 void ceph_get_cap_refs(struct ceph_inode_info
*ci
, int caps
)
2669 spin_lock(&ci
->i_ceph_lock
);
2670 __take_cap_refs(ci
, caps
, false);
2671 spin_unlock(&ci
->i_ceph_lock
);
2676 * drop cap_snap that is not associated with any snapshot.
2677 * we don't need to send FLUSHSNAP message for it.
2679 static int ceph_try_drop_cap_snap(struct ceph_inode_info
*ci
,
2680 struct ceph_cap_snap
*capsnap
)
2682 if (!capsnap
->need_flush
&&
2683 !capsnap
->writing
&& !capsnap
->dirty_pages
) {
2684 dout("dropping cap_snap %p follows %llu\n",
2685 capsnap
, capsnap
->follows
);
2686 BUG_ON(capsnap
->cap_flush
.tid
> 0);
2687 ceph_put_snap_context(capsnap
->context
);
2688 if (!list_is_last(&capsnap
->ci_item
, &ci
->i_cap_snaps
))
2689 ci
->i_ceph_flags
|= CEPH_I_FLUSH_SNAPS
;
2691 list_del(&capsnap
->ci_item
);
2692 ceph_put_cap_snap(capsnap
);
2701 * If we released the last ref on any given cap, call ceph_check_caps
2702 * to release (or schedule a release).
2704 * If we are releasing a WR cap (from a sync write), finalize any affected
2705 * cap_snap, and wake up any waiters.
2707 void ceph_put_cap_refs(struct ceph_inode_info
*ci
, int had
)
2709 struct inode
*inode
= &ci
->vfs_inode
;
2710 int last
= 0, put
= 0, flushsnaps
= 0, wake
= 0;
2712 spin_lock(&ci
->i_ceph_lock
);
2713 if (had
& CEPH_CAP_PIN
)
2715 if (had
& CEPH_CAP_FILE_RD
)
2716 if (--ci
->i_rd_ref
== 0)
2718 if (had
& CEPH_CAP_FILE_CACHE
)
2719 if (--ci
->i_rdcache_ref
== 0)
2721 if (had
& CEPH_CAP_FILE_BUFFER
) {
2722 if (--ci
->i_wb_ref
== 0) {
2726 dout("put_cap_refs %p wb %d -> %d (?)\n",
2727 inode
, ci
->i_wb_ref
+1, ci
->i_wb_ref
);
2729 if (had
& CEPH_CAP_FILE_WR
)
2730 if (--ci
->i_wr_ref
== 0) {
2732 if (__ceph_have_pending_cap_snap(ci
)) {
2733 struct ceph_cap_snap
*capsnap
=
2734 list_last_entry(&ci
->i_cap_snaps
,
2735 struct ceph_cap_snap
,
2737 capsnap
->writing
= 0;
2738 if (ceph_try_drop_cap_snap(ci
, capsnap
))
2740 else if (__ceph_finish_cap_snap(ci
, capsnap
))
2744 if (ci
->i_wrbuffer_ref_head
== 0 &&
2745 ci
->i_dirty_caps
== 0 &&
2746 ci
->i_flushing_caps
== 0) {
2747 BUG_ON(!ci
->i_head_snapc
);
2748 ceph_put_snap_context(ci
->i_head_snapc
);
2749 ci
->i_head_snapc
= NULL
;
2751 /* see comment in __ceph_remove_cap() */
2752 if (!__ceph_is_any_caps(ci
) && ci
->i_snap_realm
)
2753 drop_inode_snap_realm(ci
);
2755 spin_unlock(&ci
->i_ceph_lock
);
2757 dout("put_cap_refs %p had %s%s%s\n", inode
, ceph_cap_string(had
),
2758 last
? " last" : "", put
? " put" : "");
2760 if (last
&& !flushsnaps
)
2761 ceph_check_caps(ci
, 0, NULL
);
2762 else if (flushsnaps
)
2763 ceph_flush_snaps(ci
, NULL
);
2765 wake_up_all(&ci
->i_cap_wq
);
2771 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2772 * context. Adjust per-snap dirty page accounting as appropriate.
2773 * Once all dirty data for a cap_snap is flushed, flush snapped file
2774 * metadata back to the MDS. If we dropped the last ref, call
2777 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info
*ci
, int nr
,
2778 struct ceph_snap_context
*snapc
)
2780 struct inode
*inode
= &ci
->vfs_inode
;
2781 struct ceph_cap_snap
*capsnap
= NULL
;
2785 bool flush_snaps
= false;
2786 bool complete_capsnap
= false;
2788 spin_lock(&ci
->i_ceph_lock
);
2789 ci
->i_wrbuffer_ref
-= nr
;
2790 if (ci
->i_wrbuffer_ref
== 0) {
2795 if (ci
->i_head_snapc
== snapc
) {
2796 ci
->i_wrbuffer_ref_head
-= nr
;
2797 if (ci
->i_wrbuffer_ref_head
== 0 &&
2798 ci
->i_wr_ref
== 0 &&
2799 ci
->i_dirty_caps
== 0 &&
2800 ci
->i_flushing_caps
== 0) {
2801 BUG_ON(!ci
->i_head_snapc
);
2802 ceph_put_snap_context(ci
->i_head_snapc
);
2803 ci
->i_head_snapc
= NULL
;
2805 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2807 ci
->i_wrbuffer_ref
+nr
, ci
->i_wrbuffer_ref_head
+nr
,
2808 ci
->i_wrbuffer_ref
, ci
->i_wrbuffer_ref_head
,
2809 last
? " LAST" : "");
2811 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
2812 if (capsnap
->context
== snapc
) {
2818 capsnap
->dirty_pages
-= nr
;
2819 if (capsnap
->dirty_pages
== 0) {
2820 complete_capsnap
= true;
2821 if (!capsnap
->writing
) {
2822 if (ceph_try_drop_cap_snap(ci
, capsnap
)) {
2825 ci
->i_ceph_flags
|= CEPH_I_FLUSH_SNAPS
;
2830 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2831 " snap %lld %d/%d -> %d/%d %s%s\n",
2832 inode
, capsnap
, capsnap
->context
->seq
,
2833 ci
->i_wrbuffer_ref
+nr
, capsnap
->dirty_pages
+ nr
,
2834 ci
->i_wrbuffer_ref
, capsnap
->dirty_pages
,
2835 last
? " (wrbuffer last)" : "",
2836 complete_capsnap
? " (complete capsnap)" : "");
2839 spin_unlock(&ci
->i_ceph_lock
);
2842 ceph_check_caps(ci
, CHECK_CAPS_AUTHONLY
, NULL
);
2843 } else if (flush_snaps
) {
2844 ceph_flush_snaps(ci
, NULL
);
2846 if (complete_capsnap
)
2847 wake_up_all(&ci
->i_cap_wq
);
2853 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2855 static void invalidate_aliases(struct inode
*inode
)
2857 struct dentry
*dn
, *prev
= NULL
;
2859 dout("invalidate_aliases inode %p\n", inode
);
2860 d_prune_aliases(inode
);
2862 * For non-directory inode, d_find_alias() only returns
2863 * hashed dentry. After calling d_invalidate(), the
2864 * dentry becomes unhashed.
2866 * For directory inode, d_find_alias() can return
2867 * unhashed dentry. But directory inode should have
2868 * one alias at most.
2870 while ((dn
= d_find_alias(inode
))) {
2885 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2886 * actually be a revocation if it specifies a smaller cap set.)
2888 * caller holds s_mutex and i_ceph_lock, we drop both.
2890 static void handle_cap_grant(struct ceph_mds_client
*mdsc
,
2891 struct inode
*inode
, struct ceph_mds_caps
*grant
,
2892 struct ceph_string
**pns
, u64 inline_version
,
2893 void *inline_data
, u32 inline_len
,
2894 struct ceph_buffer
*xattr_buf
,
2895 struct ceph_mds_session
*session
,
2896 struct ceph_cap
*cap
, int issued
)
2897 __releases(ci
->i_ceph_lock
)
2898 __releases(mdsc
->snap_rwsem
)
2900 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2901 int mds
= session
->s_mds
;
2902 int seq
= le32_to_cpu(grant
->seq
);
2903 int newcaps
= le32_to_cpu(grant
->caps
);
2904 int used
, wanted
, dirty
;
2905 u64 size
= le64_to_cpu(grant
->size
);
2906 u64 max_size
= le64_to_cpu(grant
->max_size
);
2907 struct timespec mtime
, atime
, ctime
;
2910 bool writeback
= false;
2911 bool queue_trunc
= false;
2912 bool queue_invalidate
= false;
2913 bool deleted_inode
= false;
2914 bool fill_inline
= false;
2916 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2917 inode
, cap
, mds
, seq
, ceph_cap_string(newcaps
));
2918 dout(" size %llu max_size %llu, i_size %llu\n", size
, max_size
,
2923 * auth mds of the inode changed. we received the cap export message,
2924 * but still haven't received the cap import message. handle_cap_export
2925 * updated the new auth MDS' cap.
2927 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2928 * that was sent before the cap import message. So don't remove caps.
2930 if (ceph_seq_cmp(seq
, cap
->seq
) <= 0) {
2931 WARN_ON(cap
!= ci
->i_auth_cap
);
2932 WARN_ON(cap
->cap_id
!= le64_to_cpu(grant
->cap_id
));
2934 newcaps
|= cap
->issued
;
2938 * If CACHE is being revoked, and we have no dirty buffers,
2939 * try to invalidate (once). (If there are dirty buffers, we
2940 * will invalidate _after_ writeback.)
2942 if (!S_ISDIR(inode
->i_mode
) && /* don't invalidate readdir cache */
2943 ((cap
->issued
& ~newcaps
) & CEPH_CAP_FILE_CACHE
) &&
2944 (newcaps
& CEPH_CAP_FILE_LAZYIO
) == 0 &&
2945 !(ci
->i_wrbuffer_ref
|| ci
->i_wb_ref
)) {
2946 if (try_nonblocking_invalidate(inode
)) {
2947 /* there were locked pages.. invalidate later
2948 in a separate thread. */
2949 if (ci
->i_rdcache_revoking
!= ci
->i_rdcache_gen
) {
2950 queue_invalidate
= true;
2951 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
;
2956 /* side effects now are allowed */
2957 cap
->cap_gen
= session
->s_cap_gen
;
2960 __check_cap_issue(ci
, cap
, newcaps
);
2962 if ((newcaps
& CEPH_CAP_AUTH_SHARED
) &&
2963 (issued
& CEPH_CAP_AUTH_EXCL
) == 0) {
2964 inode
->i_mode
= le32_to_cpu(grant
->mode
);
2965 inode
->i_uid
= make_kuid(&init_user_ns
, le32_to_cpu(grant
->uid
));
2966 inode
->i_gid
= make_kgid(&init_user_ns
, le32_to_cpu(grant
->gid
));
2967 dout("%p mode 0%o uid.gid %d.%d\n", inode
, inode
->i_mode
,
2968 from_kuid(&init_user_ns
, inode
->i_uid
),
2969 from_kgid(&init_user_ns
, inode
->i_gid
));
2972 if ((newcaps
& CEPH_CAP_AUTH_SHARED
) &&
2973 (issued
& CEPH_CAP_LINK_EXCL
) == 0) {
2974 set_nlink(inode
, le32_to_cpu(grant
->nlink
));
2975 if (inode
->i_nlink
== 0 &&
2976 (newcaps
& (CEPH_CAP_LINK_SHARED
| CEPH_CAP_LINK_EXCL
)))
2977 deleted_inode
= true;
2980 if ((issued
& CEPH_CAP_XATTR_EXCL
) == 0 && grant
->xattr_len
) {
2981 int len
= le32_to_cpu(grant
->xattr_len
);
2982 u64 version
= le64_to_cpu(grant
->xattr_version
);
2984 if (version
> ci
->i_xattrs
.version
) {
2985 dout(" got new xattrs v%llu on %p len %d\n",
2986 version
, inode
, len
);
2987 if (ci
->i_xattrs
.blob
)
2988 ceph_buffer_put(ci
->i_xattrs
.blob
);
2989 ci
->i_xattrs
.blob
= ceph_buffer_get(xattr_buf
);
2990 ci
->i_xattrs
.version
= version
;
2991 ceph_forget_all_cached_acls(inode
);
2995 if (newcaps
& CEPH_CAP_ANY_RD
) {
2996 /* ctime/mtime/atime? */
2997 ceph_decode_timespec(&mtime
, &grant
->mtime
);
2998 ceph_decode_timespec(&atime
, &grant
->atime
);
2999 ceph_decode_timespec(&ctime
, &grant
->ctime
);
3000 ceph_fill_file_time(inode
, issued
,
3001 le32_to_cpu(grant
->time_warp_seq
),
3002 &ctime
, &mtime
, &atime
);
3005 if (newcaps
& (CEPH_CAP_ANY_FILE_RD
| CEPH_CAP_ANY_FILE_WR
)) {
3006 /* file layout may have changed */
3007 s64 old_pool
= ci
->i_layout
.pool_id
;
3008 struct ceph_string
*old_ns
;
3010 ceph_file_layout_from_legacy(&ci
->i_layout
, &grant
->layout
);
3011 old_ns
= rcu_dereference_protected(ci
->i_layout
.pool_ns
,
3012 lockdep_is_held(&ci
->i_ceph_lock
));
3013 rcu_assign_pointer(ci
->i_layout
.pool_ns
, *pns
);
3015 if (ci
->i_layout
.pool_id
!= old_pool
|| *pns
!= old_ns
)
3016 ci
->i_ceph_flags
&= ~CEPH_I_POOL_PERM
;
3020 /* size/truncate_seq? */
3021 queue_trunc
= ceph_fill_file_size(inode
, issued
,
3022 le32_to_cpu(grant
->truncate_seq
),
3023 le64_to_cpu(grant
->truncate_size
),
3025 /* max size increase? */
3026 if (ci
->i_auth_cap
== cap
&& max_size
!= ci
->i_max_size
) {
3027 dout("max_size %lld -> %llu\n",
3028 ci
->i_max_size
, max_size
);
3029 ci
->i_max_size
= max_size
;
3030 if (max_size
>= ci
->i_wanted_max_size
) {
3031 ci
->i_wanted_max_size
= 0; /* reset */
3032 ci
->i_requested_max_size
= 0;
3038 /* check cap bits */
3039 wanted
= __ceph_caps_wanted(ci
);
3040 used
= __ceph_caps_used(ci
);
3041 dirty
= __ceph_caps_dirty(ci
);
3042 dout(" my wanted = %s, used = %s, dirty %s\n",
3043 ceph_cap_string(wanted
),
3044 ceph_cap_string(used
),
3045 ceph_cap_string(dirty
));
3046 if (wanted
!= le32_to_cpu(grant
->wanted
)) {
3047 dout("mds wanted %s -> %s\n",
3048 ceph_cap_string(le32_to_cpu(grant
->wanted
)),
3049 ceph_cap_string(wanted
));
3050 /* imported cap may not have correct mds_wanted */
3051 if (le32_to_cpu(grant
->op
) == CEPH_CAP_OP_IMPORT
)
3055 /* revocation, grant, or no-op? */
3056 if (cap
->issued
& ~newcaps
) {
3057 int revoking
= cap
->issued
& ~newcaps
;
3059 dout("revocation: %s -> %s (revoking %s)\n",
3060 ceph_cap_string(cap
->issued
),
3061 ceph_cap_string(newcaps
),
3062 ceph_cap_string(revoking
));
3063 if (revoking
& used
& CEPH_CAP_FILE_BUFFER
)
3064 writeback
= true; /* initiate writeback; will delay ack */
3065 else if (revoking
== CEPH_CAP_FILE_CACHE
&&
3066 (newcaps
& CEPH_CAP_FILE_LAZYIO
) == 0 &&
3068 ; /* do nothing yet, invalidation will be queued */
3069 else if (cap
== ci
->i_auth_cap
)
3070 check_caps
= 1; /* check auth cap only */
3072 check_caps
= 2; /* check all caps */
3073 cap
->issued
= newcaps
;
3074 cap
->implemented
|= newcaps
;
3075 } else if (cap
->issued
== newcaps
) {
3076 dout("caps unchanged: %s -> %s\n",
3077 ceph_cap_string(cap
->issued
), ceph_cap_string(newcaps
));
3079 dout("grant: %s -> %s\n", ceph_cap_string(cap
->issued
),
3080 ceph_cap_string(newcaps
));
3081 /* non-auth MDS is revoking the newly grant caps ? */
3082 if (cap
== ci
->i_auth_cap
&&
3083 __ceph_caps_revoking_other(ci
, cap
, newcaps
))
3086 cap
->issued
= newcaps
;
3087 cap
->implemented
|= newcaps
; /* add bits only, to
3088 * avoid stepping on a
3089 * pending revocation */
3092 BUG_ON(cap
->issued
& ~cap
->implemented
);
3094 if (inline_version
> 0 && inline_version
>= ci
->i_inline_version
) {
3095 ci
->i_inline_version
= inline_version
;
3096 if (ci
->i_inline_version
!= CEPH_INLINE_NONE
&&
3097 (newcaps
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)))
3101 if (le32_to_cpu(grant
->op
) == CEPH_CAP_OP_IMPORT
) {
3102 if (newcaps
& ~issued
)
3104 kick_flushing_inode_caps(mdsc
, session
, inode
);
3105 up_read(&mdsc
->snap_rwsem
);
3107 spin_unlock(&ci
->i_ceph_lock
);
3111 ceph_fill_inline_data(inode
, NULL
, inline_data
, inline_len
);
3114 ceph_queue_vmtruncate(inode
);
3118 * queue inode for writeback: we can't actually call
3119 * filemap_write_and_wait, etc. from message handler
3122 ceph_queue_writeback(inode
);
3123 if (queue_invalidate
)
3124 ceph_queue_invalidate(inode
);
3126 invalidate_aliases(inode
);
3128 wake_up_all(&ci
->i_cap_wq
);
3130 if (check_caps
== 1)
3131 ceph_check_caps(ci
, CHECK_CAPS_NODELAY
|CHECK_CAPS_AUTHONLY
,
3133 else if (check_caps
== 2)
3134 ceph_check_caps(ci
, CHECK_CAPS_NODELAY
, session
);
3136 mutex_unlock(&session
->s_mutex
);
3140 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3141 * MDS has been safely committed.
3143 static void handle_cap_flush_ack(struct inode
*inode
, u64 flush_tid
,
3144 struct ceph_mds_caps
*m
,
3145 struct ceph_mds_session
*session
,
3146 struct ceph_cap
*cap
)
3147 __releases(ci
->i_ceph_lock
)
3149 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3150 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
3151 struct ceph_cap_flush
*cf
, *tmp_cf
;
3152 LIST_HEAD(to_remove
);
3153 unsigned seq
= le32_to_cpu(m
->seq
);
3154 int dirty
= le32_to_cpu(m
->dirty
);
3160 list_for_each_entry_safe(cf
, tmp_cf
, &ci
->i_cap_flush_list
, i_list
) {
3161 if (cf
->tid
== flush_tid
)
3163 if (cf
->caps
== 0) /* capsnap */
3165 if (cf
->tid
<= flush_tid
) {
3166 if (__finish_cap_flush(NULL
, ci
, cf
))
3168 list_add_tail(&cf
->i_list
, &to_remove
);
3170 cleaned
&= ~cf
->caps
;
3176 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3177 " flushing %s -> %s\n",
3178 inode
, session
->s_mds
, seq
, ceph_cap_string(dirty
),
3179 ceph_cap_string(cleaned
), ceph_cap_string(ci
->i_flushing_caps
),
3180 ceph_cap_string(ci
->i_flushing_caps
& ~cleaned
));
3182 if (list_empty(&to_remove
) && !cleaned
)
3185 ci
->i_flushing_caps
&= ~cleaned
;
3187 spin_lock(&mdsc
->cap_dirty_lock
);
3189 list_for_each_entry(cf
, &to_remove
, i_list
) {
3190 if (__finish_cap_flush(mdsc
, NULL
, cf
))
3194 if (ci
->i_flushing_caps
== 0) {
3195 if (list_empty(&ci
->i_cap_flush_list
)) {
3196 list_del_init(&ci
->i_flushing_item
);
3197 if (!list_empty(&session
->s_cap_flushing
)) {
3198 dout(" mds%d still flushing cap on %p\n",
3200 &list_first_entry(&session
->s_cap_flushing
,
3201 struct ceph_inode_info
,
3202 i_flushing_item
)->vfs_inode
);
3205 mdsc
->num_cap_flushing
--;
3206 dout(" inode %p now !flushing\n", inode
);
3208 if (ci
->i_dirty_caps
== 0) {
3209 dout(" inode %p now clean\n", inode
);
3210 BUG_ON(!list_empty(&ci
->i_dirty_item
));
3212 if (ci
->i_wr_ref
== 0 &&
3213 ci
->i_wrbuffer_ref_head
== 0) {
3214 BUG_ON(!ci
->i_head_snapc
);
3215 ceph_put_snap_context(ci
->i_head_snapc
);
3216 ci
->i_head_snapc
= NULL
;
3219 BUG_ON(list_empty(&ci
->i_dirty_item
));
3222 spin_unlock(&mdsc
->cap_dirty_lock
);
3225 spin_unlock(&ci
->i_ceph_lock
);
3227 while (!list_empty(&to_remove
)) {
3228 cf
= list_first_entry(&to_remove
,
3229 struct ceph_cap_flush
, i_list
);
3230 list_del(&cf
->i_list
);
3231 ceph_free_cap_flush(cf
);
3235 wake_up_all(&ci
->i_cap_wq
);
3237 wake_up_all(&mdsc
->cap_flushing_wq
);
3243 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3244 * throw away our cap_snap.
3246 * Caller hold s_mutex.
3248 static void handle_cap_flushsnap_ack(struct inode
*inode
, u64 flush_tid
,
3249 struct ceph_mds_caps
*m
,
3250 struct ceph_mds_session
*session
)
3252 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3253 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
3254 u64 follows
= le64_to_cpu(m
->snap_follows
);
3255 struct ceph_cap_snap
*capsnap
;
3256 bool flushed
= false;
3257 bool wake_ci
= false;
3258 bool wake_mdsc
= false;
3260 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3261 inode
, ci
, session
->s_mds
, follows
);
3263 spin_lock(&ci
->i_ceph_lock
);
3264 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
3265 if (capsnap
->follows
== follows
) {
3266 if (capsnap
->cap_flush
.tid
!= flush_tid
) {
3267 dout(" cap_snap %p follows %lld tid %lld !="
3268 " %lld\n", capsnap
, follows
,
3269 flush_tid
, capsnap
->cap_flush
.tid
);
3275 dout(" skipping cap_snap %p follows %lld\n",
3276 capsnap
, capsnap
->follows
);
3280 WARN_ON(capsnap
->dirty_pages
|| capsnap
->writing
);
3281 dout(" removing %p cap_snap %p follows %lld\n",
3282 inode
, capsnap
, follows
);
3283 list_del(&capsnap
->ci_item
);
3284 if (__finish_cap_flush(NULL
, ci
, &capsnap
->cap_flush
))
3287 spin_lock(&mdsc
->cap_dirty_lock
);
3289 if (list_empty(&ci
->i_cap_flush_list
))
3290 list_del_init(&ci
->i_flushing_item
);
3292 if (__finish_cap_flush(mdsc
, NULL
, &capsnap
->cap_flush
))
3295 spin_unlock(&mdsc
->cap_dirty_lock
);
3297 spin_unlock(&ci
->i_ceph_lock
);
3299 ceph_put_snap_context(capsnap
->context
);
3300 ceph_put_cap_snap(capsnap
);
3302 wake_up_all(&ci
->i_cap_wq
);
3304 wake_up_all(&mdsc
->cap_flushing_wq
);
3310 * Handle TRUNC from MDS, indicating file truncation.
3312 * caller hold s_mutex.
3314 static void handle_cap_trunc(struct inode
*inode
,
3315 struct ceph_mds_caps
*trunc
,
3316 struct ceph_mds_session
*session
)
3317 __releases(ci
->i_ceph_lock
)
3319 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3320 int mds
= session
->s_mds
;
3321 int seq
= le32_to_cpu(trunc
->seq
);
3322 u32 truncate_seq
= le32_to_cpu(trunc
->truncate_seq
);
3323 u64 truncate_size
= le64_to_cpu(trunc
->truncate_size
);
3324 u64 size
= le64_to_cpu(trunc
->size
);
3325 int implemented
= 0;
3326 int dirty
= __ceph_caps_dirty(ci
);
3327 int issued
= __ceph_caps_issued(ceph_inode(inode
), &implemented
);
3328 int queue_trunc
= 0;
3330 issued
|= implemented
| dirty
;
3332 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3333 inode
, mds
, seq
, truncate_size
, truncate_seq
);
3334 queue_trunc
= ceph_fill_file_size(inode
, issued
,
3335 truncate_seq
, truncate_size
, size
);
3336 spin_unlock(&ci
->i_ceph_lock
);
3339 ceph_queue_vmtruncate(inode
);
3343 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3344 * different one. If we are the most recent migration we've seen (as
3345 * indicated by mseq), make note of the migrating cap bits for the
3346 * duration (until we see the corresponding IMPORT).
3348 * caller holds s_mutex
3350 static void handle_cap_export(struct inode
*inode
, struct ceph_mds_caps
*ex
,
3351 struct ceph_mds_cap_peer
*ph
,
3352 struct ceph_mds_session
*session
)
3354 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
3355 struct ceph_mds_session
*tsession
= NULL
;
3356 struct ceph_cap
*cap
, *tcap
, *new_cap
= NULL
;
3357 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3359 unsigned mseq
= le32_to_cpu(ex
->migrate_seq
);
3360 unsigned t_seq
, t_mseq
;
3362 int mds
= session
->s_mds
;
3365 t_cap_id
= le64_to_cpu(ph
->cap_id
);
3366 t_seq
= le32_to_cpu(ph
->seq
);
3367 t_mseq
= le32_to_cpu(ph
->mseq
);
3368 target
= le32_to_cpu(ph
->mds
);
3370 t_cap_id
= t_seq
= t_mseq
= 0;
3374 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3375 inode
, ci
, mds
, mseq
, target
);
3377 spin_lock(&ci
->i_ceph_lock
);
3378 cap
= __get_cap_for_mds(ci
, mds
);
3379 if (!cap
|| cap
->cap_id
!= le64_to_cpu(ex
->cap_id
))
3383 __ceph_remove_cap(cap
, false);
3384 if (!ci
->i_auth_cap
)
3385 ci
->i_ceph_flags
|= CEPH_I_CAP_DROPPED
;
3390 * now we know we haven't received the cap import message yet
3391 * because the exported cap still exist.
3394 issued
= cap
->issued
;
3395 WARN_ON(issued
!= cap
->implemented
);
3397 tcap
= __get_cap_for_mds(ci
, target
);
3399 /* already have caps from the target */
3400 if (tcap
->cap_id
!= t_cap_id
||
3401 ceph_seq_cmp(tcap
->seq
, t_seq
) < 0) {
3402 dout(" updating import cap %p mds%d\n", tcap
, target
);
3403 tcap
->cap_id
= t_cap_id
;
3404 tcap
->seq
= t_seq
- 1;
3405 tcap
->issue_seq
= t_seq
- 1;
3406 tcap
->mseq
= t_mseq
;
3407 tcap
->issued
|= issued
;
3408 tcap
->implemented
|= issued
;
3409 if (cap
== ci
->i_auth_cap
)
3410 ci
->i_auth_cap
= tcap
;
3412 if (!list_empty(&ci
->i_cap_flush_list
) &&
3413 ci
->i_auth_cap
== tcap
) {
3414 spin_lock(&mdsc
->cap_dirty_lock
);
3415 list_move_tail(&ci
->i_flushing_item
,
3416 &tcap
->session
->s_cap_flushing
);
3417 spin_unlock(&mdsc
->cap_dirty_lock
);
3420 __ceph_remove_cap(cap
, false);
3422 } else if (tsession
) {
3423 /* add placeholder for the export tagert */
3424 int flag
= (cap
== ci
->i_auth_cap
) ? CEPH_CAP_FLAG_AUTH
: 0;
3426 ceph_add_cap(inode
, tsession
, t_cap_id
, -1, issued
, 0,
3427 t_seq
- 1, t_mseq
, (u64
)-1, flag
, &new_cap
);
3429 if (!list_empty(&ci
->i_cap_flush_list
) &&
3430 ci
->i_auth_cap
== tcap
) {
3431 spin_lock(&mdsc
->cap_dirty_lock
);
3432 list_move_tail(&ci
->i_flushing_item
,
3433 &tcap
->session
->s_cap_flushing
);
3434 spin_unlock(&mdsc
->cap_dirty_lock
);
3437 __ceph_remove_cap(cap
, false);
3441 spin_unlock(&ci
->i_ceph_lock
);
3442 mutex_unlock(&session
->s_mutex
);
3444 /* open target session */
3445 tsession
= ceph_mdsc_open_export_target_session(mdsc
, target
);
3446 if (!IS_ERR(tsession
)) {
3448 mutex_lock(&session
->s_mutex
);
3449 mutex_lock_nested(&tsession
->s_mutex
,
3450 SINGLE_DEPTH_NESTING
);
3452 mutex_lock(&tsession
->s_mutex
);
3453 mutex_lock_nested(&session
->s_mutex
,
3454 SINGLE_DEPTH_NESTING
);
3456 new_cap
= ceph_get_cap(mdsc
, NULL
);
3465 spin_unlock(&ci
->i_ceph_lock
);
3466 mutex_unlock(&session
->s_mutex
);
3468 mutex_unlock(&tsession
->s_mutex
);
3469 ceph_put_mds_session(tsession
);
3472 ceph_put_cap(mdsc
, new_cap
);
3476 * Handle cap IMPORT.
3478 * caller holds s_mutex. acquires i_ceph_lock
3480 static void handle_cap_import(struct ceph_mds_client
*mdsc
,
3481 struct inode
*inode
, struct ceph_mds_caps
*im
,
3482 struct ceph_mds_cap_peer
*ph
,
3483 struct ceph_mds_session
*session
,
3484 struct ceph_cap
**target_cap
, int *old_issued
)
3485 __acquires(ci
->i_ceph_lock
)
3487 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3488 struct ceph_cap
*cap
, *ocap
, *new_cap
= NULL
;
3489 int mds
= session
->s_mds
;
3491 unsigned caps
= le32_to_cpu(im
->caps
);
3492 unsigned wanted
= le32_to_cpu(im
->wanted
);
3493 unsigned seq
= le32_to_cpu(im
->seq
);
3494 unsigned mseq
= le32_to_cpu(im
->migrate_seq
);
3495 u64 realmino
= le64_to_cpu(im
->realm
);
3496 u64 cap_id
= le64_to_cpu(im
->cap_id
);
3501 p_cap_id
= le64_to_cpu(ph
->cap_id
);
3502 peer
= le32_to_cpu(ph
->mds
);
3508 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3509 inode
, ci
, mds
, mseq
, peer
);
3512 spin_lock(&ci
->i_ceph_lock
);
3513 cap
= __get_cap_for_mds(ci
, mds
);
3516 spin_unlock(&ci
->i_ceph_lock
);
3517 new_cap
= ceph_get_cap(mdsc
, NULL
);
3523 ceph_put_cap(mdsc
, new_cap
);
3528 __ceph_caps_issued(ci
, &issued
);
3529 issued
|= __ceph_caps_dirty(ci
);
3531 ceph_add_cap(inode
, session
, cap_id
, -1, caps
, wanted
, seq
, mseq
,
3532 realmino
, CEPH_CAP_FLAG_AUTH
, &new_cap
);
3534 ocap
= peer
>= 0 ? __get_cap_for_mds(ci
, peer
) : NULL
;
3535 if (ocap
&& ocap
->cap_id
== p_cap_id
) {
3536 dout(" remove export cap %p mds%d flags %d\n",
3537 ocap
, peer
, ph
->flags
);
3538 if ((ph
->flags
& CEPH_CAP_FLAG_AUTH
) &&
3539 (ocap
->seq
!= le32_to_cpu(ph
->seq
) ||
3540 ocap
->mseq
!= le32_to_cpu(ph
->mseq
))) {
3541 pr_err("handle_cap_import: mismatched seq/mseq: "
3542 "ino (%llx.%llx) mds%d seq %d mseq %d "
3543 "importer mds%d has peer seq %d mseq %d\n",
3544 ceph_vinop(inode
), peer
, ocap
->seq
,
3545 ocap
->mseq
, mds
, le32_to_cpu(ph
->seq
),
3546 le32_to_cpu(ph
->mseq
));
3548 __ceph_remove_cap(ocap
, (ph
->flags
& CEPH_CAP_FLAG_RELEASE
));
3551 /* make sure we re-request max_size, if necessary */
3552 ci
->i_wanted_max_size
= 0;
3553 ci
->i_requested_max_size
= 0;
3555 *old_issued
= issued
;
3560 * Handle a caps message from the MDS.
3562 * Identify the appropriate session, inode, and call the right handler
3563 * based on the cap op.
3565 void ceph_handle_caps(struct ceph_mds_session
*session
,
3566 struct ceph_msg
*msg
)
3568 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
3569 struct super_block
*sb
= mdsc
->fsc
->sb
;
3570 struct inode
*inode
;
3571 struct ceph_inode_info
*ci
;
3572 struct ceph_cap
*cap
;
3573 struct ceph_mds_caps
*h
;
3574 struct ceph_mds_cap_peer
*peer
= NULL
;
3575 struct ceph_snap_realm
*realm
= NULL
;
3576 struct ceph_string
*pool_ns
= NULL
;
3577 int mds
= session
->s_mds
;
3580 struct ceph_vino vino
;
3582 u64 inline_version
= 0;
3583 void *inline_data
= NULL
;
3586 size_t snaptrace_len
;
3589 dout("handle_caps from mds%d\n", mds
);
3592 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
3593 tid
= le64_to_cpu(msg
->hdr
.tid
);
3594 if (msg
->front
.iov_len
< sizeof(*h
))
3596 h
= msg
->front
.iov_base
;
3597 op
= le32_to_cpu(h
->op
);
3598 vino
.ino
= le64_to_cpu(h
->ino
);
3599 vino
.snap
= CEPH_NOSNAP
;
3600 seq
= le32_to_cpu(h
->seq
);
3601 mseq
= le32_to_cpu(h
->migrate_seq
);
3604 snaptrace_len
= le32_to_cpu(h
->snap_trace_len
);
3605 p
= snaptrace
+ snaptrace_len
;
3607 if (le16_to_cpu(msg
->hdr
.version
) >= 2) {
3609 ceph_decode_32_safe(&p
, end
, flock_len
, bad
);
3610 if (p
+ flock_len
> end
)
3615 if (le16_to_cpu(msg
->hdr
.version
) >= 3) {
3616 if (op
== CEPH_CAP_OP_IMPORT
) {
3617 if (p
+ sizeof(*peer
) > end
)
3621 } else if (op
== CEPH_CAP_OP_EXPORT
) {
3622 /* recorded in unused fields */
3623 peer
= (void *)&h
->size
;
3627 if (le16_to_cpu(msg
->hdr
.version
) >= 4) {
3628 ceph_decode_64_safe(&p
, end
, inline_version
, bad
);
3629 ceph_decode_32_safe(&p
, end
, inline_len
, bad
);
3630 if (p
+ inline_len
> end
)
3636 if (le16_to_cpu(msg
->hdr
.version
) >= 8) {
3638 u32 caller_uid
, caller_gid
;
3639 u32 osd_epoch_barrier
;
3642 ceph_decode_32_safe(&p
, end
, osd_epoch_barrier
, bad
);
3644 ceph_decode_64_safe(&p
, end
, flush_tid
, bad
);
3646 ceph_decode_32_safe(&p
, end
, caller_uid
, bad
);
3647 ceph_decode_32_safe(&p
, end
, caller_gid
, bad
);
3649 ceph_decode_32_safe(&p
, end
, pool_ns_len
, bad
);
3650 if (pool_ns_len
> 0) {
3651 ceph_decode_need(&p
, end
, pool_ns_len
, bad
);
3652 pool_ns
= ceph_find_or_create_string(p
, pool_ns_len
);
3658 inode
= ceph_find_inode(sb
, vino
);
3659 ci
= ceph_inode(inode
);
3660 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op
), vino
.ino
,
3663 mutex_lock(&session
->s_mutex
);
3665 dout(" mds%d seq %lld cap seq %u\n", session
->s_mds
, session
->s_seq
,
3669 dout(" i don't have ino %llx\n", vino
.ino
);
3671 if (op
== CEPH_CAP_OP_IMPORT
) {
3672 cap
= ceph_get_cap(mdsc
, NULL
);
3673 cap
->cap_ino
= vino
.ino
;
3674 cap
->queue_release
= 1;
3675 cap
->cap_id
= le64_to_cpu(h
->cap_id
);
3678 cap
->issue_seq
= seq
;
3679 spin_lock(&session
->s_cap_lock
);
3680 list_add_tail(&cap
->session_caps
,
3681 &session
->s_cap_releases
);
3682 session
->s_num_cap_releases
++;
3683 spin_unlock(&session
->s_cap_lock
);
3685 goto flush_cap_releases
;
3688 /* these will work even if we don't have a cap yet */
3690 case CEPH_CAP_OP_FLUSHSNAP_ACK
:
3691 handle_cap_flushsnap_ack(inode
, tid
, h
, session
);
3694 case CEPH_CAP_OP_EXPORT
:
3695 handle_cap_export(inode
, h
, peer
, session
);
3698 case CEPH_CAP_OP_IMPORT
:
3700 if (snaptrace_len
) {
3701 down_write(&mdsc
->snap_rwsem
);
3702 ceph_update_snap_trace(mdsc
, snaptrace
,
3703 snaptrace
+ snaptrace_len
,
3705 downgrade_write(&mdsc
->snap_rwsem
);
3707 down_read(&mdsc
->snap_rwsem
);
3709 handle_cap_import(mdsc
, inode
, h
, peer
, session
,
3711 handle_cap_grant(mdsc
, inode
, h
, &pool_ns
,
3712 inline_version
, inline_data
, inline_len
,
3713 msg
->middle
, session
, cap
, issued
);
3715 ceph_put_snap_realm(mdsc
, realm
);
3719 /* the rest require a cap */
3720 spin_lock(&ci
->i_ceph_lock
);
3721 cap
= __get_cap_for_mds(ceph_inode(inode
), mds
);
3723 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3724 inode
, ceph_ino(inode
), ceph_snap(inode
), mds
);
3725 spin_unlock(&ci
->i_ceph_lock
);
3726 goto flush_cap_releases
;
3729 /* note that each of these drops i_ceph_lock for us */
3731 case CEPH_CAP_OP_REVOKE
:
3732 case CEPH_CAP_OP_GRANT
:
3733 __ceph_caps_issued(ci
, &issued
);
3734 issued
|= __ceph_caps_dirty(ci
);
3735 handle_cap_grant(mdsc
, inode
, h
, &pool_ns
,
3736 inline_version
, inline_data
, inline_len
,
3737 msg
->middle
, session
, cap
, issued
);
3740 case CEPH_CAP_OP_FLUSH_ACK
:
3741 handle_cap_flush_ack(inode
, tid
, h
, session
, cap
);
3744 case CEPH_CAP_OP_TRUNC
:
3745 handle_cap_trunc(inode
, h
, session
);
3749 spin_unlock(&ci
->i_ceph_lock
);
3750 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op
,
3751 ceph_cap_op_name(op
));
3758 * send any cap release message to try to move things
3759 * along for the mds (who clearly thinks we still have this
3762 ceph_send_cap_releases(mdsc
, session
);
3765 mutex_unlock(&session
->s_mutex
);
3768 ceph_put_string(pool_ns
);
3772 pr_err("ceph_handle_caps: corrupt message\n");
3778 * Delayed work handler to process end of delayed cap release LRU list.
3780 void ceph_check_delayed_caps(struct ceph_mds_client
*mdsc
)
3782 struct ceph_inode_info
*ci
;
3783 int flags
= CHECK_CAPS_NODELAY
;
3785 dout("check_delayed_caps\n");
3787 spin_lock(&mdsc
->cap_delay_lock
);
3788 if (list_empty(&mdsc
->cap_delay_list
))
3790 ci
= list_first_entry(&mdsc
->cap_delay_list
,
3791 struct ceph_inode_info
,
3793 if ((ci
->i_ceph_flags
& CEPH_I_FLUSH
) == 0 &&
3794 time_before(jiffies
, ci
->i_hold_caps_max
))
3796 list_del_init(&ci
->i_cap_delay_list
);
3797 spin_unlock(&mdsc
->cap_delay_lock
);
3798 dout("check_delayed_caps on %p\n", &ci
->vfs_inode
);
3799 ceph_check_caps(ci
, flags
, NULL
);
3801 spin_unlock(&mdsc
->cap_delay_lock
);
3805 * Flush all dirty caps to the mds
3807 void ceph_flush_dirty_caps(struct ceph_mds_client
*mdsc
)
3809 struct ceph_inode_info
*ci
;
3810 struct inode
*inode
;
3812 dout("flush_dirty_caps\n");
3813 spin_lock(&mdsc
->cap_dirty_lock
);
3814 while (!list_empty(&mdsc
->cap_dirty
)) {
3815 ci
= list_first_entry(&mdsc
->cap_dirty
, struct ceph_inode_info
,
3817 inode
= &ci
->vfs_inode
;
3819 dout("flush_dirty_caps %p\n", inode
);
3820 spin_unlock(&mdsc
->cap_dirty_lock
);
3821 ceph_check_caps(ci
, CHECK_CAPS_NODELAY
|CHECK_CAPS_FLUSH
, NULL
);
3823 spin_lock(&mdsc
->cap_dirty_lock
);
3825 spin_unlock(&mdsc
->cap_dirty_lock
);
3826 dout("flush_dirty_caps done\n");
3829 void __ceph_get_fmode(struct ceph_inode_info
*ci
, int fmode
)
3832 int bits
= (fmode
<< 1) | 1;
3833 for (i
= 0; i
< CEPH_FILE_MODE_BITS
; i
++) {
3834 if (bits
& (1 << i
))
3835 ci
->i_nr_by_mode
[i
]++;
3840 * Drop open file reference. If we were the last open file,
3841 * we may need to release capabilities to the MDS (or schedule
3842 * their delayed release).
3844 void ceph_put_fmode(struct ceph_inode_info
*ci
, int fmode
)
3847 int bits
= (fmode
<< 1) | 1;
3848 spin_lock(&ci
->i_ceph_lock
);
3849 for (i
= 0; i
< CEPH_FILE_MODE_BITS
; i
++) {
3850 if (bits
& (1 << i
)) {
3851 BUG_ON(ci
->i_nr_by_mode
[i
] == 0);
3852 if (--ci
->i_nr_by_mode
[i
] == 0)
3856 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
3857 &ci
->vfs_inode
, fmode
,
3858 ci
->i_nr_by_mode
[0], ci
->i_nr_by_mode
[1],
3859 ci
->i_nr_by_mode
[2], ci
->i_nr_by_mode
[3]);
3860 spin_unlock(&ci
->i_ceph_lock
);
3862 if (last
&& ci
->i_vino
.snap
== CEPH_NOSNAP
)
3863 ceph_check_caps(ci
, 0, NULL
);
3867 * Helpers for embedding cap and dentry lease releases into mds
3870 * @force is used by dentry_release (below) to force inclusion of a
3871 * record for the directory inode, even when there aren't any caps to
3874 int ceph_encode_inode_release(void **p
, struct inode
*inode
,
3875 int mds
, int drop
, int unless
, int force
)
3877 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3878 struct ceph_cap
*cap
;
3879 struct ceph_mds_request_release
*rel
= *p
;
3883 spin_lock(&ci
->i_ceph_lock
);
3884 used
= __ceph_caps_used(ci
);
3885 dirty
= __ceph_caps_dirty(ci
);
3887 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3888 inode
, mds
, ceph_cap_string(used
|dirty
), ceph_cap_string(drop
),
3889 ceph_cap_string(unless
));
3891 /* only drop unused, clean caps */
3892 drop
&= ~(used
| dirty
);
3894 cap
= __get_cap_for_mds(ci
, mds
);
3895 if (cap
&& __cap_is_valid(cap
)) {
3897 ((cap
->issued
& drop
) &&
3898 (cap
->issued
& unless
) == 0)) {
3899 if ((cap
->issued
& drop
) &&
3900 (cap
->issued
& unless
) == 0) {
3901 int wanted
= __ceph_caps_wanted(ci
);
3902 if ((ci
->i_ceph_flags
& CEPH_I_NODELAY
) == 0)
3903 wanted
|= cap
->mds_wanted
;
3904 dout("encode_inode_release %p cap %p "
3905 "%s -> %s, wanted %s -> %s\n", inode
, cap
,
3906 ceph_cap_string(cap
->issued
),
3907 ceph_cap_string(cap
->issued
& ~drop
),
3908 ceph_cap_string(cap
->mds_wanted
),
3909 ceph_cap_string(wanted
));
3911 cap
->issued
&= ~drop
;
3912 cap
->implemented
&= ~drop
;
3913 cap
->mds_wanted
= wanted
;
3915 dout("encode_inode_release %p cap %p %s"
3916 " (force)\n", inode
, cap
,
3917 ceph_cap_string(cap
->issued
));
3920 rel
->ino
= cpu_to_le64(ceph_ino(inode
));
3921 rel
->cap_id
= cpu_to_le64(cap
->cap_id
);
3922 rel
->seq
= cpu_to_le32(cap
->seq
);
3923 rel
->issue_seq
= cpu_to_le32(cap
->issue_seq
);
3924 rel
->mseq
= cpu_to_le32(cap
->mseq
);
3925 rel
->caps
= cpu_to_le32(cap
->implemented
);
3926 rel
->wanted
= cpu_to_le32(cap
->mds_wanted
);
3932 dout("encode_inode_release %p cap %p %s\n",
3933 inode
, cap
, ceph_cap_string(cap
->issued
));
3936 spin_unlock(&ci
->i_ceph_lock
);
3940 int ceph_encode_dentry_release(void **p
, struct dentry
*dentry
,
3942 int mds
, int drop
, int unless
)
3944 struct dentry
*parent
= NULL
;
3945 struct ceph_mds_request_release
*rel
= *p
;
3946 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
3951 * force an record for the directory caps if we have a dentry lease.
3952 * this is racy (can't take i_ceph_lock and d_lock together), but it
3953 * doesn't have to be perfect; the mds will revoke anything we don't
3956 spin_lock(&dentry
->d_lock
);
3957 if (di
->lease_session
&& di
->lease_session
->s_mds
== mds
)
3960 parent
= dget(dentry
->d_parent
);
3961 dir
= d_inode(parent
);
3963 spin_unlock(&dentry
->d_lock
);
3965 ret
= ceph_encode_inode_release(p
, dir
, mds
, drop
, unless
, force
);
3968 spin_lock(&dentry
->d_lock
);
3969 if (ret
&& di
->lease_session
&& di
->lease_session
->s_mds
== mds
) {
3970 dout("encode_dentry_release %p mds%d seq %d\n",
3971 dentry
, mds
, (int)di
->lease_seq
);
3972 rel
->dname_len
= cpu_to_le32(dentry
->d_name
.len
);
3973 memcpy(*p
, dentry
->d_name
.name
, dentry
->d_name
.len
);
3974 *p
+= dentry
->d_name
.len
;
3975 rel
->dname_seq
= cpu_to_le32(di
->lease_seq
);
3976 __ceph_mdsc_drop_dentry_lease(dentry
);
3978 spin_unlock(&dentry
->d_lock
);