1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/kernel.h>
5 #include <linux/sched.h>
6 #include <linux/slab.h>
7 #include <linux/vmalloc.h>
8 #include <linux/wait.h>
9 #include <linux/writeback.h>
12 #include "mds_client.h"
14 #include <linux/ceph/decode.h>
15 #include <linux/ceph/messenger.h>
18 * Capability management
20 * The Ceph metadata servers control client access to inode metadata
21 * and file data by issuing capabilities, granting clients permission
22 * to read and/or write both inode field and file data to OSDs
23 * (storage nodes). Each capability consists of a set of bits
24 * indicating which operations are allowed.
26 * If the client holds a *_SHARED cap, the client has a coherent value
27 * that can be safely read from the cached inode.
29 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
30 * client is allowed to change inode attributes (e.g., file size,
31 * mtime), note its dirty state in the ceph_cap, and asynchronously
32 * flush that metadata change to the MDS.
34 * In the event of a conflicting operation (perhaps by another
35 * client), the MDS will revoke the conflicting client capabilities.
37 * In order for a client to cache an inode, it must hold a capability
38 * with at least one MDS server. When inodes are released, release
39 * notifications are batched and periodically sent en masse to the MDS
40 * cluster to release server state.
43 static u64
__get_oldest_flush_tid(struct ceph_mds_client
*mdsc
);
44 static void __kick_flushing_caps(struct ceph_mds_client
*mdsc
,
45 struct ceph_mds_session
*session
,
46 struct ceph_inode_info
*ci
,
47 u64 oldest_flush_tid
);
50 * Generate readable cap strings for debugging output.
52 #define MAX_CAP_STR 20
53 static char cap_str
[MAX_CAP_STR
][40];
54 static DEFINE_SPINLOCK(cap_str_lock
);
55 static int last_cap_str
;
57 static char *gcap_string(char *s
, int c
)
59 if (c
& CEPH_CAP_GSHARED
)
61 if (c
& CEPH_CAP_GEXCL
)
63 if (c
& CEPH_CAP_GCACHE
)
69 if (c
& CEPH_CAP_GBUFFER
)
71 if (c
& CEPH_CAP_GLAZYIO
)
76 const char *ceph_cap_string(int caps
)
82 spin_lock(&cap_str_lock
);
84 if (last_cap_str
== MAX_CAP_STR
)
86 spin_unlock(&cap_str_lock
);
90 if (caps
& CEPH_CAP_PIN
)
93 c
= (caps
>> CEPH_CAP_SAUTH
) & 3;
96 s
= gcap_string(s
, c
);
99 c
= (caps
>> CEPH_CAP_SLINK
) & 3;
102 s
= gcap_string(s
, c
);
105 c
= (caps
>> CEPH_CAP_SXATTR
) & 3;
108 s
= gcap_string(s
, c
);
111 c
= caps
>> CEPH_CAP_SFILE
;
114 s
= gcap_string(s
, c
);
123 void ceph_caps_init(struct ceph_mds_client
*mdsc
)
125 INIT_LIST_HEAD(&mdsc
->caps_list
);
126 spin_lock_init(&mdsc
->caps_list_lock
);
129 void ceph_caps_finalize(struct ceph_mds_client
*mdsc
)
131 struct ceph_cap
*cap
;
133 spin_lock(&mdsc
->caps_list_lock
);
134 while (!list_empty(&mdsc
->caps_list
)) {
135 cap
= list_first_entry(&mdsc
->caps_list
,
136 struct ceph_cap
, caps_item
);
137 list_del(&cap
->caps_item
);
138 kmem_cache_free(ceph_cap_cachep
, cap
);
140 mdsc
->caps_total_count
= 0;
141 mdsc
->caps_avail_count
= 0;
142 mdsc
->caps_use_count
= 0;
143 mdsc
->caps_reserve_count
= 0;
144 mdsc
->caps_min_count
= 0;
145 spin_unlock(&mdsc
->caps_list_lock
);
148 void ceph_adjust_min_caps(struct ceph_mds_client
*mdsc
, int delta
)
150 spin_lock(&mdsc
->caps_list_lock
);
151 mdsc
->caps_min_count
+= delta
;
152 BUG_ON(mdsc
->caps_min_count
< 0);
153 spin_unlock(&mdsc
->caps_list_lock
);
156 void ceph_reserve_caps(struct ceph_mds_client
*mdsc
,
157 struct ceph_cap_reservation
*ctx
, int need
)
160 struct ceph_cap
*cap
;
165 dout("reserve caps ctx=%p need=%d\n", ctx
, need
);
167 /* first reserve any caps that are already allocated */
168 spin_lock(&mdsc
->caps_list_lock
);
169 if (mdsc
->caps_avail_count
>= need
)
172 have
= mdsc
->caps_avail_count
;
173 mdsc
->caps_avail_count
-= have
;
174 mdsc
->caps_reserve_count
+= have
;
175 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
176 mdsc
->caps_reserve_count
+
177 mdsc
->caps_avail_count
);
178 spin_unlock(&mdsc
->caps_list_lock
);
180 for (i
= have
; i
< need
; i
++) {
181 cap
= kmem_cache_alloc(ceph_cap_cachep
, GFP_NOFS
);
184 list_add(&cap
->caps_item
, &newcaps
);
187 /* we didn't manage to reserve as much as we needed */
188 if (have
+ alloc
!= need
)
189 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
190 ctx
, need
, have
+ alloc
);
192 spin_lock(&mdsc
->caps_list_lock
);
193 mdsc
->caps_total_count
+= alloc
;
194 mdsc
->caps_reserve_count
+= alloc
;
195 list_splice(&newcaps
, &mdsc
->caps_list
);
197 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
198 mdsc
->caps_reserve_count
+
199 mdsc
->caps_avail_count
);
200 spin_unlock(&mdsc
->caps_list_lock
);
203 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
204 ctx
, mdsc
->caps_total_count
, mdsc
->caps_use_count
,
205 mdsc
->caps_reserve_count
, mdsc
->caps_avail_count
);
208 int ceph_unreserve_caps(struct ceph_mds_client
*mdsc
,
209 struct ceph_cap_reservation
*ctx
)
211 dout("unreserve caps ctx=%p count=%d\n", ctx
, ctx
->count
);
213 spin_lock(&mdsc
->caps_list_lock
);
214 BUG_ON(mdsc
->caps_reserve_count
< ctx
->count
);
215 mdsc
->caps_reserve_count
-= ctx
->count
;
216 mdsc
->caps_avail_count
+= ctx
->count
;
218 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
219 mdsc
->caps_total_count
, mdsc
->caps_use_count
,
220 mdsc
->caps_reserve_count
, mdsc
->caps_avail_count
);
221 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
222 mdsc
->caps_reserve_count
+
223 mdsc
->caps_avail_count
);
224 spin_unlock(&mdsc
->caps_list_lock
);
229 struct ceph_cap
*ceph_get_cap(struct ceph_mds_client
*mdsc
,
230 struct ceph_cap_reservation
*ctx
)
232 struct ceph_cap
*cap
= NULL
;
234 /* temporary, until we do something about cap import/export */
236 cap
= kmem_cache_alloc(ceph_cap_cachep
, GFP_NOFS
);
238 spin_lock(&mdsc
->caps_list_lock
);
239 mdsc
->caps_use_count
++;
240 mdsc
->caps_total_count
++;
241 spin_unlock(&mdsc
->caps_list_lock
);
246 spin_lock(&mdsc
->caps_list_lock
);
247 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
248 ctx
, ctx
->count
, mdsc
->caps_total_count
, mdsc
->caps_use_count
,
249 mdsc
->caps_reserve_count
, mdsc
->caps_avail_count
);
251 BUG_ON(ctx
->count
> mdsc
->caps_reserve_count
);
252 BUG_ON(list_empty(&mdsc
->caps_list
));
255 mdsc
->caps_reserve_count
--;
256 mdsc
->caps_use_count
++;
258 cap
= list_first_entry(&mdsc
->caps_list
, struct ceph_cap
, caps_item
);
259 list_del(&cap
->caps_item
);
261 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
262 mdsc
->caps_reserve_count
+ mdsc
->caps_avail_count
);
263 spin_unlock(&mdsc
->caps_list_lock
);
267 void ceph_put_cap(struct ceph_mds_client
*mdsc
, struct ceph_cap
*cap
)
269 spin_lock(&mdsc
->caps_list_lock
);
270 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
271 cap
, mdsc
->caps_total_count
, mdsc
->caps_use_count
,
272 mdsc
->caps_reserve_count
, mdsc
->caps_avail_count
);
273 mdsc
->caps_use_count
--;
275 * Keep some preallocated caps around (ceph_min_count), to
276 * avoid lots of free/alloc churn.
278 if (mdsc
->caps_avail_count
>= mdsc
->caps_reserve_count
+
279 mdsc
->caps_min_count
) {
280 mdsc
->caps_total_count
--;
281 kmem_cache_free(ceph_cap_cachep
, cap
);
283 mdsc
->caps_avail_count
++;
284 list_add(&cap
->caps_item
, &mdsc
->caps_list
);
287 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
288 mdsc
->caps_reserve_count
+ mdsc
->caps_avail_count
);
289 spin_unlock(&mdsc
->caps_list_lock
);
292 void ceph_reservation_status(struct ceph_fs_client
*fsc
,
293 int *total
, int *avail
, int *used
, int *reserved
,
296 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
299 *total
= mdsc
->caps_total_count
;
301 *avail
= mdsc
->caps_avail_count
;
303 *used
= mdsc
->caps_use_count
;
305 *reserved
= mdsc
->caps_reserve_count
;
307 *min
= mdsc
->caps_min_count
;
311 * Find ceph_cap for given mds, if any.
313 * Called with i_ceph_lock held.
315 static struct ceph_cap
*__get_cap_for_mds(struct ceph_inode_info
*ci
, int mds
)
317 struct ceph_cap
*cap
;
318 struct rb_node
*n
= ci
->i_caps
.rb_node
;
321 cap
= rb_entry(n
, struct ceph_cap
, ci_node
);
324 else if (mds
> cap
->mds
)
332 struct ceph_cap
*ceph_get_cap_for_mds(struct ceph_inode_info
*ci
, int mds
)
334 struct ceph_cap
*cap
;
336 spin_lock(&ci
->i_ceph_lock
);
337 cap
= __get_cap_for_mds(ci
, mds
);
338 spin_unlock(&ci
->i_ceph_lock
);
343 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
345 static int __ceph_get_cap_mds(struct ceph_inode_info
*ci
)
347 struct ceph_cap
*cap
;
351 /* prefer mds with WR|BUFFER|EXCL caps */
352 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
353 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
355 if (cap
->issued
& (CEPH_CAP_FILE_WR
|
356 CEPH_CAP_FILE_BUFFER
|
363 int ceph_get_cap_mds(struct inode
*inode
)
365 struct ceph_inode_info
*ci
= ceph_inode(inode
);
367 spin_lock(&ci
->i_ceph_lock
);
368 mds
= __ceph_get_cap_mds(ceph_inode(inode
));
369 spin_unlock(&ci
->i_ceph_lock
);
374 * Called under i_ceph_lock.
376 static void __insert_cap_node(struct ceph_inode_info
*ci
,
377 struct ceph_cap
*new)
379 struct rb_node
**p
= &ci
->i_caps
.rb_node
;
380 struct rb_node
*parent
= NULL
;
381 struct ceph_cap
*cap
= NULL
;
385 cap
= rb_entry(parent
, struct ceph_cap
, ci_node
);
386 if (new->mds
< cap
->mds
)
388 else if (new->mds
> cap
->mds
)
394 rb_link_node(&new->ci_node
, parent
, p
);
395 rb_insert_color(&new->ci_node
, &ci
->i_caps
);
399 * (re)set cap hold timeouts, which control the delayed release
400 * of unused caps back to the MDS. Should be called on cap use.
402 static void __cap_set_timeouts(struct ceph_mds_client
*mdsc
,
403 struct ceph_inode_info
*ci
)
405 struct ceph_mount_options
*ma
= mdsc
->fsc
->mount_options
;
407 ci
->i_hold_caps_min
= round_jiffies(jiffies
+
408 ma
->caps_wanted_delay_min
* HZ
);
409 ci
->i_hold_caps_max
= round_jiffies(jiffies
+
410 ma
->caps_wanted_delay_max
* HZ
);
411 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci
->vfs_inode
,
412 ci
->i_hold_caps_min
- jiffies
, ci
->i_hold_caps_max
- jiffies
);
416 * (Re)queue cap at the end of the delayed cap release list.
418 * If I_FLUSH is set, leave the inode at the front of the list.
420 * Caller holds i_ceph_lock
421 * -> we take mdsc->cap_delay_lock
423 static void __cap_delay_requeue(struct ceph_mds_client
*mdsc
,
424 struct ceph_inode_info
*ci
)
426 __cap_set_timeouts(mdsc
, ci
);
427 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci
->vfs_inode
,
428 ci
->i_ceph_flags
, ci
->i_hold_caps_max
);
429 if (!mdsc
->stopping
) {
430 spin_lock(&mdsc
->cap_delay_lock
);
431 if (!list_empty(&ci
->i_cap_delay_list
)) {
432 if (ci
->i_ceph_flags
& CEPH_I_FLUSH
)
434 list_del_init(&ci
->i_cap_delay_list
);
436 list_add_tail(&ci
->i_cap_delay_list
, &mdsc
->cap_delay_list
);
438 spin_unlock(&mdsc
->cap_delay_lock
);
443 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
444 * indicating we should send a cap message to flush dirty metadata
445 * asap, and move to the front of the delayed cap list.
447 static void __cap_delay_requeue_front(struct ceph_mds_client
*mdsc
,
448 struct ceph_inode_info
*ci
)
450 dout("__cap_delay_requeue_front %p\n", &ci
->vfs_inode
);
451 spin_lock(&mdsc
->cap_delay_lock
);
452 ci
->i_ceph_flags
|= CEPH_I_FLUSH
;
453 if (!list_empty(&ci
->i_cap_delay_list
))
454 list_del_init(&ci
->i_cap_delay_list
);
455 list_add(&ci
->i_cap_delay_list
, &mdsc
->cap_delay_list
);
456 spin_unlock(&mdsc
->cap_delay_lock
);
460 * Cancel delayed work on cap.
462 * Caller must hold i_ceph_lock.
464 static void __cap_delay_cancel(struct ceph_mds_client
*mdsc
,
465 struct ceph_inode_info
*ci
)
467 dout("__cap_delay_cancel %p\n", &ci
->vfs_inode
);
468 if (list_empty(&ci
->i_cap_delay_list
))
470 spin_lock(&mdsc
->cap_delay_lock
);
471 list_del_init(&ci
->i_cap_delay_list
);
472 spin_unlock(&mdsc
->cap_delay_lock
);
476 * Common issue checks for add_cap, handle_cap_grant.
478 static void __check_cap_issue(struct ceph_inode_info
*ci
, struct ceph_cap
*cap
,
481 unsigned had
= __ceph_caps_issued(ci
, NULL
);
484 * Each time we receive FILE_CACHE anew, we increment
487 if ((issued
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)) &&
488 (had
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)) == 0) {
493 * if we are newly issued FILE_SHARED, mark dir not complete; we
494 * don't know what happened to this directory while we didn't
497 if ((issued
& CEPH_CAP_FILE_SHARED
) &&
498 (had
& CEPH_CAP_FILE_SHARED
) == 0) {
500 if (S_ISDIR(ci
->vfs_inode
.i_mode
)) {
501 dout(" marking %p NOT complete\n", &ci
->vfs_inode
);
502 __ceph_dir_clear_complete(ci
);
508 * Add a capability under the given MDS session.
510 * Caller should hold session snap_rwsem (read) and s_mutex.
512 * @fmode is the open file mode, if we are opening a file, otherwise
513 * it is < 0. (This is so we can atomically add the cap and add an
514 * open file reference to it.)
516 void ceph_add_cap(struct inode
*inode
,
517 struct ceph_mds_session
*session
, u64 cap_id
,
518 int fmode
, unsigned issued
, unsigned wanted
,
519 unsigned seq
, unsigned mseq
, u64 realmino
, int flags
,
520 struct ceph_cap
**new_cap
)
522 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
523 struct ceph_inode_info
*ci
= ceph_inode(inode
);
524 struct ceph_cap
*cap
;
525 int mds
= session
->s_mds
;
528 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode
,
529 session
->s_mds
, cap_id
, ceph_cap_string(issued
), seq
);
532 * If we are opening the file, include file mode wanted bits
536 wanted
|= ceph_caps_for_mode(fmode
);
538 cap
= __get_cap_for_mds(ci
, mds
);
544 cap
->implemented
= 0;
550 __insert_cap_node(ci
, cap
);
552 /* add to session cap list */
553 cap
->session
= session
;
554 spin_lock(&session
->s_cap_lock
);
555 list_add_tail(&cap
->session_caps
, &session
->s_caps
);
556 session
->s_nr_caps
++;
557 spin_unlock(&session
->s_cap_lock
);
560 * auth mds of the inode changed. we received the cap export
561 * message, but still haven't received the cap import message.
562 * handle_cap_export() updated the new auth MDS' cap.
564 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
565 * a message that was send before the cap import message. So
568 if (ceph_seq_cmp(seq
, cap
->seq
) <= 0) {
569 WARN_ON(cap
!= ci
->i_auth_cap
);
570 WARN_ON(cap
->cap_id
!= cap_id
);
573 issued
|= cap
->issued
;
574 flags
|= CEPH_CAP_FLAG_AUTH
;
578 if (!ci
->i_snap_realm
) {
580 * add this inode to the appropriate snap realm
582 struct ceph_snap_realm
*realm
= ceph_lookup_snap_realm(mdsc
,
585 spin_lock(&realm
->inodes_with_caps_lock
);
586 ci
->i_snap_realm
= realm
;
587 list_add(&ci
->i_snap_realm_item
,
588 &realm
->inodes_with_caps
);
589 spin_unlock(&realm
->inodes_with_caps_lock
);
591 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
597 __check_cap_issue(ci
, cap
, issued
);
600 * If we are issued caps we don't want, or the mds' wanted
601 * value appears to be off, queue a check so we'll release
602 * later and/or update the mds wanted value.
604 actual_wanted
= __ceph_caps_wanted(ci
);
605 if ((wanted
& ~actual_wanted
) ||
606 (issued
& ~actual_wanted
& CEPH_CAP_ANY_WR
)) {
607 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
608 ceph_cap_string(issued
), ceph_cap_string(wanted
),
609 ceph_cap_string(actual_wanted
));
610 __cap_delay_requeue(mdsc
, ci
);
613 if (flags
& CEPH_CAP_FLAG_AUTH
) {
614 if (ci
->i_auth_cap
== NULL
||
615 ceph_seq_cmp(ci
->i_auth_cap
->mseq
, mseq
) < 0) {
616 ci
->i_auth_cap
= cap
;
617 cap
->mds_wanted
= wanted
;
620 WARN_ON(ci
->i_auth_cap
== cap
);
623 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
624 inode
, ceph_vinop(inode
), cap
, ceph_cap_string(issued
),
625 ceph_cap_string(issued
|cap
->issued
), seq
, mds
);
626 cap
->cap_id
= cap_id
;
627 cap
->issued
= issued
;
628 cap
->implemented
|= issued
;
629 if (ceph_seq_cmp(mseq
, cap
->mseq
) > 0)
630 cap
->mds_wanted
= wanted
;
632 cap
->mds_wanted
|= wanted
;
634 cap
->issue_seq
= seq
;
636 cap
->cap_gen
= session
->s_cap_gen
;
639 __ceph_get_fmode(ci
, fmode
);
643 * Return true if cap has not timed out and belongs to the current
644 * generation of the MDS session (i.e. has not gone 'stale' due to
645 * us losing touch with the mds).
647 static int __cap_is_valid(struct ceph_cap
*cap
)
652 spin_lock(&cap
->session
->s_gen_ttl_lock
);
653 gen
= cap
->session
->s_cap_gen
;
654 ttl
= cap
->session
->s_cap_ttl
;
655 spin_unlock(&cap
->session
->s_gen_ttl_lock
);
657 if (cap
->cap_gen
< gen
|| time_after_eq(jiffies
, ttl
)) {
658 dout("__cap_is_valid %p cap %p issued %s "
659 "but STALE (gen %u vs %u)\n", &cap
->ci
->vfs_inode
,
660 cap
, ceph_cap_string(cap
->issued
), cap
->cap_gen
, gen
);
668 * Return set of valid cap bits issued to us. Note that caps time
669 * out, and may be invalidated in bulk if the client session times out
670 * and session->s_cap_gen is bumped.
672 int __ceph_caps_issued(struct ceph_inode_info
*ci
, int *implemented
)
674 int have
= ci
->i_snap_caps
;
675 struct ceph_cap
*cap
;
680 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
681 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
682 if (!__cap_is_valid(cap
))
684 dout("__ceph_caps_issued %p cap %p issued %s\n",
685 &ci
->vfs_inode
, cap
, ceph_cap_string(cap
->issued
));
688 *implemented
|= cap
->implemented
;
691 * exclude caps issued by non-auth MDS, but are been revoking
692 * by the auth MDS. The non-auth MDS should be revoking/exporting
693 * these caps, but the message is delayed.
695 if (ci
->i_auth_cap
) {
696 cap
= ci
->i_auth_cap
;
697 have
&= ~cap
->implemented
| cap
->issued
;
703 * Get cap bits issued by caps other than @ocap
705 int __ceph_caps_issued_other(struct ceph_inode_info
*ci
, struct ceph_cap
*ocap
)
707 int have
= ci
->i_snap_caps
;
708 struct ceph_cap
*cap
;
711 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
712 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
715 if (!__cap_is_valid(cap
))
723 * Move a cap to the end of the LRU (oldest caps at list head, newest
726 static void __touch_cap(struct ceph_cap
*cap
)
728 struct ceph_mds_session
*s
= cap
->session
;
730 spin_lock(&s
->s_cap_lock
);
731 if (s
->s_cap_iterator
== NULL
) {
732 dout("__touch_cap %p cap %p mds%d\n", &cap
->ci
->vfs_inode
, cap
,
734 list_move_tail(&cap
->session_caps
, &s
->s_caps
);
736 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
737 &cap
->ci
->vfs_inode
, cap
, s
->s_mds
);
739 spin_unlock(&s
->s_cap_lock
);
743 * Check if we hold the given mask. If so, move the cap(s) to the
744 * front of their respective LRUs. (This is the preferred way for
745 * callers to check for caps they want.)
747 int __ceph_caps_issued_mask(struct ceph_inode_info
*ci
, int mask
, int touch
)
749 struct ceph_cap
*cap
;
751 int have
= ci
->i_snap_caps
;
753 if ((have
& mask
) == mask
) {
754 dout("__ceph_caps_issued_mask %p snap issued %s"
755 " (mask %s)\n", &ci
->vfs_inode
,
756 ceph_cap_string(have
),
757 ceph_cap_string(mask
));
761 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
762 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
763 if (!__cap_is_valid(cap
))
765 if ((cap
->issued
& mask
) == mask
) {
766 dout("__ceph_caps_issued_mask %p cap %p issued %s"
767 " (mask %s)\n", &ci
->vfs_inode
, cap
,
768 ceph_cap_string(cap
->issued
),
769 ceph_cap_string(mask
));
775 /* does a combination of caps satisfy mask? */
777 if ((have
& mask
) == mask
) {
778 dout("__ceph_caps_issued_mask %p combo issued %s"
779 " (mask %s)\n", &ci
->vfs_inode
,
780 ceph_cap_string(cap
->issued
),
781 ceph_cap_string(mask
));
785 /* touch this + preceding caps */
787 for (q
= rb_first(&ci
->i_caps
); q
!= p
;
789 cap
= rb_entry(q
, struct ceph_cap
,
791 if (!__cap_is_valid(cap
))
804 * Return true if mask caps are currently being revoked by an MDS.
806 int __ceph_caps_revoking_other(struct ceph_inode_info
*ci
,
807 struct ceph_cap
*ocap
, int mask
)
809 struct ceph_cap
*cap
;
812 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
813 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
815 (cap
->implemented
& ~cap
->issued
& mask
))
821 int ceph_caps_revoking(struct ceph_inode_info
*ci
, int mask
)
823 struct inode
*inode
= &ci
->vfs_inode
;
826 spin_lock(&ci
->i_ceph_lock
);
827 ret
= __ceph_caps_revoking_other(ci
, NULL
, mask
);
828 spin_unlock(&ci
->i_ceph_lock
);
829 dout("ceph_caps_revoking %p %s = %d\n", inode
,
830 ceph_cap_string(mask
), ret
);
834 int __ceph_caps_used(struct ceph_inode_info
*ci
)
838 used
|= CEPH_CAP_PIN
;
840 used
|= CEPH_CAP_FILE_RD
;
841 if (ci
->i_rdcache_ref
||
842 (!S_ISDIR(ci
->vfs_inode
.i_mode
) && /* ignore readdir cache */
843 ci
->vfs_inode
.i_data
.nrpages
))
844 used
|= CEPH_CAP_FILE_CACHE
;
846 used
|= CEPH_CAP_FILE_WR
;
847 if (ci
->i_wb_ref
|| ci
->i_wrbuffer_ref
)
848 used
|= CEPH_CAP_FILE_BUFFER
;
853 * wanted, by virtue of open file modes
855 int __ceph_caps_file_wanted(struct ceph_inode_info
*ci
)
858 for (i
= 0; i
< CEPH_FILE_MODE_BITS
; i
++) {
859 if (ci
->i_nr_by_mode
[i
])
864 return ceph_caps_for_mode(bits
>> 1);
868 * Return caps we have registered with the MDS(s) as 'wanted'.
870 int __ceph_caps_mds_wanted(struct ceph_inode_info
*ci
)
872 struct ceph_cap
*cap
;
876 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
877 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
878 if (!__cap_is_valid(cap
))
880 if (cap
== ci
->i_auth_cap
)
881 mds_wanted
|= cap
->mds_wanted
;
883 mds_wanted
|= (cap
->mds_wanted
& ~CEPH_CAP_ANY_FILE_WR
);
889 * called under i_ceph_lock
891 static int __ceph_is_any_caps(struct ceph_inode_info
*ci
)
893 return !RB_EMPTY_ROOT(&ci
->i_caps
);
896 int ceph_is_any_caps(struct inode
*inode
)
898 struct ceph_inode_info
*ci
= ceph_inode(inode
);
901 spin_lock(&ci
->i_ceph_lock
);
902 ret
= __ceph_is_any_caps(ci
);
903 spin_unlock(&ci
->i_ceph_lock
);
908 static void drop_inode_snap_realm(struct ceph_inode_info
*ci
)
910 struct ceph_snap_realm
*realm
= ci
->i_snap_realm
;
911 spin_lock(&realm
->inodes_with_caps_lock
);
912 list_del_init(&ci
->i_snap_realm_item
);
913 ci
->i_snap_realm_counter
++;
914 ci
->i_snap_realm
= NULL
;
915 spin_unlock(&realm
->inodes_with_caps_lock
);
916 ceph_put_snap_realm(ceph_sb_to_client(ci
->vfs_inode
.i_sb
)->mdsc
,
921 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
923 * caller should hold i_ceph_lock.
924 * caller will not hold session s_mutex if called from destroy_inode.
926 void __ceph_remove_cap(struct ceph_cap
*cap
, bool queue_release
)
928 struct ceph_mds_session
*session
= cap
->session
;
929 struct ceph_inode_info
*ci
= cap
->ci
;
930 struct ceph_mds_client
*mdsc
=
931 ceph_sb_to_client(ci
->vfs_inode
.i_sb
)->mdsc
;
934 dout("__ceph_remove_cap %p from %p\n", cap
, &ci
->vfs_inode
);
936 /* remove from session list */
937 spin_lock(&session
->s_cap_lock
);
938 if (session
->s_cap_iterator
== cap
) {
939 /* not yet, we are iterating over this very cap */
940 dout("__ceph_remove_cap delaying %p removal from session %p\n",
943 list_del_init(&cap
->session_caps
);
944 session
->s_nr_caps
--;
948 /* protect backpointer with s_cap_lock: see iterate_session_caps */
952 * s_cap_reconnect is protected by s_cap_lock. no one changes
953 * s_cap_gen while session is in the reconnect state.
956 (!session
->s_cap_reconnect
|| cap
->cap_gen
== session
->s_cap_gen
)) {
957 cap
->queue_release
= 1;
959 list_add_tail(&cap
->session_caps
,
960 &session
->s_cap_releases
);
961 session
->s_num_cap_releases
++;
965 cap
->queue_release
= 0;
967 cap
->cap_ino
= ci
->i_vino
.ino
;
969 spin_unlock(&session
->s_cap_lock
);
971 /* remove from inode list */
972 rb_erase(&cap
->ci_node
, &ci
->i_caps
);
973 if (ci
->i_auth_cap
== cap
)
974 ci
->i_auth_cap
= NULL
;
977 ceph_put_cap(mdsc
, cap
);
979 /* when reconnect denied, we remove session caps forcibly,
980 * i_wr_ref can be non-zero. If there are ongoing write,
983 if (!__ceph_is_any_caps(ci
) && ci
->i_wr_ref
== 0 && ci
->i_snap_realm
)
984 drop_inode_snap_realm(ci
);
986 if (!__ceph_is_any_real_caps(ci
))
987 __cap_delay_cancel(mdsc
, ci
);
990 struct cap_msg_args
{
991 struct ceph_mds_session
*session
;
992 u64 ino
, cid
, follows
;
993 u64 flush_tid
, oldest_flush_tid
, size
, max_size
;
995 struct ceph_buffer
*xattr_buf
;
996 struct timespec atime
, mtime
, ctime
;
997 int op
, caps
, wanted
, dirty
;
998 u32 seq
, issue_seq
, mseq
, time_warp_seq
;
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
);
1188 cap
->issued
&= retain
; /* drop bits we don't want */
1189 if (cap
->implemented
& ~cap
->issued
) {
1191 * Wake up any waiters on wanted -> needed transition.
1192 * This is due to the weird transition from buffered
1193 * to sync IO... we need to flush dirty pages _before_
1194 * allowing sync writes to avoid reordering.
1198 cap
->implemented
&= cap
->issued
| used
;
1199 cap
->mds_wanted
= want
;
1201 arg
.ino
= ceph_vino(inode
).ino
;
1202 arg
.cid
= cap
->cap_id
;
1203 arg
.follows
= flushing
? ci
->i_head_snapc
->seq
: 0;
1204 arg
.flush_tid
= flush_tid
;
1205 arg
.oldest_flush_tid
= oldest_flush_tid
;
1207 arg
.size
= inode
->i_size
;
1208 ci
->i_reported_size
= arg
.size
;
1209 arg
.max_size
= ci
->i_wanted_max_size
;
1210 ci
->i_requested_max_size
= arg
.max_size
;
1212 if (flushing
& CEPH_CAP_XATTR_EXCL
) {
1213 __ceph_build_xattrs_blob(ci
);
1214 arg
.xattr_version
= ci
->i_xattrs
.version
;
1215 arg
.xattr_buf
= ci
->i_xattrs
.blob
;
1217 arg
.xattr_buf
= NULL
;
1220 arg
.mtime
= inode
->i_mtime
;
1221 arg
.atime
= inode
->i_atime
;
1222 arg
.ctime
= inode
->i_ctime
;
1225 arg
.caps
= cap
->implemented
;
1227 arg
.dirty
= flushing
;
1230 arg
.issue_seq
= cap
->issue_seq
;
1231 arg
.mseq
= cap
->mseq
;
1232 arg
.time_warp_seq
= ci
->i_time_warp_seq
;
1234 arg
.uid
= inode
->i_uid
;
1235 arg
.gid
= inode
->i_gid
;
1236 arg
.mode
= inode
->i_mode
;
1238 arg
.inline_data
= ci
->i_inline_version
!= CEPH_INLINE_NONE
;
1241 arg
.flags
|= CEPH_CLIENT_CAPS_SYNC
;
1243 spin_unlock(&ci
->i_ceph_lock
);
1245 ret
= send_cap_msg(&arg
);
1247 dout("error sending cap msg, must requeue %p\n", inode
);
1252 wake_up_all(&ci
->i_cap_wq
);
1257 static inline int __send_flush_snap(struct inode
*inode
,
1258 struct ceph_mds_session
*session
,
1259 struct ceph_cap_snap
*capsnap
,
1260 u32 mseq
, u64 oldest_flush_tid
)
1262 struct cap_msg_args arg
;
1264 arg
.session
= session
;
1265 arg
.ino
= ceph_vino(inode
).ino
;
1267 arg
.follows
= capsnap
->follows
;
1268 arg
.flush_tid
= capsnap
->cap_flush
.tid
;
1269 arg
.oldest_flush_tid
= oldest_flush_tid
;
1271 arg
.size
= capsnap
->size
;
1273 arg
.xattr_version
= capsnap
->xattr_version
;
1274 arg
.xattr_buf
= capsnap
->xattr_blob
;
1276 arg
.atime
= capsnap
->atime
;
1277 arg
.mtime
= capsnap
->mtime
;
1278 arg
.ctime
= capsnap
->ctime
;
1280 arg
.op
= CEPH_CAP_OP_FLUSHSNAP
;
1281 arg
.caps
= capsnap
->issued
;
1283 arg
.dirty
= capsnap
->dirty
;
1288 arg
.time_warp_seq
= capsnap
->time_warp_seq
;
1290 arg
.uid
= capsnap
->uid
;
1291 arg
.gid
= capsnap
->gid
;
1292 arg
.mode
= capsnap
->mode
;
1294 arg
.inline_data
= capsnap
->inline_data
;
1297 return send_cap_msg(&arg
);
1301 * When a snapshot is taken, clients accumulate dirty metadata on
1302 * inodes with capabilities in ceph_cap_snaps to describe the file
1303 * state at the time the snapshot was taken. This must be flushed
1304 * asynchronously back to the MDS once sync writes complete and dirty
1305 * data is written out.
1307 * Called under i_ceph_lock. Takes s_mutex as needed.
1309 static void __ceph_flush_snaps(struct ceph_inode_info
*ci
,
1310 struct ceph_mds_session
*session
)
1311 __releases(ci
->i_ceph_lock
)
1312 __acquires(ci
->i_ceph_lock
)
1314 struct inode
*inode
= &ci
->vfs_inode
;
1315 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
1316 struct ceph_cap_snap
*capsnap
;
1317 u64 oldest_flush_tid
= 0;
1318 u64 first_tid
= 1, last_tid
= 0;
1320 dout("__flush_snaps %p session %p\n", inode
, session
);
1322 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
1324 * we need to wait for sync writes to complete and for dirty
1325 * pages to be written out.
1327 if (capsnap
->dirty_pages
|| capsnap
->writing
)
1330 /* should be removed by ceph_try_drop_cap_snap() */
1331 BUG_ON(!capsnap
->need_flush
);
1333 /* only flush each capsnap once */
1334 if (capsnap
->cap_flush
.tid
> 0) {
1335 dout(" already flushed %p, skipping\n", capsnap
);
1339 spin_lock(&mdsc
->cap_dirty_lock
);
1340 capsnap
->cap_flush
.tid
= ++mdsc
->last_cap_flush_tid
;
1341 list_add_tail(&capsnap
->cap_flush
.g_list
,
1342 &mdsc
->cap_flush_list
);
1343 if (oldest_flush_tid
== 0)
1344 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1345 if (list_empty(&ci
->i_flushing_item
)) {
1346 list_add_tail(&ci
->i_flushing_item
,
1347 &session
->s_cap_flushing
);
1349 spin_unlock(&mdsc
->cap_dirty_lock
);
1351 list_add_tail(&capsnap
->cap_flush
.i_list
,
1352 &ci
->i_cap_flush_list
);
1355 first_tid
= capsnap
->cap_flush
.tid
;
1356 last_tid
= capsnap
->cap_flush
.tid
;
1359 ci
->i_ceph_flags
&= ~CEPH_I_FLUSH_SNAPS
;
1361 while (first_tid
<= last_tid
) {
1362 struct ceph_cap
*cap
= ci
->i_auth_cap
;
1363 struct ceph_cap_flush
*cf
;
1366 if (!(cap
&& cap
->session
== session
)) {
1367 dout("__flush_snaps %p auth cap %p not mds%d, "
1368 "stop\n", inode
, cap
, session
->s_mds
);
1373 list_for_each_entry(cf
, &ci
->i_cap_flush_list
, i_list
) {
1374 if (cf
->tid
>= first_tid
) {
1382 first_tid
= cf
->tid
+ 1;
1384 capsnap
= container_of(cf
, struct ceph_cap_snap
, cap_flush
);
1385 atomic_inc(&capsnap
->nref
);
1386 spin_unlock(&ci
->i_ceph_lock
);
1388 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1389 inode
, capsnap
, cf
->tid
, ceph_cap_string(capsnap
->dirty
));
1391 ret
= __send_flush_snap(inode
, session
, capsnap
, cap
->mseq
,
1394 pr_err("__flush_snaps: error sending cap flushsnap, "
1395 "ino (%llx.%llx) tid %llu follows %llu\n",
1396 ceph_vinop(inode
), cf
->tid
, capsnap
->follows
);
1399 ceph_put_cap_snap(capsnap
);
1400 spin_lock(&ci
->i_ceph_lock
);
1404 void ceph_flush_snaps(struct ceph_inode_info
*ci
,
1405 struct ceph_mds_session
**psession
)
1407 struct inode
*inode
= &ci
->vfs_inode
;
1408 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
1409 struct ceph_mds_session
*session
= NULL
;
1412 dout("ceph_flush_snaps %p\n", inode
);
1414 session
= *psession
;
1416 spin_lock(&ci
->i_ceph_lock
);
1417 if (!(ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
)) {
1418 dout(" no capsnap needs flush, doing nothing\n");
1421 if (!ci
->i_auth_cap
) {
1422 dout(" no auth cap (migrating?), doing nothing\n");
1426 mds
= ci
->i_auth_cap
->session
->s_mds
;
1427 if (session
&& session
->s_mds
!= mds
) {
1428 dout(" oops, wrong session %p mutex\n", session
);
1429 mutex_unlock(&session
->s_mutex
);
1430 ceph_put_mds_session(session
);
1434 spin_unlock(&ci
->i_ceph_lock
);
1435 mutex_lock(&mdsc
->mutex
);
1436 session
= __ceph_lookup_mds_session(mdsc
, mds
);
1437 mutex_unlock(&mdsc
->mutex
);
1439 dout(" inverting session/ino locks on %p\n", session
);
1440 mutex_lock(&session
->s_mutex
);
1445 __ceph_flush_snaps(ci
, session
);
1447 spin_unlock(&ci
->i_ceph_lock
);
1450 *psession
= session
;
1452 mutex_unlock(&session
->s_mutex
);
1453 ceph_put_mds_session(session
);
1455 /* we flushed them all; remove this inode from the queue */
1456 spin_lock(&mdsc
->snap_flush_lock
);
1457 list_del_init(&ci
->i_snap_flush_item
);
1458 spin_unlock(&mdsc
->snap_flush_lock
);
1462 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1463 * Caller is then responsible for calling __mark_inode_dirty with the
1464 * returned flags value.
1466 int __ceph_mark_dirty_caps(struct ceph_inode_info
*ci
, int mask
,
1467 struct ceph_cap_flush
**pcf
)
1469 struct ceph_mds_client
*mdsc
=
1470 ceph_sb_to_client(ci
->vfs_inode
.i_sb
)->mdsc
;
1471 struct inode
*inode
= &ci
->vfs_inode
;
1472 int was
= ci
->i_dirty_caps
;
1475 if (!ci
->i_auth_cap
) {
1476 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1477 "but no auth cap (session was closed?)\n",
1478 inode
, ceph_ino(inode
), ceph_cap_string(mask
));
1482 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci
->vfs_inode
,
1483 ceph_cap_string(mask
), ceph_cap_string(was
),
1484 ceph_cap_string(was
| mask
));
1485 ci
->i_dirty_caps
|= mask
;
1487 WARN_ON_ONCE(ci
->i_prealloc_cap_flush
);
1488 swap(ci
->i_prealloc_cap_flush
, *pcf
);
1490 if (!ci
->i_head_snapc
) {
1491 WARN_ON_ONCE(!rwsem_is_locked(&mdsc
->snap_rwsem
));
1492 ci
->i_head_snapc
= ceph_get_snap_context(
1493 ci
->i_snap_realm
->cached_context
);
1495 dout(" inode %p now dirty snapc %p auth cap %p\n",
1496 &ci
->vfs_inode
, ci
->i_head_snapc
, ci
->i_auth_cap
);
1497 BUG_ON(!list_empty(&ci
->i_dirty_item
));
1498 spin_lock(&mdsc
->cap_dirty_lock
);
1499 list_add(&ci
->i_dirty_item
, &mdsc
->cap_dirty
);
1500 spin_unlock(&mdsc
->cap_dirty_lock
);
1501 if (ci
->i_flushing_caps
== 0) {
1503 dirty
|= I_DIRTY_SYNC
;
1506 WARN_ON_ONCE(!ci
->i_prealloc_cap_flush
);
1508 BUG_ON(list_empty(&ci
->i_dirty_item
));
1509 if (((was
| ci
->i_flushing_caps
) & CEPH_CAP_FILE_BUFFER
) &&
1510 (mask
& CEPH_CAP_FILE_BUFFER
))
1511 dirty
|= I_DIRTY_DATASYNC
;
1512 __cap_delay_requeue(mdsc
, ci
);
1516 struct ceph_cap_flush
*ceph_alloc_cap_flush(void)
1518 return kmem_cache_alloc(ceph_cap_flush_cachep
, GFP_KERNEL
);
1521 void ceph_free_cap_flush(struct ceph_cap_flush
*cf
)
1524 kmem_cache_free(ceph_cap_flush_cachep
, cf
);
1527 static u64
__get_oldest_flush_tid(struct ceph_mds_client
*mdsc
)
1529 if (!list_empty(&mdsc
->cap_flush_list
)) {
1530 struct ceph_cap_flush
*cf
=
1531 list_first_entry(&mdsc
->cap_flush_list
,
1532 struct ceph_cap_flush
, g_list
);
1539 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1540 * Return true if caller needs to wake up flush waiters.
1542 static bool __finish_cap_flush(struct ceph_mds_client
*mdsc
,
1543 struct ceph_inode_info
*ci
,
1544 struct ceph_cap_flush
*cf
)
1546 struct ceph_cap_flush
*prev
;
1547 bool wake
= cf
->wake
;
1549 /* are there older pending cap flushes? */
1550 if (wake
&& cf
->g_list
.prev
!= &mdsc
->cap_flush_list
) {
1551 prev
= list_prev_entry(cf
, g_list
);
1555 list_del(&cf
->g_list
);
1557 if (wake
&& cf
->i_list
.prev
!= &ci
->i_cap_flush_list
) {
1558 prev
= list_prev_entry(cf
, i_list
);
1562 list_del(&cf
->i_list
);
1570 * Add dirty inode to the flushing list. Assigned a seq number so we
1571 * can wait for caps to flush without starving.
1573 * Called under i_ceph_lock.
1575 static int __mark_caps_flushing(struct inode
*inode
,
1576 struct ceph_mds_session
*session
, bool wake
,
1577 u64
*flush_tid
, u64
*oldest_flush_tid
)
1579 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
1580 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1581 struct ceph_cap_flush
*cf
= NULL
;
1584 BUG_ON(ci
->i_dirty_caps
== 0);
1585 BUG_ON(list_empty(&ci
->i_dirty_item
));
1586 BUG_ON(!ci
->i_prealloc_cap_flush
);
1588 flushing
= ci
->i_dirty_caps
;
1589 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1590 ceph_cap_string(flushing
),
1591 ceph_cap_string(ci
->i_flushing_caps
),
1592 ceph_cap_string(ci
->i_flushing_caps
| flushing
));
1593 ci
->i_flushing_caps
|= flushing
;
1594 ci
->i_dirty_caps
= 0;
1595 dout(" inode %p now !dirty\n", inode
);
1597 swap(cf
, ci
->i_prealloc_cap_flush
);
1598 cf
->caps
= flushing
;
1601 spin_lock(&mdsc
->cap_dirty_lock
);
1602 list_del_init(&ci
->i_dirty_item
);
1604 cf
->tid
= ++mdsc
->last_cap_flush_tid
;
1605 list_add_tail(&cf
->g_list
, &mdsc
->cap_flush_list
);
1606 *oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1608 if (list_empty(&ci
->i_flushing_item
)) {
1609 list_add_tail(&ci
->i_flushing_item
, &session
->s_cap_flushing
);
1610 mdsc
->num_cap_flushing
++;
1612 spin_unlock(&mdsc
->cap_dirty_lock
);
1614 list_add_tail(&cf
->i_list
, &ci
->i_cap_flush_list
);
1616 *flush_tid
= cf
->tid
;
1621 * try to invalidate mapping pages without blocking.
1623 static int try_nonblocking_invalidate(struct inode
*inode
)
1625 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1626 u32 invalidating_gen
= ci
->i_rdcache_gen
;
1628 spin_unlock(&ci
->i_ceph_lock
);
1629 invalidate_mapping_pages(&inode
->i_data
, 0, -1);
1630 spin_lock(&ci
->i_ceph_lock
);
1632 if (inode
->i_data
.nrpages
== 0 &&
1633 invalidating_gen
== ci
->i_rdcache_gen
) {
1635 dout("try_nonblocking_invalidate %p success\n", inode
);
1636 /* save any racing async invalidate some trouble */
1637 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
- 1;
1640 dout("try_nonblocking_invalidate %p failed\n", inode
);
1645 * Swiss army knife function to examine currently used and wanted
1646 * versus held caps. Release, flush, ack revoked caps to mds as
1649 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1650 * cap release further.
1651 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1652 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1655 void ceph_check_caps(struct ceph_inode_info
*ci
, int flags
,
1656 struct ceph_mds_session
*session
)
1658 struct ceph_fs_client
*fsc
= ceph_inode_to_client(&ci
->vfs_inode
);
1659 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
1660 struct inode
*inode
= &ci
->vfs_inode
;
1661 struct ceph_cap
*cap
;
1662 u64 flush_tid
, oldest_flush_tid
;
1663 int file_wanted
, used
, cap_used
;
1664 int took_snap_rwsem
= 0; /* true if mdsc->snap_rwsem held */
1665 int issued
, implemented
, want
, retain
, revoking
, flushing
= 0;
1666 int mds
= -1; /* keep track of how far we've gone through i_caps list
1667 to avoid an infinite loop on retry */
1669 int delayed
= 0, sent
= 0, num
;
1670 bool is_delayed
= flags
& CHECK_CAPS_NODELAY
;
1671 bool queue_invalidate
= false;
1672 bool force_requeue
= false;
1673 bool tried_invalidate
= false;
1675 /* if we are unmounting, flush any unused caps immediately. */
1679 spin_lock(&ci
->i_ceph_lock
);
1681 if (ci
->i_ceph_flags
& CEPH_I_FLUSH
)
1682 flags
|= CHECK_CAPS_FLUSH
;
1686 spin_lock(&ci
->i_ceph_lock
);
1688 file_wanted
= __ceph_caps_file_wanted(ci
);
1689 used
= __ceph_caps_used(ci
);
1690 issued
= __ceph_caps_issued(ci
, &implemented
);
1691 revoking
= implemented
& ~issued
;
1694 retain
= file_wanted
| used
| CEPH_CAP_PIN
;
1695 if (!mdsc
->stopping
&& inode
->i_nlink
> 0) {
1697 retain
|= CEPH_CAP_ANY
; /* be greedy */
1698 } else if (S_ISDIR(inode
->i_mode
) &&
1699 (issued
& CEPH_CAP_FILE_SHARED
) &&
1700 __ceph_dir_is_complete(ci
)) {
1702 * If a directory is complete, we want to keep
1703 * the exclusive cap. So that MDS does not end up
1704 * revoking the shared cap on every create/unlink
1707 want
= CEPH_CAP_ANY_SHARED
| CEPH_CAP_FILE_EXCL
;
1711 retain
|= CEPH_CAP_ANY_SHARED
;
1713 * keep RD only if we didn't have the file open RW,
1714 * because then the mds would revoke it anyway to
1715 * journal max_size=0.
1717 if (ci
->i_max_size
== 0)
1718 retain
|= CEPH_CAP_ANY_RD
;
1722 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1723 " issued %s revoking %s retain %s %s%s%s\n", inode
,
1724 ceph_cap_string(file_wanted
),
1725 ceph_cap_string(used
), ceph_cap_string(ci
->i_dirty_caps
),
1726 ceph_cap_string(ci
->i_flushing_caps
),
1727 ceph_cap_string(issued
), ceph_cap_string(revoking
),
1728 ceph_cap_string(retain
),
1729 (flags
& CHECK_CAPS_AUTHONLY
) ? " AUTHONLY" : "",
1730 (flags
& CHECK_CAPS_NODELAY
) ? " NODELAY" : "",
1731 (flags
& CHECK_CAPS_FLUSH
) ? " FLUSH" : "");
1734 * If we no longer need to hold onto old our caps, and we may
1735 * have cached pages, but don't want them, then try to invalidate.
1736 * If we fail, it's because pages are locked.... try again later.
1738 if ((!is_delayed
|| mdsc
->stopping
) &&
1739 !S_ISDIR(inode
->i_mode
) && /* ignore readdir cache */
1740 !(ci
->i_wb_ref
|| ci
->i_wrbuffer_ref
) && /* no dirty pages... */
1741 inode
->i_data
.nrpages
&& /* have cached pages */
1742 (revoking
& (CEPH_CAP_FILE_CACHE
|
1743 CEPH_CAP_FILE_LAZYIO
)) && /* or revoking cache */
1744 !tried_invalidate
) {
1745 dout("check_caps trying to invalidate on %p\n", inode
);
1746 if (try_nonblocking_invalidate(inode
) < 0) {
1747 if (revoking
& (CEPH_CAP_FILE_CACHE
|
1748 CEPH_CAP_FILE_LAZYIO
)) {
1749 dout("check_caps queuing invalidate\n");
1750 queue_invalidate
= true;
1751 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
;
1753 dout("check_caps failed to invalidate pages\n");
1754 /* we failed to invalidate pages. check these
1755 caps again later. */
1756 force_requeue
= true;
1757 __cap_set_timeouts(mdsc
, ci
);
1760 tried_invalidate
= true;
1765 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
1766 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
1769 /* avoid looping forever */
1770 if (mds
>= cap
->mds
||
1771 ((flags
& CHECK_CAPS_AUTHONLY
) && cap
!= ci
->i_auth_cap
))
1774 /* NOTE: no side-effects allowed, until we take s_mutex */
1777 if (ci
->i_auth_cap
&& cap
!= ci
->i_auth_cap
)
1778 cap_used
&= ~ci
->i_auth_cap
->issued
;
1780 revoking
= cap
->implemented
& ~cap
->issued
;
1781 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1782 cap
->mds
, cap
, ceph_cap_string(cap_used
),
1783 ceph_cap_string(cap
->issued
),
1784 ceph_cap_string(cap
->implemented
),
1785 ceph_cap_string(revoking
));
1787 if (cap
== ci
->i_auth_cap
&&
1788 (cap
->issued
& CEPH_CAP_FILE_WR
)) {
1789 /* request larger max_size from MDS? */
1790 if (ci
->i_wanted_max_size
> ci
->i_max_size
&&
1791 ci
->i_wanted_max_size
> ci
->i_requested_max_size
) {
1792 dout("requesting new max_size\n");
1796 /* approaching file_max? */
1797 if ((inode
->i_size
<< 1) >= ci
->i_max_size
&&
1798 (ci
->i_reported_size
<< 1) < ci
->i_max_size
) {
1799 dout("i_size approaching max_size\n");
1803 /* flush anything dirty? */
1804 if (cap
== ci
->i_auth_cap
) {
1805 if ((flags
& CHECK_CAPS_FLUSH
) && ci
->i_dirty_caps
) {
1806 dout("flushing dirty caps\n");
1809 if (ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
) {
1810 dout("flushing snap caps\n");
1815 /* completed revocation? going down and there are no caps? */
1816 if (revoking
&& (revoking
& cap_used
) == 0) {
1817 dout("completed revocation of %s\n",
1818 ceph_cap_string(cap
->implemented
& ~cap
->issued
));
1822 /* want more caps from mds? */
1823 if (want
& ~(cap
->mds_wanted
| cap
->issued
))
1826 /* things we might delay */
1827 if ((cap
->issued
& ~retain
) == 0 &&
1828 cap
->mds_wanted
== want
)
1829 continue; /* nope, all good */
1835 if ((ci
->i_ceph_flags
& CEPH_I_NODELAY
) == 0 &&
1836 time_before(jiffies
, ci
->i_hold_caps_max
)) {
1837 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1838 ceph_cap_string(cap
->issued
),
1839 ceph_cap_string(cap
->issued
& retain
),
1840 ceph_cap_string(cap
->mds_wanted
),
1841 ceph_cap_string(want
));
1847 if (ci
->i_ceph_flags
& CEPH_I_NOFLUSH
) {
1848 dout(" skipping %p I_NOFLUSH set\n", inode
);
1852 if (session
&& session
!= cap
->session
) {
1853 dout("oops, wrong session %p mutex\n", session
);
1854 mutex_unlock(&session
->s_mutex
);
1858 session
= cap
->session
;
1859 if (mutex_trylock(&session
->s_mutex
) == 0) {
1860 dout("inverting session/ino locks on %p\n",
1862 spin_unlock(&ci
->i_ceph_lock
);
1863 if (took_snap_rwsem
) {
1864 up_read(&mdsc
->snap_rwsem
);
1865 took_snap_rwsem
= 0;
1867 mutex_lock(&session
->s_mutex
);
1872 /* kick flushing and flush snaps before sending normal
1874 if (cap
== ci
->i_auth_cap
&&
1876 (CEPH_I_KICK_FLUSH
| CEPH_I_FLUSH_SNAPS
))) {
1877 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
) {
1878 spin_lock(&mdsc
->cap_dirty_lock
);
1879 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1880 spin_unlock(&mdsc
->cap_dirty_lock
);
1881 __kick_flushing_caps(mdsc
, session
, ci
,
1883 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
1885 if (ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
)
1886 __ceph_flush_snaps(ci
, session
);
1891 /* take snap_rwsem after session mutex */
1892 if (!took_snap_rwsem
) {
1893 if (down_read_trylock(&mdsc
->snap_rwsem
) == 0) {
1894 dout("inverting snap/in locks on %p\n",
1896 spin_unlock(&ci
->i_ceph_lock
);
1897 down_read(&mdsc
->snap_rwsem
);
1898 took_snap_rwsem
= 1;
1901 took_snap_rwsem
= 1;
1904 if (cap
== ci
->i_auth_cap
&& ci
->i_dirty_caps
) {
1905 flushing
= __mark_caps_flushing(inode
, session
, false,
1911 spin_lock(&mdsc
->cap_dirty_lock
);
1912 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1913 spin_unlock(&mdsc
->cap_dirty_lock
);
1916 mds
= cap
->mds
; /* remember mds, so we don't repeat */
1919 /* __send_cap drops i_ceph_lock */
1920 delayed
+= __send_cap(mdsc
, cap
, CEPH_CAP_OP_UPDATE
, false,
1921 cap_used
, want
, retain
, flushing
,
1922 flush_tid
, oldest_flush_tid
);
1923 goto retry
; /* retake i_ceph_lock and restart our cap scan. */
1927 * Reschedule delayed caps release if we delayed anything,
1930 if (delayed
&& is_delayed
)
1931 force_requeue
= true; /* __send_cap delayed release; requeue */
1932 if (!delayed
&& !is_delayed
)
1933 __cap_delay_cancel(mdsc
, ci
);
1934 else if (!is_delayed
|| force_requeue
)
1935 __cap_delay_requeue(mdsc
, ci
);
1937 spin_unlock(&ci
->i_ceph_lock
);
1939 if (queue_invalidate
)
1940 ceph_queue_invalidate(inode
);
1943 mutex_unlock(&session
->s_mutex
);
1944 if (took_snap_rwsem
)
1945 up_read(&mdsc
->snap_rwsem
);
1949 * Try to flush dirty caps back to the auth mds.
1951 static int try_flush_caps(struct inode
*inode
, u64
*ptid
)
1953 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
1954 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1955 struct ceph_mds_session
*session
= NULL
;
1957 u64 flush_tid
= 0, oldest_flush_tid
= 0;
1960 spin_lock(&ci
->i_ceph_lock
);
1961 if (ci
->i_ceph_flags
& CEPH_I_NOFLUSH
) {
1962 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode
);
1965 if (ci
->i_dirty_caps
&& ci
->i_auth_cap
) {
1966 struct ceph_cap
*cap
= ci
->i_auth_cap
;
1967 int used
= __ceph_caps_used(ci
);
1968 int want
= __ceph_caps_wanted(ci
);
1971 if (!session
|| session
!= cap
->session
) {
1972 spin_unlock(&ci
->i_ceph_lock
);
1974 mutex_unlock(&session
->s_mutex
);
1975 session
= cap
->session
;
1976 mutex_lock(&session
->s_mutex
);
1979 if (cap
->session
->s_state
< CEPH_MDS_SESSION_OPEN
)
1982 flushing
= __mark_caps_flushing(inode
, session
, true,
1983 &flush_tid
, &oldest_flush_tid
);
1985 /* __send_cap drops i_ceph_lock */
1986 delayed
= __send_cap(mdsc
, cap
, CEPH_CAP_OP_FLUSH
, true,
1987 used
, want
, (cap
->issued
| cap
->implemented
),
1988 flushing
, flush_tid
, oldest_flush_tid
);
1991 spin_lock(&ci
->i_ceph_lock
);
1992 __cap_delay_requeue(mdsc
, ci
);
1993 spin_unlock(&ci
->i_ceph_lock
);
1996 if (!list_empty(&ci
->i_cap_flush_list
)) {
1997 struct ceph_cap_flush
*cf
=
1998 list_last_entry(&ci
->i_cap_flush_list
,
1999 struct ceph_cap_flush
, i_list
);
2001 flush_tid
= cf
->tid
;
2003 flushing
= ci
->i_flushing_caps
;
2004 spin_unlock(&ci
->i_ceph_lock
);
2008 mutex_unlock(&session
->s_mutex
);
2015 * Return true if we've flushed caps through the given flush_tid.
2017 static int caps_are_flushed(struct inode
*inode
, u64 flush_tid
)
2019 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2022 spin_lock(&ci
->i_ceph_lock
);
2023 if (!list_empty(&ci
->i_cap_flush_list
)) {
2024 struct ceph_cap_flush
* cf
=
2025 list_first_entry(&ci
->i_cap_flush_list
,
2026 struct ceph_cap_flush
, i_list
);
2027 if (cf
->tid
<= flush_tid
)
2030 spin_unlock(&ci
->i_ceph_lock
);
2035 * wait for any unsafe requests to complete.
2037 static int unsafe_request_wait(struct inode
*inode
)
2039 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2040 struct ceph_mds_request
*req1
= NULL
, *req2
= NULL
;
2043 spin_lock(&ci
->i_unsafe_lock
);
2044 if (S_ISDIR(inode
->i_mode
) && !list_empty(&ci
->i_unsafe_dirops
)) {
2045 req1
= list_last_entry(&ci
->i_unsafe_dirops
,
2046 struct ceph_mds_request
,
2048 ceph_mdsc_get_request(req1
);
2050 if (!list_empty(&ci
->i_unsafe_iops
)) {
2051 req2
= list_last_entry(&ci
->i_unsafe_iops
,
2052 struct ceph_mds_request
,
2053 r_unsafe_target_item
);
2054 ceph_mdsc_get_request(req2
);
2056 spin_unlock(&ci
->i_unsafe_lock
);
2058 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2059 inode
, req1
? req1
->r_tid
: 0ULL, req2
? req2
->r_tid
: 0ULL);
2061 ret
= !wait_for_completion_timeout(&req1
->r_safe_completion
,
2062 ceph_timeout_jiffies(req1
->r_timeout
));
2065 ceph_mdsc_put_request(req1
);
2068 ret
= !wait_for_completion_timeout(&req2
->r_safe_completion
,
2069 ceph_timeout_jiffies(req2
->r_timeout
));
2072 ceph_mdsc_put_request(req2
);
2077 int ceph_fsync(struct file
*file
, loff_t start
, loff_t end
, int datasync
)
2079 struct inode
*inode
= file
->f_mapping
->host
;
2080 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2085 dout("fsync %p%s\n", inode
, datasync
? " datasync" : "");
2087 ceph_sync_write_wait(inode
);
2089 ret
= filemap_write_and_wait_range(inode
->i_mapping
, start
, end
);
2098 dirty
= try_flush_caps(inode
, &flush_tid
);
2099 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty
));
2101 ret
= unsafe_request_wait(inode
);
2104 * only wait on non-file metadata writeback (the mds
2105 * can recover size and mtime, so we don't need to
2108 if (!ret
&& (dirty
& ~CEPH_CAP_ANY_FILE_WR
)) {
2109 ret
= wait_event_interruptible(ci
->i_cap_wq
,
2110 caps_are_flushed(inode
, flush_tid
));
2112 inode_unlock(inode
);
2114 dout("fsync %p%s result=%d\n", inode
, datasync
? " datasync" : "", ret
);
2119 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2120 * queue inode for flush but don't do so immediately, because we can
2121 * get by with fewer MDS messages if we wait for data writeback to
2124 int ceph_write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
2126 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2130 int wait
= wbc
->sync_mode
== WB_SYNC_ALL
;
2132 dout("write_inode %p wait=%d\n", inode
, wait
);
2134 dirty
= try_flush_caps(inode
, &flush_tid
);
2136 err
= wait_event_interruptible(ci
->i_cap_wq
,
2137 caps_are_flushed(inode
, flush_tid
));
2139 struct ceph_mds_client
*mdsc
=
2140 ceph_sb_to_client(inode
->i_sb
)->mdsc
;
2142 spin_lock(&ci
->i_ceph_lock
);
2143 if (__ceph_caps_dirty(ci
))
2144 __cap_delay_requeue_front(mdsc
, ci
);
2145 spin_unlock(&ci
->i_ceph_lock
);
2150 static void __kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2151 struct ceph_mds_session
*session
,
2152 struct ceph_inode_info
*ci
,
2153 u64 oldest_flush_tid
)
2154 __releases(ci
->i_ceph_lock
)
2155 __acquires(ci
->i_ceph_lock
)
2157 struct inode
*inode
= &ci
->vfs_inode
;
2158 struct ceph_cap
*cap
;
2159 struct ceph_cap_flush
*cf
;
2163 list_for_each_entry(cf
, &ci
->i_cap_flush_list
, i_list
) {
2164 if (cf
->tid
< first_tid
)
2167 cap
= ci
->i_auth_cap
;
2168 if (!(cap
&& cap
->session
== session
)) {
2169 pr_err("%p auth cap %p not mds%d ???\n",
2170 inode
, cap
, session
->s_mds
);
2174 first_tid
= cf
->tid
+ 1;
2177 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2178 inode
, cap
, cf
->tid
, ceph_cap_string(cf
->caps
));
2179 ci
->i_ceph_flags
|= CEPH_I_NODELAY
;
2180 ret
= __send_cap(mdsc
, cap
, CEPH_CAP_OP_FLUSH
,
2181 false, __ceph_caps_used(ci
),
2182 __ceph_caps_wanted(ci
),
2183 cap
->issued
| cap
->implemented
,
2184 cf
->caps
, cf
->tid
, oldest_flush_tid
);
2186 pr_err("kick_flushing_caps: error sending "
2187 "cap flush, ino (%llx.%llx) "
2188 "tid %llu flushing %s\n",
2189 ceph_vinop(inode
), cf
->tid
,
2190 ceph_cap_string(cf
->caps
));
2193 struct ceph_cap_snap
*capsnap
=
2194 container_of(cf
, struct ceph_cap_snap
,
2196 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2197 inode
, capsnap
, cf
->tid
,
2198 ceph_cap_string(capsnap
->dirty
));
2200 atomic_inc(&capsnap
->nref
);
2201 spin_unlock(&ci
->i_ceph_lock
);
2203 ret
= __send_flush_snap(inode
, session
, capsnap
, cap
->mseq
,
2206 pr_err("kick_flushing_caps: error sending "
2207 "cap flushsnap, ino (%llx.%llx) "
2208 "tid %llu follows %llu\n",
2209 ceph_vinop(inode
), cf
->tid
,
2213 ceph_put_cap_snap(capsnap
);
2216 spin_lock(&ci
->i_ceph_lock
);
2220 void ceph_early_kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2221 struct ceph_mds_session
*session
)
2223 struct ceph_inode_info
*ci
;
2224 struct ceph_cap
*cap
;
2225 u64 oldest_flush_tid
;
2227 dout("early_kick_flushing_caps mds%d\n", session
->s_mds
);
2229 spin_lock(&mdsc
->cap_dirty_lock
);
2230 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2231 spin_unlock(&mdsc
->cap_dirty_lock
);
2233 list_for_each_entry(ci
, &session
->s_cap_flushing
, i_flushing_item
) {
2234 spin_lock(&ci
->i_ceph_lock
);
2235 cap
= ci
->i_auth_cap
;
2236 if (!(cap
&& cap
->session
== session
)) {
2237 pr_err("%p auth cap %p not mds%d ???\n",
2238 &ci
->vfs_inode
, cap
, session
->s_mds
);
2239 spin_unlock(&ci
->i_ceph_lock
);
2245 * if flushing caps were revoked, we re-send the cap flush
2246 * in client reconnect stage. This guarantees MDS * processes
2247 * the cap flush message before issuing the flushing caps to
2250 if ((cap
->issued
& ci
->i_flushing_caps
) !=
2251 ci
->i_flushing_caps
) {
2252 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
2253 __kick_flushing_caps(mdsc
, session
, ci
,
2256 ci
->i_ceph_flags
|= CEPH_I_KICK_FLUSH
;
2259 spin_unlock(&ci
->i_ceph_lock
);
2263 void ceph_kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2264 struct ceph_mds_session
*session
)
2266 struct ceph_inode_info
*ci
;
2267 struct ceph_cap
*cap
;
2268 u64 oldest_flush_tid
;
2270 dout("kick_flushing_caps mds%d\n", session
->s_mds
);
2272 spin_lock(&mdsc
->cap_dirty_lock
);
2273 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2274 spin_unlock(&mdsc
->cap_dirty_lock
);
2276 list_for_each_entry(ci
, &session
->s_cap_flushing
, i_flushing_item
) {
2277 spin_lock(&ci
->i_ceph_lock
);
2278 cap
= ci
->i_auth_cap
;
2279 if (!(cap
&& cap
->session
== session
)) {
2280 pr_err("%p auth cap %p not mds%d ???\n",
2281 &ci
->vfs_inode
, cap
, session
->s_mds
);
2282 spin_unlock(&ci
->i_ceph_lock
);
2285 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
) {
2286 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
2287 __kick_flushing_caps(mdsc
, session
, ci
,
2290 spin_unlock(&ci
->i_ceph_lock
);
2294 static void kick_flushing_inode_caps(struct ceph_mds_client
*mdsc
,
2295 struct ceph_mds_session
*session
,
2296 struct inode
*inode
)
2297 __releases(ci
->i_ceph_lock
)
2299 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2300 struct ceph_cap
*cap
;
2302 cap
= ci
->i_auth_cap
;
2303 dout("kick_flushing_inode_caps %p flushing %s\n", inode
,
2304 ceph_cap_string(ci
->i_flushing_caps
));
2306 if (!list_empty(&ci
->i_cap_flush_list
)) {
2307 u64 oldest_flush_tid
;
2308 spin_lock(&mdsc
->cap_dirty_lock
);
2309 list_move_tail(&ci
->i_flushing_item
,
2310 &cap
->session
->s_cap_flushing
);
2311 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2312 spin_unlock(&mdsc
->cap_dirty_lock
);
2314 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
2315 __kick_flushing_caps(mdsc
, session
, ci
, oldest_flush_tid
);
2316 spin_unlock(&ci
->i_ceph_lock
);
2318 spin_unlock(&ci
->i_ceph_lock
);
2324 * Take references to capabilities we hold, so that we don't release
2325 * them to the MDS prematurely.
2327 * Protected by i_ceph_lock.
2329 static void __take_cap_refs(struct ceph_inode_info
*ci
, int got
,
2330 bool snap_rwsem_locked
)
2332 if (got
& CEPH_CAP_PIN
)
2334 if (got
& CEPH_CAP_FILE_RD
)
2336 if (got
& CEPH_CAP_FILE_CACHE
)
2337 ci
->i_rdcache_ref
++;
2338 if (got
& CEPH_CAP_FILE_WR
) {
2339 if (ci
->i_wr_ref
== 0 && !ci
->i_head_snapc
) {
2340 BUG_ON(!snap_rwsem_locked
);
2341 ci
->i_head_snapc
= ceph_get_snap_context(
2342 ci
->i_snap_realm
->cached_context
);
2346 if (got
& CEPH_CAP_FILE_BUFFER
) {
2347 if (ci
->i_wb_ref
== 0)
2348 ihold(&ci
->vfs_inode
);
2350 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2351 &ci
->vfs_inode
, ci
->i_wb_ref
-1, ci
->i_wb_ref
);
2356 * Try to grab cap references. Specify those refs we @want, and the
2357 * minimal set we @need. Also include the larger offset we are writing
2358 * to (when applicable), and check against max_size here as well.
2359 * Note that caller is responsible for ensuring max_size increases are
2360 * requested from the MDS.
2362 static int try_get_cap_refs(struct ceph_inode_info
*ci
, int need
, int want
,
2363 loff_t endoff
, bool nonblock
, int *got
, int *err
)
2365 struct inode
*inode
= &ci
->vfs_inode
;
2366 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
2368 int have
, implemented
;
2370 bool snap_rwsem_locked
= false;
2372 dout("get_cap_refs %p need %s want %s\n", inode
,
2373 ceph_cap_string(need
), ceph_cap_string(want
));
2376 spin_lock(&ci
->i_ceph_lock
);
2378 /* make sure file is actually open */
2379 file_wanted
= __ceph_caps_file_wanted(ci
);
2380 if ((file_wanted
& need
) != need
) {
2381 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2382 ceph_cap_string(need
), ceph_cap_string(file_wanted
));
2388 /* finish pending truncate */
2389 while (ci
->i_truncate_pending
) {
2390 spin_unlock(&ci
->i_ceph_lock
);
2391 if (snap_rwsem_locked
) {
2392 up_read(&mdsc
->snap_rwsem
);
2393 snap_rwsem_locked
= false;
2395 __ceph_do_pending_vmtruncate(inode
);
2396 spin_lock(&ci
->i_ceph_lock
);
2399 have
= __ceph_caps_issued(ci
, &implemented
);
2401 if (have
& need
& CEPH_CAP_FILE_WR
) {
2402 if (endoff
>= 0 && endoff
> (loff_t
)ci
->i_max_size
) {
2403 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2404 inode
, endoff
, ci
->i_max_size
);
2405 if (endoff
> ci
->i_requested_max_size
) {
2412 * If a sync write is in progress, we must wait, so that we
2413 * can get a final snapshot value for size+mtime.
2415 if (__ceph_have_pending_cap_snap(ci
)) {
2416 dout("get_cap_refs %p cap_snap_pending\n", inode
);
2421 if ((have
& need
) == need
) {
2423 * Look at (implemented & ~have & not) so that we keep waiting
2424 * on transition from wanted -> needed caps. This is needed
2425 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2426 * going before a prior buffered writeback happens.
2428 int not = want
& ~(have
& need
);
2429 int revoking
= implemented
& ~have
;
2430 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2431 inode
, ceph_cap_string(have
), ceph_cap_string(not),
2432 ceph_cap_string(revoking
));
2433 if ((revoking
& not) == 0) {
2434 if (!snap_rwsem_locked
&&
2435 !ci
->i_head_snapc
&&
2436 (need
& CEPH_CAP_FILE_WR
)) {
2437 if (!down_read_trylock(&mdsc
->snap_rwsem
)) {
2439 * we can not call down_read() when
2440 * task isn't in TASK_RUNNING state
2448 spin_unlock(&ci
->i_ceph_lock
);
2449 down_read(&mdsc
->snap_rwsem
);
2450 snap_rwsem_locked
= true;
2453 snap_rwsem_locked
= true;
2455 *got
= need
| (have
& want
);
2456 if ((need
& CEPH_CAP_FILE_RD
) &&
2457 !(*got
& CEPH_CAP_FILE_CACHE
))
2458 ceph_disable_fscache_readpage(ci
);
2459 __take_cap_refs(ci
, *got
, true);
2463 int session_readonly
= false;
2464 if ((need
& CEPH_CAP_FILE_WR
) && ci
->i_auth_cap
) {
2465 struct ceph_mds_session
*s
= ci
->i_auth_cap
->session
;
2466 spin_lock(&s
->s_cap_lock
);
2467 session_readonly
= s
->s_readonly
;
2468 spin_unlock(&s
->s_cap_lock
);
2470 if (session_readonly
) {
2471 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2472 inode
, ceph_cap_string(need
), ci
->i_auth_cap
->mds
);
2478 if (ci
->i_ceph_flags
& CEPH_I_CAP_DROPPED
) {
2480 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) ==
2481 CEPH_MOUNT_SHUTDOWN
) {
2482 dout("get_cap_refs %p forced umount\n", inode
);
2487 mds_wanted
= __ceph_caps_mds_wanted(ci
);
2488 if ((mds_wanted
& need
) != need
) {
2489 dout("get_cap_refs %p caps were dropped"
2490 " (session killed?)\n", inode
);
2495 if ((mds_wanted
& file_wanted
) ==
2496 (file_wanted
& (CEPH_CAP_FILE_RD
|CEPH_CAP_FILE_WR
)))
2497 ci
->i_ceph_flags
&= ~CEPH_I_CAP_DROPPED
;
2500 dout("get_cap_refs %p have %s needed %s\n", inode
,
2501 ceph_cap_string(have
), ceph_cap_string(need
));
2504 spin_unlock(&ci
->i_ceph_lock
);
2505 if (snap_rwsem_locked
)
2506 up_read(&mdsc
->snap_rwsem
);
2508 dout("get_cap_refs %p ret %d got %s\n", inode
,
2509 ret
, ceph_cap_string(*got
));
2514 * Check the offset we are writing up to against our current
2515 * max_size. If necessary, tell the MDS we want to write to
2518 static void check_max_size(struct inode
*inode
, loff_t endoff
)
2520 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2523 /* do we need to explicitly request a larger max_size? */
2524 spin_lock(&ci
->i_ceph_lock
);
2525 if (endoff
>= ci
->i_max_size
&& endoff
> ci
->i_wanted_max_size
) {
2526 dout("write %p at large endoff %llu, req max_size\n",
2528 ci
->i_wanted_max_size
= endoff
;
2530 /* duplicate ceph_check_caps()'s logic */
2531 if (ci
->i_auth_cap
&&
2532 (ci
->i_auth_cap
->issued
& CEPH_CAP_FILE_WR
) &&
2533 ci
->i_wanted_max_size
> ci
->i_max_size
&&
2534 ci
->i_wanted_max_size
> ci
->i_requested_max_size
)
2536 spin_unlock(&ci
->i_ceph_lock
);
2538 ceph_check_caps(ci
, CHECK_CAPS_AUTHONLY
, NULL
);
2541 int ceph_try_get_caps(struct ceph_inode_info
*ci
, int need
, int want
, int *got
)
2545 BUG_ON(need
& ~CEPH_CAP_FILE_RD
);
2546 BUG_ON(want
& ~(CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
));
2547 ret
= ceph_pool_perm_check(ci
, need
);
2551 ret
= try_get_cap_refs(ci
, need
, want
, 0, true, got
, &err
);
2553 if (err
== -EAGAIN
) {
2555 } else if (err
< 0) {
2563 * Wait for caps, and take cap references. If we can't get a WR cap
2564 * due to a small max_size, make sure we check_max_size (and possibly
2565 * ask the mds) so we don't get hung up indefinitely.
2567 int ceph_get_caps(struct ceph_inode_info
*ci
, int need
, int want
,
2568 loff_t endoff
, int *got
, struct page
**pinned_page
)
2570 int _got
, ret
, err
= 0;
2572 ret
= ceph_pool_perm_check(ci
, need
);
2578 check_max_size(&ci
->vfs_inode
, endoff
);
2582 ret
= try_get_cap_refs(ci
, need
, want
, endoff
,
2583 false, &_got
, &err
);
2590 DEFINE_WAIT_FUNC(wait
, woken_wake_function
);
2591 add_wait_queue(&ci
->i_cap_wq
, &wait
);
2593 while (!try_get_cap_refs(ci
, need
, want
, endoff
,
2594 true, &_got
, &err
)) {
2595 if (signal_pending(current
)) {
2599 wait_woken(&wait
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
2602 remove_wait_queue(&ci
->i_cap_wq
, &wait
);
2610 if (err
== -ESTALE
) {
2611 /* session was killed, try renew caps */
2612 ret
= ceph_renew_caps(&ci
->vfs_inode
);
2619 if (ci
->i_inline_version
!= CEPH_INLINE_NONE
&&
2620 (_got
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)) &&
2621 i_size_read(&ci
->vfs_inode
) > 0) {
2623 find_get_page(ci
->vfs_inode
.i_mapping
, 0);
2625 if (PageUptodate(page
)) {
2626 *pinned_page
= page
;
2632 * drop cap refs first because getattr while
2633 * holding * caps refs can cause deadlock.
2635 ceph_put_cap_refs(ci
, _got
);
2639 * getattr request will bring inline data into
2642 ret
= __ceph_do_getattr(&ci
->vfs_inode
, NULL
,
2643 CEPH_STAT_CAP_INLINE_DATA
,
2652 if ((_got
& CEPH_CAP_FILE_RD
) && (_got
& CEPH_CAP_FILE_CACHE
))
2653 ceph_fscache_revalidate_cookie(ci
);
2660 * Take cap refs. Caller must already know we hold at least one ref
2661 * on the caps in question or we don't know this is safe.
2663 void ceph_get_cap_refs(struct ceph_inode_info
*ci
, int caps
)
2665 spin_lock(&ci
->i_ceph_lock
);
2666 __take_cap_refs(ci
, caps
, false);
2667 spin_unlock(&ci
->i_ceph_lock
);
2672 * drop cap_snap that is not associated with any snapshot.
2673 * we don't need to send FLUSHSNAP message for it.
2675 static int ceph_try_drop_cap_snap(struct ceph_inode_info
*ci
,
2676 struct ceph_cap_snap
*capsnap
)
2678 if (!capsnap
->need_flush
&&
2679 !capsnap
->writing
&& !capsnap
->dirty_pages
) {
2680 dout("dropping cap_snap %p follows %llu\n",
2681 capsnap
, capsnap
->follows
);
2682 BUG_ON(capsnap
->cap_flush
.tid
> 0);
2683 ceph_put_snap_context(capsnap
->context
);
2684 if (!list_is_last(&capsnap
->ci_item
, &ci
->i_cap_snaps
))
2685 ci
->i_ceph_flags
|= CEPH_I_FLUSH_SNAPS
;
2687 list_del(&capsnap
->ci_item
);
2688 ceph_put_cap_snap(capsnap
);
2697 * If we released the last ref on any given cap, call ceph_check_caps
2698 * to release (or schedule a release).
2700 * If we are releasing a WR cap (from a sync write), finalize any affected
2701 * cap_snap, and wake up any waiters.
2703 void ceph_put_cap_refs(struct ceph_inode_info
*ci
, int had
)
2705 struct inode
*inode
= &ci
->vfs_inode
;
2706 int last
= 0, put
= 0, flushsnaps
= 0, wake
= 0;
2708 spin_lock(&ci
->i_ceph_lock
);
2709 if (had
& CEPH_CAP_PIN
)
2711 if (had
& CEPH_CAP_FILE_RD
)
2712 if (--ci
->i_rd_ref
== 0)
2714 if (had
& CEPH_CAP_FILE_CACHE
)
2715 if (--ci
->i_rdcache_ref
== 0)
2717 if (had
& CEPH_CAP_FILE_BUFFER
) {
2718 if (--ci
->i_wb_ref
== 0) {
2722 dout("put_cap_refs %p wb %d -> %d (?)\n",
2723 inode
, ci
->i_wb_ref
+1, ci
->i_wb_ref
);
2725 if (had
& CEPH_CAP_FILE_WR
)
2726 if (--ci
->i_wr_ref
== 0) {
2728 if (__ceph_have_pending_cap_snap(ci
)) {
2729 struct ceph_cap_snap
*capsnap
=
2730 list_last_entry(&ci
->i_cap_snaps
,
2731 struct ceph_cap_snap
,
2733 capsnap
->writing
= 0;
2734 if (ceph_try_drop_cap_snap(ci
, capsnap
))
2736 else if (__ceph_finish_cap_snap(ci
, capsnap
))
2740 if (ci
->i_wrbuffer_ref_head
== 0 &&
2741 ci
->i_dirty_caps
== 0 &&
2742 ci
->i_flushing_caps
== 0) {
2743 BUG_ON(!ci
->i_head_snapc
);
2744 ceph_put_snap_context(ci
->i_head_snapc
);
2745 ci
->i_head_snapc
= NULL
;
2747 /* see comment in __ceph_remove_cap() */
2748 if (!__ceph_is_any_caps(ci
) && ci
->i_snap_realm
)
2749 drop_inode_snap_realm(ci
);
2751 spin_unlock(&ci
->i_ceph_lock
);
2753 dout("put_cap_refs %p had %s%s%s\n", inode
, ceph_cap_string(had
),
2754 last
? " last" : "", put
? " put" : "");
2756 if (last
&& !flushsnaps
)
2757 ceph_check_caps(ci
, 0, NULL
);
2758 else if (flushsnaps
)
2759 ceph_flush_snaps(ci
, NULL
);
2761 wake_up_all(&ci
->i_cap_wq
);
2767 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2768 * context. Adjust per-snap dirty page accounting as appropriate.
2769 * Once all dirty data for a cap_snap is flushed, flush snapped file
2770 * metadata back to the MDS. If we dropped the last ref, call
2773 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info
*ci
, int nr
,
2774 struct ceph_snap_context
*snapc
)
2776 struct inode
*inode
= &ci
->vfs_inode
;
2777 struct ceph_cap_snap
*capsnap
= NULL
;
2781 bool flush_snaps
= false;
2782 bool complete_capsnap
= false;
2784 spin_lock(&ci
->i_ceph_lock
);
2785 ci
->i_wrbuffer_ref
-= nr
;
2786 if (ci
->i_wrbuffer_ref
== 0) {
2791 if (ci
->i_head_snapc
== snapc
) {
2792 ci
->i_wrbuffer_ref_head
-= nr
;
2793 if (ci
->i_wrbuffer_ref_head
== 0 &&
2794 ci
->i_wr_ref
== 0 &&
2795 ci
->i_dirty_caps
== 0 &&
2796 ci
->i_flushing_caps
== 0) {
2797 BUG_ON(!ci
->i_head_snapc
);
2798 ceph_put_snap_context(ci
->i_head_snapc
);
2799 ci
->i_head_snapc
= NULL
;
2801 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2803 ci
->i_wrbuffer_ref
+nr
, ci
->i_wrbuffer_ref_head
+nr
,
2804 ci
->i_wrbuffer_ref
, ci
->i_wrbuffer_ref_head
,
2805 last
? " LAST" : "");
2807 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
2808 if (capsnap
->context
== snapc
) {
2814 capsnap
->dirty_pages
-= nr
;
2815 if (capsnap
->dirty_pages
== 0) {
2816 complete_capsnap
= true;
2817 if (!capsnap
->writing
) {
2818 if (ceph_try_drop_cap_snap(ci
, capsnap
)) {
2821 ci
->i_ceph_flags
|= CEPH_I_FLUSH_SNAPS
;
2826 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2827 " snap %lld %d/%d -> %d/%d %s%s\n",
2828 inode
, capsnap
, capsnap
->context
->seq
,
2829 ci
->i_wrbuffer_ref
+nr
, capsnap
->dirty_pages
+ nr
,
2830 ci
->i_wrbuffer_ref
, capsnap
->dirty_pages
,
2831 last
? " (wrbuffer last)" : "",
2832 complete_capsnap
? " (complete capsnap)" : "");
2835 spin_unlock(&ci
->i_ceph_lock
);
2838 ceph_check_caps(ci
, CHECK_CAPS_AUTHONLY
, NULL
);
2839 } else if (flush_snaps
) {
2840 ceph_flush_snaps(ci
, NULL
);
2842 if (complete_capsnap
)
2843 wake_up_all(&ci
->i_cap_wq
);
2849 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2851 static void invalidate_aliases(struct inode
*inode
)
2853 struct dentry
*dn
, *prev
= NULL
;
2855 dout("invalidate_aliases inode %p\n", inode
);
2856 d_prune_aliases(inode
);
2858 * For non-directory inode, d_find_alias() only returns
2859 * hashed dentry. After calling d_invalidate(), the
2860 * dentry becomes unhashed.
2862 * For directory inode, d_find_alias() can return
2863 * unhashed dentry. But directory inode should have
2864 * one alias at most.
2866 while ((dn
= d_find_alias(inode
))) {
2881 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2882 * actually be a revocation if it specifies a smaller cap set.)
2884 * caller holds s_mutex and i_ceph_lock, we drop both.
2886 static void handle_cap_grant(struct ceph_mds_client
*mdsc
,
2887 struct inode
*inode
, struct ceph_mds_caps
*grant
,
2888 struct ceph_string
**pns
, u64 inline_version
,
2889 void *inline_data
, u32 inline_len
,
2890 struct ceph_buffer
*xattr_buf
,
2891 struct ceph_mds_session
*session
,
2892 struct ceph_cap
*cap
, int issued
)
2893 __releases(ci
->i_ceph_lock
)
2894 __releases(mdsc
->snap_rwsem
)
2896 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2897 int mds
= session
->s_mds
;
2898 int seq
= le32_to_cpu(grant
->seq
);
2899 int newcaps
= le32_to_cpu(grant
->caps
);
2900 int used
, wanted
, dirty
;
2901 u64 size
= le64_to_cpu(grant
->size
);
2902 u64 max_size
= le64_to_cpu(grant
->max_size
);
2903 struct timespec mtime
, atime
, ctime
;
2906 bool writeback
= false;
2907 bool queue_trunc
= false;
2908 bool queue_invalidate
= false;
2909 bool deleted_inode
= false;
2910 bool fill_inline
= false;
2912 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2913 inode
, cap
, mds
, seq
, ceph_cap_string(newcaps
));
2914 dout(" size %llu max_size %llu, i_size %llu\n", size
, max_size
,
2919 * auth mds of the inode changed. we received the cap export message,
2920 * but still haven't received the cap import message. handle_cap_export
2921 * updated the new auth MDS' cap.
2923 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2924 * that was sent before the cap import message. So don't remove caps.
2926 if (ceph_seq_cmp(seq
, cap
->seq
) <= 0) {
2927 WARN_ON(cap
!= ci
->i_auth_cap
);
2928 WARN_ON(cap
->cap_id
!= le64_to_cpu(grant
->cap_id
));
2930 newcaps
|= cap
->issued
;
2934 * If CACHE is being revoked, and we have no dirty buffers,
2935 * try to invalidate (once). (If there are dirty buffers, we
2936 * will invalidate _after_ writeback.)
2938 if (!S_ISDIR(inode
->i_mode
) && /* don't invalidate readdir cache */
2939 ((cap
->issued
& ~newcaps
) & CEPH_CAP_FILE_CACHE
) &&
2940 (newcaps
& CEPH_CAP_FILE_LAZYIO
) == 0 &&
2941 !(ci
->i_wrbuffer_ref
|| ci
->i_wb_ref
)) {
2942 if (try_nonblocking_invalidate(inode
)) {
2943 /* there were locked pages.. invalidate later
2944 in a separate thread. */
2945 if (ci
->i_rdcache_revoking
!= ci
->i_rdcache_gen
) {
2946 queue_invalidate
= true;
2947 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
;
2952 /* side effects now are allowed */
2953 cap
->cap_gen
= session
->s_cap_gen
;
2956 __check_cap_issue(ci
, cap
, newcaps
);
2958 if ((newcaps
& CEPH_CAP_AUTH_SHARED
) &&
2959 (issued
& CEPH_CAP_AUTH_EXCL
) == 0) {
2960 inode
->i_mode
= le32_to_cpu(grant
->mode
);
2961 inode
->i_uid
= make_kuid(&init_user_ns
, le32_to_cpu(grant
->uid
));
2962 inode
->i_gid
= make_kgid(&init_user_ns
, le32_to_cpu(grant
->gid
));
2963 dout("%p mode 0%o uid.gid %d.%d\n", inode
, inode
->i_mode
,
2964 from_kuid(&init_user_ns
, inode
->i_uid
),
2965 from_kgid(&init_user_ns
, inode
->i_gid
));
2968 if ((newcaps
& CEPH_CAP_AUTH_SHARED
) &&
2969 (issued
& CEPH_CAP_LINK_EXCL
) == 0) {
2970 set_nlink(inode
, le32_to_cpu(grant
->nlink
));
2971 if (inode
->i_nlink
== 0 &&
2972 (newcaps
& (CEPH_CAP_LINK_SHARED
| CEPH_CAP_LINK_EXCL
)))
2973 deleted_inode
= true;
2976 if ((issued
& CEPH_CAP_XATTR_EXCL
) == 0 && grant
->xattr_len
) {
2977 int len
= le32_to_cpu(grant
->xattr_len
);
2978 u64 version
= le64_to_cpu(grant
->xattr_version
);
2980 if (version
> ci
->i_xattrs
.version
) {
2981 dout(" got new xattrs v%llu on %p len %d\n",
2982 version
, inode
, len
);
2983 if (ci
->i_xattrs
.blob
)
2984 ceph_buffer_put(ci
->i_xattrs
.blob
);
2985 ci
->i_xattrs
.blob
= ceph_buffer_get(xattr_buf
);
2986 ci
->i_xattrs
.version
= version
;
2987 ceph_forget_all_cached_acls(inode
);
2991 if (newcaps
& CEPH_CAP_ANY_RD
) {
2992 /* ctime/mtime/atime? */
2993 ceph_decode_timespec(&mtime
, &grant
->mtime
);
2994 ceph_decode_timespec(&atime
, &grant
->atime
);
2995 ceph_decode_timespec(&ctime
, &grant
->ctime
);
2996 ceph_fill_file_time(inode
, issued
,
2997 le32_to_cpu(grant
->time_warp_seq
),
2998 &ctime
, &mtime
, &atime
);
3001 if (newcaps
& (CEPH_CAP_ANY_FILE_RD
| CEPH_CAP_ANY_FILE_WR
)) {
3002 /* file layout may have changed */
3003 s64 old_pool
= ci
->i_layout
.pool_id
;
3004 struct ceph_string
*old_ns
;
3006 ceph_file_layout_from_legacy(&ci
->i_layout
, &grant
->layout
);
3007 old_ns
= rcu_dereference_protected(ci
->i_layout
.pool_ns
,
3008 lockdep_is_held(&ci
->i_ceph_lock
));
3009 rcu_assign_pointer(ci
->i_layout
.pool_ns
, *pns
);
3011 if (ci
->i_layout
.pool_id
!= old_pool
|| *pns
!= old_ns
)
3012 ci
->i_ceph_flags
&= ~CEPH_I_POOL_PERM
;
3016 /* size/truncate_seq? */
3017 queue_trunc
= ceph_fill_file_size(inode
, issued
,
3018 le32_to_cpu(grant
->truncate_seq
),
3019 le64_to_cpu(grant
->truncate_size
),
3021 /* max size increase? */
3022 if (ci
->i_auth_cap
== cap
&& max_size
!= ci
->i_max_size
) {
3023 dout("max_size %lld -> %llu\n",
3024 ci
->i_max_size
, max_size
);
3025 ci
->i_max_size
= max_size
;
3026 if (max_size
>= ci
->i_wanted_max_size
) {
3027 ci
->i_wanted_max_size
= 0; /* reset */
3028 ci
->i_requested_max_size
= 0;
3034 /* check cap bits */
3035 wanted
= __ceph_caps_wanted(ci
);
3036 used
= __ceph_caps_used(ci
);
3037 dirty
= __ceph_caps_dirty(ci
);
3038 dout(" my wanted = %s, used = %s, dirty %s\n",
3039 ceph_cap_string(wanted
),
3040 ceph_cap_string(used
),
3041 ceph_cap_string(dirty
));
3042 if (wanted
!= le32_to_cpu(grant
->wanted
)) {
3043 dout("mds wanted %s -> %s\n",
3044 ceph_cap_string(le32_to_cpu(grant
->wanted
)),
3045 ceph_cap_string(wanted
));
3046 /* imported cap may not have correct mds_wanted */
3047 if (le32_to_cpu(grant
->op
) == CEPH_CAP_OP_IMPORT
)
3051 /* revocation, grant, or no-op? */
3052 if (cap
->issued
& ~newcaps
) {
3053 int revoking
= cap
->issued
& ~newcaps
;
3055 dout("revocation: %s -> %s (revoking %s)\n",
3056 ceph_cap_string(cap
->issued
),
3057 ceph_cap_string(newcaps
),
3058 ceph_cap_string(revoking
));
3059 if (revoking
& used
& CEPH_CAP_FILE_BUFFER
)
3060 writeback
= true; /* initiate writeback; will delay ack */
3061 else if (revoking
== CEPH_CAP_FILE_CACHE
&&
3062 (newcaps
& CEPH_CAP_FILE_LAZYIO
) == 0 &&
3064 ; /* do nothing yet, invalidation will be queued */
3065 else if (cap
== ci
->i_auth_cap
)
3066 check_caps
= 1; /* check auth cap only */
3068 check_caps
= 2; /* check all caps */
3069 cap
->issued
= newcaps
;
3070 cap
->implemented
|= newcaps
;
3071 } else if (cap
->issued
== newcaps
) {
3072 dout("caps unchanged: %s -> %s\n",
3073 ceph_cap_string(cap
->issued
), ceph_cap_string(newcaps
));
3075 dout("grant: %s -> %s\n", ceph_cap_string(cap
->issued
),
3076 ceph_cap_string(newcaps
));
3077 /* non-auth MDS is revoking the newly grant caps ? */
3078 if (cap
== ci
->i_auth_cap
&&
3079 __ceph_caps_revoking_other(ci
, cap
, newcaps
))
3082 cap
->issued
= newcaps
;
3083 cap
->implemented
|= newcaps
; /* add bits only, to
3084 * avoid stepping on a
3085 * pending revocation */
3088 BUG_ON(cap
->issued
& ~cap
->implemented
);
3090 if (inline_version
> 0 && inline_version
>= ci
->i_inline_version
) {
3091 ci
->i_inline_version
= inline_version
;
3092 if (ci
->i_inline_version
!= CEPH_INLINE_NONE
&&
3093 (newcaps
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)))
3097 if (le32_to_cpu(grant
->op
) == CEPH_CAP_OP_IMPORT
) {
3098 if (newcaps
& ~issued
)
3100 kick_flushing_inode_caps(mdsc
, session
, inode
);
3101 up_read(&mdsc
->snap_rwsem
);
3103 spin_unlock(&ci
->i_ceph_lock
);
3107 ceph_fill_inline_data(inode
, NULL
, inline_data
, inline_len
);
3110 ceph_queue_vmtruncate(inode
);
3114 * queue inode for writeback: we can't actually call
3115 * filemap_write_and_wait, etc. from message handler
3118 ceph_queue_writeback(inode
);
3119 if (queue_invalidate
)
3120 ceph_queue_invalidate(inode
);
3122 invalidate_aliases(inode
);
3124 wake_up_all(&ci
->i_cap_wq
);
3126 if (check_caps
== 1)
3127 ceph_check_caps(ci
, CHECK_CAPS_NODELAY
|CHECK_CAPS_AUTHONLY
,
3129 else if (check_caps
== 2)
3130 ceph_check_caps(ci
, CHECK_CAPS_NODELAY
, session
);
3132 mutex_unlock(&session
->s_mutex
);
3136 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3137 * MDS has been safely committed.
3139 static void handle_cap_flush_ack(struct inode
*inode
, u64 flush_tid
,
3140 struct ceph_mds_caps
*m
,
3141 struct ceph_mds_session
*session
,
3142 struct ceph_cap
*cap
)
3143 __releases(ci
->i_ceph_lock
)
3145 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3146 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
3147 struct ceph_cap_flush
*cf
, *tmp_cf
;
3148 LIST_HEAD(to_remove
);
3149 unsigned seq
= le32_to_cpu(m
->seq
);
3150 int dirty
= le32_to_cpu(m
->dirty
);
3156 list_for_each_entry_safe(cf
, tmp_cf
, &ci
->i_cap_flush_list
, i_list
) {
3157 if (cf
->tid
== flush_tid
)
3159 if (cf
->caps
== 0) /* capsnap */
3161 if (cf
->tid
<= flush_tid
) {
3162 if (__finish_cap_flush(NULL
, ci
, cf
))
3164 list_add_tail(&cf
->i_list
, &to_remove
);
3166 cleaned
&= ~cf
->caps
;
3172 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3173 " flushing %s -> %s\n",
3174 inode
, session
->s_mds
, seq
, ceph_cap_string(dirty
),
3175 ceph_cap_string(cleaned
), ceph_cap_string(ci
->i_flushing_caps
),
3176 ceph_cap_string(ci
->i_flushing_caps
& ~cleaned
));
3178 if (list_empty(&to_remove
) && !cleaned
)
3181 ci
->i_flushing_caps
&= ~cleaned
;
3183 spin_lock(&mdsc
->cap_dirty_lock
);
3185 list_for_each_entry(cf
, &to_remove
, i_list
) {
3186 if (__finish_cap_flush(mdsc
, NULL
, cf
))
3190 if (ci
->i_flushing_caps
== 0) {
3191 if (list_empty(&ci
->i_cap_flush_list
)) {
3192 list_del_init(&ci
->i_flushing_item
);
3193 if (!list_empty(&session
->s_cap_flushing
)) {
3194 dout(" mds%d still flushing cap on %p\n",
3196 &list_first_entry(&session
->s_cap_flushing
,
3197 struct ceph_inode_info
,
3198 i_flushing_item
)->vfs_inode
);
3201 mdsc
->num_cap_flushing
--;
3202 dout(" inode %p now !flushing\n", inode
);
3204 if (ci
->i_dirty_caps
== 0) {
3205 dout(" inode %p now clean\n", inode
);
3206 BUG_ON(!list_empty(&ci
->i_dirty_item
));
3208 if (ci
->i_wr_ref
== 0 &&
3209 ci
->i_wrbuffer_ref_head
== 0) {
3210 BUG_ON(!ci
->i_head_snapc
);
3211 ceph_put_snap_context(ci
->i_head_snapc
);
3212 ci
->i_head_snapc
= NULL
;
3215 BUG_ON(list_empty(&ci
->i_dirty_item
));
3218 spin_unlock(&mdsc
->cap_dirty_lock
);
3221 spin_unlock(&ci
->i_ceph_lock
);
3223 while (!list_empty(&to_remove
)) {
3224 cf
= list_first_entry(&to_remove
,
3225 struct ceph_cap_flush
, i_list
);
3226 list_del(&cf
->i_list
);
3227 ceph_free_cap_flush(cf
);
3231 wake_up_all(&ci
->i_cap_wq
);
3233 wake_up_all(&mdsc
->cap_flushing_wq
);
3239 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3240 * throw away our cap_snap.
3242 * Caller hold s_mutex.
3244 static void handle_cap_flushsnap_ack(struct inode
*inode
, u64 flush_tid
,
3245 struct ceph_mds_caps
*m
,
3246 struct ceph_mds_session
*session
)
3248 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3249 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
3250 u64 follows
= le64_to_cpu(m
->snap_follows
);
3251 struct ceph_cap_snap
*capsnap
;
3252 bool flushed
= false;
3253 bool wake_ci
= false;
3254 bool wake_mdsc
= false;
3256 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3257 inode
, ci
, session
->s_mds
, follows
);
3259 spin_lock(&ci
->i_ceph_lock
);
3260 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
3261 if (capsnap
->follows
== follows
) {
3262 if (capsnap
->cap_flush
.tid
!= flush_tid
) {
3263 dout(" cap_snap %p follows %lld tid %lld !="
3264 " %lld\n", capsnap
, follows
,
3265 flush_tid
, capsnap
->cap_flush
.tid
);
3271 dout(" skipping cap_snap %p follows %lld\n",
3272 capsnap
, capsnap
->follows
);
3276 WARN_ON(capsnap
->dirty_pages
|| capsnap
->writing
);
3277 dout(" removing %p cap_snap %p follows %lld\n",
3278 inode
, capsnap
, follows
);
3279 list_del(&capsnap
->ci_item
);
3280 if (__finish_cap_flush(NULL
, ci
, &capsnap
->cap_flush
))
3283 spin_lock(&mdsc
->cap_dirty_lock
);
3285 if (list_empty(&ci
->i_cap_flush_list
))
3286 list_del_init(&ci
->i_flushing_item
);
3288 if (__finish_cap_flush(mdsc
, NULL
, &capsnap
->cap_flush
))
3291 spin_unlock(&mdsc
->cap_dirty_lock
);
3293 spin_unlock(&ci
->i_ceph_lock
);
3295 ceph_put_snap_context(capsnap
->context
);
3296 ceph_put_cap_snap(capsnap
);
3298 wake_up_all(&ci
->i_cap_wq
);
3300 wake_up_all(&mdsc
->cap_flushing_wq
);
3306 * Handle TRUNC from MDS, indicating file truncation.
3308 * caller hold s_mutex.
3310 static void handle_cap_trunc(struct inode
*inode
,
3311 struct ceph_mds_caps
*trunc
,
3312 struct ceph_mds_session
*session
)
3313 __releases(ci
->i_ceph_lock
)
3315 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3316 int mds
= session
->s_mds
;
3317 int seq
= le32_to_cpu(trunc
->seq
);
3318 u32 truncate_seq
= le32_to_cpu(trunc
->truncate_seq
);
3319 u64 truncate_size
= le64_to_cpu(trunc
->truncate_size
);
3320 u64 size
= le64_to_cpu(trunc
->size
);
3321 int implemented
= 0;
3322 int dirty
= __ceph_caps_dirty(ci
);
3323 int issued
= __ceph_caps_issued(ceph_inode(inode
), &implemented
);
3324 int queue_trunc
= 0;
3326 issued
|= implemented
| dirty
;
3328 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3329 inode
, mds
, seq
, truncate_size
, truncate_seq
);
3330 queue_trunc
= ceph_fill_file_size(inode
, issued
,
3331 truncate_seq
, truncate_size
, size
);
3332 spin_unlock(&ci
->i_ceph_lock
);
3335 ceph_queue_vmtruncate(inode
);
3339 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3340 * different one. If we are the most recent migration we've seen (as
3341 * indicated by mseq), make note of the migrating cap bits for the
3342 * duration (until we see the corresponding IMPORT).
3344 * caller holds s_mutex
3346 static void handle_cap_export(struct inode
*inode
, struct ceph_mds_caps
*ex
,
3347 struct ceph_mds_cap_peer
*ph
,
3348 struct ceph_mds_session
*session
)
3350 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
3351 struct ceph_mds_session
*tsession
= NULL
;
3352 struct ceph_cap
*cap
, *tcap
, *new_cap
= NULL
;
3353 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3355 unsigned mseq
= le32_to_cpu(ex
->migrate_seq
);
3356 unsigned t_seq
, t_mseq
;
3358 int mds
= session
->s_mds
;
3361 t_cap_id
= le64_to_cpu(ph
->cap_id
);
3362 t_seq
= le32_to_cpu(ph
->seq
);
3363 t_mseq
= le32_to_cpu(ph
->mseq
);
3364 target
= le32_to_cpu(ph
->mds
);
3366 t_cap_id
= t_seq
= t_mseq
= 0;
3370 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3371 inode
, ci
, mds
, mseq
, target
);
3373 spin_lock(&ci
->i_ceph_lock
);
3374 cap
= __get_cap_for_mds(ci
, mds
);
3375 if (!cap
|| cap
->cap_id
!= le64_to_cpu(ex
->cap_id
))
3379 __ceph_remove_cap(cap
, false);
3380 if (!ci
->i_auth_cap
)
3381 ci
->i_ceph_flags
|= CEPH_I_CAP_DROPPED
;
3386 * now we know we haven't received the cap import message yet
3387 * because the exported cap still exist.
3390 issued
= cap
->issued
;
3391 WARN_ON(issued
!= cap
->implemented
);
3393 tcap
= __get_cap_for_mds(ci
, target
);
3395 /* already have caps from the target */
3396 if (tcap
->cap_id
!= t_cap_id
||
3397 ceph_seq_cmp(tcap
->seq
, t_seq
) < 0) {
3398 dout(" updating import cap %p mds%d\n", tcap
, target
);
3399 tcap
->cap_id
= t_cap_id
;
3400 tcap
->seq
= t_seq
- 1;
3401 tcap
->issue_seq
= t_seq
- 1;
3402 tcap
->mseq
= t_mseq
;
3403 tcap
->issued
|= issued
;
3404 tcap
->implemented
|= issued
;
3405 if (cap
== ci
->i_auth_cap
)
3406 ci
->i_auth_cap
= tcap
;
3407 if (!list_empty(&ci
->i_cap_flush_list
) &&
3408 ci
->i_auth_cap
== tcap
) {
3409 spin_lock(&mdsc
->cap_dirty_lock
);
3410 list_move_tail(&ci
->i_flushing_item
,
3411 &tcap
->session
->s_cap_flushing
);
3412 spin_unlock(&mdsc
->cap_dirty_lock
);
3415 __ceph_remove_cap(cap
, false);
3417 } else if (tsession
) {
3418 /* add placeholder for the export tagert */
3419 int flag
= (cap
== ci
->i_auth_cap
) ? CEPH_CAP_FLAG_AUTH
: 0;
3420 ceph_add_cap(inode
, tsession
, t_cap_id
, -1, issued
, 0,
3421 t_seq
- 1, t_mseq
, (u64
)-1, flag
, &new_cap
);
3423 __ceph_remove_cap(cap
, false);
3427 spin_unlock(&ci
->i_ceph_lock
);
3428 mutex_unlock(&session
->s_mutex
);
3430 /* open target session */
3431 tsession
= ceph_mdsc_open_export_target_session(mdsc
, target
);
3432 if (!IS_ERR(tsession
)) {
3434 mutex_lock(&session
->s_mutex
);
3435 mutex_lock_nested(&tsession
->s_mutex
,
3436 SINGLE_DEPTH_NESTING
);
3438 mutex_lock(&tsession
->s_mutex
);
3439 mutex_lock_nested(&session
->s_mutex
,
3440 SINGLE_DEPTH_NESTING
);
3442 new_cap
= ceph_get_cap(mdsc
, NULL
);
3451 spin_unlock(&ci
->i_ceph_lock
);
3452 mutex_unlock(&session
->s_mutex
);
3454 mutex_unlock(&tsession
->s_mutex
);
3455 ceph_put_mds_session(tsession
);
3458 ceph_put_cap(mdsc
, new_cap
);
3462 * Handle cap IMPORT.
3464 * caller holds s_mutex. acquires i_ceph_lock
3466 static void handle_cap_import(struct ceph_mds_client
*mdsc
,
3467 struct inode
*inode
, struct ceph_mds_caps
*im
,
3468 struct ceph_mds_cap_peer
*ph
,
3469 struct ceph_mds_session
*session
,
3470 struct ceph_cap
**target_cap
, int *old_issued
)
3471 __acquires(ci
->i_ceph_lock
)
3473 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3474 struct ceph_cap
*cap
, *ocap
, *new_cap
= NULL
;
3475 int mds
= session
->s_mds
;
3477 unsigned caps
= le32_to_cpu(im
->caps
);
3478 unsigned wanted
= le32_to_cpu(im
->wanted
);
3479 unsigned seq
= le32_to_cpu(im
->seq
);
3480 unsigned mseq
= le32_to_cpu(im
->migrate_seq
);
3481 u64 realmino
= le64_to_cpu(im
->realm
);
3482 u64 cap_id
= le64_to_cpu(im
->cap_id
);
3487 p_cap_id
= le64_to_cpu(ph
->cap_id
);
3488 peer
= le32_to_cpu(ph
->mds
);
3494 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3495 inode
, ci
, mds
, mseq
, peer
);
3498 spin_lock(&ci
->i_ceph_lock
);
3499 cap
= __get_cap_for_mds(ci
, mds
);
3502 spin_unlock(&ci
->i_ceph_lock
);
3503 new_cap
= ceph_get_cap(mdsc
, NULL
);
3509 ceph_put_cap(mdsc
, new_cap
);
3514 __ceph_caps_issued(ci
, &issued
);
3515 issued
|= __ceph_caps_dirty(ci
);
3517 ceph_add_cap(inode
, session
, cap_id
, -1, caps
, wanted
, seq
, mseq
,
3518 realmino
, CEPH_CAP_FLAG_AUTH
, &new_cap
);
3520 ocap
= peer
>= 0 ? __get_cap_for_mds(ci
, peer
) : NULL
;
3521 if (ocap
&& ocap
->cap_id
== p_cap_id
) {
3522 dout(" remove export cap %p mds%d flags %d\n",
3523 ocap
, peer
, ph
->flags
);
3524 if ((ph
->flags
& CEPH_CAP_FLAG_AUTH
) &&
3525 (ocap
->seq
!= le32_to_cpu(ph
->seq
) ||
3526 ocap
->mseq
!= le32_to_cpu(ph
->mseq
))) {
3527 pr_err("handle_cap_import: mismatched seq/mseq: "
3528 "ino (%llx.%llx) mds%d seq %d mseq %d "
3529 "importer mds%d has peer seq %d mseq %d\n",
3530 ceph_vinop(inode
), peer
, ocap
->seq
,
3531 ocap
->mseq
, mds
, le32_to_cpu(ph
->seq
),
3532 le32_to_cpu(ph
->mseq
));
3534 __ceph_remove_cap(ocap
, (ph
->flags
& CEPH_CAP_FLAG_RELEASE
));
3537 /* make sure we re-request max_size, if necessary */
3538 ci
->i_wanted_max_size
= 0;
3539 ci
->i_requested_max_size
= 0;
3541 *old_issued
= issued
;
3546 * Handle a caps message from the MDS.
3548 * Identify the appropriate session, inode, and call the right handler
3549 * based on the cap op.
3551 void ceph_handle_caps(struct ceph_mds_session
*session
,
3552 struct ceph_msg
*msg
)
3554 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
3555 struct super_block
*sb
= mdsc
->fsc
->sb
;
3556 struct inode
*inode
;
3557 struct ceph_inode_info
*ci
;
3558 struct ceph_cap
*cap
;
3559 struct ceph_mds_caps
*h
;
3560 struct ceph_mds_cap_peer
*peer
= NULL
;
3561 struct ceph_snap_realm
*realm
= NULL
;
3562 struct ceph_string
*pool_ns
= NULL
;
3563 int mds
= session
->s_mds
;
3566 struct ceph_vino vino
;
3568 u64 inline_version
= 0;
3569 void *inline_data
= NULL
;
3572 size_t snaptrace_len
;
3575 dout("handle_caps from mds%d\n", mds
);
3578 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
3579 tid
= le64_to_cpu(msg
->hdr
.tid
);
3580 if (msg
->front
.iov_len
< sizeof(*h
))
3582 h
= msg
->front
.iov_base
;
3583 op
= le32_to_cpu(h
->op
);
3584 vino
.ino
= le64_to_cpu(h
->ino
);
3585 vino
.snap
= CEPH_NOSNAP
;
3586 seq
= le32_to_cpu(h
->seq
);
3587 mseq
= le32_to_cpu(h
->migrate_seq
);
3590 snaptrace_len
= le32_to_cpu(h
->snap_trace_len
);
3591 p
= snaptrace
+ snaptrace_len
;
3593 if (le16_to_cpu(msg
->hdr
.version
) >= 2) {
3595 ceph_decode_32_safe(&p
, end
, flock_len
, bad
);
3596 if (p
+ flock_len
> end
)
3601 if (le16_to_cpu(msg
->hdr
.version
) >= 3) {
3602 if (op
== CEPH_CAP_OP_IMPORT
) {
3603 if (p
+ sizeof(*peer
) > end
)
3607 } else if (op
== CEPH_CAP_OP_EXPORT
) {
3608 /* recorded in unused fields */
3609 peer
= (void *)&h
->size
;
3613 if (le16_to_cpu(msg
->hdr
.version
) >= 4) {
3614 ceph_decode_64_safe(&p
, end
, inline_version
, bad
);
3615 ceph_decode_32_safe(&p
, end
, inline_len
, bad
);
3616 if (p
+ inline_len
> end
)
3622 if (le16_to_cpu(msg
->hdr
.version
) >= 8) {
3624 u32 caller_uid
, caller_gid
;
3625 u32 osd_epoch_barrier
;
3628 ceph_decode_32_safe(&p
, end
, osd_epoch_barrier
, bad
);
3630 ceph_decode_64_safe(&p
, end
, flush_tid
, bad
);
3632 ceph_decode_32_safe(&p
, end
, caller_uid
, bad
);
3633 ceph_decode_32_safe(&p
, end
, caller_gid
, bad
);
3635 ceph_decode_32_safe(&p
, end
, pool_ns_len
, bad
);
3636 if (pool_ns_len
> 0) {
3637 ceph_decode_need(&p
, end
, pool_ns_len
, bad
);
3638 pool_ns
= ceph_find_or_create_string(p
, pool_ns_len
);
3644 inode
= ceph_find_inode(sb
, vino
);
3645 ci
= ceph_inode(inode
);
3646 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op
), vino
.ino
,
3649 mutex_lock(&session
->s_mutex
);
3651 dout(" mds%d seq %lld cap seq %u\n", session
->s_mds
, session
->s_seq
,
3655 dout(" i don't have ino %llx\n", vino
.ino
);
3657 if (op
== CEPH_CAP_OP_IMPORT
) {
3658 cap
= ceph_get_cap(mdsc
, NULL
);
3659 cap
->cap_ino
= vino
.ino
;
3660 cap
->queue_release
= 1;
3661 cap
->cap_id
= le64_to_cpu(h
->cap_id
);
3664 cap
->issue_seq
= seq
;
3665 spin_lock(&session
->s_cap_lock
);
3666 list_add_tail(&cap
->session_caps
,
3667 &session
->s_cap_releases
);
3668 session
->s_num_cap_releases
++;
3669 spin_unlock(&session
->s_cap_lock
);
3671 goto flush_cap_releases
;
3674 /* these will work even if we don't have a cap yet */
3676 case CEPH_CAP_OP_FLUSHSNAP_ACK
:
3677 handle_cap_flushsnap_ack(inode
, tid
, h
, session
);
3680 case CEPH_CAP_OP_EXPORT
:
3681 handle_cap_export(inode
, h
, peer
, session
);
3684 case CEPH_CAP_OP_IMPORT
:
3686 if (snaptrace_len
) {
3687 down_write(&mdsc
->snap_rwsem
);
3688 ceph_update_snap_trace(mdsc
, snaptrace
,
3689 snaptrace
+ snaptrace_len
,
3691 downgrade_write(&mdsc
->snap_rwsem
);
3693 down_read(&mdsc
->snap_rwsem
);
3695 handle_cap_import(mdsc
, inode
, h
, peer
, session
,
3697 handle_cap_grant(mdsc
, inode
, h
, &pool_ns
,
3698 inline_version
, inline_data
, inline_len
,
3699 msg
->middle
, session
, cap
, issued
);
3701 ceph_put_snap_realm(mdsc
, realm
);
3705 /* the rest require a cap */
3706 spin_lock(&ci
->i_ceph_lock
);
3707 cap
= __get_cap_for_mds(ceph_inode(inode
), mds
);
3709 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3710 inode
, ceph_ino(inode
), ceph_snap(inode
), mds
);
3711 spin_unlock(&ci
->i_ceph_lock
);
3712 goto flush_cap_releases
;
3715 /* note that each of these drops i_ceph_lock for us */
3717 case CEPH_CAP_OP_REVOKE
:
3718 case CEPH_CAP_OP_GRANT
:
3719 __ceph_caps_issued(ci
, &issued
);
3720 issued
|= __ceph_caps_dirty(ci
);
3721 handle_cap_grant(mdsc
, inode
, h
, &pool_ns
,
3722 inline_version
, inline_data
, inline_len
,
3723 msg
->middle
, session
, cap
, issued
);
3726 case CEPH_CAP_OP_FLUSH_ACK
:
3727 handle_cap_flush_ack(inode
, tid
, h
, session
, cap
);
3730 case CEPH_CAP_OP_TRUNC
:
3731 handle_cap_trunc(inode
, h
, session
);
3735 spin_unlock(&ci
->i_ceph_lock
);
3736 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op
,
3737 ceph_cap_op_name(op
));
3744 * send any cap release message to try to move things
3745 * along for the mds (who clearly thinks we still have this
3748 ceph_send_cap_releases(mdsc
, session
);
3751 mutex_unlock(&session
->s_mutex
);
3754 ceph_put_string(pool_ns
);
3758 pr_err("ceph_handle_caps: corrupt message\n");
3764 * Delayed work handler to process end of delayed cap release LRU list.
3766 void ceph_check_delayed_caps(struct ceph_mds_client
*mdsc
)
3768 struct ceph_inode_info
*ci
;
3769 int flags
= CHECK_CAPS_NODELAY
;
3771 dout("check_delayed_caps\n");
3773 spin_lock(&mdsc
->cap_delay_lock
);
3774 if (list_empty(&mdsc
->cap_delay_list
))
3776 ci
= list_first_entry(&mdsc
->cap_delay_list
,
3777 struct ceph_inode_info
,
3779 if ((ci
->i_ceph_flags
& CEPH_I_FLUSH
) == 0 &&
3780 time_before(jiffies
, ci
->i_hold_caps_max
))
3782 list_del_init(&ci
->i_cap_delay_list
);
3783 spin_unlock(&mdsc
->cap_delay_lock
);
3784 dout("check_delayed_caps on %p\n", &ci
->vfs_inode
);
3785 ceph_check_caps(ci
, flags
, NULL
);
3787 spin_unlock(&mdsc
->cap_delay_lock
);
3791 * Flush all dirty caps to the mds
3793 void ceph_flush_dirty_caps(struct ceph_mds_client
*mdsc
)
3795 struct ceph_inode_info
*ci
;
3796 struct inode
*inode
;
3798 dout("flush_dirty_caps\n");
3799 spin_lock(&mdsc
->cap_dirty_lock
);
3800 while (!list_empty(&mdsc
->cap_dirty
)) {
3801 ci
= list_first_entry(&mdsc
->cap_dirty
, struct ceph_inode_info
,
3803 inode
= &ci
->vfs_inode
;
3805 dout("flush_dirty_caps %p\n", inode
);
3806 spin_unlock(&mdsc
->cap_dirty_lock
);
3807 ceph_check_caps(ci
, CHECK_CAPS_NODELAY
|CHECK_CAPS_FLUSH
, NULL
);
3809 spin_lock(&mdsc
->cap_dirty_lock
);
3811 spin_unlock(&mdsc
->cap_dirty_lock
);
3812 dout("flush_dirty_caps done\n");
3815 void __ceph_get_fmode(struct ceph_inode_info
*ci
, int fmode
)
3818 int bits
= (fmode
<< 1) | 1;
3819 for (i
= 0; i
< CEPH_FILE_MODE_BITS
; i
++) {
3820 if (bits
& (1 << i
))
3821 ci
->i_nr_by_mode
[i
]++;
3826 * Drop open file reference. If we were the last open file,
3827 * we may need to release capabilities to the MDS (or schedule
3828 * their delayed release).
3830 void ceph_put_fmode(struct ceph_inode_info
*ci
, int fmode
)
3833 int bits
= (fmode
<< 1) | 1;
3834 spin_lock(&ci
->i_ceph_lock
);
3835 for (i
= 0; i
< CEPH_FILE_MODE_BITS
; i
++) {
3836 if (bits
& (1 << i
)) {
3837 BUG_ON(ci
->i_nr_by_mode
[i
] == 0);
3838 if (--ci
->i_nr_by_mode
[i
] == 0)
3842 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
3843 &ci
->vfs_inode
, fmode
,
3844 ci
->i_nr_by_mode
[0], ci
->i_nr_by_mode
[1],
3845 ci
->i_nr_by_mode
[2], ci
->i_nr_by_mode
[3]);
3846 spin_unlock(&ci
->i_ceph_lock
);
3848 if (last
&& ci
->i_vino
.snap
== CEPH_NOSNAP
)
3849 ceph_check_caps(ci
, 0, NULL
);
3853 * Helpers for embedding cap and dentry lease releases into mds
3856 * @force is used by dentry_release (below) to force inclusion of a
3857 * record for the directory inode, even when there aren't any caps to
3860 int ceph_encode_inode_release(void **p
, struct inode
*inode
,
3861 int mds
, int drop
, int unless
, int force
)
3863 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3864 struct ceph_cap
*cap
;
3865 struct ceph_mds_request_release
*rel
= *p
;
3869 spin_lock(&ci
->i_ceph_lock
);
3870 used
= __ceph_caps_used(ci
);
3871 dirty
= __ceph_caps_dirty(ci
);
3873 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3874 inode
, mds
, ceph_cap_string(used
|dirty
), ceph_cap_string(drop
),
3875 ceph_cap_string(unless
));
3877 /* only drop unused, clean caps */
3878 drop
&= ~(used
| dirty
);
3880 cap
= __get_cap_for_mds(ci
, mds
);
3881 if (cap
&& __cap_is_valid(cap
)) {
3883 ((cap
->issued
& drop
) &&
3884 (cap
->issued
& unless
) == 0)) {
3885 if ((cap
->issued
& drop
) &&
3886 (cap
->issued
& unless
) == 0) {
3887 int wanted
= __ceph_caps_wanted(ci
);
3888 if ((ci
->i_ceph_flags
& CEPH_I_NODELAY
) == 0)
3889 wanted
|= cap
->mds_wanted
;
3890 dout("encode_inode_release %p cap %p "
3891 "%s -> %s, wanted %s -> %s\n", inode
, cap
,
3892 ceph_cap_string(cap
->issued
),
3893 ceph_cap_string(cap
->issued
& ~drop
),
3894 ceph_cap_string(cap
->mds_wanted
),
3895 ceph_cap_string(wanted
));
3897 cap
->issued
&= ~drop
;
3898 cap
->implemented
&= ~drop
;
3899 cap
->mds_wanted
= wanted
;
3901 dout("encode_inode_release %p cap %p %s"
3902 " (force)\n", inode
, cap
,
3903 ceph_cap_string(cap
->issued
));
3906 rel
->ino
= cpu_to_le64(ceph_ino(inode
));
3907 rel
->cap_id
= cpu_to_le64(cap
->cap_id
);
3908 rel
->seq
= cpu_to_le32(cap
->seq
);
3909 rel
->issue_seq
= cpu_to_le32(cap
->issue_seq
);
3910 rel
->mseq
= cpu_to_le32(cap
->mseq
);
3911 rel
->caps
= cpu_to_le32(cap
->implemented
);
3912 rel
->wanted
= cpu_to_le32(cap
->mds_wanted
);
3918 dout("encode_inode_release %p cap %p %s\n",
3919 inode
, cap
, ceph_cap_string(cap
->issued
));
3922 spin_unlock(&ci
->i_ceph_lock
);
3926 int ceph_encode_dentry_release(void **p
, struct dentry
*dentry
,
3927 int mds
, int drop
, int unless
)
3929 struct inode
*dir
= d_inode(dentry
->d_parent
);
3930 struct ceph_mds_request_release
*rel
= *p
;
3931 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
3936 * force an record for the directory caps if we have a dentry lease.
3937 * this is racy (can't take i_ceph_lock and d_lock together), but it
3938 * doesn't have to be perfect; the mds will revoke anything we don't
3941 spin_lock(&dentry
->d_lock
);
3942 if (di
->lease_session
&& di
->lease_session
->s_mds
== mds
)
3944 spin_unlock(&dentry
->d_lock
);
3946 ret
= ceph_encode_inode_release(p
, dir
, mds
, drop
, unless
, force
);
3948 spin_lock(&dentry
->d_lock
);
3949 if (ret
&& di
->lease_session
&& di
->lease_session
->s_mds
== mds
) {
3950 dout("encode_dentry_release %p mds%d seq %d\n",
3951 dentry
, mds
, (int)di
->lease_seq
);
3952 rel
->dname_len
= cpu_to_le32(dentry
->d_name
.len
);
3953 memcpy(*p
, dentry
->d_name
.name
, dentry
->d_name
.len
);
3954 *p
+= dentry
->d_name
.len
;
3955 rel
->dname_seq
= cpu_to_le32(di
->lease_seq
);
3956 __ceph_mdsc_drop_dentry_lease(dentry
);
3958 spin_unlock(&dentry
->d_lock
);