1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
5 #include <linux/kernel.h>
6 #include <linux/sched/signal.h>
7 #include <linux/slab.h>
8 #include <linux/vmalloc.h>
9 #include <linux/wait.h>
10 #include <linux/writeback.h>
11 #include <linux/iversion.h>
14 #include "mds_client.h"
16 #include <linux/ceph/decode.h>
17 #include <linux/ceph/messenger.h>
20 * Capability management
22 * The Ceph metadata servers control client access to inode metadata
23 * and file data by issuing capabilities, granting clients permission
24 * to read and/or write both inode field and file data to OSDs
25 * (storage nodes). Each capability consists of a set of bits
26 * indicating which operations are allowed.
28 * If the client holds a *_SHARED cap, the client has a coherent value
29 * that can be safely read from the cached inode.
31 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
32 * client is allowed to change inode attributes (e.g., file size,
33 * mtime), note its dirty state in the ceph_cap, and asynchronously
34 * flush that metadata change to the MDS.
36 * In the event of a conflicting operation (perhaps by another
37 * client), the MDS will revoke the conflicting client capabilities.
39 * In order for a client to cache an inode, it must hold a capability
40 * with at least one MDS server. When inodes are released, release
41 * notifications are batched and periodically sent en masse to the MDS
42 * cluster to release server state.
45 static u64
__get_oldest_flush_tid(struct ceph_mds_client
*mdsc
);
46 static void __kick_flushing_caps(struct ceph_mds_client
*mdsc
,
47 struct ceph_mds_session
*session
,
48 struct ceph_inode_info
*ci
,
49 u64 oldest_flush_tid
);
52 * Generate readable cap strings for debugging output.
54 #define MAX_CAP_STR 20
55 static char cap_str
[MAX_CAP_STR
][40];
56 static DEFINE_SPINLOCK(cap_str_lock
);
57 static int last_cap_str
;
59 static char *gcap_string(char *s
, int c
)
61 if (c
& CEPH_CAP_GSHARED
)
63 if (c
& CEPH_CAP_GEXCL
)
65 if (c
& CEPH_CAP_GCACHE
)
71 if (c
& CEPH_CAP_GBUFFER
)
73 if (c
& CEPH_CAP_GWREXTEND
)
75 if (c
& CEPH_CAP_GLAZYIO
)
80 const char *ceph_cap_string(int caps
)
86 spin_lock(&cap_str_lock
);
88 if (last_cap_str
== MAX_CAP_STR
)
90 spin_unlock(&cap_str_lock
);
94 if (caps
& CEPH_CAP_PIN
)
97 c
= (caps
>> CEPH_CAP_SAUTH
) & 3;
100 s
= gcap_string(s
, c
);
103 c
= (caps
>> CEPH_CAP_SLINK
) & 3;
106 s
= gcap_string(s
, c
);
109 c
= (caps
>> CEPH_CAP_SXATTR
) & 3;
112 s
= gcap_string(s
, c
);
115 c
= caps
>> CEPH_CAP_SFILE
;
118 s
= gcap_string(s
, c
);
127 void ceph_caps_init(struct ceph_mds_client
*mdsc
)
129 INIT_LIST_HEAD(&mdsc
->caps_list
);
130 spin_lock_init(&mdsc
->caps_list_lock
);
133 void ceph_caps_finalize(struct ceph_mds_client
*mdsc
)
135 struct ceph_cap
*cap
;
137 spin_lock(&mdsc
->caps_list_lock
);
138 while (!list_empty(&mdsc
->caps_list
)) {
139 cap
= list_first_entry(&mdsc
->caps_list
,
140 struct ceph_cap
, caps_item
);
141 list_del(&cap
->caps_item
);
142 kmem_cache_free(ceph_cap_cachep
, cap
);
144 mdsc
->caps_total_count
= 0;
145 mdsc
->caps_avail_count
= 0;
146 mdsc
->caps_use_count
= 0;
147 mdsc
->caps_reserve_count
= 0;
148 mdsc
->caps_min_count
= 0;
149 spin_unlock(&mdsc
->caps_list_lock
);
152 void ceph_adjust_caps_max_min(struct ceph_mds_client
*mdsc
,
153 struct ceph_mount_options
*fsopt
)
155 spin_lock(&mdsc
->caps_list_lock
);
156 mdsc
->caps_min_count
= fsopt
->max_readdir
;
157 if (mdsc
->caps_min_count
< 1024)
158 mdsc
->caps_min_count
= 1024;
159 mdsc
->caps_use_max
= fsopt
->caps_max
;
160 if (mdsc
->caps_use_max
> 0 &&
161 mdsc
->caps_use_max
< mdsc
->caps_min_count
)
162 mdsc
->caps_use_max
= mdsc
->caps_min_count
;
163 spin_unlock(&mdsc
->caps_list_lock
);
166 static void __ceph_unreserve_caps(struct ceph_mds_client
*mdsc
, int nr_caps
)
168 struct ceph_cap
*cap
;
172 BUG_ON(mdsc
->caps_reserve_count
< nr_caps
);
173 mdsc
->caps_reserve_count
-= nr_caps
;
174 if (mdsc
->caps_avail_count
>=
175 mdsc
->caps_reserve_count
+ mdsc
->caps_min_count
) {
176 mdsc
->caps_total_count
-= nr_caps
;
177 for (i
= 0; i
< nr_caps
; i
++) {
178 cap
= list_first_entry(&mdsc
->caps_list
,
179 struct ceph_cap
, caps_item
);
180 list_del(&cap
->caps_item
);
181 kmem_cache_free(ceph_cap_cachep
, cap
);
184 mdsc
->caps_avail_count
+= nr_caps
;
187 dout("%s: caps %d = %d used + %d resv + %d avail\n",
189 mdsc
->caps_total_count
, mdsc
->caps_use_count
,
190 mdsc
->caps_reserve_count
, mdsc
->caps_avail_count
);
191 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
192 mdsc
->caps_reserve_count
+
193 mdsc
->caps_avail_count
);
198 * Called under mdsc->mutex.
200 int ceph_reserve_caps(struct ceph_mds_client
*mdsc
,
201 struct ceph_cap_reservation
*ctx
, int need
)
204 struct ceph_cap
*cap
;
209 bool trimmed
= false;
210 struct ceph_mds_session
*s
;
213 dout("reserve caps ctx=%p need=%d\n", ctx
, need
);
215 /* first reserve any caps that are already allocated */
216 spin_lock(&mdsc
->caps_list_lock
);
217 if (mdsc
->caps_avail_count
>= need
)
220 have
= mdsc
->caps_avail_count
;
221 mdsc
->caps_avail_count
-= have
;
222 mdsc
->caps_reserve_count
+= have
;
223 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
224 mdsc
->caps_reserve_count
+
225 mdsc
->caps_avail_count
);
226 spin_unlock(&mdsc
->caps_list_lock
);
228 for (i
= have
; i
< need
; ) {
229 cap
= kmem_cache_alloc(ceph_cap_cachep
, GFP_NOFS
);
231 list_add(&cap
->caps_item
, &newcaps
);
238 for (j
= 0; j
< mdsc
->max_sessions
; j
++) {
239 s
= __ceph_lookup_mds_session(mdsc
, j
);
242 mutex_unlock(&mdsc
->mutex
);
244 mutex_lock(&s
->s_mutex
);
245 max_caps
= s
->s_nr_caps
- (need
- i
);
246 ceph_trim_caps(mdsc
, s
, max_caps
);
247 mutex_unlock(&s
->s_mutex
);
249 ceph_put_mds_session(s
);
250 mutex_lock(&mdsc
->mutex
);
254 spin_lock(&mdsc
->caps_list_lock
);
255 if (mdsc
->caps_avail_count
) {
257 if (mdsc
->caps_avail_count
>= need
- i
)
258 more_have
= need
- i
;
260 more_have
= mdsc
->caps_avail_count
;
264 mdsc
->caps_avail_count
-= more_have
;
265 mdsc
->caps_reserve_count
+= more_have
;
268 spin_unlock(&mdsc
->caps_list_lock
);
273 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
274 ctx
, need
, have
+ alloc
);
280 BUG_ON(have
+ alloc
!= need
);
285 spin_lock(&mdsc
->caps_list_lock
);
286 mdsc
->caps_total_count
+= alloc
;
287 mdsc
->caps_reserve_count
+= alloc
;
288 list_splice(&newcaps
, &mdsc
->caps_list
);
290 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
291 mdsc
->caps_reserve_count
+
292 mdsc
->caps_avail_count
);
295 __ceph_unreserve_caps(mdsc
, have
+ alloc
);
297 spin_unlock(&mdsc
->caps_list_lock
);
299 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
300 ctx
, mdsc
->caps_total_count
, mdsc
->caps_use_count
,
301 mdsc
->caps_reserve_count
, mdsc
->caps_avail_count
);
305 void ceph_unreserve_caps(struct ceph_mds_client
*mdsc
,
306 struct ceph_cap_reservation
*ctx
)
308 bool reclaim
= false;
312 dout("unreserve caps ctx=%p count=%d\n", ctx
, ctx
->count
);
313 spin_lock(&mdsc
->caps_list_lock
);
314 __ceph_unreserve_caps(mdsc
, ctx
->count
);
317 if (mdsc
->caps_use_max
> 0 &&
318 mdsc
->caps_use_count
> mdsc
->caps_use_max
)
320 spin_unlock(&mdsc
->caps_list_lock
);
323 ceph_reclaim_caps_nr(mdsc
, ctx
->used
);
326 struct ceph_cap
*ceph_get_cap(struct ceph_mds_client
*mdsc
,
327 struct ceph_cap_reservation
*ctx
)
329 struct ceph_cap
*cap
= NULL
;
331 /* temporary, until we do something about cap import/export */
333 cap
= kmem_cache_alloc(ceph_cap_cachep
, GFP_NOFS
);
335 spin_lock(&mdsc
->caps_list_lock
);
336 mdsc
->caps_use_count
++;
337 mdsc
->caps_total_count
++;
338 spin_unlock(&mdsc
->caps_list_lock
);
340 spin_lock(&mdsc
->caps_list_lock
);
341 if (mdsc
->caps_avail_count
) {
342 BUG_ON(list_empty(&mdsc
->caps_list
));
344 mdsc
->caps_avail_count
--;
345 mdsc
->caps_use_count
++;
346 cap
= list_first_entry(&mdsc
->caps_list
,
347 struct ceph_cap
, caps_item
);
348 list_del(&cap
->caps_item
);
350 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
351 mdsc
->caps_reserve_count
+ mdsc
->caps_avail_count
);
353 spin_unlock(&mdsc
->caps_list_lock
);
359 spin_lock(&mdsc
->caps_list_lock
);
360 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
361 ctx
, ctx
->count
, mdsc
->caps_total_count
, mdsc
->caps_use_count
,
362 mdsc
->caps_reserve_count
, mdsc
->caps_avail_count
);
364 BUG_ON(ctx
->count
> mdsc
->caps_reserve_count
);
365 BUG_ON(list_empty(&mdsc
->caps_list
));
369 mdsc
->caps_reserve_count
--;
370 mdsc
->caps_use_count
++;
372 cap
= list_first_entry(&mdsc
->caps_list
, struct ceph_cap
, caps_item
);
373 list_del(&cap
->caps_item
);
375 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
376 mdsc
->caps_reserve_count
+ mdsc
->caps_avail_count
);
377 spin_unlock(&mdsc
->caps_list_lock
);
381 void ceph_put_cap(struct ceph_mds_client
*mdsc
, struct ceph_cap
*cap
)
383 spin_lock(&mdsc
->caps_list_lock
);
384 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
385 cap
, mdsc
->caps_total_count
, mdsc
->caps_use_count
,
386 mdsc
->caps_reserve_count
, mdsc
->caps_avail_count
);
387 mdsc
->caps_use_count
--;
389 * Keep some preallocated caps around (ceph_min_count), to
390 * avoid lots of free/alloc churn.
392 if (mdsc
->caps_avail_count
>= mdsc
->caps_reserve_count
+
393 mdsc
->caps_min_count
) {
394 mdsc
->caps_total_count
--;
395 kmem_cache_free(ceph_cap_cachep
, cap
);
397 mdsc
->caps_avail_count
++;
398 list_add(&cap
->caps_item
, &mdsc
->caps_list
);
401 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
402 mdsc
->caps_reserve_count
+ mdsc
->caps_avail_count
);
403 spin_unlock(&mdsc
->caps_list_lock
);
406 void ceph_reservation_status(struct ceph_fs_client
*fsc
,
407 int *total
, int *avail
, int *used
, int *reserved
,
410 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
412 spin_lock(&mdsc
->caps_list_lock
);
415 *total
= mdsc
->caps_total_count
;
417 *avail
= mdsc
->caps_avail_count
;
419 *used
= mdsc
->caps_use_count
;
421 *reserved
= mdsc
->caps_reserve_count
;
423 *min
= mdsc
->caps_min_count
;
425 spin_unlock(&mdsc
->caps_list_lock
);
429 * Find ceph_cap for given mds, if any.
431 * Called with i_ceph_lock held.
433 static struct ceph_cap
*__get_cap_for_mds(struct ceph_inode_info
*ci
, int mds
)
435 struct ceph_cap
*cap
;
436 struct rb_node
*n
= ci
->i_caps
.rb_node
;
439 cap
= rb_entry(n
, struct ceph_cap
, ci_node
);
442 else if (mds
> cap
->mds
)
450 struct ceph_cap
*ceph_get_cap_for_mds(struct ceph_inode_info
*ci
, int mds
)
452 struct ceph_cap
*cap
;
454 spin_lock(&ci
->i_ceph_lock
);
455 cap
= __get_cap_for_mds(ci
, mds
);
456 spin_unlock(&ci
->i_ceph_lock
);
461 * Called under i_ceph_lock.
463 static void __insert_cap_node(struct ceph_inode_info
*ci
,
464 struct ceph_cap
*new)
466 struct rb_node
**p
= &ci
->i_caps
.rb_node
;
467 struct rb_node
*parent
= NULL
;
468 struct ceph_cap
*cap
= NULL
;
472 cap
= rb_entry(parent
, struct ceph_cap
, ci_node
);
473 if (new->mds
< cap
->mds
)
475 else if (new->mds
> cap
->mds
)
481 rb_link_node(&new->ci_node
, parent
, p
);
482 rb_insert_color(&new->ci_node
, &ci
->i_caps
);
486 * (re)set cap hold timeouts, which control the delayed release
487 * of unused caps back to the MDS. Should be called on cap use.
489 static void __cap_set_timeouts(struct ceph_mds_client
*mdsc
,
490 struct ceph_inode_info
*ci
)
492 struct ceph_mount_options
*opt
= mdsc
->fsc
->mount_options
;
493 ci
->i_hold_caps_max
= round_jiffies(jiffies
+
494 opt
->caps_wanted_delay_max
* HZ
);
495 dout("__cap_set_timeouts %p %lu\n", &ci
->vfs_inode
,
496 ci
->i_hold_caps_max
- jiffies
);
500 * (Re)queue cap at the end of the delayed cap release list.
502 * If I_FLUSH is set, leave the inode at the front of the list.
504 * Caller holds i_ceph_lock
505 * -> we take mdsc->cap_delay_lock
507 static void __cap_delay_requeue(struct ceph_mds_client
*mdsc
,
508 struct ceph_inode_info
*ci
)
510 dout("__cap_delay_requeue %p flags 0x%lx at %lu\n", &ci
->vfs_inode
,
511 ci
->i_ceph_flags
, ci
->i_hold_caps_max
);
512 if (!mdsc
->stopping
) {
513 spin_lock(&mdsc
->cap_delay_lock
);
514 if (!list_empty(&ci
->i_cap_delay_list
)) {
515 if (ci
->i_ceph_flags
& CEPH_I_FLUSH
)
517 list_del_init(&ci
->i_cap_delay_list
);
519 __cap_set_timeouts(mdsc
, ci
);
520 list_add_tail(&ci
->i_cap_delay_list
, &mdsc
->cap_delay_list
);
522 spin_unlock(&mdsc
->cap_delay_lock
);
527 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
528 * indicating we should send a cap message to flush dirty metadata
529 * asap, and move to the front of the delayed cap list.
531 static void __cap_delay_requeue_front(struct ceph_mds_client
*mdsc
,
532 struct ceph_inode_info
*ci
)
534 dout("__cap_delay_requeue_front %p\n", &ci
->vfs_inode
);
535 spin_lock(&mdsc
->cap_delay_lock
);
536 ci
->i_ceph_flags
|= CEPH_I_FLUSH
;
537 if (!list_empty(&ci
->i_cap_delay_list
))
538 list_del_init(&ci
->i_cap_delay_list
);
539 list_add(&ci
->i_cap_delay_list
, &mdsc
->cap_delay_list
);
540 spin_unlock(&mdsc
->cap_delay_lock
);
544 * Cancel delayed work on cap.
546 * Caller must hold i_ceph_lock.
548 static void __cap_delay_cancel(struct ceph_mds_client
*mdsc
,
549 struct ceph_inode_info
*ci
)
551 dout("__cap_delay_cancel %p\n", &ci
->vfs_inode
);
552 if (list_empty(&ci
->i_cap_delay_list
))
554 spin_lock(&mdsc
->cap_delay_lock
);
555 list_del_init(&ci
->i_cap_delay_list
);
556 spin_unlock(&mdsc
->cap_delay_lock
);
559 /* Common issue checks for add_cap, handle_cap_grant. */
560 static void __check_cap_issue(struct ceph_inode_info
*ci
, struct ceph_cap
*cap
,
563 unsigned had
= __ceph_caps_issued(ci
, NULL
);
565 lockdep_assert_held(&ci
->i_ceph_lock
);
568 * Each time we receive FILE_CACHE anew, we increment
571 if (S_ISREG(ci
->vfs_inode
.i_mode
) &&
572 (issued
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)) &&
573 (had
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)) == 0) {
578 * If FILE_SHARED is newly issued, mark dir not complete. We don't
579 * know what happened to this directory while we didn't have the cap.
580 * If FILE_SHARED is being revoked, also mark dir not complete. It
581 * stops on-going cached readdir.
583 if ((issued
& CEPH_CAP_FILE_SHARED
) != (had
& CEPH_CAP_FILE_SHARED
)) {
584 if (issued
& CEPH_CAP_FILE_SHARED
)
585 atomic_inc(&ci
->i_shared_gen
);
586 if (S_ISDIR(ci
->vfs_inode
.i_mode
)) {
587 dout(" marking %p NOT complete\n", &ci
->vfs_inode
);
588 __ceph_dir_clear_complete(ci
);
592 /* Wipe saved layout if we're losing DIR_CREATE caps */
593 if (S_ISDIR(ci
->vfs_inode
.i_mode
) && (had
& CEPH_CAP_DIR_CREATE
) &&
594 !(issued
& CEPH_CAP_DIR_CREATE
)) {
595 ceph_put_string(rcu_dereference_raw(ci
->i_cached_layout
.pool_ns
));
596 memset(&ci
->i_cached_layout
, 0, sizeof(ci
->i_cached_layout
));
601 * change_auth_cap_ses - move inode to appropriate lists when auth caps change
602 * @ci: inode to be moved
603 * @session: new auth caps session
605 static void change_auth_cap_ses(struct ceph_inode_info
*ci
,
606 struct ceph_mds_session
*session
)
608 lockdep_assert_held(&ci
->i_ceph_lock
);
610 if (list_empty(&ci
->i_dirty_item
) && list_empty(&ci
->i_flushing_item
))
613 spin_lock(&session
->s_mdsc
->cap_dirty_lock
);
614 if (!list_empty(&ci
->i_dirty_item
))
615 list_move(&ci
->i_dirty_item
, &session
->s_cap_dirty
);
616 if (!list_empty(&ci
->i_flushing_item
))
617 list_move_tail(&ci
->i_flushing_item
, &session
->s_cap_flushing
);
618 spin_unlock(&session
->s_mdsc
->cap_dirty_lock
);
622 * Add a capability under the given MDS session.
624 * Caller should hold session snap_rwsem (read) and ci->i_ceph_lock
626 * @fmode is the open file mode, if we are opening a file, otherwise
627 * it is < 0. (This is so we can atomically add the cap and add an
628 * open file reference to it.)
630 void ceph_add_cap(struct inode
*inode
,
631 struct ceph_mds_session
*session
, u64 cap_id
,
632 unsigned issued
, unsigned wanted
,
633 unsigned seq
, unsigned mseq
, u64 realmino
, int flags
,
634 struct ceph_cap
**new_cap
)
636 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
637 struct ceph_inode_info
*ci
= ceph_inode(inode
);
638 struct ceph_cap
*cap
;
639 int mds
= session
->s_mds
;
643 lockdep_assert_held(&ci
->i_ceph_lock
);
645 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode
,
646 session
->s_mds
, cap_id
, ceph_cap_string(issued
), seq
);
648 spin_lock(&session
->s_gen_ttl_lock
);
649 gen
= session
->s_cap_gen
;
650 spin_unlock(&session
->s_gen_ttl_lock
);
652 cap
= __get_cap_for_mds(ci
, mds
);
658 cap
->implemented
= 0;
664 __insert_cap_node(ci
, cap
);
666 /* add to session cap list */
667 cap
->session
= session
;
668 spin_lock(&session
->s_cap_lock
);
669 list_add_tail(&cap
->session_caps
, &session
->s_caps
);
670 session
->s_nr_caps
++;
671 atomic64_inc(&mdsc
->metric
.total_caps
);
672 spin_unlock(&session
->s_cap_lock
);
674 spin_lock(&session
->s_cap_lock
);
675 list_move_tail(&cap
->session_caps
, &session
->s_caps
);
676 spin_unlock(&session
->s_cap_lock
);
678 if (cap
->cap_gen
< gen
)
679 cap
->issued
= cap
->implemented
= CEPH_CAP_PIN
;
682 * auth mds of the inode changed. we received the cap export
683 * message, but still haven't received the cap import message.
684 * handle_cap_export() updated the new auth MDS' cap.
686 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
687 * a message that was send before the cap import message. So
690 if (ceph_seq_cmp(seq
, cap
->seq
) <= 0) {
691 WARN_ON(cap
!= ci
->i_auth_cap
);
692 WARN_ON(cap
->cap_id
!= cap_id
);
695 issued
|= cap
->issued
;
696 flags
|= CEPH_CAP_FLAG_AUTH
;
700 if (!ci
->i_snap_realm
||
701 ((flags
& CEPH_CAP_FLAG_AUTH
) &&
702 realmino
!= (u64
)-1 && ci
->i_snap_realm
->ino
!= realmino
)) {
704 * add this inode to the appropriate snap realm
706 struct ceph_snap_realm
*realm
= ceph_lookup_snap_realm(mdsc
,
709 struct ceph_snap_realm
*oldrealm
= ci
->i_snap_realm
;
711 spin_lock(&oldrealm
->inodes_with_caps_lock
);
712 list_del_init(&ci
->i_snap_realm_item
);
713 spin_unlock(&oldrealm
->inodes_with_caps_lock
);
716 spin_lock(&realm
->inodes_with_caps_lock
);
717 list_add(&ci
->i_snap_realm_item
,
718 &realm
->inodes_with_caps
);
719 ci
->i_snap_realm
= realm
;
720 if (realm
->ino
== ci
->i_vino
.ino
)
721 realm
->inode
= inode
;
722 spin_unlock(&realm
->inodes_with_caps_lock
);
725 ceph_put_snap_realm(mdsc
, oldrealm
);
727 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
733 __check_cap_issue(ci
, cap
, issued
);
736 * If we are issued caps we don't want, or the mds' wanted
737 * value appears to be off, queue a check so we'll release
738 * later and/or update the mds wanted value.
740 actual_wanted
= __ceph_caps_wanted(ci
);
741 if ((wanted
& ~actual_wanted
) ||
742 (issued
& ~actual_wanted
& CEPH_CAP_ANY_WR
)) {
743 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
744 ceph_cap_string(issued
), ceph_cap_string(wanted
),
745 ceph_cap_string(actual_wanted
));
746 __cap_delay_requeue(mdsc
, ci
);
749 if (flags
& CEPH_CAP_FLAG_AUTH
) {
750 if (!ci
->i_auth_cap
||
751 ceph_seq_cmp(ci
->i_auth_cap
->mseq
, mseq
) < 0) {
752 if (ci
->i_auth_cap
&&
753 ci
->i_auth_cap
->session
!= cap
->session
)
754 change_auth_cap_ses(ci
, cap
->session
);
755 ci
->i_auth_cap
= cap
;
756 cap
->mds_wanted
= wanted
;
759 WARN_ON(ci
->i_auth_cap
== cap
);
762 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
763 inode
, ceph_vinop(inode
), cap
, ceph_cap_string(issued
),
764 ceph_cap_string(issued
|cap
->issued
), seq
, mds
);
765 cap
->cap_id
= cap_id
;
766 cap
->issued
= issued
;
767 cap
->implemented
|= issued
;
768 if (ceph_seq_cmp(mseq
, cap
->mseq
) > 0)
769 cap
->mds_wanted
= wanted
;
771 cap
->mds_wanted
|= wanted
;
773 cap
->issue_seq
= seq
;
779 * Return true if cap has not timed out and belongs to the current
780 * generation of the MDS session (i.e. has not gone 'stale' due to
781 * us losing touch with the mds).
783 static int __cap_is_valid(struct ceph_cap
*cap
)
788 spin_lock(&cap
->session
->s_gen_ttl_lock
);
789 gen
= cap
->session
->s_cap_gen
;
790 ttl
= cap
->session
->s_cap_ttl
;
791 spin_unlock(&cap
->session
->s_gen_ttl_lock
);
793 if (cap
->cap_gen
< gen
|| time_after_eq(jiffies
, ttl
)) {
794 dout("__cap_is_valid %p cap %p issued %s "
795 "but STALE (gen %u vs %u)\n", &cap
->ci
->vfs_inode
,
796 cap
, ceph_cap_string(cap
->issued
), cap
->cap_gen
, gen
);
804 * Return set of valid cap bits issued to us. Note that caps time
805 * out, and may be invalidated in bulk if the client session times out
806 * and session->s_cap_gen is bumped.
808 int __ceph_caps_issued(struct ceph_inode_info
*ci
, int *implemented
)
810 int have
= ci
->i_snap_caps
;
811 struct ceph_cap
*cap
;
816 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
817 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
818 if (!__cap_is_valid(cap
))
820 dout("__ceph_caps_issued %p cap %p issued %s\n",
821 &ci
->vfs_inode
, cap
, ceph_cap_string(cap
->issued
));
824 *implemented
|= cap
->implemented
;
827 * exclude caps issued by non-auth MDS, but are been revoking
828 * by the auth MDS. The non-auth MDS should be revoking/exporting
829 * these caps, but the message is delayed.
831 if (ci
->i_auth_cap
) {
832 cap
= ci
->i_auth_cap
;
833 have
&= ~cap
->implemented
| cap
->issued
;
839 * Get cap bits issued by caps other than @ocap
841 int __ceph_caps_issued_other(struct ceph_inode_info
*ci
, struct ceph_cap
*ocap
)
843 int have
= ci
->i_snap_caps
;
844 struct ceph_cap
*cap
;
847 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
848 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
851 if (!__cap_is_valid(cap
))
859 * Move a cap to the end of the LRU (oldest caps at list head, newest
862 static void __touch_cap(struct ceph_cap
*cap
)
864 struct ceph_mds_session
*s
= cap
->session
;
866 spin_lock(&s
->s_cap_lock
);
867 if (!s
->s_cap_iterator
) {
868 dout("__touch_cap %p cap %p mds%d\n", &cap
->ci
->vfs_inode
, cap
,
870 list_move_tail(&cap
->session_caps
, &s
->s_caps
);
872 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
873 &cap
->ci
->vfs_inode
, cap
, s
->s_mds
);
875 spin_unlock(&s
->s_cap_lock
);
879 * Check if we hold the given mask. If so, move the cap(s) to the
880 * front of their respective LRUs. (This is the preferred way for
881 * callers to check for caps they want.)
883 int __ceph_caps_issued_mask(struct ceph_inode_info
*ci
, int mask
, int touch
)
885 struct ceph_cap
*cap
;
887 int have
= ci
->i_snap_caps
;
889 if ((have
& mask
) == mask
) {
890 dout("__ceph_caps_issued_mask ino 0x%lx snap issued %s"
891 " (mask %s)\n", ci
->vfs_inode
.i_ino
,
892 ceph_cap_string(have
),
893 ceph_cap_string(mask
));
897 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
898 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
899 if (!__cap_is_valid(cap
))
901 if ((cap
->issued
& mask
) == mask
) {
902 dout("__ceph_caps_issued_mask ino 0x%lx cap %p issued %s"
903 " (mask %s)\n", ci
->vfs_inode
.i_ino
, cap
,
904 ceph_cap_string(cap
->issued
),
905 ceph_cap_string(mask
));
911 /* does a combination of caps satisfy mask? */
913 if ((have
& mask
) == mask
) {
914 dout("__ceph_caps_issued_mask ino 0x%lx combo issued %s"
915 " (mask %s)\n", ci
->vfs_inode
.i_ino
,
916 ceph_cap_string(cap
->issued
),
917 ceph_cap_string(mask
));
921 /* touch this + preceding caps */
923 for (q
= rb_first(&ci
->i_caps
); q
!= p
;
925 cap
= rb_entry(q
, struct ceph_cap
,
927 if (!__cap_is_valid(cap
))
929 if (cap
->issued
& mask
)
940 int __ceph_caps_issued_mask_metric(struct ceph_inode_info
*ci
, int mask
,
943 struct ceph_fs_client
*fsc
= ceph_sb_to_client(ci
->vfs_inode
.i_sb
);
946 r
= __ceph_caps_issued_mask(ci
, mask
, touch
);
948 ceph_update_cap_hit(&fsc
->mdsc
->metric
);
950 ceph_update_cap_mis(&fsc
->mdsc
->metric
);
955 * Return true if mask caps are currently being revoked by an MDS.
957 int __ceph_caps_revoking_other(struct ceph_inode_info
*ci
,
958 struct ceph_cap
*ocap
, int mask
)
960 struct ceph_cap
*cap
;
963 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
964 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
966 (cap
->implemented
& ~cap
->issued
& mask
))
972 int ceph_caps_revoking(struct ceph_inode_info
*ci
, int mask
)
974 struct inode
*inode
= &ci
->vfs_inode
;
977 spin_lock(&ci
->i_ceph_lock
);
978 ret
= __ceph_caps_revoking_other(ci
, NULL
, mask
);
979 spin_unlock(&ci
->i_ceph_lock
);
980 dout("ceph_caps_revoking %p %s = %d\n", inode
,
981 ceph_cap_string(mask
), ret
);
985 int __ceph_caps_used(struct ceph_inode_info
*ci
)
989 used
|= CEPH_CAP_PIN
;
991 used
|= CEPH_CAP_FILE_RD
;
992 if (ci
->i_rdcache_ref
||
993 (S_ISREG(ci
->vfs_inode
.i_mode
) &&
994 ci
->vfs_inode
.i_data
.nrpages
))
995 used
|= CEPH_CAP_FILE_CACHE
;
997 used
|= CEPH_CAP_FILE_WR
;
998 if (ci
->i_wb_ref
|| ci
->i_wrbuffer_ref
)
999 used
|= CEPH_CAP_FILE_BUFFER
;
1001 used
|= CEPH_CAP_FILE_EXCL
;
1005 #define FMODE_WAIT_BIAS 1000
1008 * wanted, by virtue of open file modes
1010 int __ceph_caps_file_wanted(struct ceph_inode_info
*ci
)
1012 const int PIN_SHIFT
= ffs(CEPH_FILE_MODE_PIN
);
1013 const int RD_SHIFT
= ffs(CEPH_FILE_MODE_RD
);
1014 const int WR_SHIFT
= ffs(CEPH_FILE_MODE_WR
);
1015 const int LAZY_SHIFT
= ffs(CEPH_FILE_MODE_LAZY
);
1016 struct ceph_mount_options
*opt
=
1017 ceph_inode_to_client(&ci
->vfs_inode
)->mount_options
;
1018 unsigned long used_cutoff
= jiffies
- opt
->caps_wanted_delay_max
* HZ
;
1019 unsigned long idle_cutoff
= jiffies
- opt
->caps_wanted_delay_min
* HZ
;
1021 if (S_ISDIR(ci
->vfs_inode
.i_mode
)) {
1024 /* use used_cutoff here, to keep dir's wanted caps longer */
1025 if (ci
->i_nr_by_mode
[RD_SHIFT
] > 0 ||
1026 time_after(ci
->i_last_rd
, used_cutoff
))
1027 want
|= CEPH_CAP_ANY_SHARED
;
1029 if (ci
->i_nr_by_mode
[WR_SHIFT
] > 0 ||
1030 time_after(ci
->i_last_wr
, used_cutoff
)) {
1031 want
|= CEPH_CAP_ANY_SHARED
| CEPH_CAP_FILE_EXCL
;
1032 if (opt
->flags
& CEPH_MOUNT_OPT_ASYNC_DIROPS
)
1033 want
|= CEPH_CAP_ANY_DIR_OPS
;
1036 if (want
|| ci
->i_nr_by_mode
[PIN_SHIFT
] > 0)
1037 want
|= CEPH_CAP_PIN
;
1043 if (ci
->i_nr_by_mode
[RD_SHIFT
] > 0) {
1044 if (ci
->i_nr_by_mode
[RD_SHIFT
] >= FMODE_WAIT_BIAS
||
1045 time_after(ci
->i_last_rd
, used_cutoff
))
1046 bits
|= 1 << RD_SHIFT
;
1047 } else if (time_after(ci
->i_last_rd
, idle_cutoff
)) {
1048 bits
|= 1 << RD_SHIFT
;
1051 if (ci
->i_nr_by_mode
[WR_SHIFT
] > 0) {
1052 if (ci
->i_nr_by_mode
[WR_SHIFT
] >= FMODE_WAIT_BIAS
||
1053 time_after(ci
->i_last_wr
, used_cutoff
))
1054 bits
|= 1 << WR_SHIFT
;
1055 } else if (time_after(ci
->i_last_wr
, idle_cutoff
)) {
1056 bits
|= 1 << WR_SHIFT
;
1059 /* check lazyio only when read/write is wanted */
1060 if ((bits
& (CEPH_FILE_MODE_RDWR
<< 1)) &&
1061 ci
->i_nr_by_mode
[LAZY_SHIFT
] > 0)
1062 bits
|= 1 << LAZY_SHIFT
;
1064 return bits
? ceph_caps_for_mode(bits
>> 1) : 0;
1069 * wanted, by virtue of open file modes AND cap refs (buffered/cached data)
1071 int __ceph_caps_wanted(struct ceph_inode_info
*ci
)
1073 int w
= __ceph_caps_file_wanted(ci
) | __ceph_caps_used(ci
);
1074 if (S_ISDIR(ci
->vfs_inode
.i_mode
)) {
1075 /* we want EXCL if holding caps of dir ops */
1076 if (w
& CEPH_CAP_ANY_DIR_OPS
)
1077 w
|= CEPH_CAP_FILE_EXCL
;
1079 /* we want EXCL if dirty data */
1080 if (w
& CEPH_CAP_FILE_BUFFER
)
1081 w
|= CEPH_CAP_FILE_EXCL
;
1087 * Return caps we have registered with the MDS(s) as 'wanted'.
1089 int __ceph_caps_mds_wanted(struct ceph_inode_info
*ci
, bool check
)
1091 struct ceph_cap
*cap
;
1095 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
1096 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
1097 if (check
&& !__cap_is_valid(cap
))
1099 if (cap
== ci
->i_auth_cap
)
1100 mds_wanted
|= cap
->mds_wanted
;
1102 mds_wanted
|= (cap
->mds_wanted
& ~CEPH_CAP_ANY_FILE_WR
);
1107 int ceph_is_any_caps(struct inode
*inode
)
1109 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1112 spin_lock(&ci
->i_ceph_lock
);
1113 ret
= __ceph_is_any_real_caps(ci
);
1114 spin_unlock(&ci
->i_ceph_lock
);
1119 static void drop_inode_snap_realm(struct ceph_inode_info
*ci
)
1121 struct ceph_snap_realm
*realm
= ci
->i_snap_realm
;
1122 spin_lock(&realm
->inodes_with_caps_lock
);
1123 list_del_init(&ci
->i_snap_realm_item
);
1124 ci
->i_snap_realm_counter
++;
1125 ci
->i_snap_realm
= NULL
;
1126 if (realm
->ino
== ci
->i_vino
.ino
)
1127 realm
->inode
= NULL
;
1128 spin_unlock(&realm
->inodes_with_caps_lock
);
1129 ceph_put_snap_realm(ceph_sb_to_client(ci
->vfs_inode
.i_sb
)->mdsc
,
1134 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
1136 * caller should hold i_ceph_lock.
1137 * caller will not hold session s_mutex if called from destroy_inode.
1139 void __ceph_remove_cap(struct ceph_cap
*cap
, bool queue_release
)
1141 struct ceph_mds_session
*session
= cap
->session
;
1142 struct ceph_inode_info
*ci
= cap
->ci
;
1143 struct ceph_mds_client
*mdsc
=
1144 ceph_sb_to_client(ci
->vfs_inode
.i_sb
)->mdsc
;
1147 dout("__ceph_remove_cap %p from %p\n", cap
, &ci
->vfs_inode
);
1149 /* remove from inode's cap rbtree, and clear auth cap */
1150 rb_erase(&cap
->ci_node
, &ci
->i_caps
);
1151 if (ci
->i_auth_cap
== cap
) {
1152 WARN_ON_ONCE(!list_empty(&ci
->i_dirty_item
));
1153 ci
->i_auth_cap
= NULL
;
1156 /* remove from session list */
1157 spin_lock(&session
->s_cap_lock
);
1158 if (session
->s_cap_iterator
== cap
) {
1159 /* not yet, we are iterating over this very cap */
1160 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1163 list_del_init(&cap
->session_caps
);
1164 session
->s_nr_caps
--;
1165 atomic64_dec(&mdsc
->metric
.total_caps
);
1166 cap
->session
= NULL
;
1169 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1173 * s_cap_reconnect is protected by s_cap_lock. no one changes
1174 * s_cap_gen while session is in the reconnect state.
1176 if (queue_release
&&
1177 (!session
->s_cap_reconnect
|| cap
->cap_gen
== session
->s_cap_gen
)) {
1178 cap
->queue_release
= 1;
1180 __ceph_queue_cap_release(session
, cap
);
1184 cap
->queue_release
= 0;
1186 cap
->cap_ino
= ci
->i_vino
.ino
;
1188 spin_unlock(&session
->s_cap_lock
);
1191 ceph_put_cap(mdsc
, cap
);
1193 if (!__ceph_is_any_real_caps(ci
)) {
1194 /* when reconnect denied, we remove session caps forcibly,
1195 * i_wr_ref can be non-zero. If there are ongoing write,
1196 * keep i_snap_realm.
1198 if (ci
->i_wr_ref
== 0 && ci
->i_snap_realm
)
1199 drop_inode_snap_realm(ci
);
1201 __cap_delay_cancel(mdsc
, ci
);
1205 struct cap_msg_args
{
1206 struct ceph_mds_session
*session
;
1207 u64 ino
, cid
, follows
;
1208 u64 flush_tid
, oldest_flush_tid
, size
, max_size
;
1211 struct ceph_buffer
*xattr_buf
;
1212 struct ceph_buffer
*old_xattr_buf
;
1213 struct timespec64 atime
, mtime
, ctime
, btime
;
1214 int op
, caps
, wanted
, dirty
;
1215 u32 seq
, issue_seq
, mseq
, time_warp_seq
;
1225 * Build and send a cap message to the given MDS.
1227 * Caller should be holding s_mutex.
1229 static int send_cap_msg(struct cap_msg_args
*arg
)
1231 struct ceph_mds_caps
*fc
;
1232 struct ceph_msg
*msg
;
1235 struct ceph_osd_client
*osdc
= &arg
->session
->s_mdsc
->fsc
->client
->osdc
;
1237 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1238 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1239 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(arg
->op
),
1240 arg
->cid
, arg
->ino
, ceph_cap_string(arg
->caps
),
1241 ceph_cap_string(arg
->wanted
), ceph_cap_string(arg
->dirty
),
1242 arg
->seq
, arg
->issue_seq
, arg
->flush_tid
, arg
->oldest_flush_tid
,
1243 arg
->mseq
, arg
->follows
, arg
->size
, arg
->max_size
,
1245 arg
->xattr_buf
? (int)arg
->xattr_buf
->vec
.iov_len
: 0);
1247 /* flock buffer size + inline version + inline data size +
1248 * osd_epoch_barrier + oldest_flush_tid */
1249 extra_len
= 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4;
1250 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPS
, sizeof(*fc
) + extra_len
,
1255 msg
->hdr
.version
= cpu_to_le16(10);
1256 msg
->hdr
.tid
= cpu_to_le64(arg
->flush_tid
);
1258 fc
= msg
->front
.iov_base
;
1259 memset(fc
, 0, sizeof(*fc
));
1261 fc
->cap_id
= cpu_to_le64(arg
->cid
);
1262 fc
->op
= cpu_to_le32(arg
->op
);
1263 fc
->seq
= cpu_to_le32(arg
->seq
);
1264 fc
->issue_seq
= cpu_to_le32(arg
->issue_seq
);
1265 fc
->migrate_seq
= cpu_to_le32(arg
->mseq
);
1266 fc
->caps
= cpu_to_le32(arg
->caps
);
1267 fc
->wanted
= cpu_to_le32(arg
->wanted
);
1268 fc
->dirty
= cpu_to_le32(arg
->dirty
);
1269 fc
->ino
= cpu_to_le64(arg
->ino
);
1270 fc
->snap_follows
= cpu_to_le64(arg
->follows
);
1272 fc
->size
= cpu_to_le64(arg
->size
);
1273 fc
->max_size
= cpu_to_le64(arg
->max_size
);
1274 ceph_encode_timespec64(&fc
->mtime
, &arg
->mtime
);
1275 ceph_encode_timespec64(&fc
->atime
, &arg
->atime
);
1276 ceph_encode_timespec64(&fc
->ctime
, &arg
->ctime
);
1277 fc
->time_warp_seq
= cpu_to_le32(arg
->time_warp_seq
);
1279 fc
->uid
= cpu_to_le32(from_kuid(&init_user_ns
, arg
->uid
));
1280 fc
->gid
= cpu_to_le32(from_kgid(&init_user_ns
, arg
->gid
));
1281 fc
->mode
= cpu_to_le32(arg
->mode
);
1283 fc
->xattr_version
= cpu_to_le64(arg
->xattr_version
);
1284 if (arg
->xattr_buf
) {
1285 msg
->middle
= ceph_buffer_get(arg
->xattr_buf
);
1286 fc
->xattr_len
= cpu_to_le32(arg
->xattr_buf
->vec
.iov_len
);
1287 msg
->hdr
.middle_len
= cpu_to_le32(arg
->xattr_buf
->vec
.iov_len
);
1291 /* flock buffer size (version 2) */
1292 ceph_encode_32(&p
, 0);
1293 /* inline version (version 4) */
1294 ceph_encode_64(&p
, arg
->inline_data
? 0 : CEPH_INLINE_NONE
);
1295 /* inline data size */
1296 ceph_encode_32(&p
, 0);
1298 * osd_epoch_barrier (version 5)
1299 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1300 * case it was recently changed
1302 ceph_encode_32(&p
, READ_ONCE(osdc
->epoch_barrier
));
1303 /* oldest_flush_tid (version 6) */
1304 ceph_encode_64(&p
, arg
->oldest_flush_tid
);
1307 * caller_uid/caller_gid (version 7)
1309 * Currently, we don't properly track which caller dirtied the caps
1310 * last, and force a flush of them when there is a conflict. For now,
1311 * just set this to 0:0, to emulate how the MDS has worked up to now.
1313 ceph_encode_32(&p
, 0);
1314 ceph_encode_32(&p
, 0);
1316 /* pool namespace (version 8) (mds always ignores this) */
1317 ceph_encode_32(&p
, 0);
1319 /* btime and change_attr (version 9) */
1320 ceph_encode_timespec64(p
, &arg
->btime
);
1321 p
+= sizeof(struct ceph_timespec
);
1322 ceph_encode_64(&p
, arg
->change_attr
);
1324 /* Advisory flags (version 10) */
1325 ceph_encode_32(&p
, arg
->flags
);
1327 ceph_con_send(&arg
->session
->s_con
, msg
);
1332 * Queue cap releases when an inode is dropped from our cache.
1334 void __ceph_remove_caps(struct ceph_inode_info
*ci
)
1338 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1339 * may call __ceph_caps_issued_mask() on a freeing inode. */
1340 spin_lock(&ci
->i_ceph_lock
);
1341 p
= rb_first(&ci
->i_caps
);
1343 struct ceph_cap
*cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
1345 __ceph_remove_cap(cap
, true);
1347 spin_unlock(&ci
->i_ceph_lock
);
1351 * Prepare to send a cap message to an MDS. Update the cap state, and populate
1352 * the arg struct with the parameters that will need to be sent. This should
1353 * be done under the i_ceph_lock to guard against changes to cap state.
1355 * Make note of max_size reported/requested from mds, revoked caps
1356 * that have now been implemented.
1358 static void __prep_cap(struct cap_msg_args
*arg
, struct ceph_cap
*cap
,
1359 int op
, int flags
, int used
, int want
, int retain
,
1360 int flushing
, u64 flush_tid
, u64 oldest_flush_tid
)
1362 struct ceph_inode_info
*ci
= cap
->ci
;
1363 struct inode
*inode
= &ci
->vfs_inode
;
1366 lockdep_assert_held(&ci
->i_ceph_lock
);
1368 held
= cap
->issued
| cap
->implemented
;
1369 revoking
= cap
->implemented
& ~cap
->issued
;
1370 retain
&= ~revoking
;
1372 dout("%s %p cap %p session %p %s -> %s (revoking %s)\n",
1373 __func__
, inode
, cap
, cap
->session
,
1374 ceph_cap_string(held
), ceph_cap_string(held
& retain
),
1375 ceph_cap_string(revoking
));
1376 BUG_ON((retain
& CEPH_CAP_PIN
) == 0);
1378 ci
->i_ceph_flags
&= ~CEPH_I_FLUSH
;
1380 cap
->issued
&= retain
; /* drop bits we don't want */
1382 * Wake up any waiters on wanted -> needed transition. This is due to
1383 * the weird transition from buffered to sync IO... we need to flush
1384 * dirty pages _before_ allowing sync writes to avoid reordering.
1386 arg
->wake
= cap
->implemented
& ~cap
->issued
;
1387 cap
->implemented
&= cap
->issued
| used
;
1388 cap
->mds_wanted
= want
;
1390 arg
->session
= cap
->session
;
1391 arg
->ino
= ceph_vino(inode
).ino
;
1392 arg
->cid
= cap
->cap_id
;
1393 arg
->follows
= flushing
? ci
->i_head_snapc
->seq
: 0;
1394 arg
->flush_tid
= flush_tid
;
1395 arg
->oldest_flush_tid
= oldest_flush_tid
;
1397 arg
->size
= inode
->i_size
;
1398 ci
->i_reported_size
= arg
->size
;
1399 arg
->max_size
= ci
->i_wanted_max_size
;
1400 if (cap
== ci
->i_auth_cap
) {
1401 if (want
& CEPH_CAP_ANY_FILE_WR
)
1402 ci
->i_requested_max_size
= arg
->max_size
;
1404 ci
->i_requested_max_size
= 0;
1407 if (flushing
& CEPH_CAP_XATTR_EXCL
) {
1408 arg
->old_xattr_buf
= __ceph_build_xattrs_blob(ci
);
1409 arg
->xattr_version
= ci
->i_xattrs
.version
;
1410 arg
->xattr_buf
= ci
->i_xattrs
.blob
;
1412 arg
->xattr_buf
= NULL
;
1413 arg
->old_xattr_buf
= NULL
;
1416 arg
->mtime
= inode
->i_mtime
;
1417 arg
->atime
= inode
->i_atime
;
1418 arg
->ctime
= inode
->i_ctime
;
1419 arg
->btime
= ci
->i_btime
;
1420 arg
->change_attr
= inode_peek_iversion_raw(inode
);
1423 arg
->caps
= cap
->implemented
;
1425 arg
->dirty
= flushing
;
1427 arg
->seq
= cap
->seq
;
1428 arg
->issue_seq
= cap
->issue_seq
;
1429 arg
->mseq
= cap
->mseq
;
1430 arg
->time_warp_seq
= ci
->i_time_warp_seq
;
1432 arg
->uid
= inode
->i_uid
;
1433 arg
->gid
= inode
->i_gid
;
1434 arg
->mode
= inode
->i_mode
;
1436 arg
->inline_data
= ci
->i_inline_version
!= CEPH_INLINE_NONE
;
1437 if (!(flags
& CEPH_CLIENT_CAPS_PENDING_CAPSNAP
) &&
1438 !list_empty(&ci
->i_cap_snaps
)) {
1439 struct ceph_cap_snap
*capsnap
;
1440 list_for_each_entry_reverse(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
1441 if (capsnap
->cap_flush
.tid
)
1443 if (capsnap
->need_flush
) {
1444 flags
|= CEPH_CLIENT_CAPS_PENDING_CAPSNAP
;
1453 * Send a cap msg on the given inode.
1455 * Caller should hold snap_rwsem (read), s_mutex.
1457 static void __send_cap(struct ceph_mds_client
*mdsc
, struct cap_msg_args
*arg
,
1458 struct ceph_inode_info
*ci
)
1460 struct inode
*inode
= &ci
->vfs_inode
;
1463 ret
= send_cap_msg(arg
);
1465 pr_err("error sending cap msg, ino (%llx.%llx) "
1466 "flushing %s tid %llu, requeue\n",
1467 ceph_vinop(inode
), ceph_cap_string(arg
->dirty
),
1469 spin_lock(&ci
->i_ceph_lock
);
1470 __cap_delay_requeue(mdsc
, ci
);
1471 spin_unlock(&ci
->i_ceph_lock
);
1474 ceph_buffer_put(arg
->old_xattr_buf
);
1477 wake_up_all(&ci
->i_cap_wq
);
1480 static inline int __send_flush_snap(struct inode
*inode
,
1481 struct ceph_mds_session
*session
,
1482 struct ceph_cap_snap
*capsnap
,
1483 u32 mseq
, u64 oldest_flush_tid
)
1485 struct cap_msg_args arg
;
1487 arg
.session
= session
;
1488 arg
.ino
= ceph_vino(inode
).ino
;
1490 arg
.follows
= capsnap
->follows
;
1491 arg
.flush_tid
= capsnap
->cap_flush
.tid
;
1492 arg
.oldest_flush_tid
= oldest_flush_tid
;
1494 arg
.size
= capsnap
->size
;
1496 arg
.xattr_version
= capsnap
->xattr_version
;
1497 arg
.xattr_buf
= capsnap
->xattr_blob
;
1498 arg
.old_xattr_buf
= NULL
;
1500 arg
.atime
= capsnap
->atime
;
1501 arg
.mtime
= capsnap
->mtime
;
1502 arg
.ctime
= capsnap
->ctime
;
1503 arg
.btime
= capsnap
->btime
;
1504 arg
.change_attr
= capsnap
->change_attr
;
1506 arg
.op
= CEPH_CAP_OP_FLUSHSNAP
;
1507 arg
.caps
= capsnap
->issued
;
1509 arg
.dirty
= capsnap
->dirty
;
1514 arg
.time_warp_seq
= capsnap
->time_warp_seq
;
1516 arg
.uid
= capsnap
->uid
;
1517 arg
.gid
= capsnap
->gid
;
1518 arg
.mode
= capsnap
->mode
;
1520 arg
.inline_data
= capsnap
->inline_data
;
1524 return send_cap_msg(&arg
);
1528 * When a snapshot is taken, clients accumulate dirty metadata on
1529 * inodes with capabilities in ceph_cap_snaps to describe the file
1530 * state at the time the snapshot was taken. This must be flushed
1531 * asynchronously back to the MDS once sync writes complete and dirty
1532 * data is written out.
1534 * Called under i_ceph_lock. Takes s_mutex as needed.
1536 static void __ceph_flush_snaps(struct ceph_inode_info
*ci
,
1537 struct ceph_mds_session
*session
)
1538 __releases(ci
->i_ceph_lock
)
1539 __acquires(ci
->i_ceph_lock
)
1541 struct inode
*inode
= &ci
->vfs_inode
;
1542 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
1543 struct ceph_cap_snap
*capsnap
;
1544 u64 oldest_flush_tid
= 0;
1545 u64 first_tid
= 1, last_tid
= 0;
1547 dout("__flush_snaps %p session %p\n", inode
, session
);
1549 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
1551 * we need to wait for sync writes to complete and for dirty
1552 * pages to be written out.
1554 if (capsnap
->dirty_pages
|| capsnap
->writing
)
1557 /* should be removed by ceph_try_drop_cap_snap() */
1558 BUG_ON(!capsnap
->need_flush
);
1560 /* only flush each capsnap once */
1561 if (capsnap
->cap_flush
.tid
> 0) {
1562 dout(" already flushed %p, skipping\n", capsnap
);
1566 spin_lock(&mdsc
->cap_dirty_lock
);
1567 capsnap
->cap_flush
.tid
= ++mdsc
->last_cap_flush_tid
;
1568 list_add_tail(&capsnap
->cap_flush
.g_list
,
1569 &mdsc
->cap_flush_list
);
1570 if (oldest_flush_tid
== 0)
1571 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1572 if (list_empty(&ci
->i_flushing_item
)) {
1573 list_add_tail(&ci
->i_flushing_item
,
1574 &session
->s_cap_flushing
);
1576 spin_unlock(&mdsc
->cap_dirty_lock
);
1578 list_add_tail(&capsnap
->cap_flush
.i_list
,
1579 &ci
->i_cap_flush_list
);
1582 first_tid
= capsnap
->cap_flush
.tid
;
1583 last_tid
= capsnap
->cap_flush
.tid
;
1586 ci
->i_ceph_flags
&= ~CEPH_I_FLUSH_SNAPS
;
1588 while (first_tid
<= last_tid
) {
1589 struct ceph_cap
*cap
= ci
->i_auth_cap
;
1590 struct ceph_cap_flush
*cf
;
1593 if (!(cap
&& cap
->session
== session
)) {
1594 dout("__flush_snaps %p auth cap %p not mds%d, "
1595 "stop\n", inode
, cap
, session
->s_mds
);
1600 list_for_each_entry(cf
, &ci
->i_cap_flush_list
, i_list
) {
1601 if (cf
->tid
>= first_tid
) {
1609 first_tid
= cf
->tid
+ 1;
1611 capsnap
= container_of(cf
, struct ceph_cap_snap
, cap_flush
);
1612 refcount_inc(&capsnap
->nref
);
1613 spin_unlock(&ci
->i_ceph_lock
);
1615 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1616 inode
, capsnap
, cf
->tid
, ceph_cap_string(capsnap
->dirty
));
1618 ret
= __send_flush_snap(inode
, session
, capsnap
, cap
->mseq
,
1621 pr_err("__flush_snaps: error sending cap flushsnap, "
1622 "ino (%llx.%llx) tid %llu follows %llu\n",
1623 ceph_vinop(inode
), cf
->tid
, capsnap
->follows
);
1626 ceph_put_cap_snap(capsnap
);
1627 spin_lock(&ci
->i_ceph_lock
);
1631 void ceph_flush_snaps(struct ceph_inode_info
*ci
,
1632 struct ceph_mds_session
**psession
)
1634 struct inode
*inode
= &ci
->vfs_inode
;
1635 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
1636 struct ceph_mds_session
*session
= NULL
;
1639 dout("ceph_flush_snaps %p\n", inode
);
1641 session
= *psession
;
1643 spin_lock(&ci
->i_ceph_lock
);
1644 if (!(ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
)) {
1645 dout(" no capsnap needs flush, doing nothing\n");
1648 if (!ci
->i_auth_cap
) {
1649 dout(" no auth cap (migrating?), doing nothing\n");
1653 mds
= ci
->i_auth_cap
->session
->s_mds
;
1654 if (session
&& session
->s_mds
!= mds
) {
1655 dout(" oops, wrong session %p mutex\n", session
);
1656 mutex_unlock(&session
->s_mutex
);
1657 ceph_put_mds_session(session
);
1661 spin_unlock(&ci
->i_ceph_lock
);
1662 mutex_lock(&mdsc
->mutex
);
1663 session
= __ceph_lookup_mds_session(mdsc
, mds
);
1664 mutex_unlock(&mdsc
->mutex
);
1666 dout(" inverting session/ino locks on %p\n", session
);
1667 mutex_lock(&session
->s_mutex
);
1672 // make sure flushsnap messages are sent in proper order.
1673 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
)
1674 __kick_flushing_caps(mdsc
, session
, ci
, 0);
1676 __ceph_flush_snaps(ci
, session
);
1678 spin_unlock(&ci
->i_ceph_lock
);
1681 *psession
= session
;
1682 } else if (session
) {
1683 mutex_unlock(&session
->s_mutex
);
1684 ceph_put_mds_session(session
);
1686 /* we flushed them all; remove this inode from the queue */
1687 spin_lock(&mdsc
->snap_flush_lock
);
1688 list_del_init(&ci
->i_snap_flush_item
);
1689 spin_unlock(&mdsc
->snap_flush_lock
);
1693 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1694 * Caller is then responsible for calling __mark_inode_dirty with the
1695 * returned flags value.
1697 int __ceph_mark_dirty_caps(struct ceph_inode_info
*ci
, int mask
,
1698 struct ceph_cap_flush
**pcf
)
1700 struct ceph_mds_client
*mdsc
=
1701 ceph_sb_to_client(ci
->vfs_inode
.i_sb
)->mdsc
;
1702 struct inode
*inode
= &ci
->vfs_inode
;
1703 int was
= ci
->i_dirty_caps
;
1706 lockdep_assert_held(&ci
->i_ceph_lock
);
1708 if (!ci
->i_auth_cap
) {
1709 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1710 "but no auth cap (session was closed?)\n",
1711 inode
, ceph_ino(inode
), ceph_cap_string(mask
));
1715 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci
->vfs_inode
,
1716 ceph_cap_string(mask
), ceph_cap_string(was
),
1717 ceph_cap_string(was
| mask
));
1718 ci
->i_dirty_caps
|= mask
;
1720 struct ceph_mds_session
*session
= ci
->i_auth_cap
->session
;
1722 WARN_ON_ONCE(ci
->i_prealloc_cap_flush
);
1723 swap(ci
->i_prealloc_cap_flush
, *pcf
);
1725 if (!ci
->i_head_snapc
) {
1726 WARN_ON_ONCE(!rwsem_is_locked(&mdsc
->snap_rwsem
));
1727 ci
->i_head_snapc
= ceph_get_snap_context(
1728 ci
->i_snap_realm
->cached_context
);
1730 dout(" inode %p now dirty snapc %p auth cap %p\n",
1731 &ci
->vfs_inode
, ci
->i_head_snapc
, ci
->i_auth_cap
);
1732 BUG_ON(!list_empty(&ci
->i_dirty_item
));
1733 spin_lock(&mdsc
->cap_dirty_lock
);
1734 list_add(&ci
->i_dirty_item
, &session
->s_cap_dirty
);
1735 spin_unlock(&mdsc
->cap_dirty_lock
);
1736 if (ci
->i_flushing_caps
== 0) {
1738 dirty
|= I_DIRTY_SYNC
;
1741 WARN_ON_ONCE(!ci
->i_prealloc_cap_flush
);
1743 BUG_ON(list_empty(&ci
->i_dirty_item
));
1744 if (((was
| ci
->i_flushing_caps
) & CEPH_CAP_FILE_BUFFER
) &&
1745 (mask
& CEPH_CAP_FILE_BUFFER
))
1746 dirty
|= I_DIRTY_DATASYNC
;
1747 __cap_delay_requeue(mdsc
, ci
);
1751 struct ceph_cap_flush
*ceph_alloc_cap_flush(void)
1753 return kmem_cache_alloc(ceph_cap_flush_cachep
, GFP_KERNEL
);
1756 void ceph_free_cap_flush(struct ceph_cap_flush
*cf
)
1759 kmem_cache_free(ceph_cap_flush_cachep
, cf
);
1762 static u64
__get_oldest_flush_tid(struct ceph_mds_client
*mdsc
)
1764 if (!list_empty(&mdsc
->cap_flush_list
)) {
1765 struct ceph_cap_flush
*cf
=
1766 list_first_entry(&mdsc
->cap_flush_list
,
1767 struct ceph_cap_flush
, g_list
);
1774 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1775 * Return true if caller needs to wake up flush waiters.
1777 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client
*mdsc
,
1778 struct ceph_cap_flush
*cf
)
1780 struct ceph_cap_flush
*prev
;
1781 bool wake
= cf
->wake
;
1783 if (wake
&& cf
->g_list
.prev
!= &mdsc
->cap_flush_list
) {
1784 prev
= list_prev_entry(cf
, g_list
);
1788 list_del(&cf
->g_list
);
1792 static bool __detach_cap_flush_from_ci(struct ceph_inode_info
*ci
,
1793 struct ceph_cap_flush
*cf
)
1795 struct ceph_cap_flush
*prev
;
1796 bool wake
= cf
->wake
;
1798 if (wake
&& cf
->i_list
.prev
!= &ci
->i_cap_flush_list
) {
1799 prev
= list_prev_entry(cf
, i_list
);
1803 list_del(&cf
->i_list
);
1808 * Add dirty inode to the flushing list. Assigned a seq number so we
1809 * can wait for caps to flush without starving.
1811 * Called under i_ceph_lock. Returns the flush tid.
1813 static u64
__mark_caps_flushing(struct inode
*inode
,
1814 struct ceph_mds_session
*session
, bool wake
,
1815 u64
*oldest_flush_tid
)
1817 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
1818 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1819 struct ceph_cap_flush
*cf
= NULL
;
1822 lockdep_assert_held(&ci
->i_ceph_lock
);
1823 BUG_ON(ci
->i_dirty_caps
== 0);
1824 BUG_ON(list_empty(&ci
->i_dirty_item
));
1825 BUG_ON(!ci
->i_prealloc_cap_flush
);
1827 flushing
= ci
->i_dirty_caps
;
1828 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1829 ceph_cap_string(flushing
),
1830 ceph_cap_string(ci
->i_flushing_caps
),
1831 ceph_cap_string(ci
->i_flushing_caps
| flushing
));
1832 ci
->i_flushing_caps
|= flushing
;
1833 ci
->i_dirty_caps
= 0;
1834 dout(" inode %p now !dirty\n", inode
);
1836 swap(cf
, ci
->i_prealloc_cap_flush
);
1837 cf
->caps
= flushing
;
1840 spin_lock(&mdsc
->cap_dirty_lock
);
1841 list_del_init(&ci
->i_dirty_item
);
1843 cf
->tid
= ++mdsc
->last_cap_flush_tid
;
1844 list_add_tail(&cf
->g_list
, &mdsc
->cap_flush_list
);
1845 *oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1847 if (list_empty(&ci
->i_flushing_item
)) {
1848 list_add_tail(&ci
->i_flushing_item
, &session
->s_cap_flushing
);
1849 mdsc
->num_cap_flushing
++;
1851 spin_unlock(&mdsc
->cap_dirty_lock
);
1853 list_add_tail(&cf
->i_list
, &ci
->i_cap_flush_list
);
1859 * try to invalidate mapping pages without blocking.
1861 static int try_nonblocking_invalidate(struct inode
*inode
)
1863 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1864 u32 invalidating_gen
= ci
->i_rdcache_gen
;
1866 spin_unlock(&ci
->i_ceph_lock
);
1867 invalidate_mapping_pages(&inode
->i_data
, 0, -1);
1868 spin_lock(&ci
->i_ceph_lock
);
1870 if (inode
->i_data
.nrpages
== 0 &&
1871 invalidating_gen
== ci
->i_rdcache_gen
) {
1873 dout("try_nonblocking_invalidate %p success\n", inode
);
1874 /* save any racing async invalidate some trouble */
1875 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
- 1;
1878 dout("try_nonblocking_invalidate %p failed\n", inode
);
1882 bool __ceph_should_report_size(struct ceph_inode_info
*ci
)
1884 loff_t size
= ci
->vfs_inode
.i_size
;
1885 /* mds will adjust max size according to the reported size */
1886 if (ci
->i_flushing_caps
& CEPH_CAP_FILE_WR
)
1888 if (size
>= ci
->i_max_size
)
1890 /* half of previous max_size increment has been used */
1891 if (ci
->i_max_size
> ci
->i_reported_size
&&
1892 (size
<< 1) >= ci
->i_max_size
+ ci
->i_reported_size
)
1898 * Swiss army knife function to examine currently used and wanted
1899 * versus held caps. Release, flush, ack revoked caps to mds as
1902 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1903 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1906 void ceph_check_caps(struct ceph_inode_info
*ci
, int flags
,
1907 struct ceph_mds_session
*session
)
1909 struct ceph_fs_client
*fsc
= ceph_inode_to_client(&ci
->vfs_inode
);
1910 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
1911 struct inode
*inode
= &ci
->vfs_inode
;
1912 struct ceph_cap
*cap
;
1913 u64 flush_tid
, oldest_flush_tid
;
1914 int file_wanted
, used
, cap_used
;
1915 int took_snap_rwsem
= 0; /* true if mdsc->snap_rwsem held */
1916 int issued
, implemented
, want
, retain
, revoking
, flushing
= 0;
1917 int mds
= -1; /* keep track of how far we've gone through i_caps list
1918 to avoid an infinite loop on retry */
1920 bool queue_invalidate
= false;
1921 bool tried_invalidate
= false;
1923 spin_lock(&ci
->i_ceph_lock
);
1924 if (ci
->i_ceph_flags
& CEPH_I_FLUSH
)
1925 flags
|= CHECK_CAPS_FLUSH
;
1929 spin_lock(&ci
->i_ceph_lock
);
1931 file_wanted
= __ceph_caps_file_wanted(ci
);
1932 used
= __ceph_caps_used(ci
);
1933 issued
= __ceph_caps_issued(ci
, &implemented
);
1934 revoking
= implemented
& ~issued
;
1937 retain
= file_wanted
| used
| CEPH_CAP_PIN
;
1938 if (!mdsc
->stopping
&& inode
->i_nlink
> 0) {
1940 retain
|= CEPH_CAP_ANY
; /* be greedy */
1941 } else if (S_ISDIR(inode
->i_mode
) &&
1942 (issued
& CEPH_CAP_FILE_SHARED
) &&
1943 __ceph_dir_is_complete(ci
)) {
1945 * If a directory is complete, we want to keep
1946 * the exclusive cap. So that MDS does not end up
1947 * revoking the shared cap on every create/unlink
1950 if (IS_RDONLY(inode
)) {
1951 want
= CEPH_CAP_ANY_SHARED
;
1953 want
|= CEPH_CAP_ANY_SHARED
| CEPH_CAP_FILE_EXCL
;
1958 retain
|= CEPH_CAP_ANY_SHARED
;
1960 * keep RD only if we didn't have the file open RW,
1961 * because then the mds would revoke it anyway to
1962 * journal max_size=0.
1964 if (ci
->i_max_size
== 0)
1965 retain
|= CEPH_CAP_ANY_RD
;
1969 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1970 " issued %s revoking %s retain %s %s%s\n", inode
,
1971 ceph_cap_string(file_wanted
),
1972 ceph_cap_string(used
), ceph_cap_string(ci
->i_dirty_caps
),
1973 ceph_cap_string(ci
->i_flushing_caps
),
1974 ceph_cap_string(issued
), ceph_cap_string(revoking
),
1975 ceph_cap_string(retain
),
1976 (flags
& CHECK_CAPS_AUTHONLY
) ? " AUTHONLY" : "",
1977 (flags
& CHECK_CAPS_FLUSH
) ? " FLUSH" : "");
1980 * If we no longer need to hold onto old our caps, and we may
1981 * have cached pages, but don't want them, then try to invalidate.
1982 * If we fail, it's because pages are locked.... try again later.
1984 if ((!(flags
& CHECK_CAPS_NOINVAL
) || mdsc
->stopping
) &&
1985 S_ISREG(inode
->i_mode
) &&
1986 !(ci
->i_wb_ref
|| ci
->i_wrbuffer_ref
) && /* no dirty pages... */
1987 inode
->i_data
.nrpages
&& /* have cached pages */
1988 (revoking
& (CEPH_CAP_FILE_CACHE
|
1989 CEPH_CAP_FILE_LAZYIO
)) && /* or revoking cache */
1990 !tried_invalidate
) {
1991 dout("check_caps trying to invalidate on %p\n", inode
);
1992 if (try_nonblocking_invalidate(inode
) < 0) {
1993 dout("check_caps queuing invalidate\n");
1994 queue_invalidate
= true;
1995 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
;
1997 tried_invalidate
= true;
2001 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
2003 struct cap_msg_args arg
;
2005 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
2007 /* avoid looping forever */
2008 if (mds
>= cap
->mds
||
2009 ((flags
& CHECK_CAPS_AUTHONLY
) && cap
!= ci
->i_auth_cap
))
2012 /* NOTE: no side-effects allowed, until we take s_mutex */
2015 if (ci
->i_auth_cap
&& cap
!= ci
->i_auth_cap
)
2016 cap_used
&= ~ci
->i_auth_cap
->issued
;
2018 revoking
= cap
->implemented
& ~cap
->issued
;
2019 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
2020 cap
->mds
, cap
, ceph_cap_string(cap_used
),
2021 ceph_cap_string(cap
->issued
),
2022 ceph_cap_string(cap
->implemented
),
2023 ceph_cap_string(revoking
));
2025 if (cap
== ci
->i_auth_cap
&&
2026 (cap
->issued
& CEPH_CAP_FILE_WR
)) {
2027 /* request larger max_size from MDS? */
2028 if (ci
->i_wanted_max_size
> ci
->i_max_size
&&
2029 ci
->i_wanted_max_size
> ci
->i_requested_max_size
) {
2030 dout("requesting new max_size\n");
2034 /* approaching file_max? */
2035 if (__ceph_should_report_size(ci
)) {
2036 dout("i_size approaching max_size\n");
2040 /* flush anything dirty? */
2041 if (cap
== ci
->i_auth_cap
) {
2042 if ((flags
& CHECK_CAPS_FLUSH
) && ci
->i_dirty_caps
) {
2043 dout("flushing dirty caps\n");
2046 if (ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
) {
2047 dout("flushing snap caps\n");
2052 /* completed revocation? going down and there are no caps? */
2053 if (revoking
&& (revoking
& cap_used
) == 0) {
2054 dout("completed revocation of %s\n",
2055 ceph_cap_string(cap
->implemented
& ~cap
->issued
));
2059 /* want more caps from mds? */
2060 if (want
& ~cap
->mds_wanted
) {
2061 if (want
& ~(cap
->mds_wanted
| cap
->issued
))
2063 if (!__cap_is_valid(cap
))
2067 /* things we might delay */
2068 if ((cap
->issued
& ~retain
) == 0)
2069 continue; /* nope, all good */
2072 if (session
&& session
!= cap
->session
) {
2073 dout("oops, wrong session %p mutex\n", session
);
2074 mutex_unlock(&session
->s_mutex
);
2078 session
= cap
->session
;
2079 if (mutex_trylock(&session
->s_mutex
) == 0) {
2080 dout("inverting session/ino locks on %p\n",
2082 session
= ceph_get_mds_session(session
);
2083 spin_unlock(&ci
->i_ceph_lock
);
2084 if (took_snap_rwsem
) {
2085 up_read(&mdsc
->snap_rwsem
);
2086 took_snap_rwsem
= 0;
2089 mutex_lock(&session
->s_mutex
);
2090 ceph_put_mds_session(session
);
2093 * Because we take the reference while
2094 * holding the i_ceph_lock, it should
2095 * never be NULL. Throw a warning if it
2104 /* kick flushing and flush snaps before sending normal
2106 if (cap
== ci
->i_auth_cap
&&
2108 (CEPH_I_KICK_FLUSH
| CEPH_I_FLUSH_SNAPS
))) {
2109 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
)
2110 __kick_flushing_caps(mdsc
, session
, ci
, 0);
2111 if (ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
)
2112 __ceph_flush_snaps(ci
, session
);
2117 /* take snap_rwsem after session mutex */
2118 if (!took_snap_rwsem
) {
2119 if (down_read_trylock(&mdsc
->snap_rwsem
) == 0) {
2120 dout("inverting snap/in locks on %p\n",
2122 spin_unlock(&ci
->i_ceph_lock
);
2123 down_read(&mdsc
->snap_rwsem
);
2124 took_snap_rwsem
= 1;
2127 took_snap_rwsem
= 1;
2130 if (cap
== ci
->i_auth_cap
&& ci
->i_dirty_caps
) {
2131 flushing
= ci
->i_dirty_caps
;
2132 flush_tid
= __mark_caps_flushing(inode
, session
, false,
2134 if (flags
& CHECK_CAPS_FLUSH
&&
2135 list_empty(&session
->s_cap_dirty
))
2136 mflags
|= CEPH_CLIENT_CAPS_SYNC
;
2140 spin_lock(&mdsc
->cap_dirty_lock
);
2141 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2142 spin_unlock(&mdsc
->cap_dirty_lock
);
2145 mds
= cap
->mds
; /* remember mds, so we don't repeat */
2147 __prep_cap(&arg
, cap
, CEPH_CAP_OP_UPDATE
, mflags
, cap_used
,
2148 want
, retain
, flushing
, flush_tid
, oldest_flush_tid
);
2149 spin_unlock(&ci
->i_ceph_lock
);
2151 __send_cap(mdsc
, &arg
, ci
);
2153 goto retry
; /* retake i_ceph_lock and restart our cap scan. */
2156 /* periodically re-calculate caps wanted by open files */
2157 if (__ceph_is_any_real_caps(ci
) &&
2158 list_empty(&ci
->i_cap_delay_list
) &&
2159 (file_wanted
& ~CEPH_CAP_PIN
) &&
2160 !(used
& (CEPH_CAP_FILE_RD
| CEPH_CAP_ANY_FILE_WR
))) {
2161 __cap_delay_requeue(mdsc
, ci
);
2164 spin_unlock(&ci
->i_ceph_lock
);
2166 if (queue_invalidate
)
2167 ceph_queue_invalidate(inode
);
2170 mutex_unlock(&session
->s_mutex
);
2171 if (took_snap_rwsem
)
2172 up_read(&mdsc
->snap_rwsem
);
2176 * Try to flush dirty caps back to the auth mds.
2178 static int try_flush_caps(struct inode
*inode
, u64
*ptid
)
2180 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
2181 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2182 struct ceph_mds_session
*session
= NULL
;
2184 u64 flush_tid
= 0, oldest_flush_tid
= 0;
2187 spin_lock(&ci
->i_ceph_lock
);
2189 if (ci
->i_dirty_caps
&& ci
->i_auth_cap
) {
2190 struct ceph_cap
*cap
= ci
->i_auth_cap
;
2191 struct cap_msg_args arg
;
2193 if (session
!= cap
->session
) {
2194 spin_unlock(&ci
->i_ceph_lock
);
2196 mutex_unlock(&session
->s_mutex
);
2197 session
= cap
->session
;
2198 mutex_lock(&session
->s_mutex
);
2201 if (cap
->session
->s_state
< CEPH_MDS_SESSION_OPEN
) {
2202 spin_unlock(&ci
->i_ceph_lock
);
2206 if (ci
->i_ceph_flags
&
2207 (CEPH_I_KICK_FLUSH
| CEPH_I_FLUSH_SNAPS
)) {
2208 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
)
2209 __kick_flushing_caps(mdsc
, session
, ci
, 0);
2210 if (ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
)
2211 __ceph_flush_snaps(ci
, session
);
2215 flushing
= ci
->i_dirty_caps
;
2216 flush_tid
= __mark_caps_flushing(inode
, session
, true,
2219 __prep_cap(&arg
, cap
, CEPH_CAP_OP_FLUSH
, CEPH_CLIENT_CAPS_SYNC
,
2220 __ceph_caps_used(ci
), __ceph_caps_wanted(ci
),
2221 (cap
->issued
| cap
->implemented
),
2222 flushing
, flush_tid
, oldest_flush_tid
);
2223 spin_unlock(&ci
->i_ceph_lock
);
2225 __send_cap(mdsc
, &arg
, ci
);
2227 if (!list_empty(&ci
->i_cap_flush_list
)) {
2228 struct ceph_cap_flush
*cf
=
2229 list_last_entry(&ci
->i_cap_flush_list
,
2230 struct ceph_cap_flush
, i_list
);
2232 flush_tid
= cf
->tid
;
2234 flushing
= ci
->i_flushing_caps
;
2235 spin_unlock(&ci
->i_ceph_lock
);
2239 mutex_unlock(&session
->s_mutex
);
2246 * Return true if we've flushed caps through the given flush_tid.
2248 static int caps_are_flushed(struct inode
*inode
, u64 flush_tid
)
2250 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2253 spin_lock(&ci
->i_ceph_lock
);
2254 if (!list_empty(&ci
->i_cap_flush_list
)) {
2255 struct ceph_cap_flush
* cf
=
2256 list_first_entry(&ci
->i_cap_flush_list
,
2257 struct ceph_cap_flush
, i_list
);
2258 if (cf
->tid
<= flush_tid
)
2261 spin_unlock(&ci
->i_ceph_lock
);
2266 * wait for any unsafe requests to complete.
2268 static int unsafe_request_wait(struct inode
*inode
)
2270 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2271 struct ceph_mds_request
*req1
= NULL
, *req2
= NULL
;
2274 spin_lock(&ci
->i_unsafe_lock
);
2275 if (S_ISDIR(inode
->i_mode
) && !list_empty(&ci
->i_unsafe_dirops
)) {
2276 req1
= list_last_entry(&ci
->i_unsafe_dirops
,
2277 struct ceph_mds_request
,
2279 ceph_mdsc_get_request(req1
);
2281 if (!list_empty(&ci
->i_unsafe_iops
)) {
2282 req2
= list_last_entry(&ci
->i_unsafe_iops
,
2283 struct ceph_mds_request
,
2284 r_unsafe_target_item
);
2285 ceph_mdsc_get_request(req2
);
2287 spin_unlock(&ci
->i_unsafe_lock
);
2289 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2290 inode
, req1
? req1
->r_tid
: 0ULL, req2
? req2
->r_tid
: 0ULL);
2292 ret
= !wait_for_completion_timeout(&req1
->r_safe_completion
,
2293 ceph_timeout_jiffies(req1
->r_timeout
));
2296 ceph_mdsc_put_request(req1
);
2299 ret
= !wait_for_completion_timeout(&req2
->r_safe_completion
,
2300 ceph_timeout_jiffies(req2
->r_timeout
));
2303 ceph_mdsc_put_request(req2
);
2308 int ceph_fsync(struct file
*file
, loff_t start
, loff_t end
, int datasync
)
2310 struct ceph_file_info
*fi
= file
->private_data
;
2311 struct inode
*inode
= file
->f_mapping
->host
;
2312 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2317 dout("fsync %p%s\n", inode
, datasync
? " datasync" : "");
2319 ret
= file_write_and_wait_range(file
, start
, end
);
2323 ret
= ceph_wait_on_async_create(inode
);
2327 dirty
= try_flush_caps(inode
, &flush_tid
);
2328 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty
));
2330 err
= unsafe_request_wait(inode
);
2333 * only wait on non-file metadata writeback (the mds
2334 * can recover size and mtime, so we don't need to
2337 if (!err
&& (dirty
& ~CEPH_CAP_ANY_FILE_WR
)) {
2338 err
= wait_event_interruptible(ci
->i_cap_wq
,
2339 caps_are_flushed(inode
, flush_tid
));
2345 if (errseq_check(&ci
->i_meta_err
, READ_ONCE(fi
->meta_err
))) {
2346 spin_lock(&file
->f_lock
);
2347 err
= errseq_check_and_advance(&ci
->i_meta_err
,
2349 spin_unlock(&file
->f_lock
);
2354 dout("fsync %p%s result=%d\n", inode
, datasync
? " datasync" : "", ret
);
2359 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2360 * queue inode for flush but don't do so immediately, because we can
2361 * get by with fewer MDS messages if we wait for data writeback to
2364 int ceph_write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
2366 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2370 int wait
= (wbc
->sync_mode
== WB_SYNC_ALL
&& !wbc
->for_sync
);
2372 dout("write_inode %p wait=%d\n", inode
, wait
);
2374 dirty
= try_flush_caps(inode
, &flush_tid
);
2376 err
= wait_event_interruptible(ci
->i_cap_wq
,
2377 caps_are_flushed(inode
, flush_tid
));
2379 struct ceph_mds_client
*mdsc
=
2380 ceph_sb_to_client(inode
->i_sb
)->mdsc
;
2382 spin_lock(&ci
->i_ceph_lock
);
2383 if (__ceph_caps_dirty(ci
))
2384 __cap_delay_requeue_front(mdsc
, ci
);
2385 spin_unlock(&ci
->i_ceph_lock
);
2390 static void __kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2391 struct ceph_mds_session
*session
,
2392 struct ceph_inode_info
*ci
,
2393 u64 oldest_flush_tid
)
2394 __releases(ci
->i_ceph_lock
)
2395 __acquires(ci
->i_ceph_lock
)
2397 struct inode
*inode
= &ci
->vfs_inode
;
2398 struct ceph_cap
*cap
;
2399 struct ceph_cap_flush
*cf
;
2402 u64 last_snap_flush
= 0;
2404 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
2406 list_for_each_entry_reverse(cf
, &ci
->i_cap_flush_list
, i_list
) {
2408 last_snap_flush
= cf
->tid
;
2413 list_for_each_entry(cf
, &ci
->i_cap_flush_list
, i_list
) {
2414 if (cf
->tid
< first_tid
)
2417 cap
= ci
->i_auth_cap
;
2418 if (!(cap
&& cap
->session
== session
)) {
2419 pr_err("%p auth cap %p not mds%d ???\n",
2420 inode
, cap
, session
->s_mds
);
2424 first_tid
= cf
->tid
+ 1;
2427 struct cap_msg_args arg
;
2429 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2430 inode
, cap
, cf
->tid
, ceph_cap_string(cf
->caps
));
2431 __prep_cap(&arg
, cap
, CEPH_CAP_OP_FLUSH
,
2432 (cf
->tid
< last_snap_flush
?
2433 CEPH_CLIENT_CAPS_PENDING_CAPSNAP
: 0),
2434 __ceph_caps_used(ci
),
2435 __ceph_caps_wanted(ci
),
2436 (cap
->issued
| cap
->implemented
),
2437 cf
->caps
, cf
->tid
, oldest_flush_tid
);
2438 spin_unlock(&ci
->i_ceph_lock
);
2439 __send_cap(mdsc
, &arg
, ci
);
2441 struct ceph_cap_snap
*capsnap
=
2442 container_of(cf
, struct ceph_cap_snap
,
2444 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2445 inode
, capsnap
, cf
->tid
,
2446 ceph_cap_string(capsnap
->dirty
));
2448 refcount_inc(&capsnap
->nref
);
2449 spin_unlock(&ci
->i_ceph_lock
);
2451 ret
= __send_flush_snap(inode
, session
, capsnap
, cap
->mseq
,
2454 pr_err("kick_flushing_caps: error sending "
2455 "cap flushsnap, ino (%llx.%llx) "
2456 "tid %llu follows %llu\n",
2457 ceph_vinop(inode
), cf
->tid
,
2461 ceph_put_cap_snap(capsnap
);
2464 spin_lock(&ci
->i_ceph_lock
);
2468 void ceph_early_kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2469 struct ceph_mds_session
*session
)
2471 struct ceph_inode_info
*ci
;
2472 struct ceph_cap
*cap
;
2473 u64 oldest_flush_tid
;
2475 dout("early_kick_flushing_caps mds%d\n", session
->s_mds
);
2477 spin_lock(&mdsc
->cap_dirty_lock
);
2478 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2479 spin_unlock(&mdsc
->cap_dirty_lock
);
2481 list_for_each_entry(ci
, &session
->s_cap_flushing
, i_flushing_item
) {
2482 spin_lock(&ci
->i_ceph_lock
);
2483 cap
= ci
->i_auth_cap
;
2484 if (!(cap
&& cap
->session
== session
)) {
2485 pr_err("%p auth cap %p not mds%d ???\n",
2486 &ci
->vfs_inode
, cap
, session
->s_mds
);
2487 spin_unlock(&ci
->i_ceph_lock
);
2493 * if flushing caps were revoked, we re-send the cap flush
2494 * in client reconnect stage. This guarantees MDS * processes
2495 * the cap flush message before issuing the flushing caps to
2498 if ((cap
->issued
& ci
->i_flushing_caps
) !=
2499 ci
->i_flushing_caps
) {
2500 /* encode_caps_cb() also will reset these sequence
2501 * numbers. make sure sequence numbers in cap flush
2502 * message match later reconnect message */
2506 __kick_flushing_caps(mdsc
, session
, ci
,
2509 ci
->i_ceph_flags
|= CEPH_I_KICK_FLUSH
;
2512 spin_unlock(&ci
->i_ceph_lock
);
2516 void ceph_kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2517 struct ceph_mds_session
*session
)
2519 struct ceph_inode_info
*ci
;
2520 struct ceph_cap
*cap
;
2521 u64 oldest_flush_tid
;
2523 lockdep_assert_held(&session
->s_mutex
);
2525 dout("kick_flushing_caps mds%d\n", session
->s_mds
);
2527 spin_lock(&mdsc
->cap_dirty_lock
);
2528 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2529 spin_unlock(&mdsc
->cap_dirty_lock
);
2531 list_for_each_entry(ci
, &session
->s_cap_flushing
, i_flushing_item
) {
2532 spin_lock(&ci
->i_ceph_lock
);
2533 cap
= ci
->i_auth_cap
;
2534 if (!(cap
&& cap
->session
== session
)) {
2535 pr_err("%p auth cap %p not mds%d ???\n",
2536 &ci
->vfs_inode
, cap
, session
->s_mds
);
2537 spin_unlock(&ci
->i_ceph_lock
);
2540 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
) {
2541 __kick_flushing_caps(mdsc
, session
, ci
,
2544 spin_unlock(&ci
->i_ceph_lock
);
2548 void ceph_kick_flushing_inode_caps(struct ceph_mds_session
*session
,
2549 struct ceph_inode_info
*ci
)
2551 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2552 struct ceph_cap
*cap
= ci
->i_auth_cap
;
2554 lockdep_assert_held(&ci
->i_ceph_lock
);
2556 dout("%s %p flushing %s\n", __func__
, &ci
->vfs_inode
,
2557 ceph_cap_string(ci
->i_flushing_caps
));
2559 if (!list_empty(&ci
->i_cap_flush_list
)) {
2560 u64 oldest_flush_tid
;
2561 spin_lock(&mdsc
->cap_dirty_lock
);
2562 list_move_tail(&ci
->i_flushing_item
,
2563 &cap
->session
->s_cap_flushing
);
2564 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2565 spin_unlock(&mdsc
->cap_dirty_lock
);
2567 __kick_flushing_caps(mdsc
, session
, ci
, oldest_flush_tid
);
2573 * Take references to capabilities we hold, so that we don't release
2574 * them to the MDS prematurely.
2576 void ceph_take_cap_refs(struct ceph_inode_info
*ci
, int got
,
2577 bool snap_rwsem_locked
)
2579 lockdep_assert_held(&ci
->i_ceph_lock
);
2581 if (got
& CEPH_CAP_PIN
)
2583 if (got
& CEPH_CAP_FILE_RD
)
2585 if (got
& CEPH_CAP_FILE_CACHE
)
2586 ci
->i_rdcache_ref
++;
2587 if (got
& CEPH_CAP_FILE_EXCL
)
2589 if (got
& CEPH_CAP_FILE_WR
) {
2590 if (ci
->i_wr_ref
== 0 && !ci
->i_head_snapc
) {
2591 BUG_ON(!snap_rwsem_locked
);
2592 ci
->i_head_snapc
= ceph_get_snap_context(
2593 ci
->i_snap_realm
->cached_context
);
2597 if (got
& CEPH_CAP_FILE_BUFFER
) {
2598 if (ci
->i_wb_ref
== 0)
2599 ihold(&ci
->vfs_inode
);
2601 dout("%s %p wb %d -> %d (?)\n", __func__
,
2602 &ci
->vfs_inode
, ci
->i_wb_ref
-1, ci
->i_wb_ref
);
2607 * Try to grab cap references. Specify those refs we @want, and the
2608 * minimal set we @need. Also include the larger offset we are writing
2609 * to (when applicable), and check against max_size here as well.
2610 * Note that caller is responsible for ensuring max_size increases are
2611 * requested from the MDS.
2613 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2614 * or a negative error code. There are 3 speical error codes:
2615 * -EAGAIN: need to sleep but non-blocking is specified
2616 * -EFBIG: ask caller to call check_max_size() and try again.
2617 * -ESTALE: ask caller to call ceph_renew_caps() and try again.
2620 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2621 NON_BLOCKING
= (1 << 8),
2622 CHECK_FILELOCK
= (1 << 9),
2625 static int try_get_cap_refs(struct inode
*inode
, int need
, int want
,
2626 loff_t endoff
, int flags
, int *got
)
2628 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2629 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
2631 int have
, implemented
;
2632 bool snap_rwsem_locked
= false;
2634 dout("get_cap_refs %p need %s want %s\n", inode
,
2635 ceph_cap_string(need
), ceph_cap_string(want
));
2638 spin_lock(&ci
->i_ceph_lock
);
2640 if ((flags
& CHECK_FILELOCK
) &&
2641 (ci
->i_ceph_flags
& CEPH_I_ERROR_FILELOCK
)) {
2642 dout("try_get_cap_refs %p error filelock\n", inode
);
2647 /* finish pending truncate */
2648 while (ci
->i_truncate_pending
) {
2649 spin_unlock(&ci
->i_ceph_lock
);
2650 if (snap_rwsem_locked
) {
2651 up_read(&mdsc
->snap_rwsem
);
2652 snap_rwsem_locked
= false;
2654 __ceph_do_pending_vmtruncate(inode
);
2655 spin_lock(&ci
->i_ceph_lock
);
2658 have
= __ceph_caps_issued(ci
, &implemented
);
2660 if (have
& need
& CEPH_CAP_FILE_WR
) {
2661 if (endoff
>= 0 && endoff
> (loff_t
)ci
->i_max_size
) {
2662 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2663 inode
, endoff
, ci
->i_max_size
);
2664 if (endoff
> ci
->i_requested_max_size
)
2665 ret
= ci
->i_auth_cap
? -EFBIG
: -ESTALE
;
2669 * If a sync write is in progress, we must wait, so that we
2670 * can get a final snapshot value for size+mtime.
2672 if (__ceph_have_pending_cap_snap(ci
)) {
2673 dout("get_cap_refs %p cap_snap_pending\n", inode
);
2678 if ((have
& need
) == need
) {
2680 * Look at (implemented & ~have & not) so that we keep waiting
2681 * on transition from wanted -> needed caps. This is needed
2682 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2683 * going before a prior buffered writeback happens.
2685 int not = want
& ~(have
& need
);
2686 int revoking
= implemented
& ~have
;
2687 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2688 inode
, ceph_cap_string(have
), ceph_cap_string(not),
2689 ceph_cap_string(revoking
));
2690 if ((revoking
& not) == 0) {
2691 if (!snap_rwsem_locked
&&
2692 !ci
->i_head_snapc
&&
2693 (need
& CEPH_CAP_FILE_WR
)) {
2694 if (!down_read_trylock(&mdsc
->snap_rwsem
)) {
2696 * we can not call down_read() when
2697 * task isn't in TASK_RUNNING state
2699 if (flags
& NON_BLOCKING
) {
2704 spin_unlock(&ci
->i_ceph_lock
);
2705 down_read(&mdsc
->snap_rwsem
);
2706 snap_rwsem_locked
= true;
2709 snap_rwsem_locked
= true;
2711 if ((have
& want
) == want
)
2715 if (S_ISREG(inode
->i_mode
) &&
2716 (need
& CEPH_CAP_FILE_RD
) &&
2717 !(*got
& CEPH_CAP_FILE_CACHE
))
2718 ceph_disable_fscache_readpage(ci
);
2719 ceph_take_cap_refs(ci
, *got
, true);
2723 int session_readonly
= false;
2725 if (ci
->i_auth_cap
&&
2726 (need
& (CEPH_CAP_FILE_WR
| CEPH_CAP_FILE_EXCL
))) {
2727 struct ceph_mds_session
*s
= ci
->i_auth_cap
->session
;
2728 spin_lock(&s
->s_cap_lock
);
2729 session_readonly
= s
->s_readonly
;
2730 spin_unlock(&s
->s_cap_lock
);
2732 if (session_readonly
) {
2733 dout("get_cap_refs %p need %s but mds%d readonly\n",
2734 inode
, ceph_cap_string(need
), ci
->i_auth_cap
->mds
);
2739 if (READ_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
) {
2740 dout("get_cap_refs %p forced umount\n", inode
);
2744 mds_wanted
= __ceph_caps_mds_wanted(ci
, false);
2745 if (need
& ~mds_wanted
) {
2746 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2747 inode
, ceph_cap_string(need
),
2748 ceph_cap_string(mds_wanted
));
2753 dout("get_cap_refs %p have %s need %s\n", inode
,
2754 ceph_cap_string(have
), ceph_cap_string(need
));
2758 __ceph_touch_fmode(ci
, mdsc
, flags
);
2760 spin_unlock(&ci
->i_ceph_lock
);
2761 if (snap_rwsem_locked
)
2762 up_read(&mdsc
->snap_rwsem
);
2765 ceph_update_cap_mis(&mdsc
->metric
);
2767 ceph_update_cap_hit(&mdsc
->metric
);
2769 dout("get_cap_refs %p ret %d got %s\n", inode
,
2770 ret
, ceph_cap_string(*got
));
2775 * Check the offset we are writing up to against our current
2776 * max_size. If necessary, tell the MDS we want to write to
2779 static void check_max_size(struct inode
*inode
, loff_t endoff
)
2781 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2784 /* do we need to explicitly request a larger max_size? */
2785 spin_lock(&ci
->i_ceph_lock
);
2786 if (endoff
>= ci
->i_max_size
&& endoff
> ci
->i_wanted_max_size
) {
2787 dout("write %p at large endoff %llu, req max_size\n",
2789 ci
->i_wanted_max_size
= endoff
;
2791 /* duplicate ceph_check_caps()'s logic */
2792 if (ci
->i_auth_cap
&&
2793 (ci
->i_auth_cap
->issued
& CEPH_CAP_FILE_WR
) &&
2794 ci
->i_wanted_max_size
> ci
->i_max_size
&&
2795 ci
->i_wanted_max_size
> ci
->i_requested_max_size
)
2797 spin_unlock(&ci
->i_ceph_lock
);
2799 ceph_check_caps(ci
, CHECK_CAPS_AUTHONLY
, NULL
);
2802 static inline int get_used_fmode(int caps
)
2805 if (caps
& CEPH_CAP_FILE_RD
)
2806 fmode
|= CEPH_FILE_MODE_RD
;
2807 if (caps
& CEPH_CAP_FILE_WR
)
2808 fmode
|= CEPH_FILE_MODE_WR
;
2812 int ceph_try_get_caps(struct inode
*inode
, int need
, int want
,
2813 bool nonblock
, int *got
)
2817 BUG_ON(need
& ~CEPH_CAP_FILE_RD
);
2818 BUG_ON(want
& ~(CEPH_CAP_FILE_CACHE
| CEPH_CAP_FILE_LAZYIO
|
2819 CEPH_CAP_FILE_SHARED
| CEPH_CAP_FILE_EXCL
|
2820 CEPH_CAP_ANY_DIR_OPS
));
2822 ret
= ceph_pool_perm_check(inode
, need
);
2827 flags
= get_used_fmode(need
| want
);
2829 flags
|= NON_BLOCKING
;
2831 ret
= try_get_cap_refs(inode
, need
, want
, 0, flags
, got
);
2832 /* three special error codes */
2833 if (ret
== -EAGAIN
|| ret
== -EFBIG
|| ret
== -ESTALE
)
2839 * Wait for caps, and take cap references. If we can't get a WR cap
2840 * due to a small max_size, make sure we check_max_size (and possibly
2841 * ask the mds) so we don't get hung up indefinitely.
2843 int ceph_get_caps(struct file
*filp
, int need
, int want
,
2844 loff_t endoff
, int *got
, struct page
**pinned_page
)
2846 struct ceph_file_info
*fi
= filp
->private_data
;
2847 struct inode
*inode
= file_inode(filp
);
2848 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2849 struct ceph_fs_client
*fsc
= ceph_inode_to_client(inode
);
2850 int ret
, _got
, flags
;
2852 ret
= ceph_pool_perm_check(inode
, need
);
2856 if ((fi
->fmode
& CEPH_FILE_MODE_WR
) &&
2857 fi
->filp_gen
!= READ_ONCE(fsc
->filp_gen
))
2860 flags
= get_used_fmode(need
| want
);
2863 flags
&= CEPH_FILE_MODE_MASK
;
2864 if (atomic_read(&fi
->num_locks
))
2865 flags
|= CHECK_FILELOCK
;
2867 ret
= try_get_cap_refs(inode
, need
, want
, endoff
,
2869 WARN_ON_ONCE(ret
== -EAGAIN
);
2871 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
2873 DEFINE_WAIT_FUNC(wait
, woken_wake_function
);
2875 cw
.ino
= inode
->i_ino
;
2876 cw
.tgid
= current
->tgid
;
2880 spin_lock(&mdsc
->caps_list_lock
);
2881 list_add(&cw
.list
, &mdsc
->cap_wait_list
);
2882 spin_unlock(&mdsc
->caps_list_lock
);
2884 /* make sure used fmode not timeout */
2885 ceph_get_fmode(ci
, flags
, FMODE_WAIT_BIAS
);
2886 add_wait_queue(&ci
->i_cap_wq
, &wait
);
2888 flags
|= NON_BLOCKING
;
2889 while (!(ret
= try_get_cap_refs(inode
, need
, want
,
2890 endoff
, flags
, &_got
))) {
2891 if (signal_pending(current
)) {
2895 wait_woken(&wait
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
2898 remove_wait_queue(&ci
->i_cap_wq
, &wait
);
2899 ceph_put_fmode(ci
, flags
, FMODE_WAIT_BIAS
);
2901 spin_lock(&mdsc
->caps_list_lock
);
2903 spin_unlock(&mdsc
->caps_list_lock
);
2909 if ((fi
->fmode
& CEPH_FILE_MODE_WR
) &&
2910 fi
->filp_gen
!= READ_ONCE(fsc
->filp_gen
)) {
2911 if (ret
>= 0 && _got
)
2912 ceph_put_cap_refs(ci
, _got
);
2917 if (ret
== -EFBIG
|| ret
== -ESTALE
) {
2918 int ret2
= ceph_wait_on_async_create(inode
);
2922 if (ret
== -EFBIG
) {
2923 check_max_size(inode
, endoff
);
2926 if (ret
== -ESTALE
) {
2927 /* session was killed, try renew caps */
2928 ret
= ceph_renew_caps(inode
, flags
);
2935 if (S_ISREG(ci
->vfs_inode
.i_mode
) &&
2936 ci
->i_inline_version
!= CEPH_INLINE_NONE
&&
2937 (_got
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)) &&
2938 i_size_read(inode
) > 0) {
2940 find_get_page(inode
->i_mapping
, 0);
2942 if (PageUptodate(page
)) {
2943 *pinned_page
= page
;
2949 * drop cap refs first because getattr while
2950 * holding * caps refs can cause deadlock.
2952 ceph_put_cap_refs(ci
, _got
);
2956 * getattr request will bring inline data into
2959 ret
= __ceph_do_getattr(inode
, NULL
,
2960 CEPH_STAT_CAP_INLINE_DATA
,
2969 if (S_ISREG(ci
->vfs_inode
.i_mode
) &&
2970 (_got
& CEPH_CAP_FILE_RD
) && (_got
& CEPH_CAP_FILE_CACHE
))
2971 ceph_fscache_revalidate_cookie(ci
);
2978 * Take cap refs. Caller must already know we hold at least one ref
2979 * on the caps in question or we don't know this is safe.
2981 void ceph_get_cap_refs(struct ceph_inode_info
*ci
, int caps
)
2983 spin_lock(&ci
->i_ceph_lock
);
2984 ceph_take_cap_refs(ci
, caps
, false);
2985 spin_unlock(&ci
->i_ceph_lock
);
2990 * drop cap_snap that is not associated with any snapshot.
2991 * we don't need to send FLUSHSNAP message for it.
2993 static int ceph_try_drop_cap_snap(struct ceph_inode_info
*ci
,
2994 struct ceph_cap_snap
*capsnap
)
2996 if (!capsnap
->need_flush
&&
2997 !capsnap
->writing
&& !capsnap
->dirty_pages
) {
2998 dout("dropping cap_snap %p follows %llu\n",
2999 capsnap
, capsnap
->follows
);
3000 BUG_ON(capsnap
->cap_flush
.tid
> 0);
3001 ceph_put_snap_context(capsnap
->context
);
3002 if (!list_is_last(&capsnap
->ci_item
, &ci
->i_cap_snaps
))
3003 ci
->i_ceph_flags
|= CEPH_I_FLUSH_SNAPS
;
3005 list_del(&capsnap
->ci_item
);
3006 ceph_put_cap_snap(capsnap
);
3015 * If we released the last ref on any given cap, call ceph_check_caps
3016 * to release (or schedule a release).
3018 * If we are releasing a WR cap (from a sync write), finalize any affected
3019 * cap_snap, and wake up any waiters.
3021 static void __ceph_put_cap_refs(struct ceph_inode_info
*ci
, int had
,
3022 bool skip_checking_caps
)
3024 struct inode
*inode
= &ci
->vfs_inode
;
3025 int last
= 0, put
= 0, flushsnaps
= 0, wake
= 0;
3027 spin_lock(&ci
->i_ceph_lock
);
3028 if (had
& CEPH_CAP_PIN
)
3030 if (had
& CEPH_CAP_FILE_RD
)
3031 if (--ci
->i_rd_ref
== 0)
3033 if (had
& CEPH_CAP_FILE_CACHE
)
3034 if (--ci
->i_rdcache_ref
== 0)
3036 if (had
& CEPH_CAP_FILE_EXCL
)
3037 if (--ci
->i_fx_ref
== 0)
3039 if (had
& CEPH_CAP_FILE_BUFFER
) {
3040 if (--ci
->i_wb_ref
== 0) {
3044 dout("put_cap_refs %p wb %d -> %d (?)\n",
3045 inode
, ci
->i_wb_ref
+1, ci
->i_wb_ref
);
3047 if (had
& CEPH_CAP_FILE_WR
)
3048 if (--ci
->i_wr_ref
== 0) {
3050 if (__ceph_have_pending_cap_snap(ci
)) {
3051 struct ceph_cap_snap
*capsnap
=
3052 list_last_entry(&ci
->i_cap_snaps
,
3053 struct ceph_cap_snap
,
3055 capsnap
->writing
= 0;
3056 if (ceph_try_drop_cap_snap(ci
, capsnap
))
3058 else if (__ceph_finish_cap_snap(ci
, capsnap
))
3062 if (ci
->i_wrbuffer_ref_head
== 0 &&
3063 ci
->i_dirty_caps
== 0 &&
3064 ci
->i_flushing_caps
== 0) {
3065 BUG_ON(!ci
->i_head_snapc
);
3066 ceph_put_snap_context(ci
->i_head_snapc
);
3067 ci
->i_head_snapc
= NULL
;
3069 /* see comment in __ceph_remove_cap() */
3070 if (!__ceph_is_any_real_caps(ci
) && ci
->i_snap_realm
)
3071 drop_inode_snap_realm(ci
);
3073 spin_unlock(&ci
->i_ceph_lock
);
3075 dout("put_cap_refs %p had %s%s%s\n", inode
, ceph_cap_string(had
),
3076 last
? " last" : "", put
? " put" : "");
3078 if (last
&& !skip_checking_caps
)
3079 ceph_check_caps(ci
, 0, NULL
);
3080 else if (flushsnaps
)
3081 ceph_flush_snaps(ci
, NULL
);
3083 wake_up_all(&ci
->i_cap_wq
);
3088 void ceph_put_cap_refs(struct ceph_inode_info
*ci
, int had
)
3090 __ceph_put_cap_refs(ci
, had
, false);
3093 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info
*ci
, int had
)
3095 __ceph_put_cap_refs(ci
, had
, true);
3099 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3100 * context. Adjust per-snap dirty page accounting as appropriate.
3101 * Once all dirty data for a cap_snap is flushed, flush snapped file
3102 * metadata back to the MDS. If we dropped the last ref, call
3105 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info
*ci
, int nr
,
3106 struct ceph_snap_context
*snapc
)
3108 struct inode
*inode
= &ci
->vfs_inode
;
3109 struct ceph_cap_snap
*capsnap
= NULL
;
3113 bool flush_snaps
= false;
3114 bool complete_capsnap
= false;
3116 spin_lock(&ci
->i_ceph_lock
);
3117 ci
->i_wrbuffer_ref
-= nr
;
3118 if (ci
->i_wrbuffer_ref
== 0) {
3123 if (ci
->i_head_snapc
== snapc
) {
3124 ci
->i_wrbuffer_ref_head
-= nr
;
3125 if (ci
->i_wrbuffer_ref_head
== 0 &&
3126 ci
->i_wr_ref
== 0 &&
3127 ci
->i_dirty_caps
== 0 &&
3128 ci
->i_flushing_caps
== 0) {
3129 BUG_ON(!ci
->i_head_snapc
);
3130 ceph_put_snap_context(ci
->i_head_snapc
);
3131 ci
->i_head_snapc
= NULL
;
3133 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3135 ci
->i_wrbuffer_ref
+nr
, ci
->i_wrbuffer_ref_head
+nr
,
3136 ci
->i_wrbuffer_ref
, ci
->i_wrbuffer_ref_head
,
3137 last
? " LAST" : "");
3139 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
3140 if (capsnap
->context
== snapc
) {
3146 capsnap
->dirty_pages
-= nr
;
3147 if (capsnap
->dirty_pages
== 0) {
3148 complete_capsnap
= true;
3149 if (!capsnap
->writing
) {
3150 if (ceph_try_drop_cap_snap(ci
, capsnap
)) {
3153 ci
->i_ceph_flags
|= CEPH_I_FLUSH_SNAPS
;
3158 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3159 " snap %lld %d/%d -> %d/%d %s%s\n",
3160 inode
, capsnap
, capsnap
->context
->seq
,
3161 ci
->i_wrbuffer_ref
+nr
, capsnap
->dirty_pages
+ nr
,
3162 ci
->i_wrbuffer_ref
, capsnap
->dirty_pages
,
3163 last
? " (wrbuffer last)" : "",
3164 complete_capsnap
? " (complete capsnap)" : "");
3167 spin_unlock(&ci
->i_ceph_lock
);
3170 ceph_check_caps(ci
, 0, NULL
);
3171 } else if (flush_snaps
) {
3172 ceph_flush_snaps(ci
, NULL
);
3174 if (complete_capsnap
)
3175 wake_up_all(&ci
->i_cap_wq
);
3177 /* avoid calling iput_final() in osd dispatch threads */
3178 ceph_async_iput(inode
);
3183 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3185 static void invalidate_aliases(struct inode
*inode
)
3187 struct dentry
*dn
, *prev
= NULL
;
3189 dout("invalidate_aliases inode %p\n", inode
);
3190 d_prune_aliases(inode
);
3192 * For non-directory inode, d_find_alias() only returns
3193 * hashed dentry. After calling d_invalidate(), the
3194 * dentry becomes unhashed.
3196 * For directory inode, d_find_alias() can return
3197 * unhashed dentry. But directory inode should have
3198 * one alias at most.
3200 while ((dn
= d_find_alias(inode
))) {
3214 struct cap_extra_info
{
3215 struct ceph_string
*pool_ns
;
3225 /* currently issued */
3227 struct timespec64 btime
;
3231 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3232 * actually be a revocation if it specifies a smaller cap set.)
3234 * caller holds s_mutex and i_ceph_lock, we drop both.
3236 static void handle_cap_grant(struct inode
*inode
,
3237 struct ceph_mds_session
*session
,
3238 struct ceph_cap
*cap
,
3239 struct ceph_mds_caps
*grant
,
3240 struct ceph_buffer
*xattr_buf
,
3241 struct cap_extra_info
*extra_info
)
3242 __releases(ci
->i_ceph_lock
)
3243 __releases(session
->s_mdsc
->snap_rwsem
)
3245 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3246 int seq
= le32_to_cpu(grant
->seq
);
3247 int newcaps
= le32_to_cpu(grant
->caps
);
3248 int used
, wanted
, dirty
;
3249 u64 size
= le64_to_cpu(grant
->size
);
3250 u64 max_size
= le64_to_cpu(grant
->max_size
);
3251 unsigned char check_caps
= 0;
3252 bool was_stale
= cap
->cap_gen
< session
->s_cap_gen
;
3254 bool writeback
= false;
3255 bool queue_trunc
= false;
3256 bool queue_invalidate
= false;
3257 bool deleted_inode
= false;
3258 bool fill_inline
= false;
3260 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3261 inode
, cap
, session
->s_mds
, seq
, ceph_cap_string(newcaps
));
3262 dout(" size %llu max_size %llu, i_size %llu\n", size
, max_size
,
3267 * If CACHE is being revoked, and we have no dirty buffers,
3268 * try to invalidate (once). (If there are dirty buffers, we
3269 * will invalidate _after_ writeback.)
3271 if (S_ISREG(inode
->i_mode
) && /* don't invalidate readdir cache */
3272 ((cap
->issued
& ~newcaps
) & CEPH_CAP_FILE_CACHE
) &&
3273 (newcaps
& CEPH_CAP_FILE_LAZYIO
) == 0 &&
3274 !(ci
->i_wrbuffer_ref
|| ci
->i_wb_ref
)) {
3275 if (try_nonblocking_invalidate(inode
)) {
3276 /* there were locked pages.. invalidate later
3277 in a separate thread. */
3278 if (ci
->i_rdcache_revoking
!= ci
->i_rdcache_gen
) {
3279 queue_invalidate
= true;
3280 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
;
3286 cap
->issued
= cap
->implemented
= CEPH_CAP_PIN
;
3289 * auth mds of the inode changed. we received the cap export message,
3290 * but still haven't received the cap import message. handle_cap_export
3291 * updated the new auth MDS' cap.
3293 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3294 * that was sent before the cap import message. So don't remove caps.
3296 if (ceph_seq_cmp(seq
, cap
->seq
) <= 0) {
3297 WARN_ON(cap
!= ci
->i_auth_cap
);
3298 WARN_ON(cap
->cap_id
!= le64_to_cpu(grant
->cap_id
));
3300 newcaps
|= cap
->issued
;
3303 /* side effects now are allowed */
3304 cap
->cap_gen
= session
->s_cap_gen
;
3307 __check_cap_issue(ci
, cap
, newcaps
);
3309 inode_set_max_iversion_raw(inode
, extra_info
->change_attr
);
3311 if ((newcaps
& CEPH_CAP_AUTH_SHARED
) &&
3312 (extra_info
->issued
& CEPH_CAP_AUTH_EXCL
) == 0) {
3313 inode
->i_mode
= le32_to_cpu(grant
->mode
);
3314 inode
->i_uid
= make_kuid(&init_user_ns
, le32_to_cpu(grant
->uid
));
3315 inode
->i_gid
= make_kgid(&init_user_ns
, le32_to_cpu(grant
->gid
));
3316 ci
->i_btime
= extra_info
->btime
;
3317 dout("%p mode 0%o uid.gid %d.%d\n", inode
, inode
->i_mode
,
3318 from_kuid(&init_user_ns
, inode
->i_uid
),
3319 from_kgid(&init_user_ns
, inode
->i_gid
));
3322 if ((newcaps
& CEPH_CAP_LINK_SHARED
) &&
3323 (extra_info
->issued
& CEPH_CAP_LINK_EXCL
) == 0) {
3324 set_nlink(inode
, le32_to_cpu(grant
->nlink
));
3325 if (inode
->i_nlink
== 0 &&
3326 (newcaps
& (CEPH_CAP_LINK_SHARED
| CEPH_CAP_LINK_EXCL
)))
3327 deleted_inode
= true;
3330 if ((extra_info
->issued
& CEPH_CAP_XATTR_EXCL
) == 0 &&
3332 int len
= le32_to_cpu(grant
->xattr_len
);
3333 u64 version
= le64_to_cpu(grant
->xattr_version
);
3335 if (version
> ci
->i_xattrs
.version
) {
3336 dout(" got new xattrs v%llu on %p len %d\n",
3337 version
, inode
, len
);
3338 if (ci
->i_xattrs
.blob
)
3339 ceph_buffer_put(ci
->i_xattrs
.blob
);
3340 ci
->i_xattrs
.blob
= ceph_buffer_get(xattr_buf
);
3341 ci
->i_xattrs
.version
= version
;
3342 ceph_forget_all_cached_acls(inode
);
3343 ceph_security_invalidate_secctx(inode
);
3347 if (newcaps
& CEPH_CAP_ANY_RD
) {
3348 struct timespec64 mtime
, atime
, ctime
;
3349 /* ctime/mtime/atime? */
3350 ceph_decode_timespec64(&mtime
, &grant
->mtime
);
3351 ceph_decode_timespec64(&atime
, &grant
->atime
);
3352 ceph_decode_timespec64(&ctime
, &grant
->ctime
);
3353 ceph_fill_file_time(inode
, extra_info
->issued
,
3354 le32_to_cpu(grant
->time_warp_seq
),
3355 &ctime
, &mtime
, &atime
);
3358 if ((newcaps
& CEPH_CAP_FILE_SHARED
) && extra_info
->dirstat_valid
) {
3359 ci
->i_files
= extra_info
->nfiles
;
3360 ci
->i_subdirs
= extra_info
->nsubdirs
;
3363 if (newcaps
& (CEPH_CAP_ANY_FILE_RD
| CEPH_CAP_ANY_FILE_WR
)) {
3364 /* file layout may have changed */
3365 s64 old_pool
= ci
->i_layout
.pool_id
;
3366 struct ceph_string
*old_ns
;
3368 ceph_file_layout_from_legacy(&ci
->i_layout
, &grant
->layout
);
3369 old_ns
= rcu_dereference_protected(ci
->i_layout
.pool_ns
,
3370 lockdep_is_held(&ci
->i_ceph_lock
));
3371 rcu_assign_pointer(ci
->i_layout
.pool_ns
, extra_info
->pool_ns
);
3373 if (ci
->i_layout
.pool_id
!= old_pool
||
3374 extra_info
->pool_ns
!= old_ns
)
3375 ci
->i_ceph_flags
&= ~CEPH_I_POOL_PERM
;
3377 extra_info
->pool_ns
= old_ns
;
3379 /* size/truncate_seq? */
3380 queue_trunc
= ceph_fill_file_size(inode
, extra_info
->issued
,
3381 le32_to_cpu(grant
->truncate_seq
),
3382 le64_to_cpu(grant
->truncate_size
),
3386 if (ci
->i_auth_cap
== cap
&& (newcaps
& CEPH_CAP_ANY_FILE_WR
)) {
3387 if (max_size
!= ci
->i_max_size
) {
3388 dout("max_size %lld -> %llu\n",
3389 ci
->i_max_size
, max_size
);
3390 ci
->i_max_size
= max_size
;
3391 if (max_size
>= ci
->i_wanted_max_size
) {
3392 ci
->i_wanted_max_size
= 0; /* reset */
3393 ci
->i_requested_max_size
= 0;
3399 /* check cap bits */
3400 wanted
= __ceph_caps_wanted(ci
);
3401 used
= __ceph_caps_used(ci
);
3402 dirty
= __ceph_caps_dirty(ci
);
3403 dout(" my wanted = %s, used = %s, dirty %s\n",
3404 ceph_cap_string(wanted
),
3405 ceph_cap_string(used
),
3406 ceph_cap_string(dirty
));
3408 if ((was_stale
|| le32_to_cpu(grant
->op
) == CEPH_CAP_OP_IMPORT
) &&
3409 (wanted
& ~(cap
->mds_wanted
| newcaps
))) {
3411 * If mds is importing cap, prior cap messages that update
3412 * 'wanted' may get dropped by mds (migrate seq mismatch).
3414 * We don't send cap message to update 'wanted' if what we
3415 * want are already issued. If mds revokes caps, cap message
3416 * that releases caps also tells mds what we want. But if
3417 * caps got revoked by mds forcedly (session stale). We may
3418 * haven't told mds what we want.
3423 /* revocation, grant, or no-op? */
3424 if (cap
->issued
& ~newcaps
) {
3425 int revoking
= cap
->issued
& ~newcaps
;
3427 dout("revocation: %s -> %s (revoking %s)\n",
3428 ceph_cap_string(cap
->issued
),
3429 ceph_cap_string(newcaps
),
3430 ceph_cap_string(revoking
));
3431 if (S_ISREG(inode
->i_mode
) &&
3432 (revoking
& used
& CEPH_CAP_FILE_BUFFER
))
3433 writeback
= true; /* initiate writeback; will delay ack */
3434 else if (queue_invalidate
&&
3435 revoking
== CEPH_CAP_FILE_CACHE
&&
3436 (newcaps
& CEPH_CAP_FILE_LAZYIO
) == 0)
3437 ; /* do nothing yet, invalidation will be queued */
3438 else if (cap
== ci
->i_auth_cap
)
3439 check_caps
= 1; /* check auth cap only */
3441 check_caps
= 2; /* check all caps */
3442 cap
->issued
= newcaps
;
3443 cap
->implemented
|= newcaps
;
3444 } else if (cap
->issued
== newcaps
) {
3445 dout("caps unchanged: %s -> %s\n",
3446 ceph_cap_string(cap
->issued
), ceph_cap_string(newcaps
));
3448 dout("grant: %s -> %s\n", ceph_cap_string(cap
->issued
),
3449 ceph_cap_string(newcaps
));
3450 /* non-auth MDS is revoking the newly grant caps ? */
3451 if (cap
== ci
->i_auth_cap
&&
3452 __ceph_caps_revoking_other(ci
, cap
, newcaps
))
3455 cap
->issued
= newcaps
;
3456 cap
->implemented
|= newcaps
; /* add bits only, to
3457 * avoid stepping on a
3458 * pending revocation */
3461 BUG_ON(cap
->issued
& ~cap
->implemented
);
3463 if (extra_info
->inline_version
> 0 &&
3464 extra_info
->inline_version
>= ci
->i_inline_version
) {
3465 ci
->i_inline_version
= extra_info
->inline_version
;
3466 if (ci
->i_inline_version
!= CEPH_INLINE_NONE
&&
3467 (newcaps
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)))
3471 if (ci
->i_auth_cap
== cap
&&
3472 le32_to_cpu(grant
->op
) == CEPH_CAP_OP_IMPORT
) {
3473 if (newcaps
& ~extra_info
->issued
)
3476 if (ci
->i_requested_max_size
> max_size
||
3477 !(le32_to_cpu(grant
->wanted
) & CEPH_CAP_ANY_FILE_WR
)) {
3478 /* re-request max_size if necessary */
3479 ci
->i_requested_max_size
= 0;
3483 ceph_kick_flushing_inode_caps(session
, ci
);
3484 spin_unlock(&ci
->i_ceph_lock
);
3485 up_read(&session
->s_mdsc
->snap_rwsem
);
3487 spin_unlock(&ci
->i_ceph_lock
);
3491 ceph_fill_inline_data(inode
, NULL
, extra_info
->inline_data
,
3492 extra_info
->inline_len
);
3495 ceph_queue_vmtruncate(inode
);
3499 * queue inode for writeback: we can't actually call
3500 * filemap_write_and_wait, etc. from message handler
3503 ceph_queue_writeback(inode
);
3504 if (queue_invalidate
)
3505 ceph_queue_invalidate(inode
);
3507 invalidate_aliases(inode
);
3509 wake_up_all(&ci
->i_cap_wq
);
3511 if (check_caps
== 1)
3512 ceph_check_caps(ci
, CHECK_CAPS_AUTHONLY
| CHECK_CAPS_NOINVAL
,
3514 else if (check_caps
== 2)
3515 ceph_check_caps(ci
, CHECK_CAPS_NOINVAL
, session
);
3517 mutex_unlock(&session
->s_mutex
);
3521 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3522 * MDS has been safely committed.
3524 static void handle_cap_flush_ack(struct inode
*inode
, u64 flush_tid
,
3525 struct ceph_mds_caps
*m
,
3526 struct ceph_mds_session
*session
,
3527 struct ceph_cap
*cap
)
3528 __releases(ci
->i_ceph_lock
)
3530 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3531 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
3532 struct ceph_cap_flush
*cf
, *tmp_cf
;
3533 LIST_HEAD(to_remove
);
3534 unsigned seq
= le32_to_cpu(m
->seq
);
3535 int dirty
= le32_to_cpu(m
->dirty
);
3538 bool wake_ci
= false;
3539 bool wake_mdsc
= false;
3541 list_for_each_entry_safe(cf
, tmp_cf
, &ci
->i_cap_flush_list
, i_list
) {
3542 /* Is this the one that was flushed? */
3543 if (cf
->tid
== flush_tid
)
3546 /* Is this a capsnap? */
3550 if (cf
->tid
<= flush_tid
) {
3552 * An earlier or current tid. The FLUSH_ACK should
3553 * represent a superset of this flush's caps.
3555 wake_ci
|= __detach_cap_flush_from_ci(ci
, cf
);
3556 list_add_tail(&cf
->i_list
, &to_remove
);
3559 * This is a later one. Any caps in it are still dirty
3560 * so don't count them as cleaned.
3562 cleaned
&= ~cf
->caps
;
3568 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3569 " flushing %s -> %s\n",
3570 inode
, session
->s_mds
, seq
, ceph_cap_string(dirty
),
3571 ceph_cap_string(cleaned
), ceph_cap_string(ci
->i_flushing_caps
),
3572 ceph_cap_string(ci
->i_flushing_caps
& ~cleaned
));
3574 if (list_empty(&to_remove
) && !cleaned
)
3577 ci
->i_flushing_caps
&= ~cleaned
;
3579 spin_lock(&mdsc
->cap_dirty_lock
);
3581 list_for_each_entry(cf
, &to_remove
, i_list
)
3582 wake_mdsc
|= __detach_cap_flush_from_mdsc(mdsc
, cf
);
3584 if (ci
->i_flushing_caps
== 0) {
3585 if (list_empty(&ci
->i_cap_flush_list
)) {
3586 list_del_init(&ci
->i_flushing_item
);
3587 if (!list_empty(&session
->s_cap_flushing
)) {
3588 dout(" mds%d still flushing cap on %p\n",
3590 &list_first_entry(&session
->s_cap_flushing
,
3591 struct ceph_inode_info
,
3592 i_flushing_item
)->vfs_inode
);
3595 mdsc
->num_cap_flushing
--;
3596 dout(" inode %p now !flushing\n", inode
);
3598 if (ci
->i_dirty_caps
== 0) {
3599 dout(" inode %p now clean\n", inode
);
3600 BUG_ON(!list_empty(&ci
->i_dirty_item
));
3602 if (ci
->i_wr_ref
== 0 &&
3603 ci
->i_wrbuffer_ref_head
== 0) {
3604 BUG_ON(!ci
->i_head_snapc
);
3605 ceph_put_snap_context(ci
->i_head_snapc
);
3606 ci
->i_head_snapc
= NULL
;
3609 BUG_ON(list_empty(&ci
->i_dirty_item
));
3612 spin_unlock(&mdsc
->cap_dirty_lock
);
3615 spin_unlock(&ci
->i_ceph_lock
);
3617 while (!list_empty(&to_remove
)) {
3618 cf
= list_first_entry(&to_remove
,
3619 struct ceph_cap_flush
, i_list
);
3620 list_del(&cf
->i_list
);
3621 ceph_free_cap_flush(cf
);
3625 wake_up_all(&ci
->i_cap_wq
);
3627 wake_up_all(&mdsc
->cap_flushing_wq
);
3633 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3634 * throw away our cap_snap.
3636 * Caller hold s_mutex.
3638 static void handle_cap_flushsnap_ack(struct inode
*inode
, u64 flush_tid
,
3639 struct ceph_mds_caps
*m
,
3640 struct ceph_mds_session
*session
)
3642 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3643 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
3644 u64 follows
= le64_to_cpu(m
->snap_follows
);
3645 struct ceph_cap_snap
*capsnap
;
3646 bool flushed
= false;
3647 bool wake_ci
= false;
3648 bool wake_mdsc
= false;
3650 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3651 inode
, ci
, session
->s_mds
, follows
);
3653 spin_lock(&ci
->i_ceph_lock
);
3654 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
3655 if (capsnap
->follows
== follows
) {
3656 if (capsnap
->cap_flush
.tid
!= flush_tid
) {
3657 dout(" cap_snap %p follows %lld tid %lld !="
3658 " %lld\n", capsnap
, follows
,
3659 flush_tid
, capsnap
->cap_flush
.tid
);
3665 dout(" skipping cap_snap %p follows %lld\n",
3666 capsnap
, capsnap
->follows
);
3670 WARN_ON(capsnap
->dirty_pages
|| capsnap
->writing
);
3671 dout(" removing %p cap_snap %p follows %lld\n",
3672 inode
, capsnap
, follows
);
3673 list_del(&capsnap
->ci_item
);
3674 wake_ci
|= __detach_cap_flush_from_ci(ci
, &capsnap
->cap_flush
);
3676 spin_lock(&mdsc
->cap_dirty_lock
);
3678 if (list_empty(&ci
->i_cap_flush_list
))
3679 list_del_init(&ci
->i_flushing_item
);
3681 wake_mdsc
|= __detach_cap_flush_from_mdsc(mdsc
,
3682 &capsnap
->cap_flush
);
3683 spin_unlock(&mdsc
->cap_dirty_lock
);
3685 spin_unlock(&ci
->i_ceph_lock
);
3687 ceph_put_snap_context(capsnap
->context
);
3688 ceph_put_cap_snap(capsnap
);
3690 wake_up_all(&ci
->i_cap_wq
);
3692 wake_up_all(&mdsc
->cap_flushing_wq
);
3698 * Handle TRUNC from MDS, indicating file truncation.
3700 * caller hold s_mutex.
3702 static bool handle_cap_trunc(struct inode
*inode
,
3703 struct ceph_mds_caps
*trunc
,
3704 struct ceph_mds_session
*session
)
3706 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3707 int mds
= session
->s_mds
;
3708 int seq
= le32_to_cpu(trunc
->seq
);
3709 u32 truncate_seq
= le32_to_cpu(trunc
->truncate_seq
);
3710 u64 truncate_size
= le64_to_cpu(trunc
->truncate_size
);
3711 u64 size
= le64_to_cpu(trunc
->size
);
3712 int implemented
= 0;
3713 int dirty
= __ceph_caps_dirty(ci
);
3714 int issued
= __ceph_caps_issued(ceph_inode(inode
), &implemented
);
3715 bool queue_trunc
= false;
3717 lockdep_assert_held(&ci
->i_ceph_lock
);
3719 issued
|= implemented
| dirty
;
3721 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3722 inode
, mds
, seq
, truncate_size
, truncate_seq
);
3723 queue_trunc
= ceph_fill_file_size(inode
, issued
,
3724 truncate_seq
, truncate_size
, size
);
3729 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3730 * different one. If we are the most recent migration we've seen (as
3731 * indicated by mseq), make note of the migrating cap bits for the
3732 * duration (until we see the corresponding IMPORT).
3734 * caller holds s_mutex
3736 static void handle_cap_export(struct inode
*inode
, struct ceph_mds_caps
*ex
,
3737 struct ceph_mds_cap_peer
*ph
,
3738 struct ceph_mds_session
*session
)
3740 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
3741 struct ceph_mds_session
*tsession
= NULL
;
3742 struct ceph_cap
*cap
, *tcap
, *new_cap
= NULL
;
3743 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3745 unsigned mseq
= le32_to_cpu(ex
->migrate_seq
);
3746 unsigned t_seq
, t_mseq
;
3748 int mds
= session
->s_mds
;
3751 t_cap_id
= le64_to_cpu(ph
->cap_id
);
3752 t_seq
= le32_to_cpu(ph
->seq
);
3753 t_mseq
= le32_to_cpu(ph
->mseq
);
3754 target
= le32_to_cpu(ph
->mds
);
3756 t_cap_id
= t_seq
= t_mseq
= 0;
3760 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3761 inode
, ci
, mds
, mseq
, target
);
3763 spin_lock(&ci
->i_ceph_lock
);
3764 cap
= __get_cap_for_mds(ci
, mds
);
3765 if (!cap
|| cap
->cap_id
!= le64_to_cpu(ex
->cap_id
))
3769 __ceph_remove_cap(cap
, false);
3774 * now we know we haven't received the cap import message yet
3775 * because the exported cap still exist.
3778 issued
= cap
->issued
;
3779 if (issued
!= cap
->implemented
)
3780 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3781 "ino (%llx.%llx) mds%d seq %d mseq %d "
3782 "issued %s implemented %s\n",
3783 ceph_vinop(inode
), mds
, cap
->seq
, cap
->mseq
,
3784 ceph_cap_string(issued
),
3785 ceph_cap_string(cap
->implemented
));
3788 tcap
= __get_cap_for_mds(ci
, target
);
3790 /* already have caps from the target */
3791 if (tcap
->cap_id
== t_cap_id
&&
3792 ceph_seq_cmp(tcap
->seq
, t_seq
) < 0) {
3793 dout(" updating import cap %p mds%d\n", tcap
, target
);
3794 tcap
->cap_id
= t_cap_id
;
3795 tcap
->seq
= t_seq
- 1;
3796 tcap
->issue_seq
= t_seq
- 1;
3797 tcap
->issued
|= issued
;
3798 tcap
->implemented
|= issued
;
3799 if (cap
== ci
->i_auth_cap
) {
3800 ci
->i_auth_cap
= tcap
;
3801 change_auth_cap_ses(ci
, tcap
->session
);
3804 __ceph_remove_cap(cap
, false);
3806 } else if (tsession
) {
3807 /* add placeholder for the export tagert */
3808 int flag
= (cap
== ci
->i_auth_cap
) ? CEPH_CAP_FLAG_AUTH
: 0;
3810 ceph_add_cap(inode
, tsession
, t_cap_id
, issued
, 0,
3811 t_seq
- 1, t_mseq
, (u64
)-1, flag
, &new_cap
);
3813 if (!list_empty(&ci
->i_cap_flush_list
) &&
3814 ci
->i_auth_cap
== tcap
) {
3815 spin_lock(&mdsc
->cap_dirty_lock
);
3816 list_move_tail(&ci
->i_flushing_item
,
3817 &tcap
->session
->s_cap_flushing
);
3818 spin_unlock(&mdsc
->cap_dirty_lock
);
3821 __ceph_remove_cap(cap
, false);
3825 spin_unlock(&ci
->i_ceph_lock
);
3826 mutex_unlock(&session
->s_mutex
);
3828 /* open target session */
3829 tsession
= ceph_mdsc_open_export_target_session(mdsc
, target
);
3830 if (!IS_ERR(tsession
)) {
3832 mutex_lock(&session
->s_mutex
);
3833 mutex_lock_nested(&tsession
->s_mutex
,
3834 SINGLE_DEPTH_NESTING
);
3836 mutex_lock(&tsession
->s_mutex
);
3837 mutex_lock_nested(&session
->s_mutex
,
3838 SINGLE_DEPTH_NESTING
);
3840 new_cap
= ceph_get_cap(mdsc
, NULL
);
3845 mutex_lock(&session
->s_mutex
);
3850 spin_unlock(&ci
->i_ceph_lock
);
3851 mutex_unlock(&session
->s_mutex
);
3853 mutex_unlock(&tsession
->s_mutex
);
3854 ceph_put_mds_session(tsession
);
3857 ceph_put_cap(mdsc
, new_cap
);
3861 * Handle cap IMPORT.
3863 * caller holds s_mutex. acquires i_ceph_lock
3865 static void handle_cap_import(struct ceph_mds_client
*mdsc
,
3866 struct inode
*inode
, struct ceph_mds_caps
*im
,
3867 struct ceph_mds_cap_peer
*ph
,
3868 struct ceph_mds_session
*session
,
3869 struct ceph_cap
**target_cap
, int *old_issued
)
3871 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3872 struct ceph_cap
*cap
, *ocap
, *new_cap
= NULL
;
3873 int mds
= session
->s_mds
;
3875 unsigned caps
= le32_to_cpu(im
->caps
);
3876 unsigned wanted
= le32_to_cpu(im
->wanted
);
3877 unsigned seq
= le32_to_cpu(im
->seq
);
3878 unsigned mseq
= le32_to_cpu(im
->migrate_seq
);
3879 u64 realmino
= le64_to_cpu(im
->realm
);
3880 u64 cap_id
= le64_to_cpu(im
->cap_id
);
3885 p_cap_id
= le64_to_cpu(ph
->cap_id
);
3886 peer
= le32_to_cpu(ph
->mds
);
3892 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3893 inode
, ci
, mds
, mseq
, peer
);
3895 cap
= __get_cap_for_mds(ci
, mds
);
3898 spin_unlock(&ci
->i_ceph_lock
);
3899 new_cap
= ceph_get_cap(mdsc
, NULL
);
3900 spin_lock(&ci
->i_ceph_lock
);
3906 ceph_put_cap(mdsc
, new_cap
);
3911 __ceph_caps_issued(ci
, &issued
);
3912 issued
|= __ceph_caps_dirty(ci
);
3914 ceph_add_cap(inode
, session
, cap_id
, caps
, wanted
, seq
, mseq
,
3915 realmino
, CEPH_CAP_FLAG_AUTH
, &new_cap
);
3917 ocap
= peer
>= 0 ? __get_cap_for_mds(ci
, peer
) : NULL
;
3918 if (ocap
&& ocap
->cap_id
== p_cap_id
) {
3919 dout(" remove export cap %p mds%d flags %d\n",
3920 ocap
, peer
, ph
->flags
);
3921 if ((ph
->flags
& CEPH_CAP_FLAG_AUTH
) &&
3922 (ocap
->seq
!= le32_to_cpu(ph
->seq
) ||
3923 ocap
->mseq
!= le32_to_cpu(ph
->mseq
))) {
3924 pr_err_ratelimited("handle_cap_import: "
3925 "mismatched seq/mseq: ino (%llx.%llx) "
3926 "mds%d seq %d mseq %d importer mds%d "
3927 "has peer seq %d mseq %d\n",
3928 ceph_vinop(inode
), peer
, ocap
->seq
,
3929 ocap
->mseq
, mds
, le32_to_cpu(ph
->seq
),
3930 le32_to_cpu(ph
->mseq
));
3932 __ceph_remove_cap(ocap
, (ph
->flags
& CEPH_CAP_FLAG_RELEASE
));
3935 *old_issued
= issued
;
3940 * Handle a caps message from the MDS.
3942 * Identify the appropriate session, inode, and call the right handler
3943 * based on the cap op.
3945 void ceph_handle_caps(struct ceph_mds_session
*session
,
3946 struct ceph_msg
*msg
)
3948 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
3949 struct inode
*inode
;
3950 struct ceph_inode_info
*ci
;
3951 struct ceph_cap
*cap
;
3952 struct ceph_mds_caps
*h
;
3953 struct ceph_mds_cap_peer
*peer
= NULL
;
3954 struct ceph_snap_realm
*realm
= NULL
;
3956 int msg_version
= le16_to_cpu(msg
->hdr
.version
);
3958 struct ceph_vino vino
;
3960 size_t snaptrace_len
;
3962 struct cap_extra_info extra_info
= {};
3965 dout("handle_caps from mds%d\n", session
->s_mds
);
3968 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
3969 if (msg
->front
.iov_len
< sizeof(*h
))
3971 h
= msg
->front
.iov_base
;
3972 op
= le32_to_cpu(h
->op
);
3973 vino
.ino
= le64_to_cpu(h
->ino
);
3974 vino
.snap
= CEPH_NOSNAP
;
3975 seq
= le32_to_cpu(h
->seq
);
3976 mseq
= le32_to_cpu(h
->migrate_seq
);
3979 snaptrace_len
= le32_to_cpu(h
->snap_trace_len
);
3980 p
= snaptrace
+ snaptrace_len
;
3982 if (msg_version
>= 2) {
3984 ceph_decode_32_safe(&p
, end
, flock_len
, bad
);
3985 if (p
+ flock_len
> end
)
3990 if (msg_version
>= 3) {
3991 if (op
== CEPH_CAP_OP_IMPORT
) {
3992 if (p
+ sizeof(*peer
) > end
)
3996 } else if (op
== CEPH_CAP_OP_EXPORT
) {
3997 /* recorded in unused fields */
3998 peer
= (void *)&h
->size
;
4002 if (msg_version
>= 4) {
4003 ceph_decode_64_safe(&p
, end
, extra_info
.inline_version
, bad
);
4004 ceph_decode_32_safe(&p
, end
, extra_info
.inline_len
, bad
);
4005 if (p
+ extra_info
.inline_len
> end
)
4007 extra_info
.inline_data
= p
;
4008 p
+= extra_info
.inline_len
;
4011 if (msg_version
>= 5) {
4012 struct ceph_osd_client
*osdc
= &mdsc
->fsc
->client
->osdc
;
4015 ceph_decode_32_safe(&p
, end
, epoch_barrier
, bad
);
4016 ceph_osdc_update_epoch_barrier(osdc
, epoch_barrier
);
4019 if (msg_version
>= 8) {
4021 u32 caller_uid
, caller_gid
;
4025 ceph_decode_64_safe(&p
, end
, flush_tid
, bad
);
4027 ceph_decode_32_safe(&p
, end
, caller_uid
, bad
);
4028 ceph_decode_32_safe(&p
, end
, caller_gid
, bad
);
4030 ceph_decode_32_safe(&p
, end
, pool_ns_len
, bad
);
4031 if (pool_ns_len
> 0) {
4032 ceph_decode_need(&p
, end
, pool_ns_len
, bad
);
4033 extra_info
.pool_ns
=
4034 ceph_find_or_create_string(p
, pool_ns_len
);
4039 if (msg_version
>= 9) {
4040 struct ceph_timespec
*btime
;
4042 if (p
+ sizeof(*btime
) > end
)
4045 ceph_decode_timespec64(&extra_info
.btime
, btime
);
4046 p
+= sizeof(*btime
);
4047 ceph_decode_64_safe(&p
, end
, extra_info
.change_attr
, bad
);
4050 if (msg_version
>= 11) {
4053 ceph_decode_32_safe(&p
, end
, flags
, bad
);
4055 extra_info
.dirstat_valid
= true;
4056 ceph_decode_64_safe(&p
, end
, extra_info
.nfiles
, bad
);
4057 ceph_decode_64_safe(&p
, end
, extra_info
.nsubdirs
, bad
);
4061 inode
= ceph_find_inode(mdsc
->fsc
->sb
, vino
);
4062 ci
= ceph_inode(inode
);
4063 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op
), vino
.ino
,
4066 mutex_lock(&session
->s_mutex
);
4068 dout(" mds%d seq %lld cap seq %u\n", session
->s_mds
, session
->s_seq
,
4072 dout(" i don't have ino %llx\n", vino
.ino
);
4074 if (op
== CEPH_CAP_OP_IMPORT
) {
4075 cap
= ceph_get_cap(mdsc
, NULL
);
4076 cap
->cap_ino
= vino
.ino
;
4077 cap
->queue_release
= 1;
4078 cap
->cap_id
= le64_to_cpu(h
->cap_id
);
4081 cap
->issue_seq
= seq
;
4082 spin_lock(&session
->s_cap_lock
);
4083 __ceph_queue_cap_release(session
, cap
);
4084 spin_unlock(&session
->s_cap_lock
);
4086 goto flush_cap_releases
;
4089 /* these will work even if we don't have a cap yet */
4091 case CEPH_CAP_OP_FLUSHSNAP_ACK
:
4092 handle_cap_flushsnap_ack(inode
, le64_to_cpu(msg
->hdr
.tid
),
4096 case CEPH_CAP_OP_EXPORT
:
4097 handle_cap_export(inode
, h
, peer
, session
);
4100 case CEPH_CAP_OP_IMPORT
:
4102 if (snaptrace_len
) {
4103 down_write(&mdsc
->snap_rwsem
);
4104 ceph_update_snap_trace(mdsc
, snaptrace
,
4105 snaptrace
+ snaptrace_len
,
4107 downgrade_write(&mdsc
->snap_rwsem
);
4109 down_read(&mdsc
->snap_rwsem
);
4111 spin_lock(&ci
->i_ceph_lock
);
4112 handle_cap_import(mdsc
, inode
, h
, peer
, session
,
4113 &cap
, &extra_info
.issued
);
4114 handle_cap_grant(inode
, session
, cap
,
4115 h
, msg
->middle
, &extra_info
);
4117 ceph_put_snap_realm(mdsc
, realm
);
4121 /* the rest require a cap */
4122 spin_lock(&ci
->i_ceph_lock
);
4123 cap
= __get_cap_for_mds(ceph_inode(inode
), session
->s_mds
);
4125 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4126 inode
, ceph_ino(inode
), ceph_snap(inode
),
4128 spin_unlock(&ci
->i_ceph_lock
);
4129 goto flush_cap_releases
;
4132 /* note that each of these drops i_ceph_lock for us */
4134 case CEPH_CAP_OP_REVOKE
:
4135 case CEPH_CAP_OP_GRANT
:
4136 __ceph_caps_issued(ci
, &extra_info
.issued
);
4137 extra_info
.issued
|= __ceph_caps_dirty(ci
);
4138 handle_cap_grant(inode
, session
, cap
,
4139 h
, msg
->middle
, &extra_info
);
4142 case CEPH_CAP_OP_FLUSH_ACK
:
4143 handle_cap_flush_ack(inode
, le64_to_cpu(msg
->hdr
.tid
),
4147 case CEPH_CAP_OP_TRUNC
:
4148 queue_trunc
= handle_cap_trunc(inode
, h
, session
);
4149 spin_unlock(&ci
->i_ceph_lock
);
4151 ceph_queue_vmtruncate(inode
);
4155 spin_unlock(&ci
->i_ceph_lock
);
4156 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op
,
4157 ceph_cap_op_name(op
));
4161 mutex_unlock(&session
->s_mutex
);
4163 ceph_put_string(extra_info
.pool_ns
);
4164 /* avoid calling iput_final() in mds dispatch threads */
4165 ceph_async_iput(inode
);
4170 * send any cap release message to try to move things
4171 * along for the mds (who clearly thinks we still have this
4174 ceph_flush_cap_releases(mdsc
, session
);
4178 pr_err("ceph_handle_caps: corrupt message\n");
4184 * Delayed work handler to process end of delayed cap release LRU list.
4186 void ceph_check_delayed_caps(struct ceph_mds_client
*mdsc
)
4188 struct inode
*inode
;
4189 struct ceph_inode_info
*ci
;
4191 dout("check_delayed_caps\n");
4192 spin_lock(&mdsc
->cap_delay_lock
);
4193 while (!list_empty(&mdsc
->cap_delay_list
)) {
4194 ci
= list_first_entry(&mdsc
->cap_delay_list
,
4195 struct ceph_inode_info
,
4197 if ((ci
->i_ceph_flags
& CEPH_I_FLUSH
) == 0 &&
4198 time_before(jiffies
, ci
->i_hold_caps_max
))
4200 list_del_init(&ci
->i_cap_delay_list
);
4202 inode
= igrab(&ci
->vfs_inode
);
4204 spin_unlock(&mdsc
->cap_delay_lock
);
4205 dout("check_delayed_caps on %p\n", inode
);
4206 ceph_check_caps(ci
, 0, NULL
);
4207 /* avoid calling iput_final() in tick thread */
4208 ceph_async_iput(inode
);
4209 spin_lock(&mdsc
->cap_delay_lock
);
4212 spin_unlock(&mdsc
->cap_delay_lock
);
4216 * Flush all dirty caps to the mds
4218 static void flush_dirty_session_caps(struct ceph_mds_session
*s
)
4220 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4221 struct ceph_inode_info
*ci
;
4222 struct inode
*inode
;
4224 dout("flush_dirty_caps\n");
4225 spin_lock(&mdsc
->cap_dirty_lock
);
4226 while (!list_empty(&s
->s_cap_dirty
)) {
4227 ci
= list_first_entry(&s
->s_cap_dirty
, struct ceph_inode_info
,
4229 inode
= &ci
->vfs_inode
;
4231 dout("flush_dirty_caps %p\n", inode
);
4232 spin_unlock(&mdsc
->cap_dirty_lock
);
4233 ceph_check_caps(ci
, CHECK_CAPS_FLUSH
, NULL
);
4235 spin_lock(&mdsc
->cap_dirty_lock
);
4237 spin_unlock(&mdsc
->cap_dirty_lock
);
4238 dout("flush_dirty_caps done\n");
4241 static void iterate_sessions(struct ceph_mds_client
*mdsc
,
4242 void (*cb
)(struct ceph_mds_session
*))
4246 mutex_lock(&mdsc
->mutex
);
4247 for (mds
= 0; mds
< mdsc
->max_sessions
; ++mds
) {
4248 struct ceph_mds_session
*s
;
4250 if (!mdsc
->sessions
[mds
])
4253 s
= ceph_get_mds_session(mdsc
->sessions
[mds
]);
4257 mutex_unlock(&mdsc
->mutex
);
4259 ceph_put_mds_session(s
);
4260 mutex_lock(&mdsc
->mutex
);
4262 mutex_unlock(&mdsc
->mutex
);
4265 void ceph_flush_dirty_caps(struct ceph_mds_client
*mdsc
)
4267 iterate_sessions(mdsc
, flush_dirty_session_caps
);
4270 void __ceph_touch_fmode(struct ceph_inode_info
*ci
,
4271 struct ceph_mds_client
*mdsc
, int fmode
)
4273 unsigned long now
= jiffies
;
4274 if (fmode
& CEPH_FILE_MODE_RD
)
4275 ci
->i_last_rd
= now
;
4276 if (fmode
& CEPH_FILE_MODE_WR
)
4277 ci
->i_last_wr
= now
;
4278 /* queue periodic check */
4280 __ceph_is_any_real_caps(ci
) &&
4281 list_empty(&ci
->i_cap_delay_list
))
4282 __cap_delay_requeue(mdsc
, ci
);
4285 void ceph_get_fmode(struct ceph_inode_info
*ci
, int fmode
, int count
)
4288 int bits
= (fmode
<< 1) | 1;
4289 spin_lock(&ci
->i_ceph_lock
);
4290 for (i
= 0; i
< CEPH_FILE_MODE_BITS
; i
++) {
4291 if (bits
& (1 << i
))
4292 ci
->i_nr_by_mode
[i
] += count
;
4294 spin_unlock(&ci
->i_ceph_lock
);
4298 * Drop open file reference. If we were the last open file,
4299 * we may need to release capabilities to the MDS (or schedule
4300 * their delayed release).
4302 void ceph_put_fmode(struct ceph_inode_info
*ci
, int fmode
, int count
)
4305 int bits
= (fmode
<< 1) | 1;
4306 spin_lock(&ci
->i_ceph_lock
);
4307 for (i
= 0; i
< CEPH_FILE_MODE_BITS
; i
++) {
4308 if (bits
& (1 << i
)) {
4309 BUG_ON(ci
->i_nr_by_mode
[i
] < count
);
4310 ci
->i_nr_by_mode
[i
] -= count
;
4313 spin_unlock(&ci
->i_ceph_lock
);
4317 * For a soon-to-be unlinked file, drop the LINK caps. If it
4318 * looks like the link count will hit 0, drop any other caps (other
4319 * than PIN) we don't specifically want (due to the file still being
4322 int ceph_drop_caps_for_unlink(struct inode
*inode
)
4324 struct ceph_inode_info
*ci
= ceph_inode(inode
);
4325 int drop
= CEPH_CAP_LINK_SHARED
| CEPH_CAP_LINK_EXCL
;
4327 spin_lock(&ci
->i_ceph_lock
);
4328 if (inode
->i_nlink
== 1) {
4329 drop
|= ~(__ceph_caps_wanted(ci
) | CEPH_CAP_PIN
);
4331 if (__ceph_caps_dirty(ci
)) {
4332 struct ceph_mds_client
*mdsc
=
4333 ceph_inode_to_client(inode
)->mdsc
;
4334 __cap_delay_requeue_front(mdsc
, ci
);
4337 spin_unlock(&ci
->i_ceph_lock
);
4342 * Helpers for embedding cap and dentry lease releases into mds
4345 * @force is used by dentry_release (below) to force inclusion of a
4346 * record for the directory inode, even when there aren't any caps to
4349 int ceph_encode_inode_release(void **p
, struct inode
*inode
,
4350 int mds
, int drop
, int unless
, int force
)
4352 struct ceph_inode_info
*ci
= ceph_inode(inode
);
4353 struct ceph_cap
*cap
;
4354 struct ceph_mds_request_release
*rel
= *p
;
4358 spin_lock(&ci
->i_ceph_lock
);
4359 used
= __ceph_caps_used(ci
);
4360 dirty
= __ceph_caps_dirty(ci
);
4362 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4363 inode
, mds
, ceph_cap_string(used
|dirty
), ceph_cap_string(drop
),
4364 ceph_cap_string(unless
));
4366 /* only drop unused, clean caps */
4367 drop
&= ~(used
| dirty
);
4369 cap
= __get_cap_for_mds(ci
, mds
);
4370 if (cap
&& __cap_is_valid(cap
)) {
4371 unless
&= cap
->issued
;
4373 if (unless
& CEPH_CAP_AUTH_EXCL
)
4374 drop
&= ~CEPH_CAP_AUTH_SHARED
;
4375 if (unless
& CEPH_CAP_LINK_EXCL
)
4376 drop
&= ~CEPH_CAP_LINK_SHARED
;
4377 if (unless
& CEPH_CAP_XATTR_EXCL
)
4378 drop
&= ~CEPH_CAP_XATTR_SHARED
;
4379 if (unless
& CEPH_CAP_FILE_EXCL
)
4380 drop
&= ~CEPH_CAP_FILE_SHARED
;
4383 if (force
|| (cap
->issued
& drop
)) {
4384 if (cap
->issued
& drop
) {
4385 int wanted
= __ceph_caps_wanted(ci
);
4386 dout("encode_inode_release %p cap %p "
4387 "%s -> %s, wanted %s -> %s\n", inode
, cap
,
4388 ceph_cap_string(cap
->issued
),
4389 ceph_cap_string(cap
->issued
& ~drop
),
4390 ceph_cap_string(cap
->mds_wanted
),
4391 ceph_cap_string(wanted
));
4393 cap
->issued
&= ~drop
;
4394 cap
->implemented
&= ~drop
;
4395 cap
->mds_wanted
= wanted
;
4396 if (cap
== ci
->i_auth_cap
&&
4397 !(wanted
& CEPH_CAP_ANY_FILE_WR
))
4398 ci
->i_requested_max_size
= 0;
4400 dout("encode_inode_release %p cap %p %s"
4401 " (force)\n", inode
, cap
,
4402 ceph_cap_string(cap
->issued
));
4405 rel
->ino
= cpu_to_le64(ceph_ino(inode
));
4406 rel
->cap_id
= cpu_to_le64(cap
->cap_id
);
4407 rel
->seq
= cpu_to_le32(cap
->seq
);
4408 rel
->issue_seq
= cpu_to_le32(cap
->issue_seq
);
4409 rel
->mseq
= cpu_to_le32(cap
->mseq
);
4410 rel
->caps
= cpu_to_le32(cap
->implemented
);
4411 rel
->wanted
= cpu_to_le32(cap
->mds_wanted
);
4417 dout("encode_inode_release %p cap %p %s (noop)\n",
4418 inode
, cap
, ceph_cap_string(cap
->issued
));
4421 spin_unlock(&ci
->i_ceph_lock
);
4425 int ceph_encode_dentry_release(void **p
, struct dentry
*dentry
,
4427 int mds
, int drop
, int unless
)
4429 struct dentry
*parent
= NULL
;
4430 struct ceph_mds_request_release
*rel
= *p
;
4431 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
4436 * force an record for the directory caps if we have a dentry lease.
4437 * this is racy (can't take i_ceph_lock and d_lock together), but it
4438 * doesn't have to be perfect; the mds will revoke anything we don't
4441 spin_lock(&dentry
->d_lock
);
4442 if (di
->lease_session
&& di
->lease_session
->s_mds
== mds
)
4445 parent
= dget(dentry
->d_parent
);
4446 dir
= d_inode(parent
);
4448 spin_unlock(&dentry
->d_lock
);
4450 ret
= ceph_encode_inode_release(p
, dir
, mds
, drop
, unless
, force
);
4453 spin_lock(&dentry
->d_lock
);
4454 if (ret
&& di
->lease_session
&& di
->lease_session
->s_mds
== mds
) {
4455 dout("encode_dentry_release %p mds%d seq %d\n",
4456 dentry
, mds
, (int)di
->lease_seq
);
4457 rel
->dname_len
= cpu_to_le32(dentry
->d_name
.len
);
4458 memcpy(*p
, dentry
->d_name
.name
, dentry
->d_name
.len
);
4459 *p
+= dentry
->d_name
.len
;
4460 rel
->dname_seq
= cpu_to_le32(di
->lease_seq
);
4461 __ceph_mdsc_drop_dentry_lease(dentry
);
4463 spin_unlock(&dentry
->d_lock
);