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%llx snap issued %s"
891 " (mask %s)\n", ceph_ino(&ci
->vfs_inode
),
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%llx cap %p issued %s"
903 " (mask %s)\n", ceph_ino(&ci
->vfs_inode
), 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%llx combo issued %s"
915 " (mask %s)\n", ceph_ino(&ci
->vfs_inode
),
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
;
1146 /* 'ci' being NULL means the remove have already occurred */
1148 dout("%s: cap inode is NULL\n", __func__
);
1152 dout("__ceph_remove_cap %p from %p\n", cap
, &ci
->vfs_inode
);
1154 mdsc
= ceph_inode_to_client(&ci
->vfs_inode
)->mdsc
;
1156 /* remove from inode's cap rbtree, and clear auth cap */
1157 rb_erase(&cap
->ci_node
, &ci
->i_caps
);
1158 if (ci
->i_auth_cap
== cap
) {
1159 WARN_ON_ONCE(!list_empty(&ci
->i_dirty_item
) &&
1160 !mdsc
->fsc
->blocklisted
);
1161 ci
->i_auth_cap
= NULL
;
1164 /* remove from session list */
1165 spin_lock(&session
->s_cap_lock
);
1166 if (session
->s_cap_iterator
== cap
) {
1167 /* not yet, we are iterating over this very cap */
1168 dout("__ceph_remove_cap delaying %p removal from session %p\n",
1171 list_del_init(&cap
->session_caps
);
1172 session
->s_nr_caps
--;
1173 atomic64_dec(&mdsc
->metric
.total_caps
);
1174 cap
->session
= NULL
;
1177 /* protect backpointer with s_cap_lock: see iterate_session_caps */
1181 * s_cap_reconnect is protected by s_cap_lock. no one changes
1182 * s_cap_gen while session is in the reconnect state.
1184 if (queue_release
&&
1185 (!session
->s_cap_reconnect
|| cap
->cap_gen
== session
->s_cap_gen
)) {
1186 cap
->queue_release
= 1;
1188 __ceph_queue_cap_release(session
, cap
);
1192 cap
->queue_release
= 0;
1194 cap
->cap_ino
= ci
->i_vino
.ino
;
1196 spin_unlock(&session
->s_cap_lock
);
1199 ceph_put_cap(mdsc
, cap
);
1201 if (!__ceph_is_any_real_caps(ci
)) {
1202 /* when reconnect denied, we remove session caps forcibly,
1203 * i_wr_ref can be non-zero. If there are ongoing write,
1204 * keep i_snap_realm.
1206 if (ci
->i_wr_ref
== 0 && ci
->i_snap_realm
)
1207 drop_inode_snap_realm(ci
);
1209 __cap_delay_cancel(mdsc
, ci
);
1213 struct cap_msg_args
{
1214 struct ceph_mds_session
*session
;
1215 u64 ino
, cid
, follows
;
1216 u64 flush_tid
, oldest_flush_tid
, size
, max_size
;
1219 struct ceph_buffer
*xattr_buf
;
1220 struct ceph_buffer
*old_xattr_buf
;
1221 struct timespec64 atime
, mtime
, ctime
, btime
;
1222 int op
, caps
, wanted
, dirty
;
1223 u32 seq
, issue_seq
, mseq
, time_warp_seq
;
1233 * cap struct size + flock buffer size + inline version + inline data size +
1234 * osd_epoch_barrier + oldest_flush_tid
1236 #define CAP_MSG_SIZE (sizeof(struct ceph_mds_caps) + \
1237 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4)
1239 /* Marshal up the cap msg to the MDS */
1240 static void encode_cap_msg(struct ceph_msg
*msg
, struct cap_msg_args
*arg
)
1242 struct ceph_mds_caps
*fc
;
1244 struct ceph_osd_client
*osdc
= &arg
->session
->s_mdsc
->fsc
->client
->osdc
;
1246 dout("%s %s %llx %llx caps %s wanted %s dirty %s seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu xattr_ver %llu xattr_len %d\n",
1247 __func__
, ceph_cap_op_name(arg
->op
), arg
->cid
, arg
->ino
,
1248 ceph_cap_string(arg
->caps
), ceph_cap_string(arg
->wanted
),
1249 ceph_cap_string(arg
->dirty
), arg
->seq
, arg
->issue_seq
,
1250 arg
->flush_tid
, arg
->oldest_flush_tid
, arg
->mseq
, arg
->follows
,
1251 arg
->size
, arg
->max_size
, arg
->xattr_version
,
1252 arg
->xattr_buf
? (int)arg
->xattr_buf
->vec
.iov_len
: 0);
1254 msg
->hdr
.version
= cpu_to_le16(10);
1255 msg
->hdr
.tid
= cpu_to_le64(arg
->flush_tid
);
1257 fc
= msg
->front
.iov_base
;
1258 memset(fc
, 0, sizeof(*fc
));
1260 fc
->cap_id
= cpu_to_le64(arg
->cid
);
1261 fc
->op
= cpu_to_le32(arg
->op
);
1262 fc
->seq
= cpu_to_le32(arg
->seq
);
1263 fc
->issue_seq
= cpu_to_le32(arg
->issue_seq
);
1264 fc
->migrate_seq
= cpu_to_le32(arg
->mseq
);
1265 fc
->caps
= cpu_to_le32(arg
->caps
);
1266 fc
->wanted
= cpu_to_le32(arg
->wanted
);
1267 fc
->dirty
= cpu_to_le32(arg
->dirty
);
1268 fc
->ino
= cpu_to_le64(arg
->ino
);
1269 fc
->snap_follows
= cpu_to_le64(arg
->follows
);
1271 fc
->size
= cpu_to_le64(arg
->size
);
1272 fc
->max_size
= cpu_to_le64(arg
->max_size
);
1273 ceph_encode_timespec64(&fc
->mtime
, &arg
->mtime
);
1274 ceph_encode_timespec64(&fc
->atime
, &arg
->atime
);
1275 ceph_encode_timespec64(&fc
->ctime
, &arg
->ctime
);
1276 fc
->time_warp_seq
= cpu_to_le32(arg
->time_warp_seq
);
1278 fc
->uid
= cpu_to_le32(from_kuid(&init_user_ns
, arg
->uid
));
1279 fc
->gid
= cpu_to_le32(from_kgid(&init_user_ns
, arg
->gid
));
1280 fc
->mode
= cpu_to_le32(arg
->mode
);
1282 fc
->xattr_version
= cpu_to_le64(arg
->xattr_version
);
1283 if (arg
->xattr_buf
) {
1284 msg
->middle
= ceph_buffer_get(arg
->xattr_buf
);
1285 fc
->xattr_len
= cpu_to_le32(arg
->xattr_buf
->vec
.iov_len
);
1286 msg
->hdr
.middle_len
= cpu_to_le32(arg
->xattr_buf
->vec
.iov_len
);
1290 /* flock buffer size (version 2) */
1291 ceph_encode_32(&p
, 0);
1292 /* inline version (version 4) */
1293 ceph_encode_64(&p
, arg
->inline_data
? 0 : CEPH_INLINE_NONE
);
1294 /* inline data size */
1295 ceph_encode_32(&p
, 0);
1297 * osd_epoch_barrier (version 5)
1298 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1299 * case it was recently changed
1301 ceph_encode_32(&p
, READ_ONCE(osdc
->epoch_barrier
));
1302 /* oldest_flush_tid (version 6) */
1303 ceph_encode_64(&p
, arg
->oldest_flush_tid
);
1306 * caller_uid/caller_gid (version 7)
1308 * Currently, we don't properly track which caller dirtied the caps
1309 * last, and force a flush of them when there is a conflict. For now,
1310 * just set this to 0:0, to emulate how the MDS has worked up to now.
1312 ceph_encode_32(&p
, 0);
1313 ceph_encode_32(&p
, 0);
1315 /* pool namespace (version 8) (mds always ignores this) */
1316 ceph_encode_32(&p
, 0);
1318 /* btime and change_attr (version 9) */
1319 ceph_encode_timespec64(p
, &arg
->btime
);
1320 p
+= sizeof(struct ceph_timespec
);
1321 ceph_encode_64(&p
, arg
->change_attr
);
1323 /* Advisory flags (version 10) */
1324 ceph_encode_32(&p
, arg
->flags
);
1328 * Queue cap releases when an inode is dropped from our cache.
1330 void __ceph_remove_caps(struct ceph_inode_info
*ci
)
1334 /* lock i_ceph_lock, because ceph_d_revalidate(..., LOOKUP_RCU)
1335 * may call __ceph_caps_issued_mask() on a freeing inode. */
1336 spin_lock(&ci
->i_ceph_lock
);
1337 p
= rb_first(&ci
->i_caps
);
1339 struct ceph_cap
*cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
1341 __ceph_remove_cap(cap
, true);
1343 spin_unlock(&ci
->i_ceph_lock
);
1347 * Prepare to send a cap message to an MDS. Update the cap state, and populate
1348 * the arg struct with the parameters that will need to be sent. This should
1349 * be done under the i_ceph_lock to guard against changes to cap state.
1351 * Make note of max_size reported/requested from mds, revoked caps
1352 * that have now been implemented.
1354 static void __prep_cap(struct cap_msg_args
*arg
, struct ceph_cap
*cap
,
1355 int op
, int flags
, int used
, int want
, int retain
,
1356 int flushing
, u64 flush_tid
, u64 oldest_flush_tid
)
1358 struct ceph_inode_info
*ci
= cap
->ci
;
1359 struct inode
*inode
= &ci
->vfs_inode
;
1362 lockdep_assert_held(&ci
->i_ceph_lock
);
1364 held
= cap
->issued
| cap
->implemented
;
1365 revoking
= cap
->implemented
& ~cap
->issued
;
1366 retain
&= ~revoking
;
1368 dout("%s %p cap %p session %p %s -> %s (revoking %s)\n",
1369 __func__
, inode
, cap
, cap
->session
,
1370 ceph_cap_string(held
), ceph_cap_string(held
& retain
),
1371 ceph_cap_string(revoking
));
1372 BUG_ON((retain
& CEPH_CAP_PIN
) == 0);
1374 ci
->i_ceph_flags
&= ~CEPH_I_FLUSH
;
1376 cap
->issued
&= retain
; /* drop bits we don't want */
1378 * Wake up any waiters on wanted -> needed transition. This is due to
1379 * the weird transition from buffered to sync IO... we need to flush
1380 * dirty pages _before_ allowing sync writes to avoid reordering.
1382 arg
->wake
= cap
->implemented
& ~cap
->issued
;
1383 cap
->implemented
&= cap
->issued
| used
;
1384 cap
->mds_wanted
= want
;
1386 arg
->session
= cap
->session
;
1387 arg
->ino
= ceph_vino(inode
).ino
;
1388 arg
->cid
= cap
->cap_id
;
1389 arg
->follows
= flushing
? ci
->i_head_snapc
->seq
: 0;
1390 arg
->flush_tid
= flush_tid
;
1391 arg
->oldest_flush_tid
= oldest_flush_tid
;
1393 arg
->size
= inode
->i_size
;
1394 ci
->i_reported_size
= arg
->size
;
1395 arg
->max_size
= ci
->i_wanted_max_size
;
1396 if (cap
== ci
->i_auth_cap
) {
1397 if (want
& CEPH_CAP_ANY_FILE_WR
)
1398 ci
->i_requested_max_size
= arg
->max_size
;
1400 ci
->i_requested_max_size
= 0;
1403 if (flushing
& CEPH_CAP_XATTR_EXCL
) {
1404 arg
->old_xattr_buf
= __ceph_build_xattrs_blob(ci
);
1405 arg
->xattr_version
= ci
->i_xattrs
.version
;
1406 arg
->xattr_buf
= ci
->i_xattrs
.blob
;
1408 arg
->xattr_buf
= NULL
;
1409 arg
->old_xattr_buf
= NULL
;
1412 arg
->mtime
= inode
->i_mtime
;
1413 arg
->atime
= inode
->i_atime
;
1414 arg
->ctime
= inode
->i_ctime
;
1415 arg
->btime
= ci
->i_btime
;
1416 arg
->change_attr
= inode_peek_iversion_raw(inode
);
1419 arg
->caps
= cap
->implemented
;
1421 arg
->dirty
= flushing
;
1423 arg
->seq
= cap
->seq
;
1424 arg
->issue_seq
= cap
->issue_seq
;
1425 arg
->mseq
= cap
->mseq
;
1426 arg
->time_warp_seq
= ci
->i_time_warp_seq
;
1428 arg
->uid
= inode
->i_uid
;
1429 arg
->gid
= inode
->i_gid
;
1430 arg
->mode
= inode
->i_mode
;
1432 arg
->inline_data
= ci
->i_inline_version
!= CEPH_INLINE_NONE
;
1433 if (!(flags
& CEPH_CLIENT_CAPS_PENDING_CAPSNAP
) &&
1434 !list_empty(&ci
->i_cap_snaps
)) {
1435 struct ceph_cap_snap
*capsnap
;
1436 list_for_each_entry_reverse(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
1437 if (capsnap
->cap_flush
.tid
)
1439 if (capsnap
->need_flush
) {
1440 flags
|= CEPH_CLIENT_CAPS_PENDING_CAPSNAP
;
1449 * Send a cap msg on the given inode.
1451 * Caller should hold snap_rwsem (read), s_mutex.
1453 static void __send_cap(struct cap_msg_args
*arg
, struct ceph_inode_info
*ci
)
1455 struct ceph_msg
*msg
;
1456 struct inode
*inode
= &ci
->vfs_inode
;
1458 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPS
, CAP_MSG_SIZE
, GFP_NOFS
, false);
1460 pr_err("error allocating cap msg: ino (%llx.%llx) flushing %s tid %llu, requeuing cap.\n",
1461 ceph_vinop(inode
), ceph_cap_string(arg
->dirty
),
1463 spin_lock(&ci
->i_ceph_lock
);
1464 __cap_delay_requeue(arg
->session
->s_mdsc
, ci
);
1465 spin_unlock(&ci
->i_ceph_lock
);
1469 encode_cap_msg(msg
, arg
);
1470 ceph_con_send(&arg
->session
->s_con
, msg
);
1471 ceph_buffer_put(arg
->old_xattr_buf
);
1473 wake_up_all(&ci
->i_cap_wq
);
1476 static inline int __send_flush_snap(struct inode
*inode
,
1477 struct ceph_mds_session
*session
,
1478 struct ceph_cap_snap
*capsnap
,
1479 u32 mseq
, u64 oldest_flush_tid
)
1481 struct cap_msg_args arg
;
1482 struct ceph_msg
*msg
;
1484 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPS
, CAP_MSG_SIZE
, GFP_NOFS
, false);
1488 arg
.session
= session
;
1489 arg
.ino
= ceph_vino(inode
).ino
;
1491 arg
.follows
= capsnap
->follows
;
1492 arg
.flush_tid
= capsnap
->cap_flush
.tid
;
1493 arg
.oldest_flush_tid
= oldest_flush_tid
;
1495 arg
.size
= capsnap
->size
;
1497 arg
.xattr_version
= capsnap
->xattr_version
;
1498 arg
.xattr_buf
= capsnap
->xattr_blob
;
1499 arg
.old_xattr_buf
= NULL
;
1501 arg
.atime
= capsnap
->atime
;
1502 arg
.mtime
= capsnap
->mtime
;
1503 arg
.ctime
= capsnap
->ctime
;
1504 arg
.btime
= capsnap
->btime
;
1505 arg
.change_attr
= capsnap
->change_attr
;
1507 arg
.op
= CEPH_CAP_OP_FLUSHSNAP
;
1508 arg
.caps
= capsnap
->issued
;
1510 arg
.dirty
= capsnap
->dirty
;
1515 arg
.time_warp_seq
= capsnap
->time_warp_seq
;
1517 arg
.uid
= capsnap
->uid
;
1518 arg
.gid
= capsnap
->gid
;
1519 arg
.mode
= capsnap
->mode
;
1521 arg
.inline_data
= capsnap
->inline_data
;
1525 encode_cap_msg(msg
, &arg
);
1526 ceph_con_send(&arg
.session
->s_con
, msg
);
1531 * When a snapshot is taken, clients accumulate dirty metadata on
1532 * inodes with capabilities in ceph_cap_snaps to describe the file
1533 * state at the time the snapshot was taken. This must be flushed
1534 * asynchronously back to the MDS once sync writes complete and dirty
1535 * data is written out.
1537 * Called under i_ceph_lock. Takes s_mutex as needed.
1539 static void __ceph_flush_snaps(struct ceph_inode_info
*ci
,
1540 struct ceph_mds_session
*session
)
1541 __releases(ci
->i_ceph_lock
)
1542 __acquires(ci
->i_ceph_lock
)
1544 struct inode
*inode
= &ci
->vfs_inode
;
1545 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
1546 struct ceph_cap_snap
*capsnap
;
1547 u64 oldest_flush_tid
= 0;
1548 u64 first_tid
= 1, last_tid
= 0;
1550 dout("__flush_snaps %p session %p\n", inode
, session
);
1552 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
1554 * we need to wait for sync writes to complete and for dirty
1555 * pages to be written out.
1557 if (capsnap
->dirty_pages
|| capsnap
->writing
)
1560 /* should be removed by ceph_try_drop_cap_snap() */
1561 BUG_ON(!capsnap
->need_flush
);
1563 /* only flush each capsnap once */
1564 if (capsnap
->cap_flush
.tid
> 0) {
1565 dout(" already flushed %p, skipping\n", capsnap
);
1569 spin_lock(&mdsc
->cap_dirty_lock
);
1570 capsnap
->cap_flush
.tid
= ++mdsc
->last_cap_flush_tid
;
1571 list_add_tail(&capsnap
->cap_flush
.g_list
,
1572 &mdsc
->cap_flush_list
);
1573 if (oldest_flush_tid
== 0)
1574 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1575 if (list_empty(&ci
->i_flushing_item
)) {
1576 list_add_tail(&ci
->i_flushing_item
,
1577 &session
->s_cap_flushing
);
1579 spin_unlock(&mdsc
->cap_dirty_lock
);
1581 list_add_tail(&capsnap
->cap_flush
.i_list
,
1582 &ci
->i_cap_flush_list
);
1585 first_tid
= capsnap
->cap_flush
.tid
;
1586 last_tid
= capsnap
->cap_flush
.tid
;
1589 ci
->i_ceph_flags
&= ~CEPH_I_FLUSH_SNAPS
;
1591 while (first_tid
<= last_tid
) {
1592 struct ceph_cap
*cap
= ci
->i_auth_cap
;
1593 struct ceph_cap_flush
*cf
;
1596 if (!(cap
&& cap
->session
== session
)) {
1597 dout("__flush_snaps %p auth cap %p not mds%d, "
1598 "stop\n", inode
, cap
, session
->s_mds
);
1603 list_for_each_entry(cf
, &ci
->i_cap_flush_list
, i_list
) {
1604 if (cf
->tid
>= first_tid
) {
1612 first_tid
= cf
->tid
+ 1;
1614 capsnap
= container_of(cf
, struct ceph_cap_snap
, cap_flush
);
1615 refcount_inc(&capsnap
->nref
);
1616 spin_unlock(&ci
->i_ceph_lock
);
1618 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1619 inode
, capsnap
, cf
->tid
, ceph_cap_string(capsnap
->dirty
));
1621 ret
= __send_flush_snap(inode
, session
, capsnap
, cap
->mseq
,
1624 pr_err("__flush_snaps: error sending cap flushsnap, "
1625 "ino (%llx.%llx) tid %llu follows %llu\n",
1626 ceph_vinop(inode
), cf
->tid
, capsnap
->follows
);
1629 ceph_put_cap_snap(capsnap
);
1630 spin_lock(&ci
->i_ceph_lock
);
1634 void ceph_flush_snaps(struct ceph_inode_info
*ci
,
1635 struct ceph_mds_session
**psession
)
1637 struct inode
*inode
= &ci
->vfs_inode
;
1638 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
1639 struct ceph_mds_session
*session
= NULL
;
1642 dout("ceph_flush_snaps %p\n", inode
);
1644 session
= *psession
;
1646 spin_lock(&ci
->i_ceph_lock
);
1647 if (!(ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
)) {
1648 dout(" no capsnap needs flush, doing nothing\n");
1651 if (!ci
->i_auth_cap
) {
1652 dout(" no auth cap (migrating?), doing nothing\n");
1656 mds
= ci
->i_auth_cap
->session
->s_mds
;
1657 if (session
&& session
->s_mds
!= mds
) {
1658 dout(" oops, wrong session %p mutex\n", session
);
1659 mutex_unlock(&session
->s_mutex
);
1660 ceph_put_mds_session(session
);
1664 spin_unlock(&ci
->i_ceph_lock
);
1665 mutex_lock(&mdsc
->mutex
);
1666 session
= __ceph_lookup_mds_session(mdsc
, mds
);
1667 mutex_unlock(&mdsc
->mutex
);
1669 dout(" inverting session/ino locks on %p\n", session
);
1670 mutex_lock(&session
->s_mutex
);
1675 // make sure flushsnap messages are sent in proper order.
1676 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
)
1677 __kick_flushing_caps(mdsc
, session
, ci
, 0);
1679 __ceph_flush_snaps(ci
, session
);
1681 spin_unlock(&ci
->i_ceph_lock
);
1684 *psession
= session
;
1685 } else if (session
) {
1686 mutex_unlock(&session
->s_mutex
);
1687 ceph_put_mds_session(session
);
1689 /* we flushed them all; remove this inode from the queue */
1690 spin_lock(&mdsc
->snap_flush_lock
);
1691 list_del_init(&ci
->i_snap_flush_item
);
1692 spin_unlock(&mdsc
->snap_flush_lock
);
1696 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1697 * Caller is then responsible for calling __mark_inode_dirty with the
1698 * returned flags value.
1700 int __ceph_mark_dirty_caps(struct ceph_inode_info
*ci
, int mask
,
1701 struct ceph_cap_flush
**pcf
)
1703 struct ceph_mds_client
*mdsc
=
1704 ceph_sb_to_client(ci
->vfs_inode
.i_sb
)->mdsc
;
1705 struct inode
*inode
= &ci
->vfs_inode
;
1706 int was
= ci
->i_dirty_caps
;
1709 lockdep_assert_held(&ci
->i_ceph_lock
);
1711 if (!ci
->i_auth_cap
) {
1712 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1713 "but no auth cap (session was closed?)\n",
1714 inode
, ceph_ino(inode
), ceph_cap_string(mask
));
1718 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci
->vfs_inode
,
1719 ceph_cap_string(mask
), ceph_cap_string(was
),
1720 ceph_cap_string(was
| mask
));
1721 ci
->i_dirty_caps
|= mask
;
1723 struct ceph_mds_session
*session
= ci
->i_auth_cap
->session
;
1725 WARN_ON_ONCE(ci
->i_prealloc_cap_flush
);
1726 swap(ci
->i_prealloc_cap_flush
, *pcf
);
1728 if (!ci
->i_head_snapc
) {
1729 WARN_ON_ONCE(!rwsem_is_locked(&mdsc
->snap_rwsem
));
1730 ci
->i_head_snapc
= ceph_get_snap_context(
1731 ci
->i_snap_realm
->cached_context
);
1733 dout(" inode %p now dirty snapc %p auth cap %p\n",
1734 &ci
->vfs_inode
, ci
->i_head_snapc
, ci
->i_auth_cap
);
1735 BUG_ON(!list_empty(&ci
->i_dirty_item
));
1736 spin_lock(&mdsc
->cap_dirty_lock
);
1737 list_add(&ci
->i_dirty_item
, &session
->s_cap_dirty
);
1738 spin_unlock(&mdsc
->cap_dirty_lock
);
1739 if (ci
->i_flushing_caps
== 0) {
1741 dirty
|= I_DIRTY_SYNC
;
1744 WARN_ON_ONCE(!ci
->i_prealloc_cap_flush
);
1746 BUG_ON(list_empty(&ci
->i_dirty_item
));
1747 if (((was
| ci
->i_flushing_caps
) & CEPH_CAP_FILE_BUFFER
) &&
1748 (mask
& CEPH_CAP_FILE_BUFFER
))
1749 dirty
|= I_DIRTY_DATASYNC
;
1750 __cap_delay_requeue(mdsc
, ci
);
1754 struct ceph_cap_flush
*ceph_alloc_cap_flush(void)
1756 return kmem_cache_alloc(ceph_cap_flush_cachep
, GFP_KERNEL
);
1759 void ceph_free_cap_flush(struct ceph_cap_flush
*cf
)
1762 kmem_cache_free(ceph_cap_flush_cachep
, cf
);
1765 static u64
__get_oldest_flush_tid(struct ceph_mds_client
*mdsc
)
1767 if (!list_empty(&mdsc
->cap_flush_list
)) {
1768 struct ceph_cap_flush
*cf
=
1769 list_first_entry(&mdsc
->cap_flush_list
,
1770 struct ceph_cap_flush
, g_list
);
1777 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1778 * Return true if caller needs to wake up flush waiters.
1780 static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client
*mdsc
,
1781 struct ceph_cap_flush
*cf
)
1783 struct ceph_cap_flush
*prev
;
1784 bool wake
= cf
->wake
;
1786 if (wake
&& cf
->g_list
.prev
!= &mdsc
->cap_flush_list
) {
1787 prev
= list_prev_entry(cf
, g_list
);
1791 list_del(&cf
->g_list
);
1795 static bool __detach_cap_flush_from_ci(struct ceph_inode_info
*ci
,
1796 struct ceph_cap_flush
*cf
)
1798 struct ceph_cap_flush
*prev
;
1799 bool wake
= cf
->wake
;
1801 if (wake
&& cf
->i_list
.prev
!= &ci
->i_cap_flush_list
) {
1802 prev
= list_prev_entry(cf
, i_list
);
1806 list_del(&cf
->i_list
);
1811 * Add dirty inode to the flushing list. Assigned a seq number so we
1812 * can wait for caps to flush without starving.
1814 * Called under i_ceph_lock. Returns the flush tid.
1816 static u64
__mark_caps_flushing(struct inode
*inode
,
1817 struct ceph_mds_session
*session
, bool wake
,
1818 u64
*oldest_flush_tid
)
1820 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
1821 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1822 struct ceph_cap_flush
*cf
= NULL
;
1825 lockdep_assert_held(&ci
->i_ceph_lock
);
1826 BUG_ON(ci
->i_dirty_caps
== 0);
1827 BUG_ON(list_empty(&ci
->i_dirty_item
));
1828 BUG_ON(!ci
->i_prealloc_cap_flush
);
1830 flushing
= ci
->i_dirty_caps
;
1831 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1832 ceph_cap_string(flushing
),
1833 ceph_cap_string(ci
->i_flushing_caps
),
1834 ceph_cap_string(ci
->i_flushing_caps
| flushing
));
1835 ci
->i_flushing_caps
|= flushing
;
1836 ci
->i_dirty_caps
= 0;
1837 dout(" inode %p now !dirty\n", inode
);
1839 swap(cf
, ci
->i_prealloc_cap_flush
);
1840 cf
->caps
= flushing
;
1843 spin_lock(&mdsc
->cap_dirty_lock
);
1844 list_del_init(&ci
->i_dirty_item
);
1846 cf
->tid
= ++mdsc
->last_cap_flush_tid
;
1847 list_add_tail(&cf
->g_list
, &mdsc
->cap_flush_list
);
1848 *oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1850 if (list_empty(&ci
->i_flushing_item
)) {
1851 list_add_tail(&ci
->i_flushing_item
, &session
->s_cap_flushing
);
1852 mdsc
->num_cap_flushing
++;
1854 spin_unlock(&mdsc
->cap_dirty_lock
);
1856 list_add_tail(&cf
->i_list
, &ci
->i_cap_flush_list
);
1862 * try to invalidate mapping pages without blocking.
1864 static int try_nonblocking_invalidate(struct inode
*inode
)
1866 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1867 u32 invalidating_gen
= ci
->i_rdcache_gen
;
1869 spin_unlock(&ci
->i_ceph_lock
);
1870 invalidate_mapping_pages(&inode
->i_data
, 0, -1);
1871 spin_lock(&ci
->i_ceph_lock
);
1873 if (inode
->i_data
.nrpages
== 0 &&
1874 invalidating_gen
== ci
->i_rdcache_gen
) {
1876 dout("try_nonblocking_invalidate %p success\n", inode
);
1877 /* save any racing async invalidate some trouble */
1878 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
- 1;
1881 dout("try_nonblocking_invalidate %p failed\n", inode
);
1885 bool __ceph_should_report_size(struct ceph_inode_info
*ci
)
1887 loff_t size
= ci
->vfs_inode
.i_size
;
1888 /* mds will adjust max size according to the reported size */
1889 if (ci
->i_flushing_caps
& CEPH_CAP_FILE_WR
)
1891 if (size
>= ci
->i_max_size
)
1893 /* half of previous max_size increment has been used */
1894 if (ci
->i_max_size
> ci
->i_reported_size
&&
1895 (size
<< 1) >= ci
->i_max_size
+ ci
->i_reported_size
)
1901 * Swiss army knife function to examine currently used and wanted
1902 * versus held caps. Release, flush, ack revoked caps to mds as
1905 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1906 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1909 void ceph_check_caps(struct ceph_inode_info
*ci
, int flags
,
1910 struct ceph_mds_session
*session
)
1912 struct inode
*inode
= &ci
->vfs_inode
;
1913 struct ceph_mds_client
*mdsc
= ceph_sb_to_mdsc(inode
->i_sb
);
1914 struct ceph_cap
*cap
;
1915 u64 flush_tid
, oldest_flush_tid
;
1916 int file_wanted
, used
, cap_used
;
1917 int took_snap_rwsem
= 0; /* true if mdsc->snap_rwsem held */
1918 int issued
, implemented
, want
, retain
, revoking
, flushing
= 0;
1919 int mds
= -1; /* keep track of how far we've gone through i_caps list
1920 to avoid an infinite loop on retry */
1922 bool queue_invalidate
= false;
1923 bool tried_invalidate
= false;
1925 spin_lock(&ci
->i_ceph_lock
);
1926 if (ci
->i_ceph_flags
& CEPH_I_FLUSH
)
1927 flags
|= CHECK_CAPS_FLUSH
;
1931 spin_lock(&ci
->i_ceph_lock
);
1933 /* Caps wanted by virtue of active open files. */
1934 file_wanted
= __ceph_caps_file_wanted(ci
);
1936 /* Caps which have active references against them */
1937 used
= __ceph_caps_used(ci
);
1940 * "issued" represents the current caps that the MDS wants us to have.
1941 * "implemented" is the set that we have been granted, and includes the
1942 * ones that have not yet been returned to the MDS (the "revoking" set,
1943 * usually because they have outstanding references).
1945 issued
= __ceph_caps_issued(ci
, &implemented
);
1946 revoking
= implemented
& ~issued
;
1950 /* The ones we currently want to retain (may be adjusted below) */
1951 retain
= file_wanted
| used
| CEPH_CAP_PIN
;
1952 if (!mdsc
->stopping
&& inode
->i_nlink
> 0) {
1954 retain
|= CEPH_CAP_ANY
; /* be greedy */
1955 } else if (S_ISDIR(inode
->i_mode
) &&
1956 (issued
& CEPH_CAP_FILE_SHARED
) &&
1957 __ceph_dir_is_complete(ci
)) {
1959 * If a directory is complete, we want to keep
1960 * the exclusive cap. So that MDS does not end up
1961 * revoking the shared cap on every create/unlink
1964 if (IS_RDONLY(inode
)) {
1965 want
= CEPH_CAP_ANY_SHARED
;
1967 want
|= CEPH_CAP_ANY_SHARED
| CEPH_CAP_FILE_EXCL
;
1972 retain
|= CEPH_CAP_ANY_SHARED
;
1974 * keep RD only if we didn't have the file open RW,
1975 * because then the mds would revoke it anyway to
1976 * journal max_size=0.
1978 if (ci
->i_max_size
== 0)
1979 retain
|= CEPH_CAP_ANY_RD
;
1983 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1984 " issued %s revoking %s retain %s %s%s\n", inode
,
1985 ceph_cap_string(file_wanted
),
1986 ceph_cap_string(used
), ceph_cap_string(ci
->i_dirty_caps
),
1987 ceph_cap_string(ci
->i_flushing_caps
),
1988 ceph_cap_string(issued
), ceph_cap_string(revoking
),
1989 ceph_cap_string(retain
),
1990 (flags
& CHECK_CAPS_AUTHONLY
) ? " AUTHONLY" : "",
1991 (flags
& CHECK_CAPS_FLUSH
) ? " FLUSH" : "");
1994 * If we no longer need to hold onto old our caps, and we may
1995 * have cached pages, but don't want them, then try to invalidate.
1996 * If we fail, it's because pages are locked.... try again later.
1998 if ((!(flags
& CHECK_CAPS_NOINVAL
) || mdsc
->stopping
) &&
1999 S_ISREG(inode
->i_mode
) &&
2000 !(ci
->i_wb_ref
|| ci
->i_wrbuffer_ref
) && /* no dirty pages... */
2001 inode
->i_data
.nrpages
&& /* have cached pages */
2002 (revoking
& (CEPH_CAP_FILE_CACHE
|
2003 CEPH_CAP_FILE_LAZYIO
)) && /* or revoking cache */
2004 !tried_invalidate
) {
2005 dout("check_caps trying to invalidate on %p\n", inode
);
2006 if (try_nonblocking_invalidate(inode
) < 0) {
2007 dout("check_caps queuing invalidate\n");
2008 queue_invalidate
= true;
2009 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
;
2011 tried_invalidate
= true;
2015 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
2017 struct cap_msg_args arg
;
2019 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
2021 /* avoid looping forever */
2022 if (mds
>= cap
->mds
||
2023 ((flags
& CHECK_CAPS_AUTHONLY
) && cap
!= ci
->i_auth_cap
))
2026 /* NOTE: no side-effects allowed, until we take s_mutex */
2029 * If we have an auth cap, we don't need to consider any
2030 * overlapping caps as used.
2033 if (ci
->i_auth_cap
&& cap
!= ci
->i_auth_cap
)
2034 cap_used
&= ~ci
->i_auth_cap
->issued
;
2036 revoking
= cap
->implemented
& ~cap
->issued
;
2037 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
2038 cap
->mds
, cap
, ceph_cap_string(cap_used
),
2039 ceph_cap_string(cap
->issued
),
2040 ceph_cap_string(cap
->implemented
),
2041 ceph_cap_string(revoking
));
2043 if (cap
== ci
->i_auth_cap
&&
2044 (cap
->issued
& CEPH_CAP_FILE_WR
)) {
2045 /* request larger max_size from MDS? */
2046 if (ci
->i_wanted_max_size
> ci
->i_max_size
&&
2047 ci
->i_wanted_max_size
> ci
->i_requested_max_size
) {
2048 dout("requesting new max_size\n");
2052 /* approaching file_max? */
2053 if (__ceph_should_report_size(ci
)) {
2054 dout("i_size approaching max_size\n");
2058 /* flush anything dirty? */
2059 if (cap
== ci
->i_auth_cap
) {
2060 if ((flags
& CHECK_CAPS_FLUSH
) && ci
->i_dirty_caps
) {
2061 dout("flushing dirty caps\n");
2064 if (ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
) {
2065 dout("flushing snap caps\n");
2070 /* completed revocation? going down and there are no caps? */
2071 if (revoking
&& (revoking
& cap_used
) == 0) {
2072 dout("completed revocation of %s\n",
2073 ceph_cap_string(cap
->implemented
& ~cap
->issued
));
2077 /* want more caps from mds? */
2078 if (want
& ~cap
->mds_wanted
) {
2079 if (want
& ~(cap
->mds_wanted
| cap
->issued
))
2081 if (!__cap_is_valid(cap
))
2085 /* things we might delay */
2086 if ((cap
->issued
& ~retain
) == 0)
2087 continue; /* nope, all good */
2090 if (session
&& session
!= cap
->session
) {
2091 dout("oops, wrong session %p mutex\n", session
);
2092 mutex_unlock(&session
->s_mutex
);
2096 session
= cap
->session
;
2097 if (mutex_trylock(&session
->s_mutex
) == 0) {
2098 dout("inverting session/ino locks on %p\n",
2100 session
= ceph_get_mds_session(session
);
2101 spin_unlock(&ci
->i_ceph_lock
);
2102 if (took_snap_rwsem
) {
2103 up_read(&mdsc
->snap_rwsem
);
2104 took_snap_rwsem
= 0;
2107 mutex_lock(&session
->s_mutex
);
2108 ceph_put_mds_session(session
);
2111 * Because we take the reference while
2112 * holding the i_ceph_lock, it should
2113 * never be NULL. Throw a warning if it
2122 /* kick flushing and flush snaps before sending normal
2124 if (cap
== ci
->i_auth_cap
&&
2126 (CEPH_I_KICK_FLUSH
| CEPH_I_FLUSH_SNAPS
))) {
2127 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
)
2128 __kick_flushing_caps(mdsc
, session
, ci
, 0);
2129 if (ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
)
2130 __ceph_flush_snaps(ci
, session
);
2135 /* take snap_rwsem after session mutex */
2136 if (!took_snap_rwsem
) {
2137 if (down_read_trylock(&mdsc
->snap_rwsem
) == 0) {
2138 dout("inverting snap/in locks on %p\n",
2140 spin_unlock(&ci
->i_ceph_lock
);
2141 down_read(&mdsc
->snap_rwsem
);
2142 took_snap_rwsem
= 1;
2145 took_snap_rwsem
= 1;
2148 if (cap
== ci
->i_auth_cap
&& ci
->i_dirty_caps
) {
2149 flushing
= ci
->i_dirty_caps
;
2150 flush_tid
= __mark_caps_flushing(inode
, session
, false,
2152 if (flags
& CHECK_CAPS_FLUSH
&&
2153 list_empty(&session
->s_cap_dirty
))
2154 mflags
|= CEPH_CLIENT_CAPS_SYNC
;
2158 spin_lock(&mdsc
->cap_dirty_lock
);
2159 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2160 spin_unlock(&mdsc
->cap_dirty_lock
);
2163 mds
= cap
->mds
; /* remember mds, so we don't repeat */
2165 __prep_cap(&arg
, cap
, CEPH_CAP_OP_UPDATE
, mflags
, cap_used
,
2166 want
, retain
, flushing
, flush_tid
, oldest_flush_tid
);
2167 spin_unlock(&ci
->i_ceph_lock
);
2169 __send_cap(&arg
, ci
);
2171 goto retry
; /* retake i_ceph_lock and restart our cap scan. */
2174 /* periodically re-calculate caps wanted by open files */
2175 if (__ceph_is_any_real_caps(ci
) &&
2176 list_empty(&ci
->i_cap_delay_list
) &&
2177 (file_wanted
& ~CEPH_CAP_PIN
) &&
2178 !(used
& (CEPH_CAP_FILE_RD
| CEPH_CAP_ANY_FILE_WR
))) {
2179 __cap_delay_requeue(mdsc
, ci
);
2182 spin_unlock(&ci
->i_ceph_lock
);
2184 if (queue_invalidate
)
2185 ceph_queue_invalidate(inode
);
2188 mutex_unlock(&session
->s_mutex
);
2189 if (took_snap_rwsem
)
2190 up_read(&mdsc
->snap_rwsem
);
2194 * Try to flush dirty caps back to the auth mds.
2196 static int try_flush_caps(struct inode
*inode
, u64
*ptid
)
2198 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
2199 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2200 struct ceph_mds_session
*session
= NULL
;
2202 u64 flush_tid
= 0, oldest_flush_tid
= 0;
2205 spin_lock(&ci
->i_ceph_lock
);
2207 if (ci
->i_dirty_caps
&& ci
->i_auth_cap
) {
2208 struct ceph_cap
*cap
= ci
->i_auth_cap
;
2209 struct cap_msg_args arg
;
2211 if (session
!= cap
->session
) {
2212 spin_unlock(&ci
->i_ceph_lock
);
2214 mutex_unlock(&session
->s_mutex
);
2215 session
= cap
->session
;
2216 mutex_lock(&session
->s_mutex
);
2219 if (cap
->session
->s_state
< CEPH_MDS_SESSION_OPEN
) {
2220 spin_unlock(&ci
->i_ceph_lock
);
2224 if (ci
->i_ceph_flags
&
2225 (CEPH_I_KICK_FLUSH
| CEPH_I_FLUSH_SNAPS
)) {
2226 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
)
2227 __kick_flushing_caps(mdsc
, session
, ci
, 0);
2228 if (ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
)
2229 __ceph_flush_snaps(ci
, session
);
2233 flushing
= ci
->i_dirty_caps
;
2234 flush_tid
= __mark_caps_flushing(inode
, session
, true,
2237 __prep_cap(&arg
, cap
, CEPH_CAP_OP_FLUSH
, CEPH_CLIENT_CAPS_SYNC
,
2238 __ceph_caps_used(ci
), __ceph_caps_wanted(ci
),
2239 (cap
->issued
| cap
->implemented
),
2240 flushing
, flush_tid
, oldest_flush_tid
);
2241 spin_unlock(&ci
->i_ceph_lock
);
2243 __send_cap(&arg
, ci
);
2245 if (!list_empty(&ci
->i_cap_flush_list
)) {
2246 struct ceph_cap_flush
*cf
=
2247 list_last_entry(&ci
->i_cap_flush_list
,
2248 struct ceph_cap_flush
, i_list
);
2250 flush_tid
= cf
->tid
;
2252 flushing
= ci
->i_flushing_caps
;
2253 spin_unlock(&ci
->i_ceph_lock
);
2257 mutex_unlock(&session
->s_mutex
);
2264 * Return true if we've flushed caps through the given flush_tid.
2266 static int caps_are_flushed(struct inode
*inode
, u64 flush_tid
)
2268 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2271 spin_lock(&ci
->i_ceph_lock
);
2272 if (!list_empty(&ci
->i_cap_flush_list
)) {
2273 struct ceph_cap_flush
* cf
=
2274 list_first_entry(&ci
->i_cap_flush_list
,
2275 struct ceph_cap_flush
, i_list
);
2276 if (cf
->tid
<= flush_tid
)
2279 spin_unlock(&ci
->i_ceph_lock
);
2284 * wait for any unsafe requests to complete.
2286 static int unsafe_request_wait(struct inode
*inode
)
2288 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2289 struct ceph_mds_request
*req1
= NULL
, *req2
= NULL
;
2292 spin_lock(&ci
->i_unsafe_lock
);
2293 if (S_ISDIR(inode
->i_mode
) && !list_empty(&ci
->i_unsafe_dirops
)) {
2294 req1
= list_last_entry(&ci
->i_unsafe_dirops
,
2295 struct ceph_mds_request
,
2297 ceph_mdsc_get_request(req1
);
2299 if (!list_empty(&ci
->i_unsafe_iops
)) {
2300 req2
= list_last_entry(&ci
->i_unsafe_iops
,
2301 struct ceph_mds_request
,
2302 r_unsafe_target_item
);
2303 ceph_mdsc_get_request(req2
);
2305 spin_unlock(&ci
->i_unsafe_lock
);
2307 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2308 inode
, req1
? req1
->r_tid
: 0ULL, req2
? req2
->r_tid
: 0ULL);
2310 ret
= !wait_for_completion_timeout(&req1
->r_safe_completion
,
2311 ceph_timeout_jiffies(req1
->r_timeout
));
2314 ceph_mdsc_put_request(req1
);
2317 ret
= !wait_for_completion_timeout(&req2
->r_safe_completion
,
2318 ceph_timeout_jiffies(req2
->r_timeout
));
2321 ceph_mdsc_put_request(req2
);
2326 int ceph_fsync(struct file
*file
, loff_t start
, loff_t end
, int datasync
)
2328 struct ceph_file_info
*fi
= file
->private_data
;
2329 struct inode
*inode
= file
->f_mapping
->host
;
2330 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2335 dout("fsync %p%s\n", inode
, datasync
? " datasync" : "");
2337 ret
= file_write_and_wait_range(file
, start
, end
);
2341 ret
= ceph_wait_on_async_create(inode
);
2345 dirty
= try_flush_caps(inode
, &flush_tid
);
2346 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty
));
2348 err
= unsafe_request_wait(inode
);
2351 * only wait on non-file metadata writeback (the mds
2352 * can recover size and mtime, so we don't need to
2355 if (!err
&& (dirty
& ~CEPH_CAP_ANY_FILE_WR
)) {
2356 err
= wait_event_interruptible(ci
->i_cap_wq
,
2357 caps_are_flushed(inode
, flush_tid
));
2363 if (errseq_check(&ci
->i_meta_err
, READ_ONCE(fi
->meta_err
))) {
2364 spin_lock(&file
->f_lock
);
2365 err
= errseq_check_and_advance(&ci
->i_meta_err
,
2367 spin_unlock(&file
->f_lock
);
2372 dout("fsync %p%s result=%d\n", inode
, datasync
? " datasync" : "", ret
);
2377 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2378 * queue inode for flush but don't do so immediately, because we can
2379 * get by with fewer MDS messages if we wait for data writeback to
2382 int ceph_write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
2384 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2388 int wait
= (wbc
->sync_mode
== WB_SYNC_ALL
&& !wbc
->for_sync
);
2390 dout("write_inode %p wait=%d\n", inode
, wait
);
2392 dirty
= try_flush_caps(inode
, &flush_tid
);
2394 err
= wait_event_interruptible(ci
->i_cap_wq
,
2395 caps_are_flushed(inode
, flush_tid
));
2397 struct ceph_mds_client
*mdsc
=
2398 ceph_sb_to_client(inode
->i_sb
)->mdsc
;
2400 spin_lock(&ci
->i_ceph_lock
);
2401 if (__ceph_caps_dirty(ci
))
2402 __cap_delay_requeue_front(mdsc
, ci
);
2403 spin_unlock(&ci
->i_ceph_lock
);
2408 static void __kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2409 struct ceph_mds_session
*session
,
2410 struct ceph_inode_info
*ci
,
2411 u64 oldest_flush_tid
)
2412 __releases(ci
->i_ceph_lock
)
2413 __acquires(ci
->i_ceph_lock
)
2415 struct inode
*inode
= &ci
->vfs_inode
;
2416 struct ceph_cap
*cap
;
2417 struct ceph_cap_flush
*cf
;
2420 u64 last_snap_flush
= 0;
2422 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
2424 list_for_each_entry_reverse(cf
, &ci
->i_cap_flush_list
, i_list
) {
2426 last_snap_flush
= cf
->tid
;
2431 list_for_each_entry(cf
, &ci
->i_cap_flush_list
, i_list
) {
2432 if (cf
->tid
< first_tid
)
2435 cap
= ci
->i_auth_cap
;
2436 if (!(cap
&& cap
->session
== session
)) {
2437 pr_err("%p auth cap %p not mds%d ???\n",
2438 inode
, cap
, session
->s_mds
);
2442 first_tid
= cf
->tid
+ 1;
2445 struct cap_msg_args arg
;
2447 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2448 inode
, cap
, cf
->tid
, ceph_cap_string(cf
->caps
));
2449 __prep_cap(&arg
, cap
, CEPH_CAP_OP_FLUSH
,
2450 (cf
->tid
< last_snap_flush
?
2451 CEPH_CLIENT_CAPS_PENDING_CAPSNAP
: 0),
2452 __ceph_caps_used(ci
),
2453 __ceph_caps_wanted(ci
),
2454 (cap
->issued
| cap
->implemented
),
2455 cf
->caps
, cf
->tid
, oldest_flush_tid
);
2456 spin_unlock(&ci
->i_ceph_lock
);
2457 __send_cap(&arg
, ci
);
2459 struct ceph_cap_snap
*capsnap
=
2460 container_of(cf
, struct ceph_cap_snap
,
2462 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2463 inode
, capsnap
, cf
->tid
,
2464 ceph_cap_string(capsnap
->dirty
));
2466 refcount_inc(&capsnap
->nref
);
2467 spin_unlock(&ci
->i_ceph_lock
);
2469 ret
= __send_flush_snap(inode
, session
, capsnap
, cap
->mseq
,
2472 pr_err("kick_flushing_caps: error sending "
2473 "cap flushsnap, ino (%llx.%llx) "
2474 "tid %llu follows %llu\n",
2475 ceph_vinop(inode
), cf
->tid
,
2479 ceph_put_cap_snap(capsnap
);
2482 spin_lock(&ci
->i_ceph_lock
);
2486 void ceph_early_kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2487 struct ceph_mds_session
*session
)
2489 struct ceph_inode_info
*ci
;
2490 struct ceph_cap
*cap
;
2491 u64 oldest_flush_tid
;
2493 dout("early_kick_flushing_caps mds%d\n", session
->s_mds
);
2495 spin_lock(&mdsc
->cap_dirty_lock
);
2496 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2497 spin_unlock(&mdsc
->cap_dirty_lock
);
2499 list_for_each_entry(ci
, &session
->s_cap_flushing
, i_flushing_item
) {
2500 spin_lock(&ci
->i_ceph_lock
);
2501 cap
= ci
->i_auth_cap
;
2502 if (!(cap
&& cap
->session
== session
)) {
2503 pr_err("%p auth cap %p not mds%d ???\n",
2504 &ci
->vfs_inode
, cap
, session
->s_mds
);
2505 spin_unlock(&ci
->i_ceph_lock
);
2511 * if flushing caps were revoked, we re-send the cap flush
2512 * in client reconnect stage. This guarantees MDS * processes
2513 * the cap flush message before issuing the flushing caps to
2516 if ((cap
->issued
& ci
->i_flushing_caps
) !=
2517 ci
->i_flushing_caps
) {
2518 /* encode_caps_cb() also will reset these sequence
2519 * numbers. make sure sequence numbers in cap flush
2520 * message match later reconnect message */
2524 __kick_flushing_caps(mdsc
, session
, ci
,
2527 ci
->i_ceph_flags
|= CEPH_I_KICK_FLUSH
;
2530 spin_unlock(&ci
->i_ceph_lock
);
2534 void ceph_kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2535 struct ceph_mds_session
*session
)
2537 struct ceph_inode_info
*ci
;
2538 struct ceph_cap
*cap
;
2539 u64 oldest_flush_tid
;
2541 lockdep_assert_held(&session
->s_mutex
);
2543 dout("kick_flushing_caps mds%d\n", session
->s_mds
);
2545 spin_lock(&mdsc
->cap_dirty_lock
);
2546 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2547 spin_unlock(&mdsc
->cap_dirty_lock
);
2549 list_for_each_entry(ci
, &session
->s_cap_flushing
, i_flushing_item
) {
2550 spin_lock(&ci
->i_ceph_lock
);
2551 cap
= ci
->i_auth_cap
;
2552 if (!(cap
&& cap
->session
== session
)) {
2553 pr_err("%p auth cap %p not mds%d ???\n",
2554 &ci
->vfs_inode
, cap
, session
->s_mds
);
2555 spin_unlock(&ci
->i_ceph_lock
);
2558 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
) {
2559 __kick_flushing_caps(mdsc
, session
, ci
,
2562 spin_unlock(&ci
->i_ceph_lock
);
2566 void ceph_kick_flushing_inode_caps(struct ceph_mds_session
*session
,
2567 struct ceph_inode_info
*ci
)
2569 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2570 struct ceph_cap
*cap
= ci
->i_auth_cap
;
2572 lockdep_assert_held(&ci
->i_ceph_lock
);
2574 dout("%s %p flushing %s\n", __func__
, &ci
->vfs_inode
,
2575 ceph_cap_string(ci
->i_flushing_caps
));
2577 if (!list_empty(&ci
->i_cap_flush_list
)) {
2578 u64 oldest_flush_tid
;
2579 spin_lock(&mdsc
->cap_dirty_lock
);
2580 list_move_tail(&ci
->i_flushing_item
,
2581 &cap
->session
->s_cap_flushing
);
2582 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2583 spin_unlock(&mdsc
->cap_dirty_lock
);
2585 __kick_flushing_caps(mdsc
, session
, ci
, oldest_flush_tid
);
2591 * Take references to capabilities we hold, so that we don't release
2592 * them to the MDS prematurely.
2594 void ceph_take_cap_refs(struct ceph_inode_info
*ci
, int got
,
2595 bool snap_rwsem_locked
)
2597 lockdep_assert_held(&ci
->i_ceph_lock
);
2599 if (got
& CEPH_CAP_PIN
)
2601 if (got
& CEPH_CAP_FILE_RD
)
2603 if (got
& CEPH_CAP_FILE_CACHE
)
2604 ci
->i_rdcache_ref
++;
2605 if (got
& CEPH_CAP_FILE_EXCL
)
2607 if (got
& CEPH_CAP_FILE_WR
) {
2608 if (ci
->i_wr_ref
== 0 && !ci
->i_head_snapc
) {
2609 BUG_ON(!snap_rwsem_locked
);
2610 ci
->i_head_snapc
= ceph_get_snap_context(
2611 ci
->i_snap_realm
->cached_context
);
2615 if (got
& CEPH_CAP_FILE_BUFFER
) {
2616 if (ci
->i_wb_ref
== 0)
2617 ihold(&ci
->vfs_inode
);
2619 dout("%s %p wb %d -> %d (?)\n", __func__
,
2620 &ci
->vfs_inode
, ci
->i_wb_ref
-1, ci
->i_wb_ref
);
2625 * Try to grab cap references. Specify those refs we @want, and the
2626 * minimal set we @need. Also include the larger offset we are writing
2627 * to (when applicable), and check against max_size here as well.
2628 * Note that caller is responsible for ensuring max_size increases are
2629 * requested from the MDS.
2631 * Returns 0 if caps were not able to be acquired (yet), 1 if succeed,
2632 * or a negative error code. There are 3 speical error codes:
2633 * -EAGAIN: need to sleep but non-blocking is specified
2634 * -EFBIG: ask caller to call check_max_size() and try again.
2635 * -ESTALE: ask caller to call ceph_renew_caps() and try again.
2638 /* first 8 bits are reserved for CEPH_FILE_MODE_FOO */
2639 NON_BLOCKING
= (1 << 8),
2640 CHECK_FILELOCK
= (1 << 9),
2643 static int try_get_cap_refs(struct inode
*inode
, int need
, int want
,
2644 loff_t endoff
, int flags
, int *got
)
2646 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2647 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
2649 int have
, implemented
;
2650 bool snap_rwsem_locked
= false;
2652 dout("get_cap_refs %p need %s want %s\n", inode
,
2653 ceph_cap_string(need
), ceph_cap_string(want
));
2656 spin_lock(&ci
->i_ceph_lock
);
2658 if ((flags
& CHECK_FILELOCK
) &&
2659 (ci
->i_ceph_flags
& CEPH_I_ERROR_FILELOCK
)) {
2660 dout("try_get_cap_refs %p error filelock\n", inode
);
2665 /* finish pending truncate */
2666 while (ci
->i_truncate_pending
) {
2667 spin_unlock(&ci
->i_ceph_lock
);
2668 if (snap_rwsem_locked
) {
2669 up_read(&mdsc
->snap_rwsem
);
2670 snap_rwsem_locked
= false;
2672 __ceph_do_pending_vmtruncate(inode
);
2673 spin_lock(&ci
->i_ceph_lock
);
2676 have
= __ceph_caps_issued(ci
, &implemented
);
2678 if (have
& need
& CEPH_CAP_FILE_WR
) {
2679 if (endoff
>= 0 && endoff
> (loff_t
)ci
->i_max_size
) {
2680 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2681 inode
, endoff
, ci
->i_max_size
);
2682 if (endoff
> ci
->i_requested_max_size
)
2683 ret
= ci
->i_auth_cap
? -EFBIG
: -ESTALE
;
2687 * If a sync write is in progress, we must wait, so that we
2688 * can get a final snapshot value for size+mtime.
2690 if (__ceph_have_pending_cap_snap(ci
)) {
2691 dout("get_cap_refs %p cap_snap_pending\n", inode
);
2696 if ((have
& need
) == need
) {
2698 * Look at (implemented & ~have & not) so that we keep waiting
2699 * on transition from wanted -> needed caps. This is needed
2700 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2701 * going before a prior buffered writeback happens.
2703 int not = want
& ~(have
& need
);
2704 int revoking
= implemented
& ~have
;
2705 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2706 inode
, ceph_cap_string(have
), ceph_cap_string(not),
2707 ceph_cap_string(revoking
));
2708 if ((revoking
& not) == 0) {
2709 if (!snap_rwsem_locked
&&
2710 !ci
->i_head_snapc
&&
2711 (need
& CEPH_CAP_FILE_WR
)) {
2712 if (!down_read_trylock(&mdsc
->snap_rwsem
)) {
2714 * we can not call down_read() when
2715 * task isn't in TASK_RUNNING state
2717 if (flags
& NON_BLOCKING
) {
2722 spin_unlock(&ci
->i_ceph_lock
);
2723 down_read(&mdsc
->snap_rwsem
);
2724 snap_rwsem_locked
= true;
2727 snap_rwsem_locked
= true;
2729 if ((have
& want
) == want
)
2733 if (S_ISREG(inode
->i_mode
) &&
2734 (need
& CEPH_CAP_FILE_RD
) &&
2735 !(*got
& CEPH_CAP_FILE_CACHE
))
2736 ceph_disable_fscache_readpage(ci
);
2737 ceph_take_cap_refs(ci
, *got
, true);
2741 int session_readonly
= false;
2743 if (ci
->i_auth_cap
&&
2744 (need
& (CEPH_CAP_FILE_WR
| CEPH_CAP_FILE_EXCL
))) {
2745 struct ceph_mds_session
*s
= ci
->i_auth_cap
->session
;
2746 spin_lock(&s
->s_cap_lock
);
2747 session_readonly
= s
->s_readonly
;
2748 spin_unlock(&s
->s_cap_lock
);
2750 if (session_readonly
) {
2751 dout("get_cap_refs %p need %s but mds%d readonly\n",
2752 inode
, ceph_cap_string(need
), ci
->i_auth_cap
->mds
);
2757 if (READ_ONCE(mdsc
->fsc
->mount_state
) >= CEPH_MOUNT_SHUTDOWN
) {
2758 dout("get_cap_refs %p forced umount\n", inode
);
2762 mds_wanted
= __ceph_caps_mds_wanted(ci
, false);
2763 if (need
& ~mds_wanted
) {
2764 dout("get_cap_refs %p need %s > mds_wanted %s\n",
2765 inode
, ceph_cap_string(need
),
2766 ceph_cap_string(mds_wanted
));
2771 dout("get_cap_refs %p have %s need %s\n", inode
,
2772 ceph_cap_string(have
), ceph_cap_string(need
));
2776 __ceph_touch_fmode(ci
, mdsc
, flags
);
2778 spin_unlock(&ci
->i_ceph_lock
);
2779 if (snap_rwsem_locked
)
2780 up_read(&mdsc
->snap_rwsem
);
2783 ceph_update_cap_mis(&mdsc
->metric
);
2785 ceph_update_cap_hit(&mdsc
->metric
);
2787 dout("get_cap_refs %p ret %d got %s\n", inode
,
2788 ret
, ceph_cap_string(*got
));
2793 * Check the offset we are writing up to against our current
2794 * max_size. If necessary, tell the MDS we want to write to
2797 static void check_max_size(struct inode
*inode
, loff_t endoff
)
2799 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2802 /* do we need to explicitly request a larger max_size? */
2803 spin_lock(&ci
->i_ceph_lock
);
2804 if (endoff
>= ci
->i_max_size
&& endoff
> ci
->i_wanted_max_size
) {
2805 dout("write %p at large endoff %llu, req max_size\n",
2807 ci
->i_wanted_max_size
= endoff
;
2809 /* duplicate ceph_check_caps()'s logic */
2810 if (ci
->i_auth_cap
&&
2811 (ci
->i_auth_cap
->issued
& CEPH_CAP_FILE_WR
) &&
2812 ci
->i_wanted_max_size
> ci
->i_max_size
&&
2813 ci
->i_wanted_max_size
> ci
->i_requested_max_size
)
2815 spin_unlock(&ci
->i_ceph_lock
);
2817 ceph_check_caps(ci
, CHECK_CAPS_AUTHONLY
, NULL
);
2820 static inline int get_used_fmode(int caps
)
2823 if (caps
& CEPH_CAP_FILE_RD
)
2824 fmode
|= CEPH_FILE_MODE_RD
;
2825 if (caps
& CEPH_CAP_FILE_WR
)
2826 fmode
|= CEPH_FILE_MODE_WR
;
2830 int ceph_try_get_caps(struct inode
*inode
, int need
, int want
,
2831 bool nonblock
, int *got
)
2835 BUG_ON(need
& ~CEPH_CAP_FILE_RD
);
2836 BUG_ON(want
& ~(CEPH_CAP_FILE_CACHE
| CEPH_CAP_FILE_LAZYIO
|
2837 CEPH_CAP_FILE_SHARED
| CEPH_CAP_FILE_EXCL
|
2838 CEPH_CAP_ANY_DIR_OPS
));
2840 ret
= ceph_pool_perm_check(inode
, need
);
2845 flags
= get_used_fmode(need
| want
);
2847 flags
|= NON_BLOCKING
;
2849 ret
= try_get_cap_refs(inode
, need
, want
, 0, flags
, got
);
2850 /* three special error codes */
2851 if (ret
== -EAGAIN
|| ret
== -EFBIG
|| ret
== -ESTALE
)
2857 * Wait for caps, and take cap references. If we can't get a WR cap
2858 * due to a small max_size, make sure we check_max_size (and possibly
2859 * ask the mds) so we don't get hung up indefinitely.
2861 int ceph_get_caps(struct file
*filp
, int need
, int want
,
2862 loff_t endoff
, int *got
, struct page
**pinned_page
)
2864 struct ceph_file_info
*fi
= filp
->private_data
;
2865 struct inode
*inode
= file_inode(filp
);
2866 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2867 struct ceph_fs_client
*fsc
= ceph_inode_to_client(inode
);
2868 int ret
, _got
, flags
;
2870 ret
= ceph_pool_perm_check(inode
, need
);
2874 if ((fi
->fmode
& CEPH_FILE_MODE_WR
) &&
2875 fi
->filp_gen
!= READ_ONCE(fsc
->filp_gen
))
2878 flags
= get_used_fmode(need
| want
);
2881 flags
&= CEPH_FILE_MODE_MASK
;
2882 if (atomic_read(&fi
->num_locks
))
2883 flags
|= CHECK_FILELOCK
;
2885 ret
= try_get_cap_refs(inode
, need
, want
, endoff
,
2887 WARN_ON_ONCE(ret
== -EAGAIN
);
2889 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
2891 DEFINE_WAIT_FUNC(wait
, woken_wake_function
);
2893 cw
.ino
= ceph_ino(inode
);
2894 cw
.tgid
= current
->tgid
;
2898 spin_lock(&mdsc
->caps_list_lock
);
2899 list_add(&cw
.list
, &mdsc
->cap_wait_list
);
2900 spin_unlock(&mdsc
->caps_list_lock
);
2902 /* make sure used fmode not timeout */
2903 ceph_get_fmode(ci
, flags
, FMODE_WAIT_BIAS
);
2904 add_wait_queue(&ci
->i_cap_wq
, &wait
);
2906 flags
|= NON_BLOCKING
;
2907 while (!(ret
= try_get_cap_refs(inode
, need
, want
,
2908 endoff
, flags
, &_got
))) {
2909 if (signal_pending(current
)) {
2913 wait_woken(&wait
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
2916 remove_wait_queue(&ci
->i_cap_wq
, &wait
);
2917 ceph_put_fmode(ci
, flags
, FMODE_WAIT_BIAS
);
2919 spin_lock(&mdsc
->caps_list_lock
);
2921 spin_unlock(&mdsc
->caps_list_lock
);
2927 if ((fi
->fmode
& CEPH_FILE_MODE_WR
) &&
2928 fi
->filp_gen
!= READ_ONCE(fsc
->filp_gen
)) {
2929 if (ret
>= 0 && _got
)
2930 ceph_put_cap_refs(ci
, _got
);
2935 if (ret
== -EFBIG
|| ret
== -ESTALE
) {
2936 int ret2
= ceph_wait_on_async_create(inode
);
2940 if (ret
== -EFBIG
) {
2941 check_max_size(inode
, endoff
);
2944 if (ret
== -ESTALE
) {
2945 /* session was killed, try renew caps */
2946 ret
= ceph_renew_caps(inode
, flags
);
2953 if (S_ISREG(ci
->vfs_inode
.i_mode
) &&
2954 ci
->i_inline_version
!= CEPH_INLINE_NONE
&&
2955 (_got
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)) &&
2956 i_size_read(inode
) > 0) {
2958 find_get_page(inode
->i_mapping
, 0);
2960 if (PageUptodate(page
)) {
2961 *pinned_page
= page
;
2967 * drop cap refs first because getattr while
2968 * holding * caps refs can cause deadlock.
2970 ceph_put_cap_refs(ci
, _got
);
2974 * getattr request will bring inline data into
2977 ret
= __ceph_do_getattr(inode
, NULL
,
2978 CEPH_STAT_CAP_INLINE_DATA
,
2987 if (S_ISREG(ci
->vfs_inode
.i_mode
) &&
2988 (_got
& CEPH_CAP_FILE_RD
) && (_got
& CEPH_CAP_FILE_CACHE
))
2989 ceph_fscache_revalidate_cookie(ci
);
2996 * Take cap refs. Caller must already know we hold at least one ref
2997 * on the caps in question or we don't know this is safe.
2999 void ceph_get_cap_refs(struct ceph_inode_info
*ci
, int caps
)
3001 spin_lock(&ci
->i_ceph_lock
);
3002 ceph_take_cap_refs(ci
, caps
, false);
3003 spin_unlock(&ci
->i_ceph_lock
);
3008 * drop cap_snap that is not associated with any snapshot.
3009 * we don't need to send FLUSHSNAP message for it.
3011 static int ceph_try_drop_cap_snap(struct ceph_inode_info
*ci
,
3012 struct ceph_cap_snap
*capsnap
)
3014 if (!capsnap
->need_flush
&&
3015 !capsnap
->writing
&& !capsnap
->dirty_pages
) {
3016 dout("dropping cap_snap %p follows %llu\n",
3017 capsnap
, capsnap
->follows
);
3018 BUG_ON(capsnap
->cap_flush
.tid
> 0);
3019 ceph_put_snap_context(capsnap
->context
);
3020 if (!list_is_last(&capsnap
->ci_item
, &ci
->i_cap_snaps
))
3021 ci
->i_ceph_flags
|= CEPH_I_FLUSH_SNAPS
;
3023 list_del(&capsnap
->ci_item
);
3024 ceph_put_cap_snap(capsnap
);
3033 * If we released the last ref on any given cap, call ceph_check_caps
3034 * to release (or schedule a release).
3036 * If we are releasing a WR cap (from a sync write), finalize any affected
3037 * cap_snap, and wake up any waiters.
3039 static void __ceph_put_cap_refs(struct ceph_inode_info
*ci
, int had
,
3040 bool skip_checking_caps
)
3042 struct inode
*inode
= &ci
->vfs_inode
;
3043 int last
= 0, put
= 0, flushsnaps
= 0, wake
= 0;
3045 spin_lock(&ci
->i_ceph_lock
);
3046 if (had
& CEPH_CAP_PIN
)
3048 if (had
& CEPH_CAP_FILE_RD
)
3049 if (--ci
->i_rd_ref
== 0)
3051 if (had
& CEPH_CAP_FILE_CACHE
)
3052 if (--ci
->i_rdcache_ref
== 0)
3054 if (had
& CEPH_CAP_FILE_EXCL
)
3055 if (--ci
->i_fx_ref
== 0)
3057 if (had
& CEPH_CAP_FILE_BUFFER
) {
3058 if (--ci
->i_wb_ref
== 0) {
3062 dout("put_cap_refs %p wb %d -> %d (?)\n",
3063 inode
, ci
->i_wb_ref
+1, ci
->i_wb_ref
);
3065 if (had
& CEPH_CAP_FILE_WR
)
3066 if (--ci
->i_wr_ref
== 0) {
3068 if (__ceph_have_pending_cap_snap(ci
)) {
3069 struct ceph_cap_snap
*capsnap
=
3070 list_last_entry(&ci
->i_cap_snaps
,
3071 struct ceph_cap_snap
,
3073 capsnap
->writing
= 0;
3074 if (ceph_try_drop_cap_snap(ci
, capsnap
))
3076 else if (__ceph_finish_cap_snap(ci
, capsnap
))
3080 if (ci
->i_wrbuffer_ref_head
== 0 &&
3081 ci
->i_dirty_caps
== 0 &&
3082 ci
->i_flushing_caps
== 0) {
3083 BUG_ON(!ci
->i_head_snapc
);
3084 ceph_put_snap_context(ci
->i_head_snapc
);
3085 ci
->i_head_snapc
= NULL
;
3087 /* see comment in __ceph_remove_cap() */
3088 if (!__ceph_is_any_real_caps(ci
) && ci
->i_snap_realm
)
3089 drop_inode_snap_realm(ci
);
3091 spin_unlock(&ci
->i_ceph_lock
);
3093 dout("put_cap_refs %p had %s%s%s\n", inode
, ceph_cap_string(had
),
3094 last
? " last" : "", put
? " put" : "");
3096 if (last
&& !skip_checking_caps
)
3097 ceph_check_caps(ci
, 0, NULL
);
3098 else if (flushsnaps
)
3099 ceph_flush_snaps(ci
, NULL
);
3101 wake_up_all(&ci
->i_cap_wq
);
3106 void ceph_put_cap_refs(struct ceph_inode_info
*ci
, int had
)
3108 __ceph_put_cap_refs(ci
, had
, false);
3111 void ceph_put_cap_refs_no_check_caps(struct ceph_inode_info
*ci
, int had
)
3113 __ceph_put_cap_refs(ci
, had
, true);
3117 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
3118 * context. Adjust per-snap dirty page accounting as appropriate.
3119 * Once all dirty data for a cap_snap is flushed, flush snapped file
3120 * metadata back to the MDS. If we dropped the last ref, call
3123 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info
*ci
, int nr
,
3124 struct ceph_snap_context
*snapc
)
3126 struct inode
*inode
= &ci
->vfs_inode
;
3127 struct ceph_cap_snap
*capsnap
= NULL
;
3131 bool flush_snaps
= false;
3132 bool complete_capsnap
= false;
3134 spin_lock(&ci
->i_ceph_lock
);
3135 ci
->i_wrbuffer_ref
-= nr
;
3136 if (ci
->i_wrbuffer_ref
== 0) {
3141 if (ci
->i_head_snapc
== snapc
) {
3142 ci
->i_wrbuffer_ref_head
-= nr
;
3143 if (ci
->i_wrbuffer_ref_head
== 0 &&
3144 ci
->i_wr_ref
== 0 &&
3145 ci
->i_dirty_caps
== 0 &&
3146 ci
->i_flushing_caps
== 0) {
3147 BUG_ON(!ci
->i_head_snapc
);
3148 ceph_put_snap_context(ci
->i_head_snapc
);
3149 ci
->i_head_snapc
= NULL
;
3151 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
3153 ci
->i_wrbuffer_ref
+nr
, ci
->i_wrbuffer_ref_head
+nr
,
3154 ci
->i_wrbuffer_ref
, ci
->i_wrbuffer_ref_head
,
3155 last
? " LAST" : "");
3157 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
3158 if (capsnap
->context
== snapc
) {
3164 capsnap
->dirty_pages
-= nr
;
3165 if (capsnap
->dirty_pages
== 0) {
3166 complete_capsnap
= true;
3167 if (!capsnap
->writing
) {
3168 if (ceph_try_drop_cap_snap(ci
, capsnap
)) {
3171 ci
->i_ceph_flags
|= CEPH_I_FLUSH_SNAPS
;
3176 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
3177 " snap %lld %d/%d -> %d/%d %s%s\n",
3178 inode
, capsnap
, capsnap
->context
->seq
,
3179 ci
->i_wrbuffer_ref
+nr
, capsnap
->dirty_pages
+ nr
,
3180 ci
->i_wrbuffer_ref
, capsnap
->dirty_pages
,
3181 last
? " (wrbuffer last)" : "",
3182 complete_capsnap
? " (complete capsnap)" : "");
3185 spin_unlock(&ci
->i_ceph_lock
);
3188 ceph_check_caps(ci
, 0, NULL
);
3189 } else if (flush_snaps
) {
3190 ceph_flush_snaps(ci
, NULL
);
3192 if (complete_capsnap
)
3193 wake_up_all(&ci
->i_cap_wq
);
3195 /* avoid calling iput_final() in osd dispatch threads */
3196 ceph_async_iput(inode
);
3201 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
3203 static void invalidate_aliases(struct inode
*inode
)
3205 struct dentry
*dn
, *prev
= NULL
;
3207 dout("invalidate_aliases inode %p\n", inode
);
3208 d_prune_aliases(inode
);
3210 * For non-directory inode, d_find_alias() only returns
3211 * hashed dentry. After calling d_invalidate(), the
3212 * dentry becomes unhashed.
3214 * For directory inode, d_find_alias() can return
3215 * unhashed dentry. But directory inode should have
3216 * one alias at most.
3218 while ((dn
= d_find_alias(inode
))) {
3232 struct cap_extra_info
{
3233 struct ceph_string
*pool_ns
;
3243 /* currently issued */
3245 struct timespec64 btime
;
3249 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
3250 * actually be a revocation if it specifies a smaller cap set.)
3252 * caller holds s_mutex and i_ceph_lock, we drop both.
3254 static void handle_cap_grant(struct inode
*inode
,
3255 struct ceph_mds_session
*session
,
3256 struct ceph_cap
*cap
,
3257 struct ceph_mds_caps
*grant
,
3258 struct ceph_buffer
*xattr_buf
,
3259 struct cap_extra_info
*extra_info
)
3260 __releases(ci
->i_ceph_lock
)
3261 __releases(session
->s_mdsc
->snap_rwsem
)
3263 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3264 int seq
= le32_to_cpu(grant
->seq
);
3265 int newcaps
= le32_to_cpu(grant
->caps
);
3266 int used
, wanted
, dirty
;
3267 u64 size
= le64_to_cpu(grant
->size
);
3268 u64 max_size
= le64_to_cpu(grant
->max_size
);
3269 unsigned char check_caps
= 0;
3270 bool was_stale
= cap
->cap_gen
< session
->s_cap_gen
;
3272 bool writeback
= false;
3273 bool queue_trunc
= false;
3274 bool queue_invalidate
= false;
3275 bool deleted_inode
= false;
3276 bool fill_inline
= false;
3278 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
3279 inode
, cap
, session
->s_mds
, seq
, ceph_cap_string(newcaps
));
3280 dout(" size %llu max_size %llu, i_size %llu\n", size
, max_size
,
3285 * If CACHE is being revoked, and we have no dirty buffers,
3286 * try to invalidate (once). (If there are dirty buffers, we
3287 * will invalidate _after_ writeback.)
3289 if (S_ISREG(inode
->i_mode
) && /* don't invalidate readdir cache */
3290 ((cap
->issued
& ~newcaps
) & CEPH_CAP_FILE_CACHE
) &&
3291 (newcaps
& CEPH_CAP_FILE_LAZYIO
) == 0 &&
3292 !(ci
->i_wrbuffer_ref
|| ci
->i_wb_ref
)) {
3293 if (try_nonblocking_invalidate(inode
)) {
3294 /* there were locked pages.. invalidate later
3295 in a separate thread. */
3296 if (ci
->i_rdcache_revoking
!= ci
->i_rdcache_gen
) {
3297 queue_invalidate
= true;
3298 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
;
3304 cap
->issued
= cap
->implemented
= CEPH_CAP_PIN
;
3307 * auth mds of the inode changed. we received the cap export message,
3308 * but still haven't received the cap import message. handle_cap_export
3309 * updated the new auth MDS' cap.
3311 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
3312 * that was sent before the cap import message. So don't remove caps.
3314 if (ceph_seq_cmp(seq
, cap
->seq
) <= 0) {
3315 WARN_ON(cap
!= ci
->i_auth_cap
);
3316 WARN_ON(cap
->cap_id
!= le64_to_cpu(grant
->cap_id
));
3318 newcaps
|= cap
->issued
;
3321 /* side effects now are allowed */
3322 cap
->cap_gen
= session
->s_cap_gen
;
3325 __check_cap_issue(ci
, cap
, newcaps
);
3327 inode_set_max_iversion_raw(inode
, extra_info
->change_attr
);
3329 if ((newcaps
& CEPH_CAP_AUTH_SHARED
) &&
3330 (extra_info
->issued
& CEPH_CAP_AUTH_EXCL
) == 0) {
3331 inode
->i_mode
= le32_to_cpu(grant
->mode
);
3332 inode
->i_uid
= make_kuid(&init_user_ns
, le32_to_cpu(grant
->uid
));
3333 inode
->i_gid
= make_kgid(&init_user_ns
, le32_to_cpu(grant
->gid
));
3334 ci
->i_btime
= extra_info
->btime
;
3335 dout("%p mode 0%o uid.gid %d.%d\n", inode
, inode
->i_mode
,
3336 from_kuid(&init_user_ns
, inode
->i_uid
),
3337 from_kgid(&init_user_ns
, inode
->i_gid
));
3340 if ((newcaps
& CEPH_CAP_LINK_SHARED
) &&
3341 (extra_info
->issued
& CEPH_CAP_LINK_EXCL
) == 0) {
3342 set_nlink(inode
, le32_to_cpu(grant
->nlink
));
3343 if (inode
->i_nlink
== 0 &&
3344 (newcaps
& (CEPH_CAP_LINK_SHARED
| CEPH_CAP_LINK_EXCL
)))
3345 deleted_inode
= true;
3348 if ((extra_info
->issued
& CEPH_CAP_XATTR_EXCL
) == 0 &&
3350 int len
= le32_to_cpu(grant
->xattr_len
);
3351 u64 version
= le64_to_cpu(grant
->xattr_version
);
3353 if (version
> ci
->i_xattrs
.version
) {
3354 dout(" got new xattrs v%llu on %p len %d\n",
3355 version
, inode
, len
);
3356 if (ci
->i_xattrs
.blob
)
3357 ceph_buffer_put(ci
->i_xattrs
.blob
);
3358 ci
->i_xattrs
.blob
= ceph_buffer_get(xattr_buf
);
3359 ci
->i_xattrs
.version
= version
;
3360 ceph_forget_all_cached_acls(inode
);
3361 ceph_security_invalidate_secctx(inode
);
3365 if (newcaps
& CEPH_CAP_ANY_RD
) {
3366 struct timespec64 mtime
, atime
, ctime
;
3367 /* ctime/mtime/atime? */
3368 ceph_decode_timespec64(&mtime
, &grant
->mtime
);
3369 ceph_decode_timespec64(&atime
, &grant
->atime
);
3370 ceph_decode_timespec64(&ctime
, &grant
->ctime
);
3371 ceph_fill_file_time(inode
, extra_info
->issued
,
3372 le32_to_cpu(grant
->time_warp_seq
),
3373 &ctime
, &mtime
, &atime
);
3376 if ((newcaps
& CEPH_CAP_FILE_SHARED
) && extra_info
->dirstat_valid
) {
3377 ci
->i_files
= extra_info
->nfiles
;
3378 ci
->i_subdirs
= extra_info
->nsubdirs
;
3381 if (newcaps
& (CEPH_CAP_ANY_FILE_RD
| CEPH_CAP_ANY_FILE_WR
)) {
3382 /* file layout may have changed */
3383 s64 old_pool
= ci
->i_layout
.pool_id
;
3384 struct ceph_string
*old_ns
;
3386 ceph_file_layout_from_legacy(&ci
->i_layout
, &grant
->layout
);
3387 old_ns
= rcu_dereference_protected(ci
->i_layout
.pool_ns
,
3388 lockdep_is_held(&ci
->i_ceph_lock
));
3389 rcu_assign_pointer(ci
->i_layout
.pool_ns
, extra_info
->pool_ns
);
3391 if (ci
->i_layout
.pool_id
!= old_pool
||
3392 extra_info
->pool_ns
!= old_ns
)
3393 ci
->i_ceph_flags
&= ~CEPH_I_POOL_PERM
;
3395 extra_info
->pool_ns
= old_ns
;
3397 /* size/truncate_seq? */
3398 queue_trunc
= ceph_fill_file_size(inode
, extra_info
->issued
,
3399 le32_to_cpu(grant
->truncate_seq
),
3400 le64_to_cpu(grant
->truncate_size
),
3404 if (ci
->i_auth_cap
== cap
&& (newcaps
& CEPH_CAP_ANY_FILE_WR
)) {
3405 if (max_size
!= ci
->i_max_size
) {
3406 dout("max_size %lld -> %llu\n",
3407 ci
->i_max_size
, max_size
);
3408 ci
->i_max_size
= max_size
;
3409 if (max_size
>= ci
->i_wanted_max_size
) {
3410 ci
->i_wanted_max_size
= 0; /* reset */
3411 ci
->i_requested_max_size
= 0;
3417 /* check cap bits */
3418 wanted
= __ceph_caps_wanted(ci
);
3419 used
= __ceph_caps_used(ci
);
3420 dirty
= __ceph_caps_dirty(ci
);
3421 dout(" my wanted = %s, used = %s, dirty %s\n",
3422 ceph_cap_string(wanted
),
3423 ceph_cap_string(used
),
3424 ceph_cap_string(dirty
));
3426 if ((was_stale
|| le32_to_cpu(grant
->op
) == CEPH_CAP_OP_IMPORT
) &&
3427 (wanted
& ~(cap
->mds_wanted
| newcaps
))) {
3429 * If mds is importing cap, prior cap messages that update
3430 * 'wanted' may get dropped by mds (migrate seq mismatch).
3432 * We don't send cap message to update 'wanted' if what we
3433 * want are already issued. If mds revokes caps, cap message
3434 * that releases caps also tells mds what we want. But if
3435 * caps got revoked by mds forcedly (session stale). We may
3436 * haven't told mds what we want.
3441 /* revocation, grant, or no-op? */
3442 if (cap
->issued
& ~newcaps
) {
3443 int revoking
= cap
->issued
& ~newcaps
;
3445 dout("revocation: %s -> %s (revoking %s)\n",
3446 ceph_cap_string(cap
->issued
),
3447 ceph_cap_string(newcaps
),
3448 ceph_cap_string(revoking
));
3449 if (S_ISREG(inode
->i_mode
) &&
3450 (revoking
& used
& CEPH_CAP_FILE_BUFFER
))
3451 writeback
= true; /* initiate writeback; will delay ack */
3452 else if (queue_invalidate
&&
3453 revoking
== CEPH_CAP_FILE_CACHE
&&
3454 (newcaps
& CEPH_CAP_FILE_LAZYIO
) == 0)
3455 ; /* do nothing yet, invalidation will be queued */
3456 else if (cap
== ci
->i_auth_cap
)
3457 check_caps
= 1; /* check auth cap only */
3459 check_caps
= 2; /* check all caps */
3460 cap
->issued
= newcaps
;
3461 cap
->implemented
|= newcaps
;
3462 } else if (cap
->issued
== newcaps
) {
3463 dout("caps unchanged: %s -> %s\n",
3464 ceph_cap_string(cap
->issued
), ceph_cap_string(newcaps
));
3466 dout("grant: %s -> %s\n", ceph_cap_string(cap
->issued
),
3467 ceph_cap_string(newcaps
));
3468 /* non-auth MDS is revoking the newly grant caps ? */
3469 if (cap
== ci
->i_auth_cap
&&
3470 __ceph_caps_revoking_other(ci
, cap
, newcaps
))
3473 cap
->issued
= newcaps
;
3474 cap
->implemented
|= newcaps
; /* add bits only, to
3475 * avoid stepping on a
3476 * pending revocation */
3479 BUG_ON(cap
->issued
& ~cap
->implemented
);
3481 if (extra_info
->inline_version
> 0 &&
3482 extra_info
->inline_version
>= ci
->i_inline_version
) {
3483 ci
->i_inline_version
= extra_info
->inline_version
;
3484 if (ci
->i_inline_version
!= CEPH_INLINE_NONE
&&
3485 (newcaps
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)))
3489 if (ci
->i_auth_cap
== cap
&&
3490 le32_to_cpu(grant
->op
) == CEPH_CAP_OP_IMPORT
) {
3491 if (newcaps
& ~extra_info
->issued
)
3494 if (ci
->i_requested_max_size
> max_size
||
3495 !(le32_to_cpu(grant
->wanted
) & CEPH_CAP_ANY_FILE_WR
)) {
3496 /* re-request max_size if necessary */
3497 ci
->i_requested_max_size
= 0;
3501 ceph_kick_flushing_inode_caps(session
, ci
);
3502 spin_unlock(&ci
->i_ceph_lock
);
3503 up_read(&session
->s_mdsc
->snap_rwsem
);
3505 spin_unlock(&ci
->i_ceph_lock
);
3509 ceph_fill_inline_data(inode
, NULL
, extra_info
->inline_data
,
3510 extra_info
->inline_len
);
3513 ceph_queue_vmtruncate(inode
);
3517 * queue inode for writeback: we can't actually call
3518 * filemap_write_and_wait, etc. from message handler
3521 ceph_queue_writeback(inode
);
3522 if (queue_invalidate
)
3523 ceph_queue_invalidate(inode
);
3525 invalidate_aliases(inode
);
3527 wake_up_all(&ci
->i_cap_wq
);
3529 if (check_caps
== 1)
3530 ceph_check_caps(ci
, CHECK_CAPS_AUTHONLY
| CHECK_CAPS_NOINVAL
,
3532 else if (check_caps
== 2)
3533 ceph_check_caps(ci
, CHECK_CAPS_NOINVAL
, session
);
3535 mutex_unlock(&session
->s_mutex
);
3539 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3540 * MDS has been safely committed.
3542 static void handle_cap_flush_ack(struct inode
*inode
, u64 flush_tid
,
3543 struct ceph_mds_caps
*m
,
3544 struct ceph_mds_session
*session
,
3545 struct ceph_cap
*cap
)
3546 __releases(ci
->i_ceph_lock
)
3548 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3549 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
3550 struct ceph_cap_flush
*cf
, *tmp_cf
;
3551 LIST_HEAD(to_remove
);
3552 unsigned seq
= le32_to_cpu(m
->seq
);
3553 int dirty
= le32_to_cpu(m
->dirty
);
3556 bool wake_ci
= false;
3557 bool wake_mdsc
= false;
3559 list_for_each_entry_safe(cf
, tmp_cf
, &ci
->i_cap_flush_list
, i_list
) {
3560 /* Is this the one that was flushed? */
3561 if (cf
->tid
== flush_tid
)
3564 /* Is this a capsnap? */
3568 if (cf
->tid
<= flush_tid
) {
3570 * An earlier or current tid. The FLUSH_ACK should
3571 * represent a superset of this flush's caps.
3573 wake_ci
|= __detach_cap_flush_from_ci(ci
, cf
);
3574 list_add_tail(&cf
->i_list
, &to_remove
);
3577 * This is a later one. Any caps in it are still dirty
3578 * so don't count them as cleaned.
3580 cleaned
&= ~cf
->caps
;
3586 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3587 " flushing %s -> %s\n",
3588 inode
, session
->s_mds
, seq
, ceph_cap_string(dirty
),
3589 ceph_cap_string(cleaned
), ceph_cap_string(ci
->i_flushing_caps
),
3590 ceph_cap_string(ci
->i_flushing_caps
& ~cleaned
));
3592 if (list_empty(&to_remove
) && !cleaned
)
3595 ci
->i_flushing_caps
&= ~cleaned
;
3597 spin_lock(&mdsc
->cap_dirty_lock
);
3599 list_for_each_entry(cf
, &to_remove
, i_list
)
3600 wake_mdsc
|= __detach_cap_flush_from_mdsc(mdsc
, cf
);
3602 if (ci
->i_flushing_caps
== 0) {
3603 if (list_empty(&ci
->i_cap_flush_list
)) {
3604 list_del_init(&ci
->i_flushing_item
);
3605 if (!list_empty(&session
->s_cap_flushing
)) {
3606 dout(" mds%d still flushing cap on %p\n",
3608 &list_first_entry(&session
->s_cap_flushing
,
3609 struct ceph_inode_info
,
3610 i_flushing_item
)->vfs_inode
);
3613 mdsc
->num_cap_flushing
--;
3614 dout(" inode %p now !flushing\n", inode
);
3616 if (ci
->i_dirty_caps
== 0) {
3617 dout(" inode %p now clean\n", inode
);
3618 BUG_ON(!list_empty(&ci
->i_dirty_item
));
3620 if (ci
->i_wr_ref
== 0 &&
3621 ci
->i_wrbuffer_ref_head
== 0) {
3622 BUG_ON(!ci
->i_head_snapc
);
3623 ceph_put_snap_context(ci
->i_head_snapc
);
3624 ci
->i_head_snapc
= NULL
;
3627 BUG_ON(list_empty(&ci
->i_dirty_item
));
3630 spin_unlock(&mdsc
->cap_dirty_lock
);
3633 spin_unlock(&ci
->i_ceph_lock
);
3635 while (!list_empty(&to_remove
)) {
3636 cf
= list_first_entry(&to_remove
,
3637 struct ceph_cap_flush
, i_list
);
3638 list_del(&cf
->i_list
);
3639 ceph_free_cap_flush(cf
);
3643 wake_up_all(&ci
->i_cap_wq
);
3645 wake_up_all(&mdsc
->cap_flushing_wq
);
3651 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3652 * throw away our cap_snap.
3654 * Caller hold s_mutex.
3656 static void handle_cap_flushsnap_ack(struct inode
*inode
, u64 flush_tid
,
3657 struct ceph_mds_caps
*m
,
3658 struct ceph_mds_session
*session
)
3660 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3661 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
3662 u64 follows
= le64_to_cpu(m
->snap_follows
);
3663 struct ceph_cap_snap
*capsnap
;
3664 bool flushed
= false;
3665 bool wake_ci
= false;
3666 bool wake_mdsc
= false;
3668 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3669 inode
, ci
, session
->s_mds
, follows
);
3671 spin_lock(&ci
->i_ceph_lock
);
3672 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
3673 if (capsnap
->follows
== follows
) {
3674 if (capsnap
->cap_flush
.tid
!= flush_tid
) {
3675 dout(" cap_snap %p follows %lld tid %lld !="
3676 " %lld\n", capsnap
, follows
,
3677 flush_tid
, capsnap
->cap_flush
.tid
);
3683 dout(" skipping cap_snap %p follows %lld\n",
3684 capsnap
, capsnap
->follows
);
3688 WARN_ON(capsnap
->dirty_pages
|| capsnap
->writing
);
3689 dout(" removing %p cap_snap %p follows %lld\n",
3690 inode
, capsnap
, follows
);
3691 list_del(&capsnap
->ci_item
);
3692 wake_ci
|= __detach_cap_flush_from_ci(ci
, &capsnap
->cap_flush
);
3694 spin_lock(&mdsc
->cap_dirty_lock
);
3696 if (list_empty(&ci
->i_cap_flush_list
))
3697 list_del_init(&ci
->i_flushing_item
);
3699 wake_mdsc
|= __detach_cap_flush_from_mdsc(mdsc
,
3700 &capsnap
->cap_flush
);
3701 spin_unlock(&mdsc
->cap_dirty_lock
);
3703 spin_unlock(&ci
->i_ceph_lock
);
3705 ceph_put_snap_context(capsnap
->context
);
3706 ceph_put_cap_snap(capsnap
);
3708 wake_up_all(&ci
->i_cap_wq
);
3710 wake_up_all(&mdsc
->cap_flushing_wq
);
3716 * Handle TRUNC from MDS, indicating file truncation.
3718 * caller hold s_mutex.
3720 static bool handle_cap_trunc(struct inode
*inode
,
3721 struct ceph_mds_caps
*trunc
,
3722 struct ceph_mds_session
*session
)
3724 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3725 int mds
= session
->s_mds
;
3726 int seq
= le32_to_cpu(trunc
->seq
);
3727 u32 truncate_seq
= le32_to_cpu(trunc
->truncate_seq
);
3728 u64 truncate_size
= le64_to_cpu(trunc
->truncate_size
);
3729 u64 size
= le64_to_cpu(trunc
->size
);
3730 int implemented
= 0;
3731 int dirty
= __ceph_caps_dirty(ci
);
3732 int issued
= __ceph_caps_issued(ceph_inode(inode
), &implemented
);
3733 bool queue_trunc
= false;
3735 lockdep_assert_held(&ci
->i_ceph_lock
);
3737 issued
|= implemented
| dirty
;
3739 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3740 inode
, mds
, seq
, truncate_size
, truncate_seq
);
3741 queue_trunc
= ceph_fill_file_size(inode
, issued
,
3742 truncate_seq
, truncate_size
, size
);
3747 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3748 * different one. If we are the most recent migration we've seen (as
3749 * indicated by mseq), make note of the migrating cap bits for the
3750 * duration (until we see the corresponding IMPORT).
3752 * caller holds s_mutex
3754 static void handle_cap_export(struct inode
*inode
, struct ceph_mds_caps
*ex
,
3755 struct ceph_mds_cap_peer
*ph
,
3756 struct ceph_mds_session
*session
)
3758 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
3759 struct ceph_mds_session
*tsession
= NULL
;
3760 struct ceph_cap
*cap
, *tcap
, *new_cap
= NULL
;
3761 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3763 unsigned mseq
= le32_to_cpu(ex
->migrate_seq
);
3764 unsigned t_seq
, t_mseq
;
3766 int mds
= session
->s_mds
;
3769 t_cap_id
= le64_to_cpu(ph
->cap_id
);
3770 t_seq
= le32_to_cpu(ph
->seq
);
3771 t_mseq
= le32_to_cpu(ph
->mseq
);
3772 target
= le32_to_cpu(ph
->mds
);
3774 t_cap_id
= t_seq
= t_mseq
= 0;
3778 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3779 inode
, ci
, mds
, mseq
, target
);
3781 spin_lock(&ci
->i_ceph_lock
);
3782 cap
= __get_cap_for_mds(ci
, mds
);
3783 if (!cap
|| cap
->cap_id
!= le64_to_cpu(ex
->cap_id
))
3787 __ceph_remove_cap(cap
, false);
3792 * now we know we haven't received the cap import message yet
3793 * because the exported cap still exist.
3796 issued
= cap
->issued
;
3797 if (issued
!= cap
->implemented
)
3798 pr_err_ratelimited("handle_cap_export: issued != implemented: "
3799 "ino (%llx.%llx) mds%d seq %d mseq %d "
3800 "issued %s implemented %s\n",
3801 ceph_vinop(inode
), mds
, cap
->seq
, cap
->mseq
,
3802 ceph_cap_string(issued
),
3803 ceph_cap_string(cap
->implemented
));
3806 tcap
= __get_cap_for_mds(ci
, target
);
3808 /* already have caps from the target */
3809 if (tcap
->cap_id
== t_cap_id
&&
3810 ceph_seq_cmp(tcap
->seq
, t_seq
) < 0) {
3811 dout(" updating import cap %p mds%d\n", tcap
, target
);
3812 tcap
->cap_id
= t_cap_id
;
3813 tcap
->seq
= t_seq
- 1;
3814 tcap
->issue_seq
= t_seq
- 1;
3815 tcap
->issued
|= issued
;
3816 tcap
->implemented
|= issued
;
3817 if (cap
== ci
->i_auth_cap
) {
3818 ci
->i_auth_cap
= tcap
;
3819 change_auth_cap_ses(ci
, tcap
->session
);
3822 __ceph_remove_cap(cap
, false);
3824 } else if (tsession
) {
3825 /* add placeholder for the export tagert */
3826 int flag
= (cap
== ci
->i_auth_cap
) ? CEPH_CAP_FLAG_AUTH
: 0;
3828 ceph_add_cap(inode
, tsession
, t_cap_id
, issued
, 0,
3829 t_seq
- 1, t_mseq
, (u64
)-1, flag
, &new_cap
);
3831 if (!list_empty(&ci
->i_cap_flush_list
) &&
3832 ci
->i_auth_cap
== tcap
) {
3833 spin_lock(&mdsc
->cap_dirty_lock
);
3834 list_move_tail(&ci
->i_flushing_item
,
3835 &tcap
->session
->s_cap_flushing
);
3836 spin_unlock(&mdsc
->cap_dirty_lock
);
3839 __ceph_remove_cap(cap
, false);
3843 spin_unlock(&ci
->i_ceph_lock
);
3844 mutex_unlock(&session
->s_mutex
);
3846 /* open target session */
3847 tsession
= ceph_mdsc_open_export_target_session(mdsc
, target
);
3848 if (!IS_ERR(tsession
)) {
3850 mutex_lock(&session
->s_mutex
);
3851 mutex_lock_nested(&tsession
->s_mutex
,
3852 SINGLE_DEPTH_NESTING
);
3854 mutex_lock(&tsession
->s_mutex
);
3855 mutex_lock_nested(&session
->s_mutex
,
3856 SINGLE_DEPTH_NESTING
);
3858 new_cap
= ceph_get_cap(mdsc
, NULL
);
3863 mutex_lock(&session
->s_mutex
);
3868 spin_unlock(&ci
->i_ceph_lock
);
3869 mutex_unlock(&session
->s_mutex
);
3871 mutex_unlock(&tsession
->s_mutex
);
3872 ceph_put_mds_session(tsession
);
3875 ceph_put_cap(mdsc
, new_cap
);
3879 * Handle cap IMPORT.
3881 * caller holds s_mutex. acquires i_ceph_lock
3883 static void handle_cap_import(struct ceph_mds_client
*mdsc
,
3884 struct inode
*inode
, struct ceph_mds_caps
*im
,
3885 struct ceph_mds_cap_peer
*ph
,
3886 struct ceph_mds_session
*session
,
3887 struct ceph_cap
**target_cap
, int *old_issued
)
3889 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3890 struct ceph_cap
*cap
, *ocap
, *new_cap
= NULL
;
3891 int mds
= session
->s_mds
;
3893 unsigned caps
= le32_to_cpu(im
->caps
);
3894 unsigned wanted
= le32_to_cpu(im
->wanted
);
3895 unsigned seq
= le32_to_cpu(im
->seq
);
3896 unsigned mseq
= le32_to_cpu(im
->migrate_seq
);
3897 u64 realmino
= le64_to_cpu(im
->realm
);
3898 u64 cap_id
= le64_to_cpu(im
->cap_id
);
3903 p_cap_id
= le64_to_cpu(ph
->cap_id
);
3904 peer
= le32_to_cpu(ph
->mds
);
3910 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3911 inode
, ci
, mds
, mseq
, peer
);
3913 cap
= __get_cap_for_mds(ci
, mds
);
3916 spin_unlock(&ci
->i_ceph_lock
);
3917 new_cap
= ceph_get_cap(mdsc
, NULL
);
3918 spin_lock(&ci
->i_ceph_lock
);
3924 ceph_put_cap(mdsc
, new_cap
);
3929 __ceph_caps_issued(ci
, &issued
);
3930 issued
|= __ceph_caps_dirty(ci
);
3932 ceph_add_cap(inode
, session
, cap_id
, caps
, wanted
, seq
, mseq
,
3933 realmino
, CEPH_CAP_FLAG_AUTH
, &new_cap
);
3935 ocap
= peer
>= 0 ? __get_cap_for_mds(ci
, peer
) : NULL
;
3936 if (ocap
&& ocap
->cap_id
== p_cap_id
) {
3937 dout(" remove export cap %p mds%d flags %d\n",
3938 ocap
, peer
, ph
->flags
);
3939 if ((ph
->flags
& CEPH_CAP_FLAG_AUTH
) &&
3940 (ocap
->seq
!= le32_to_cpu(ph
->seq
) ||
3941 ocap
->mseq
!= le32_to_cpu(ph
->mseq
))) {
3942 pr_err_ratelimited("handle_cap_import: "
3943 "mismatched seq/mseq: ino (%llx.%llx) "
3944 "mds%d seq %d mseq %d importer mds%d "
3945 "has peer seq %d mseq %d\n",
3946 ceph_vinop(inode
), peer
, ocap
->seq
,
3947 ocap
->mseq
, mds
, le32_to_cpu(ph
->seq
),
3948 le32_to_cpu(ph
->mseq
));
3950 __ceph_remove_cap(ocap
, (ph
->flags
& CEPH_CAP_FLAG_RELEASE
));
3953 *old_issued
= issued
;
3958 * Handle a caps message from the MDS.
3960 * Identify the appropriate session, inode, and call the right handler
3961 * based on the cap op.
3963 void ceph_handle_caps(struct ceph_mds_session
*session
,
3964 struct ceph_msg
*msg
)
3966 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
3967 struct inode
*inode
;
3968 struct ceph_inode_info
*ci
;
3969 struct ceph_cap
*cap
;
3970 struct ceph_mds_caps
*h
;
3971 struct ceph_mds_cap_peer
*peer
= NULL
;
3972 struct ceph_snap_realm
*realm
= NULL
;
3974 int msg_version
= le16_to_cpu(msg
->hdr
.version
);
3976 struct ceph_vino vino
;
3978 size_t snaptrace_len
;
3980 struct cap_extra_info extra_info
= {};
3983 dout("handle_caps from mds%d\n", session
->s_mds
);
3986 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
3987 if (msg
->front
.iov_len
< sizeof(*h
))
3989 h
= msg
->front
.iov_base
;
3990 op
= le32_to_cpu(h
->op
);
3991 vino
.ino
= le64_to_cpu(h
->ino
);
3992 vino
.snap
= CEPH_NOSNAP
;
3993 seq
= le32_to_cpu(h
->seq
);
3994 mseq
= le32_to_cpu(h
->migrate_seq
);
3997 snaptrace_len
= le32_to_cpu(h
->snap_trace_len
);
3998 p
= snaptrace
+ snaptrace_len
;
4000 if (msg_version
>= 2) {
4002 ceph_decode_32_safe(&p
, end
, flock_len
, bad
);
4003 if (p
+ flock_len
> end
)
4008 if (msg_version
>= 3) {
4009 if (op
== CEPH_CAP_OP_IMPORT
) {
4010 if (p
+ sizeof(*peer
) > end
)
4014 } else if (op
== CEPH_CAP_OP_EXPORT
) {
4015 /* recorded in unused fields */
4016 peer
= (void *)&h
->size
;
4020 if (msg_version
>= 4) {
4021 ceph_decode_64_safe(&p
, end
, extra_info
.inline_version
, bad
);
4022 ceph_decode_32_safe(&p
, end
, extra_info
.inline_len
, bad
);
4023 if (p
+ extra_info
.inline_len
> end
)
4025 extra_info
.inline_data
= p
;
4026 p
+= extra_info
.inline_len
;
4029 if (msg_version
>= 5) {
4030 struct ceph_osd_client
*osdc
= &mdsc
->fsc
->client
->osdc
;
4033 ceph_decode_32_safe(&p
, end
, epoch_barrier
, bad
);
4034 ceph_osdc_update_epoch_barrier(osdc
, epoch_barrier
);
4037 if (msg_version
>= 8) {
4041 ceph_decode_skip_64(&p
, end
, bad
); // flush_tid
4043 ceph_decode_skip_32(&p
, end
, bad
); // caller_uid
4044 ceph_decode_skip_32(&p
, end
, bad
); // caller_gid
4046 ceph_decode_32_safe(&p
, end
, pool_ns_len
, bad
);
4047 if (pool_ns_len
> 0) {
4048 ceph_decode_need(&p
, end
, pool_ns_len
, bad
);
4049 extra_info
.pool_ns
=
4050 ceph_find_or_create_string(p
, pool_ns_len
);
4055 if (msg_version
>= 9) {
4056 struct ceph_timespec
*btime
;
4058 if (p
+ sizeof(*btime
) > end
)
4061 ceph_decode_timespec64(&extra_info
.btime
, btime
);
4062 p
+= sizeof(*btime
);
4063 ceph_decode_64_safe(&p
, end
, extra_info
.change_attr
, bad
);
4066 if (msg_version
>= 11) {
4068 ceph_decode_skip_32(&p
, end
, bad
); // flags
4070 extra_info
.dirstat_valid
= true;
4071 ceph_decode_64_safe(&p
, end
, extra_info
.nfiles
, bad
);
4072 ceph_decode_64_safe(&p
, end
, extra_info
.nsubdirs
, bad
);
4076 inode
= ceph_find_inode(mdsc
->fsc
->sb
, vino
);
4077 ci
= ceph_inode(inode
);
4078 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op
), vino
.ino
,
4081 mutex_lock(&session
->s_mutex
);
4082 inc_session_sequence(session
);
4083 dout(" mds%d seq %lld cap seq %u\n", session
->s_mds
, session
->s_seq
,
4087 dout(" i don't have ino %llx\n", vino
.ino
);
4089 if (op
== CEPH_CAP_OP_IMPORT
) {
4090 cap
= ceph_get_cap(mdsc
, NULL
);
4091 cap
->cap_ino
= vino
.ino
;
4092 cap
->queue_release
= 1;
4093 cap
->cap_id
= le64_to_cpu(h
->cap_id
);
4096 cap
->issue_seq
= seq
;
4097 spin_lock(&session
->s_cap_lock
);
4098 __ceph_queue_cap_release(session
, cap
);
4099 spin_unlock(&session
->s_cap_lock
);
4101 goto flush_cap_releases
;
4104 /* these will work even if we don't have a cap yet */
4106 case CEPH_CAP_OP_FLUSHSNAP_ACK
:
4107 handle_cap_flushsnap_ack(inode
, le64_to_cpu(msg
->hdr
.tid
),
4111 case CEPH_CAP_OP_EXPORT
:
4112 handle_cap_export(inode
, h
, peer
, session
);
4115 case CEPH_CAP_OP_IMPORT
:
4117 if (snaptrace_len
) {
4118 down_write(&mdsc
->snap_rwsem
);
4119 ceph_update_snap_trace(mdsc
, snaptrace
,
4120 snaptrace
+ snaptrace_len
,
4122 downgrade_write(&mdsc
->snap_rwsem
);
4124 down_read(&mdsc
->snap_rwsem
);
4126 spin_lock(&ci
->i_ceph_lock
);
4127 handle_cap_import(mdsc
, inode
, h
, peer
, session
,
4128 &cap
, &extra_info
.issued
);
4129 handle_cap_grant(inode
, session
, cap
,
4130 h
, msg
->middle
, &extra_info
);
4132 ceph_put_snap_realm(mdsc
, realm
);
4136 /* the rest require a cap */
4137 spin_lock(&ci
->i_ceph_lock
);
4138 cap
= __get_cap_for_mds(ceph_inode(inode
), session
->s_mds
);
4140 dout(" no cap on %p ino %llx.%llx from mds%d\n",
4141 inode
, ceph_ino(inode
), ceph_snap(inode
),
4143 spin_unlock(&ci
->i_ceph_lock
);
4144 goto flush_cap_releases
;
4147 /* note that each of these drops i_ceph_lock for us */
4149 case CEPH_CAP_OP_REVOKE
:
4150 case CEPH_CAP_OP_GRANT
:
4151 __ceph_caps_issued(ci
, &extra_info
.issued
);
4152 extra_info
.issued
|= __ceph_caps_dirty(ci
);
4153 handle_cap_grant(inode
, session
, cap
,
4154 h
, msg
->middle
, &extra_info
);
4157 case CEPH_CAP_OP_FLUSH_ACK
:
4158 handle_cap_flush_ack(inode
, le64_to_cpu(msg
->hdr
.tid
),
4162 case CEPH_CAP_OP_TRUNC
:
4163 queue_trunc
= handle_cap_trunc(inode
, h
, session
);
4164 spin_unlock(&ci
->i_ceph_lock
);
4166 ceph_queue_vmtruncate(inode
);
4170 spin_unlock(&ci
->i_ceph_lock
);
4171 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op
,
4172 ceph_cap_op_name(op
));
4176 mutex_unlock(&session
->s_mutex
);
4178 ceph_put_string(extra_info
.pool_ns
);
4179 /* avoid calling iput_final() in mds dispatch threads */
4180 ceph_async_iput(inode
);
4185 * send any cap release message to try to move things
4186 * along for the mds (who clearly thinks we still have this
4189 ceph_flush_cap_releases(mdsc
, session
);
4193 pr_err("ceph_handle_caps: corrupt message\n");
4199 * Delayed work handler to process end of delayed cap release LRU list.
4201 void ceph_check_delayed_caps(struct ceph_mds_client
*mdsc
)
4203 struct inode
*inode
;
4204 struct ceph_inode_info
*ci
;
4206 dout("check_delayed_caps\n");
4207 spin_lock(&mdsc
->cap_delay_lock
);
4208 while (!list_empty(&mdsc
->cap_delay_list
)) {
4209 ci
= list_first_entry(&mdsc
->cap_delay_list
,
4210 struct ceph_inode_info
,
4212 if ((ci
->i_ceph_flags
& CEPH_I_FLUSH
) == 0 &&
4213 time_before(jiffies
, ci
->i_hold_caps_max
))
4215 list_del_init(&ci
->i_cap_delay_list
);
4217 inode
= igrab(&ci
->vfs_inode
);
4219 spin_unlock(&mdsc
->cap_delay_lock
);
4220 dout("check_delayed_caps on %p\n", inode
);
4221 ceph_check_caps(ci
, 0, NULL
);
4222 /* avoid calling iput_final() in tick thread */
4223 ceph_async_iput(inode
);
4224 spin_lock(&mdsc
->cap_delay_lock
);
4227 spin_unlock(&mdsc
->cap_delay_lock
);
4231 * Flush all dirty caps to the mds
4233 static void flush_dirty_session_caps(struct ceph_mds_session
*s
)
4235 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4236 struct ceph_inode_info
*ci
;
4237 struct inode
*inode
;
4239 dout("flush_dirty_caps\n");
4240 spin_lock(&mdsc
->cap_dirty_lock
);
4241 while (!list_empty(&s
->s_cap_dirty
)) {
4242 ci
= list_first_entry(&s
->s_cap_dirty
, struct ceph_inode_info
,
4244 inode
= &ci
->vfs_inode
;
4246 dout("flush_dirty_caps %p\n", inode
);
4247 spin_unlock(&mdsc
->cap_dirty_lock
);
4248 ceph_check_caps(ci
, CHECK_CAPS_FLUSH
, NULL
);
4250 spin_lock(&mdsc
->cap_dirty_lock
);
4252 spin_unlock(&mdsc
->cap_dirty_lock
);
4253 dout("flush_dirty_caps done\n");
4256 static void iterate_sessions(struct ceph_mds_client
*mdsc
,
4257 void (*cb
)(struct ceph_mds_session
*))
4261 mutex_lock(&mdsc
->mutex
);
4262 for (mds
= 0; mds
< mdsc
->max_sessions
; ++mds
) {
4263 struct ceph_mds_session
*s
;
4265 if (!mdsc
->sessions
[mds
])
4268 s
= ceph_get_mds_session(mdsc
->sessions
[mds
]);
4272 mutex_unlock(&mdsc
->mutex
);
4274 ceph_put_mds_session(s
);
4275 mutex_lock(&mdsc
->mutex
);
4277 mutex_unlock(&mdsc
->mutex
);
4280 void ceph_flush_dirty_caps(struct ceph_mds_client
*mdsc
)
4282 iterate_sessions(mdsc
, flush_dirty_session_caps
);
4285 void __ceph_touch_fmode(struct ceph_inode_info
*ci
,
4286 struct ceph_mds_client
*mdsc
, int fmode
)
4288 unsigned long now
= jiffies
;
4289 if (fmode
& CEPH_FILE_MODE_RD
)
4290 ci
->i_last_rd
= now
;
4291 if (fmode
& CEPH_FILE_MODE_WR
)
4292 ci
->i_last_wr
= now
;
4293 /* queue periodic check */
4295 __ceph_is_any_real_caps(ci
) &&
4296 list_empty(&ci
->i_cap_delay_list
))
4297 __cap_delay_requeue(mdsc
, ci
);
4300 void ceph_get_fmode(struct ceph_inode_info
*ci
, int fmode
, int count
)
4302 struct ceph_mds_client
*mdsc
= ceph_sb_to_mdsc(ci
->vfs_inode
.i_sb
);
4303 int bits
= (fmode
<< 1) | 1;
4304 bool is_opened
= false;
4308 atomic64_inc(&mdsc
->metric
.opened_files
);
4310 spin_lock(&ci
->i_ceph_lock
);
4311 for (i
= 0; i
< CEPH_FILE_MODE_BITS
; i
++) {
4312 if (bits
& (1 << i
))
4313 ci
->i_nr_by_mode
[i
] += count
;
4316 * If any of the mode ref is larger than 1,
4317 * that means it has been already opened by
4318 * others. Just skip checking the PIN ref.
4320 if (i
&& ci
->i_nr_by_mode
[i
] > 1)
4325 percpu_counter_inc(&mdsc
->metric
.opened_inodes
);
4326 spin_unlock(&ci
->i_ceph_lock
);
4330 * Drop open file reference. If we were the last open file,
4331 * we may need to release capabilities to the MDS (or schedule
4332 * their delayed release).
4334 void ceph_put_fmode(struct ceph_inode_info
*ci
, int fmode
, int count
)
4336 struct ceph_mds_client
*mdsc
= ceph_sb_to_mdsc(ci
->vfs_inode
.i_sb
);
4337 int bits
= (fmode
<< 1) | 1;
4338 bool is_closed
= true;
4342 atomic64_dec(&mdsc
->metric
.opened_files
);
4344 spin_lock(&ci
->i_ceph_lock
);
4345 for (i
= 0; i
< CEPH_FILE_MODE_BITS
; i
++) {
4346 if (bits
& (1 << i
)) {
4347 BUG_ON(ci
->i_nr_by_mode
[i
] < count
);
4348 ci
->i_nr_by_mode
[i
] -= count
;
4352 * If any of the mode ref is not 0 after
4353 * decreased, that means it is still opened
4354 * by others. Just skip checking the PIN ref.
4356 if (i
&& ci
->i_nr_by_mode
[i
])
4361 percpu_counter_dec(&mdsc
->metric
.opened_inodes
);
4362 spin_unlock(&ci
->i_ceph_lock
);
4366 * For a soon-to-be unlinked file, drop the LINK caps. If it
4367 * looks like the link count will hit 0, drop any other caps (other
4368 * than PIN) we don't specifically want (due to the file still being
4371 int ceph_drop_caps_for_unlink(struct inode
*inode
)
4373 struct ceph_inode_info
*ci
= ceph_inode(inode
);
4374 int drop
= CEPH_CAP_LINK_SHARED
| CEPH_CAP_LINK_EXCL
;
4376 spin_lock(&ci
->i_ceph_lock
);
4377 if (inode
->i_nlink
== 1) {
4378 drop
|= ~(__ceph_caps_wanted(ci
) | CEPH_CAP_PIN
);
4380 if (__ceph_caps_dirty(ci
)) {
4381 struct ceph_mds_client
*mdsc
=
4382 ceph_inode_to_client(inode
)->mdsc
;
4383 __cap_delay_requeue_front(mdsc
, ci
);
4386 spin_unlock(&ci
->i_ceph_lock
);
4391 * Helpers for embedding cap and dentry lease releases into mds
4394 * @force is used by dentry_release (below) to force inclusion of a
4395 * record for the directory inode, even when there aren't any caps to
4398 int ceph_encode_inode_release(void **p
, struct inode
*inode
,
4399 int mds
, int drop
, int unless
, int force
)
4401 struct ceph_inode_info
*ci
= ceph_inode(inode
);
4402 struct ceph_cap
*cap
;
4403 struct ceph_mds_request_release
*rel
= *p
;
4407 spin_lock(&ci
->i_ceph_lock
);
4408 used
= __ceph_caps_used(ci
);
4409 dirty
= __ceph_caps_dirty(ci
);
4411 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
4412 inode
, mds
, ceph_cap_string(used
|dirty
), ceph_cap_string(drop
),
4413 ceph_cap_string(unless
));
4415 /* only drop unused, clean caps */
4416 drop
&= ~(used
| dirty
);
4418 cap
= __get_cap_for_mds(ci
, mds
);
4419 if (cap
&& __cap_is_valid(cap
)) {
4420 unless
&= cap
->issued
;
4422 if (unless
& CEPH_CAP_AUTH_EXCL
)
4423 drop
&= ~CEPH_CAP_AUTH_SHARED
;
4424 if (unless
& CEPH_CAP_LINK_EXCL
)
4425 drop
&= ~CEPH_CAP_LINK_SHARED
;
4426 if (unless
& CEPH_CAP_XATTR_EXCL
)
4427 drop
&= ~CEPH_CAP_XATTR_SHARED
;
4428 if (unless
& CEPH_CAP_FILE_EXCL
)
4429 drop
&= ~CEPH_CAP_FILE_SHARED
;
4432 if (force
|| (cap
->issued
& drop
)) {
4433 if (cap
->issued
& drop
) {
4434 int wanted
= __ceph_caps_wanted(ci
);
4435 dout("encode_inode_release %p cap %p "
4436 "%s -> %s, wanted %s -> %s\n", inode
, cap
,
4437 ceph_cap_string(cap
->issued
),
4438 ceph_cap_string(cap
->issued
& ~drop
),
4439 ceph_cap_string(cap
->mds_wanted
),
4440 ceph_cap_string(wanted
));
4442 cap
->issued
&= ~drop
;
4443 cap
->implemented
&= ~drop
;
4444 cap
->mds_wanted
= wanted
;
4445 if (cap
== ci
->i_auth_cap
&&
4446 !(wanted
& CEPH_CAP_ANY_FILE_WR
))
4447 ci
->i_requested_max_size
= 0;
4449 dout("encode_inode_release %p cap %p %s"
4450 " (force)\n", inode
, cap
,
4451 ceph_cap_string(cap
->issued
));
4454 rel
->ino
= cpu_to_le64(ceph_ino(inode
));
4455 rel
->cap_id
= cpu_to_le64(cap
->cap_id
);
4456 rel
->seq
= cpu_to_le32(cap
->seq
);
4457 rel
->issue_seq
= cpu_to_le32(cap
->issue_seq
);
4458 rel
->mseq
= cpu_to_le32(cap
->mseq
);
4459 rel
->caps
= cpu_to_le32(cap
->implemented
);
4460 rel
->wanted
= cpu_to_le32(cap
->mds_wanted
);
4466 dout("encode_inode_release %p cap %p %s (noop)\n",
4467 inode
, cap
, ceph_cap_string(cap
->issued
));
4470 spin_unlock(&ci
->i_ceph_lock
);
4474 int ceph_encode_dentry_release(void **p
, struct dentry
*dentry
,
4476 int mds
, int drop
, int unless
)
4478 struct dentry
*parent
= NULL
;
4479 struct ceph_mds_request_release
*rel
= *p
;
4480 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
4485 * force an record for the directory caps if we have a dentry lease.
4486 * this is racy (can't take i_ceph_lock and d_lock together), but it
4487 * doesn't have to be perfect; the mds will revoke anything we don't
4490 spin_lock(&dentry
->d_lock
);
4491 if (di
->lease_session
&& di
->lease_session
->s_mds
== mds
)
4494 parent
= dget(dentry
->d_parent
);
4495 dir
= d_inode(parent
);
4497 spin_unlock(&dentry
->d_lock
);
4499 ret
= ceph_encode_inode_release(p
, dir
, mds
, drop
, unless
, force
);
4502 spin_lock(&dentry
->d_lock
);
4503 if (ret
&& di
->lease_session
&& di
->lease_session
->s_mds
== mds
) {
4504 dout("encode_dentry_release %p mds%d seq %d\n",
4505 dentry
, mds
, (int)di
->lease_seq
);
4506 rel
->dname_len
= cpu_to_le32(dentry
->d_name
.len
);
4507 memcpy(*p
, dentry
->d_name
.name
, dentry
->d_name
.len
);
4508 *p
+= dentry
->d_name
.len
;
4509 rel
->dname_seq
= cpu_to_le32(di
->lease_seq
);
4510 __ceph_mdsc_drop_dentry_lease(dentry
);
4512 spin_unlock(&dentry
->d_lock
);