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>
13 #include "mds_client.h"
15 #include <linux/ceph/decode.h>
16 #include <linux/ceph/messenger.h>
19 * Capability management
21 * The Ceph metadata servers control client access to inode metadata
22 * and file data by issuing capabilities, granting clients permission
23 * to read and/or write both inode field and file data to OSDs
24 * (storage nodes). Each capability consists of a set of bits
25 * indicating which operations are allowed.
27 * If the client holds a *_SHARED cap, the client has a coherent value
28 * that can be safely read from the cached inode.
30 * In the case of a *_EXCL (exclusive) or FILE_WR capabilities, the
31 * client is allowed to change inode attributes (e.g., file size,
32 * mtime), note its dirty state in the ceph_cap, and asynchronously
33 * flush that metadata change to the MDS.
35 * In the event of a conflicting operation (perhaps by another
36 * client), the MDS will revoke the conflicting client capabilities.
38 * In order for a client to cache an inode, it must hold a capability
39 * with at least one MDS server. When inodes are released, release
40 * notifications are batched and periodically sent en masse to the MDS
41 * cluster to release server state.
44 static u64
__get_oldest_flush_tid(struct ceph_mds_client
*mdsc
);
45 static void __kick_flushing_caps(struct ceph_mds_client
*mdsc
,
46 struct ceph_mds_session
*session
,
47 struct ceph_inode_info
*ci
,
48 u64 oldest_flush_tid
);
51 * Generate readable cap strings for debugging output.
53 #define MAX_CAP_STR 20
54 static char cap_str
[MAX_CAP_STR
][40];
55 static DEFINE_SPINLOCK(cap_str_lock
);
56 static int last_cap_str
;
58 static char *gcap_string(char *s
, int c
)
60 if (c
& CEPH_CAP_GSHARED
)
62 if (c
& CEPH_CAP_GEXCL
)
64 if (c
& CEPH_CAP_GCACHE
)
70 if (c
& CEPH_CAP_GBUFFER
)
72 if (c
& CEPH_CAP_GLAZYIO
)
77 const char *ceph_cap_string(int caps
)
83 spin_lock(&cap_str_lock
);
85 if (last_cap_str
== MAX_CAP_STR
)
87 spin_unlock(&cap_str_lock
);
91 if (caps
& CEPH_CAP_PIN
)
94 c
= (caps
>> CEPH_CAP_SAUTH
) & 3;
97 s
= gcap_string(s
, c
);
100 c
= (caps
>> CEPH_CAP_SLINK
) & 3;
103 s
= gcap_string(s
, c
);
106 c
= (caps
>> CEPH_CAP_SXATTR
) & 3;
109 s
= gcap_string(s
, c
);
112 c
= caps
>> CEPH_CAP_SFILE
;
115 s
= gcap_string(s
, c
);
124 void ceph_caps_init(struct ceph_mds_client
*mdsc
)
126 INIT_LIST_HEAD(&mdsc
->caps_list
);
127 spin_lock_init(&mdsc
->caps_list_lock
);
130 void ceph_caps_finalize(struct ceph_mds_client
*mdsc
)
132 struct ceph_cap
*cap
;
134 spin_lock(&mdsc
->caps_list_lock
);
135 while (!list_empty(&mdsc
->caps_list
)) {
136 cap
= list_first_entry(&mdsc
->caps_list
,
137 struct ceph_cap
, caps_item
);
138 list_del(&cap
->caps_item
);
139 kmem_cache_free(ceph_cap_cachep
, cap
);
141 mdsc
->caps_total_count
= 0;
142 mdsc
->caps_avail_count
= 0;
143 mdsc
->caps_use_count
= 0;
144 mdsc
->caps_reserve_count
= 0;
145 mdsc
->caps_min_count
= 0;
146 spin_unlock(&mdsc
->caps_list_lock
);
149 void ceph_adjust_min_caps(struct ceph_mds_client
*mdsc
, int delta
)
151 spin_lock(&mdsc
->caps_list_lock
);
152 mdsc
->caps_min_count
+= delta
;
153 BUG_ON(mdsc
->caps_min_count
< 0);
154 spin_unlock(&mdsc
->caps_list_lock
);
157 void ceph_reserve_caps(struct ceph_mds_client
*mdsc
,
158 struct ceph_cap_reservation
*ctx
, int need
)
161 struct ceph_cap
*cap
;
166 dout("reserve caps ctx=%p need=%d\n", ctx
, need
);
168 /* first reserve any caps that are already allocated */
169 spin_lock(&mdsc
->caps_list_lock
);
170 if (mdsc
->caps_avail_count
>= need
)
173 have
= mdsc
->caps_avail_count
;
174 mdsc
->caps_avail_count
-= have
;
175 mdsc
->caps_reserve_count
+= have
;
176 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
177 mdsc
->caps_reserve_count
+
178 mdsc
->caps_avail_count
);
179 spin_unlock(&mdsc
->caps_list_lock
);
181 for (i
= have
; i
< need
; i
++) {
182 cap
= kmem_cache_alloc(ceph_cap_cachep
, GFP_NOFS
);
185 list_add(&cap
->caps_item
, &newcaps
);
188 /* we didn't manage to reserve as much as we needed */
189 if (have
+ alloc
!= need
)
190 pr_warn("reserve caps ctx=%p ENOMEM need=%d got=%d\n",
191 ctx
, need
, have
+ alloc
);
193 spin_lock(&mdsc
->caps_list_lock
);
194 mdsc
->caps_total_count
+= alloc
;
195 mdsc
->caps_reserve_count
+= alloc
;
196 list_splice(&newcaps
, &mdsc
->caps_list
);
198 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
199 mdsc
->caps_reserve_count
+
200 mdsc
->caps_avail_count
);
201 spin_unlock(&mdsc
->caps_list_lock
);
204 dout("reserve caps ctx=%p %d = %d used + %d resv + %d avail\n",
205 ctx
, mdsc
->caps_total_count
, mdsc
->caps_use_count
,
206 mdsc
->caps_reserve_count
, mdsc
->caps_avail_count
);
209 int ceph_unreserve_caps(struct ceph_mds_client
*mdsc
,
210 struct ceph_cap_reservation
*ctx
)
212 dout("unreserve caps ctx=%p count=%d\n", ctx
, ctx
->count
);
214 spin_lock(&mdsc
->caps_list_lock
);
215 BUG_ON(mdsc
->caps_reserve_count
< ctx
->count
);
216 mdsc
->caps_reserve_count
-= ctx
->count
;
217 mdsc
->caps_avail_count
+= ctx
->count
;
219 dout("unreserve caps %d = %d used + %d resv + %d avail\n",
220 mdsc
->caps_total_count
, mdsc
->caps_use_count
,
221 mdsc
->caps_reserve_count
, mdsc
->caps_avail_count
);
222 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
223 mdsc
->caps_reserve_count
+
224 mdsc
->caps_avail_count
);
225 spin_unlock(&mdsc
->caps_list_lock
);
230 struct ceph_cap
*ceph_get_cap(struct ceph_mds_client
*mdsc
,
231 struct ceph_cap_reservation
*ctx
)
233 struct ceph_cap
*cap
= NULL
;
235 /* temporary, until we do something about cap import/export */
237 cap
= kmem_cache_alloc(ceph_cap_cachep
, GFP_NOFS
);
239 spin_lock(&mdsc
->caps_list_lock
);
240 mdsc
->caps_use_count
++;
241 mdsc
->caps_total_count
++;
242 spin_unlock(&mdsc
->caps_list_lock
);
247 spin_lock(&mdsc
->caps_list_lock
);
248 dout("get_cap ctx=%p (%d) %d = %d used + %d resv + %d avail\n",
249 ctx
, ctx
->count
, mdsc
->caps_total_count
, mdsc
->caps_use_count
,
250 mdsc
->caps_reserve_count
, mdsc
->caps_avail_count
);
252 BUG_ON(ctx
->count
> mdsc
->caps_reserve_count
);
253 BUG_ON(list_empty(&mdsc
->caps_list
));
256 mdsc
->caps_reserve_count
--;
257 mdsc
->caps_use_count
++;
259 cap
= list_first_entry(&mdsc
->caps_list
, struct ceph_cap
, caps_item
);
260 list_del(&cap
->caps_item
);
262 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
263 mdsc
->caps_reserve_count
+ mdsc
->caps_avail_count
);
264 spin_unlock(&mdsc
->caps_list_lock
);
268 void ceph_put_cap(struct ceph_mds_client
*mdsc
, struct ceph_cap
*cap
)
270 spin_lock(&mdsc
->caps_list_lock
);
271 dout("put_cap %p %d = %d used + %d resv + %d avail\n",
272 cap
, mdsc
->caps_total_count
, mdsc
->caps_use_count
,
273 mdsc
->caps_reserve_count
, mdsc
->caps_avail_count
);
274 mdsc
->caps_use_count
--;
276 * Keep some preallocated caps around (ceph_min_count), to
277 * avoid lots of free/alloc churn.
279 if (mdsc
->caps_avail_count
>= mdsc
->caps_reserve_count
+
280 mdsc
->caps_min_count
) {
281 mdsc
->caps_total_count
--;
282 kmem_cache_free(ceph_cap_cachep
, cap
);
284 mdsc
->caps_avail_count
++;
285 list_add(&cap
->caps_item
, &mdsc
->caps_list
);
288 BUG_ON(mdsc
->caps_total_count
!= mdsc
->caps_use_count
+
289 mdsc
->caps_reserve_count
+ mdsc
->caps_avail_count
);
290 spin_unlock(&mdsc
->caps_list_lock
);
293 void ceph_reservation_status(struct ceph_fs_client
*fsc
,
294 int *total
, int *avail
, int *used
, int *reserved
,
297 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
300 *total
= mdsc
->caps_total_count
;
302 *avail
= mdsc
->caps_avail_count
;
304 *used
= mdsc
->caps_use_count
;
306 *reserved
= mdsc
->caps_reserve_count
;
308 *min
= mdsc
->caps_min_count
;
312 * Find ceph_cap for given mds, if any.
314 * Called with i_ceph_lock held.
316 static struct ceph_cap
*__get_cap_for_mds(struct ceph_inode_info
*ci
, int mds
)
318 struct ceph_cap
*cap
;
319 struct rb_node
*n
= ci
->i_caps
.rb_node
;
322 cap
= rb_entry(n
, struct ceph_cap
, ci_node
);
325 else if (mds
> cap
->mds
)
333 struct ceph_cap
*ceph_get_cap_for_mds(struct ceph_inode_info
*ci
, int mds
)
335 struct ceph_cap
*cap
;
337 spin_lock(&ci
->i_ceph_lock
);
338 cap
= __get_cap_for_mds(ci
, mds
);
339 spin_unlock(&ci
->i_ceph_lock
);
344 * Return id of any MDS with a cap, preferably FILE_WR|BUFFER|EXCL, else -1.
346 static int __ceph_get_cap_mds(struct ceph_inode_info
*ci
)
348 struct ceph_cap
*cap
;
352 /* prefer mds with WR|BUFFER|EXCL caps */
353 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
354 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
356 if (cap
->issued
& (CEPH_CAP_FILE_WR
|
357 CEPH_CAP_FILE_BUFFER
|
364 int ceph_get_cap_mds(struct inode
*inode
)
366 struct ceph_inode_info
*ci
= ceph_inode(inode
);
368 spin_lock(&ci
->i_ceph_lock
);
369 mds
= __ceph_get_cap_mds(ceph_inode(inode
));
370 spin_unlock(&ci
->i_ceph_lock
);
375 * Called under i_ceph_lock.
377 static void __insert_cap_node(struct ceph_inode_info
*ci
,
378 struct ceph_cap
*new)
380 struct rb_node
**p
= &ci
->i_caps
.rb_node
;
381 struct rb_node
*parent
= NULL
;
382 struct ceph_cap
*cap
= NULL
;
386 cap
= rb_entry(parent
, struct ceph_cap
, ci_node
);
387 if (new->mds
< cap
->mds
)
389 else if (new->mds
> cap
->mds
)
395 rb_link_node(&new->ci_node
, parent
, p
);
396 rb_insert_color(&new->ci_node
, &ci
->i_caps
);
400 * (re)set cap hold timeouts, which control the delayed release
401 * of unused caps back to the MDS. Should be called on cap use.
403 static void __cap_set_timeouts(struct ceph_mds_client
*mdsc
,
404 struct ceph_inode_info
*ci
)
406 struct ceph_mount_options
*ma
= mdsc
->fsc
->mount_options
;
408 ci
->i_hold_caps_min
= round_jiffies(jiffies
+
409 ma
->caps_wanted_delay_min
* HZ
);
410 ci
->i_hold_caps_max
= round_jiffies(jiffies
+
411 ma
->caps_wanted_delay_max
* HZ
);
412 dout("__cap_set_timeouts %p min %lu max %lu\n", &ci
->vfs_inode
,
413 ci
->i_hold_caps_min
- jiffies
, ci
->i_hold_caps_max
- jiffies
);
417 * (Re)queue cap at the end of the delayed cap release list.
419 * If I_FLUSH is set, leave the inode at the front of the list.
421 * Caller holds i_ceph_lock
422 * -> we take mdsc->cap_delay_lock
424 static void __cap_delay_requeue(struct ceph_mds_client
*mdsc
,
425 struct ceph_inode_info
*ci
)
427 __cap_set_timeouts(mdsc
, ci
);
428 dout("__cap_delay_requeue %p flags %d at %lu\n", &ci
->vfs_inode
,
429 ci
->i_ceph_flags
, ci
->i_hold_caps_max
);
430 if (!mdsc
->stopping
) {
431 spin_lock(&mdsc
->cap_delay_lock
);
432 if (!list_empty(&ci
->i_cap_delay_list
)) {
433 if (ci
->i_ceph_flags
& CEPH_I_FLUSH
)
435 list_del_init(&ci
->i_cap_delay_list
);
437 list_add_tail(&ci
->i_cap_delay_list
, &mdsc
->cap_delay_list
);
439 spin_unlock(&mdsc
->cap_delay_lock
);
444 * Queue an inode for immediate writeback. Mark inode with I_FLUSH,
445 * indicating we should send a cap message to flush dirty metadata
446 * asap, and move to the front of the delayed cap list.
448 static void __cap_delay_requeue_front(struct ceph_mds_client
*mdsc
,
449 struct ceph_inode_info
*ci
)
451 dout("__cap_delay_requeue_front %p\n", &ci
->vfs_inode
);
452 spin_lock(&mdsc
->cap_delay_lock
);
453 ci
->i_ceph_flags
|= CEPH_I_FLUSH
;
454 if (!list_empty(&ci
->i_cap_delay_list
))
455 list_del_init(&ci
->i_cap_delay_list
);
456 list_add(&ci
->i_cap_delay_list
, &mdsc
->cap_delay_list
);
457 spin_unlock(&mdsc
->cap_delay_lock
);
461 * Cancel delayed work on cap.
463 * Caller must hold i_ceph_lock.
465 static void __cap_delay_cancel(struct ceph_mds_client
*mdsc
,
466 struct ceph_inode_info
*ci
)
468 dout("__cap_delay_cancel %p\n", &ci
->vfs_inode
);
469 if (list_empty(&ci
->i_cap_delay_list
))
471 spin_lock(&mdsc
->cap_delay_lock
);
472 list_del_init(&ci
->i_cap_delay_list
);
473 spin_unlock(&mdsc
->cap_delay_lock
);
477 * Common issue checks for add_cap, handle_cap_grant.
479 static void __check_cap_issue(struct ceph_inode_info
*ci
, struct ceph_cap
*cap
,
482 unsigned had
= __ceph_caps_issued(ci
, NULL
);
485 * Each time we receive FILE_CACHE anew, we increment
488 if ((issued
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)) &&
489 (had
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)) == 0) {
494 * If FILE_SHARED is newly issued, mark dir not complete. We don't
495 * know what happened to this directory while we didn't have the cap.
496 * If FILE_SHARED is being revoked, also mark dir not complete. It
497 * stops on-going cached readdir.
499 if ((issued
& CEPH_CAP_FILE_SHARED
) != (had
& CEPH_CAP_FILE_SHARED
)) {
500 if (issued
& CEPH_CAP_FILE_SHARED
)
502 if (S_ISDIR(ci
->vfs_inode
.i_mode
)) {
503 dout(" marking %p NOT complete\n", &ci
->vfs_inode
);
504 __ceph_dir_clear_complete(ci
);
510 * Add a capability under the given MDS session.
512 * Caller should hold session snap_rwsem (read) and s_mutex.
514 * @fmode is the open file mode, if we are opening a file, otherwise
515 * it is < 0. (This is so we can atomically add the cap and add an
516 * open file reference to it.)
518 void ceph_add_cap(struct inode
*inode
,
519 struct ceph_mds_session
*session
, u64 cap_id
,
520 int fmode
, unsigned issued
, unsigned wanted
,
521 unsigned seq
, unsigned mseq
, u64 realmino
, int flags
,
522 struct ceph_cap
**new_cap
)
524 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
525 struct ceph_inode_info
*ci
= ceph_inode(inode
);
526 struct ceph_cap
*cap
;
527 int mds
= session
->s_mds
;
530 dout("add_cap %p mds%d cap %llx %s seq %d\n", inode
,
531 session
->s_mds
, cap_id
, ceph_cap_string(issued
), seq
);
534 * If we are opening the file, include file mode wanted bits
538 wanted
|= ceph_caps_for_mode(fmode
);
540 cap
= __get_cap_for_mds(ci
, mds
);
546 cap
->implemented
= 0;
552 __insert_cap_node(ci
, cap
);
554 /* add to session cap list */
555 cap
->session
= session
;
556 spin_lock(&session
->s_cap_lock
);
557 list_add_tail(&cap
->session_caps
, &session
->s_caps
);
558 session
->s_nr_caps
++;
559 spin_unlock(&session
->s_cap_lock
);
562 * auth mds of the inode changed. we received the cap export
563 * message, but still haven't received the cap import message.
564 * handle_cap_export() updated the new auth MDS' cap.
566 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
567 * a message that was send before the cap import message. So
570 if (ceph_seq_cmp(seq
, cap
->seq
) <= 0) {
571 WARN_ON(cap
!= ci
->i_auth_cap
);
572 WARN_ON(cap
->cap_id
!= cap_id
);
575 issued
|= cap
->issued
;
576 flags
|= CEPH_CAP_FLAG_AUTH
;
580 if (!ci
->i_snap_realm
) {
582 * add this inode to the appropriate snap realm
584 struct ceph_snap_realm
*realm
= ceph_lookup_snap_realm(mdsc
,
587 spin_lock(&realm
->inodes_with_caps_lock
);
588 ci
->i_snap_realm
= realm
;
589 list_add(&ci
->i_snap_realm_item
,
590 &realm
->inodes_with_caps
);
591 spin_unlock(&realm
->inodes_with_caps_lock
);
593 pr_err("ceph_add_cap: couldn't find snap realm %llx\n",
599 __check_cap_issue(ci
, cap
, issued
);
602 * If we are issued caps we don't want, or the mds' wanted
603 * value appears to be off, queue a check so we'll release
604 * later and/or update the mds wanted value.
606 actual_wanted
= __ceph_caps_wanted(ci
);
607 if ((wanted
& ~actual_wanted
) ||
608 (issued
& ~actual_wanted
& CEPH_CAP_ANY_WR
)) {
609 dout(" issued %s, mds wanted %s, actual %s, queueing\n",
610 ceph_cap_string(issued
), ceph_cap_string(wanted
),
611 ceph_cap_string(actual_wanted
));
612 __cap_delay_requeue(mdsc
, ci
);
615 if (flags
& CEPH_CAP_FLAG_AUTH
) {
616 if (!ci
->i_auth_cap
||
617 ceph_seq_cmp(ci
->i_auth_cap
->mseq
, mseq
) < 0) {
618 ci
->i_auth_cap
= cap
;
619 cap
->mds_wanted
= wanted
;
622 WARN_ON(ci
->i_auth_cap
== cap
);
625 dout("add_cap inode %p (%llx.%llx) cap %p %s now %s seq %d mds%d\n",
626 inode
, ceph_vinop(inode
), cap
, ceph_cap_string(issued
),
627 ceph_cap_string(issued
|cap
->issued
), seq
, mds
);
628 cap
->cap_id
= cap_id
;
629 cap
->issued
= issued
;
630 cap
->implemented
|= issued
;
631 if (ceph_seq_cmp(mseq
, cap
->mseq
) > 0)
632 cap
->mds_wanted
= wanted
;
634 cap
->mds_wanted
|= wanted
;
636 cap
->issue_seq
= seq
;
638 cap
->cap_gen
= session
->s_cap_gen
;
641 __ceph_get_fmode(ci
, fmode
);
645 * Return true if cap has not timed out and belongs to the current
646 * generation of the MDS session (i.e. has not gone 'stale' due to
647 * us losing touch with the mds).
649 static int __cap_is_valid(struct ceph_cap
*cap
)
654 spin_lock(&cap
->session
->s_gen_ttl_lock
);
655 gen
= cap
->session
->s_cap_gen
;
656 ttl
= cap
->session
->s_cap_ttl
;
657 spin_unlock(&cap
->session
->s_gen_ttl_lock
);
659 if (cap
->cap_gen
< gen
|| time_after_eq(jiffies
, ttl
)) {
660 dout("__cap_is_valid %p cap %p issued %s "
661 "but STALE (gen %u vs %u)\n", &cap
->ci
->vfs_inode
,
662 cap
, ceph_cap_string(cap
->issued
), cap
->cap_gen
, gen
);
670 * Return set of valid cap bits issued to us. Note that caps time
671 * out, and may be invalidated in bulk if the client session times out
672 * and session->s_cap_gen is bumped.
674 int __ceph_caps_issued(struct ceph_inode_info
*ci
, int *implemented
)
676 int have
= ci
->i_snap_caps
;
677 struct ceph_cap
*cap
;
682 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
683 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
684 if (!__cap_is_valid(cap
))
686 dout("__ceph_caps_issued %p cap %p issued %s\n",
687 &ci
->vfs_inode
, cap
, ceph_cap_string(cap
->issued
));
690 *implemented
|= cap
->implemented
;
693 * exclude caps issued by non-auth MDS, but are been revoking
694 * by the auth MDS. The non-auth MDS should be revoking/exporting
695 * these caps, but the message is delayed.
697 if (ci
->i_auth_cap
) {
698 cap
= ci
->i_auth_cap
;
699 have
&= ~cap
->implemented
| cap
->issued
;
705 * Get cap bits issued by caps other than @ocap
707 int __ceph_caps_issued_other(struct ceph_inode_info
*ci
, struct ceph_cap
*ocap
)
709 int have
= ci
->i_snap_caps
;
710 struct ceph_cap
*cap
;
713 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
714 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
717 if (!__cap_is_valid(cap
))
725 * Move a cap to the end of the LRU (oldest caps at list head, newest
728 static void __touch_cap(struct ceph_cap
*cap
)
730 struct ceph_mds_session
*s
= cap
->session
;
732 spin_lock(&s
->s_cap_lock
);
733 if (!s
->s_cap_iterator
) {
734 dout("__touch_cap %p cap %p mds%d\n", &cap
->ci
->vfs_inode
, cap
,
736 list_move_tail(&cap
->session_caps
, &s
->s_caps
);
738 dout("__touch_cap %p cap %p mds%d NOP, iterating over caps\n",
739 &cap
->ci
->vfs_inode
, cap
, s
->s_mds
);
741 spin_unlock(&s
->s_cap_lock
);
745 * Check if we hold the given mask. If so, move the cap(s) to the
746 * front of their respective LRUs. (This is the preferred way for
747 * callers to check for caps they want.)
749 int __ceph_caps_issued_mask(struct ceph_inode_info
*ci
, int mask
, int touch
)
751 struct ceph_cap
*cap
;
753 int have
= ci
->i_snap_caps
;
755 if ((have
& mask
) == mask
) {
756 dout("__ceph_caps_issued_mask %p snap issued %s"
757 " (mask %s)\n", &ci
->vfs_inode
,
758 ceph_cap_string(have
),
759 ceph_cap_string(mask
));
763 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
764 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
765 if (!__cap_is_valid(cap
))
767 if ((cap
->issued
& mask
) == mask
) {
768 dout("__ceph_caps_issued_mask %p cap %p issued %s"
769 " (mask %s)\n", &ci
->vfs_inode
, cap
,
770 ceph_cap_string(cap
->issued
),
771 ceph_cap_string(mask
));
777 /* does a combination of caps satisfy mask? */
779 if ((have
& mask
) == mask
) {
780 dout("__ceph_caps_issued_mask %p combo issued %s"
781 " (mask %s)\n", &ci
->vfs_inode
,
782 ceph_cap_string(cap
->issued
),
783 ceph_cap_string(mask
));
787 /* touch this + preceding caps */
789 for (q
= rb_first(&ci
->i_caps
); q
!= p
;
791 cap
= rb_entry(q
, struct ceph_cap
,
793 if (!__cap_is_valid(cap
))
806 * Return true if mask caps are currently being revoked by an MDS.
808 int __ceph_caps_revoking_other(struct ceph_inode_info
*ci
,
809 struct ceph_cap
*ocap
, int mask
)
811 struct ceph_cap
*cap
;
814 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
815 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
817 (cap
->implemented
& ~cap
->issued
& mask
))
823 int ceph_caps_revoking(struct ceph_inode_info
*ci
, int mask
)
825 struct inode
*inode
= &ci
->vfs_inode
;
828 spin_lock(&ci
->i_ceph_lock
);
829 ret
= __ceph_caps_revoking_other(ci
, NULL
, mask
);
830 spin_unlock(&ci
->i_ceph_lock
);
831 dout("ceph_caps_revoking %p %s = %d\n", inode
,
832 ceph_cap_string(mask
), ret
);
836 int __ceph_caps_used(struct ceph_inode_info
*ci
)
840 used
|= CEPH_CAP_PIN
;
842 used
|= CEPH_CAP_FILE_RD
;
843 if (ci
->i_rdcache_ref
||
844 (!S_ISDIR(ci
->vfs_inode
.i_mode
) && /* ignore readdir cache */
845 ci
->vfs_inode
.i_data
.nrpages
))
846 used
|= CEPH_CAP_FILE_CACHE
;
848 used
|= CEPH_CAP_FILE_WR
;
849 if (ci
->i_wb_ref
|| ci
->i_wrbuffer_ref
)
850 used
|= CEPH_CAP_FILE_BUFFER
;
855 * wanted, by virtue of open file modes
857 int __ceph_caps_file_wanted(struct ceph_inode_info
*ci
)
860 for (i
= 0; i
< CEPH_FILE_MODE_BITS
; i
++) {
861 if (ci
->i_nr_by_mode
[i
])
866 return ceph_caps_for_mode(bits
>> 1);
870 * Return caps we have registered with the MDS(s) as 'wanted'.
872 int __ceph_caps_mds_wanted(struct ceph_inode_info
*ci
, bool check
)
874 struct ceph_cap
*cap
;
878 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
879 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
880 if (check
&& !__cap_is_valid(cap
))
882 if (cap
== ci
->i_auth_cap
)
883 mds_wanted
|= cap
->mds_wanted
;
885 mds_wanted
|= (cap
->mds_wanted
& ~CEPH_CAP_ANY_FILE_WR
);
891 * called under i_ceph_lock
893 static int __ceph_is_any_caps(struct ceph_inode_info
*ci
)
895 return !RB_EMPTY_ROOT(&ci
->i_caps
);
898 int ceph_is_any_caps(struct inode
*inode
)
900 struct ceph_inode_info
*ci
= ceph_inode(inode
);
903 spin_lock(&ci
->i_ceph_lock
);
904 ret
= __ceph_is_any_caps(ci
);
905 spin_unlock(&ci
->i_ceph_lock
);
910 static void drop_inode_snap_realm(struct ceph_inode_info
*ci
)
912 struct ceph_snap_realm
*realm
= ci
->i_snap_realm
;
913 spin_lock(&realm
->inodes_with_caps_lock
);
914 list_del_init(&ci
->i_snap_realm_item
);
915 ci
->i_snap_realm_counter
++;
916 ci
->i_snap_realm
= NULL
;
917 spin_unlock(&realm
->inodes_with_caps_lock
);
918 ceph_put_snap_realm(ceph_sb_to_client(ci
->vfs_inode
.i_sb
)->mdsc
,
923 * Remove a cap. Take steps to deal with a racing iterate_session_caps.
925 * caller should hold i_ceph_lock.
926 * caller will not hold session s_mutex if called from destroy_inode.
928 void __ceph_remove_cap(struct ceph_cap
*cap
, bool queue_release
)
930 struct ceph_mds_session
*session
= cap
->session
;
931 struct ceph_inode_info
*ci
= cap
->ci
;
932 struct ceph_mds_client
*mdsc
=
933 ceph_sb_to_client(ci
->vfs_inode
.i_sb
)->mdsc
;
936 dout("__ceph_remove_cap %p from %p\n", cap
, &ci
->vfs_inode
);
938 /* remove from session list */
939 spin_lock(&session
->s_cap_lock
);
940 if (session
->s_cap_iterator
== cap
) {
941 /* not yet, we are iterating over this very cap */
942 dout("__ceph_remove_cap delaying %p removal from session %p\n",
945 list_del_init(&cap
->session_caps
);
946 session
->s_nr_caps
--;
950 /* protect backpointer with s_cap_lock: see iterate_session_caps */
954 * s_cap_reconnect is protected by s_cap_lock. no one changes
955 * s_cap_gen while session is in the reconnect state.
958 (!session
->s_cap_reconnect
|| cap
->cap_gen
== session
->s_cap_gen
)) {
959 cap
->queue_release
= 1;
961 list_add_tail(&cap
->session_caps
,
962 &session
->s_cap_releases
);
963 session
->s_num_cap_releases
++;
967 cap
->queue_release
= 0;
969 cap
->cap_ino
= ci
->i_vino
.ino
;
971 spin_unlock(&session
->s_cap_lock
);
973 /* remove from inode list */
974 rb_erase(&cap
->ci_node
, &ci
->i_caps
);
975 if (ci
->i_auth_cap
== cap
)
976 ci
->i_auth_cap
= NULL
;
979 ceph_put_cap(mdsc
, cap
);
981 /* when reconnect denied, we remove session caps forcibly,
982 * i_wr_ref can be non-zero. If there are ongoing write,
985 if (!__ceph_is_any_caps(ci
) && ci
->i_wr_ref
== 0 && ci
->i_snap_realm
)
986 drop_inode_snap_realm(ci
);
988 if (!__ceph_is_any_real_caps(ci
))
989 __cap_delay_cancel(mdsc
, ci
);
992 struct cap_msg_args
{
993 struct ceph_mds_session
*session
;
994 u64 ino
, cid
, follows
;
995 u64 flush_tid
, oldest_flush_tid
, size
, max_size
;
997 struct ceph_buffer
*xattr_buf
;
998 struct timespec atime
, mtime
, ctime
;
999 int op
, caps
, wanted
, dirty
;
1000 u32 seq
, issue_seq
, mseq
, time_warp_seq
;
1009 * Build and send a cap message to the given MDS.
1011 * Caller should be holding s_mutex.
1013 static int send_cap_msg(struct cap_msg_args
*arg
)
1015 struct ceph_mds_caps
*fc
;
1016 struct ceph_msg
*msg
;
1019 struct timespec zerotime
= {0};
1020 struct ceph_osd_client
*osdc
= &arg
->session
->s_mdsc
->fsc
->client
->osdc
;
1022 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
1023 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
1024 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(arg
->op
),
1025 arg
->cid
, arg
->ino
, ceph_cap_string(arg
->caps
),
1026 ceph_cap_string(arg
->wanted
), ceph_cap_string(arg
->dirty
),
1027 arg
->seq
, arg
->issue_seq
, arg
->flush_tid
, arg
->oldest_flush_tid
,
1028 arg
->mseq
, arg
->follows
, arg
->size
, arg
->max_size
,
1030 arg
->xattr_buf
? (int)arg
->xattr_buf
->vec
.iov_len
: 0);
1032 /* flock buffer size + inline version + inline data size +
1033 * osd_epoch_barrier + oldest_flush_tid */
1034 extra_len
= 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 4;
1035 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPS
, sizeof(*fc
) + extra_len
,
1040 msg
->hdr
.version
= cpu_to_le16(10);
1041 msg
->hdr
.tid
= cpu_to_le64(arg
->flush_tid
);
1043 fc
= msg
->front
.iov_base
;
1044 memset(fc
, 0, sizeof(*fc
));
1046 fc
->cap_id
= cpu_to_le64(arg
->cid
);
1047 fc
->op
= cpu_to_le32(arg
->op
);
1048 fc
->seq
= cpu_to_le32(arg
->seq
);
1049 fc
->issue_seq
= cpu_to_le32(arg
->issue_seq
);
1050 fc
->migrate_seq
= cpu_to_le32(arg
->mseq
);
1051 fc
->caps
= cpu_to_le32(arg
->caps
);
1052 fc
->wanted
= cpu_to_le32(arg
->wanted
);
1053 fc
->dirty
= cpu_to_le32(arg
->dirty
);
1054 fc
->ino
= cpu_to_le64(arg
->ino
);
1055 fc
->snap_follows
= cpu_to_le64(arg
->follows
);
1057 fc
->size
= cpu_to_le64(arg
->size
);
1058 fc
->max_size
= cpu_to_le64(arg
->max_size
);
1059 ceph_encode_timespec(&fc
->mtime
, &arg
->mtime
);
1060 ceph_encode_timespec(&fc
->atime
, &arg
->atime
);
1061 ceph_encode_timespec(&fc
->ctime
, &arg
->ctime
);
1062 fc
->time_warp_seq
= cpu_to_le32(arg
->time_warp_seq
);
1064 fc
->uid
= cpu_to_le32(from_kuid(&init_user_ns
, arg
->uid
));
1065 fc
->gid
= cpu_to_le32(from_kgid(&init_user_ns
, arg
->gid
));
1066 fc
->mode
= cpu_to_le32(arg
->mode
);
1068 fc
->xattr_version
= cpu_to_le64(arg
->xattr_version
);
1069 if (arg
->xattr_buf
) {
1070 msg
->middle
= ceph_buffer_get(arg
->xattr_buf
);
1071 fc
->xattr_len
= cpu_to_le32(arg
->xattr_buf
->vec
.iov_len
);
1072 msg
->hdr
.middle_len
= cpu_to_le32(arg
->xattr_buf
->vec
.iov_len
);
1076 /* flock buffer size (version 2) */
1077 ceph_encode_32(&p
, 0);
1078 /* inline version (version 4) */
1079 ceph_encode_64(&p
, arg
->inline_data
? 0 : CEPH_INLINE_NONE
);
1080 /* inline data size */
1081 ceph_encode_32(&p
, 0);
1083 * osd_epoch_barrier (version 5)
1084 * The epoch_barrier is protected osdc->lock, so READ_ONCE here in
1085 * case it was recently changed
1087 ceph_encode_32(&p
, READ_ONCE(osdc
->epoch_barrier
));
1088 /* oldest_flush_tid (version 6) */
1089 ceph_encode_64(&p
, arg
->oldest_flush_tid
);
1092 * caller_uid/caller_gid (version 7)
1094 * Currently, we don't properly track which caller dirtied the caps
1095 * last, and force a flush of them when there is a conflict. For now,
1096 * just set this to 0:0, to emulate how the MDS has worked up to now.
1098 ceph_encode_32(&p
, 0);
1099 ceph_encode_32(&p
, 0);
1101 /* pool namespace (version 8) (mds always ignores this) */
1102 ceph_encode_32(&p
, 0);
1105 * btime and change_attr (version 9)
1107 * We just zero these out for now, as the MDS ignores them unless
1108 * the requisite feature flags are set (which we don't do yet).
1110 ceph_encode_timespec(p
, &zerotime
);
1111 p
+= sizeof(struct ceph_timespec
);
1112 ceph_encode_64(&p
, 0);
1114 /* Advisory flags (version 10) */
1115 ceph_encode_32(&p
, arg
->flags
);
1117 ceph_con_send(&arg
->session
->s_con
, msg
);
1122 * Queue cap releases when an inode is dropped from our cache. Since
1123 * inode is about to be destroyed, there is no need for i_ceph_lock.
1125 void ceph_queue_caps_release(struct inode
*inode
)
1127 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1130 p
= rb_first(&ci
->i_caps
);
1132 struct ceph_cap
*cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
1134 __ceph_remove_cap(cap
, true);
1139 * Send a cap msg on the given inode. Update our caps state, then
1140 * drop i_ceph_lock and send the message.
1142 * Make note of max_size reported/requested from mds, revoked caps
1143 * that have now been implemented.
1145 * Make half-hearted attempt ot to invalidate page cache if we are
1146 * dropping RDCACHE. Note that this will leave behind locked pages
1147 * that we'll then need to deal with elsewhere.
1149 * Return non-zero if delayed release, or we experienced an error
1150 * such that the caller should requeue + retry later.
1152 * called with i_ceph_lock, then drops it.
1153 * caller should hold snap_rwsem (read), s_mutex.
1155 static int __send_cap(struct ceph_mds_client
*mdsc
, struct ceph_cap
*cap
,
1156 int op
, bool sync
, int used
, int want
, int retain
,
1157 int flushing
, u64 flush_tid
, u64 oldest_flush_tid
)
1158 __releases(cap
->ci
->i_ceph_lock
)
1160 struct ceph_inode_info
*ci
= cap
->ci
;
1161 struct inode
*inode
= &ci
->vfs_inode
;
1162 struct cap_msg_args arg
;
1168 held
= cap
->issued
| cap
->implemented
;
1169 revoking
= cap
->implemented
& ~cap
->issued
;
1170 retain
&= ~revoking
;
1172 dout("__send_cap %p cap %p session %p %s -> %s (revoking %s)\n",
1173 inode
, cap
, cap
->session
,
1174 ceph_cap_string(held
), ceph_cap_string(held
& retain
),
1175 ceph_cap_string(revoking
));
1176 BUG_ON((retain
& CEPH_CAP_PIN
) == 0);
1178 arg
.session
= cap
->session
;
1180 /* don't release wanted unless we've waited a bit. */
1181 if ((ci
->i_ceph_flags
& CEPH_I_NODELAY
) == 0 &&
1182 time_before(jiffies
, ci
->i_hold_caps_min
)) {
1183 dout(" delaying issued %s -> %s, wanted %s -> %s on send\n",
1184 ceph_cap_string(cap
->issued
),
1185 ceph_cap_string(cap
->issued
& retain
),
1186 ceph_cap_string(cap
->mds_wanted
),
1187 ceph_cap_string(want
));
1188 want
|= cap
->mds_wanted
;
1189 retain
|= cap
->issued
;
1192 ci
->i_ceph_flags
&= ~(CEPH_I_NODELAY
| CEPH_I_FLUSH
);
1193 if (want
& ~cap
->mds_wanted
) {
1194 /* user space may open/close single file frequently.
1195 * This avoids droping mds_wanted immediately after
1196 * requesting new mds_wanted.
1198 __cap_set_timeouts(mdsc
, ci
);
1201 cap
->issued
&= retain
; /* drop bits we don't want */
1202 if (cap
->implemented
& ~cap
->issued
) {
1204 * Wake up any waiters on wanted -> needed transition.
1205 * This is due to the weird transition from buffered
1206 * to sync IO... we need to flush dirty pages _before_
1207 * allowing sync writes to avoid reordering.
1211 cap
->implemented
&= cap
->issued
| used
;
1212 cap
->mds_wanted
= want
;
1214 arg
.ino
= ceph_vino(inode
).ino
;
1215 arg
.cid
= cap
->cap_id
;
1216 arg
.follows
= flushing
? ci
->i_head_snapc
->seq
: 0;
1217 arg
.flush_tid
= flush_tid
;
1218 arg
.oldest_flush_tid
= oldest_flush_tid
;
1220 arg
.size
= inode
->i_size
;
1221 ci
->i_reported_size
= arg
.size
;
1222 arg
.max_size
= ci
->i_wanted_max_size
;
1223 ci
->i_requested_max_size
= arg
.max_size
;
1225 if (flushing
& CEPH_CAP_XATTR_EXCL
) {
1226 __ceph_build_xattrs_blob(ci
);
1227 arg
.xattr_version
= ci
->i_xattrs
.version
;
1228 arg
.xattr_buf
= ci
->i_xattrs
.blob
;
1230 arg
.xattr_buf
= NULL
;
1233 arg
.mtime
= inode
->i_mtime
;
1234 arg
.atime
= inode
->i_atime
;
1235 arg
.ctime
= inode
->i_ctime
;
1238 arg
.caps
= cap
->implemented
;
1240 arg
.dirty
= flushing
;
1243 arg
.issue_seq
= cap
->issue_seq
;
1244 arg
.mseq
= cap
->mseq
;
1245 arg
.time_warp_seq
= ci
->i_time_warp_seq
;
1247 arg
.uid
= inode
->i_uid
;
1248 arg
.gid
= inode
->i_gid
;
1249 arg
.mode
= inode
->i_mode
;
1251 arg
.inline_data
= ci
->i_inline_version
!= CEPH_INLINE_NONE
;
1252 if (list_empty(&ci
->i_cap_snaps
))
1253 arg
.flags
= CEPH_CLIENT_CAPS_NO_CAPSNAP
;
1255 arg
.flags
= CEPH_CLIENT_CAPS_PENDING_CAPSNAP
;
1257 arg
.flags
|= CEPH_CLIENT_CAPS_SYNC
;
1259 spin_unlock(&ci
->i_ceph_lock
);
1261 ret
= send_cap_msg(&arg
);
1263 dout("error sending cap msg, must requeue %p\n", inode
);
1268 wake_up_all(&ci
->i_cap_wq
);
1273 static inline int __send_flush_snap(struct inode
*inode
,
1274 struct ceph_mds_session
*session
,
1275 struct ceph_cap_snap
*capsnap
,
1276 u32 mseq
, u64 oldest_flush_tid
)
1278 struct cap_msg_args arg
;
1280 arg
.session
= session
;
1281 arg
.ino
= ceph_vino(inode
).ino
;
1283 arg
.follows
= capsnap
->follows
;
1284 arg
.flush_tid
= capsnap
->cap_flush
.tid
;
1285 arg
.oldest_flush_tid
= oldest_flush_tid
;
1287 arg
.size
= capsnap
->size
;
1289 arg
.xattr_version
= capsnap
->xattr_version
;
1290 arg
.xattr_buf
= capsnap
->xattr_blob
;
1292 arg
.atime
= capsnap
->atime
;
1293 arg
.mtime
= capsnap
->mtime
;
1294 arg
.ctime
= capsnap
->ctime
;
1296 arg
.op
= CEPH_CAP_OP_FLUSHSNAP
;
1297 arg
.caps
= capsnap
->issued
;
1299 arg
.dirty
= capsnap
->dirty
;
1304 arg
.time_warp_seq
= capsnap
->time_warp_seq
;
1306 arg
.uid
= capsnap
->uid
;
1307 arg
.gid
= capsnap
->gid
;
1308 arg
.mode
= capsnap
->mode
;
1310 arg
.inline_data
= capsnap
->inline_data
;
1313 return send_cap_msg(&arg
);
1317 * When a snapshot is taken, clients accumulate dirty metadata on
1318 * inodes with capabilities in ceph_cap_snaps to describe the file
1319 * state at the time the snapshot was taken. This must be flushed
1320 * asynchronously back to the MDS once sync writes complete and dirty
1321 * data is written out.
1323 * Called under i_ceph_lock. Takes s_mutex as needed.
1325 static void __ceph_flush_snaps(struct ceph_inode_info
*ci
,
1326 struct ceph_mds_session
*session
)
1327 __releases(ci
->i_ceph_lock
)
1328 __acquires(ci
->i_ceph_lock
)
1330 struct inode
*inode
= &ci
->vfs_inode
;
1331 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
1332 struct ceph_cap_snap
*capsnap
;
1333 u64 oldest_flush_tid
= 0;
1334 u64 first_tid
= 1, last_tid
= 0;
1336 dout("__flush_snaps %p session %p\n", inode
, session
);
1338 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
1340 * we need to wait for sync writes to complete and for dirty
1341 * pages to be written out.
1343 if (capsnap
->dirty_pages
|| capsnap
->writing
)
1346 /* should be removed by ceph_try_drop_cap_snap() */
1347 BUG_ON(!capsnap
->need_flush
);
1349 /* only flush each capsnap once */
1350 if (capsnap
->cap_flush
.tid
> 0) {
1351 dout(" already flushed %p, skipping\n", capsnap
);
1355 spin_lock(&mdsc
->cap_dirty_lock
);
1356 capsnap
->cap_flush
.tid
= ++mdsc
->last_cap_flush_tid
;
1357 list_add_tail(&capsnap
->cap_flush
.g_list
,
1358 &mdsc
->cap_flush_list
);
1359 if (oldest_flush_tid
== 0)
1360 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1361 if (list_empty(&ci
->i_flushing_item
)) {
1362 list_add_tail(&ci
->i_flushing_item
,
1363 &session
->s_cap_flushing
);
1365 spin_unlock(&mdsc
->cap_dirty_lock
);
1367 list_add_tail(&capsnap
->cap_flush
.i_list
,
1368 &ci
->i_cap_flush_list
);
1371 first_tid
= capsnap
->cap_flush
.tid
;
1372 last_tid
= capsnap
->cap_flush
.tid
;
1375 ci
->i_ceph_flags
&= ~CEPH_I_FLUSH_SNAPS
;
1377 while (first_tid
<= last_tid
) {
1378 struct ceph_cap
*cap
= ci
->i_auth_cap
;
1379 struct ceph_cap_flush
*cf
;
1382 if (!(cap
&& cap
->session
== session
)) {
1383 dout("__flush_snaps %p auth cap %p not mds%d, "
1384 "stop\n", inode
, cap
, session
->s_mds
);
1389 list_for_each_entry(cf
, &ci
->i_cap_flush_list
, i_list
) {
1390 if (cf
->tid
>= first_tid
) {
1398 first_tid
= cf
->tid
+ 1;
1400 capsnap
= container_of(cf
, struct ceph_cap_snap
, cap_flush
);
1401 refcount_inc(&capsnap
->nref
);
1402 spin_unlock(&ci
->i_ceph_lock
);
1404 dout("__flush_snaps %p capsnap %p tid %llu %s\n",
1405 inode
, capsnap
, cf
->tid
, ceph_cap_string(capsnap
->dirty
));
1407 ret
= __send_flush_snap(inode
, session
, capsnap
, cap
->mseq
,
1410 pr_err("__flush_snaps: error sending cap flushsnap, "
1411 "ino (%llx.%llx) tid %llu follows %llu\n",
1412 ceph_vinop(inode
), cf
->tid
, capsnap
->follows
);
1415 ceph_put_cap_snap(capsnap
);
1416 spin_lock(&ci
->i_ceph_lock
);
1420 void ceph_flush_snaps(struct ceph_inode_info
*ci
,
1421 struct ceph_mds_session
**psession
)
1423 struct inode
*inode
= &ci
->vfs_inode
;
1424 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
1425 struct ceph_mds_session
*session
= NULL
;
1428 dout("ceph_flush_snaps %p\n", inode
);
1430 session
= *psession
;
1432 spin_lock(&ci
->i_ceph_lock
);
1433 if (!(ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
)) {
1434 dout(" no capsnap needs flush, doing nothing\n");
1437 if (!ci
->i_auth_cap
) {
1438 dout(" no auth cap (migrating?), doing nothing\n");
1442 mds
= ci
->i_auth_cap
->session
->s_mds
;
1443 if (session
&& session
->s_mds
!= mds
) {
1444 dout(" oops, wrong session %p mutex\n", session
);
1445 mutex_unlock(&session
->s_mutex
);
1446 ceph_put_mds_session(session
);
1450 spin_unlock(&ci
->i_ceph_lock
);
1451 mutex_lock(&mdsc
->mutex
);
1452 session
= __ceph_lookup_mds_session(mdsc
, mds
);
1453 mutex_unlock(&mdsc
->mutex
);
1455 dout(" inverting session/ino locks on %p\n", session
);
1456 mutex_lock(&session
->s_mutex
);
1461 // make sure flushsnap messages are sent in proper order.
1462 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
) {
1463 __kick_flushing_caps(mdsc
, session
, ci
, 0);
1464 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
1467 __ceph_flush_snaps(ci
, session
);
1469 spin_unlock(&ci
->i_ceph_lock
);
1472 *psession
= session
;
1473 } else if (session
) {
1474 mutex_unlock(&session
->s_mutex
);
1475 ceph_put_mds_session(session
);
1477 /* we flushed them all; remove this inode from the queue */
1478 spin_lock(&mdsc
->snap_flush_lock
);
1479 list_del_init(&ci
->i_snap_flush_item
);
1480 spin_unlock(&mdsc
->snap_flush_lock
);
1484 * Mark caps dirty. If inode is newly dirty, return the dirty flags.
1485 * Caller is then responsible for calling __mark_inode_dirty with the
1486 * returned flags value.
1488 int __ceph_mark_dirty_caps(struct ceph_inode_info
*ci
, int mask
,
1489 struct ceph_cap_flush
**pcf
)
1491 struct ceph_mds_client
*mdsc
=
1492 ceph_sb_to_client(ci
->vfs_inode
.i_sb
)->mdsc
;
1493 struct inode
*inode
= &ci
->vfs_inode
;
1494 int was
= ci
->i_dirty_caps
;
1497 if (!ci
->i_auth_cap
) {
1498 pr_warn("__mark_dirty_caps %p %llx mask %s, "
1499 "but no auth cap (session was closed?)\n",
1500 inode
, ceph_ino(inode
), ceph_cap_string(mask
));
1504 dout("__mark_dirty_caps %p %s dirty %s -> %s\n", &ci
->vfs_inode
,
1505 ceph_cap_string(mask
), ceph_cap_string(was
),
1506 ceph_cap_string(was
| mask
));
1507 ci
->i_dirty_caps
|= mask
;
1509 WARN_ON_ONCE(ci
->i_prealloc_cap_flush
);
1510 swap(ci
->i_prealloc_cap_flush
, *pcf
);
1512 if (!ci
->i_head_snapc
) {
1513 WARN_ON_ONCE(!rwsem_is_locked(&mdsc
->snap_rwsem
));
1514 ci
->i_head_snapc
= ceph_get_snap_context(
1515 ci
->i_snap_realm
->cached_context
);
1517 dout(" inode %p now dirty snapc %p auth cap %p\n",
1518 &ci
->vfs_inode
, ci
->i_head_snapc
, ci
->i_auth_cap
);
1519 BUG_ON(!list_empty(&ci
->i_dirty_item
));
1520 spin_lock(&mdsc
->cap_dirty_lock
);
1521 list_add(&ci
->i_dirty_item
, &mdsc
->cap_dirty
);
1522 spin_unlock(&mdsc
->cap_dirty_lock
);
1523 if (ci
->i_flushing_caps
== 0) {
1525 dirty
|= I_DIRTY_SYNC
;
1528 WARN_ON_ONCE(!ci
->i_prealloc_cap_flush
);
1530 BUG_ON(list_empty(&ci
->i_dirty_item
));
1531 if (((was
| ci
->i_flushing_caps
) & CEPH_CAP_FILE_BUFFER
) &&
1532 (mask
& CEPH_CAP_FILE_BUFFER
))
1533 dirty
|= I_DIRTY_DATASYNC
;
1534 __cap_delay_requeue(mdsc
, ci
);
1538 struct ceph_cap_flush
*ceph_alloc_cap_flush(void)
1540 return kmem_cache_alloc(ceph_cap_flush_cachep
, GFP_KERNEL
);
1543 void ceph_free_cap_flush(struct ceph_cap_flush
*cf
)
1546 kmem_cache_free(ceph_cap_flush_cachep
, cf
);
1549 static u64
__get_oldest_flush_tid(struct ceph_mds_client
*mdsc
)
1551 if (!list_empty(&mdsc
->cap_flush_list
)) {
1552 struct ceph_cap_flush
*cf
=
1553 list_first_entry(&mdsc
->cap_flush_list
,
1554 struct ceph_cap_flush
, g_list
);
1561 * Remove cap_flush from the mdsc's or inode's flushing cap list.
1562 * Return true if caller needs to wake up flush waiters.
1564 static bool __finish_cap_flush(struct ceph_mds_client
*mdsc
,
1565 struct ceph_inode_info
*ci
,
1566 struct ceph_cap_flush
*cf
)
1568 struct ceph_cap_flush
*prev
;
1569 bool wake
= cf
->wake
;
1571 /* are there older pending cap flushes? */
1572 if (wake
&& cf
->g_list
.prev
!= &mdsc
->cap_flush_list
) {
1573 prev
= list_prev_entry(cf
, g_list
);
1577 list_del(&cf
->g_list
);
1579 if (wake
&& cf
->i_list
.prev
!= &ci
->i_cap_flush_list
) {
1580 prev
= list_prev_entry(cf
, i_list
);
1584 list_del(&cf
->i_list
);
1592 * Add dirty inode to the flushing list. Assigned a seq number so we
1593 * can wait for caps to flush without starving.
1595 * Called under i_ceph_lock.
1597 static int __mark_caps_flushing(struct inode
*inode
,
1598 struct ceph_mds_session
*session
, bool wake
,
1599 u64
*flush_tid
, u64
*oldest_flush_tid
)
1601 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
1602 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1603 struct ceph_cap_flush
*cf
= NULL
;
1606 BUG_ON(ci
->i_dirty_caps
== 0);
1607 BUG_ON(list_empty(&ci
->i_dirty_item
));
1608 BUG_ON(!ci
->i_prealloc_cap_flush
);
1610 flushing
= ci
->i_dirty_caps
;
1611 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
1612 ceph_cap_string(flushing
),
1613 ceph_cap_string(ci
->i_flushing_caps
),
1614 ceph_cap_string(ci
->i_flushing_caps
| flushing
));
1615 ci
->i_flushing_caps
|= flushing
;
1616 ci
->i_dirty_caps
= 0;
1617 dout(" inode %p now !dirty\n", inode
);
1619 swap(cf
, ci
->i_prealloc_cap_flush
);
1620 cf
->caps
= flushing
;
1623 spin_lock(&mdsc
->cap_dirty_lock
);
1624 list_del_init(&ci
->i_dirty_item
);
1626 cf
->tid
= ++mdsc
->last_cap_flush_tid
;
1627 list_add_tail(&cf
->g_list
, &mdsc
->cap_flush_list
);
1628 *oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1630 if (list_empty(&ci
->i_flushing_item
)) {
1631 list_add_tail(&ci
->i_flushing_item
, &session
->s_cap_flushing
);
1632 mdsc
->num_cap_flushing
++;
1634 spin_unlock(&mdsc
->cap_dirty_lock
);
1636 list_add_tail(&cf
->i_list
, &ci
->i_cap_flush_list
);
1638 *flush_tid
= cf
->tid
;
1643 * try to invalidate mapping pages without blocking.
1645 static int try_nonblocking_invalidate(struct inode
*inode
)
1647 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1648 u32 invalidating_gen
= ci
->i_rdcache_gen
;
1650 spin_unlock(&ci
->i_ceph_lock
);
1651 invalidate_mapping_pages(&inode
->i_data
, 0, -1);
1652 spin_lock(&ci
->i_ceph_lock
);
1654 if (inode
->i_data
.nrpages
== 0 &&
1655 invalidating_gen
== ci
->i_rdcache_gen
) {
1657 dout("try_nonblocking_invalidate %p success\n", inode
);
1658 /* save any racing async invalidate some trouble */
1659 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
- 1;
1662 dout("try_nonblocking_invalidate %p failed\n", inode
);
1666 bool __ceph_should_report_size(struct ceph_inode_info
*ci
)
1668 loff_t size
= ci
->vfs_inode
.i_size
;
1669 /* mds will adjust max size according to the reported size */
1670 if (ci
->i_flushing_caps
& CEPH_CAP_FILE_WR
)
1672 if (size
>= ci
->i_max_size
)
1674 /* half of previous max_size increment has been used */
1675 if (ci
->i_max_size
> ci
->i_reported_size
&&
1676 (size
<< 1) >= ci
->i_max_size
+ ci
->i_reported_size
)
1682 * Swiss army knife function to examine currently used and wanted
1683 * versus held caps. Release, flush, ack revoked caps to mds as
1686 * CHECK_CAPS_NODELAY - caller is delayed work and we should not delay
1687 * cap release further.
1688 * CHECK_CAPS_AUTHONLY - we should only check the auth cap
1689 * CHECK_CAPS_FLUSH - we should flush any dirty caps immediately, without
1692 void ceph_check_caps(struct ceph_inode_info
*ci
, int flags
,
1693 struct ceph_mds_session
*session
)
1695 struct ceph_fs_client
*fsc
= ceph_inode_to_client(&ci
->vfs_inode
);
1696 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
1697 struct inode
*inode
= &ci
->vfs_inode
;
1698 struct ceph_cap
*cap
;
1699 u64 flush_tid
, oldest_flush_tid
;
1700 int file_wanted
, used
, cap_used
;
1701 int took_snap_rwsem
= 0; /* true if mdsc->snap_rwsem held */
1702 int issued
, implemented
, want
, retain
, revoking
, flushing
= 0;
1703 int mds
= -1; /* keep track of how far we've gone through i_caps list
1704 to avoid an infinite loop on retry */
1706 int delayed
= 0, sent
= 0, num
;
1707 bool is_delayed
= flags
& CHECK_CAPS_NODELAY
;
1708 bool queue_invalidate
= false;
1709 bool force_requeue
= false;
1710 bool tried_invalidate
= false;
1712 /* if we are unmounting, flush any unused caps immediately. */
1716 spin_lock(&ci
->i_ceph_lock
);
1718 if (ci
->i_ceph_flags
& CEPH_I_FLUSH
)
1719 flags
|= CHECK_CAPS_FLUSH
;
1723 spin_lock(&ci
->i_ceph_lock
);
1725 file_wanted
= __ceph_caps_file_wanted(ci
);
1726 used
= __ceph_caps_used(ci
);
1727 issued
= __ceph_caps_issued(ci
, &implemented
);
1728 revoking
= implemented
& ~issued
;
1731 retain
= file_wanted
| used
| CEPH_CAP_PIN
;
1732 if (!mdsc
->stopping
&& inode
->i_nlink
> 0) {
1734 retain
|= CEPH_CAP_ANY
; /* be greedy */
1735 } else if (S_ISDIR(inode
->i_mode
) &&
1736 (issued
& CEPH_CAP_FILE_SHARED
) &&
1737 __ceph_dir_is_complete(ci
)) {
1739 * If a directory is complete, we want to keep
1740 * the exclusive cap. So that MDS does not end up
1741 * revoking the shared cap on every create/unlink
1744 want
= CEPH_CAP_ANY_SHARED
| CEPH_CAP_FILE_EXCL
;
1748 retain
|= CEPH_CAP_ANY_SHARED
;
1750 * keep RD only if we didn't have the file open RW,
1751 * because then the mds would revoke it anyway to
1752 * journal max_size=0.
1754 if (ci
->i_max_size
== 0)
1755 retain
|= CEPH_CAP_ANY_RD
;
1759 dout("check_caps %p file_want %s used %s dirty %s flushing %s"
1760 " issued %s revoking %s retain %s %s%s%s\n", inode
,
1761 ceph_cap_string(file_wanted
),
1762 ceph_cap_string(used
), ceph_cap_string(ci
->i_dirty_caps
),
1763 ceph_cap_string(ci
->i_flushing_caps
),
1764 ceph_cap_string(issued
), ceph_cap_string(revoking
),
1765 ceph_cap_string(retain
),
1766 (flags
& CHECK_CAPS_AUTHONLY
) ? " AUTHONLY" : "",
1767 (flags
& CHECK_CAPS_NODELAY
) ? " NODELAY" : "",
1768 (flags
& CHECK_CAPS_FLUSH
) ? " FLUSH" : "");
1771 * If we no longer need to hold onto old our caps, and we may
1772 * have cached pages, but don't want them, then try to invalidate.
1773 * If we fail, it's because pages are locked.... try again later.
1775 if ((!is_delayed
|| mdsc
->stopping
) &&
1776 !S_ISDIR(inode
->i_mode
) && /* ignore readdir cache */
1777 !(ci
->i_wb_ref
|| ci
->i_wrbuffer_ref
) && /* no dirty pages... */
1778 inode
->i_data
.nrpages
&& /* have cached pages */
1779 (revoking
& (CEPH_CAP_FILE_CACHE
|
1780 CEPH_CAP_FILE_LAZYIO
)) && /* or revoking cache */
1781 !tried_invalidate
) {
1782 dout("check_caps trying to invalidate on %p\n", inode
);
1783 if (try_nonblocking_invalidate(inode
) < 0) {
1784 if (revoking
& (CEPH_CAP_FILE_CACHE
|
1785 CEPH_CAP_FILE_LAZYIO
)) {
1786 dout("check_caps queuing invalidate\n");
1787 queue_invalidate
= true;
1788 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
;
1790 dout("check_caps failed to invalidate pages\n");
1791 /* we failed to invalidate pages. check these
1792 caps again later. */
1793 force_requeue
= true;
1794 __cap_set_timeouts(mdsc
, ci
);
1797 tried_invalidate
= true;
1802 for (p
= rb_first(&ci
->i_caps
); p
; p
= rb_next(p
)) {
1803 cap
= rb_entry(p
, struct ceph_cap
, ci_node
);
1806 /* avoid looping forever */
1807 if (mds
>= cap
->mds
||
1808 ((flags
& CHECK_CAPS_AUTHONLY
) && cap
!= ci
->i_auth_cap
))
1811 /* NOTE: no side-effects allowed, until we take s_mutex */
1814 if (ci
->i_auth_cap
&& cap
!= ci
->i_auth_cap
)
1815 cap_used
&= ~ci
->i_auth_cap
->issued
;
1817 revoking
= cap
->implemented
& ~cap
->issued
;
1818 dout(" mds%d cap %p used %s issued %s implemented %s revoking %s\n",
1819 cap
->mds
, cap
, ceph_cap_string(cap_used
),
1820 ceph_cap_string(cap
->issued
),
1821 ceph_cap_string(cap
->implemented
),
1822 ceph_cap_string(revoking
));
1824 if (cap
== ci
->i_auth_cap
&&
1825 (cap
->issued
& CEPH_CAP_FILE_WR
)) {
1826 /* request larger max_size from MDS? */
1827 if (ci
->i_wanted_max_size
> ci
->i_max_size
&&
1828 ci
->i_wanted_max_size
> ci
->i_requested_max_size
) {
1829 dout("requesting new max_size\n");
1833 /* approaching file_max? */
1834 if (__ceph_should_report_size(ci
)) {
1835 dout("i_size approaching max_size\n");
1839 /* flush anything dirty? */
1840 if (cap
== ci
->i_auth_cap
) {
1841 if ((flags
& CHECK_CAPS_FLUSH
) && ci
->i_dirty_caps
) {
1842 dout("flushing dirty caps\n");
1845 if (ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
) {
1846 dout("flushing snap caps\n");
1851 /* completed revocation? going down and there are no caps? */
1852 if (revoking
&& (revoking
& cap_used
) == 0) {
1853 dout("completed revocation of %s\n",
1854 ceph_cap_string(cap
->implemented
& ~cap
->issued
));
1858 /* want more caps from mds? */
1859 if (want
& ~(cap
->mds_wanted
| cap
->issued
))
1862 /* things we might delay */
1863 if ((cap
->issued
& ~retain
) == 0 &&
1864 cap
->mds_wanted
== want
)
1865 continue; /* nope, all good */
1871 if ((ci
->i_ceph_flags
& CEPH_I_NODELAY
) == 0 &&
1872 time_before(jiffies
, ci
->i_hold_caps_max
)) {
1873 dout(" delaying issued %s -> %s, wanted %s -> %s\n",
1874 ceph_cap_string(cap
->issued
),
1875 ceph_cap_string(cap
->issued
& retain
),
1876 ceph_cap_string(cap
->mds_wanted
),
1877 ceph_cap_string(want
));
1883 if (ci
->i_ceph_flags
& CEPH_I_NOFLUSH
) {
1884 dout(" skipping %p I_NOFLUSH set\n", inode
);
1888 if (session
&& session
!= cap
->session
) {
1889 dout("oops, wrong session %p mutex\n", session
);
1890 mutex_unlock(&session
->s_mutex
);
1894 session
= cap
->session
;
1895 if (mutex_trylock(&session
->s_mutex
) == 0) {
1896 dout("inverting session/ino locks on %p\n",
1898 spin_unlock(&ci
->i_ceph_lock
);
1899 if (took_snap_rwsem
) {
1900 up_read(&mdsc
->snap_rwsem
);
1901 took_snap_rwsem
= 0;
1903 mutex_lock(&session
->s_mutex
);
1908 /* kick flushing and flush snaps before sending normal
1910 if (cap
== ci
->i_auth_cap
&&
1912 (CEPH_I_KICK_FLUSH
| CEPH_I_FLUSH_SNAPS
))) {
1913 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
) {
1914 __kick_flushing_caps(mdsc
, session
, ci
, 0);
1915 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
1917 if (ci
->i_ceph_flags
& CEPH_I_FLUSH_SNAPS
)
1918 __ceph_flush_snaps(ci
, session
);
1923 /* take snap_rwsem after session mutex */
1924 if (!took_snap_rwsem
) {
1925 if (down_read_trylock(&mdsc
->snap_rwsem
) == 0) {
1926 dout("inverting snap/in locks on %p\n",
1928 spin_unlock(&ci
->i_ceph_lock
);
1929 down_read(&mdsc
->snap_rwsem
);
1930 took_snap_rwsem
= 1;
1933 took_snap_rwsem
= 1;
1936 if (cap
== ci
->i_auth_cap
&& ci
->i_dirty_caps
) {
1937 flushing
= __mark_caps_flushing(inode
, session
, false,
1943 spin_lock(&mdsc
->cap_dirty_lock
);
1944 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
1945 spin_unlock(&mdsc
->cap_dirty_lock
);
1948 mds
= cap
->mds
; /* remember mds, so we don't repeat */
1951 /* __send_cap drops i_ceph_lock */
1952 delayed
+= __send_cap(mdsc
, cap
, CEPH_CAP_OP_UPDATE
, false,
1953 cap_used
, want
, retain
, flushing
,
1954 flush_tid
, oldest_flush_tid
);
1955 goto retry
; /* retake i_ceph_lock and restart our cap scan. */
1959 * Reschedule delayed caps release if we delayed anything,
1962 if (delayed
&& is_delayed
)
1963 force_requeue
= true; /* __send_cap delayed release; requeue */
1964 if (!delayed
&& !is_delayed
)
1965 __cap_delay_cancel(mdsc
, ci
);
1966 else if (!is_delayed
|| force_requeue
)
1967 __cap_delay_requeue(mdsc
, ci
);
1969 spin_unlock(&ci
->i_ceph_lock
);
1971 if (queue_invalidate
)
1972 ceph_queue_invalidate(inode
);
1975 mutex_unlock(&session
->s_mutex
);
1976 if (took_snap_rwsem
)
1977 up_read(&mdsc
->snap_rwsem
);
1981 * Try to flush dirty caps back to the auth mds.
1983 static int try_flush_caps(struct inode
*inode
, u64
*ptid
)
1985 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
1986 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1987 struct ceph_mds_session
*session
= NULL
;
1989 u64 flush_tid
= 0, oldest_flush_tid
= 0;
1992 spin_lock(&ci
->i_ceph_lock
);
1993 if (ci
->i_ceph_flags
& CEPH_I_NOFLUSH
) {
1994 spin_unlock(&ci
->i_ceph_lock
);
1995 dout("try_flush_caps skipping %p I_NOFLUSH set\n", inode
);
1998 if (ci
->i_dirty_caps
&& ci
->i_auth_cap
) {
1999 struct ceph_cap
*cap
= ci
->i_auth_cap
;
2000 int used
= __ceph_caps_used(ci
);
2001 int want
= __ceph_caps_wanted(ci
);
2004 if (!session
|| session
!= cap
->session
) {
2005 spin_unlock(&ci
->i_ceph_lock
);
2007 mutex_unlock(&session
->s_mutex
);
2008 session
= cap
->session
;
2009 mutex_lock(&session
->s_mutex
);
2012 if (cap
->session
->s_state
< CEPH_MDS_SESSION_OPEN
) {
2013 spin_unlock(&ci
->i_ceph_lock
);
2017 flushing
= __mark_caps_flushing(inode
, session
, true,
2018 &flush_tid
, &oldest_flush_tid
);
2020 /* __send_cap drops i_ceph_lock */
2021 delayed
= __send_cap(mdsc
, cap
, CEPH_CAP_OP_FLUSH
, true,
2022 used
, want
, (cap
->issued
| cap
->implemented
),
2023 flushing
, flush_tid
, oldest_flush_tid
);
2026 spin_lock(&ci
->i_ceph_lock
);
2027 __cap_delay_requeue(mdsc
, ci
);
2028 spin_unlock(&ci
->i_ceph_lock
);
2031 if (!list_empty(&ci
->i_cap_flush_list
)) {
2032 struct ceph_cap_flush
*cf
=
2033 list_last_entry(&ci
->i_cap_flush_list
,
2034 struct ceph_cap_flush
, i_list
);
2036 flush_tid
= cf
->tid
;
2038 flushing
= ci
->i_flushing_caps
;
2039 spin_unlock(&ci
->i_ceph_lock
);
2043 mutex_unlock(&session
->s_mutex
);
2050 * Return true if we've flushed caps through the given flush_tid.
2052 static int caps_are_flushed(struct inode
*inode
, u64 flush_tid
)
2054 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2057 spin_lock(&ci
->i_ceph_lock
);
2058 if (!list_empty(&ci
->i_cap_flush_list
)) {
2059 struct ceph_cap_flush
* cf
=
2060 list_first_entry(&ci
->i_cap_flush_list
,
2061 struct ceph_cap_flush
, i_list
);
2062 if (cf
->tid
<= flush_tid
)
2065 spin_unlock(&ci
->i_ceph_lock
);
2070 * wait for any unsafe requests to complete.
2072 static int unsafe_request_wait(struct inode
*inode
)
2074 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2075 struct ceph_mds_request
*req1
= NULL
, *req2
= NULL
;
2078 spin_lock(&ci
->i_unsafe_lock
);
2079 if (S_ISDIR(inode
->i_mode
) && !list_empty(&ci
->i_unsafe_dirops
)) {
2080 req1
= list_last_entry(&ci
->i_unsafe_dirops
,
2081 struct ceph_mds_request
,
2083 ceph_mdsc_get_request(req1
);
2085 if (!list_empty(&ci
->i_unsafe_iops
)) {
2086 req2
= list_last_entry(&ci
->i_unsafe_iops
,
2087 struct ceph_mds_request
,
2088 r_unsafe_target_item
);
2089 ceph_mdsc_get_request(req2
);
2091 spin_unlock(&ci
->i_unsafe_lock
);
2093 dout("unsafe_request_wait %p wait on tid %llu %llu\n",
2094 inode
, req1
? req1
->r_tid
: 0ULL, req2
? req2
->r_tid
: 0ULL);
2096 ret
= !wait_for_completion_timeout(&req1
->r_safe_completion
,
2097 ceph_timeout_jiffies(req1
->r_timeout
));
2100 ceph_mdsc_put_request(req1
);
2103 ret
= !wait_for_completion_timeout(&req2
->r_safe_completion
,
2104 ceph_timeout_jiffies(req2
->r_timeout
));
2107 ceph_mdsc_put_request(req2
);
2112 int ceph_fsync(struct file
*file
, loff_t start
, loff_t end
, int datasync
)
2114 struct inode
*inode
= file
->f_mapping
->host
;
2115 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2120 dout("fsync %p%s\n", inode
, datasync
? " datasync" : "");
2122 ret
= file_write_and_wait_range(file
, start
, end
);
2131 dirty
= try_flush_caps(inode
, &flush_tid
);
2132 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty
));
2134 ret
= unsafe_request_wait(inode
);
2137 * only wait on non-file metadata writeback (the mds
2138 * can recover size and mtime, so we don't need to
2141 if (!ret
&& (dirty
& ~CEPH_CAP_ANY_FILE_WR
)) {
2142 ret
= wait_event_interruptible(ci
->i_cap_wq
,
2143 caps_are_flushed(inode
, flush_tid
));
2145 inode_unlock(inode
);
2147 dout("fsync %p%s result=%d\n", inode
, datasync
? " datasync" : "", ret
);
2152 * Flush any dirty caps back to the mds. If we aren't asked to wait,
2153 * queue inode for flush but don't do so immediately, because we can
2154 * get by with fewer MDS messages if we wait for data writeback to
2157 int ceph_write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
2159 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2163 int wait
= wbc
->sync_mode
== WB_SYNC_ALL
;
2165 dout("write_inode %p wait=%d\n", inode
, wait
);
2167 dirty
= try_flush_caps(inode
, &flush_tid
);
2169 err
= wait_event_interruptible(ci
->i_cap_wq
,
2170 caps_are_flushed(inode
, flush_tid
));
2172 struct ceph_mds_client
*mdsc
=
2173 ceph_sb_to_client(inode
->i_sb
)->mdsc
;
2175 spin_lock(&ci
->i_ceph_lock
);
2176 if (__ceph_caps_dirty(ci
))
2177 __cap_delay_requeue_front(mdsc
, ci
);
2178 spin_unlock(&ci
->i_ceph_lock
);
2183 static void __kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2184 struct ceph_mds_session
*session
,
2185 struct ceph_inode_info
*ci
,
2186 u64 oldest_flush_tid
)
2187 __releases(ci
->i_ceph_lock
)
2188 __acquires(ci
->i_ceph_lock
)
2190 struct inode
*inode
= &ci
->vfs_inode
;
2191 struct ceph_cap
*cap
;
2192 struct ceph_cap_flush
*cf
;
2196 list_for_each_entry(cf
, &ci
->i_cap_flush_list
, i_list
) {
2197 if (cf
->tid
< first_tid
)
2200 cap
= ci
->i_auth_cap
;
2201 if (!(cap
&& cap
->session
== session
)) {
2202 pr_err("%p auth cap %p not mds%d ???\n",
2203 inode
, cap
, session
->s_mds
);
2207 first_tid
= cf
->tid
+ 1;
2210 dout("kick_flushing_caps %p cap %p tid %llu %s\n",
2211 inode
, cap
, cf
->tid
, ceph_cap_string(cf
->caps
));
2212 ci
->i_ceph_flags
|= CEPH_I_NODELAY
;
2213 ret
= __send_cap(mdsc
, cap
, CEPH_CAP_OP_FLUSH
,
2214 false, __ceph_caps_used(ci
),
2215 __ceph_caps_wanted(ci
),
2216 cap
->issued
| cap
->implemented
,
2217 cf
->caps
, cf
->tid
, oldest_flush_tid
);
2219 pr_err("kick_flushing_caps: error sending "
2220 "cap flush, ino (%llx.%llx) "
2221 "tid %llu flushing %s\n",
2222 ceph_vinop(inode
), cf
->tid
,
2223 ceph_cap_string(cf
->caps
));
2226 struct ceph_cap_snap
*capsnap
=
2227 container_of(cf
, struct ceph_cap_snap
,
2229 dout("kick_flushing_caps %p capsnap %p tid %llu %s\n",
2230 inode
, capsnap
, cf
->tid
,
2231 ceph_cap_string(capsnap
->dirty
));
2233 refcount_inc(&capsnap
->nref
);
2234 spin_unlock(&ci
->i_ceph_lock
);
2236 ret
= __send_flush_snap(inode
, session
, capsnap
, cap
->mseq
,
2239 pr_err("kick_flushing_caps: error sending "
2240 "cap flushsnap, ino (%llx.%llx) "
2241 "tid %llu follows %llu\n",
2242 ceph_vinop(inode
), cf
->tid
,
2246 ceph_put_cap_snap(capsnap
);
2249 spin_lock(&ci
->i_ceph_lock
);
2253 void ceph_early_kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2254 struct ceph_mds_session
*session
)
2256 struct ceph_inode_info
*ci
;
2257 struct ceph_cap
*cap
;
2258 u64 oldest_flush_tid
;
2260 dout("early_kick_flushing_caps mds%d\n", session
->s_mds
);
2262 spin_lock(&mdsc
->cap_dirty_lock
);
2263 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2264 spin_unlock(&mdsc
->cap_dirty_lock
);
2266 list_for_each_entry(ci
, &session
->s_cap_flushing
, i_flushing_item
) {
2267 spin_lock(&ci
->i_ceph_lock
);
2268 cap
= ci
->i_auth_cap
;
2269 if (!(cap
&& cap
->session
== session
)) {
2270 pr_err("%p auth cap %p not mds%d ???\n",
2271 &ci
->vfs_inode
, cap
, session
->s_mds
);
2272 spin_unlock(&ci
->i_ceph_lock
);
2278 * if flushing caps were revoked, we re-send the cap flush
2279 * in client reconnect stage. This guarantees MDS * processes
2280 * the cap flush message before issuing the flushing caps to
2283 if ((cap
->issued
& ci
->i_flushing_caps
) !=
2284 ci
->i_flushing_caps
) {
2285 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
2286 __kick_flushing_caps(mdsc
, session
, ci
,
2289 ci
->i_ceph_flags
|= CEPH_I_KICK_FLUSH
;
2292 spin_unlock(&ci
->i_ceph_lock
);
2296 void ceph_kick_flushing_caps(struct ceph_mds_client
*mdsc
,
2297 struct ceph_mds_session
*session
)
2299 struct ceph_inode_info
*ci
;
2300 struct ceph_cap
*cap
;
2301 u64 oldest_flush_tid
;
2303 dout("kick_flushing_caps mds%d\n", session
->s_mds
);
2305 spin_lock(&mdsc
->cap_dirty_lock
);
2306 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2307 spin_unlock(&mdsc
->cap_dirty_lock
);
2309 list_for_each_entry(ci
, &session
->s_cap_flushing
, i_flushing_item
) {
2310 spin_lock(&ci
->i_ceph_lock
);
2311 cap
= ci
->i_auth_cap
;
2312 if (!(cap
&& cap
->session
== session
)) {
2313 pr_err("%p auth cap %p not mds%d ???\n",
2314 &ci
->vfs_inode
, cap
, session
->s_mds
);
2315 spin_unlock(&ci
->i_ceph_lock
);
2318 if (ci
->i_ceph_flags
& CEPH_I_KICK_FLUSH
) {
2319 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
2320 __kick_flushing_caps(mdsc
, session
, ci
,
2323 spin_unlock(&ci
->i_ceph_lock
);
2327 static void kick_flushing_inode_caps(struct ceph_mds_client
*mdsc
,
2328 struct ceph_mds_session
*session
,
2329 struct inode
*inode
)
2330 __releases(ci
->i_ceph_lock
)
2332 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2333 struct ceph_cap
*cap
;
2335 cap
= ci
->i_auth_cap
;
2336 dout("kick_flushing_inode_caps %p flushing %s\n", inode
,
2337 ceph_cap_string(ci
->i_flushing_caps
));
2339 if (!list_empty(&ci
->i_cap_flush_list
)) {
2340 u64 oldest_flush_tid
;
2341 spin_lock(&mdsc
->cap_dirty_lock
);
2342 list_move_tail(&ci
->i_flushing_item
,
2343 &cap
->session
->s_cap_flushing
);
2344 oldest_flush_tid
= __get_oldest_flush_tid(mdsc
);
2345 spin_unlock(&mdsc
->cap_dirty_lock
);
2347 ci
->i_ceph_flags
&= ~CEPH_I_KICK_FLUSH
;
2348 __kick_flushing_caps(mdsc
, session
, ci
, oldest_flush_tid
);
2349 spin_unlock(&ci
->i_ceph_lock
);
2351 spin_unlock(&ci
->i_ceph_lock
);
2357 * Take references to capabilities we hold, so that we don't release
2358 * them to the MDS prematurely.
2360 * Protected by i_ceph_lock.
2362 static void __take_cap_refs(struct ceph_inode_info
*ci
, int got
,
2363 bool snap_rwsem_locked
)
2365 if (got
& CEPH_CAP_PIN
)
2367 if (got
& CEPH_CAP_FILE_RD
)
2369 if (got
& CEPH_CAP_FILE_CACHE
)
2370 ci
->i_rdcache_ref
++;
2371 if (got
& CEPH_CAP_FILE_WR
) {
2372 if (ci
->i_wr_ref
== 0 && !ci
->i_head_snapc
) {
2373 BUG_ON(!snap_rwsem_locked
);
2374 ci
->i_head_snapc
= ceph_get_snap_context(
2375 ci
->i_snap_realm
->cached_context
);
2379 if (got
& CEPH_CAP_FILE_BUFFER
) {
2380 if (ci
->i_wb_ref
== 0)
2381 ihold(&ci
->vfs_inode
);
2383 dout("__take_cap_refs %p wb %d -> %d (?)\n",
2384 &ci
->vfs_inode
, ci
->i_wb_ref
-1, ci
->i_wb_ref
);
2389 * Try to grab cap references. Specify those refs we @want, and the
2390 * minimal set we @need. Also include the larger offset we are writing
2391 * to (when applicable), and check against max_size here as well.
2392 * Note that caller is responsible for ensuring max_size increases are
2393 * requested from the MDS.
2395 static int try_get_cap_refs(struct ceph_inode_info
*ci
, int need
, int want
,
2396 loff_t endoff
, bool nonblock
, int *got
, int *err
)
2398 struct inode
*inode
= &ci
->vfs_inode
;
2399 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
2401 int have
, implemented
;
2403 bool snap_rwsem_locked
= false;
2405 dout("get_cap_refs %p need %s want %s\n", inode
,
2406 ceph_cap_string(need
), ceph_cap_string(want
));
2409 spin_lock(&ci
->i_ceph_lock
);
2411 /* make sure file is actually open */
2412 file_wanted
= __ceph_caps_file_wanted(ci
);
2413 if ((file_wanted
& need
) != need
) {
2414 dout("try_get_cap_refs need %s file_wanted %s, EBADF\n",
2415 ceph_cap_string(need
), ceph_cap_string(file_wanted
));
2421 /* finish pending truncate */
2422 while (ci
->i_truncate_pending
) {
2423 spin_unlock(&ci
->i_ceph_lock
);
2424 if (snap_rwsem_locked
) {
2425 up_read(&mdsc
->snap_rwsem
);
2426 snap_rwsem_locked
= false;
2428 __ceph_do_pending_vmtruncate(inode
);
2429 spin_lock(&ci
->i_ceph_lock
);
2432 have
= __ceph_caps_issued(ci
, &implemented
);
2434 if (have
& need
& CEPH_CAP_FILE_WR
) {
2435 if (endoff
>= 0 && endoff
> (loff_t
)ci
->i_max_size
) {
2436 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2437 inode
, endoff
, ci
->i_max_size
);
2438 if (endoff
> ci
->i_requested_max_size
) {
2445 * If a sync write is in progress, we must wait, so that we
2446 * can get a final snapshot value for size+mtime.
2448 if (__ceph_have_pending_cap_snap(ci
)) {
2449 dout("get_cap_refs %p cap_snap_pending\n", inode
);
2454 if ((have
& need
) == need
) {
2456 * Look at (implemented & ~have & not) so that we keep waiting
2457 * on transition from wanted -> needed caps. This is needed
2458 * for WRBUFFER|WR -> WR to avoid a new WR sync write from
2459 * going before a prior buffered writeback happens.
2461 int not = want
& ~(have
& need
);
2462 int revoking
= implemented
& ~have
;
2463 dout("get_cap_refs %p have %s but not %s (revoking %s)\n",
2464 inode
, ceph_cap_string(have
), ceph_cap_string(not),
2465 ceph_cap_string(revoking
));
2466 if ((revoking
& not) == 0) {
2467 if (!snap_rwsem_locked
&&
2468 !ci
->i_head_snapc
&&
2469 (need
& CEPH_CAP_FILE_WR
)) {
2470 if (!down_read_trylock(&mdsc
->snap_rwsem
)) {
2472 * we can not call down_read() when
2473 * task isn't in TASK_RUNNING state
2481 spin_unlock(&ci
->i_ceph_lock
);
2482 down_read(&mdsc
->snap_rwsem
);
2483 snap_rwsem_locked
= true;
2486 snap_rwsem_locked
= true;
2488 *got
= need
| (have
& want
);
2489 if ((need
& CEPH_CAP_FILE_RD
) &&
2490 !(*got
& CEPH_CAP_FILE_CACHE
))
2491 ceph_disable_fscache_readpage(ci
);
2492 __take_cap_refs(ci
, *got
, true);
2496 int session_readonly
= false;
2497 if ((need
& CEPH_CAP_FILE_WR
) && ci
->i_auth_cap
) {
2498 struct ceph_mds_session
*s
= ci
->i_auth_cap
->session
;
2499 spin_lock(&s
->s_cap_lock
);
2500 session_readonly
= s
->s_readonly
;
2501 spin_unlock(&s
->s_cap_lock
);
2503 if (session_readonly
) {
2504 dout("get_cap_refs %p needed %s but mds%d readonly\n",
2505 inode
, ceph_cap_string(need
), ci
->i_auth_cap
->mds
);
2511 if (ci
->i_ceph_flags
& CEPH_I_CAP_DROPPED
) {
2513 if (READ_ONCE(mdsc
->fsc
->mount_state
) ==
2514 CEPH_MOUNT_SHUTDOWN
) {
2515 dout("get_cap_refs %p forced umount\n", inode
);
2520 mds_wanted
= __ceph_caps_mds_wanted(ci
, false);
2521 if (need
& ~(mds_wanted
& need
)) {
2522 dout("get_cap_refs %p caps were dropped"
2523 " (session killed?)\n", inode
);
2528 if (!(file_wanted
& ~mds_wanted
))
2529 ci
->i_ceph_flags
&= ~CEPH_I_CAP_DROPPED
;
2532 dout("get_cap_refs %p have %s needed %s\n", inode
,
2533 ceph_cap_string(have
), ceph_cap_string(need
));
2536 spin_unlock(&ci
->i_ceph_lock
);
2537 if (snap_rwsem_locked
)
2538 up_read(&mdsc
->snap_rwsem
);
2540 dout("get_cap_refs %p ret %d got %s\n", inode
,
2541 ret
, ceph_cap_string(*got
));
2546 * Check the offset we are writing up to against our current
2547 * max_size. If necessary, tell the MDS we want to write to
2550 static void check_max_size(struct inode
*inode
, loff_t endoff
)
2552 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2555 /* do we need to explicitly request a larger max_size? */
2556 spin_lock(&ci
->i_ceph_lock
);
2557 if (endoff
>= ci
->i_max_size
&& endoff
> ci
->i_wanted_max_size
) {
2558 dout("write %p at large endoff %llu, req max_size\n",
2560 ci
->i_wanted_max_size
= endoff
;
2562 /* duplicate ceph_check_caps()'s logic */
2563 if (ci
->i_auth_cap
&&
2564 (ci
->i_auth_cap
->issued
& CEPH_CAP_FILE_WR
) &&
2565 ci
->i_wanted_max_size
> ci
->i_max_size
&&
2566 ci
->i_wanted_max_size
> ci
->i_requested_max_size
)
2568 spin_unlock(&ci
->i_ceph_lock
);
2570 ceph_check_caps(ci
, CHECK_CAPS_AUTHONLY
, NULL
);
2573 int ceph_try_get_caps(struct ceph_inode_info
*ci
, int need
, int want
, int *got
)
2577 BUG_ON(need
& ~CEPH_CAP_FILE_RD
);
2578 BUG_ON(want
& ~(CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
));
2579 ret
= ceph_pool_perm_check(ci
, need
);
2583 ret
= try_get_cap_refs(ci
, need
, want
, 0, true, got
, &err
);
2585 if (err
== -EAGAIN
) {
2587 } else if (err
< 0) {
2595 * Wait for caps, and take cap references. If we can't get a WR cap
2596 * due to a small max_size, make sure we check_max_size (and possibly
2597 * ask the mds) so we don't get hung up indefinitely.
2599 int ceph_get_caps(struct ceph_inode_info
*ci
, int need
, int want
,
2600 loff_t endoff
, int *got
, struct page
**pinned_page
)
2602 int _got
, ret
, err
= 0;
2604 ret
= ceph_pool_perm_check(ci
, need
);
2610 check_max_size(&ci
->vfs_inode
, endoff
);
2614 ret
= try_get_cap_refs(ci
, need
, want
, endoff
,
2615 false, &_got
, &err
);
2622 DEFINE_WAIT_FUNC(wait
, woken_wake_function
);
2623 add_wait_queue(&ci
->i_cap_wq
, &wait
);
2625 while (!try_get_cap_refs(ci
, need
, want
, endoff
,
2626 true, &_got
, &err
)) {
2627 if (signal_pending(current
)) {
2631 wait_woken(&wait
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
2634 remove_wait_queue(&ci
->i_cap_wq
, &wait
);
2642 if (err
== -ESTALE
) {
2643 /* session was killed, try renew caps */
2644 ret
= ceph_renew_caps(&ci
->vfs_inode
);
2651 if (ci
->i_inline_version
!= CEPH_INLINE_NONE
&&
2652 (_got
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)) &&
2653 i_size_read(&ci
->vfs_inode
) > 0) {
2655 find_get_page(ci
->vfs_inode
.i_mapping
, 0);
2657 if (PageUptodate(page
)) {
2658 *pinned_page
= page
;
2664 * drop cap refs first because getattr while
2665 * holding * caps refs can cause deadlock.
2667 ceph_put_cap_refs(ci
, _got
);
2671 * getattr request will bring inline data into
2674 ret
= __ceph_do_getattr(&ci
->vfs_inode
, NULL
,
2675 CEPH_STAT_CAP_INLINE_DATA
,
2684 if ((_got
& CEPH_CAP_FILE_RD
) && (_got
& CEPH_CAP_FILE_CACHE
))
2685 ceph_fscache_revalidate_cookie(ci
);
2692 * Take cap refs. Caller must already know we hold at least one ref
2693 * on the caps in question or we don't know this is safe.
2695 void ceph_get_cap_refs(struct ceph_inode_info
*ci
, int caps
)
2697 spin_lock(&ci
->i_ceph_lock
);
2698 __take_cap_refs(ci
, caps
, false);
2699 spin_unlock(&ci
->i_ceph_lock
);
2704 * drop cap_snap that is not associated with any snapshot.
2705 * we don't need to send FLUSHSNAP message for it.
2707 static int ceph_try_drop_cap_snap(struct ceph_inode_info
*ci
,
2708 struct ceph_cap_snap
*capsnap
)
2710 if (!capsnap
->need_flush
&&
2711 !capsnap
->writing
&& !capsnap
->dirty_pages
) {
2712 dout("dropping cap_snap %p follows %llu\n",
2713 capsnap
, capsnap
->follows
);
2714 BUG_ON(capsnap
->cap_flush
.tid
> 0);
2715 ceph_put_snap_context(capsnap
->context
);
2716 if (!list_is_last(&capsnap
->ci_item
, &ci
->i_cap_snaps
))
2717 ci
->i_ceph_flags
|= CEPH_I_FLUSH_SNAPS
;
2719 list_del(&capsnap
->ci_item
);
2720 ceph_put_cap_snap(capsnap
);
2729 * If we released the last ref on any given cap, call ceph_check_caps
2730 * to release (or schedule a release).
2732 * If we are releasing a WR cap (from a sync write), finalize any affected
2733 * cap_snap, and wake up any waiters.
2735 void ceph_put_cap_refs(struct ceph_inode_info
*ci
, int had
)
2737 struct inode
*inode
= &ci
->vfs_inode
;
2738 int last
= 0, put
= 0, flushsnaps
= 0, wake
= 0;
2740 spin_lock(&ci
->i_ceph_lock
);
2741 if (had
& CEPH_CAP_PIN
)
2743 if (had
& CEPH_CAP_FILE_RD
)
2744 if (--ci
->i_rd_ref
== 0)
2746 if (had
& CEPH_CAP_FILE_CACHE
)
2747 if (--ci
->i_rdcache_ref
== 0)
2749 if (had
& CEPH_CAP_FILE_BUFFER
) {
2750 if (--ci
->i_wb_ref
== 0) {
2754 dout("put_cap_refs %p wb %d -> %d (?)\n",
2755 inode
, ci
->i_wb_ref
+1, ci
->i_wb_ref
);
2757 if (had
& CEPH_CAP_FILE_WR
)
2758 if (--ci
->i_wr_ref
== 0) {
2760 if (__ceph_have_pending_cap_snap(ci
)) {
2761 struct ceph_cap_snap
*capsnap
=
2762 list_last_entry(&ci
->i_cap_snaps
,
2763 struct ceph_cap_snap
,
2765 capsnap
->writing
= 0;
2766 if (ceph_try_drop_cap_snap(ci
, capsnap
))
2768 else if (__ceph_finish_cap_snap(ci
, capsnap
))
2772 if (ci
->i_wrbuffer_ref_head
== 0 &&
2773 ci
->i_dirty_caps
== 0 &&
2774 ci
->i_flushing_caps
== 0) {
2775 BUG_ON(!ci
->i_head_snapc
);
2776 ceph_put_snap_context(ci
->i_head_snapc
);
2777 ci
->i_head_snapc
= NULL
;
2779 /* see comment in __ceph_remove_cap() */
2780 if (!__ceph_is_any_caps(ci
) && ci
->i_snap_realm
)
2781 drop_inode_snap_realm(ci
);
2783 spin_unlock(&ci
->i_ceph_lock
);
2785 dout("put_cap_refs %p had %s%s%s\n", inode
, ceph_cap_string(had
),
2786 last
? " last" : "", put
? " put" : "");
2788 if (last
&& !flushsnaps
)
2789 ceph_check_caps(ci
, 0, NULL
);
2790 else if (flushsnaps
)
2791 ceph_flush_snaps(ci
, NULL
);
2793 wake_up_all(&ci
->i_cap_wq
);
2799 * Release @nr WRBUFFER refs on dirty pages for the given @snapc snap
2800 * context. Adjust per-snap dirty page accounting as appropriate.
2801 * Once all dirty data for a cap_snap is flushed, flush snapped file
2802 * metadata back to the MDS. If we dropped the last ref, call
2805 void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info
*ci
, int nr
,
2806 struct ceph_snap_context
*snapc
)
2808 struct inode
*inode
= &ci
->vfs_inode
;
2809 struct ceph_cap_snap
*capsnap
= NULL
;
2813 bool flush_snaps
= false;
2814 bool complete_capsnap
= false;
2816 spin_lock(&ci
->i_ceph_lock
);
2817 ci
->i_wrbuffer_ref
-= nr
;
2818 if (ci
->i_wrbuffer_ref
== 0) {
2823 if (ci
->i_head_snapc
== snapc
) {
2824 ci
->i_wrbuffer_ref_head
-= nr
;
2825 if (ci
->i_wrbuffer_ref_head
== 0 &&
2826 ci
->i_wr_ref
== 0 &&
2827 ci
->i_dirty_caps
== 0 &&
2828 ci
->i_flushing_caps
== 0) {
2829 BUG_ON(!ci
->i_head_snapc
);
2830 ceph_put_snap_context(ci
->i_head_snapc
);
2831 ci
->i_head_snapc
= NULL
;
2833 dout("put_wrbuffer_cap_refs on %p head %d/%d -> %d/%d %s\n",
2835 ci
->i_wrbuffer_ref
+nr
, ci
->i_wrbuffer_ref_head
+nr
,
2836 ci
->i_wrbuffer_ref
, ci
->i_wrbuffer_ref_head
,
2837 last
? " LAST" : "");
2839 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
2840 if (capsnap
->context
== snapc
) {
2846 capsnap
->dirty_pages
-= nr
;
2847 if (capsnap
->dirty_pages
== 0) {
2848 complete_capsnap
= true;
2849 if (!capsnap
->writing
) {
2850 if (ceph_try_drop_cap_snap(ci
, capsnap
)) {
2853 ci
->i_ceph_flags
|= CEPH_I_FLUSH_SNAPS
;
2858 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2859 " snap %lld %d/%d -> %d/%d %s%s\n",
2860 inode
, capsnap
, capsnap
->context
->seq
,
2861 ci
->i_wrbuffer_ref
+nr
, capsnap
->dirty_pages
+ nr
,
2862 ci
->i_wrbuffer_ref
, capsnap
->dirty_pages
,
2863 last
? " (wrbuffer last)" : "",
2864 complete_capsnap
? " (complete capsnap)" : "");
2867 spin_unlock(&ci
->i_ceph_lock
);
2870 ceph_check_caps(ci
, CHECK_CAPS_AUTHONLY
, NULL
);
2871 } else if (flush_snaps
) {
2872 ceph_flush_snaps(ci
, NULL
);
2874 if (complete_capsnap
)
2875 wake_up_all(&ci
->i_cap_wq
);
2881 * Invalidate unlinked inode's aliases, so we can drop the inode ASAP.
2883 static void invalidate_aliases(struct inode
*inode
)
2885 struct dentry
*dn
, *prev
= NULL
;
2887 dout("invalidate_aliases inode %p\n", inode
);
2888 d_prune_aliases(inode
);
2890 * For non-directory inode, d_find_alias() only returns
2891 * hashed dentry. After calling d_invalidate(), the
2892 * dentry becomes unhashed.
2894 * For directory inode, d_find_alias() can return
2895 * unhashed dentry. But directory inode should have
2896 * one alias at most.
2898 while ((dn
= d_find_alias(inode
))) {
2913 * Handle a cap GRANT message from the MDS. (Note that a GRANT may
2914 * actually be a revocation if it specifies a smaller cap set.)
2916 * caller holds s_mutex and i_ceph_lock, we drop both.
2918 static void handle_cap_grant(struct ceph_mds_client
*mdsc
,
2919 struct inode
*inode
, struct ceph_mds_caps
*grant
,
2920 struct ceph_string
**pns
, u64 inline_version
,
2921 void *inline_data
, u32 inline_len
,
2922 struct ceph_buffer
*xattr_buf
,
2923 struct ceph_mds_session
*session
,
2924 struct ceph_cap
*cap
, int issued
)
2925 __releases(ci
->i_ceph_lock
)
2926 __releases(mdsc
->snap_rwsem
)
2928 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2929 int mds
= session
->s_mds
;
2930 int seq
= le32_to_cpu(grant
->seq
);
2931 int newcaps
= le32_to_cpu(grant
->caps
);
2932 int used
, wanted
, dirty
;
2933 u64 size
= le64_to_cpu(grant
->size
);
2934 u64 max_size
= le64_to_cpu(grant
->max_size
);
2935 struct timespec mtime
, atime
, ctime
;
2938 bool writeback
= false;
2939 bool queue_trunc
= false;
2940 bool queue_invalidate
= false;
2941 bool deleted_inode
= false;
2942 bool fill_inline
= false;
2944 dout("handle_cap_grant inode %p cap %p mds%d seq %d %s\n",
2945 inode
, cap
, mds
, seq
, ceph_cap_string(newcaps
));
2946 dout(" size %llu max_size %llu, i_size %llu\n", size
, max_size
,
2951 * auth mds of the inode changed. we received the cap export message,
2952 * but still haven't received the cap import message. handle_cap_export
2953 * updated the new auth MDS' cap.
2955 * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing a message
2956 * that was sent before the cap import message. So don't remove caps.
2958 if (ceph_seq_cmp(seq
, cap
->seq
) <= 0) {
2959 WARN_ON(cap
!= ci
->i_auth_cap
);
2960 WARN_ON(cap
->cap_id
!= le64_to_cpu(grant
->cap_id
));
2962 newcaps
|= cap
->issued
;
2966 * If CACHE is being revoked, and we have no dirty buffers,
2967 * try to invalidate (once). (If there are dirty buffers, we
2968 * will invalidate _after_ writeback.)
2970 if (!S_ISDIR(inode
->i_mode
) && /* don't invalidate readdir cache */
2971 ((cap
->issued
& ~newcaps
) & CEPH_CAP_FILE_CACHE
) &&
2972 (newcaps
& CEPH_CAP_FILE_LAZYIO
) == 0 &&
2973 !(ci
->i_wrbuffer_ref
|| ci
->i_wb_ref
)) {
2974 if (try_nonblocking_invalidate(inode
)) {
2975 /* there were locked pages.. invalidate later
2976 in a separate thread. */
2977 if (ci
->i_rdcache_revoking
!= ci
->i_rdcache_gen
) {
2978 queue_invalidate
= true;
2979 ci
->i_rdcache_revoking
= ci
->i_rdcache_gen
;
2984 /* side effects now are allowed */
2985 cap
->cap_gen
= session
->s_cap_gen
;
2988 __check_cap_issue(ci
, cap
, newcaps
);
2990 if ((newcaps
& CEPH_CAP_AUTH_SHARED
) &&
2991 (issued
& CEPH_CAP_AUTH_EXCL
) == 0) {
2992 inode
->i_mode
= le32_to_cpu(grant
->mode
);
2993 inode
->i_uid
= make_kuid(&init_user_ns
, le32_to_cpu(grant
->uid
));
2994 inode
->i_gid
= make_kgid(&init_user_ns
, le32_to_cpu(grant
->gid
));
2995 dout("%p mode 0%o uid.gid %d.%d\n", inode
, inode
->i_mode
,
2996 from_kuid(&init_user_ns
, inode
->i_uid
),
2997 from_kgid(&init_user_ns
, inode
->i_gid
));
3000 if ((newcaps
& CEPH_CAP_AUTH_SHARED
) &&
3001 (issued
& CEPH_CAP_LINK_EXCL
) == 0) {
3002 set_nlink(inode
, le32_to_cpu(grant
->nlink
));
3003 if (inode
->i_nlink
== 0 &&
3004 (newcaps
& (CEPH_CAP_LINK_SHARED
| CEPH_CAP_LINK_EXCL
)))
3005 deleted_inode
= true;
3008 if ((issued
& CEPH_CAP_XATTR_EXCL
) == 0 && grant
->xattr_len
) {
3009 int len
= le32_to_cpu(grant
->xattr_len
);
3010 u64 version
= le64_to_cpu(grant
->xattr_version
);
3012 if (version
> ci
->i_xattrs
.version
) {
3013 dout(" got new xattrs v%llu on %p len %d\n",
3014 version
, inode
, len
);
3015 if (ci
->i_xattrs
.blob
)
3016 ceph_buffer_put(ci
->i_xattrs
.blob
);
3017 ci
->i_xattrs
.blob
= ceph_buffer_get(xattr_buf
);
3018 ci
->i_xattrs
.version
= version
;
3019 ceph_forget_all_cached_acls(inode
);
3023 if (newcaps
& CEPH_CAP_ANY_RD
) {
3024 /* ctime/mtime/atime? */
3025 ceph_decode_timespec(&mtime
, &grant
->mtime
);
3026 ceph_decode_timespec(&atime
, &grant
->atime
);
3027 ceph_decode_timespec(&ctime
, &grant
->ctime
);
3028 ceph_fill_file_time(inode
, issued
,
3029 le32_to_cpu(grant
->time_warp_seq
),
3030 &ctime
, &mtime
, &atime
);
3033 if (newcaps
& (CEPH_CAP_ANY_FILE_RD
| CEPH_CAP_ANY_FILE_WR
)) {
3034 /* file layout may have changed */
3035 s64 old_pool
= ci
->i_layout
.pool_id
;
3036 struct ceph_string
*old_ns
;
3038 ceph_file_layout_from_legacy(&ci
->i_layout
, &grant
->layout
);
3039 old_ns
= rcu_dereference_protected(ci
->i_layout
.pool_ns
,
3040 lockdep_is_held(&ci
->i_ceph_lock
));
3041 rcu_assign_pointer(ci
->i_layout
.pool_ns
, *pns
);
3043 if (ci
->i_layout
.pool_id
!= old_pool
|| *pns
!= old_ns
)
3044 ci
->i_ceph_flags
&= ~CEPH_I_POOL_PERM
;
3048 /* size/truncate_seq? */
3049 queue_trunc
= ceph_fill_file_size(inode
, issued
,
3050 le32_to_cpu(grant
->truncate_seq
),
3051 le64_to_cpu(grant
->truncate_size
),
3055 if (ci
->i_auth_cap
== cap
&& (newcaps
& CEPH_CAP_ANY_FILE_WR
)) {
3056 if (max_size
!= ci
->i_max_size
) {
3057 dout("max_size %lld -> %llu\n",
3058 ci
->i_max_size
, max_size
);
3059 ci
->i_max_size
= max_size
;
3060 if (max_size
>= ci
->i_wanted_max_size
) {
3061 ci
->i_wanted_max_size
= 0; /* reset */
3062 ci
->i_requested_max_size
= 0;
3065 } else if (ci
->i_wanted_max_size
> ci
->i_max_size
&&
3066 ci
->i_wanted_max_size
> ci
->i_requested_max_size
) {
3067 /* CEPH_CAP_OP_IMPORT */
3072 /* check cap bits */
3073 wanted
= __ceph_caps_wanted(ci
);
3074 used
= __ceph_caps_used(ci
);
3075 dirty
= __ceph_caps_dirty(ci
);
3076 dout(" my wanted = %s, used = %s, dirty %s\n",
3077 ceph_cap_string(wanted
),
3078 ceph_cap_string(used
),
3079 ceph_cap_string(dirty
));
3080 if (wanted
!= le32_to_cpu(grant
->wanted
)) {
3081 dout("mds wanted %s -> %s\n",
3082 ceph_cap_string(le32_to_cpu(grant
->wanted
)),
3083 ceph_cap_string(wanted
));
3084 /* imported cap may not have correct mds_wanted */
3085 if (le32_to_cpu(grant
->op
) == CEPH_CAP_OP_IMPORT
)
3089 /* revocation, grant, or no-op? */
3090 if (cap
->issued
& ~newcaps
) {
3091 int revoking
= cap
->issued
& ~newcaps
;
3093 dout("revocation: %s -> %s (revoking %s)\n",
3094 ceph_cap_string(cap
->issued
),
3095 ceph_cap_string(newcaps
),
3096 ceph_cap_string(revoking
));
3097 if (revoking
& used
& CEPH_CAP_FILE_BUFFER
)
3098 writeback
= true; /* initiate writeback; will delay ack */
3099 else if (revoking
== CEPH_CAP_FILE_CACHE
&&
3100 (newcaps
& CEPH_CAP_FILE_LAZYIO
) == 0 &&
3102 ; /* do nothing yet, invalidation will be queued */
3103 else if (cap
== ci
->i_auth_cap
)
3104 check_caps
= 1; /* check auth cap only */
3106 check_caps
= 2; /* check all caps */
3107 cap
->issued
= newcaps
;
3108 cap
->implemented
|= newcaps
;
3109 } else if (cap
->issued
== newcaps
) {
3110 dout("caps unchanged: %s -> %s\n",
3111 ceph_cap_string(cap
->issued
), ceph_cap_string(newcaps
));
3113 dout("grant: %s -> %s\n", ceph_cap_string(cap
->issued
),
3114 ceph_cap_string(newcaps
));
3115 /* non-auth MDS is revoking the newly grant caps ? */
3116 if (cap
== ci
->i_auth_cap
&&
3117 __ceph_caps_revoking_other(ci
, cap
, newcaps
))
3120 cap
->issued
= newcaps
;
3121 cap
->implemented
|= newcaps
; /* add bits only, to
3122 * avoid stepping on a
3123 * pending revocation */
3126 BUG_ON(cap
->issued
& ~cap
->implemented
);
3128 if (inline_version
> 0 && inline_version
>= ci
->i_inline_version
) {
3129 ci
->i_inline_version
= inline_version
;
3130 if (ci
->i_inline_version
!= CEPH_INLINE_NONE
&&
3131 (newcaps
& (CEPH_CAP_FILE_CACHE
|CEPH_CAP_FILE_LAZYIO
)))
3135 if (le32_to_cpu(grant
->op
) == CEPH_CAP_OP_IMPORT
) {
3136 if (newcaps
& ~issued
)
3138 kick_flushing_inode_caps(mdsc
, session
, inode
);
3139 up_read(&mdsc
->snap_rwsem
);
3141 spin_unlock(&ci
->i_ceph_lock
);
3145 ceph_fill_inline_data(inode
, NULL
, inline_data
, inline_len
);
3148 ceph_queue_vmtruncate(inode
);
3152 * queue inode for writeback: we can't actually call
3153 * filemap_write_and_wait, etc. from message handler
3156 ceph_queue_writeback(inode
);
3157 if (queue_invalidate
)
3158 ceph_queue_invalidate(inode
);
3160 invalidate_aliases(inode
);
3162 wake_up_all(&ci
->i_cap_wq
);
3164 if (check_caps
== 1)
3165 ceph_check_caps(ci
, CHECK_CAPS_NODELAY
|CHECK_CAPS_AUTHONLY
,
3167 else if (check_caps
== 2)
3168 ceph_check_caps(ci
, CHECK_CAPS_NODELAY
, session
);
3170 mutex_unlock(&session
->s_mutex
);
3174 * Handle FLUSH_ACK from MDS, indicating that metadata we sent to the
3175 * MDS has been safely committed.
3177 static void handle_cap_flush_ack(struct inode
*inode
, u64 flush_tid
,
3178 struct ceph_mds_caps
*m
,
3179 struct ceph_mds_session
*session
,
3180 struct ceph_cap
*cap
)
3181 __releases(ci
->i_ceph_lock
)
3183 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3184 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
3185 struct ceph_cap_flush
*cf
, *tmp_cf
;
3186 LIST_HEAD(to_remove
);
3187 unsigned seq
= le32_to_cpu(m
->seq
);
3188 int dirty
= le32_to_cpu(m
->dirty
);
3191 bool wake_ci
= false;
3192 bool wake_mdsc
= false;
3194 list_for_each_entry_safe(cf
, tmp_cf
, &ci
->i_cap_flush_list
, i_list
) {
3195 if (cf
->tid
== flush_tid
)
3197 if (cf
->caps
== 0) /* capsnap */
3199 if (cf
->tid
<= flush_tid
) {
3200 if (__finish_cap_flush(NULL
, ci
, cf
))
3202 list_add_tail(&cf
->i_list
, &to_remove
);
3204 cleaned
&= ~cf
->caps
;
3210 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
3211 " flushing %s -> %s\n",
3212 inode
, session
->s_mds
, seq
, ceph_cap_string(dirty
),
3213 ceph_cap_string(cleaned
), ceph_cap_string(ci
->i_flushing_caps
),
3214 ceph_cap_string(ci
->i_flushing_caps
& ~cleaned
));
3216 if (list_empty(&to_remove
) && !cleaned
)
3219 ci
->i_flushing_caps
&= ~cleaned
;
3221 spin_lock(&mdsc
->cap_dirty_lock
);
3223 list_for_each_entry(cf
, &to_remove
, i_list
) {
3224 if (__finish_cap_flush(mdsc
, NULL
, cf
))
3228 if (ci
->i_flushing_caps
== 0) {
3229 if (list_empty(&ci
->i_cap_flush_list
)) {
3230 list_del_init(&ci
->i_flushing_item
);
3231 if (!list_empty(&session
->s_cap_flushing
)) {
3232 dout(" mds%d still flushing cap on %p\n",
3234 &list_first_entry(&session
->s_cap_flushing
,
3235 struct ceph_inode_info
,
3236 i_flushing_item
)->vfs_inode
);
3239 mdsc
->num_cap_flushing
--;
3240 dout(" inode %p now !flushing\n", inode
);
3242 if (ci
->i_dirty_caps
== 0) {
3243 dout(" inode %p now clean\n", inode
);
3244 BUG_ON(!list_empty(&ci
->i_dirty_item
));
3246 if (ci
->i_wr_ref
== 0 &&
3247 ci
->i_wrbuffer_ref_head
== 0) {
3248 BUG_ON(!ci
->i_head_snapc
);
3249 ceph_put_snap_context(ci
->i_head_snapc
);
3250 ci
->i_head_snapc
= NULL
;
3253 BUG_ON(list_empty(&ci
->i_dirty_item
));
3256 spin_unlock(&mdsc
->cap_dirty_lock
);
3259 spin_unlock(&ci
->i_ceph_lock
);
3261 while (!list_empty(&to_remove
)) {
3262 cf
= list_first_entry(&to_remove
,
3263 struct ceph_cap_flush
, i_list
);
3264 list_del(&cf
->i_list
);
3265 ceph_free_cap_flush(cf
);
3269 wake_up_all(&ci
->i_cap_wq
);
3271 wake_up_all(&mdsc
->cap_flushing_wq
);
3277 * Handle FLUSHSNAP_ACK. MDS has flushed snap data to disk and we can
3278 * throw away our cap_snap.
3280 * Caller hold s_mutex.
3282 static void handle_cap_flushsnap_ack(struct inode
*inode
, u64 flush_tid
,
3283 struct ceph_mds_caps
*m
,
3284 struct ceph_mds_session
*session
)
3286 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3287 struct ceph_mds_client
*mdsc
= ceph_sb_to_client(inode
->i_sb
)->mdsc
;
3288 u64 follows
= le64_to_cpu(m
->snap_follows
);
3289 struct ceph_cap_snap
*capsnap
;
3290 bool flushed
= false;
3291 bool wake_ci
= false;
3292 bool wake_mdsc
= false;
3294 dout("handle_cap_flushsnap_ack inode %p ci %p mds%d follows %lld\n",
3295 inode
, ci
, session
->s_mds
, follows
);
3297 spin_lock(&ci
->i_ceph_lock
);
3298 list_for_each_entry(capsnap
, &ci
->i_cap_snaps
, ci_item
) {
3299 if (capsnap
->follows
== follows
) {
3300 if (capsnap
->cap_flush
.tid
!= flush_tid
) {
3301 dout(" cap_snap %p follows %lld tid %lld !="
3302 " %lld\n", capsnap
, follows
,
3303 flush_tid
, capsnap
->cap_flush
.tid
);
3309 dout(" skipping cap_snap %p follows %lld\n",
3310 capsnap
, capsnap
->follows
);
3314 WARN_ON(capsnap
->dirty_pages
|| capsnap
->writing
);
3315 dout(" removing %p cap_snap %p follows %lld\n",
3316 inode
, capsnap
, follows
);
3317 list_del(&capsnap
->ci_item
);
3318 if (__finish_cap_flush(NULL
, ci
, &capsnap
->cap_flush
))
3321 spin_lock(&mdsc
->cap_dirty_lock
);
3323 if (list_empty(&ci
->i_cap_flush_list
))
3324 list_del_init(&ci
->i_flushing_item
);
3326 if (__finish_cap_flush(mdsc
, NULL
, &capsnap
->cap_flush
))
3329 spin_unlock(&mdsc
->cap_dirty_lock
);
3331 spin_unlock(&ci
->i_ceph_lock
);
3333 ceph_put_snap_context(capsnap
->context
);
3334 ceph_put_cap_snap(capsnap
);
3336 wake_up_all(&ci
->i_cap_wq
);
3338 wake_up_all(&mdsc
->cap_flushing_wq
);
3344 * Handle TRUNC from MDS, indicating file truncation.
3346 * caller hold s_mutex.
3348 static void handle_cap_trunc(struct inode
*inode
,
3349 struct ceph_mds_caps
*trunc
,
3350 struct ceph_mds_session
*session
)
3351 __releases(ci
->i_ceph_lock
)
3353 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3354 int mds
= session
->s_mds
;
3355 int seq
= le32_to_cpu(trunc
->seq
);
3356 u32 truncate_seq
= le32_to_cpu(trunc
->truncate_seq
);
3357 u64 truncate_size
= le64_to_cpu(trunc
->truncate_size
);
3358 u64 size
= le64_to_cpu(trunc
->size
);
3359 int implemented
= 0;
3360 int dirty
= __ceph_caps_dirty(ci
);
3361 int issued
= __ceph_caps_issued(ceph_inode(inode
), &implemented
);
3362 int queue_trunc
= 0;
3364 issued
|= implemented
| dirty
;
3366 dout("handle_cap_trunc inode %p mds%d seq %d to %lld seq %d\n",
3367 inode
, mds
, seq
, truncate_size
, truncate_seq
);
3368 queue_trunc
= ceph_fill_file_size(inode
, issued
,
3369 truncate_seq
, truncate_size
, size
);
3370 spin_unlock(&ci
->i_ceph_lock
);
3373 ceph_queue_vmtruncate(inode
);
3377 * Handle EXPORT from MDS. Cap is being migrated _from_ this mds to a
3378 * different one. If we are the most recent migration we've seen (as
3379 * indicated by mseq), make note of the migrating cap bits for the
3380 * duration (until we see the corresponding IMPORT).
3382 * caller holds s_mutex
3384 static void handle_cap_export(struct inode
*inode
, struct ceph_mds_caps
*ex
,
3385 struct ceph_mds_cap_peer
*ph
,
3386 struct ceph_mds_session
*session
)
3388 struct ceph_mds_client
*mdsc
= ceph_inode_to_client(inode
)->mdsc
;
3389 struct ceph_mds_session
*tsession
= NULL
;
3390 struct ceph_cap
*cap
, *tcap
, *new_cap
= NULL
;
3391 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3393 unsigned mseq
= le32_to_cpu(ex
->migrate_seq
);
3394 unsigned t_seq
, t_mseq
;
3396 int mds
= session
->s_mds
;
3399 t_cap_id
= le64_to_cpu(ph
->cap_id
);
3400 t_seq
= le32_to_cpu(ph
->seq
);
3401 t_mseq
= le32_to_cpu(ph
->mseq
);
3402 target
= le32_to_cpu(ph
->mds
);
3404 t_cap_id
= t_seq
= t_mseq
= 0;
3408 dout("handle_cap_export inode %p ci %p mds%d mseq %d target %d\n",
3409 inode
, ci
, mds
, mseq
, target
);
3411 spin_lock(&ci
->i_ceph_lock
);
3412 cap
= __get_cap_for_mds(ci
, mds
);
3413 if (!cap
|| cap
->cap_id
!= le64_to_cpu(ex
->cap_id
))
3417 __ceph_remove_cap(cap
, false);
3418 if (!ci
->i_auth_cap
)
3419 ci
->i_ceph_flags
|= CEPH_I_CAP_DROPPED
;
3424 * now we know we haven't received the cap import message yet
3425 * because the exported cap still exist.
3428 issued
= cap
->issued
;
3429 WARN_ON(issued
!= cap
->implemented
);
3431 tcap
= __get_cap_for_mds(ci
, target
);
3433 /* already have caps from the target */
3434 if (tcap
->cap_id
== t_cap_id
&&
3435 ceph_seq_cmp(tcap
->seq
, t_seq
) < 0) {
3436 dout(" updating import cap %p mds%d\n", tcap
, target
);
3437 tcap
->cap_id
= t_cap_id
;
3438 tcap
->seq
= t_seq
- 1;
3439 tcap
->issue_seq
= t_seq
- 1;
3440 tcap
->mseq
= t_mseq
;
3441 tcap
->issued
|= issued
;
3442 tcap
->implemented
|= issued
;
3443 if (cap
== ci
->i_auth_cap
)
3444 ci
->i_auth_cap
= tcap
;
3446 if (!list_empty(&ci
->i_cap_flush_list
) &&
3447 ci
->i_auth_cap
== tcap
) {
3448 spin_lock(&mdsc
->cap_dirty_lock
);
3449 list_move_tail(&ci
->i_flushing_item
,
3450 &tcap
->session
->s_cap_flushing
);
3451 spin_unlock(&mdsc
->cap_dirty_lock
);
3454 __ceph_remove_cap(cap
, false);
3456 } else if (tsession
) {
3457 /* add placeholder for the export tagert */
3458 int flag
= (cap
== ci
->i_auth_cap
) ? CEPH_CAP_FLAG_AUTH
: 0;
3460 ceph_add_cap(inode
, tsession
, t_cap_id
, -1, issued
, 0,
3461 t_seq
- 1, t_mseq
, (u64
)-1, flag
, &new_cap
);
3463 if (!list_empty(&ci
->i_cap_flush_list
) &&
3464 ci
->i_auth_cap
== tcap
) {
3465 spin_lock(&mdsc
->cap_dirty_lock
);
3466 list_move_tail(&ci
->i_flushing_item
,
3467 &tcap
->session
->s_cap_flushing
);
3468 spin_unlock(&mdsc
->cap_dirty_lock
);
3471 __ceph_remove_cap(cap
, false);
3475 spin_unlock(&ci
->i_ceph_lock
);
3476 mutex_unlock(&session
->s_mutex
);
3478 /* open target session */
3479 tsession
= ceph_mdsc_open_export_target_session(mdsc
, target
);
3480 if (!IS_ERR(tsession
)) {
3482 mutex_lock(&session
->s_mutex
);
3483 mutex_lock_nested(&tsession
->s_mutex
,
3484 SINGLE_DEPTH_NESTING
);
3486 mutex_lock(&tsession
->s_mutex
);
3487 mutex_lock_nested(&session
->s_mutex
,
3488 SINGLE_DEPTH_NESTING
);
3490 new_cap
= ceph_get_cap(mdsc
, NULL
);
3499 spin_unlock(&ci
->i_ceph_lock
);
3500 mutex_unlock(&session
->s_mutex
);
3502 mutex_unlock(&tsession
->s_mutex
);
3503 ceph_put_mds_session(tsession
);
3506 ceph_put_cap(mdsc
, new_cap
);
3510 * Handle cap IMPORT.
3512 * caller holds s_mutex. acquires i_ceph_lock
3514 static void handle_cap_import(struct ceph_mds_client
*mdsc
,
3515 struct inode
*inode
, struct ceph_mds_caps
*im
,
3516 struct ceph_mds_cap_peer
*ph
,
3517 struct ceph_mds_session
*session
,
3518 struct ceph_cap
**target_cap
, int *old_issued
)
3519 __acquires(ci
->i_ceph_lock
)
3521 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3522 struct ceph_cap
*cap
, *ocap
, *new_cap
= NULL
;
3523 int mds
= session
->s_mds
;
3525 unsigned caps
= le32_to_cpu(im
->caps
);
3526 unsigned wanted
= le32_to_cpu(im
->wanted
);
3527 unsigned seq
= le32_to_cpu(im
->seq
);
3528 unsigned mseq
= le32_to_cpu(im
->migrate_seq
);
3529 u64 realmino
= le64_to_cpu(im
->realm
);
3530 u64 cap_id
= le64_to_cpu(im
->cap_id
);
3535 p_cap_id
= le64_to_cpu(ph
->cap_id
);
3536 peer
= le32_to_cpu(ph
->mds
);
3542 dout("handle_cap_import inode %p ci %p mds%d mseq %d peer %d\n",
3543 inode
, ci
, mds
, mseq
, peer
);
3546 spin_lock(&ci
->i_ceph_lock
);
3547 cap
= __get_cap_for_mds(ci
, mds
);
3550 spin_unlock(&ci
->i_ceph_lock
);
3551 new_cap
= ceph_get_cap(mdsc
, NULL
);
3557 ceph_put_cap(mdsc
, new_cap
);
3562 __ceph_caps_issued(ci
, &issued
);
3563 issued
|= __ceph_caps_dirty(ci
);
3565 ceph_add_cap(inode
, session
, cap_id
, -1, caps
, wanted
, seq
, mseq
,
3566 realmino
, CEPH_CAP_FLAG_AUTH
, &new_cap
);
3568 ocap
= peer
>= 0 ? __get_cap_for_mds(ci
, peer
) : NULL
;
3569 if (ocap
&& ocap
->cap_id
== p_cap_id
) {
3570 dout(" remove export cap %p mds%d flags %d\n",
3571 ocap
, peer
, ph
->flags
);
3572 if ((ph
->flags
& CEPH_CAP_FLAG_AUTH
) &&
3573 (ocap
->seq
!= le32_to_cpu(ph
->seq
) ||
3574 ocap
->mseq
!= le32_to_cpu(ph
->mseq
))) {
3575 pr_err("handle_cap_import: mismatched seq/mseq: "
3576 "ino (%llx.%llx) mds%d seq %d mseq %d "
3577 "importer mds%d has peer seq %d mseq %d\n",
3578 ceph_vinop(inode
), peer
, ocap
->seq
,
3579 ocap
->mseq
, mds
, le32_to_cpu(ph
->seq
),
3580 le32_to_cpu(ph
->mseq
));
3582 __ceph_remove_cap(ocap
, (ph
->flags
& CEPH_CAP_FLAG_RELEASE
));
3585 /* make sure we re-request max_size, if necessary */
3586 ci
->i_requested_max_size
= 0;
3588 *old_issued
= issued
;
3593 * Handle a caps message from the MDS.
3595 * Identify the appropriate session, inode, and call the right handler
3596 * based on the cap op.
3598 void ceph_handle_caps(struct ceph_mds_session
*session
,
3599 struct ceph_msg
*msg
)
3601 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
3602 struct super_block
*sb
= mdsc
->fsc
->sb
;
3603 struct inode
*inode
;
3604 struct ceph_inode_info
*ci
;
3605 struct ceph_cap
*cap
;
3606 struct ceph_mds_caps
*h
;
3607 struct ceph_mds_cap_peer
*peer
= NULL
;
3608 struct ceph_snap_realm
*realm
= NULL
;
3609 struct ceph_string
*pool_ns
= NULL
;
3610 int mds
= session
->s_mds
;
3613 struct ceph_vino vino
;
3615 u64 inline_version
= 0;
3616 void *inline_data
= NULL
;
3619 size_t snaptrace_len
;
3622 dout("handle_caps from mds%d\n", mds
);
3625 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
3626 tid
= le64_to_cpu(msg
->hdr
.tid
);
3627 if (msg
->front
.iov_len
< sizeof(*h
))
3629 h
= msg
->front
.iov_base
;
3630 op
= le32_to_cpu(h
->op
);
3631 vino
.ino
= le64_to_cpu(h
->ino
);
3632 vino
.snap
= CEPH_NOSNAP
;
3633 seq
= le32_to_cpu(h
->seq
);
3634 mseq
= le32_to_cpu(h
->migrate_seq
);
3637 snaptrace_len
= le32_to_cpu(h
->snap_trace_len
);
3638 p
= snaptrace
+ snaptrace_len
;
3640 if (le16_to_cpu(msg
->hdr
.version
) >= 2) {
3642 ceph_decode_32_safe(&p
, end
, flock_len
, bad
);
3643 if (p
+ flock_len
> end
)
3648 if (le16_to_cpu(msg
->hdr
.version
) >= 3) {
3649 if (op
== CEPH_CAP_OP_IMPORT
) {
3650 if (p
+ sizeof(*peer
) > end
)
3654 } else if (op
== CEPH_CAP_OP_EXPORT
) {
3655 /* recorded in unused fields */
3656 peer
= (void *)&h
->size
;
3660 if (le16_to_cpu(msg
->hdr
.version
) >= 4) {
3661 ceph_decode_64_safe(&p
, end
, inline_version
, bad
);
3662 ceph_decode_32_safe(&p
, end
, inline_len
, bad
);
3663 if (p
+ inline_len
> end
)
3669 if (le16_to_cpu(msg
->hdr
.version
) >= 5) {
3670 struct ceph_osd_client
*osdc
= &mdsc
->fsc
->client
->osdc
;
3673 ceph_decode_32_safe(&p
, end
, epoch_barrier
, bad
);
3674 ceph_osdc_update_epoch_barrier(osdc
, epoch_barrier
);
3677 if (le16_to_cpu(msg
->hdr
.version
) >= 8) {
3679 u32 caller_uid
, caller_gid
;
3683 ceph_decode_64_safe(&p
, end
, flush_tid
, bad
);
3685 ceph_decode_32_safe(&p
, end
, caller_uid
, bad
);
3686 ceph_decode_32_safe(&p
, end
, caller_gid
, bad
);
3688 ceph_decode_32_safe(&p
, end
, pool_ns_len
, bad
);
3689 if (pool_ns_len
> 0) {
3690 ceph_decode_need(&p
, end
, pool_ns_len
, bad
);
3691 pool_ns
= ceph_find_or_create_string(p
, pool_ns_len
);
3697 inode
= ceph_find_inode(sb
, vino
);
3698 ci
= ceph_inode(inode
);
3699 dout(" op %s ino %llx.%llx inode %p\n", ceph_cap_op_name(op
), vino
.ino
,
3702 mutex_lock(&session
->s_mutex
);
3704 dout(" mds%d seq %lld cap seq %u\n", session
->s_mds
, session
->s_seq
,
3708 dout(" i don't have ino %llx\n", vino
.ino
);
3710 if (op
== CEPH_CAP_OP_IMPORT
) {
3711 cap
= ceph_get_cap(mdsc
, NULL
);
3712 cap
->cap_ino
= vino
.ino
;
3713 cap
->queue_release
= 1;
3714 cap
->cap_id
= le64_to_cpu(h
->cap_id
);
3717 cap
->issue_seq
= seq
;
3718 spin_lock(&session
->s_cap_lock
);
3719 list_add_tail(&cap
->session_caps
,
3720 &session
->s_cap_releases
);
3721 session
->s_num_cap_releases
++;
3722 spin_unlock(&session
->s_cap_lock
);
3724 goto flush_cap_releases
;
3727 /* these will work even if we don't have a cap yet */
3729 case CEPH_CAP_OP_FLUSHSNAP_ACK
:
3730 handle_cap_flushsnap_ack(inode
, tid
, h
, session
);
3733 case CEPH_CAP_OP_EXPORT
:
3734 handle_cap_export(inode
, h
, peer
, session
);
3737 case CEPH_CAP_OP_IMPORT
:
3739 if (snaptrace_len
) {
3740 down_write(&mdsc
->snap_rwsem
);
3741 ceph_update_snap_trace(mdsc
, snaptrace
,
3742 snaptrace
+ snaptrace_len
,
3744 downgrade_write(&mdsc
->snap_rwsem
);
3746 down_read(&mdsc
->snap_rwsem
);
3748 handle_cap_import(mdsc
, inode
, h
, peer
, session
,
3750 handle_cap_grant(mdsc
, inode
, h
, &pool_ns
,
3751 inline_version
, inline_data
, inline_len
,
3752 msg
->middle
, session
, cap
, issued
);
3754 ceph_put_snap_realm(mdsc
, realm
);
3758 /* the rest require a cap */
3759 spin_lock(&ci
->i_ceph_lock
);
3760 cap
= __get_cap_for_mds(ceph_inode(inode
), mds
);
3762 dout(" no cap on %p ino %llx.%llx from mds%d\n",
3763 inode
, ceph_ino(inode
), ceph_snap(inode
), mds
);
3764 spin_unlock(&ci
->i_ceph_lock
);
3765 goto flush_cap_releases
;
3768 /* note that each of these drops i_ceph_lock for us */
3770 case CEPH_CAP_OP_REVOKE
:
3771 case CEPH_CAP_OP_GRANT
:
3772 __ceph_caps_issued(ci
, &issued
);
3773 issued
|= __ceph_caps_dirty(ci
);
3774 handle_cap_grant(mdsc
, inode
, h
, &pool_ns
,
3775 inline_version
, inline_data
, inline_len
,
3776 msg
->middle
, session
, cap
, issued
);
3779 case CEPH_CAP_OP_FLUSH_ACK
:
3780 handle_cap_flush_ack(inode
, tid
, h
, session
, cap
);
3783 case CEPH_CAP_OP_TRUNC
:
3784 handle_cap_trunc(inode
, h
, session
);
3788 spin_unlock(&ci
->i_ceph_lock
);
3789 pr_err("ceph_handle_caps: unknown cap op %d %s\n", op
,
3790 ceph_cap_op_name(op
));
3797 * send any cap release message to try to move things
3798 * along for the mds (who clearly thinks we still have this
3801 ceph_send_cap_releases(mdsc
, session
);
3804 mutex_unlock(&session
->s_mutex
);
3807 ceph_put_string(pool_ns
);
3811 pr_err("ceph_handle_caps: corrupt message\n");
3817 * Delayed work handler to process end of delayed cap release LRU list.
3819 void ceph_check_delayed_caps(struct ceph_mds_client
*mdsc
)
3821 struct inode
*inode
;
3822 struct ceph_inode_info
*ci
;
3823 int flags
= CHECK_CAPS_NODELAY
;
3825 dout("check_delayed_caps\n");
3827 spin_lock(&mdsc
->cap_delay_lock
);
3828 if (list_empty(&mdsc
->cap_delay_list
))
3830 ci
= list_first_entry(&mdsc
->cap_delay_list
,
3831 struct ceph_inode_info
,
3833 if ((ci
->i_ceph_flags
& CEPH_I_FLUSH
) == 0 &&
3834 time_before(jiffies
, ci
->i_hold_caps_max
))
3836 list_del_init(&ci
->i_cap_delay_list
);
3838 inode
= igrab(&ci
->vfs_inode
);
3839 spin_unlock(&mdsc
->cap_delay_lock
);
3842 dout("check_delayed_caps on %p\n", inode
);
3843 ceph_check_caps(ci
, flags
, NULL
);
3847 spin_unlock(&mdsc
->cap_delay_lock
);
3851 * Flush all dirty caps to the mds
3853 void ceph_flush_dirty_caps(struct ceph_mds_client
*mdsc
)
3855 struct ceph_inode_info
*ci
;
3856 struct inode
*inode
;
3858 dout("flush_dirty_caps\n");
3859 spin_lock(&mdsc
->cap_dirty_lock
);
3860 while (!list_empty(&mdsc
->cap_dirty
)) {
3861 ci
= list_first_entry(&mdsc
->cap_dirty
, struct ceph_inode_info
,
3863 inode
= &ci
->vfs_inode
;
3865 dout("flush_dirty_caps %p\n", inode
);
3866 spin_unlock(&mdsc
->cap_dirty_lock
);
3867 ceph_check_caps(ci
, CHECK_CAPS_NODELAY
|CHECK_CAPS_FLUSH
, NULL
);
3869 spin_lock(&mdsc
->cap_dirty_lock
);
3871 spin_unlock(&mdsc
->cap_dirty_lock
);
3872 dout("flush_dirty_caps done\n");
3875 void __ceph_get_fmode(struct ceph_inode_info
*ci
, int fmode
)
3878 int bits
= (fmode
<< 1) | 1;
3879 for (i
= 0; i
< CEPH_FILE_MODE_BITS
; i
++) {
3880 if (bits
& (1 << i
))
3881 ci
->i_nr_by_mode
[i
]++;
3886 * Drop open file reference. If we were the last open file,
3887 * we may need to release capabilities to the MDS (or schedule
3888 * their delayed release).
3890 void ceph_put_fmode(struct ceph_inode_info
*ci
, int fmode
)
3893 int bits
= (fmode
<< 1) | 1;
3894 spin_lock(&ci
->i_ceph_lock
);
3895 for (i
= 0; i
< CEPH_FILE_MODE_BITS
; i
++) {
3896 if (bits
& (1 << i
)) {
3897 BUG_ON(ci
->i_nr_by_mode
[i
] == 0);
3898 if (--ci
->i_nr_by_mode
[i
] == 0)
3902 dout("put_fmode %p fmode %d {%d,%d,%d,%d}\n",
3903 &ci
->vfs_inode
, fmode
,
3904 ci
->i_nr_by_mode
[0], ci
->i_nr_by_mode
[1],
3905 ci
->i_nr_by_mode
[2], ci
->i_nr_by_mode
[3]);
3906 spin_unlock(&ci
->i_ceph_lock
);
3908 if (last
&& ci
->i_vino
.snap
== CEPH_NOSNAP
)
3909 ceph_check_caps(ci
, 0, NULL
);
3913 * Helpers for embedding cap and dentry lease releases into mds
3916 * @force is used by dentry_release (below) to force inclusion of a
3917 * record for the directory inode, even when there aren't any caps to
3920 int ceph_encode_inode_release(void **p
, struct inode
*inode
,
3921 int mds
, int drop
, int unless
, int force
)
3923 struct ceph_inode_info
*ci
= ceph_inode(inode
);
3924 struct ceph_cap
*cap
;
3925 struct ceph_mds_request_release
*rel
= *p
;
3929 spin_lock(&ci
->i_ceph_lock
);
3930 used
= __ceph_caps_used(ci
);
3931 dirty
= __ceph_caps_dirty(ci
);
3933 dout("encode_inode_release %p mds%d used|dirty %s drop %s unless %s\n",
3934 inode
, mds
, ceph_cap_string(used
|dirty
), ceph_cap_string(drop
),
3935 ceph_cap_string(unless
));
3937 /* only drop unused, clean caps */
3938 drop
&= ~(used
| dirty
);
3940 cap
= __get_cap_for_mds(ci
, mds
);
3941 if (cap
&& __cap_is_valid(cap
)) {
3943 ((cap
->issued
& drop
) &&
3944 (cap
->issued
& unless
) == 0)) {
3945 if ((cap
->issued
& drop
) &&
3946 (cap
->issued
& unless
) == 0) {
3947 int wanted
= __ceph_caps_wanted(ci
);
3948 if ((ci
->i_ceph_flags
& CEPH_I_NODELAY
) == 0)
3949 wanted
|= cap
->mds_wanted
;
3950 dout("encode_inode_release %p cap %p "
3951 "%s -> %s, wanted %s -> %s\n", inode
, cap
,
3952 ceph_cap_string(cap
->issued
),
3953 ceph_cap_string(cap
->issued
& ~drop
),
3954 ceph_cap_string(cap
->mds_wanted
),
3955 ceph_cap_string(wanted
));
3957 cap
->issued
&= ~drop
;
3958 cap
->implemented
&= ~drop
;
3959 cap
->mds_wanted
= wanted
;
3961 dout("encode_inode_release %p cap %p %s"
3962 " (force)\n", inode
, cap
,
3963 ceph_cap_string(cap
->issued
));
3966 rel
->ino
= cpu_to_le64(ceph_ino(inode
));
3967 rel
->cap_id
= cpu_to_le64(cap
->cap_id
);
3968 rel
->seq
= cpu_to_le32(cap
->seq
);
3969 rel
->issue_seq
= cpu_to_le32(cap
->issue_seq
);
3970 rel
->mseq
= cpu_to_le32(cap
->mseq
);
3971 rel
->caps
= cpu_to_le32(cap
->implemented
);
3972 rel
->wanted
= cpu_to_le32(cap
->mds_wanted
);
3978 dout("encode_inode_release %p cap %p %s\n",
3979 inode
, cap
, ceph_cap_string(cap
->issued
));
3982 spin_unlock(&ci
->i_ceph_lock
);
3986 int ceph_encode_dentry_release(void **p
, struct dentry
*dentry
,
3988 int mds
, int drop
, int unless
)
3990 struct dentry
*parent
= NULL
;
3991 struct ceph_mds_request_release
*rel
= *p
;
3992 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
3997 * force an record for the directory caps if we have a dentry lease.
3998 * this is racy (can't take i_ceph_lock and d_lock together), but it
3999 * doesn't have to be perfect; the mds will revoke anything we don't
4002 spin_lock(&dentry
->d_lock
);
4003 if (di
->lease_session
&& di
->lease_session
->s_mds
== mds
)
4006 parent
= dget(dentry
->d_parent
);
4007 dir
= d_inode(parent
);
4009 spin_unlock(&dentry
->d_lock
);
4011 ret
= ceph_encode_inode_release(p
, dir
, mds
, drop
, unless
, force
);
4014 spin_lock(&dentry
->d_lock
);
4015 if (ret
&& di
->lease_session
&& di
->lease_session
->s_mds
== mds
) {
4016 dout("encode_dentry_release %p mds%d seq %d\n",
4017 dentry
, mds
, (int)di
->lease_seq
);
4018 rel
->dname_len
= cpu_to_le32(dentry
->d_name
.len
);
4019 memcpy(*p
, dentry
->d_name
.name
, dentry
->d_name
.len
);
4020 *p
+= dentry
->d_name
.len
;
4021 rel
->dname_seq
= cpu_to_le32(di
->lease_seq
);
4022 __ceph_mdsc_drop_dentry_lease(dentry
);
4024 spin_unlock(&dentry
->d_lock
);