1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
7 #include <linux/sched.h>
8 #include <linux/debugfs.h>
9 #include <linux/seq_file.h>
10 #include <linux/utsname.h>
11 #include <linux/ratelimit.h>
14 #include "mds_client.h"
16 #include <linux/ceph/ceph_features.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/pagelist.h>
20 #include <linux/ceph/auth.h>
21 #include <linux/ceph/debugfs.h>
24 * A cluster of MDS (metadata server) daemons is responsible for
25 * managing the file system namespace (the directory hierarchy and
26 * inodes) and for coordinating shared access to storage. Metadata is
27 * partitioning hierarchically across a number of servers, and that
28 * partition varies over time as the cluster adjusts the distribution
29 * in order to balance load.
31 * The MDS client is primarily responsible to managing synchronous
32 * metadata requests for operations like open, unlink, and so forth.
33 * If there is a MDS failure, we find out about it when we (possibly
34 * request and) receive a new MDS map, and can resubmit affected
37 * For the most part, though, we take advantage of a lossless
38 * communications channel to the MDS, and do not need to worry about
39 * timing out or resubmitting requests.
41 * We maintain a stateful "session" with each MDS we interact with.
42 * Within each session, we sent periodic heartbeat messages to ensure
43 * any capabilities or leases we have been issues remain valid. If
44 * the session times out and goes stale, our leases and capabilities
45 * are no longer valid.
48 struct ceph_reconnect_state
{
50 struct ceph_pagelist
*pagelist
;
54 static void __wake_requests(struct ceph_mds_client
*mdsc
,
55 struct list_head
*head
);
57 static const struct ceph_connection_operations mds_con_ops
;
65 * parse individual inode info
67 static int parse_reply_info_in(void **p
, void *end
,
68 struct ceph_mds_reply_info_in
*info
,
74 *p
+= sizeof(struct ceph_mds_reply_inode
) +
75 sizeof(*info
->in
->fragtree
.splits
) *
76 le32_to_cpu(info
->in
->fragtree
.nsplits
);
78 ceph_decode_32_safe(p
, end
, info
->symlink_len
, bad
);
79 ceph_decode_need(p
, end
, info
->symlink_len
, bad
);
81 *p
+= info
->symlink_len
;
83 if (features
& CEPH_FEATURE_DIRLAYOUTHASH
)
84 ceph_decode_copy_safe(p
, end
, &info
->dir_layout
,
85 sizeof(info
->dir_layout
), bad
);
87 memset(&info
->dir_layout
, 0, sizeof(info
->dir_layout
));
89 ceph_decode_32_safe(p
, end
, info
->xattr_len
, bad
);
90 ceph_decode_need(p
, end
, info
->xattr_len
, bad
);
91 info
->xattr_data
= *p
;
92 *p
+= info
->xattr_len
;
94 if (features
& CEPH_FEATURE_MDS_INLINE_DATA
) {
95 ceph_decode_64_safe(p
, end
, info
->inline_version
, bad
);
96 ceph_decode_32_safe(p
, end
, info
->inline_len
, bad
);
97 ceph_decode_need(p
, end
, info
->inline_len
, bad
);
98 info
->inline_data
= *p
;
99 *p
+= info
->inline_len
;
101 info
->inline_version
= CEPH_INLINE_NONE
;
103 info
->pool_ns_len
= 0;
104 info
->pool_ns_data
= NULL
;
105 if (features
& CEPH_FEATURE_FS_FILE_LAYOUT_V2
) {
106 ceph_decode_32_safe(p
, end
, info
->pool_ns_len
, bad
);
107 if (info
->pool_ns_len
> 0) {
108 ceph_decode_need(p
, end
, info
->pool_ns_len
, bad
);
109 info
->pool_ns_data
= *p
;
110 *p
+= info
->pool_ns_len
;
120 * parse a normal reply, which may contain a (dir+)dentry and/or a
123 static int parse_reply_info_trace(void **p
, void *end
,
124 struct ceph_mds_reply_info_parsed
*info
,
129 if (info
->head
->is_dentry
) {
130 err
= parse_reply_info_in(p
, end
, &info
->diri
, features
);
134 if (unlikely(*p
+ sizeof(*info
->dirfrag
) > end
))
137 *p
+= sizeof(*info
->dirfrag
) +
138 sizeof(u32
)*le32_to_cpu(info
->dirfrag
->ndist
);
139 if (unlikely(*p
> end
))
142 ceph_decode_32_safe(p
, end
, info
->dname_len
, bad
);
143 ceph_decode_need(p
, end
, info
->dname_len
, bad
);
145 *p
+= info
->dname_len
;
147 *p
+= sizeof(*info
->dlease
);
150 if (info
->head
->is_target
) {
151 err
= parse_reply_info_in(p
, end
, &info
->targeti
, features
);
156 if (unlikely(*p
!= end
))
163 pr_err("problem parsing mds trace %d\n", err
);
168 * parse readdir results
170 static int parse_reply_info_dir(void **p
, void *end
,
171 struct ceph_mds_reply_info_parsed
*info
,
178 if (*p
+ sizeof(*info
->dir_dir
) > end
)
180 *p
+= sizeof(*info
->dir_dir
) +
181 sizeof(u32
)*le32_to_cpu(info
->dir_dir
->ndist
);
185 ceph_decode_need(p
, end
, sizeof(num
) + 2, bad
);
186 num
= ceph_decode_32(p
);
188 u16 flags
= ceph_decode_16(p
);
189 info
->dir_end
= !!(flags
& CEPH_READDIR_FRAG_END
);
190 info
->dir_complete
= !!(flags
& CEPH_READDIR_FRAG_COMPLETE
);
191 info
->hash_order
= !!(flags
& CEPH_READDIR_HASH_ORDER
);
196 BUG_ON(!info
->dir_entries
);
197 if ((unsigned long)(info
->dir_entries
+ num
) >
198 (unsigned long)info
->dir_entries
+ info
->dir_buf_size
) {
199 pr_err("dir contents are larger than expected\n");
206 struct ceph_mds_reply_dir_entry
*rde
= info
->dir_entries
+ i
;
208 ceph_decode_need(p
, end
, sizeof(u32
)*2, bad
);
209 rde
->name_len
= ceph_decode_32(p
);
210 ceph_decode_need(p
, end
, rde
->name_len
, bad
);
213 dout("parsed dir dname '%.*s'\n", rde
->name_len
, rde
->name
);
215 *p
+= sizeof(struct ceph_mds_reply_lease
);
218 err
= parse_reply_info_in(p
, end
, &rde
->inode
, features
);
221 /* ceph_readdir_prepopulate() will update it */
235 pr_err("problem parsing dir contents %d\n", err
);
240 * parse fcntl F_GETLK results
242 static int parse_reply_info_filelock(void **p
, void *end
,
243 struct ceph_mds_reply_info_parsed
*info
,
246 if (*p
+ sizeof(*info
->filelock_reply
) > end
)
249 info
->filelock_reply
= *p
;
250 *p
+= sizeof(*info
->filelock_reply
);
252 if (unlikely(*p
!= end
))
261 * parse create results
263 static int parse_reply_info_create(void **p
, void *end
,
264 struct ceph_mds_reply_info_parsed
*info
,
267 if (features
& CEPH_FEATURE_REPLY_CREATE_INODE
) {
269 info
->has_create_ino
= false;
271 info
->has_create_ino
= true;
272 info
->ino
= ceph_decode_64(p
);
276 if (unlikely(*p
!= end
))
285 * parse extra results
287 static int parse_reply_info_extra(void **p
, void *end
,
288 struct ceph_mds_reply_info_parsed
*info
,
291 if (info
->head
->op
== CEPH_MDS_OP_GETFILELOCK
)
292 return parse_reply_info_filelock(p
, end
, info
, features
);
293 else if (info
->head
->op
== CEPH_MDS_OP_READDIR
||
294 info
->head
->op
== CEPH_MDS_OP_LSSNAP
)
295 return parse_reply_info_dir(p
, end
, info
, features
);
296 else if (info
->head
->op
== CEPH_MDS_OP_CREATE
)
297 return parse_reply_info_create(p
, end
, info
, features
);
303 * parse entire mds reply
305 static int parse_reply_info(struct ceph_msg
*msg
,
306 struct ceph_mds_reply_info_parsed
*info
,
313 info
->head
= msg
->front
.iov_base
;
314 p
= msg
->front
.iov_base
+ sizeof(struct ceph_mds_reply_head
);
315 end
= p
+ msg
->front
.iov_len
- sizeof(struct ceph_mds_reply_head
);
318 ceph_decode_32_safe(&p
, end
, len
, bad
);
320 ceph_decode_need(&p
, end
, len
, bad
);
321 err
= parse_reply_info_trace(&p
, p
+len
, info
, features
);
327 ceph_decode_32_safe(&p
, end
, len
, bad
);
329 ceph_decode_need(&p
, end
, len
, bad
);
330 err
= parse_reply_info_extra(&p
, p
+len
, info
, features
);
336 ceph_decode_32_safe(&p
, end
, len
, bad
);
337 info
->snapblob_len
= len
;
348 pr_err("mds parse_reply err %d\n", err
);
352 static void destroy_reply_info(struct ceph_mds_reply_info_parsed
*info
)
354 if (!info
->dir_entries
)
356 free_pages((unsigned long)info
->dir_entries
, get_order(info
->dir_buf_size
));
363 const char *ceph_session_state_name(int s
)
366 case CEPH_MDS_SESSION_NEW
: return "new";
367 case CEPH_MDS_SESSION_OPENING
: return "opening";
368 case CEPH_MDS_SESSION_OPEN
: return "open";
369 case CEPH_MDS_SESSION_HUNG
: return "hung";
370 case CEPH_MDS_SESSION_CLOSING
: return "closing";
371 case CEPH_MDS_SESSION_RESTARTING
: return "restarting";
372 case CEPH_MDS_SESSION_RECONNECTING
: return "reconnecting";
373 case CEPH_MDS_SESSION_REJECTED
: return "rejected";
374 default: return "???";
378 static struct ceph_mds_session
*get_session(struct ceph_mds_session
*s
)
380 if (atomic_inc_not_zero(&s
->s_ref
)) {
381 dout("mdsc get_session %p %d -> %d\n", s
,
382 atomic_read(&s
->s_ref
)-1, atomic_read(&s
->s_ref
));
385 dout("mdsc get_session %p 0 -- FAIL", s
);
390 void ceph_put_mds_session(struct ceph_mds_session
*s
)
392 dout("mdsc put_session %p %d -> %d\n", s
,
393 atomic_read(&s
->s_ref
), atomic_read(&s
->s_ref
)-1);
394 if (atomic_dec_and_test(&s
->s_ref
)) {
395 if (s
->s_auth
.authorizer
)
396 ceph_auth_destroy_authorizer(s
->s_auth
.authorizer
);
402 * called under mdsc->mutex
404 struct ceph_mds_session
*__ceph_lookup_mds_session(struct ceph_mds_client
*mdsc
,
407 struct ceph_mds_session
*session
;
409 if (mds
>= mdsc
->max_sessions
|| mdsc
->sessions
[mds
] == NULL
)
411 session
= mdsc
->sessions
[mds
];
412 dout("lookup_mds_session %p %d\n", session
,
413 atomic_read(&session
->s_ref
));
414 get_session(session
);
418 static bool __have_session(struct ceph_mds_client
*mdsc
, int mds
)
420 if (mds
>= mdsc
->max_sessions
)
422 return mdsc
->sessions
[mds
];
425 static int __verify_registered_session(struct ceph_mds_client
*mdsc
,
426 struct ceph_mds_session
*s
)
428 if (s
->s_mds
>= mdsc
->max_sessions
||
429 mdsc
->sessions
[s
->s_mds
] != s
)
435 * create+register a new session for given mds.
436 * called under mdsc->mutex.
438 static struct ceph_mds_session
*register_session(struct ceph_mds_client
*mdsc
,
441 struct ceph_mds_session
*s
;
443 if (mds
>= mdsc
->mdsmap
->m_max_mds
)
444 return ERR_PTR(-EINVAL
);
446 s
= kzalloc(sizeof(*s
), GFP_NOFS
);
448 return ERR_PTR(-ENOMEM
);
451 s
->s_state
= CEPH_MDS_SESSION_NEW
;
454 mutex_init(&s
->s_mutex
);
456 ceph_con_init(&s
->s_con
, s
, &mds_con_ops
, &mdsc
->fsc
->client
->msgr
);
458 spin_lock_init(&s
->s_gen_ttl_lock
);
460 s
->s_cap_ttl
= jiffies
- 1;
462 spin_lock_init(&s
->s_cap_lock
);
463 s
->s_renew_requested
= 0;
465 INIT_LIST_HEAD(&s
->s_caps
);
468 atomic_set(&s
->s_ref
, 1);
469 INIT_LIST_HEAD(&s
->s_waiting
);
470 INIT_LIST_HEAD(&s
->s_unsafe
);
471 s
->s_num_cap_releases
= 0;
472 s
->s_cap_reconnect
= 0;
473 s
->s_cap_iterator
= NULL
;
474 INIT_LIST_HEAD(&s
->s_cap_releases
);
475 INIT_LIST_HEAD(&s
->s_cap_flushing
);
477 dout("register_session mds%d\n", mds
);
478 if (mds
>= mdsc
->max_sessions
) {
479 int newmax
= 1 << get_count_order(mds
+1);
480 struct ceph_mds_session
**sa
;
482 dout("register_session realloc to %d\n", newmax
);
483 sa
= kcalloc(newmax
, sizeof(void *), GFP_NOFS
);
486 if (mdsc
->sessions
) {
487 memcpy(sa
, mdsc
->sessions
,
488 mdsc
->max_sessions
* sizeof(void *));
489 kfree(mdsc
->sessions
);
492 mdsc
->max_sessions
= newmax
;
494 mdsc
->sessions
[mds
] = s
;
495 atomic_inc(&mdsc
->num_sessions
);
496 atomic_inc(&s
->s_ref
); /* one ref to sessions[], one to caller */
498 ceph_con_open(&s
->s_con
, CEPH_ENTITY_TYPE_MDS
, mds
,
499 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
505 return ERR_PTR(-ENOMEM
);
509 * called under mdsc->mutex
511 static void __unregister_session(struct ceph_mds_client
*mdsc
,
512 struct ceph_mds_session
*s
)
514 dout("__unregister_session mds%d %p\n", s
->s_mds
, s
);
515 BUG_ON(mdsc
->sessions
[s
->s_mds
] != s
);
516 mdsc
->sessions
[s
->s_mds
] = NULL
;
517 ceph_con_close(&s
->s_con
);
518 ceph_put_mds_session(s
);
519 atomic_dec(&mdsc
->num_sessions
);
523 * drop session refs in request.
525 * should be last request ref, or hold mdsc->mutex
527 static void put_request_session(struct ceph_mds_request
*req
)
529 if (req
->r_session
) {
530 ceph_put_mds_session(req
->r_session
);
531 req
->r_session
= NULL
;
535 void ceph_mdsc_release_request(struct kref
*kref
)
537 struct ceph_mds_request
*req
= container_of(kref
,
538 struct ceph_mds_request
,
540 destroy_reply_info(&req
->r_reply_info
);
542 ceph_msg_put(req
->r_request
);
544 ceph_msg_put(req
->r_reply
);
546 ceph_put_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
549 if (req
->r_locked_dir
)
550 ceph_put_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
551 iput(req
->r_target_inode
);
554 if (req
->r_old_dentry
)
555 dput(req
->r_old_dentry
);
556 if (req
->r_old_dentry_dir
) {
558 * track (and drop pins for) r_old_dentry_dir
559 * separately, since r_old_dentry's d_parent may have
560 * changed between the dir mutex being dropped and
561 * this request being freed.
563 ceph_put_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
565 iput(req
->r_old_dentry_dir
);
570 ceph_pagelist_release(req
->r_pagelist
);
571 put_request_session(req
);
572 ceph_unreserve_caps(req
->r_mdsc
, &req
->r_caps_reservation
);
576 DEFINE_RB_FUNCS(request
, struct ceph_mds_request
, r_tid
, r_node
)
579 * lookup session, bump ref if found.
581 * called under mdsc->mutex.
583 static struct ceph_mds_request
*
584 lookup_get_request(struct ceph_mds_client
*mdsc
, u64 tid
)
586 struct ceph_mds_request
*req
;
588 req
= lookup_request(&mdsc
->request_tree
, tid
);
590 ceph_mdsc_get_request(req
);
596 * Register an in-flight request, and assign a tid. Link to directory
597 * are modifying (if any).
599 * Called under mdsc->mutex.
601 static void __register_request(struct ceph_mds_client
*mdsc
,
602 struct ceph_mds_request
*req
,
605 req
->r_tid
= ++mdsc
->last_tid
;
607 ceph_reserve_caps(mdsc
, &req
->r_caps_reservation
,
609 dout("__register_request %p tid %lld\n", req
, req
->r_tid
);
610 ceph_mdsc_get_request(req
);
611 insert_request(&mdsc
->request_tree
, req
);
613 req
->r_uid
= current_fsuid();
614 req
->r_gid
= current_fsgid();
616 if (mdsc
->oldest_tid
== 0 && req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
)
617 mdsc
->oldest_tid
= req
->r_tid
;
621 req
->r_unsafe_dir
= dir
;
625 static void __unregister_request(struct ceph_mds_client
*mdsc
,
626 struct ceph_mds_request
*req
)
628 dout("__unregister_request %p tid %lld\n", req
, req
->r_tid
);
630 if (req
->r_tid
== mdsc
->oldest_tid
) {
631 struct rb_node
*p
= rb_next(&req
->r_node
);
632 mdsc
->oldest_tid
= 0;
634 struct ceph_mds_request
*next_req
=
635 rb_entry(p
, struct ceph_mds_request
, r_node
);
636 if (next_req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
) {
637 mdsc
->oldest_tid
= next_req
->r_tid
;
644 erase_request(&mdsc
->request_tree
, req
);
646 if (req
->r_unsafe_dir
&& req
->r_got_unsafe
) {
647 struct ceph_inode_info
*ci
= ceph_inode(req
->r_unsafe_dir
);
648 spin_lock(&ci
->i_unsafe_lock
);
649 list_del_init(&req
->r_unsafe_dir_item
);
650 spin_unlock(&ci
->i_unsafe_lock
);
652 if (req
->r_target_inode
&& req
->r_got_unsafe
) {
653 struct ceph_inode_info
*ci
= ceph_inode(req
->r_target_inode
);
654 spin_lock(&ci
->i_unsafe_lock
);
655 list_del_init(&req
->r_unsafe_target_item
);
656 spin_unlock(&ci
->i_unsafe_lock
);
659 if (req
->r_unsafe_dir
) {
660 iput(req
->r_unsafe_dir
);
661 req
->r_unsafe_dir
= NULL
;
664 complete_all(&req
->r_safe_completion
);
666 ceph_mdsc_put_request(req
);
670 * Choose mds to send request to next. If there is a hint set in the
671 * request (e.g., due to a prior forward hint from the mds), use that.
672 * Otherwise, consult frag tree and/or caps to identify the
673 * appropriate mds. If all else fails, choose randomly.
675 * Called under mdsc->mutex.
677 static struct dentry
*get_nonsnap_parent(struct dentry
*dentry
)
680 * we don't need to worry about protecting the d_parent access
681 * here because we never renaming inside the snapped namespace
682 * except to resplice to another snapdir, and either the old or new
683 * result is a valid result.
685 while (!IS_ROOT(dentry
) && ceph_snap(d_inode(dentry
)) != CEPH_NOSNAP
)
686 dentry
= dentry
->d_parent
;
690 static int __choose_mds(struct ceph_mds_client
*mdsc
,
691 struct ceph_mds_request
*req
)
694 struct ceph_inode_info
*ci
;
695 struct ceph_cap
*cap
;
696 int mode
= req
->r_direct_mode
;
698 u32 hash
= req
->r_direct_hash
;
699 bool is_hash
= req
->r_direct_is_hash
;
702 * is there a specific mds we should try? ignore hint if we have
703 * no session and the mds is not up (active or recovering).
705 if (req
->r_resend_mds
>= 0 &&
706 (__have_session(mdsc
, req
->r_resend_mds
) ||
707 ceph_mdsmap_get_state(mdsc
->mdsmap
, req
->r_resend_mds
) > 0)) {
708 dout("choose_mds using resend_mds mds%d\n",
710 return req
->r_resend_mds
;
713 if (mode
== USE_RANDOM_MDS
)
718 inode
= req
->r_inode
;
719 } else if (req
->r_dentry
) {
720 /* ignore race with rename; old or new d_parent is okay */
721 struct dentry
*parent
= req
->r_dentry
->d_parent
;
722 struct inode
*dir
= d_inode(parent
);
724 if (dir
->i_sb
!= mdsc
->fsc
->sb
) {
726 inode
= d_inode(req
->r_dentry
);
727 } else if (ceph_snap(dir
) != CEPH_NOSNAP
) {
728 /* direct snapped/virtual snapdir requests
729 * based on parent dir inode */
730 struct dentry
*dn
= get_nonsnap_parent(parent
);
732 dout("__choose_mds using nonsnap parent %p\n", inode
);
735 inode
= d_inode(req
->r_dentry
);
736 if (!inode
|| mode
== USE_AUTH_MDS
) {
739 hash
= ceph_dentry_hash(dir
, req
->r_dentry
);
745 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode
, (int)is_hash
,
749 ci
= ceph_inode(inode
);
751 if (is_hash
&& S_ISDIR(inode
->i_mode
)) {
752 struct ceph_inode_frag frag
;
755 ceph_choose_frag(ci
, hash
, &frag
, &found
);
757 if (mode
== USE_ANY_MDS
&& frag
.ndist
> 0) {
760 /* choose a random replica */
761 get_random_bytes(&r
, 1);
764 dout("choose_mds %p %llx.%llx "
765 "frag %u mds%d (%d/%d)\n",
766 inode
, ceph_vinop(inode
),
769 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
770 CEPH_MDS_STATE_ACTIVE
)
774 /* since this file/dir wasn't known to be
775 * replicated, then we want to look for the
776 * authoritative mds. */
779 /* choose auth mds */
781 dout("choose_mds %p %llx.%llx "
782 "frag %u mds%d (auth)\n",
783 inode
, ceph_vinop(inode
), frag
.frag
, mds
);
784 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
785 CEPH_MDS_STATE_ACTIVE
)
791 spin_lock(&ci
->i_ceph_lock
);
793 if (mode
== USE_AUTH_MDS
)
794 cap
= ci
->i_auth_cap
;
795 if (!cap
&& !RB_EMPTY_ROOT(&ci
->i_caps
))
796 cap
= rb_entry(rb_first(&ci
->i_caps
), struct ceph_cap
, ci_node
);
798 spin_unlock(&ci
->i_ceph_lock
);
801 mds
= cap
->session
->s_mds
;
802 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
803 inode
, ceph_vinop(inode
), mds
,
804 cap
== ci
->i_auth_cap
? "auth " : "", cap
);
805 spin_unlock(&ci
->i_ceph_lock
);
809 mds
= ceph_mdsmap_get_random_mds(mdsc
->mdsmap
);
810 dout("choose_mds chose random mds%d\n", mds
);
818 static struct ceph_msg
*create_session_msg(u32 op
, u64 seq
)
820 struct ceph_msg
*msg
;
821 struct ceph_mds_session_head
*h
;
823 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
), GFP_NOFS
,
826 pr_err("create_session_msg ENOMEM creating msg\n");
829 h
= msg
->front
.iov_base
;
830 h
->op
= cpu_to_le32(op
);
831 h
->seq
= cpu_to_le64(seq
);
837 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
838 * to include additional client metadata fields.
840 static struct ceph_msg
*create_session_open_msg(struct ceph_mds_client
*mdsc
, u64 seq
)
842 struct ceph_msg
*msg
;
843 struct ceph_mds_session_head
*h
;
845 int metadata_bytes
= 0;
846 int metadata_key_count
= 0;
847 struct ceph_options
*opt
= mdsc
->fsc
->client
->options
;
848 struct ceph_mount_options
*fsopt
= mdsc
->fsc
->mount_options
;
851 const char* metadata
[][2] = {
852 {"hostname", utsname()->nodename
},
853 {"kernel_version", utsname()->release
},
854 {"entity_id", opt
->name
? : ""},
855 {"root", fsopt
->server_path
? : "/"},
859 /* Calculate serialized length of metadata */
860 metadata_bytes
= 4; /* map length */
861 for (i
= 0; metadata
[i
][0] != NULL
; ++i
) {
862 metadata_bytes
+= 8 + strlen(metadata
[i
][0]) +
863 strlen(metadata
[i
][1]);
864 metadata_key_count
++;
867 /* Allocate the message */
868 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
) + metadata_bytes
,
871 pr_err("create_session_msg ENOMEM creating msg\n");
874 h
= msg
->front
.iov_base
;
875 h
->op
= cpu_to_le32(CEPH_SESSION_REQUEST_OPEN
);
876 h
->seq
= cpu_to_le64(seq
);
879 * Serialize client metadata into waiting buffer space, using
880 * the format that userspace expects for map<string, string>
882 * ClientSession messages with metadata are v2
884 msg
->hdr
.version
= cpu_to_le16(2);
885 msg
->hdr
.compat_version
= cpu_to_le16(1);
887 /* The write pointer, following the session_head structure */
888 p
= msg
->front
.iov_base
+ sizeof(*h
);
890 /* Number of entries in the map */
891 ceph_encode_32(&p
, metadata_key_count
);
893 /* Two length-prefixed strings for each entry in the map */
894 for (i
= 0; metadata
[i
][0] != NULL
; ++i
) {
895 size_t const key_len
= strlen(metadata
[i
][0]);
896 size_t const val_len
= strlen(metadata
[i
][1]);
898 ceph_encode_32(&p
, key_len
);
899 memcpy(p
, metadata
[i
][0], key_len
);
901 ceph_encode_32(&p
, val_len
);
902 memcpy(p
, metadata
[i
][1], val_len
);
910 * send session open request.
912 * called under mdsc->mutex
914 static int __open_session(struct ceph_mds_client
*mdsc
,
915 struct ceph_mds_session
*session
)
917 struct ceph_msg
*msg
;
919 int mds
= session
->s_mds
;
921 /* wait for mds to go active? */
922 mstate
= ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
);
923 dout("open_session to mds%d (%s)\n", mds
,
924 ceph_mds_state_name(mstate
));
925 session
->s_state
= CEPH_MDS_SESSION_OPENING
;
926 session
->s_renew_requested
= jiffies
;
928 /* send connect message */
929 msg
= create_session_open_msg(mdsc
, session
->s_seq
);
932 ceph_con_send(&session
->s_con
, msg
);
937 * open sessions for any export targets for the given mds
939 * called under mdsc->mutex
941 static struct ceph_mds_session
*
942 __open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
944 struct ceph_mds_session
*session
;
946 session
= __ceph_lookup_mds_session(mdsc
, target
);
948 session
= register_session(mdsc
, target
);
952 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
953 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
954 __open_session(mdsc
, session
);
959 struct ceph_mds_session
*
960 ceph_mdsc_open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
962 struct ceph_mds_session
*session
;
964 dout("open_export_target_session to mds%d\n", target
);
966 mutex_lock(&mdsc
->mutex
);
967 session
= __open_export_target_session(mdsc
, target
);
968 mutex_unlock(&mdsc
->mutex
);
973 static void __open_export_target_sessions(struct ceph_mds_client
*mdsc
,
974 struct ceph_mds_session
*session
)
976 struct ceph_mds_info
*mi
;
977 struct ceph_mds_session
*ts
;
978 int i
, mds
= session
->s_mds
;
980 if (mds
>= mdsc
->mdsmap
->m_max_mds
)
983 mi
= &mdsc
->mdsmap
->m_info
[mds
];
984 dout("open_export_target_sessions for mds%d (%d targets)\n",
985 session
->s_mds
, mi
->num_export_targets
);
987 for (i
= 0; i
< mi
->num_export_targets
; i
++) {
988 ts
= __open_export_target_session(mdsc
, mi
->export_targets
[i
]);
990 ceph_put_mds_session(ts
);
994 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client
*mdsc
,
995 struct ceph_mds_session
*session
)
997 mutex_lock(&mdsc
->mutex
);
998 __open_export_target_sessions(mdsc
, session
);
999 mutex_unlock(&mdsc
->mutex
);
1006 /* caller holds s_cap_lock, we drop it */
1007 static void cleanup_cap_releases(struct ceph_mds_client
*mdsc
,
1008 struct ceph_mds_session
*session
)
1009 __releases(session
->s_cap_lock
)
1011 LIST_HEAD(tmp_list
);
1012 list_splice_init(&session
->s_cap_releases
, &tmp_list
);
1013 session
->s_num_cap_releases
= 0;
1014 spin_unlock(&session
->s_cap_lock
);
1016 dout("cleanup_cap_releases mds%d\n", session
->s_mds
);
1017 while (!list_empty(&tmp_list
)) {
1018 struct ceph_cap
*cap
;
1019 /* zero out the in-progress message */
1020 cap
= list_first_entry(&tmp_list
,
1021 struct ceph_cap
, session_caps
);
1022 list_del(&cap
->session_caps
);
1023 ceph_put_cap(mdsc
, cap
);
1027 static void cleanup_session_requests(struct ceph_mds_client
*mdsc
,
1028 struct ceph_mds_session
*session
)
1030 struct ceph_mds_request
*req
;
1033 dout("cleanup_session_requests mds%d\n", session
->s_mds
);
1034 mutex_lock(&mdsc
->mutex
);
1035 while (!list_empty(&session
->s_unsafe
)) {
1036 req
= list_first_entry(&session
->s_unsafe
,
1037 struct ceph_mds_request
, r_unsafe_item
);
1038 list_del_init(&req
->r_unsafe_item
);
1039 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1041 __unregister_request(mdsc
, req
);
1043 /* zero r_attempts, so kick_requests() will re-send requests */
1044 p
= rb_first(&mdsc
->request_tree
);
1046 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
1048 if (req
->r_session
&&
1049 req
->r_session
->s_mds
== session
->s_mds
)
1050 req
->r_attempts
= 0;
1052 mutex_unlock(&mdsc
->mutex
);
1056 * Helper to safely iterate over all caps associated with a session, with
1057 * special care taken to handle a racing __ceph_remove_cap().
1059 * Caller must hold session s_mutex.
1061 static int iterate_session_caps(struct ceph_mds_session
*session
,
1062 int (*cb
)(struct inode
*, struct ceph_cap
*,
1065 struct list_head
*p
;
1066 struct ceph_cap
*cap
;
1067 struct inode
*inode
, *last_inode
= NULL
;
1068 struct ceph_cap
*old_cap
= NULL
;
1071 dout("iterate_session_caps %p mds%d\n", session
, session
->s_mds
);
1072 spin_lock(&session
->s_cap_lock
);
1073 p
= session
->s_caps
.next
;
1074 while (p
!= &session
->s_caps
) {
1075 cap
= list_entry(p
, struct ceph_cap
, session_caps
);
1076 inode
= igrab(&cap
->ci
->vfs_inode
);
1081 session
->s_cap_iterator
= cap
;
1082 spin_unlock(&session
->s_cap_lock
);
1089 ceph_put_cap(session
->s_mdsc
, old_cap
);
1093 ret
= cb(inode
, cap
, arg
);
1096 spin_lock(&session
->s_cap_lock
);
1098 if (cap
->ci
== NULL
) {
1099 dout("iterate_session_caps finishing cap %p removal\n",
1101 BUG_ON(cap
->session
!= session
);
1102 cap
->session
= NULL
;
1103 list_del_init(&cap
->session_caps
);
1104 session
->s_nr_caps
--;
1105 if (cap
->queue_release
) {
1106 list_add_tail(&cap
->session_caps
,
1107 &session
->s_cap_releases
);
1108 session
->s_num_cap_releases
++;
1110 old_cap
= cap
; /* put_cap it w/o locks held */
1118 session
->s_cap_iterator
= NULL
;
1119 spin_unlock(&session
->s_cap_lock
);
1123 ceph_put_cap(session
->s_mdsc
, old_cap
);
1128 static int remove_session_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1131 struct ceph_fs_client
*fsc
= (struct ceph_fs_client
*)arg
;
1132 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1133 LIST_HEAD(to_remove
);
1135 bool invalidate
= false;
1137 dout("removing cap %p, ci is %p, inode is %p\n",
1138 cap
, ci
, &ci
->vfs_inode
);
1139 spin_lock(&ci
->i_ceph_lock
);
1140 __ceph_remove_cap(cap
, false);
1141 if (!ci
->i_auth_cap
) {
1142 struct ceph_cap_flush
*cf
;
1143 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
1145 ci
->i_ceph_flags
|= CEPH_I_CAP_DROPPED
;
1147 if (ci
->i_wrbuffer_ref
> 0 &&
1148 ACCESS_ONCE(fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
1151 while (!list_empty(&ci
->i_cap_flush_list
)) {
1152 cf
= list_first_entry(&ci
->i_cap_flush_list
,
1153 struct ceph_cap_flush
, i_list
);
1154 list_move(&cf
->i_list
, &to_remove
);
1157 spin_lock(&mdsc
->cap_dirty_lock
);
1159 list_for_each_entry(cf
, &to_remove
, i_list
)
1160 list_del(&cf
->g_list
);
1162 if (!list_empty(&ci
->i_dirty_item
)) {
1163 pr_warn_ratelimited(
1164 " dropping dirty %s state for %p %lld\n",
1165 ceph_cap_string(ci
->i_dirty_caps
),
1166 inode
, ceph_ino(inode
));
1167 ci
->i_dirty_caps
= 0;
1168 list_del_init(&ci
->i_dirty_item
);
1171 if (!list_empty(&ci
->i_flushing_item
)) {
1172 pr_warn_ratelimited(
1173 " dropping dirty+flushing %s state for %p %lld\n",
1174 ceph_cap_string(ci
->i_flushing_caps
),
1175 inode
, ceph_ino(inode
));
1176 ci
->i_flushing_caps
= 0;
1177 list_del_init(&ci
->i_flushing_item
);
1178 mdsc
->num_cap_flushing
--;
1181 spin_unlock(&mdsc
->cap_dirty_lock
);
1183 if (!ci
->i_dirty_caps
&& ci
->i_prealloc_cap_flush
) {
1184 list_add(&ci
->i_prealloc_cap_flush
->i_list
, &to_remove
);
1185 ci
->i_prealloc_cap_flush
= NULL
;
1188 spin_unlock(&ci
->i_ceph_lock
);
1189 while (!list_empty(&to_remove
)) {
1190 struct ceph_cap_flush
*cf
;
1191 cf
= list_first_entry(&to_remove
,
1192 struct ceph_cap_flush
, i_list
);
1193 list_del(&cf
->i_list
);
1194 ceph_free_cap_flush(cf
);
1197 wake_up_all(&ci
->i_cap_wq
);
1199 ceph_queue_invalidate(inode
);
1206 * caller must hold session s_mutex
1208 static void remove_session_caps(struct ceph_mds_session
*session
)
1210 struct ceph_fs_client
*fsc
= session
->s_mdsc
->fsc
;
1211 struct super_block
*sb
= fsc
->sb
;
1212 dout("remove_session_caps on %p\n", session
);
1213 iterate_session_caps(session
, remove_session_caps_cb
, fsc
);
1215 wake_up_all(&fsc
->mdsc
->cap_flushing_wq
);
1217 spin_lock(&session
->s_cap_lock
);
1218 if (session
->s_nr_caps
> 0) {
1219 struct inode
*inode
;
1220 struct ceph_cap
*cap
, *prev
= NULL
;
1221 struct ceph_vino vino
;
1223 * iterate_session_caps() skips inodes that are being
1224 * deleted, we need to wait until deletions are complete.
1225 * __wait_on_freeing_inode() is designed for the job,
1226 * but it is not exported, so use lookup inode function
1229 while (!list_empty(&session
->s_caps
)) {
1230 cap
= list_entry(session
->s_caps
.next
,
1231 struct ceph_cap
, session_caps
);
1235 vino
= cap
->ci
->i_vino
;
1236 spin_unlock(&session
->s_cap_lock
);
1238 inode
= ceph_find_inode(sb
, vino
);
1241 spin_lock(&session
->s_cap_lock
);
1245 // drop cap expires and unlock s_cap_lock
1246 cleanup_cap_releases(session
->s_mdsc
, session
);
1248 BUG_ON(session
->s_nr_caps
> 0);
1249 BUG_ON(!list_empty(&session
->s_cap_flushing
));
1253 * wake up any threads waiting on this session's caps. if the cap is
1254 * old (didn't get renewed on the client reconnect), remove it now.
1256 * caller must hold s_mutex.
1258 static int wake_up_session_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1261 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1264 spin_lock(&ci
->i_ceph_lock
);
1265 ci
->i_wanted_max_size
= 0;
1266 ci
->i_requested_max_size
= 0;
1267 spin_unlock(&ci
->i_ceph_lock
);
1269 wake_up_all(&ci
->i_cap_wq
);
1273 static void wake_up_session_caps(struct ceph_mds_session
*session
,
1276 dout("wake_up_session_caps %p mds%d\n", session
, session
->s_mds
);
1277 iterate_session_caps(session
, wake_up_session_cb
,
1278 (void *)(unsigned long)reconnect
);
1282 * Send periodic message to MDS renewing all currently held caps. The
1283 * ack will reset the expiration for all caps from this session.
1285 * caller holds s_mutex
1287 static int send_renew_caps(struct ceph_mds_client
*mdsc
,
1288 struct ceph_mds_session
*session
)
1290 struct ceph_msg
*msg
;
1293 if (time_after_eq(jiffies
, session
->s_cap_ttl
) &&
1294 time_after_eq(session
->s_cap_ttl
, session
->s_renew_requested
))
1295 pr_info("mds%d caps stale\n", session
->s_mds
);
1296 session
->s_renew_requested
= jiffies
;
1298 /* do not try to renew caps until a recovering mds has reconnected
1299 * with its clients. */
1300 state
= ceph_mdsmap_get_state(mdsc
->mdsmap
, session
->s_mds
);
1301 if (state
< CEPH_MDS_STATE_RECONNECT
) {
1302 dout("send_renew_caps ignoring mds%d (%s)\n",
1303 session
->s_mds
, ceph_mds_state_name(state
));
1307 dout("send_renew_caps to mds%d (%s)\n", session
->s_mds
,
1308 ceph_mds_state_name(state
));
1309 msg
= create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS
,
1310 ++session
->s_renew_seq
);
1313 ceph_con_send(&session
->s_con
, msg
);
1317 static int send_flushmsg_ack(struct ceph_mds_client
*mdsc
,
1318 struct ceph_mds_session
*session
, u64 seq
)
1320 struct ceph_msg
*msg
;
1322 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1323 session
->s_mds
, ceph_session_state_name(session
->s_state
), seq
);
1324 msg
= create_session_msg(CEPH_SESSION_FLUSHMSG_ACK
, seq
);
1327 ceph_con_send(&session
->s_con
, msg
);
1333 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1335 * Called under session->s_mutex
1337 static void renewed_caps(struct ceph_mds_client
*mdsc
,
1338 struct ceph_mds_session
*session
, int is_renew
)
1343 spin_lock(&session
->s_cap_lock
);
1344 was_stale
= is_renew
&& time_after_eq(jiffies
, session
->s_cap_ttl
);
1346 session
->s_cap_ttl
= session
->s_renew_requested
+
1347 mdsc
->mdsmap
->m_session_timeout
*HZ
;
1350 if (time_before(jiffies
, session
->s_cap_ttl
)) {
1351 pr_info("mds%d caps renewed\n", session
->s_mds
);
1354 pr_info("mds%d caps still stale\n", session
->s_mds
);
1357 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1358 session
->s_mds
, session
->s_cap_ttl
, was_stale
? "stale" : "fresh",
1359 time_before(jiffies
, session
->s_cap_ttl
) ? "stale" : "fresh");
1360 spin_unlock(&session
->s_cap_lock
);
1363 wake_up_session_caps(session
, 0);
1367 * send a session close request
1369 static int request_close_session(struct ceph_mds_client
*mdsc
,
1370 struct ceph_mds_session
*session
)
1372 struct ceph_msg
*msg
;
1374 dout("request_close_session mds%d state %s seq %lld\n",
1375 session
->s_mds
, ceph_session_state_name(session
->s_state
),
1377 msg
= create_session_msg(CEPH_SESSION_REQUEST_CLOSE
, session
->s_seq
);
1380 ceph_con_send(&session
->s_con
, msg
);
1385 * Called with s_mutex held.
1387 static int __close_session(struct ceph_mds_client
*mdsc
,
1388 struct ceph_mds_session
*session
)
1390 if (session
->s_state
>= CEPH_MDS_SESSION_CLOSING
)
1392 session
->s_state
= CEPH_MDS_SESSION_CLOSING
;
1393 return request_close_session(mdsc
, session
);
1397 * Trim old(er) caps.
1399 * Because we can't cache an inode without one or more caps, we do
1400 * this indirectly: if a cap is unused, we prune its aliases, at which
1401 * point the inode will hopefully get dropped to.
1403 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1404 * memory pressure from the MDS, though, so it needn't be perfect.
1406 static int trim_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
, void *arg
)
1408 struct ceph_mds_session
*session
= arg
;
1409 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1410 int used
, wanted
, oissued
, mine
;
1412 if (session
->s_trim_caps
<= 0)
1415 spin_lock(&ci
->i_ceph_lock
);
1416 mine
= cap
->issued
| cap
->implemented
;
1417 used
= __ceph_caps_used(ci
);
1418 wanted
= __ceph_caps_file_wanted(ci
);
1419 oissued
= __ceph_caps_issued_other(ci
, cap
);
1421 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1422 inode
, cap
, ceph_cap_string(mine
), ceph_cap_string(oissued
),
1423 ceph_cap_string(used
), ceph_cap_string(wanted
));
1424 if (cap
== ci
->i_auth_cap
) {
1425 if (ci
->i_dirty_caps
|| ci
->i_flushing_caps
||
1426 !list_empty(&ci
->i_cap_snaps
))
1428 if ((used
| wanted
) & CEPH_CAP_ANY_WR
)
1431 /* The inode has cached pages, but it's no longer used.
1432 * we can safely drop it */
1433 if (wanted
== 0 && used
== CEPH_CAP_FILE_CACHE
&&
1434 !(oissued
& CEPH_CAP_FILE_CACHE
)) {
1438 if ((used
| wanted
) & ~oissued
& mine
)
1439 goto out
; /* we need these caps */
1441 session
->s_trim_caps
--;
1443 /* we aren't the only cap.. just remove us */
1444 __ceph_remove_cap(cap
, true);
1446 /* try dropping referring dentries */
1447 spin_unlock(&ci
->i_ceph_lock
);
1448 d_prune_aliases(inode
);
1449 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1450 inode
, cap
, atomic_read(&inode
->i_count
));
1455 spin_unlock(&ci
->i_ceph_lock
);
1460 * Trim session cap count down to some max number.
1462 static int trim_caps(struct ceph_mds_client
*mdsc
,
1463 struct ceph_mds_session
*session
,
1466 int trim_caps
= session
->s_nr_caps
- max_caps
;
1468 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1469 session
->s_mds
, session
->s_nr_caps
, max_caps
, trim_caps
);
1470 if (trim_caps
> 0) {
1471 session
->s_trim_caps
= trim_caps
;
1472 iterate_session_caps(session
, trim_caps_cb
, session
);
1473 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1474 session
->s_mds
, session
->s_nr_caps
, max_caps
,
1475 trim_caps
- session
->s_trim_caps
);
1476 session
->s_trim_caps
= 0;
1479 ceph_send_cap_releases(mdsc
, session
);
1483 static int check_caps_flush(struct ceph_mds_client
*mdsc
,
1488 spin_lock(&mdsc
->cap_dirty_lock
);
1489 if (!list_empty(&mdsc
->cap_flush_list
)) {
1490 struct ceph_cap_flush
*cf
=
1491 list_first_entry(&mdsc
->cap_flush_list
,
1492 struct ceph_cap_flush
, g_list
);
1493 if (cf
->tid
<= want_flush_tid
) {
1494 dout("check_caps_flush still flushing tid "
1495 "%llu <= %llu\n", cf
->tid
, want_flush_tid
);
1499 spin_unlock(&mdsc
->cap_dirty_lock
);
1504 * flush all dirty inode data to disk.
1506 * returns true if we've flushed through want_flush_tid
1508 static void wait_caps_flush(struct ceph_mds_client
*mdsc
,
1511 dout("check_caps_flush want %llu\n", want_flush_tid
);
1513 wait_event(mdsc
->cap_flushing_wq
,
1514 check_caps_flush(mdsc
, want_flush_tid
));
1516 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid
);
1520 * called under s_mutex
1522 void ceph_send_cap_releases(struct ceph_mds_client
*mdsc
,
1523 struct ceph_mds_session
*session
)
1525 struct ceph_msg
*msg
= NULL
;
1526 struct ceph_mds_cap_release
*head
;
1527 struct ceph_mds_cap_item
*item
;
1528 struct ceph_cap
*cap
;
1529 LIST_HEAD(tmp_list
);
1530 int num_cap_releases
;
1532 spin_lock(&session
->s_cap_lock
);
1534 list_splice_init(&session
->s_cap_releases
, &tmp_list
);
1535 num_cap_releases
= session
->s_num_cap_releases
;
1536 session
->s_num_cap_releases
= 0;
1537 spin_unlock(&session
->s_cap_lock
);
1539 while (!list_empty(&tmp_list
)) {
1541 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE
,
1542 PAGE_SIZE
, GFP_NOFS
, false);
1545 head
= msg
->front
.iov_base
;
1546 head
->num
= cpu_to_le32(0);
1547 msg
->front
.iov_len
= sizeof(*head
);
1549 cap
= list_first_entry(&tmp_list
, struct ceph_cap
,
1551 list_del(&cap
->session_caps
);
1554 head
= msg
->front
.iov_base
;
1555 le32_add_cpu(&head
->num
, 1);
1556 item
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1557 item
->ino
= cpu_to_le64(cap
->cap_ino
);
1558 item
->cap_id
= cpu_to_le64(cap
->cap_id
);
1559 item
->migrate_seq
= cpu_to_le32(cap
->mseq
);
1560 item
->seq
= cpu_to_le32(cap
->issue_seq
);
1561 msg
->front
.iov_len
+= sizeof(*item
);
1563 ceph_put_cap(mdsc
, cap
);
1565 if (le32_to_cpu(head
->num
) == CEPH_CAPS_PER_RELEASE
) {
1566 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1567 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1568 ceph_con_send(&session
->s_con
, msg
);
1573 BUG_ON(num_cap_releases
!= 0);
1575 spin_lock(&session
->s_cap_lock
);
1576 if (!list_empty(&session
->s_cap_releases
))
1578 spin_unlock(&session
->s_cap_lock
);
1581 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1582 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1583 ceph_con_send(&session
->s_con
, msg
);
1587 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1589 spin_lock(&session
->s_cap_lock
);
1590 list_splice(&tmp_list
, &session
->s_cap_releases
);
1591 session
->s_num_cap_releases
+= num_cap_releases
;
1592 spin_unlock(&session
->s_cap_lock
);
1599 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request
*req
,
1602 struct ceph_inode_info
*ci
= ceph_inode(dir
);
1603 struct ceph_mds_reply_info_parsed
*rinfo
= &req
->r_reply_info
;
1604 struct ceph_mount_options
*opt
= req
->r_mdsc
->fsc
->mount_options
;
1605 size_t size
= sizeof(struct ceph_mds_reply_dir_entry
);
1606 int order
, num_entries
;
1608 spin_lock(&ci
->i_ceph_lock
);
1609 num_entries
= ci
->i_files
+ ci
->i_subdirs
;
1610 spin_unlock(&ci
->i_ceph_lock
);
1611 num_entries
= max(num_entries
, 1);
1612 num_entries
= min(num_entries
, opt
->max_readdir
);
1614 order
= get_order(size
* num_entries
);
1615 while (order
>= 0) {
1616 rinfo
->dir_entries
= (void*)__get_free_pages(GFP_KERNEL
|
1619 if (rinfo
->dir_entries
)
1623 if (!rinfo
->dir_entries
)
1626 num_entries
= (PAGE_SIZE
<< order
) / size
;
1627 num_entries
= min(num_entries
, opt
->max_readdir
);
1629 rinfo
->dir_buf_size
= PAGE_SIZE
<< order
;
1630 req
->r_num_caps
= num_entries
+ 1;
1631 req
->r_args
.readdir
.max_entries
= cpu_to_le32(num_entries
);
1632 req
->r_args
.readdir
.max_bytes
= cpu_to_le32(opt
->max_readdir_bytes
);
1637 * Create an mds request.
1639 struct ceph_mds_request
*
1640 ceph_mdsc_create_request(struct ceph_mds_client
*mdsc
, int op
, int mode
)
1642 struct ceph_mds_request
*req
= kzalloc(sizeof(*req
), GFP_NOFS
);
1645 return ERR_PTR(-ENOMEM
);
1647 mutex_init(&req
->r_fill_mutex
);
1649 req
->r_started
= jiffies
;
1650 req
->r_resend_mds
= -1;
1651 INIT_LIST_HEAD(&req
->r_unsafe_dir_item
);
1652 INIT_LIST_HEAD(&req
->r_unsafe_target_item
);
1654 kref_init(&req
->r_kref
);
1655 RB_CLEAR_NODE(&req
->r_node
);
1656 INIT_LIST_HEAD(&req
->r_wait
);
1657 init_completion(&req
->r_completion
);
1658 init_completion(&req
->r_safe_completion
);
1659 INIT_LIST_HEAD(&req
->r_unsafe_item
);
1661 req
->r_stamp
= current_fs_time(mdsc
->fsc
->sb
);
1664 req
->r_direct_mode
= mode
;
1669 * return oldest (lowest) request, tid in request tree, 0 if none.
1671 * called under mdsc->mutex.
1673 static struct ceph_mds_request
*__get_oldest_req(struct ceph_mds_client
*mdsc
)
1675 if (RB_EMPTY_ROOT(&mdsc
->request_tree
))
1677 return rb_entry(rb_first(&mdsc
->request_tree
),
1678 struct ceph_mds_request
, r_node
);
1681 static inline u64
__get_oldest_tid(struct ceph_mds_client
*mdsc
)
1683 return mdsc
->oldest_tid
;
1687 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1688 * on build_path_from_dentry in fs/cifs/dir.c.
1690 * If @stop_on_nosnap, generate path relative to the first non-snapped
1693 * Encode hidden .snap dirs as a double /, i.e.
1694 * foo/.snap/bar -> foo//bar
1696 char *ceph_mdsc_build_path(struct dentry
*dentry
, int *plen
, u64
*base
,
1699 struct dentry
*temp
;
1705 return ERR_PTR(-EINVAL
);
1709 seq
= read_seqbegin(&rename_lock
);
1711 for (temp
= dentry
; !IS_ROOT(temp
);) {
1712 struct inode
*inode
= d_inode(temp
);
1713 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
)
1714 len
++; /* slash only */
1715 else if (stop_on_nosnap
&& inode
&&
1716 ceph_snap(inode
) == CEPH_NOSNAP
)
1719 len
+= 1 + temp
->d_name
.len
;
1720 temp
= temp
->d_parent
;
1724 len
--; /* no leading '/' */
1726 path
= kmalloc(len
+1, GFP_NOFS
);
1728 return ERR_PTR(-ENOMEM
);
1730 path
[pos
] = 0; /* trailing null */
1732 for (temp
= dentry
; !IS_ROOT(temp
) && pos
!= 0; ) {
1733 struct inode
*inode
;
1735 spin_lock(&temp
->d_lock
);
1736 inode
= d_inode(temp
);
1737 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
) {
1738 dout("build_path path+%d: %p SNAPDIR\n",
1740 } else if (stop_on_nosnap
&& inode
&&
1741 ceph_snap(inode
) == CEPH_NOSNAP
) {
1742 spin_unlock(&temp
->d_lock
);
1745 pos
-= temp
->d_name
.len
;
1747 spin_unlock(&temp
->d_lock
);
1750 strncpy(path
+ pos
, temp
->d_name
.name
,
1753 spin_unlock(&temp
->d_lock
);
1756 temp
= temp
->d_parent
;
1759 if (pos
!= 0 || read_seqretry(&rename_lock
, seq
)) {
1760 pr_err("build_path did not end path lookup where "
1761 "expected, namelen is %d, pos is %d\n", len
, pos
);
1762 /* presumably this is only possible if racing with a
1763 rename of one of the parent directories (we can not
1764 lock the dentries above us to prevent this, but
1765 retrying should be harmless) */
1770 *base
= ceph_ino(d_inode(temp
));
1772 dout("build_path on %p %d built %llx '%.*s'\n",
1773 dentry
, d_count(dentry
), *base
, len
, path
);
1777 static int build_dentry_path(struct dentry
*dentry
,
1778 const char **ppath
, int *ppathlen
, u64
*pino
,
1783 if (ceph_snap(d_inode(dentry
->d_parent
)) == CEPH_NOSNAP
) {
1784 *pino
= ceph_ino(d_inode(dentry
->d_parent
));
1785 *ppath
= dentry
->d_name
.name
;
1786 *ppathlen
= dentry
->d_name
.len
;
1789 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1791 return PTR_ERR(path
);
1797 static int build_inode_path(struct inode
*inode
,
1798 const char **ppath
, int *ppathlen
, u64
*pino
,
1801 struct dentry
*dentry
;
1804 if (ceph_snap(inode
) == CEPH_NOSNAP
) {
1805 *pino
= ceph_ino(inode
);
1809 dentry
= d_find_alias(inode
);
1810 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1813 return PTR_ERR(path
);
1820 * request arguments may be specified via an inode *, a dentry *, or
1821 * an explicit ino+path.
1823 static int set_request_path_attr(struct inode
*rinode
, struct dentry
*rdentry
,
1824 const char *rpath
, u64 rino
,
1825 const char **ppath
, int *pathlen
,
1826 u64
*ino
, int *freepath
)
1831 r
= build_inode_path(rinode
, ppath
, pathlen
, ino
, freepath
);
1832 dout(" inode %p %llx.%llx\n", rinode
, ceph_ino(rinode
),
1834 } else if (rdentry
) {
1835 r
= build_dentry_path(rdentry
, ppath
, pathlen
, ino
, freepath
);
1836 dout(" dentry %p %llx/%.*s\n", rdentry
, *ino
, *pathlen
,
1838 } else if (rpath
|| rino
) {
1841 *pathlen
= rpath
? strlen(rpath
) : 0;
1842 dout(" path %.*s\n", *pathlen
, rpath
);
1849 * called under mdsc->mutex
1851 static struct ceph_msg
*create_request_message(struct ceph_mds_client
*mdsc
,
1852 struct ceph_mds_request
*req
,
1853 int mds
, bool drop_cap_releases
)
1855 struct ceph_msg
*msg
;
1856 struct ceph_mds_request_head
*head
;
1857 const char *path1
= NULL
;
1858 const char *path2
= NULL
;
1859 u64 ino1
= 0, ino2
= 0;
1860 int pathlen1
= 0, pathlen2
= 0;
1861 int freepath1
= 0, freepath2
= 0;
1867 ret
= set_request_path_attr(req
->r_inode
, req
->r_dentry
,
1868 req
->r_path1
, req
->r_ino1
.ino
,
1869 &path1
, &pathlen1
, &ino1
, &freepath1
);
1875 ret
= set_request_path_attr(NULL
, req
->r_old_dentry
,
1876 req
->r_path2
, req
->r_ino2
.ino
,
1877 &path2
, &pathlen2
, &ino2
, &freepath2
);
1883 len
= sizeof(*head
) +
1884 pathlen1
+ pathlen2
+ 2*(1 + sizeof(u32
) + sizeof(u64
)) +
1885 sizeof(struct ceph_timespec
);
1887 /* calculate (max) length for cap releases */
1888 len
+= sizeof(struct ceph_mds_request_release
) *
1889 (!!req
->r_inode_drop
+ !!req
->r_dentry_drop
+
1890 !!req
->r_old_inode_drop
+ !!req
->r_old_dentry_drop
);
1891 if (req
->r_dentry_drop
)
1892 len
+= req
->r_dentry
->d_name
.len
;
1893 if (req
->r_old_dentry_drop
)
1894 len
+= req
->r_old_dentry
->d_name
.len
;
1896 msg
= ceph_msg_new(CEPH_MSG_CLIENT_REQUEST
, len
, GFP_NOFS
, false);
1898 msg
= ERR_PTR(-ENOMEM
);
1902 msg
->hdr
.version
= cpu_to_le16(2);
1903 msg
->hdr
.tid
= cpu_to_le64(req
->r_tid
);
1905 head
= msg
->front
.iov_base
;
1906 p
= msg
->front
.iov_base
+ sizeof(*head
);
1907 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1909 head
->mdsmap_epoch
= cpu_to_le32(mdsc
->mdsmap
->m_epoch
);
1910 head
->op
= cpu_to_le32(req
->r_op
);
1911 head
->caller_uid
= cpu_to_le32(from_kuid(&init_user_ns
, req
->r_uid
));
1912 head
->caller_gid
= cpu_to_le32(from_kgid(&init_user_ns
, req
->r_gid
));
1913 head
->args
= req
->r_args
;
1915 ceph_encode_filepath(&p
, end
, ino1
, path1
);
1916 ceph_encode_filepath(&p
, end
, ino2
, path2
);
1918 /* make note of release offset, in case we need to replay */
1919 req
->r_request_release_offset
= p
- msg
->front
.iov_base
;
1923 if (req
->r_inode_drop
)
1924 releases
+= ceph_encode_inode_release(&p
,
1925 req
->r_inode
? req
->r_inode
: d_inode(req
->r_dentry
),
1926 mds
, req
->r_inode_drop
, req
->r_inode_unless
, 0);
1927 if (req
->r_dentry_drop
)
1928 releases
+= ceph_encode_dentry_release(&p
, req
->r_dentry
,
1929 mds
, req
->r_dentry_drop
, req
->r_dentry_unless
);
1930 if (req
->r_old_dentry_drop
)
1931 releases
+= ceph_encode_dentry_release(&p
, req
->r_old_dentry
,
1932 mds
, req
->r_old_dentry_drop
, req
->r_old_dentry_unless
);
1933 if (req
->r_old_inode_drop
)
1934 releases
+= ceph_encode_inode_release(&p
,
1935 d_inode(req
->r_old_dentry
),
1936 mds
, req
->r_old_inode_drop
, req
->r_old_inode_unless
, 0);
1938 if (drop_cap_releases
) {
1940 p
= msg
->front
.iov_base
+ req
->r_request_release_offset
;
1943 head
->num_releases
= cpu_to_le16(releases
);
1947 struct ceph_timespec ts
;
1948 ceph_encode_timespec(&ts
, &req
->r_stamp
);
1949 ceph_encode_copy(&p
, &ts
, sizeof(ts
));
1953 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
1954 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1956 if (req
->r_pagelist
) {
1957 struct ceph_pagelist
*pagelist
= req
->r_pagelist
;
1958 atomic_inc(&pagelist
->refcnt
);
1959 ceph_msg_data_add_pagelist(msg
, pagelist
);
1960 msg
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
1962 msg
->hdr
.data_len
= 0;
1965 msg
->hdr
.data_off
= cpu_to_le16(0);
1969 kfree((char *)path2
);
1972 kfree((char *)path1
);
1978 * called under mdsc->mutex if error, under no mutex if
1981 static void complete_request(struct ceph_mds_client
*mdsc
,
1982 struct ceph_mds_request
*req
)
1984 if (req
->r_callback
)
1985 req
->r_callback(mdsc
, req
);
1987 complete_all(&req
->r_completion
);
1991 * called under mdsc->mutex
1993 static int __prepare_send_request(struct ceph_mds_client
*mdsc
,
1994 struct ceph_mds_request
*req
,
1995 int mds
, bool drop_cap_releases
)
1997 struct ceph_mds_request_head
*rhead
;
1998 struct ceph_msg
*msg
;
2003 struct ceph_cap
*cap
=
2004 ceph_get_cap_for_mds(ceph_inode(req
->r_inode
), mds
);
2007 req
->r_sent_on_mseq
= cap
->mseq
;
2009 req
->r_sent_on_mseq
= -1;
2011 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req
,
2012 req
->r_tid
, ceph_mds_op_name(req
->r_op
), req
->r_attempts
);
2014 if (req
->r_got_unsafe
) {
2017 * Replay. Do not regenerate message (and rebuild
2018 * paths, etc.); just use the original message.
2019 * Rebuilding paths will break for renames because
2020 * d_move mangles the src name.
2022 msg
= req
->r_request
;
2023 rhead
= msg
->front
.iov_base
;
2025 flags
= le32_to_cpu(rhead
->flags
);
2026 flags
|= CEPH_MDS_FLAG_REPLAY
;
2027 rhead
->flags
= cpu_to_le32(flags
);
2029 if (req
->r_target_inode
)
2030 rhead
->ino
= cpu_to_le64(ceph_ino(req
->r_target_inode
));
2032 rhead
->num_retry
= req
->r_attempts
- 1;
2034 /* remove cap/dentry releases from message */
2035 rhead
->num_releases
= 0;
2038 p
= msg
->front
.iov_base
+ req
->r_request_release_offset
;
2040 struct ceph_timespec ts
;
2041 ceph_encode_timespec(&ts
, &req
->r_stamp
);
2042 ceph_encode_copy(&p
, &ts
, sizeof(ts
));
2045 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
2046 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
2050 if (req
->r_request
) {
2051 ceph_msg_put(req
->r_request
);
2052 req
->r_request
= NULL
;
2054 msg
= create_request_message(mdsc
, req
, mds
, drop_cap_releases
);
2056 req
->r_err
= PTR_ERR(msg
);
2057 return PTR_ERR(msg
);
2059 req
->r_request
= msg
;
2061 rhead
= msg
->front
.iov_base
;
2062 rhead
->oldest_client_tid
= cpu_to_le64(__get_oldest_tid(mdsc
));
2063 if (req
->r_got_unsafe
)
2064 flags
|= CEPH_MDS_FLAG_REPLAY
;
2065 if (req
->r_locked_dir
)
2066 flags
|= CEPH_MDS_FLAG_WANT_DENTRY
;
2067 rhead
->flags
= cpu_to_le32(flags
);
2068 rhead
->num_fwd
= req
->r_num_fwd
;
2069 rhead
->num_retry
= req
->r_attempts
- 1;
2072 dout(" r_locked_dir = %p\n", req
->r_locked_dir
);
2077 * send request, or put it on the appropriate wait list.
2079 static int __do_request(struct ceph_mds_client
*mdsc
,
2080 struct ceph_mds_request
*req
)
2082 struct ceph_mds_session
*session
= NULL
;
2086 if (req
->r_err
|| req
->r_got_result
) {
2088 __unregister_request(mdsc
, req
);
2092 if (req
->r_timeout
&&
2093 time_after_eq(jiffies
, req
->r_started
+ req
->r_timeout
)) {
2094 dout("do_request timed out\n");
2098 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
) {
2099 dout("do_request forced umount\n");
2103 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_MOUNTING
) {
2104 if (mdsc
->mdsmap_err
) {
2105 err
= mdsc
->mdsmap_err
;
2106 dout("do_request mdsmap err %d\n", err
);
2109 if (mdsc
->mdsmap
->m_epoch
== 0) {
2110 dout("do_request no mdsmap, waiting for map\n");
2111 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
2114 if (!(mdsc
->fsc
->mount_options
->flags
&
2115 CEPH_MOUNT_OPT_MOUNTWAIT
) &&
2116 !ceph_mdsmap_is_cluster_available(mdsc
->mdsmap
)) {
2118 pr_info("probably no mds server is up\n");
2123 put_request_session(req
);
2125 mds
= __choose_mds(mdsc
, req
);
2127 ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) < CEPH_MDS_STATE_ACTIVE
) {
2128 dout("do_request no mds or not active, waiting for map\n");
2129 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
2133 /* get, open session */
2134 session
= __ceph_lookup_mds_session(mdsc
, mds
);
2136 session
= register_session(mdsc
, mds
);
2137 if (IS_ERR(session
)) {
2138 err
= PTR_ERR(session
);
2142 req
->r_session
= get_session(session
);
2144 dout("do_request mds%d session %p state %s\n", mds
, session
,
2145 ceph_session_state_name(session
->s_state
));
2146 if (session
->s_state
!= CEPH_MDS_SESSION_OPEN
&&
2147 session
->s_state
!= CEPH_MDS_SESSION_HUNG
) {
2148 if (session
->s_state
== CEPH_MDS_SESSION_REJECTED
) {
2152 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
2153 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
2154 __open_session(mdsc
, session
);
2155 list_add(&req
->r_wait
, &session
->s_waiting
);
2160 req
->r_resend_mds
= -1; /* forget any previous mds hint */
2162 if (req
->r_request_started
== 0) /* note request start time */
2163 req
->r_request_started
= jiffies
;
2165 err
= __prepare_send_request(mdsc
, req
, mds
, false);
2167 ceph_msg_get(req
->r_request
);
2168 ceph_con_send(&session
->s_con
, req
->r_request
);
2172 ceph_put_mds_session(session
);
2175 dout("__do_request early error %d\n", err
);
2177 complete_request(mdsc
, req
);
2178 __unregister_request(mdsc
, req
);
2185 * called under mdsc->mutex
2187 static void __wake_requests(struct ceph_mds_client
*mdsc
,
2188 struct list_head
*head
)
2190 struct ceph_mds_request
*req
;
2191 LIST_HEAD(tmp_list
);
2193 list_splice_init(head
, &tmp_list
);
2195 while (!list_empty(&tmp_list
)) {
2196 req
= list_entry(tmp_list
.next
,
2197 struct ceph_mds_request
, r_wait
);
2198 list_del_init(&req
->r_wait
);
2199 dout(" wake request %p tid %llu\n", req
, req
->r_tid
);
2200 __do_request(mdsc
, req
);
2205 * Wake up threads with requests pending for @mds, so that they can
2206 * resubmit their requests to a possibly different mds.
2208 static void kick_requests(struct ceph_mds_client
*mdsc
, int mds
)
2210 struct ceph_mds_request
*req
;
2211 struct rb_node
*p
= rb_first(&mdsc
->request_tree
);
2213 dout("kick_requests mds%d\n", mds
);
2215 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2217 if (req
->r_got_unsafe
)
2219 if (req
->r_attempts
> 0)
2220 continue; /* only new requests */
2221 if (req
->r_session
&&
2222 req
->r_session
->s_mds
== mds
) {
2223 dout(" kicking tid %llu\n", req
->r_tid
);
2224 list_del_init(&req
->r_wait
);
2225 __do_request(mdsc
, req
);
2230 void ceph_mdsc_submit_request(struct ceph_mds_client
*mdsc
,
2231 struct ceph_mds_request
*req
)
2233 dout("submit_request on %p\n", req
);
2234 mutex_lock(&mdsc
->mutex
);
2235 __register_request(mdsc
, req
, NULL
);
2236 __do_request(mdsc
, req
);
2237 mutex_unlock(&mdsc
->mutex
);
2241 * Synchrously perform an mds request. Take care of all of the
2242 * session setup, forwarding, retry details.
2244 int ceph_mdsc_do_request(struct ceph_mds_client
*mdsc
,
2246 struct ceph_mds_request
*req
)
2250 dout("do_request on %p\n", req
);
2252 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2254 ceph_get_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
2255 if (req
->r_locked_dir
)
2256 ceph_get_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
2257 if (req
->r_old_dentry_dir
)
2258 ceph_get_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
2262 mutex_lock(&mdsc
->mutex
);
2263 __register_request(mdsc
, req
, dir
);
2264 __do_request(mdsc
, req
);
2272 mutex_unlock(&mdsc
->mutex
);
2273 dout("do_request waiting\n");
2274 if (!req
->r_timeout
&& req
->r_wait_for_completion
) {
2275 err
= req
->r_wait_for_completion(mdsc
, req
);
2277 long timeleft
= wait_for_completion_killable_timeout(
2279 ceph_timeout_jiffies(req
->r_timeout
));
2283 err
= -EIO
; /* timed out */
2285 err
= timeleft
; /* killed */
2287 dout("do_request waited, got %d\n", err
);
2288 mutex_lock(&mdsc
->mutex
);
2290 /* only abort if we didn't race with a real reply */
2291 if (req
->r_got_result
) {
2292 err
= le32_to_cpu(req
->r_reply_info
.head
->result
);
2293 } else if (err
< 0) {
2294 dout("aborted request %lld with %d\n", req
->r_tid
, err
);
2297 * ensure we aren't running concurrently with
2298 * ceph_fill_trace or ceph_readdir_prepopulate, which
2299 * rely on locks (dir mutex) held by our caller.
2301 mutex_lock(&req
->r_fill_mutex
);
2303 req
->r_aborted
= true;
2304 mutex_unlock(&req
->r_fill_mutex
);
2306 if (req
->r_locked_dir
&&
2307 (req
->r_op
& CEPH_MDS_OP_WRITE
))
2308 ceph_invalidate_dir_request(req
);
2314 mutex_unlock(&mdsc
->mutex
);
2315 dout("do_request %p done, result %d\n", req
, err
);
2320 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2321 * namespace request.
2323 void ceph_invalidate_dir_request(struct ceph_mds_request
*req
)
2325 struct inode
*inode
= req
->r_locked_dir
;
2327 dout("invalidate_dir_request %p (complete, lease(s))\n", inode
);
2329 ceph_dir_clear_complete(inode
);
2331 ceph_invalidate_dentry_lease(req
->r_dentry
);
2332 if (req
->r_old_dentry
)
2333 ceph_invalidate_dentry_lease(req
->r_old_dentry
);
2339 * We take the session mutex and parse and process the reply immediately.
2340 * This preserves the logical ordering of replies, capabilities, etc., sent
2341 * by the MDS as they are applied to our local cache.
2343 static void handle_reply(struct ceph_mds_session
*session
, struct ceph_msg
*msg
)
2345 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2346 struct ceph_mds_request
*req
;
2347 struct ceph_mds_reply_head
*head
= msg
->front
.iov_base
;
2348 struct ceph_mds_reply_info_parsed
*rinfo
; /* parsed reply info */
2349 struct ceph_snap_realm
*realm
;
2352 int mds
= session
->s_mds
;
2354 if (msg
->front
.iov_len
< sizeof(*head
)) {
2355 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2360 /* get request, session */
2361 tid
= le64_to_cpu(msg
->hdr
.tid
);
2362 mutex_lock(&mdsc
->mutex
);
2363 req
= lookup_get_request(mdsc
, tid
);
2365 dout("handle_reply on unknown tid %llu\n", tid
);
2366 mutex_unlock(&mdsc
->mutex
);
2369 dout("handle_reply %p\n", req
);
2371 /* correct session? */
2372 if (req
->r_session
!= session
) {
2373 pr_err("mdsc_handle_reply got %llu on session mds%d"
2374 " not mds%d\n", tid
, session
->s_mds
,
2375 req
->r_session
? req
->r_session
->s_mds
: -1);
2376 mutex_unlock(&mdsc
->mutex
);
2381 if ((req
->r_got_unsafe
&& !head
->safe
) ||
2382 (req
->r_got_safe
&& head
->safe
)) {
2383 pr_warn("got a dup %s reply on %llu from mds%d\n",
2384 head
->safe
? "safe" : "unsafe", tid
, mds
);
2385 mutex_unlock(&mdsc
->mutex
);
2388 if (req
->r_got_safe
) {
2389 pr_warn("got unsafe after safe on %llu from mds%d\n",
2391 mutex_unlock(&mdsc
->mutex
);
2395 result
= le32_to_cpu(head
->result
);
2399 * if we're not talking to the authority, send to them
2400 * if the authority has changed while we weren't looking,
2401 * send to new authority
2402 * Otherwise we just have to return an ESTALE
2404 if (result
== -ESTALE
) {
2405 dout("got ESTALE on request %llu", req
->r_tid
);
2406 req
->r_resend_mds
= -1;
2407 if (req
->r_direct_mode
!= USE_AUTH_MDS
) {
2408 dout("not using auth, setting for that now");
2409 req
->r_direct_mode
= USE_AUTH_MDS
;
2410 __do_request(mdsc
, req
);
2411 mutex_unlock(&mdsc
->mutex
);
2414 int mds
= __choose_mds(mdsc
, req
);
2415 if (mds
>= 0 && mds
!= req
->r_session
->s_mds
) {
2416 dout("but auth changed, so resending");
2417 __do_request(mdsc
, req
);
2418 mutex_unlock(&mdsc
->mutex
);
2422 dout("have to return ESTALE on request %llu", req
->r_tid
);
2427 req
->r_got_safe
= true;
2428 __unregister_request(mdsc
, req
);
2430 if (req
->r_got_unsafe
) {
2432 * We already handled the unsafe response, now do the
2433 * cleanup. No need to examine the response; the MDS
2434 * doesn't include any result info in the safe
2435 * response. And even if it did, there is nothing
2436 * useful we could do with a revised return value.
2438 dout("got safe reply %llu, mds%d\n", tid
, mds
);
2439 list_del_init(&req
->r_unsafe_item
);
2441 /* last unsafe request during umount? */
2442 if (mdsc
->stopping
&& !__get_oldest_req(mdsc
))
2443 complete_all(&mdsc
->safe_umount_waiters
);
2444 mutex_unlock(&mdsc
->mutex
);
2448 req
->r_got_unsafe
= true;
2449 list_add_tail(&req
->r_unsafe_item
, &req
->r_session
->s_unsafe
);
2450 if (req
->r_unsafe_dir
) {
2451 struct ceph_inode_info
*ci
=
2452 ceph_inode(req
->r_unsafe_dir
);
2453 spin_lock(&ci
->i_unsafe_lock
);
2454 list_add_tail(&req
->r_unsafe_dir_item
,
2455 &ci
->i_unsafe_dirops
);
2456 spin_unlock(&ci
->i_unsafe_lock
);
2460 dout("handle_reply tid %lld result %d\n", tid
, result
);
2461 rinfo
= &req
->r_reply_info
;
2462 err
= parse_reply_info(msg
, rinfo
, session
->s_con
.peer_features
);
2463 mutex_unlock(&mdsc
->mutex
);
2465 mutex_lock(&session
->s_mutex
);
2467 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds
, tid
);
2474 if (rinfo
->snapblob_len
) {
2475 down_write(&mdsc
->snap_rwsem
);
2476 ceph_update_snap_trace(mdsc
, rinfo
->snapblob
,
2477 rinfo
->snapblob
+ rinfo
->snapblob_len
,
2478 le32_to_cpu(head
->op
) == CEPH_MDS_OP_RMSNAP
,
2480 downgrade_write(&mdsc
->snap_rwsem
);
2482 down_read(&mdsc
->snap_rwsem
);
2485 /* insert trace into our cache */
2486 mutex_lock(&req
->r_fill_mutex
);
2487 current
->journal_info
= req
;
2488 err
= ceph_fill_trace(mdsc
->fsc
->sb
, req
, req
->r_session
);
2490 if (result
== 0 && (req
->r_op
== CEPH_MDS_OP_READDIR
||
2491 req
->r_op
== CEPH_MDS_OP_LSSNAP
))
2492 ceph_readdir_prepopulate(req
, req
->r_session
);
2493 ceph_unreserve_caps(mdsc
, &req
->r_caps_reservation
);
2495 current
->journal_info
= NULL
;
2496 mutex_unlock(&req
->r_fill_mutex
);
2498 up_read(&mdsc
->snap_rwsem
);
2500 ceph_put_snap_realm(mdsc
, realm
);
2502 if (err
== 0 && req
->r_got_unsafe
&& req
->r_target_inode
) {
2503 struct ceph_inode_info
*ci
= ceph_inode(req
->r_target_inode
);
2504 spin_lock(&ci
->i_unsafe_lock
);
2505 list_add_tail(&req
->r_unsafe_target_item
, &ci
->i_unsafe_iops
);
2506 spin_unlock(&ci
->i_unsafe_lock
);
2509 mutex_lock(&mdsc
->mutex
);
2510 if (!req
->r_aborted
) {
2514 req
->r_reply
= ceph_msg_get(msg
);
2515 req
->r_got_result
= true;
2518 dout("reply arrived after request %lld was aborted\n", tid
);
2520 mutex_unlock(&mdsc
->mutex
);
2522 mutex_unlock(&session
->s_mutex
);
2524 /* kick calling process */
2525 complete_request(mdsc
, req
);
2527 ceph_mdsc_put_request(req
);
2534 * handle mds notification that our request has been forwarded.
2536 static void handle_forward(struct ceph_mds_client
*mdsc
,
2537 struct ceph_mds_session
*session
,
2538 struct ceph_msg
*msg
)
2540 struct ceph_mds_request
*req
;
2541 u64 tid
= le64_to_cpu(msg
->hdr
.tid
);
2545 void *p
= msg
->front
.iov_base
;
2546 void *end
= p
+ msg
->front
.iov_len
;
2548 ceph_decode_need(&p
, end
, 2*sizeof(u32
), bad
);
2549 next_mds
= ceph_decode_32(&p
);
2550 fwd_seq
= ceph_decode_32(&p
);
2552 mutex_lock(&mdsc
->mutex
);
2553 req
= lookup_get_request(mdsc
, tid
);
2555 dout("forward tid %llu to mds%d - req dne\n", tid
, next_mds
);
2556 goto out
; /* dup reply? */
2559 if (req
->r_aborted
) {
2560 dout("forward tid %llu aborted, unregistering\n", tid
);
2561 __unregister_request(mdsc
, req
);
2562 } else if (fwd_seq
<= req
->r_num_fwd
) {
2563 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2564 tid
, next_mds
, req
->r_num_fwd
, fwd_seq
);
2566 /* resend. forward race not possible; mds would drop */
2567 dout("forward tid %llu to mds%d (we resend)\n", tid
, next_mds
);
2569 BUG_ON(req
->r_got_result
);
2570 req
->r_attempts
= 0;
2571 req
->r_num_fwd
= fwd_seq
;
2572 req
->r_resend_mds
= next_mds
;
2573 put_request_session(req
);
2574 __do_request(mdsc
, req
);
2576 ceph_mdsc_put_request(req
);
2578 mutex_unlock(&mdsc
->mutex
);
2582 pr_err("mdsc_handle_forward decode error err=%d\n", err
);
2586 * handle a mds session control message
2588 static void handle_session(struct ceph_mds_session
*session
,
2589 struct ceph_msg
*msg
)
2591 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2594 int mds
= session
->s_mds
;
2595 struct ceph_mds_session_head
*h
= msg
->front
.iov_base
;
2599 if (msg
->front
.iov_len
!= sizeof(*h
))
2601 op
= le32_to_cpu(h
->op
);
2602 seq
= le64_to_cpu(h
->seq
);
2604 mutex_lock(&mdsc
->mutex
);
2605 if (op
== CEPH_SESSION_CLOSE
)
2606 __unregister_session(mdsc
, session
);
2607 /* FIXME: this ttl calculation is generous */
2608 session
->s_ttl
= jiffies
+ HZ
*mdsc
->mdsmap
->m_session_autoclose
;
2609 mutex_unlock(&mdsc
->mutex
);
2611 mutex_lock(&session
->s_mutex
);
2613 dout("handle_session mds%d %s %p state %s seq %llu\n",
2614 mds
, ceph_session_op_name(op
), session
,
2615 ceph_session_state_name(session
->s_state
), seq
);
2617 if (session
->s_state
== CEPH_MDS_SESSION_HUNG
) {
2618 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2619 pr_info("mds%d came back\n", session
->s_mds
);
2623 case CEPH_SESSION_OPEN
:
2624 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2625 pr_info("mds%d reconnect success\n", session
->s_mds
);
2626 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2627 renewed_caps(mdsc
, session
, 0);
2630 __close_session(mdsc
, session
);
2633 case CEPH_SESSION_RENEWCAPS
:
2634 if (session
->s_renew_seq
== seq
)
2635 renewed_caps(mdsc
, session
, 1);
2638 case CEPH_SESSION_CLOSE
:
2639 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2640 pr_info("mds%d reconnect denied\n", session
->s_mds
);
2641 cleanup_session_requests(mdsc
, session
);
2642 remove_session_caps(session
);
2643 wake
= 2; /* for good measure */
2644 wake_up_all(&mdsc
->session_close_wq
);
2647 case CEPH_SESSION_STALE
:
2648 pr_info("mds%d caps went stale, renewing\n",
2650 spin_lock(&session
->s_gen_ttl_lock
);
2651 session
->s_cap_gen
++;
2652 session
->s_cap_ttl
= jiffies
- 1;
2653 spin_unlock(&session
->s_gen_ttl_lock
);
2654 send_renew_caps(mdsc
, session
);
2657 case CEPH_SESSION_RECALL_STATE
:
2658 trim_caps(mdsc
, session
, le32_to_cpu(h
->max_caps
));
2661 case CEPH_SESSION_FLUSHMSG
:
2662 send_flushmsg_ack(mdsc
, session
, seq
);
2665 case CEPH_SESSION_FORCE_RO
:
2666 dout("force_session_readonly %p\n", session
);
2667 spin_lock(&session
->s_cap_lock
);
2668 session
->s_readonly
= true;
2669 spin_unlock(&session
->s_cap_lock
);
2670 wake_up_session_caps(session
, 0);
2673 case CEPH_SESSION_REJECT
:
2674 WARN_ON(session
->s_state
!= CEPH_MDS_SESSION_OPENING
);
2675 pr_info("mds%d rejected session\n", session
->s_mds
);
2676 session
->s_state
= CEPH_MDS_SESSION_REJECTED
;
2677 cleanup_session_requests(mdsc
, session
);
2678 remove_session_caps(session
);
2679 wake
= 2; /* for good measure */
2683 pr_err("mdsc_handle_session bad op %d mds%d\n", op
, mds
);
2687 mutex_unlock(&session
->s_mutex
);
2689 mutex_lock(&mdsc
->mutex
);
2690 __wake_requests(mdsc
, &session
->s_waiting
);
2692 kick_requests(mdsc
, mds
);
2693 mutex_unlock(&mdsc
->mutex
);
2698 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds
,
2699 (int)msg
->front
.iov_len
);
2706 * called under session->mutex.
2708 static void replay_unsafe_requests(struct ceph_mds_client
*mdsc
,
2709 struct ceph_mds_session
*session
)
2711 struct ceph_mds_request
*req
, *nreq
;
2715 dout("replay_unsafe_requests mds%d\n", session
->s_mds
);
2717 mutex_lock(&mdsc
->mutex
);
2718 list_for_each_entry_safe(req
, nreq
, &session
->s_unsafe
, r_unsafe_item
) {
2719 err
= __prepare_send_request(mdsc
, req
, session
->s_mds
, true);
2721 ceph_msg_get(req
->r_request
);
2722 ceph_con_send(&session
->s_con
, req
->r_request
);
2727 * also re-send old requests when MDS enters reconnect stage. So that MDS
2728 * can process completed request in clientreplay stage.
2730 p
= rb_first(&mdsc
->request_tree
);
2732 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2734 if (req
->r_got_unsafe
)
2736 if (req
->r_attempts
== 0)
2737 continue; /* only old requests */
2738 if (req
->r_session
&&
2739 req
->r_session
->s_mds
== session
->s_mds
) {
2740 err
= __prepare_send_request(mdsc
, req
,
2741 session
->s_mds
, true);
2743 ceph_msg_get(req
->r_request
);
2744 ceph_con_send(&session
->s_con
, req
->r_request
);
2748 mutex_unlock(&mdsc
->mutex
);
2752 * Encode information about a cap for a reconnect with the MDS.
2754 static int encode_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
2758 struct ceph_mds_cap_reconnect v2
;
2759 struct ceph_mds_cap_reconnect_v1 v1
;
2761 struct ceph_inode_info
*ci
;
2762 struct ceph_reconnect_state
*recon_state
= arg
;
2763 struct ceph_pagelist
*pagelist
= recon_state
->pagelist
;
2768 struct dentry
*dentry
;
2772 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2773 inode
, ceph_vinop(inode
), cap
, cap
->cap_id
,
2774 ceph_cap_string(cap
->issued
));
2775 err
= ceph_pagelist_encode_64(pagelist
, ceph_ino(inode
));
2779 dentry
= d_find_alias(inode
);
2781 path
= ceph_mdsc_build_path(dentry
, &pathlen
, &pathbase
, 0);
2783 err
= PTR_ERR(path
);
2792 spin_lock(&ci
->i_ceph_lock
);
2793 cap
->seq
= 0; /* reset cap seq */
2794 cap
->issue_seq
= 0; /* and issue_seq */
2795 cap
->mseq
= 0; /* and migrate_seq */
2796 cap
->cap_gen
= cap
->session
->s_cap_gen
;
2798 if (recon_state
->msg_version
>= 2) {
2799 rec
.v2
.cap_id
= cpu_to_le64(cap
->cap_id
);
2800 rec
.v2
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2801 rec
.v2
.issued
= cpu_to_le32(cap
->issued
);
2802 rec
.v2
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2803 rec
.v2
.pathbase
= cpu_to_le64(pathbase
);
2804 rec
.v2
.flock_len
= 0;
2806 rec
.v1
.cap_id
= cpu_to_le64(cap
->cap_id
);
2807 rec
.v1
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2808 rec
.v1
.issued
= cpu_to_le32(cap
->issued
);
2809 rec
.v1
.size
= cpu_to_le64(inode
->i_size
);
2810 ceph_encode_timespec(&rec
.v1
.mtime
, &inode
->i_mtime
);
2811 ceph_encode_timespec(&rec
.v1
.atime
, &inode
->i_atime
);
2812 rec
.v1
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2813 rec
.v1
.pathbase
= cpu_to_le64(pathbase
);
2816 if (list_empty(&ci
->i_cap_snaps
)) {
2819 struct ceph_cap_snap
*capsnap
=
2820 list_first_entry(&ci
->i_cap_snaps
,
2821 struct ceph_cap_snap
, ci_item
);
2822 snap_follows
= capsnap
->follows
;
2824 spin_unlock(&ci
->i_ceph_lock
);
2826 if (recon_state
->msg_version
>= 2) {
2827 int num_fcntl_locks
, num_flock_locks
;
2828 struct ceph_filelock
*flocks
;
2829 size_t struct_len
, total_len
= 0;
2833 ceph_count_locks(inode
, &num_fcntl_locks
, &num_flock_locks
);
2834 flocks
= kmalloc((num_fcntl_locks
+num_flock_locks
) *
2835 sizeof(struct ceph_filelock
), GFP_NOFS
);
2840 err
= ceph_encode_locks_to_buffer(inode
, flocks
,
2850 if (recon_state
->msg_version
>= 3) {
2851 /* version, compat_version and struct_len */
2852 total_len
= 2 * sizeof(u8
) + sizeof(u32
);
2856 * number of encoded locks is stable, so copy to pagelist
2858 struct_len
= 2 * sizeof(u32
) +
2859 (num_fcntl_locks
+ num_flock_locks
) *
2860 sizeof(struct ceph_filelock
);
2861 rec
.v2
.flock_len
= cpu_to_le32(struct_len
);
2863 struct_len
+= sizeof(rec
.v2
);
2864 struct_len
+= sizeof(u32
) + pathlen
;
2867 struct_len
+= sizeof(u64
); /* snap_follows */
2869 total_len
+= struct_len
;
2870 err
= ceph_pagelist_reserve(pagelist
, total_len
);
2873 if (recon_state
->msg_version
>= 3) {
2874 ceph_pagelist_encode_8(pagelist
, struct_v
);
2875 ceph_pagelist_encode_8(pagelist
, 1);
2876 ceph_pagelist_encode_32(pagelist
, struct_len
);
2878 ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
2879 ceph_pagelist_append(pagelist
, &rec
, sizeof(rec
.v2
));
2880 ceph_locks_to_pagelist(flocks
, pagelist
,
2884 ceph_pagelist_encode_64(pagelist
, snap_follows
);
2888 size_t size
= sizeof(u32
) + pathlen
+ sizeof(rec
.v1
);
2889 err
= ceph_pagelist_reserve(pagelist
, size
);
2891 ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
2892 ceph_pagelist_append(pagelist
, &rec
, sizeof(rec
.v1
));
2896 recon_state
->nr_caps
++;
2906 * If an MDS fails and recovers, clients need to reconnect in order to
2907 * reestablish shared state. This includes all caps issued through
2908 * this session _and_ the snap_realm hierarchy. Because it's not
2909 * clear which snap realms the mds cares about, we send everything we
2910 * know about.. that ensures we'll then get any new info the
2911 * recovering MDS might have.
2913 * This is a relatively heavyweight operation, but it's rare.
2915 * called with mdsc->mutex held.
2917 static void send_mds_reconnect(struct ceph_mds_client
*mdsc
,
2918 struct ceph_mds_session
*session
)
2920 struct ceph_msg
*reply
;
2922 int mds
= session
->s_mds
;
2925 struct ceph_pagelist
*pagelist
;
2926 struct ceph_reconnect_state recon_state
;
2928 pr_info("mds%d reconnect start\n", mds
);
2930 pagelist
= kmalloc(sizeof(*pagelist
), GFP_NOFS
);
2932 goto fail_nopagelist
;
2933 ceph_pagelist_init(pagelist
);
2935 reply
= ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT
, 0, GFP_NOFS
, false);
2939 mutex_lock(&session
->s_mutex
);
2940 session
->s_state
= CEPH_MDS_SESSION_RECONNECTING
;
2943 dout("session %p state %s\n", session
,
2944 ceph_session_state_name(session
->s_state
));
2946 spin_lock(&session
->s_gen_ttl_lock
);
2947 session
->s_cap_gen
++;
2948 spin_unlock(&session
->s_gen_ttl_lock
);
2950 spin_lock(&session
->s_cap_lock
);
2951 /* don't know if session is readonly */
2952 session
->s_readonly
= 0;
2954 * notify __ceph_remove_cap() that we are composing cap reconnect.
2955 * If a cap get released before being added to the cap reconnect,
2956 * __ceph_remove_cap() should skip queuing cap release.
2958 session
->s_cap_reconnect
= 1;
2959 /* drop old cap expires; we're about to reestablish that state */
2960 cleanup_cap_releases(mdsc
, session
);
2962 /* trim unused caps to reduce MDS's cache rejoin time */
2963 if (mdsc
->fsc
->sb
->s_root
)
2964 shrink_dcache_parent(mdsc
->fsc
->sb
->s_root
);
2966 ceph_con_close(&session
->s_con
);
2967 ceph_con_open(&session
->s_con
,
2968 CEPH_ENTITY_TYPE_MDS
, mds
,
2969 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
2971 /* replay unsafe requests */
2972 replay_unsafe_requests(mdsc
, session
);
2974 down_read(&mdsc
->snap_rwsem
);
2976 /* traverse this session's caps */
2977 s_nr_caps
= session
->s_nr_caps
;
2978 err
= ceph_pagelist_encode_32(pagelist
, s_nr_caps
);
2982 recon_state
.nr_caps
= 0;
2983 recon_state
.pagelist
= pagelist
;
2984 if (session
->s_con
.peer_features
& CEPH_FEATURE_MDSENC
)
2985 recon_state
.msg_version
= 3;
2986 else if (session
->s_con
.peer_features
& CEPH_FEATURE_FLOCK
)
2987 recon_state
.msg_version
= 2;
2989 recon_state
.msg_version
= 1;
2990 err
= iterate_session_caps(session
, encode_caps_cb
, &recon_state
);
2994 spin_lock(&session
->s_cap_lock
);
2995 session
->s_cap_reconnect
= 0;
2996 spin_unlock(&session
->s_cap_lock
);
2999 * snaprealms. we provide mds with the ino, seq (version), and
3000 * parent for all of our realms. If the mds has any newer info,
3003 for (p
= rb_first(&mdsc
->snap_realms
); p
; p
= rb_next(p
)) {
3004 struct ceph_snap_realm
*realm
=
3005 rb_entry(p
, struct ceph_snap_realm
, node
);
3006 struct ceph_mds_snaprealm_reconnect sr_rec
;
3008 dout(" adding snap realm %llx seq %lld parent %llx\n",
3009 realm
->ino
, realm
->seq
, realm
->parent_ino
);
3010 sr_rec
.ino
= cpu_to_le64(realm
->ino
);
3011 sr_rec
.seq
= cpu_to_le64(realm
->seq
);
3012 sr_rec
.parent
= cpu_to_le64(realm
->parent_ino
);
3013 err
= ceph_pagelist_append(pagelist
, &sr_rec
, sizeof(sr_rec
));
3018 reply
->hdr
.version
= cpu_to_le16(recon_state
.msg_version
);
3020 /* raced with cap release? */
3021 if (s_nr_caps
!= recon_state
.nr_caps
) {
3022 struct page
*page
= list_first_entry(&pagelist
->head
,
3024 __le32
*addr
= kmap_atomic(page
);
3025 *addr
= cpu_to_le32(recon_state
.nr_caps
);
3026 kunmap_atomic(addr
);
3029 reply
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
3030 ceph_msg_data_add_pagelist(reply
, pagelist
);
3032 ceph_early_kick_flushing_caps(mdsc
, session
);
3034 ceph_con_send(&session
->s_con
, reply
);
3036 mutex_unlock(&session
->s_mutex
);
3038 mutex_lock(&mdsc
->mutex
);
3039 __wake_requests(mdsc
, &session
->s_waiting
);
3040 mutex_unlock(&mdsc
->mutex
);
3042 up_read(&mdsc
->snap_rwsem
);
3046 ceph_msg_put(reply
);
3047 up_read(&mdsc
->snap_rwsem
);
3048 mutex_unlock(&session
->s_mutex
);
3050 ceph_pagelist_release(pagelist
);
3052 pr_err("error %d preparing reconnect for mds%d\n", err
, mds
);
3058 * compare old and new mdsmaps, kicking requests
3059 * and closing out old connections as necessary
3061 * called under mdsc->mutex.
3063 static void check_new_map(struct ceph_mds_client
*mdsc
,
3064 struct ceph_mdsmap
*newmap
,
3065 struct ceph_mdsmap
*oldmap
)
3068 int oldstate
, newstate
;
3069 struct ceph_mds_session
*s
;
3071 dout("check_new_map new %u old %u\n",
3072 newmap
->m_epoch
, oldmap
->m_epoch
);
3074 for (i
= 0; i
< oldmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
3075 if (mdsc
->sessions
[i
] == NULL
)
3077 s
= mdsc
->sessions
[i
];
3078 oldstate
= ceph_mdsmap_get_state(oldmap
, i
);
3079 newstate
= ceph_mdsmap_get_state(newmap
, i
);
3081 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3082 i
, ceph_mds_state_name(oldstate
),
3083 ceph_mdsmap_is_laggy(oldmap
, i
) ? " (laggy)" : "",
3084 ceph_mds_state_name(newstate
),
3085 ceph_mdsmap_is_laggy(newmap
, i
) ? " (laggy)" : "",
3086 ceph_session_state_name(s
->s_state
));
3088 if (i
>= newmap
->m_max_mds
||
3089 memcmp(ceph_mdsmap_get_addr(oldmap
, i
),
3090 ceph_mdsmap_get_addr(newmap
, i
),
3091 sizeof(struct ceph_entity_addr
))) {
3092 if (s
->s_state
== CEPH_MDS_SESSION_OPENING
) {
3093 /* the session never opened, just close it
3095 __wake_requests(mdsc
, &s
->s_waiting
);
3096 __unregister_session(mdsc
, s
);
3099 mutex_unlock(&mdsc
->mutex
);
3100 mutex_lock(&s
->s_mutex
);
3101 mutex_lock(&mdsc
->mutex
);
3102 ceph_con_close(&s
->s_con
);
3103 mutex_unlock(&s
->s_mutex
);
3104 s
->s_state
= CEPH_MDS_SESSION_RESTARTING
;
3106 } else if (oldstate
== newstate
) {
3107 continue; /* nothing new with this mds */
3113 if (s
->s_state
== CEPH_MDS_SESSION_RESTARTING
&&
3114 newstate
>= CEPH_MDS_STATE_RECONNECT
) {
3115 mutex_unlock(&mdsc
->mutex
);
3116 send_mds_reconnect(mdsc
, s
);
3117 mutex_lock(&mdsc
->mutex
);
3121 * kick request on any mds that has gone active.
3123 if (oldstate
< CEPH_MDS_STATE_ACTIVE
&&
3124 newstate
>= CEPH_MDS_STATE_ACTIVE
) {
3125 if (oldstate
!= CEPH_MDS_STATE_CREATING
&&
3126 oldstate
!= CEPH_MDS_STATE_STARTING
)
3127 pr_info("mds%d recovery completed\n", s
->s_mds
);
3128 kick_requests(mdsc
, i
);
3129 ceph_kick_flushing_caps(mdsc
, s
);
3130 wake_up_session_caps(s
, 1);
3134 for (i
= 0; i
< newmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
3135 s
= mdsc
->sessions
[i
];
3138 if (!ceph_mdsmap_is_laggy(newmap
, i
))
3140 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3141 s
->s_state
== CEPH_MDS_SESSION_HUNG
||
3142 s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3143 dout(" connecting to export targets of laggy mds%d\n",
3145 __open_export_target_sessions(mdsc
, s
);
3157 * caller must hold session s_mutex, dentry->d_lock
3159 void __ceph_mdsc_drop_dentry_lease(struct dentry
*dentry
)
3161 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
3163 ceph_put_mds_session(di
->lease_session
);
3164 di
->lease_session
= NULL
;
3167 static void handle_lease(struct ceph_mds_client
*mdsc
,
3168 struct ceph_mds_session
*session
,
3169 struct ceph_msg
*msg
)
3171 struct super_block
*sb
= mdsc
->fsc
->sb
;
3172 struct inode
*inode
;
3173 struct dentry
*parent
, *dentry
;
3174 struct ceph_dentry_info
*di
;
3175 int mds
= session
->s_mds
;
3176 struct ceph_mds_lease
*h
= msg
->front
.iov_base
;
3178 struct ceph_vino vino
;
3182 dout("handle_lease from mds%d\n", mds
);
3185 if (msg
->front
.iov_len
< sizeof(*h
) + sizeof(u32
))
3187 vino
.ino
= le64_to_cpu(h
->ino
);
3188 vino
.snap
= CEPH_NOSNAP
;
3189 seq
= le32_to_cpu(h
->seq
);
3190 dname
.name
= (void *)h
+ sizeof(*h
) + sizeof(u32
);
3191 dname
.len
= msg
->front
.iov_len
- sizeof(*h
) - sizeof(u32
);
3192 if (dname
.len
!= get_unaligned_le32(h
+1))
3196 inode
= ceph_find_inode(sb
, vino
);
3197 dout("handle_lease %s, ino %llx %p %.*s\n",
3198 ceph_lease_op_name(h
->action
), vino
.ino
, inode
,
3199 dname
.len
, dname
.name
);
3201 mutex_lock(&session
->s_mutex
);
3204 if (inode
== NULL
) {
3205 dout("handle_lease no inode %llx\n", vino
.ino
);
3210 parent
= d_find_alias(inode
);
3212 dout("no parent dentry on inode %p\n", inode
);
3214 goto release
; /* hrm... */
3216 dname
.hash
= full_name_hash(parent
, dname
.name
, dname
.len
);
3217 dentry
= d_lookup(parent
, &dname
);
3222 spin_lock(&dentry
->d_lock
);
3223 di
= ceph_dentry(dentry
);
3224 switch (h
->action
) {
3225 case CEPH_MDS_LEASE_REVOKE
:
3226 if (di
->lease_session
== session
) {
3227 if (ceph_seq_cmp(di
->lease_seq
, seq
) > 0)
3228 h
->seq
= cpu_to_le32(di
->lease_seq
);
3229 __ceph_mdsc_drop_dentry_lease(dentry
);
3234 case CEPH_MDS_LEASE_RENEW
:
3235 if (di
->lease_session
== session
&&
3236 di
->lease_gen
== session
->s_cap_gen
&&
3237 di
->lease_renew_from
&&
3238 di
->lease_renew_after
== 0) {
3239 unsigned long duration
=
3240 msecs_to_jiffies(le32_to_cpu(h
->duration_ms
));
3242 di
->lease_seq
= seq
;
3243 di
->time
= di
->lease_renew_from
+ duration
;
3244 di
->lease_renew_after
= di
->lease_renew_from
+
3246 di
->lease_renew_from
= 0;
3250 spin_unlock(&dentry
->d_lock
);
3257 /* let's just reuse the same message */
3258 h
->action
= CEPH_MDS_LEASE_REVOKE_ACK
;
3260 ceph_con_send(&session
->s_con
, msg
);
3264 mutex_unlock(&session
->s_mutex
);
3268 pr_err("corrupt lease message\n");
3272 void ceph_mdsc_lease_send_msg(struct ceph_mds_session
*session
,
3273 struct inode
*inode
,
3274 struct dentry
*dentry
, char action
,
3277 struct ceph_msg
*msg
;
3278 struct ceph_mds_lease
*lease
;
3279 int len
= sizeof(*lease
) + sizeof(u32
);
3282 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3283 inode
, dentry
, ceph_lease_op_name(action
), session
->s_mds
);
3284 dnamelen
= dentry
->d_name
.len
;
3287 msg
= ceph_msg_new(CEPH_MSG_CLIENT_LEASE
, len
, GFP_NOFS
, false);
3290 lease
= msg
->front
.iov_base
;
3291 lease
->action
= action
;
3292 lease
->ino
= cpu_to_le64(ceph_vino(inode
).ino
);
3293 lease
->first
= lease
->last
= cpu_to_le64(ceph_vino(inode
).snap
);
3294 lease
->seq
= cpu_to_le32(seq
);
3295 put_unaligned_le32(dnamelen
, lease
+ 1);
3296 memcpy((void *)(lease
+ 1) + 4, dentry
->d_name
.name
, dnamelen
);
3299 * if this is a preemptive lease RELEASE, no need to
3300 * flush request stream, since the actual request will
3303 msg
->more_to_follow
= (action
== CEPH_MDS_LEASE_RELEASE
);
3305 ceph_con_send(&session
->s_con
, msg
);
3309 * drop all leases (and dentry refs) in preparation for umount
3311 static void drop_leases(struct ceph_mds_client
*mdsc
)
3315 dout("drop_leases\n");
3316 mutex_lock(&mdsc
->mutex
);
3317 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3318 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3321 mutex_unlock(&mdsc
->mutex
);
3322 mutex_lock(&s
->s_mutex
);
3323 mutex_unlock(&s
->s_mutex
);
3324 ceph_put_mds_session(s
);
3325 mutex_lock(&mdsc
->mutex
);
3327 mutex_unlock(&mdsc
->mutex
);
3333 * delayed work -- periodically trim expired leases, renew caps with mds
3335 static void schedule_delayed(struct ceph_mds_client
*mdsc
)
3338 unsigned hz
= round_jiffies_relative(HZ
* delay
);
3339 schedule_delayed_work(&mdsc
->delayed_work
, hz
);
3342 static void delayed_work(struct work_struct
*work
)
3345 struct ceph_mds_client
*mdsc
=
3346 container_of(work
, struct ceph_mds_client
, delayed_work
.work
);
3350 dout("mdsc delayed_work\n");
3351 ceph_check_delayed_caps(mdsc
);
3353 mutex_lock(&mdsc
->mutex
);
3354 renew_interval
= mdsc
->mdsmap
->m_session_timeout
>> 2;
3355 renew_caps
= time_after_eq(jiffies
, HZ
*renew_interval
+
3356 mdsc
->last_renew_caps
);
3358 mdsc
->last_renew_caps
= jiffies
;
3360 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3361 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3364 if (s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3365 dout("resending session close request for mds%d\n",
3367 request_close_session(mdsc
, s
);
3368 ceph_put_mds_session(s
);
3371 if (s
->s_ttl
&& time_after(jiffies
, s
->s_ttl
)) {
3372 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
) {
3373 s
->s_state
= CEPH_MDS_SESSION_HUNG
;
3374 pr_info("mds%d hung\n", s
->s_mds
);
3377 if (s
->s_state
< CEPH_MDS_SESSION_OPEN
) {
3378 /* this mds is failed or recovering, just wait */
3379 ceph_put_mds_session(s
);
3382 mutex_unlock(&mdsc
->mutex
);
3384 mutex_lock(&s
->s_mutex
);
3386 send_renew_caps(mdsc
, s
);
3388 ceph_con_keepalive(&s
->s_con
);
3389 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3390 s
->s_state
== CEPH_MDS_SESSION_HUNG
)
3391 ceph_send_cap_releases(mdsc
, s
);
3392 mutex_unlock(&s
->s_mutex
);
3393 ceph_put_mds_session(s
);
3395 mutex_lock(&mdsc
->mutex
);
3397 mutex_unlock(&mdsc
->mutex
);
3399 schedule_delayed(mdsc
);
3402 int ceph_mdsc_init(struct ceph_fs_client
*fsc
)
3405 struct ceph_mds_client
*mdsc
;
3407 mdsc
= kzalloc(sizeof(struct ceph_mds_client
), GFP_NOFS
);
3412 mutex_init(&mdsc
->mutex
);
3413 mdsc
->mdsmap
= kzalloc(sizeof(*mdsc
->mdsmap
), GFP_NOFS
);
3414 if (mdsc
->mdsmap
== NULL
) {
3419 init_completion(&mdsc
->safe_umount_waiters
);
3420 init_waitqueue_head(&mdsc
->session_close_wq
);
3421 INIT_LIST_HEAD(&mdsc
->waiting_for_map
);
3422 mdsc
->sessions
= NULL
;
3423 atomic_set(&mdsc
->num_sessions
, 0);
3424 mdsc
->max_sessions
= 0;
3426 mdsc
->last_snap_seq
= 0;
3427 init_rwsem(&mdsc
->snap_rwsem
);
3428 mdsc
->snap_realms
= RB_ROOT
;
3429 INIT_LIST_HEAD(&mdsc
->snap_empty
);
3430 spin_lock_init(&mdsc
->snap_empty_lock
);
3432 mdsc
->oldest_tid
= 0;
3433 mdsc
->request_tree
= RB_ROOT
;
3434 INIT_DELAYED_WORK(&mdsc
->delayed_work
, delayed_work
);
3435 mdsc
->last_renew_caps
= jiffies
;
3436 INIT_LIST_HEAD(&mdsc
->cap_delay_list
);
3437 spin_lock_init(&mdsc
->cap_delay_lock
);
3438 INIT_LIST_HEAD(&mdsc
->snap_flush_list
);
3439 spin_lock_init(&mdsc
->snap_flush_lock
);
3440 mdsc
->last_cap_flush_tid
= 1;
3441 INIT_LIST_HEAD(&mdsc
->cap_flush_list
);
3442 INIT_LIST_HEAD(&mdsc
->cap_dirty
);
3443 INIT_LIST_HEAD(&mdsc
->cap_dirty_migrating
);
3444 mdsc
->num_cap_flushing
= 0;
3445 spin_lock_init(&mdsc
->cap_dirty_lock
);
3446 init_waitqueue_head(&mdsc
->cap_flushing_wq
);
3447 spin_lock_init(&mdsc
->dentry_lru_lock
);
3448 INIT_LIST_HEAD(&mdsc
->dentry_lru
);
3450 ceph_caps_init(mdsc
);
3451 ceph_adjust_min_caps(mdsc
, fsc
->min_caps
);
3453 init_rwsem(&mdsc
->pool_perm_rwsem
);
3454 mdsc
->pool_perm_tree
= RB_ROOT
;
3460 * Wait for safe replies on open mds requests. If we time out, drop
3461 * all requests from the tree to avoid dangling dentry refs.
3463 static void wait_requests(struct ceph_mds_client
*mdsc
)
3465 struct ceph_options
*opts
= mdsc
->fsc
->client
->options
;
3466 struct ceph_mds_request
*req
;
3468 mutex_lock(&mdsc
->mutex
);
3469 if (__get_oldest_req(mdsc
)) {
3470 mutex_unlock(&mdsc
->mutex
);
3472 dout("wait_requests waiting for requests\n");
3473 wait_for_completion_timeout(&mdsc
->safe_umount_waiters
,
3474 ceph_timeout_jiffies(opts
->mount_timeout
));
3476 /* tear down remaining requests */
3477 mutex_lock(&mdsc
->mutex
);
3478 while ((req
= __get_oldest_req(mdsc
))) {
3479 dout("wait_requests timed out on tid %llu\n",
3481 __unregister_request(mdsc
, req
);
3484 mutex_unlock(&mdsc
->mutex
);
3485 dout("wait_requests done\n");
3489 * called before mount is ro, and before dentries are torn down.
3490 * (hmm, does this still race with new lookups?)
3492 void ceph_mdsc_pre_umount(struct ceph_mds_client
*mdsc
)
3494 dout("pre_umount\n");
3498 ceph_flush_dirty_caps(mdsc
);
3499 wait_requests(mdsc
);
3502 * wait for reply handlers to drop their request refs and
3503 * their inode/dcache refs
3509 * wait for all write mds requests to flush.
3511 static void wait_unsafe_requests(struct ceph_mds_client
*mdsc
, u64 want_tid
)
3513 struct ceph_mds_request
*req
= NULL
, *nextreq
;
3516 mutex_lock(&mdsc
->mutex
);
3517 dout("wait_unsafe_requests want %lld\n", want_tid
);
3519 req
= __get_oldest_req(mdsc
);
3520 while (req
&& req
->r_tid
<= want_tid
) {
3521 /* find next request */
3522 n
= rb_next(&req
->r_node
);
3524 nextreq
= rb_entry(n
, struct ceph_mds_request
, r_node
);
3527 if (req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
&&
3528 (req
->r_op
& CEPH_MDS_OP_WRITE
)) {
3530 ceph_mdsc_get_request(req
);
3532 ceph_mdsc_get_request(nextreq
);
3533 mutex_unlock(&mdsc
->mutex
);
3534 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3535 req
->r_tid
, want_tid
);
3536 wait_for_completion(&req
->r_safe_completion
);
3537 mutex_lock(&mdsc
->mutex
);
3538 ceph_mdsc_put_request(req
);
3540 break; /* next dne before, so we're done! */
3541 if (RB_EMPTY_NODE(&nextreq
->r_node
)) {
3542 /* next request was removed from tree */
3543 ceph_mdsc_put_request(nextreq
);
3546 ceph_mdsc_put_request(nextreq
); /* won't go away */
3550 mutex_unlock(&mdsc
->mutex
);
3551 dout("wait_unsafe_requests done\n");
3554 void ceph_mdsc_sync(struct ceph_mds_client
*mdsc
)
3556 u64 want_tid
, want_flush
;
3558 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
3562 mutex_lock(&mdsc
->mutex
);
3563 want_tid
= mdsc
->last_tid
;
3564 mutex_unlock(&mdsc
->mutex
);
3566 ceph_flush_dirty_caps(mdsc
);
3567 spin_lock(&mdsc
->cap_dirty_lock
);
3568 want_flush
= mdsc
->last_cap_flush_tid
;
3569 if (!list_empty(&mdsc
->cap_flush_list
)) {
3570 struct ceph_cap_flush
*cf
=
3571 list_last_entry(&mdsc
->cap_flush_list
,
3572 struct ceph_cap_flush
, g_list
);
3575 spin_unlock(&mdsc
->cap_dirty_lock
);
3577 dout("sync want tid %lld flush_seq %lld\n",
3578 want_tid
, want_flush
);
3580 wait_unsafe_requests(mdsc
, want_tid
);
3581 wait_caps_flush(mdsc
, want_flush
);
3585 * true if all sessions are closed, or we force unmount
3587 static bool done_closing_sessions(struct ceph_mds_client
*mdsc
, int skipped
)
3589 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
3591 return atomic_read(&mdsc
->num_sessions
) <= skipped
;
3595 * called after sb is ro.
3597 void ceph_mdsc_close_sessions(struct ceph_mds_client
*mdsc
)
3599 struct ceph_options
*opts
= mdsc
->fsc
->client
->options
;
3600 struct ceph_mds_session
*session
;
3604 dout("close_sessions\n");
3606 /* close sessions */
3607 mutex_lock(&mdsc
->mutex
);
3608 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3609 session
= __ceph_lookup_mds_session(mdsc
, i
);
3612 mutex_unlock(&mdsc
->mutex
);
3613 mutex_lock(&session
->s_mutex
);
3614 if (__close_session(mdsc
, session
) <= 0)
3616 mutex_unlock(&session
->s_mutex
);
3617 ceph_put_mds_session(session
);
3618 mutex_lock(&mdsc
->mutex
);
3620 mutex_unlock(&mdsc
->mutex
);
3622 dout("waiting for sessions to close\n");
3623 wait_event_timeout(mdsc
->session_close_wq
,
3624 done_closing_sessions(mdsc
, skipped
),
3625 ceph_timeout_jiffies(opts
->mount_timeout
));
3627 /* tear down remaining sessions */
3628 mutex_lock(&mdsc
->mutex
);
3629 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3630 if (mdsc
->sessions
[i
]) {
3631 session
= get_session(mdsc
->sessions
[i
]);
3632 __unregister_session(mdsc
, session
);
3633 mutex_unlock(&mdsc
->mutex
);
3634 mutex_lock(&session
->s_mutex
);
3635 remove_session_caps(session
);
3636 mutex_unlock(&session
->s_mutex
);
3637 ceph_put_mds_session(session
);
3638 mutex_lock(&mdsc
->mutex
);
3641 WARN_ON(!list_empty(&mdsc
->cap_delay_list
));
3642 mutex_unlock(&mdsc
->mutex
);
3644 ceph_cleanup_empty_realms(mdsc
);
3646 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3651 void ceph_mdsc_force_umount(struct ceph_mds_client
*mdsc
)
3653 struct ceph_mds_session
*session
;
3656 dout("force umount\n");
3658 mutex_lock(&mdsc
->mutex
);
3659 for (mds
= 0; mds
< mdsc
->max_sessions
; mds
++) {
3660 session
= __ceph_lookup_mds_session(mdsc
, mds
);
3663 mutex_unlock(&mdsc
->mutex
);
3664 mutex_lock(&session
->s_mutex
);
3665 __close_session(mdsc
, session
);
3666 if (session
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3667 cleanup_session_requests(mdsc
, session
);
3668 remove_session_caps(session
);
3670 mutex_unlock(&session
->s_mutex
);
3671 ceph_put_mds_session(session
);
3672 mutex_lock(&mdsc
->mutex
);
3673 kick_requests(mdsc
, mds
);
3675 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3676 mutex_unlock(&mdsc
->mutex
);
3679 static void ceph_mdsc_stop(struct ceph_mds_client
*mdsc
)
3682 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3684 ceph_mdsmap_destroy(mdsc
->mdsmap
);
3685 kfree(mdsc
->sessions
);
3686 ceph_caps_finalize(mdsc
);
3687 ceph_pool_perm_destroy(mdsc
);
3690 void ceph_mdsc_destroy(struct ceph_fs_client
*fsc
)
3692 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
3694 dout("mdsc_destroy %p\n", mdsc
);
3695 ceph_mdsc_stop(mdsc
);
3697 /* flush out any connection work with references to us */
3702 dout("mdsc_destroy %p done\n", mdsc
);
3705 void ceph_mdsc_handle_fsmap(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3707 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3708 const char *mds_namespace
= fsc
->mount_options
->mds_namespace
;
3709 void *p
= msg
->front
.iov_base
;
3710 void *end
= p
+ msg
->front
.iov_len
;
3714 u32 mount_fscid
= (u32
)-1;
3715 u8 struct_v
, struct_cv
;
3718 ceph_decode_need(&p
, end
, sizeof(u32
), bad
);
3719 epoch
= ceph_decode_32(&p
);
3721 dout("handle_fsmap epoch %u\n", epoch
);
3723 ceph_decode_need(&p
, end
, 2 + sizeof(u32
), bad
);
3724 struct_v
= ceph_decode_8(&p
);
3725 struct_cv
= ceph_decode_8(&p
);
3726 map_len
= ceph_decode_32(&p
);
3728 ceph_decode_need(&p
, end
, sizeof(u32
) * 3, bad
);
3729 p
+= sizeof(u32
) * 2; /* skip epoch and legacy_client_fscid */
3731 num_fs
= ceph_decode_32(&p
);
3732 while (num_fs
-- > 0) {
3733 void *info_p
, *info_end
;
3738 ceph_decode_need(&p
, end
, 2 + sizeof(u32
), bad
);
3739 info_v
= ceph_decode_8(&p
);
3740 info_cv
= ceph_decode_8(&p
);
3741 info_len
= ceph_decode_32(&p
);
3742 ceph_decode_need(&p
, end
, info_len
, bad
);
3744 info_end
= p
+ info_len
;
3747 ceph_decode_need(&info_p
, info_end
, sizeof(u32
) * 2, bad
);
3748 fscid
= ceph_decode_32(&info_p
);
3749 namelen
= ceph_decode_32(&info_p
);
3750 ceph_decode_need(&info_p
, info_end
, namelen
, bad
);
3752 if (mds_namespace
&&
3753 strlen(mds_namespace
) == namelen
&&
3754 !strncmp(mds_namespace
, (char *)info_p
, namelen
)) {
3755 mount_fscid
= fscid
;
3760 ceph_monc_got_map(&fsc
->client
->monc
, CEPH_SUB_FSMAP
, epoch
);
3761 if (mount_fscid
!= (u32
)-1) {
3762 fsc
->client
->monc
.fs_cluster_id
= mount_fscid
;
3763 ceph_monc_want_map(&fsc
->client
->monc
, CEPH_SUB_MDSMAP
,
3765 ceph_monc_renew_subs(&fsc
->client
->monc
);
3772 pr_err("error decoding fsmap\n");
3774 mutex_lock(&mdsc
->mutex
);
3775 mdsc
->mdsmap_err
= -ENOENT
;
3776 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3777 mutex_unlock(&mdsc
->mutex
);
3782 * handle mds map update.
3784 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3788 void *p
= msg
->front
.iov_base
;
3789 void *end
= p
+ msg
->front
.iov_len
;
3790 struct ceph_mdsmap
*newmap
, *oldmap
;
3791 struct ceph_fsid fsid
;
3794 ceph_decode_need(&p
, end
, sizeof(fsid
)+2*sizeof(u32
), bad
);
3795 ceph_decode_copy(&p
, &fsid
, sizeof(fsid
));
3796 if (ceph_check_fsid(mdsc
->fsc
->client
, &fsid
) < 0)
3798 epoch
= ceph_decode_32(&p
);
3799 maplen
= ceph_decode_32(&p
);
3800 dout("handle_map epoch %u len %d\n", epoch
, (int)maplen
);
3802 /* do we need it? */
3803 mutex_lock(&mdsc
->mutex
);
3804 if (mdsc
->mdsmap
&& epoch
<= mdsc
->mdsmap
->m_epoch
) {
3805 dout("handle_map epoch %u <= our %u\n",
3806 epoch
, mdsc
->mdsmap
->m_epoch
);
3807 mutex_unlock(&mdsc
->mutex
);
3811 newmap
= ceph_mdsmap_decode(&p
, end
);
3812 if (IS_ERR(newmap
)) {
3813 err
= PTR_ERR(newmap
);
3817 /* swap into place */
3819 oldmap
= mdsc
->mdsmap
;
3820 mdsc
->mdsmap
= newmap
;
3821 check_new_map(mdsc
, newmap
, oldmap
);
3822 ceph_mdsmap_destroy(oldmap
);
3824 mdsc
->mdsmap
= newmap
; /* first mds map */
3826 mdsc
->fsc
->sb
->s_maxbytes
= mdsc
->mdsmap
->m_max_file_size
;
3828 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3829 ceph_monc_got_map(&mdsc
->fsc
->client
->monc
, CEPH_SUB_MDSMAP
,
3830 mdsc
->mdsmap
->m_epoch
);
3832 mutex_unlock(&mdsc
->mutex
);
3833 schedule_delayed(mdsc
);
3837 mutex_unlock(&mdsc
->mutex
);
3839 pr_err("error decoding mdsmap %d\n", err
);
3843 static struct ceph_connection
*con_get(struct ceph_connection
*con
)
3845 struct ceph_mds_session
*s
= con
->private;
3847 if (get_session(s
)) {
3848 dout("mdsc con_get %p ok (%d)\n", s
, atomic_read(&s
->s_ref
));
3851 dout("mdsc con_get %p FAIL\n", s
);
3855 static void con_put(struct ceph_connection
*con
)
3857 struct ceph_mds_session
*s
= con
->private;
3859 dout("mdsc con_put %p (%d)\n", s
, atomic_read(&s
->s_ref
) - 1);
3860 ceph_put_mds_session(s
);
3864 * if the client is unresponsive for long enough, the mds will kill
3865 * the session entirely.
3867 static void peer_reset(struct ceph_connection
*con
)
3869 struct ceph_mds_session
*s
= con
->private;
3870 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3872 pr_warn("mds%d closed our session\n", s
->s_mds
);
3873 send_mds_reconnect(mdsc
, s
);
3876 static void dispatch(struct ceph_connection
*con
, struct ceph_msg
*msg
)
3878 struct ceph_mds_session
*s
= con
->private;
3879 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3880 int type
= le16_to_cpu(msg
->hdr
.type
);
3882 mutex_lock(&mdsc
->mutex
);
3883 if (__verify_registered_session(mdsc
, s
) < 0) {
3884 mutex_unlock(&mdsc
->mutex
);
3887 mutex_unlock(&mdsc
->mutex
);
3890 case CEPH_MSG_MDS_MAP
:
3891 ceph_mdsc_handle_mdsmap(mdsc
, msg
);
3893 case CEPH_MSG_FS_MAP_USER
:
3894 ceph_mdsc_handle_fsmap(mdsc
, msg
);
3896 case CEPH_MSG_CLIENT_SESSION
:
3897 handle_session(s
, msg
);
3899 case CEPH_MSG_CLIENT_REPLY
:
3900 handle_reply(s
, msg
);
3902 case CEPH_MSG_CLIENT_REQUEST_FORWARD
:
3903 handle_forward(mdsc
, s
, msg
);
3905 case CEPH_MSG_CLIENT_CAPS
:
3906 ceph_handle_caps(s
, msg
);
3908 case CEPH_MSG_CLIENT_SNAP
:
3909 ceph_handle_snap(mdsc
, s
, msg
);
3911 case CEPH_MSG_CLIENT_LEASE
:
3912 handle_lease(mdsc
, s
, msg
);
3916 pr_err("received unknown message type %d %s\n", type
,
3917 ceph_msg_type_name(type
));
3928 * Note: returned pointer is the address of a structure that's
3929 * managed separately. Caller must *not* attempt to free it.
3931 static struct ceph_auth_handshake
*get_authorizer(struct ceph_connection
*con
,
3932 int *proto
, int force_new
)
3934 struct ceph_mds_session
*s
= con
->private;
3935 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3936 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3937 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
3939 if (force_new
&& auth
->authorizer
) {
3940 ceph_auth_destroy_authorizer(auth
->authorizer
);
3941 auth
->authorizer
= NULL
;
3943 if (!auth
->authorizer
) {
3944 int ret
= ceph_auth_create_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3947 return ERR_PTR(ret
);
3949 int ret
= ceph_auth_update_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3952 return ERR_PTR(ret
);
3954 *proto
= ac
->protocol
;
3960 static int verify_authorizer_reply(struct ceph_connection
*con
)
3962 struct ceph_mds_session
*s
= con
->private;
3963 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3964 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3966 return ceph_auth_verify_authorizer_reply(ac
, s
->s_auth
.authorizer
);
3969 static int invalidate_authorizer(struct ceph_connection
*con
)
3971 struct ceph_mds_session
*s
= con
->private;
3972 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3973 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3975 ceph_auth_invalidate_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
);
3977 return ceph_monc_validate_auth(&mdsc
->fsc
->client
->monc
);
3980 static struct ceph_msg
*mds_alloc_msg(struct ceph_connection
*con
,
3981 struct ceph_msg_header
*hdr
, int *skip
)
3983 struct ceph_msg
*msg
;
3984 int type
= (int) le16_to_cpu(hdr
->type
);
3985 int front_len
= (int) le32_to_cpu(hdr
->front_len
);
3991 msg
= ceph_msg_new(type
, front_len
, GFP_NOFS
, false);
3993 pr_err("unable to allocate msg type %d len %d\n",
4001 static int mds_sign_message(struct ceph_msg
*msg
)
4003 struct ceph_mds_session
*s
= msg
->con
->private;
4004 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
4006 return ceph_auth_sign_message(auth
, msg
);
4009 static int mds_check_message_signature(struct ceph_msg
*msg
)
4011 struct ceph_mds_session
*s
= msg
->con
->private;
4012 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
4014 return ceph_auth_check_message_signature(auth
, msg
);
4017 static const struct ceph_connection_operations mds_con_ops
= {
4020 .dispatch
= dispatch
,
4021 .get_authorizer
= get_authorizer
,
4022 .verify_authorizer_reply
= verify_authorizer_reply
,
4023 .invalidate_authorizer
= invalidate_authorizer
,
4024 .peer_reset
= peer_reset
,
4025 .alloc_msg
= mds_alloc_msg
,
4026 .sign_message
= mds_sign_message
,
4027 .check_message_signature
= mds_check_message_signature
,