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45 * are no longer valid.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 u32 op
= le32_to_cpu(info
->head
->op
);
293 if (op
== CEPH_MDS_OP_GETFILELOCK
)
294 return parse_reply_info_filelock(p
, end
, info
, features
);
295 else if (op
== CEPH_MDS_OP_READDIR
|| op
== CEPH_MDS_OP_LSSNAP
)
296 return parse_reply_info_dir(p
, end
, info
, features
);
297 else if (op
== CEPH_MDS_OP_CREATE
)
298 return parse_reply_info_create(p
, end
, info
, features
);
304 * parse entire mds reply
306 static int parse_reply_info(struct ceph_msg
*msg
,
307 struct ceph_mds_reply_info_parsed
*info
,
314 info
->head
= msg
->front
.iov_base
;
315 p
= msg
->front
.iov_base
+ sizeof(struct ceph_mds_reply_head
);
316 end
= p
+ msg
->front
.iov_len
- sizeof(struct ceph_mds_reply_head
);
319 ceph_decode_32_safe(&p
, end
, len
, bad
);
321 ceph_decode_need(&p
, end
, len
, bad
);
322 err
= parse_reply_info_trace(&p
, p
+len
, info
, features
);
328 ceph_decode_32_safe(&p
, end
, len
, bad
);
330 ceph_decode_need(&p
, end
, len
, bad
);
331 err
= parse_reply_info_extra(&p
, p
+len
, info
, features
);
337 ceph_decode_32_safe(&p
, end
, len
, bad
);
338 info
->snapblob_len
= len
;
349 pr_err("mds parse_reply err %d\n", err
);
353 static void destroy_reply_info(struct ceph_mds_reply_info_parsed
*info
)
355 if (!info
->dir_entries
)
357 free_pages((unsigned long)info
->dir_entries
, get_order(info
->dir_buf_size
));
364 const char *ceph_session_state_name(int s
)
367 case CEPH_MDS_SESSION_NEW
: return "new";
368 case CEPH_MDS_SESSION_OPENING
: return "opening";
369 case CEPH_MDS_SESSION_OPEN
: return "open";
370 case CEPH_MDS_SESSION_HUNG
: return "hung";
371 case CEPH_MDS_SESSION_CLOSING
: return "closing";
372 case CEPH_MDS_SESSION_RESTARTING
: return "restarting";
373 case CEPH_MDS_SESSION_RECONNECTING
: return "reconnecting";
374 case CEPH_MDS_SESSION_REJECTED
: return "rejected";
375 default: return "???";
379 static struct ceph_mds_session
*get_session(struct ceph_mds_session
*s
)
381 if (atomic_inc_not_zero(&s
->s_ref
)) {
382 dout("mdsc get_session %p %d -> %d\n", s
,
383 atomic_read(&s
->s_ref
)-1, atomic_read(&s
->s_ref
));
386 dout("mdsc get_session %p 0 -- FAIL", s
);
391 void ceph_put_mds_session(struct ceph_mds_session
*s
)
393 dout("mdsc put_session %p %d -> %d\n", s
,
394 atomic_read(&s
->s_ref
), atomic_read(&s
->s_ref
)-1);
395 if (atomic_dec_and_test(&s
->s_ref
)) {
396 if (s
->s_auth
.authorizer
)
397 ceph_auth_destroy_authorizer(s
->s_auth
.authorizer
);
403 * called under mdsc->mutex
405 struct ceph_mds_session
*__ceph_lookup_mds_session(struct ceph_mds_client
*mdsc
,
408 struct ceph_mds_session
*session
;
410 if (mds
>= mdsc
->max_sessions
|| mdsc
->sessions
[mds
] == NULL
)
412 session
= mdsc
->sessions
[mds
];
413 dout("lookup_mds_session %p %d\n", session
,
414 atomic_read(&session
->s_ref
));
415 get_session(session
);
419 static bool __have_session(struct ceph_mds_client
*mdsc
, int mds
)
421 if (mds
>= mdsc
->max_sessions
)
423 return mdsc
->sessions
[mds
];
426 static int __verify_registered_session(struct ceph_mds_client
*mdsc
,
427 struct ceph_mds_session
*s
)
429 if (s
->s_mds
>= mdsc
->max_sessions
||
430 mdsc
->sessions
[s
->s_mds
] != s
)
436 * create+register a new session for given mds.
437 * called under mdsc->mutex.
439 static struct ceph_mds_session
*register_session(struct ceph_mds_client
*mdsc
,
442 struct ceph_mds_session
*s
;
444 if (mds
>= mdsc
->mdsmap
->m_max_mds
)
445 return ERR_PTR(-EINVAL
);
447 s
= kzalloc(sizeof(*s
), GFP_NOFS
);
449 return ERR_PTR(-ENOMEM
);
452 s
->s_state
= CEPH_MDS_SESSION_NEW
;
455 mutex_init(&s
->s_mutex
);
457 ceph_con_init(&s
->s_con
, s
, &mds_con_ops
, &mdsc
->fsc
->client
->msgr
);
459 spin_lock_init(&s
->s_gen_ttl_lock
);
461 s
->s_cap_ttl
= jiffies
- 1;
463 spin_lock_init(&s
->s_cap_lock
);
464 s
->s_renew_requested
= 0;
466 INIT_LIST_HEAD(&s
->s_caps
);
469 atomic_set(&s
->s_ref
, 1);
470 INIT_LIST_HEAD(&s
->s_waiting
);
471 INIT_LIST_HEAD(&s
->s_unsafe
);
472 s
->s_num_cap_releases
= 0;
473 s
->s_cap_reconnect
= 0;
474 s
->s_cap_iterator
= NULL
;
475 INIT_LIST_HEAD(&s
->s_cap_releases
);
476 INIT_LIST_HEAD(&s
->s_cap_flushing
);
478 dout("register_session mds%d\n", mds
);
479 if (mds
>= mdsc
->max_sessions
) {
480 int newmax
= 1 << get_count_order(mds
+1);
481 struct ceph_mds_session
**sa
;
483 dout("register_session realloc to %d\n", newmax
);
484 sa
= kcalloc(newmax
, sizeof(void *), GFP_NOFS
);
487 if (mdsc
->sessions
) {
488 memcpy(sa
, mdsc
->sessions
,
489 mdsc
->max_sessions
* sizeof(void *));
490 kfree(mdsc
->sessions
);
493 mdsc
->max_sessions
= newmax
;
495 mdsc
->sessions
[mds
] = s
;
496 atomic_inc(&mdsc
->num_sessions
);
497 atomic_inc(&s
->s_ref
); /* one ref to sessions[], one to caller */
499 ceph_con_open(&s
->s_con
, CEPH_ENTITY_TYPE_MDS
, mds
,
500 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
506 return ERR_PTR(-ENOMEM
);
510 * called under mdsc->mutex
512 static void __unregister_session(struct ceph_mds_client
*mdsc
,
513 struct ceph_mds_session
*s
)
515 dout("__unregister_session mds%d %p\n", s
->s_mds
, s
);
516 BUG_ON(mdsc
->sessions
[s
->s_mds
] != s
);
517 mdsc
->sessions
[s
->s_mds
] = NULL
;
518 ceph_con_close(&s
->s_con
);
519 ceph_put_mds_session(s
);
520 atomic_dec(&mdsc
->num_sessions
);
524 * drop session refs in request.
526 * should be last request ref, or hold mdsc->mutex
528 static void put_request_session(struct ceph_mds_request
*req
)
530 if (req
->r_session
) {
531 ceph_put_mds_session(req
->r_session
);
532 req
->r_session
= NULL
;
536 void ceph_mdsc_release_request(struct kref
*kref
)
538 struct ceph_mds_request
*req
= container_of(kref
,
539 struct ceph_mds_request
,
541 destroy_reply_info(&req
->r_reply_info
);
543 ceph_msg_put(req
->r_request
);
545 ceph_msg_put(req
->r_reply
);
547 ceph_put_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
550 if (req
->r_locked_dir
)
551 ceph_put_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
552 iput(req
->r_target_inode
);
555 if (req
->r_old_dentry
)
556 dput(req
->r_old_dentry
);
557 if (req
->r_old_dentry_dir
) {
559 * track (and drop pins for) r_old_dentry_dir
560 * separately, since r_old_dentry's d_parent may have
561 * changed between the dir mutex being dropped and
562 * this request being freed.
564 ceph_put_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
566 iput(req
->r_old_dentry_dir
);
571 ceph_pagelist_release(req
->r_pagelist
);
572 put_request_session(req
);
573 ceph_unreserve_caps(req
->r_mdsc
, &req
->r_caps_reservation
);
577 DEFINE_RB_FUNCS(request
, struct ceph_mds_request
, r_tid
, r_node
)
580 * lookup session, bump ref if found.
582 * called under mdsc->mutex.
584 static struct ceph_mds_request
*
585 lookup_get_request(struct ceph_mds_client
*mdsc
, u64 tid
)
587 struct ceph_mds_request
*req
;
589 req
= lookup_request(&mdsc
->request_tree
, tid
);
591 ceph_mdsc_get_request(req
);
597 * Register an in-flight request, and assign a tid. Link to directory
598 * are modifying (if any).
600 * Called under mdsc->mutex.
602 static void __register_request(struct ceph_mds_client
*mdsc
,
603 struct ceph_mds_request
*req
,
606 req
->r_tid
= ++mdsc
->last_tid
;
608 ceph_reserve_caps(mdsc
, &req
->r_caps_reservation
,
610 dout("__register_request %p tid %lld\n", req
, req
->r_tid
);
611 ceph_mdsc_get_request(req
);
612 insert_request(&mdsc
->request_tree
, req
);
614 req
->r_uid
= current_fsuid();
615 req
->r_gid
= current_fsgid();
617 if (mdsc
->oldest_tid
== 0 && req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
)
618 mdsc
->oldest_tid
= req
->r_tid
;
622 req
->r_unsafe_dir
= dir
;
626 static void __unregister_request(struct ceph_mds_client
*mdsc
,
627 struct ceph_mds_request
*req
)
629 dout("__unregister_request %p tid %lld\n", req
, req
->r_tid
);
631 /* Never leave an unregistered request on an unsafe list! */
632 list_del_init(&req
->r_unsafe_item
);
634 if (req
->r_tid
== mdsc
->oldest_tid
) {
635 struct rb_node
*p
= rb_next(&req
->r_node
);
636 mdsc
->oldest_tid
= 0;
638 struct ceph_mds_request
*next_req
=
639 rb_entry(p
, struct ceph_mds_request
, r_node
);
640 if (next_req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
) {
641 mdsc
->oldest_tid
= next_req
->r_tid
;
648 erase_request(&mdsc
->request_tree
, req
);
650 if (req
->r_unsafe_dir
&& req
->r_got_unsafe
) {
651 struct ceph_inode_info
*ci
= ceph_inode(req
->r_unsafe_dir
);
652 spin_lock(&ci
->i_unsafe_lock
);
653 list_del_init(&req
->r_unsafe_dir_item
);
654 spin_unlock(&ci
->i_unsafe_lock
);
656 if (req
->r_target_inode
&& req
->r_got_unsafe
) {
657 struct ceph_inode_info
*ci
= ceph_inode(req
->r_target_inode
);
658 spin_lock(&ci
->i_unsafe_lock
);
659 list_del_init(&req
->r_unsafe_target_item
);
660 spin_unlock(&ci
->i_unsafe_lock
);
663 if (req
->r_unsafe_dir
) {
664 iput(req
->r_unsafe_dir
);
665 req
->r_unsafe_dir
= NULL
;
668 complete_all(&req
->r_safe_completion
);
670 ceph_mdsc_put_request(req
);
674 * Choose mds to send request to next. If there is a hint set in the
675 * request (e.g., due to a prior forward hint from the mds), use that.
676 * Otherwise, consult frag tree and/or caps to identify the
677 * appropriate mds. If all else fails, choose randomly.
679 * Called under mdsc->mutex.
681 static struct dentry
*get_nonsnap_parent(struct dentry
*dentry
)
684 * we don't need to worry about protecting the d_parent access
685 * here because we never renaming inside the snapped namespace
686 * except to resplice to another snapdir, and either the old or new
687 * result is a valid result.
689 while (!IS_ROOT(dentry
) && ceph_snap(d_inode(dentry
)) != CEPH_NOSNAP
)
690 dentry
= dentry
->d_parent
;
694 static int __choose_mds(struct ceph_mds_client
*mdsc
,
695 struct ceph_mds_request
*req
)
698 struct ceph_inode_info
*ci
;
699 struct ceph_cap
*cap
;
700 int mode
= req
->r_direct_mode
;
702 u32 hash
= req
->r_direct_hash
;
703 bool is_hash
= req
->r_direct_is_hash
;
706 * is there a specific mds we should try? ignore hint if we have
707 * no session and the mds is not up (active or recovering).
709 if (req
->r_resend_mds
>= 0 &&
710 (__have_session(mdsc
, req
->r_resend_mds
) ||
711 ceph_mdsmap_get_state(mdsc
->mdsmap
, req
->r_resend_mds
) > 0)) {
712 dout("choose_mds using resend_mds mds%d\n",
714 return req
->r_resend_mds
;
717 if (mode
== USE_RANDOM_MDS
)
722 inode
= req
->r_inode
;
723 } else if (req
->r_dentry
) {
724 /* ignore race with rename; old or new d_parent is okay */
725 struct dentry
*parent
= req
->r_dentry
->d_parent
;
726 struct inode
*dir
= d_inode(parent
);
728 if (dir
->i_sb
!= mdsc
->fsc
->sb
) {
730 inode
= d_inode(req
->r_dentry
);
731 } else if (ceph_snap(dir
) != CEPH_NOSNAP
) {
732 /* direct snapped/virtual snapdir requests
733 * based on parent dir inode */
734 struct dentry
*dn
= get_nonsnap_parent(parent
);
736 dout("__choose_mds using nonsnap parent %p\n", inode
);
739 inode
= d_inode(req
->r_dentry
);
740 if (!inode
|| mode
== USE_AUTH_MDS
) {
743 hash
= ceph_dentry_hash(dir
, req
->r_dentry
);
749 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode
, (int)is_hash
,
753 ci
= ceph_inode(inode
);
755 if (is_hash
&& S_ISDIR(inode
->i_mode
)) {
756 struct ceph_inode_frag frag
;
759 ceph_choose_frag(ci
, hash
, &frag
, &found
);
761 if (mode
== USE_ANY_MDS
&& frag
.ndist
> 0) {
764 /* choose a random replica */
765 get_random_bytes(&r
, 1);
768 dout("choose_mds %p %llx.%llx "
769 "frag %u mds%d (%d/%d)\n",
770 inode
, ceph_vinop(inode
),
773 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
774 CEPH_MDS_STATE_ACTIVE
)
778 /* since this file/dir wasn't known to be
779 * replicated, then we want to look for the
780 * authoritative mds. */
783 /* choose auth mds */
785 dout("choose_mds %p %llx.%llx "
786 "frag %u mds%d (auth)\n",
787 inode
, ceph_vinop(inode
), frag
.frag
, mds
);
788 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
789 CEPH_MDS_STATE_ACTIVE
)
795 spin_lock(&ci
->i_ceph_lock
);
797 if (mode
== USE_AUTH_MDS
)
798 cap
= ci
->i_auth_cap
;
799 if (!cap
&& !RB_EMPTY_ROOT(&ci
->i_caps
))
800 cap
= rb_entry(rb_first(&ci
->i_caps
), struct ceph_cap
, ci_node
);
802 spin_unlock(&ci
->i_ceph_lock
);
805 mds
= cap
->session
->s_mds
;
806 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
807 inode
, ceph_vinop(inode
), mds
,
808 cap
== ci
->i_auth_cap
? "auth " : "", cap
);
809 spin_unlock(&ci
->i_ceph_lock
);
813 mds
= ceph_mdsmap_get_random_mds(mdsc
->mdsmap
);
814 dout("choose_mds chose random mds%d\n", mds
);
822 static struct ceph_msg
*create_session_msg(u32 op
, u64 seq
)
824 struct ceph_msg
*msg
;
825 struct ceph_mds_session_head
*h
;
827 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
), GFP_NOFS
,
830 pr_err("create_session_msg ENOMEM creating msg\n");
833 h
= msg
->front
.iov_base
;
834 h
->op
= cpu_to_le32(op
);
835 h
->seq
= cpu_to_le64(seq
);
841 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
842 * to include additional client metadata fields.
844 static struct ceph_msg
*create_session_open_msg(struct ceph_mds_client
*mdsc
, u64 seq
)
846 struct ceph_msg
*msg
;
847 struct ceph_mds_session_head
*h
;
849 int metadata_bytes
= 0;
850 int metadata_key_count
= 0;
851 struct ceph_options
*opt
= mdsc
->fsc
->client
->options
;
852 struct ceph_mount_options
*fsopt
= mdsc
->fsc
->mount_options
;
855 const char* metadata
[][2] = {
856 {"hostname", utsname()->nodename
},
857 {"kernel_version", utsname()->release
},
858 {"entity_id", opt
->name
? : ""},
859 {"root", fsopt
->server_path
? : "/"},
863 /* Calculate serialized length of metadata */
864 metadata_bytes
= 4; /* map length */
865 for (i
= 0; metadata
[i
][0] != NULL
; ++i
) {
866 metadata_bytes
+= 8 + strlen(metadata
[i
][0]) +
867 strlen(metadata
[i
][1]);
868 metadata_key_count
++;
871 /* Allocate the message */
872 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
) + metadata_bytes
,
875 pr_err("create_session_msg ENOMEM creating msg\n");
878 h
= msg
->front
.iov_base
;
879 h
->op
= cpu_to_le32(CEPH_SESSION_REQUEST_OPEN
);
880 h
->seq
= cpu_to_le64(seq
);
883 * Serialize client metadata into waiting buffer space, using
884 * the format that userspace expects for map<string, string>
886 * ClientSession messages with metadata are v2
888 msg
->hdr
.version
= cpu_to_le16(2);
889 msg
->hdr
.compat_version
= cpu_to_le16(1);
891 /* The write pointer, following the session_head structure */
892 p
= msg
->front
.iov_base
+ sizeof(*h
);
894 /* Number of entries in the map */
895 ceph_encode_32(&p
, metadata_key_count
);
897 /* Two length-prefixed strings for each entry in the map */
898 for (i
= 0; metadata
[i
][0] != NULL
; ++i
) {
899 size_t const key_len
= strlen(metadata
[i
][0]);
900 size_t const val_len
= strlen(metadata
[i
][1]);
902 ceph_encode_32(&p
, key_len
);
903 memcpy(p
, metadata
[i
][0], key_len
);
905 ceph_encode_32(&p
, val_len
);
906 memcpy(p
, metadata
[i
][1], val_len
);
914 * send session open request.
916 * called under mdsc->mutex
918 static int __open_session(struct ceph_mds_client
*mdsc
,
919 struct ceph_mds_session
*session
)
921 struct ceph_msg
*msg
;
923 int mds
= session
->s_mds
;
925 /* wait for mds to go active? */
926 mstate
= ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
);
927 dout("open_session to mds%d (%s)\n", mds
,
928 ceph_mds_state_name(mstate
));
929 session
->s_state
= CEPH_MDS_SESSION_OPENING
;
930 session
->s_renew_requested
= jiffies
;
932 /* send connect message */
933 msg
= create_session_open_msg(mdsc
, session
->s_seq
);
936 ceph_con_send(&session
->s_con
, msg
);
941 * open sessions for any export targets for the given mds
943 * called under mdsc->mutex
945 static struct ceph_mds_session
*
946 __open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
948 struct ceph_mds_session
*session
;
950 session
= __ceph_lookup_mds_session(mdsc
, target
);
952 session
= register_session(mdsc
, target
);
956 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
957 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
958 __open_session(mdsc
, session
);
963 struct ceph_mds_session
*
964 ceph_mdsc_open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
966 struct ceph_mds_session
*session
;
968 dout("open_export_target_session to mds%d\n", target
);
970 mutex_lock(&mdsc
->mutex
);
971 session
= __open_export_target_session(mdsc
, target
);
972 mutex_unlock(&mdsc
->mutex
);
977 static void __open_export_target_sessions(struct ceph_mds_client
*mdsc
,
978 struct ceph_mds_session
*session
)
980 struct ceph_mds_info
*mi
;
981 struct ceph_mds_session
*ts
;
982 int i
, mds
= session
->s_mds
;
984 if (mds
>= mdsc
->mdsmap
->m_max_mds
)
987 mi
= &mdsc
->mdsmap
->m_info
[mds
];
988 dout("open_export_target_sessions for mds%d (%d targets)\n",
989 session
->s_mds
, mi
->num_export_targets
);
991 for (i
= 0; i
< mi
->num_export_targets
; i
++) {
992 ts
= __open_export_target_session(mdsc
, mi
->export_targets
[i
]);
994 ceph_put_mds_session(ts
);
998 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client
*mdsc
,
999 struct ceph_mds_session
*session
)
1001 mutex_lock(&mdsc
->mutex
);
1002 __open_export_target_sessions(mdsc
, session
);
1003 mutex_unlock(&mdsc
->mutex
);
1010 /* caller holds s_cap_lock, we drop it */
1011 static void cleanup_cap_releases(struct ceph_mds_client
*mdsc
,
1012 struct ceph_mds_session
*session
)
1013 __releases(session
->s_cap_lock
)
1015 LIST_HEAD(tmp_list
);
1016 list_splice_init(&session
->s_cap_releases
, &tmp_list
);
1017 session
->s_num_cap_releases
= 0;
1018 spin_unlock(&session
->s_cap_lock
);
1020 dout("cleanup_cap_releases mds%d\n", session
->s_mds
);
1021 while (!list_empty(&tmp_list
)) {
1022 struct ceph_cap
*cap
;
1023 /* zero out the in-progress message */
1024 cap
= list_first_entry(&tmp_list
,
1025 struct ceph_cap
, session_caps
);
1026 list_del(&cap
->session_caps
);
1027 ceph_put_cap(mdsc
, cap
);
1031 static void cleanup_session_requests(struct ceph_mds_client
*mdsc
,
1032 struct ceph_mds_session
*session
)
1034 struct ceph_mds_request
*req
;
1037 dout("cleanup_session_requests mds%d\n", session
->s_mds
);
1038 mutex_lock(&mdsc
->mutex
);
1039 while (!list_empty(&session
->s_unsafe
)) {
1040 req
= list_first_entry(&session
->s_unsafe
,
1041 struct ceph_mds_request
, r_unsafe_item
);
1042 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1044 __unregister_request(mdsc
, req
);
1046 /* zero r_attempts, so kick_requests() will re-send requests */
1047 p
= rb_first(&mdsc
->request_tree
);
1049 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
1051 if (req
->r_session
&&
1052 req
->r_session
->s_mds
== session
->s_mds
)
1053 req
->r_attempts
= 0;
1055 mutex_unlock(&mdsc
->mutex
);
1059 * Helper to safely iterate over all caps associated with a session, with
1060 * special care taken to handle a racing __ceph_remove_cap().
1062 * Caller must hold session s_mutex.
1064 static int iterate_session_caps(struct ceph_mds_session
*session
,
1065 int (*cb
)(struct inode
*, struct ceph_cap
*,
1068 struct list_head
*p
;
1069 struct ceph_cap
*cap
;
1070 struct inode
*inode
, *last_inode
= NULL
;
1071 struct ceph_cap
*old_cap
= NULL
;
1074 dout("iterate_session_caps %p mds%d\n", session
, session
->s_mds
);
1075 spin_lock(&session
->s_cap_lock
);
1076 p
= session
->s_caps
.next
;
1077 while (p
!= &session
->s_caps
) {
1078 cap
= list_entry(p
, struct ceph_cap
, session_caps
);
1079 inode
= igrab(&cap
->ci
->vfs_inode
);
1084 session
->s_cap_iterator
= cap
;
1085 spin_unlock(&session
->s_cap_lock
);
1092 ceph_put_cap(session
->s_mdsc
, old_cap
);
1096 ret
= cb(inode
, cap
, arg
);
1099 spin_lock(&session
->s_cap_lock
);
1101 if (cap
->ci
== NULL
) {
1102 dout("iterate_session_caps finishing cap %p removal\n",
1104 BUG_ON(cap
->session
!= session
);
1105 cap
->session
= NULL
;
1106 list_del_init(&cap
->session_caps
);
1107 session
->s_nr_caps
--;
1108 if (cap
->queue_release
) {
1109 list_add_tail(&cap
->session_caps
,
1110 &session
->s_cap_releases
);
1111 session
->s_num_cap_releases
++;
1113 old_cap
= cap
; /* put_cap it w/o locks held */
1121 session
->s_cap_iterator
= NULL
;
1122 spin_unlock(&session
->s_cap_lock
);
1126 ceph_put_cap(session
->s_mdsc
, old_cap
);
1131 static int remove_session_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1134 struct ceph_fs_client
*fsc
= (struct ceph_fs_client
*)arg
;
1135 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1136 LIST_HEAD(to_remove
);
1138 bool invalidate
= false;
1140 dout("removing cap %p, ci is %p, inode is %p\n",
1141 cap
, ci
, &ci
->vfs_inode
);
1142 spin_lock(&ci
->i_ceph_lock
);
1143 __ceph_remove_cap(cap
, false);
1144 if (!ci
->i_auth_cap
) {
1145 struct ceph_cap_flush
*cf
;
1146 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
1148 ci
->i_ceph_flags
|= CEPH_I_CAP_DROPPED
;
1150 if (ci
->i_wrbuffer_ref
> 0 &&
1151 ACCESS_ONCE(fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
1154 while (!list_empty(&ci
->i_cap_flush_list
)) {
1155 cf
= list_first_entry(&ci
->i_cap_flush_list
,
1156 struct ceph_cap_flush
, i_list
);
1157 list_move(&cf
->i_list
, &to_remove
);
1160 spin_lock(&mdsc
->cap_dirty_lock
);
1162 list_for_each_entry(cf
, &to_remove
, i_list
)
1163 list_del(&cf
->g_list
);
1165 if (!list_empty(&ci
->i_dirty_item
)) {
1166 pr_warn_ratelimited(
1167 " dropping dirty %s state for %p %lld\n",
1168 ceph_cap_string(ci
->i_dirty_caps
),
1169 inode
, ceph_ino(inode
));
1170 ci
->i_dirty_caps
= 0;
1171 list_del_init(&ci
->i_dirty_item
);
1174 if (!list_empty(&ci
->i_flushing_item
)) {
1175 pr_warn_ratelimited(
1176 " dropping dirty+flushing %s state for %p %lld\n",
1177 ceph_cap_string(ci
->i_flushing_caps
),
1178 inode
, ceph_ino(inode
));
1179 ci
->i_flushing_caps
= 0;
1180 list_del_init(&ci
->i_flushing_item
);
1181 mdsc
->num_cap_flushing
--;
1184 spin_unlock(&mdsc
->cap_dirty_lock
);
1186 if (!ci
->i_dirty_caps
&& ci
->i_prealloc_cap_flush
) {
1187 list_add(&ci
->i_prealloc_cap_flush
->i_list
, &to_remove
);
1188 ci
->i_prealloc_cap_flush
= NULL
;
1191 spin_unlock(&ci
->i_ceph_lock
);
1192 while (!list_empty(&to_remove
)) {
1193 struct ceph_cap_flush
*cf
;
1194 cf
= list_first_entry(&to_remove
,
1195 struct ceph_cap_flush
, i_list
);
1196 list_del(&cf
->i_list
);
1197 ceph_free_cap_flush(cf
);
1200 wake_up_all(&ci
->i_cap_wq
);
1202 ceph_queue_invalidate(inode
);
1209 * caller must hold session s_mutex
1211 static void remove_session_caps(struct ceph_mds_session
*session
)
1213 struct ceph_fs_client
*fsc
= session
->s_mdsc
->fsc
;
1214 struct super_block
*sb
= fsc
->sb
;
1215 dout("remove_session_caps on %p\n", session
);
1216 iterate_session_caps(session
, remove_session_caps_cb
, fsc
);
1218 wake_up_all(&fsc
->mdsc
->cap_flushing_wq
);
1220 spin_lock(&session
->s_cap_lock
);
1221 if (session
->s_nr_caps
> 0) {
1222 struct inode
*inode
;
1223 struct ceph_cap
*cap
, *prev
= NULL
;
1224 struct ceph_vino vino
;
1226 * iterate_session_caps() skips inodes that are being
1227 * deleted, we need to wait until deletions are complete.
1228 * __wait_on_freeing_inode() is designed for the job,
1229 * but it is not exported, so use lookup inode function
1232 while (!list_empty(&session
->s_caps
)) {
1233 cap
= list_entry(session
->s_caps
.next
,
1234 struct ceph_cap
, session_caps
);
1238 vino
= cap
->ci
->i_vino
;
1239 spin_unlock(&session
->s_cap_lock
);
1241 inode
= ceph_find_inode(sb
, vino
);
1244 spin_lock(&session
->s_cap_lock
);
1248 // drop cap expires and unlock s_cap_lock
1249 cleanup_cap_releases(session
->s_mdsc
, session
);
1251 BUG_ON(session
->s_nr_caps
> 0);
1252 BUG_ON(!list_empty(&session
->s_cap_flushing
));
1256 * wake up any threads waiting on this session's caps. if the cap is
1257 * old (didn't get renewed on the client reconnect), remove it now.
1259 * caller must hold s_mutex.
1261 static int wake_up_session_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1264 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1267 spin_lock(&ci
->i_ceph_lock
);
1268 ci
->i_wanted_max_size
= 0;
1269 ci
->i_requested_max_size
= 0;
1270 spin_unlock(&ci
->i_ceph_lock
);
1272 wake_up_all(&ci
->i_cap_wq
);
1276 static void wake_up_session_caps(struct ceph_mds_session
*session
,
1279 dout("wake_up_session_caps %p mds%d\n", session
, session
->s_mds
);
1280 iterate_session_caps(session
, wake_up_session_cb
,
1281 (void *)(unsigned long)reconnect
);
1285 * Send periodic message to MDS renewing all currently held caps. The
1286 * ack will reset the expiration for all caps from this session.
1288 * caller holds s_mutex
1290 static int send_renew_caps(struct ceph_mds_client
*mdsc
,
1291 struct ceph_mds_session
*session
)
1293 struct ceph_msg
*msg
;
1296 if (time_after_eq(jiffies
, session
->s_cap_ttl
) &&
1297 time_after_eq(session
->s_cap_ttl
, session
->s_renew_requested
))
1298 pr_info("mds%d caps stale\n", session
->s_mds
);
1299 session
->s_renew_requested
= jiffies
;
1301 /* do not try to renew caps until a recovering mds has reconnected
1302 * with its clients. */
1303 state
= ceph_mdsmap_get_state(mdsc
->mdsmap
, session
->s_mds
);
1304 if (state
< CEPH_MDS_STATE_RECONNECT
) {
1305 dout("send_renew_caps ignoring mds%d (%s)\n",
1306 session
->s_mds
, ceph_mds_state_name(state
));
1310 dout("send_renew_caps to mds%d (%s)\n", session
->s_mds
,
1311 ceph_mds_state_name(state
));
1312 msg
= create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS
,
1313 ++session
->s_renew_seq
);
1316 ceph_con_send(&session
->s_con
, msg
);
1320 static int send_flushmsg_ack(struct ceph_mds_client
*mdsc
,
1321 struct ceph_mds_session
*session
, u64 seq
)
1323 struct ceph_msg
*msg
;
1325 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1326 session
->s_mds
, ceph_session_state_name(session
->s_state
), seq
);
1327 msg
= create_session_msg(CEPH_SESSION_FLUSHMSG_ACK
, seq
);
1330 ceph_con_send(&session
->s_con
, msg
);
1336 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1338 * Called under session->s_mutex
1340 static void renewed_caps(struct ceph_mds_client
*mdsc
,
1341 struct ceph_mds_session
*session
, int is_renew
)
1346 spin_lock(&session
->s_cap_lock
);
1347 was_stale
= is_renew
&& time_after_eq(jiffies
, session
->s_cap_ttl
);
1349 session
->s_cap_ttl
= session
->s_renew_requested
+
1350 mdsc
->mdsmap
->m_session_timeout
*HZ
;
1353 if (time_before(jiffies
, session
->s_cap_ttl
)) {
1354 pr_info("mds%d caps renewed\n", session
->s_mds
);
1357 pr_info("mds%d caps still stale\n", session
->s_mds
);
1360 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1361 session
->s_mds
, session
->s_cap_ttl
, was_stale
? "stale" : "fresh",
1362 time_before(jiffies
, session
->s_cap_ttl
) ? "stale" : "fresh");
1363 spin_unlock(&session
->s_cap_lock
);
1366 wake_up_session_caps(session
, 0);
1370 * send a session close request
1372 static int request_close_session(struct ceph_mds_client
*mdsc
,
1373 struct ceph_mds_session
*session
)
1375 struct ceph_msg
*msg
;
1377 dout("request_close_session mds%d state %s seq %lld\n",
1378 session
->s_mds
, ceph_session_state_name(session
->s_state
),
1380 msg
= create_session_msg(CEPH_SESSION_REQUEST_CLOSE
, session
->s_seq
);
1383 ceph_con_send(&session
->s_con
, msg
);
1388 * Called with s_mutex held.
1390 static int __close_session(struct ceph_mds_client
*mdsc
,
1391 struct ceph_mds_session
*session
)
1393 if (session
->s_state
>= CEPH_MDS_SESSION_CLOSING
)
1395 session
->s_state
= CEPH_MDS_SESSION_CLOSING
;
1396 return request_close_session(mdsc
, session
);
1400 * Trim old(er) caps.
1402 * Because we can't cache an inode without one or more caps, we do
1403 * this indirectly: if a cap is unused, we prune its aliases, at which
1404 * point the inode will hopefully get dropped to.
1406 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1407 * memory pressure from the MDS, though, so it needn't be perfect.
1409 static int trim_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
, void *arg
)
1411 struct ceph_mds_session
*session
= arg
;
1412 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1413 int used
, wanted
, oissued
, mine
;
1415 if (session
->s_trim_caps
<= 0)
1418 spin_lock(&ci
->i_ceph_lock
);
1419 mine
= cap
->issued
| cap
->implemented
;
1420 used
= __ceph_caps_used(ci
);
1421 wanted
= __ceph_caps_file_wanted(ci
);
1422 oissued
= __ceph_caps_issued_other(ci
, cap
);
1424 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1425 inode
, cap
, ceph_cap_string(mine
), ceph_cap_string(oissued
),
1426 ceph_cap_string(used
), ceph_cap_string(wanted
));
1427 if (cap
== ci
->i_auth_cap
) {
1428 if (ci
->i_dirty_caps
|| ci
->i_flushing_caps
||
1429 !list_empty(&ci
->i_cap_snaps
))
1431 if ((used
| wanted
) & CEPH_CAP_ANY_WR
)
1434 /* The inode has cached pages, but it's no longer used.
1435 * we can safely drop it */
1436 if (wanted
== 0 && used
== CEPH_CAP_FILE_CACHE
&&
1437 !(oissued
& CEPH_CAP_FILE_CACHE
)) {
1441 if ((used
| wanted
) & ~oissued
& mine
)
1442 goto out
; /* we need these caps */
1444 session
->s_trim_caps
--;
1446 /* we aren't the only cap.. just remove us */
1447 __ceph_remove_cap(cap
, true);
1449 /* try dropping referring dentries */
1450 spin_unlock(&ci
->i_ceph_lock
);
1451 d_prune_aliases(inode
);
1452 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1453 inode
, cap
, atomic_read(&inode
->i_count
));
1458 spin_unlock(&ci
->i_ceph_lock
);
1463 * Trim session cap count down to some max number.
1465 static int trim_caps(struct ceph_mds_client
*mdsc
,
1466 struct ceph_mds_session
*session
,
1469 int trim_caps
= session
->s_nr_caps
- max_caps
;
1471 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1472 session
->s_mds
, session
->s_nr_caps
, max_caps
, trim_caps
);
1473 if (trim_caps
> 0) {
1474 session
->s_trim_caps
= trim_caps
;
1475 iterate_session_caps(session
, trim_caps_cb
, session
);
1476 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1477 session
->s_mds
, session
->s_nr_caps
, max_caps
,
1478 trim_caps
- session
->s_trim_caps
);
1479 session
->s_trim_caps
= 0;
1482 ceph_send_cap_releases(mdsc
, session
);
1486 static int check_caps_flush(struct ceph_mds_client
*mdsc
,
1491 spin_lock(&mdsc
->cap_dirty_lock
);
1492 if (!list_empty(&mdsc
->cap_flush_list
)) {
1493 struct ceph_cap_flush
*cf
=
1494 list_first_entry(&mdsc
->cap_flush_list
,
1495 struct ceph_cap_flush
, g_list
);
1496 if (cf
->tid
<= want_flush_tid
) {
1497 dout("check_caps_flush still flushing tid "
1498 "%llu <= %llu\n", cf
->tid
, want_flush_tid
);
1502 spin_unlock(&mdsc
->cap_dirty_lock
);
1507 * flush all dirty inode data to disk.
1509 * returns true if we've flushed through want_flush_tid
1511 static void wait_caps_flush(struct ceph_mds_client
*mdsc
,
1514 dout("check_caps_flush want %llu\n", want_flush_tid
);
1516 wait_event(mdsc
->cap_flushing_wq
,
1517 check_caps_flush(mdsc
, want_flush_tid
));
1519 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid
);
1523 * called under s_mutex
1525 void ceph_send_cap_releases(struct ceph_mds_client
*mdsc
,
1526 struct ceph_mds_session
*session
)
1528 struct ceph_msg
*msg
= NULL
;
1529 struct ceph_mds_cap_release
*head
;
1530 struct ceph_mds_cap_item
*item
;
1531 struct ceph_cap
*cap
;
1532 LIST_HEAD(tmp_list
);
1533 int num_cap_releases
;
1535 spin_lock(&session
->s_cap_lock
);
1537 list_splice_init(&session
->s_cap_releases
, &tmp_list
);
1538 num_cap_releases
= session
->s_num_cap_releases
;
1539 session
->s_num_cap_releases
= 0;
1540 spin_unlock(&session
->s_cap_lock
);
1542 while (!list_empty(&tmp_list
)) {
1544 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE
,
1545 PAGE_SIZE
, GFP_NOFS
, false);
1548 head
= msg
->front
.iov_base
;
1549 head
->num
= cpu_to_le32(0);
1550 msg
->front
.iov_len
= sizeof(*head
);
1552 cap
= list_first_entry(&tmp_list
, struct ceph_cap
,
1554 list_del(&cap
->session_caps
);
1557 head
= msg
->front
.iov_base
;
1558 le32_add_cpu(&head
->num
, 1);
1559 item
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1560 item
->ino
= cpu_to_le64(cap
->cap_ino
);
1561 item
->cap_id
= cpu_to_le64(cap
->cap_id
);
1562 item
->migrate_seq
= cpu_to_le32(cap
->mseq
);
1563 item
->seq
= cpu_to_le32(cap
->issue_seq
);
1564 msg
->front
.iov_len
+= sizeof(*item
);
1566 ceph_put_cap(mdsc
, cap
);
1568 if (le32_to_cpu(head
->num
) == CEPH_CAPS_PER_RELEASE
) {
1569 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1570 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1571 ceph_con_send(&session
->s_con
, msg
);
1576 BUG_ON(num_cap_releases
!= 0);
1578 spin_lock(&session
->s_cap_lock
);
1579 if (!list_empty(&session
->s_cap_releases
))
1581 spin_unlock(&session
->s_cap_lock
);
1584 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1585 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1586 ceph_con_send(&session
->s_con
, msg
);
1590 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1592 spin_lock(&session
->s_cap_lock
);
1593 list_splice(&tmp_list
, &session
->s_cap_releases
);
1594 session
->s_num_cap_releases
+= num_cap_releases
;
1595 spin_unlock(&session
->s_cap_lock
);
1602 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request
*req
,
1605 struct ceph_inode_info
*ci
= ceph_inode(dir
);
1606 struct ceph_mds_reply_info_parsed
*rinfo
= &req
->r_reply_info
;
1607 struct ceph_mount_options
*opt
= req
->r_mdsc
->fsc
->mount_options
;
1608 size_t size
= sizeof(struct ceph_mds_reply_dir_entry
);
1609 int order
, num_entries
;
1611 spin_lock(&ci
->i_ceph_lock
);
1612 num_entries
= ci
->i_files
+ ci
->i_subdirs
;
1613 spin_unlock(&ci
->i_ceph_lock
);
1614 num_entries
= max(num_entries
, 1);
1615 num_entries
= min(num_entries
, opt
->max_readdir
);
1617 order
= get_order(size
* num_entries
);
1618 while (order
>= 0) {
1619 rinfo
->dir_entries
= (void*)__get_free_pages(GFP_KERNEL
|
1622 if (rinfo
->dir_entries
)
1626 if (!rinfo
->dir_entries
)
1629 num_entries
= (PAGE_SIZE
<< order
) / size
;
1630 num_entries
= min(num_entries
, opt
->max_readdir
);
1632 rinfo
->dir_buf_size
= PAGE_SIZE
<< order
;
1633 req
->r_num_caps
= num_entries
+ 1;
1634 req
->r_args
.readdir
.max_entries
= cpu_to_le32(num_entries
);
1635 req
->r_args
.readdir
.max_bytes
= cpu_to_le32(opt
->max_readdir_bytes
);
1640 * Create an mds request.
1642 struct ceph_mds_request
*
1643 ceph_mdsc_create_request(struct ceph_mds_client
*mdsc
, int op
, int mode
)
1645 struct ceph_mds_request
*req
= kzalloc(sizeof(*req
), GFP_NOFS
);
1648 return ERR_PTR(-ENOMEM
);
1650 mutex_init(&req
->r_fill_mutex
);
1652 req
->r_started
= jiffies
;
1653 req
->r_resend_mds
= -1;
1654 INIT_LIST_HEAD(&req
->r_unsafe_dir_item
);
1655 INIT_LIST_HEAD(&req
->r_unsafe_target_item
);
1657 kref_init(&req
->r_kref
);
1658 RB_CLEAR_NODE(&req
->r_node
);
1659 INIT_LIST_HEAD(&req
->r_wait
);
1660 init_completion(&req
->r_completion
);
1661 init_completion(&req
->r_safe_completion
);
1662 INIT_LIST_HEAD(&req
->r_unsafe_item
);
1664 req
->r_stamp
= current_fs_time(mdsc
->fsc
->sb
);
1667 req
->r_direct_mode
= mode
;
1672 * return oldest (lowest) request, tid in request tree, 0 if none.
1674 * called under mdsc->mutex.
1676 static struct ceph_mds_request
*__get_oldest_req(struct ceph_mds_client
*mdsc
)
1678 if (RB_EMPTY_ROOT(&mdsc
->request_tree
))
1680 return rb_entry(rb_first(&mdsc
->request_tree
),
1681 struct ceph_mds_request
, r_node
);
1684 static inline u64
__get_oldest_tid(struct ceph_mds_client
*mdsc
)
1686 return mdsc
->oldest_tid
;
1690 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1691 * on build_path_from_dentry in fs/cifs/dir.c.
1693 * If @stop_on_nosnap, generate path relative to the first non-snapped
1696 * Encode hidden .snap dirs as a double /, i.e.
1697 * foo/.snap/bar -> foo//bar
1699 char *ceph_mdsc_build_path(struct dentry
*dentry
, int *plen
, u64
*base
,
1702 struct dentry
*temp
;
1708 return ERR_PTR(-EINVAL
);
1712 seq
= read_seqbegin(&rename_lock
);
1714 for (temp
= dentry
; !IS_ROOT(temp
);) {
1715 struct inode
*inode
= d_inode(temp
);
1716 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
)
1717 len
++; /* slash only */
1718 else if (stop_on_nosnap
&& inode
&&
1719 ceph_snap(inode
) == CEPH_NOSNAP
)
1722 len
+= 1 + temp
->d_name
.len
;
1723 temp
= temp
->d_parent
;
1727 len
--; /* no leading '/' */
1729 path
= kmalloc(len
+1, GFP_NOFS
);
1731 return ERR_PTR(-ENOMEM
);
1733 path
[pos
] = 0; /* trailing null */
1735 for (temp
= dentry
; !IS_ROOT(temp
) && pos
!= 0; ) {
1736 struct inode
*inode
;
1738 spin_lock(&temp
->d_lock
);
1739 inode
= d_inode(temp
);
1740 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
) {
1741 dout("build_path path+%d: %p SNAPDIR\n",
1743 } else if (stop_on_nosnap
&& inode
&&
1744 ceph_snap(inode
) == CEPH_NOSNAP
) {
1745 spin_unlock(&temp
->d_lock
);
1748 pos
-= temp
->d_name
.len
;
1750 spin_unlock(&temp
->d_lock
);
1753 strncpy(path
+ pos
, temp
->d_name
.name
,
1756 spin_unlock(&temp
->d_lock
);
1759 temp
= temp
->d_parent
;
1762 if (pos
!= 0 || read_seqretry(&rename_lock
, seq
)) {
1763 pr_err("build_path did not end path lookup where "
1764 "expected, namelen is %d, pos is %d\n", len
, pos
);
1765 /* presumably this is only possible if racing with a
1766 rename of one of the parent directories (we can not
1767 lock the dentries above us to prevent this, but
1768 retrying should be harmless) */
1773 *base
= ceph_ino(d_inode(temp
));
1775 dout("build_path on %p %d built %llx '%.*s'\n",
1776 dentry
, d_count(dentry
), *base
, len
, path
);
1780 static int build_dentry_path(struct dentry
*dentry
,
1781 const char **ppath
, int *ppathlen
, u64
*pino
,
1786 if (ceph_snap(d_inode(dentry
->d_parent
)) == CEPH_NOSNAP
) {
1787 *pino
= ceph_ino(d_inode(dentry
->d_parent
));
1788 *ppath
= dentry
->d_name
.name
;
1789 *ppathlen
= dentry
->d_name
.len
;
1792 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1794 return PTR_ERR(path
);
1800 static int build_inode_path(struct inode
*inode
,
1801 const char **ppath
, int *ppathlen
, u64
*pino
,
1804 struct dentry
*dentry
;
1807 if (ceph_snap(inode
) == CEPH_NOSNAP
) {
1808 *pino
= ceph_ino(inode
);
1812 dentry
= d_find_alias(inode
);
1813 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1816 return PTR_ERR(path
);
1823 * request arguments may be specified via an inode *, a dentry *, or
1824 * an explicit ino+path.
1826 static int set_request_path_attr(struct inode
*rinode
, struct dentry
*rdentry
,
1827 const char *rpath
, u64 rino
,
1828 const char **ppath
, int *pathlen
,
1829 u64
*ino
, int *freepath
)
1834 r
= build_inode_path(rinode
, ppath
, pathlen
, ino
, freepath
);
1835 dout(" inode %p %llx.%llx\n", rinode
, ceph_ino(rinode
),
1837 } else if (rdentry
) {
1838 r
= build_dentry_path(rdentry
, ppath
, pathlen
, ino
, freepath
);
1839 dout(" dentry %p %llx/%.*s\n", rdentry
, *ino
, *pathlen
,
1841 } else if (rpath
|| rino
) {
1844 *pathlen
= rpath
? strlen(rpath
) : 0;
1845 dout(" path %.*s\n", *pathlen
, rpath
);
1852 * called under mdsc->mutex
1854 static struct ceph_msg
*create_request_message(struct ceph_mds_client
*mdsc
,
1855 struct ceph_mds_request
*req
,
1856 int mds
, bool drop_cap_releases
)
1858 struct ceph_msg
*msg
;
1859 struct ceph_mds_request_head
*head
;
1860 const char *path1
= NULL
;
1861 const char *path2
= NULL
;
1862 u64 ino1
= 0, ino2
= 0;
1863 int pathlen1
= 0, pathlen2
= 0;
1864 int freepath1
= 0, freepath2
= 0;
1870 ret
= set_request_path_attr(req
->r_inode
, req
->r_dentry
,
1871 req
->r_path1
, req
->r_ino1
.ino
,
1872 &path1
, &pathlen1
, &ino1
, &freepath1
);
1878 ret
= set_request_path_attr(NULL
, req
->r_old_dentry
,
1879 req
->r_path2
, req
->r_ino2
.ino
,
1880 &path2
, &pathlen2
, &ino2
, &freepath2
);
1886 len
= sizeof(*head
) +
1887 pathlen1
+ pathlen2
+ 2*(1 + sizeof(u32
) + sizeof(u64
)) +
1888 sizeof(struct ceph_timespec
);
1890 /* calculate (max) length for cap releases */
1891 len
+= sizeof(struct ceph_mds_request_release
) *
1892 (!!req
->r_inode_drop
+ !!req
->r_dentry_drop
+
1893 !!req
->r_old_inode_drop
+ !!req
->r_old_dentry_drop
);
1894 if (req
->r_dentry_drop
)
1895 len
+= req
->r_dentry
->d_name
.len
;
1896 if (req
->r_old_dentry_drop
)
1897 len
+= req
->r_old_dentry
->d_name
.len
;
1899 msg
= ceph_msg_new(CEPH_MSG_CLIENT_REQUEST
, len
, GFP_NOFS
, false);
1901 msg
= ERR_PTR(-ENOMEM
);
1905 msg
->hdr
.version
= cpu_to_le16(2);
1906 msg
->hdr
.tid
= cpu_to_le64(req
->r_tid
);
1908 head
= msg
->front
.iov_base
;
1909 p
= msg
->front
.iov_base
+ sizeof(*head
);
1910 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1912 head
->mdsmap_epoch
= cpu_to_le32(mdsc
->mdsmap
->m_epoch
);
1913 head
->op
= cpu_to_le32(req
->r_op
);
1914 head
->caller_uid
= cpu_to_le32(from_kuid(&init_user_ns
, req
->r_uid
));
1915 head
->caller_gid
= cpu_to_le32(from_kgid(&init_user_ns
, req
->r_gid
));
1916 head
->args
= req
->r_args
;
1918 ceph_encode_filepath(&p
, end
, ino1
, path1
);
1919 ceph_encode_filepath(&p
, end
, ino2
, path2
);
1921 /* make note of release offset, in case we need to replay */
1922 req
->r_request_release_offset
= p
- msg
->front
.iov_base
;
1926 if (req
->r_inode_drop
)
1927 releases
+= ceph_encode_inode_release(&p
,
1928 req
->r_inode
? req
->r_inode
: d_inode(req
->r_dentry
),
1929 mds
, req
->r_inode_drop
, req
->r_inode_unless
, 0);
1930 if (req
->r_dentry_drop
)
1931 releases
+= ceph_encode_dentry_release(&p
, req
->r_dentry
,
1932 mds
, req
->r_dentry_drop
, req
->r_dentry_unless
);
1933 if (req
->r_old_dentry_drop
)
1934 releases
+= ceph_encode_dentry_release(&p
, req
->r_old_dentry
,
1935 mds
, req
->r_old_dentry_drop
, req
->r_old_dentry_unless
);
1936 if (req
->r_old_inode_drop
)
1937 releases
+= ceph_encode_inode_release(&p
,
1938 d_inode(req
->r_old_dentry
),
1939 mds
, req
->r_old_inode_drop
, req
->r_old_inode_unless
, 0);
1941 if (drop_cap_releases
) {
1943 p
= msg
->front
.iov_base
+ req
->r_request_release_offset
;
1946 head
->num_releases
= cpu_to_le16(releases
);
1950 struct ceph_timespec ts
;
1951 ceph_encode_timespec(&ts
, &req
->r_stamp
);
1952 ceph_encode_copy(&p
, &ts
, sizeof(ts
));
1956 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
1957 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1959 if (req
->r_pagelist
) {
1960 struct ceph_pagelist
*pagelist
= req
->r_pagelist
;
1961 atomic_inc(&pagelist
->refcnt
);
1962 ceph_msg_data_add_pagelist(msg
, pagelist
);
1963 msg
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
1965 msg
->hdr
.data_len
= 0;
1968 msg
->hdr
.data_off
= cpu_to_le16(0);
1972 kfree((char *)path2
);
1975 kfree((char *)path1
);
1981 * called under mdsc->mutex if error, under no mutex if
1984 static void complete_request(struct ceph_mds_client
*mdsc
,
1985 struct ceph_mds_request
*req
)
1987 if (req
->r_callback
)
1988 req
->r_callback(mdsc
, req
);
1990 complete_all(&req
->r_completion
);
1994 * called under mdsc->mutex
1996 static int __prepare_send_request(struct ceph_mds_client
*mdsc
,
1997 struct ceph_mds_request
*req
,
1998 int mds
, bool drop_cap_releases
)
2000 struct ceph_mds_request_head
*rhead
;
2001 struct ceph_msg
*msg
;
2006 struct ceph_cap
*cap
=
2007 ceph_get_cap_for_mds(ceph_inode(req
->r_inode
), mds
);
2010 req
->r_sent_on_mseq
= cap
->mseq
;
2012 req
->r_sent_on_mseq
= -1;
2014 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req
,
2015 req
->r_tid
, ceph_mds_op_name(req
->r_op
), req
->r_attempts
);
2017 if (req
->r_got_unsafe
) {
2020 * Replay. Do not regenerate message (and rebuild
2021 * paths, etc.); just use the original message.
2022 * Rebuilding paths will break for renames because
2023 * d_move mangles the src name.
2025 msg
= req
->r_request
;
2026 rhead
= msg
->front
.iov_base
;
2028 flags
= le32_to_cpu(rhead
->flags
);
2029 flags
|= CEPH_MDS_FLAG_REPLAY
;
2030 rhead
->flags
= cpu_to_le32(flags
);
2032 if (req
->r_target_inode
)
2033 rhead
->ino
= cpu_to_le64(ceph_ino(req
->r_target_inode
));
2035 rhead
->num_retry
= req
->r_attempts
- 1;
2037 /* remove cap/dentry releases from message */
2038 rhead
->num_releases
= 0;
2041 p
= msg
->front
.iov_base
+ req
->r_request_release_offset
;
2043 struct ceph_timespec ts
;
2044 ceph_encode_timespec(&ts
, &req
->r_stamp
);
2045 ceph_encode_copy(&p
, &ts
, sizeof(ts
));
2048 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
2049 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
2053 if (req
->r_request
) {
2054 ceph_msg_put(req
->r_request
);
2055 req
->r_request
= NULL
;
2057 msg
= create_request_message(mdsc
, req
, mds
, drop_cap_releases
);
2059 req
->r_err
= PTR_ERR(msg
);
2060 return PTR_ERR(msg
);
2062 req
->r_request
= msg
;
2064 rhead
= msg
->front
.iov_base
;
2065 rhead
->oldest_client_tid
= cpu_to_le64(__get_oldest_tid(mdsc
));
2066 if (req
->r_got_unsafe
)
2067 flags
|= CEPH_MDS_FLAG_REPLAY
;
2068 if (req
->r_locked_dir
)
2069 flags
|= CEPH_MDS_FLAG_WANT_DENTRY
;
2070 rhead
->flags
= cpu_to_le32(flags
);
2071 rhead
->num_fwd
= req
->r_num_fwd
;
2072 rhead
->num_retry
= req
->r_attempts
- 1;
2075 dout(" r_locked_dir = %p\n", req
->r_locked_dir
);
2080 * send request, or put it on the appropriate wait list.
2082 static int __do_request(struct ceph_mds_client
*mdsc
,
2083 struct ceph_mds_request
*req
)
2085 struct ceph_mds_session
*session
= NULL
;
2089 if (req
->r_err
|| req
->r_got_result
) {
2091 __unregister_request(mdsc
, req
);
2095 if (req
->r_timeout
&&
2096 time_after_eq(jiffies
, req
->r_started
+ req
->r_timeout
)) {
2097 dout("do_request timed out\n");
2101 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
) {
2102 dout("do_request forced umount\n");
2106 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_MOUNTING
) {
2107 if (mdsc
->mdsmap_err
) {
2108 err
= mdsc
->mdsmap_err
;
2109 dout("do_request mdsmap err %d\n", err
);
2112 if (mdsc
->mdsmap
->m_epoch
== 0) {
2113 dout("do_request no mdsmap, waiting for map\n");
2114 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
2117 if (!(mdsc
->fsc
->mount_options
->flags
&
2118 CEPH_MOUNT_OPT_MOUNTWAIT
) &&
2119 !ceph_mdsmap_is_cluster_available(mdsc
->mdsmap
)) {
2121 pr_info("probably no mds server is up\n");
2126 put_request_session(req
);
2128 mds
= __choose_mds(mdsc
, req
);
2130 ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) < CEPH_MDS_STATE_ACTIVE
) {
2131 dout("do_request no mds or not active, waiting for map\n");
2132 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
2136 /* get, open session */
2137 session
= __ceph_lookup_mds_session(mdsc
, mds
);
2139 session
= register_session(mdsc
, mds
);
2140 if (IS_ERR(session
)) {
2141 err
= PTR_ERR(session
);
2145 req
->r_session
= get_session(session
);
2147 dout("do_request mds%d session %p state %s\n", mds
, session
,
2148 ceph_session_state_name(session
->s_state
));
2149 if (session
->s_state
!= CEPH_MDS_SESSION_OPEN
&&
2150 session
->s_state
!= CEPH_MDS_SESSION_HUNG
) {
2151 if (session
->s_state
== CEPH_MDS_SESSION_REJECTED
) {
2155 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
2156 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
2157 __open_session(mdsc
, session
);
2158 list_add(&req
->r_wait
, &session
->s_waiting
);
2163 req
->r_resend_mds
= -1; /* forget any previous mds hint */
2165 if (req
->r_request_started
== 0) /* note request start time */
2166 req
->r_request_started
= jiffies
;
2168 err
= __prepare_send_request(mdsc
, req
, mds
, false);
2170 ceph_msg_get(req
->r_request
);
2171 ceph_con_send(&session
->s_con
, req
->r_request
);
2175 ceph_put_mds_session(session
);
2178 dout("__do_request early error %d\n", err
);
2180 complete_request(mdsc
, req
);
2181 __unregister_request(mdsc
, req
);
2188 * called under mdsc->mutex
2190 static void __wake_requests(struct ceph_mds_client
*mdsc
,
2191 struct list_head
*head
)
2193 struct ceph_mds_request
*req
;
2194 LIST_HEAD(tmp_list
);
2196 list_splice_init(head
, &tmp_list
);
2198 while (!list_empty(&tmp_list
)) {
2199 req
= list_entry(tmp_list
.next
,
2200 struct ceph_mds_request
, r_wait
);
2201 list_del_init(&req
->r_wait
);
2202 dout(" wake request %p tid %llu\n", req
, req
->r_tid
);
2203 __do_request(mdsc
, req
);
2208 * Wake up threads with requests pending for @mds, so that they can
2209 * resubmit their requests to a possibly different mds.
2211 static void kick_requests(struct ceph_mds_client
*mdsc
, int mds
)
2213 struct ceph_mds_request
*req
;
2214 struct rb_node
*p
= rb_first(&mdsc
->request_tree
);
2216 dout("kick_requests mds%d\n", mds
);
2218 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2220 if (req
->r_got_unsafe
)
2222 if (req
->r_attempts
> 0)
2223 continue; /* only new requests */
2224 if (req
->r_session
&&
2225 req
->r_session
->s_mds
== mds
) {
2226 dout(" kicking tid %llu\n", req
->r_tid
);
2227 list_del_init(&req
->r_wait
);
2228 __do_request(mdsc
, req
);
2233 void ceph_mdsc_submit_request(struct ceph_mds_client
*mdsc
,
2234 struct ceph_mds_request
*req
)
2236 dout("submit_request on %p\n", req
);
2237 mutex_lock(&mdsc
->mutex
);
2238 __register_request(mdsc
, req
, NULL
);
2239 __do_request(mdsc
, req
);
2240 mutex_unlock(&mdsc
->mutex
);
2244 * Synchrously perform an mds request. Take care of all of the
2245 * session setup, forwarding, retry details.
2247 int ceph_mdsc_do_request(struct ceph_mds_client
*mdsc
,
2249 struct ceph_mds_request
*req
)
2253 dout("do_request on %p\n", req
);
2255 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2257 ceph_get_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
2258 if (req
->r_locked_dir
)
2259 ceph_get_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
2260 if (req
->r_old_dentry_dir
)
2261 ceph_get_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
2265 mutex_lock(&mdsc
->mutex
);
2266 __register_request(mdsc
, req
, dir
);
2267 __do_request(mdsc
, req
);
2275 mutex_unlock(&mdsc
->mutex
);
2276 dout("do_request waiting\n");
2277 if (!req
->r_timeout
&& req
->r_wait_for_completion
) {
2278 err
= req
->r_wait_for_completion(mdsc
, req
);
2280 long timeleft
= wait_for_completion_killable_timeout(
2282 ceph_timeout_jiffies(req
->r_timeout
));
2286 err
= -EIO
; /* timed out */
2288 err
= timeleft
; /* killed */
2290 dout("do_request waited, got %d\n", err
);
2291 mutex_lock(&mdsc
->mutex
);
2293 /* only abort if we didn't race with a real reply */
2294 if (req
->r_got_result
) {
2295 err
= le32_to_cpu(req
->r_reply_info
.head
->result
);
2296 } else if (err
< 0) {
2297 dout("aborted request %lld with %d\n", req
->r_tid
, err
);
2300 * ensure we aren't running concurrently with
2301 * ceph_fill_trace or ceph_readdir_prepopulate, which
2302 * rely on locks (dir mutex) held by our caller.
2304 mutex_lock(&req
->r_fill_mutex
);
2306 req
->r_aborted
= true;
2307 mutex_unlock(&req
->r_fill_mutex
);
2309 if (req
->r_locked_dir
&&
2310 (req
->r_op
& CEPH_MDS_OP_WRITE
))
2311 ceph_invalidate_dir_request(req
);
2317 mutex_unlock(&mdsc
->mutex
);
2318 dout("do_request %p done, result %d\n", req
, err
);
2323 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2324 * namespace request.
2326 void ceph_invalidate_dir_request(struct ceph_mds_request
*req
)
2328 struct inode
*inode
= req
->r_locked_dir
;
2330 dout("invalidate_dir_request %p (complete, lease(s))\n", inode
);
2332 ceph_dir_clear_complete(inode
);
2334 ceph_invalidate_dentry_lease(req
->r_dentry
);
2335 if (req
->r_old_dentry
)
2336 ceph_invalidate_dentry_lease(req
->r_old_dentry
);
2342 * We take the session mutex and parse and process the reply immediately.
2343 * This preserves the logical ordering of replies, capabilities, etc., sent
2344 * by the MDS as they are applied to our local cache.
2346 static void handle_reply(struct ceph_mds_session
*session
, struct ceph_msg
*msg
)
2348 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2349 struct ceph_mds_request
*req
;
2350 struct ceph_mds_reply_head
*head
= msg
->front
.iov_base
;
2351 struct ceph_mds_reply_info_parsed
*rinfo
; /* parsed reply info */
2352 struct ceph_snap_realm
*realm
;
2355 int mds
= session
->s_mds
;
2357 if (msg
->front
.iov_len
< sizeof(*head
)) {
2358 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2363 /* get request, session */
2364 tid
= le64_to_cpu(msg
->hdr
.tid
);
2365 mutex_lock(&mdsc
->mutex
);
2366 req
= lookup_get_request(mdsc
, tid
);
2368 dout("handle_reply on unknown tid %llu\n", tid
);
2369 mutex_unlock(&mdsc
->mutex
);
2372 dout("handle_reply %p\n", req
);
2374 /* correct session? */
2375 if (req
->r_session
!= session
) {
2376 pr_err("mdsc_handle_reply got %llu on session mds%d"
2377 " not mds%d\n", tid
, session
->s_mds
,
2378 req
->r_session
? req
->r_session
->s_mds
: -1);
2379 mutex_unlock(&mdsc
->mutex
);
2384 if ((req
->r_got_unsafe
&& !head
->safe
) ||
2385 (req
->r_got_safe
&& head
->safe
)) {
2386 pr_warn("got a dup %s reply on %llu from mds%d\n",
2387 head
->safe
? "safe" : "unsafe", tid
, mds
);
2388 mutex_unlock(&mdsc
->mutex
);
2391 if (req
->r_got_safe
) {
2392 pr_warn("got unsafe after safe on %llu from mds%d\n",
2394 mutex_unlock(&mdsc
->mutex
);
2398 result
= le32_to_cpu(head
->result
);
2402 * if we're not talking to the authority, send to them
2403 * if the authority has changed while we weren't looking,
2404 * send to new authority
2405 * Otherwise we just have to return an ESTALE
2407 if (result
== -ESTALE
) {
2408 dout("got ESTALE on request %llu", req
->r_tid
);
2409 req
->r_resend_mds
= -1;
2410 if (req
->r_direct_mode
!= USE_AUTH_MDS
) {
2411 dout("not using auth, setting for that now");
2412 req
->r_direct_mode
= USE_AUTH_MDS
;
2413 __do_request(mdsc
, req
);
2414 mutex_unlock(&mdsc
->mutex
);
2417 int mds
= __choose_mds(mdsc
, req
);
2418 if (mds
>= 0 && mds
!= req
->r_session
->s_mds
) {
2419 dout("but auth changed, so resending");
2420 __do_request(mdsc
, req
);
2421 mutex_unlock(&mdsc
->mutex
);
2425 dout("have to return ESTALE on request %llu", req
->r_tid
);
2430 req
->r_got_safe
= true;
2431 __unregister_request(mdsc
, req
);
2433 if (req
->r_got_unsafe
) {
2435 * We already handled the unsafe response, now do the
2436 * cleanup. No need to examine the response; the MDS
2437 * doesn't include any result info in the safe
2438 * response. And even if it did, there is nothing
2439 * useful we could do with a revised return value.
2441 dout("got safe reply %llu, mds%d\n", tid
, mds
);
2443 /* last unsafe request during umount? */
2444 if (mdsc
->stopping
&& !__get_oldest_req(mdsc
))
2445 complete_all(&mdsc
->safe_umount_waiters
);
2446 mutex_unlock(&mdsc
->mutex
);
2450 req
->r_got_unsafe
= true;
2451 list_add_tail(&req
->r_unsafe_item
, &req
->r_session
->s_unsafe
);
2452 if (req
->r_unsafe_dir
) {
2453 struct ceph_inode_info
*ci
=
2454 ceph_inode(req
->r_unsafe_dir
);
2455 spin_lock(&ci
->i_unsafe_lock
);
2456 list_add_tail(&req
->r_unsafe_dir_item
,
2457 &ci
->i_unsafe_dirops
);
2458 spin_unlock(&ci
->i_unsafe_lock
);
2462 dout("handle_reply tid %lld result %d\n", tid
, result
);
2463 rinfo
= &req
->r_reply_info
;
2464 err
= parse_reply_info(msg
, rinfo
, session
->s_con
.peer_features
);
2465 mutex_unlock(&mdsc
->mutex
);
2467 mutex_lock(&session
->s_mutex
);
2469 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds
, tid
);
2476 if (rinfo
->snapblob_len
) {
2477 down_write(&mdsc
->snap_rwsem
);
2478 ceph_update_snap_trace(mdsc
, rinfo
->snapblob
,
2479 rinfo
->snapblob
+ rinfo
->snapblob_len
,
2480 le32_to_cpu(head
->op
) == CEPH_MDS_OP_RMSNAP
,
2482 downgrade_write(&mdsc
->snap_rwsem
);
2484 down_read(&mdsc
->snap_rwsem
);
2487 /* insert trace into our cache */
2488 mutex_lock(&req
->r_fill_mutex
);
2489 current
->journal_info
= req
;
2490 err
= ceph_fill_trace(mdsc
->fsc
->sb
, req
, req
->r_session
);
2492 if (result
== 0 && (req
->r_op
== CEPH_MDS_OP_READDIR
||
2493 req
->r_op
== CEPH_MDS_OP_LSSNAP
))
2494 ceph_readdir_prepopulate(req
, req
->r_session
);
2495 ceph_unreserve_caps(mdsc
, &req
->r_caps_reservation
);
2497 current
->journal_info
= NULL
;
2498 mutex_unlock(&req
->r_fill_mutex
);
2500 up_read(&mdsc
->snap_rwsem
);
2502 ceph_put_snap_realm(mdsc
, realm
);
2504 if (err
== 0 && req
->r_got_unsafe
&& req
->r_target_inode
) {
2505 struct ceph_inode_info
*ci
= ceph_inode(req
->r_target_inode
);
2506 spin_lock(&ci
->i_unsafe_lock
);
2507 list_add_tail(&req
->r_unsafe_target_item
, &ci
->i_unsafe_iops
);
2508 spin_unlock(&ci
->i_unsafe_lock
);
2511 mutex_lock(&mdsc
->mutex
);
2512 if (!req
->r_aborted
) {
2516 req
->r_reply
= ceph_msg_get(msg
);
2517 req
->r_got_result
= true;
2520 dout("reply arrived after request %lld was aborted\n", tid
);
2522 mutex_unlock(&mdsc
->mutex
);
2524 mutex_unlock(&session
->s_mutex
);
2526 /* kick calling process */
2527 complete_request(mdsc
, req
);
2529 ceph_mdsc_put_request(req
);
2536 * handle mds notification that our request has been forwarded.
2538 static void handle_forward(struct ceph_mds_client
*mdsc
,
2539 struct ceph_mds_session
*session
,
2540 struct ceph_msg
*msg
)
2542 struct ceph_mds_request
*req
;
2543 u64 tid
= le64_to_cpu(msg
->hdr
.tid
);
2547 void *p
= msg
->front
.iov_base
;
2548 void *end
= p
+ msg
->front
.iov_len
;
2550 ceph_decode_need(&p
, end
, 2*sizeof(u32
), bad
);
2551 next_mds
= ceph_decode_32(&p
);
2552 fwd_seq
= ceph_decode_32(&p
);
2554 mutex_lock(&mdsc
->mutex
);
2555 req
= lookup_get_request(mdsc
, tid
);
2557 dout("forward tid %llu to mds%d - req dne\n", tid
, next_mds
);
2558 goto out
; /* dup reply? */
2561 if (req
->r_aborted
) {
2562 dout("forward tid %llu aborted, unregistering\n", tid
);
2563 __unregister_request(mdsc
, req
);
2564 } else if (fwd_seq
<= req
->r_num_fwd
) {
2565 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2566 tid
, next_mds
, req
->r_num_fwd
, fwd_seq
);
2568 /* resend. forward race not possible; mds would drop */
2569 dout("forward tid %llu to mds%d (we resend)\n", tid
, next_mds
);
2571 BUG_ON(req
->r_got_result
);
2572 req
->r_attempts
= 0;
2573 req
->r_num_fwd
= fwd_seq
;
2574 req
->r_resend_mds
= next_mds
;
2575 put_request_session(req
);
2576 __do_request(mdsc
, req
);
2578 ceph_mdsc_put_request(req
);
2580 mutex_unlock(&mdsc
->mutex
);
2584 pr_err("mdsc_handle_forward decode error err=%d\n", err
);
2588 * handle a mds session control message
2590 static void handle_session(struct ceph_mds_session
*session
,
2591 struct ceph_msg
*msg
)
2593 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2596 int mds
= session
->s_mds
;
2597 struct ceph_mds_session_head
*h
= msg
->front
.iov_base
;
2601 if (msg
->front
.iov_len
!= sizeof(*h
))
2603 op
= le32_to_cpu(h
->op
);
2604 seq
= le64_to_cpu(h
->seq
);
2606 mutex_lock(&mdsc
->mutex
);
2607 if (op
== CEPH_SESSION_CLOSE
)
2608 __unregister_session(mdsc
, session
);
2609 /* FIXME: this ttl calculation is generous */
2610 session
->s_ttl
= jiffies
+ HZ
*mdsc
->mdsmap
->m_session_autoclose
;
2611 mutex_unlock(&mdsc
->mutex
);
2613 mutex_lock(&session
->s_mutex
);
2615 dout("handle_session mds%d %s %p state %s seq %llu\n",
2616 mds
, ceph_session_op_name(op
), session
,
2617 ceph_session_state_name(session
->s_state
), seq
);
2619 if (session
->s_state
== CEPH_MDS_SESSION_HUNG
) {
2620 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2621 pr_info("mds%d came back\n", session
->s_mds
);
2625 case CEPH_SESSION_OPEN
:
2626 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2627 pr_info("mds%d reconnect success\n", session
->s_mds
);
2628 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2629 renewed_caps(mdsc
, session
, 0);
2632 __close_session(mdsc
, session
);
2635 case CEPH_SESSION_RENEWCAPS
:
2636 if (session
->s_renew_seq
== seq
)
2637 renewed_caps(mdsc
, session
, 1);
2640 case CEPH_SESSION_CLOSE
:
2641 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2642 pr_info("mds%d reconnect denied\n", session
->s_mds
);
2643 cleanup_session_requests(mdsc
, session
);
2644 remove_session_caps(session
);
2645 wake
= 2; /* for good measure */
2646 wake_up_all(&mdsc
->session_close_wq
);
2649 case CEPH_SESSION_STALE
:
2650 pr_info("mds%d caps went stale, renewing\n",
2652 spin_lock(&session
->s_gen_ttl_lock
);
2653 session
->s_cap_gen
++;
2654 session
->s_cap_ttl
= jiffies
- 1;
2655 spin_unlock(&session
->s_gen_ttl_lock
);
2656 send_renew_caps(mdsc
, session
);
2659 case CEPH_SESSION_RECALL_STATE
:
2660 trim_caps(mdsc
, session
, le32_to_cpu(h
->max_caps
));
2663 case CEPH_SESSION_FLUSHMSG
:
2664 send_flushmsg_ack(mdsc
, session
, seq
);
2667 case CEPH_SESSION_FORCE_RO
:
2668 dout("force_session_readonly %p\n", session
);
2669 spin_lock(&session
->s_cap_lock
);
2670 session
->s_readonly
= true;
2671 spin_unlock(&session
->s_cap_lock
);
2672 wake_up_session_caps(session
, 0);
2675 case CEPH_SESSION_REJECT
:
2676 WARN_ON(session
->s_state
!= CEPH_MDS_SESSION_OPENING
);
2677 pr_info("mds%d rejected session\n", session
->s_mds
);
2678 session
->s_state
= CEPH_MDS_SESSION_REJECTED
;
2679 cleanup_session_requests(mdsc
, session
);
2680 remove_session_caps(session
);
2681 wake
= 2; /* for good measure */
2685 pr_err("mdsc_handle_session bad op %d mds%d\n", op
, mds
);
2689 mutex_unlock(&session
->s_mutex
);
2691 mutex_lock(&mdsc
->mutex
);
2692 __wake_requests(mdsc
, &session
->s_waiting
);
2694 kick_requests(mdsc
, mds
);
2695 mutex_unlock(&mdsc
->mutex
);
2700 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds
,
2701 (int)msg
->front
.iov_len
);
2708 * called under session->mutex.
2710 static void replay_unsafe_requests(struct ceph_mds_client
*mdsc
,
2711 struct ceph_mds_session
*session
)
2713 struct ceph_mds_request
*req
, *nreq
;
2717 dout("replay_unsafe_requests mds%d\n", session
->s_mds
);
2719 mutex_lock(&mdsc
->mutex
);
2720 list_for_each_entry_safe(req
, nreq
, &session
->s_unsafe
, r_unsafe_item
) {
2721 err
= __prepare_send_request(mdsc
, req
, session
->s_mds
, true);
2723 ceph_msg_get(req
->r_request
);
2724 ceph_con_send(&session
->s_con
, req
->r_request
);
2729 * also re-send old requests when MDS enters reconnect stage. So that MDS
2730 * can process completed request in clientreplay stage.
2732 p
= rb_first(&mdsc
->request_tree
);
2734 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2736 if (req
->r_got_unsafe
)
2738 if (req
->r_attempts
== 0)
2739 continue; /* only old requests */
2740 if (req
->r_session
&&
2741 req
->r_session
->s_mds
== session
->s_mds
) {
2742 err
= __prepare_send_request(mdsc
, req
,
2743 session
->s_mds
, true);
2745 ceph_msg_get(req
->r_request
);
2746 ceph_con_send(&session
->s_con
, req
->r_request
);
2750 mutex_unlock(&mdsc
->mutex
);
2754 * Encode information about a cap for a reconnect with the MDS.
2756 static int encode_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
2760 struct ceph_mds_cap_reconnect v2
;
2761 struct ceph_mds_cap_reconnect_v1 v1
;
2763 struct ceph_inode_info
*ci
;
2764 struct ceph_reconnect_state
*recon_state
= arg
;
2765 struct ceph_pagelist
*pagelist
= recon_state
->pagelist
;
2770 struct dentry
*dentry
;
2774 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2775 inode
, ceph_vinop(inode
), cap
, cap
->cap_id
,
2776 ceph_cap_string(cap
->issued
));
2777 err
= ceph_pagelist_encode_64(pagelist
, ceph_ino(inode
));
2781 dentry
= d_find_alias(inode
);
2783 path
= ceph_mdsc_build_path(dentry
, &pathlen
, &pathbase
, 0);
2785 err
= PTR_ERR(path
);
2794 spin_lock(&ci
->i_ceph_lock
);
2795 cap
->seq
= 0; /* reset cap seq */
2796 cap
->issue_seq
= 0; /* and issue_seq */
2797 cap
->mseq
= 0; /* and migrate_seq */
2798 cap
->cap_gen
= cap
->session
->s_cap_gen
;
2800 if (recon_state
->msg_version
>= 2) {
2801 rec
.v2
.cap_id
= cpu_to_le64(cap
->cap_id
);
2802 rec
.v2
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2803 rec
.v2
.issued
= cpu_to_le32(cap
->issued
);
2804 rec
.v2
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2805 rec
.v2
.pathbase
= cpu_to_le64(pathbase
);
2806 rec
.v2
.flock_len
= 0;
2808 rec
.v1
.cap_id
= cpu_to_le64(cap
->cap_id
);
2809 rec
.v1
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2810 rec
.v1
.issued
= cpu_to_le32(cap
->issued
);
2811 rec
.v1
.size
= cpu_to_le64(inode
->i_size
);
2812 ceph_encode_timespec(&rec
.v1
.mtime
, &inode
->i_mtime
);
2813 ceph_encode_timespec(&rec
.v1
.atime
, &inode
->i_atime
);
2814 rec
.v1
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2815 rec
.v1
.pathbase
= cpu_to_le64(pathbase
);
2818 if (list_empty(&ci
->i_cap_snaps
)) {
2821 struct ceph_cap_snap
*capsnap
=
2822 list_first_entry(&ci
->i_cap_snaps
,
2823 struct ceph_cap_snap
, ci_item
);
2824 snap_follows
= capsnap
->follows
;
2826 spin_unlock(&ci
->i_ceph_lock
);
2828 if (recon_state
->msg_version
>= 2) {
2829 int num_fcntl_locks
, num_flock_locks
;
2830 struct ceph_filelock
*flocks
;
2831 size_t struct_len
, total_len
= 0;
2835 ceph_count_locks(inode
, &num_fcntl_locks
, &num_flock_locks
);
2836 flocks
= kmalloc((num_fcntl_locks
+num_flock_locks
) *
2837 sizeof(struct ceph_filelock
), GFP_NOFS
);
2842 err
= ceph_encode_locks_to_buffer(inode
, flocks
,
2852 if (recon_state
->msg_version
>= 3) {
2853 /* version, compat_version and struct_len */
2854 total_len
= 2 * sizeof(u8
) + sizeof(u32
);
2858 * number of encoded locks is stable, so copy to pagelist
2860 struct_len
= 2 * sizeof(u32
) +
2861 (num_fcntl_locks
+ num_flock_locks
) *
2862 sizeof(struct ceph_filelock
);
2863 rec
.v2
.flock_len
= cpu_to_le32(struct_len
);
2865 struct_len
+= sizeof(rec
.v2
);
2866 struct_len
+= sizeof(u32
) + pathlen
;
2869 struct_len
+= sizeof(u64
); /* snap_follows */
2871 total_len
+= struct_len
;
2872 err
= ceph_pagelist_reserve(pagelist
, total_len
);
2875 if (recon_state
->msg_version
>= 3) {
2876 ceph_pagelist_encode_8(pagelist
, struct_v
);
2877 ceph_pagelist_encode_8(pagelist
, 1);
2878 ceph_pagelist_encode_32(pagelist
, struct_len
);
2880 ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
2881 ceph_pagelist_append(pagelist
, &rec
, sizeof(rec
.v2
));
2882 ceph_locks_to_pagelist(flocks
, pagelist
,
2886 ceph_pagelist_encode_64(pagelist
, snap_follows
);
2890 size_t size
= sizeof(u32
) + pathlen
+ sizeof(rec
.v1
);
2891 err
= ceph_pagelist_reserve(pagelist
, size
);
2893 ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
2894 ceph_pagelist_append(pagelist
, &rec
, sizeof(rec
.v1
));
2898 recon_state
->nr_caps
++;
2908 * If an MDS fails and recovers, clients need to reconnect in order to
2909 * reestablish shared state. This includes all caps issued through
2910 * this session _and_ the snap_realm hierarchy. Because it's not
2911 * clear which snap realms the mds cares about, we send everything we
2912 * know about.. that ensures we'll then get any new info the
2913 * recovering MDS might have.
2915 * This is a relatively heavyweight operation, but it's rare.
2917 * called with mdsc->mutex held.
2919 static void send_mds_reconnect(struct ceph_mds_client
*mdsc
,
2920 struct ceph_mds_session
*session
)
2922 struct ceph_msg
*reply
;
2924 int mds
= session
->s_mds
;
2927 struct ceph_pagelist
*pagelist
;
2928 struct ceph_reconnect_state recon_state
;
2930 pr_info("mds%d reconnect start\n", mds
);
2932 pagelist
= kmalloc(sizeof(*pagelist
), GFP_NOFS
);
2934 goto fail_nopagelist
;
2935 ceph_pagelist_init(pagelist
);
2937 reply
= ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT
, 0, GFP_NOFS
, false);
2941 mutex_lock(&session
->s_mutex
);
2942 session
->s_state
= CEPH_MDS_SESSION_RECONNECTING
;
2945 dout("session %p state %s\n", session
,
2946 ceph_session_state_name(session
->s_state
));
2948 spin_lock(&session
->s_gen_ttl_lock
);
2949 session
->s_cap_gen
++;
2950 spin_unlock(&session
->s_gen_ttl_lock
);
2952 spin_lock(&session
->s_cap_lock
);
2953 /* don't know if session is readonly */
2954 session
->s_readonly
= 0;
2956 * notify __ceph_remove_cap() that we are composing cap reconnect.
2957 * If a cap get released before being added to the cap reconnect,
2958 * __ceph_remove_cap() should skip queuing cap release.
2960 session
->s_cap_reconnect
= 1;
2961 /* drop old cap expires; we're about to reestablish that state */
2962 cleanup_cap_releases(mdsc
, session
);
2964 /* trim unused caps to reduce MDS's cache rejoin time */
2965 if (mdsc
->fsc
->sb
->s_root
)
2966 shrink_dcache_parent(mdsc
->fsc
->sb
->s_root
);
2968 ceph_con_close(&session
->s_con
);
2969 ceph_con_open(&session
->s_con
,
2970 CEPH_ENTITY_TYPE_MDS
, mds
,
2971 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
2973 /* replay unsafe requests */
2974 replay_unsafe_requests(mdsc
, session
);
2976 down_read(&mdsc
->snap_rwsem
);
2978 /* traverse this session's caps */
2979 s_nr_caps
= session
->s_nr_caps
;
2980 err
= ceph_pagelist_encode_32(pagelist
, s_nr_caps
);
2984 recon_state
.nr_caps
= 0;
2985 recon_state
.pagelist
= pagelist
;
2986 if (session
->s_con
.peer_features
& CEPH_FEATURE_MDSENC
)
2987 recon_state
.msg_version
= 3;
2988 else if (session
->s_con
.peer_features
& CEPH_FEATURE_FLOCK
)
2989 recon_state
.msg_version
= 2;
2991 recon_state
.msg_version
= 1;
2992 err
= iterate_session_caps(session
, encode_caps_cb
, &recon_state
);
2996 spin_lock(&session
->s_cap_lock
);
2997 session
->s_cap_reconnect
= 0;
2998 spin_unlock(&session
->s_cap_lock
);
3001 * snaprealms. we provide mds with the ino, seq (version), and
3002 * parent for all of our realms. If the mds has any newer info,
3005 for (p
= rb_first(&mdsc
->snap_realms
); p
; p
= rb_next(p
)) {
3006 struct ceph_snap_realm
*realm
=
3007 rb_entry(p
, struct ceph_snap_realm
, node
);
3008 struct ceph_mds_snaprealm_reconnect sr_rec
;
3010 dout(" adding snap realm %llx seq %lld parent %llx\n",
3011 realm
->ino
, realm
->seq
, realm
->parent_ino
);
3012 sr_rec
.ino
= cpu_to_le64(realm
->ino
);
3013 sr_rec
.seq
= cpu_to_le64(realm
->seq
);
3014 sr_rec
.parent
= cpu_to_le64(realm
->parent_ino
);
3015 err
= ceph_pagelist_append(pagelist
, &sr_rec
, sizeof(sr_rec
));
3020 reply
->hdr
.version
= cpu_to_le16(recon_state
.msg_version
);
3022 /* raced with cap release? */
3023 if (s_nr_caps
!= recon_state
.nr_caps
) {
3024 struct page
*page
= list_first_entry(&pagelist
->head
,
3026 __le32
*addr
= kmap_atomic(page
);
3027 *addr
= cpu_to_le32(recon_state
.nr_caps
);
3028 kunmap_atomic(addr
);
3031 reply
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
3032 ceph_msg_data_add_pagelist(reply
, pagelist
);
3034 ceph_early_kick_flushing_caps(mdsc
, session
);
3036 ceph_con_send(&session
->s_con
, reply
);
3038 mutex_unlock(&session
->s_mutex
);
3040 mutex_lock(&mdsc
->mutex
);
3041 __wake_requests(mdsc
, &session
->s_waiting
);
3042 mutex_unlock(&mdsc
->mutex
);
3044 up_read(&mdsc
->snap_rwsem
);
3048 ceph_msg_put(reply
);
3049 up_read(&mdsc
->snap_rwsem
);
3050 mutex_unlock(&session
->s_mutex
);
3052 ceph_pagelist_release(pagelist
);
3054 pr_err("error %d preparing reconnect for mds%d\n", err
, mds
);
3060 * compare old and new mdsmaps, kicking requests
3061 * and closing out old connections as necessary
3063 * called under mdsc->mutex.
3065 static void check_new_map(struct ceph_mds_client
*mdsc
,
3066 struct ceph_mdsmap
*newmap
,
3067 struct ceph_mdsmap
*oldmap
)
3070 int oldstate
, newstate
;
3071 struct ceph_mds_session
*s
;
3073 dout("check_new_map new %u old %u\n",
3074 newmap
->m_epoch
, oldmap
->m_epoch
);
3076 for (i
= 0; i
< oldmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
3077 if (mdsc
->sessions
[i
] == NULL
)
3079 s
= mdsc
->sessions
[i
];
3080 oldstate
= ceph_mdsmap_get_state(oldmap
, i
);
3081 newstate
= ceph_mdsmap_get_state(newmap
, i
);
3083 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3084 i
, ceph_mds_state_name(oldstate
),
3085 ceph_mdsmap_is_laggy(oldmap
, i
) ? " (laggy)" : "",
3086 ceph_mds_state_name(newstate
),
3087 ceph_mdsmap_is_laggy(newmap
, i
) ? " (laggy)" : "",
3088 ceph_session_state_name(s
->s_state
));
3090 if (i
>= newmap
->m_max_mds
||
3091 memcmp(ceph_mdsmap_get_addr(oldmap
, i
),
3092 ceph_mdsmap_get_addr(newmap
, i
),
3093 sizeof(struct ceph_entity_addr
))) {
3094 if (s
->s_state
== CEPH_MDS_SESSION_OPENING
) {
3095 /* the session never opened, just close it
3097 __wake_requests(mdsc
, &s
->s_waiting
);
3098 __unregister_session(mdsc
, s
);
3101 mutex_unlock(&mdsc
->mutex
);
3102 mutex_lock(&s
->s_mutex
);
3103 mutex_lock(&mdsc
->mutex
);
3104 ceph_con_close(&s
->s_con
);
3105 mutex_unlock(&s
->s_mutex
);
3106 s
->s_state
= CEPH_MDS_SESSION_RESTARTING
;
3108 } else if (oldstate
== newstate
) {
3109 continue; /* nothing new with this mds */
3115 if (s
->s_state
== CEPH_MDS_SESSION_RESTARTING
&&
3116 newstate
>= CEPH_MDS_STATE_RECONNECT
) {
3117 mutex_unlock(&mdsc
->mutex
);
3118 send_mds_reconnect(mdsc
, s
);
3119 mutex_lock(&mdsc
->mutex
);
3123 * kick request on any mds that has gone active.
3125 if (oldstate
< CEPH_MDS_STATE_ACTIVE
&&
3126 newstate
>= CEPH_MDS_STATE_ACTIVE
) {
3127 if (oldstate
!= CEPH_MDS_STATE_CREATING
&&
3128 oldstate
!= CEPH_MDS_STATE_STARTING
)
3129 pr_info("mds%d recovery completed\n", s
->s_mds
);
3130 kick_requests(mdsc
, i
);
3131 ceph_kick_flushing_caps(mdsc
, s
);
3132 wake_up_session_caps(s
, 1);
3136 for (i
= 0; i
< newmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
3137 s
= mdsc
->sessions
[i
];
3140 if (!ceph_mdsmap_is_laggy(newmap
, i
))
3142 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3143 s
->s_state
== CEPH_MDS_SESSION_HUNG
||
3144 s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3145 dout(" connecting to export targets of laggy mds%d\n",
3147 __open_export_target_sessions(mdsc
, s
);
3159 * caller must hold session s_mutex, dentry->d_lock
3161 void __ceph_mdsc_drop_dentry_lease(struct dentry
*dentry
)
3163 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
3165 ceph_put_mds_session(di
->lease_session
);
3166 di
->lease_session
= NULL
;
3169 static void handle_lease(struct ceph_mds_client
*mdsc
,
3170 struct ceph_mds_session
*session
,
3171 struct ceph_msg
*msg
)
3173 struct super_block
*sb
= mdsc
->fsc
->sb
;
3174 struct inode
*inode
;
3175 struct dentry
*parent
, *dentry
;
3176 struct ceph_dentry_info
*di
;
3177 int mds
= session
->s_mds
;
3178 struct ceph_mds_lease
*h
= msg
->front
.iov_base
;
3180 struct ceph_vino vino
;
3184 dout("handle_lease from mds%d\n", mds
);
3187 if (msg
->front
.iov_len
< sizeof(*h
) + sizeof(u32
))
3189 vino
.ino
= le64_to_cpu(h
->ino
);
3190 vino
.snap
= CEPH_NOSNAP
;
3191 seq
= le32_to_cpu(h
->seq
);
3192 dname
.name
= (void *)h
+ sizeof(*h
) + sizeof(u32
);
3193 dname
.len
= msg
->front
.iov_len
- sizeof(*h
) - sizeof(u32
);
3194 if (dname
.len
!= get_unaligned_le32(h
+1))
3198 inode
= ceph_find_inode(sb
, vino
);
3199 dout("handle_lease %s, ino %llx %p %.*s\n",
3200 ceph_lease_op_name(h
->action
), vino
.ino
, inode
,
3201 dname
.len
, dname
.name
);
3203 mutex_lock(&session
->s_mutex
);
3206 if (inode
== NULL
) {
3207 dout("handle_lease no inode %llx\n", vino
.ino
);
3212 parent
= d_find_alias(inode
);
3214 dout("no parent dentry on inode %p\n", inode
);
3216 goto release
; /* hrm... */
3218 dname
.hash
= full_name_hash(parent
, dname
.name
, dname
.len
);
3219 dentry
= d_lookup(parent
, &dname
);
3224 spin_lock(&dentry
->d_lock
);
3225 di
= ceph_dentry(dentry
);
3226 switch (h
->action
) {
3227 case CEPH_MDS_LEASE_REVOKE
:
3228 if (di
->lease_session
== session
) {
3229 if (ceph_seq_cmp(di
->lease_seq
, seq
) > 0)
3230 h
->seq
= cpu_to_le32(di
->lease_seq
);
3231 __ceph_mdsc_drop_dentry_lease(dentry
);
3236 case CEPH_MDS_LEASE_RENEW
:
3237 if (di
->lease_session
== session
&&
3238 di
->lease_gen
== session
->s_cap_gen
&&
3239 di
->lease_renew_from
&&
3240 di
->lease_renew_after
== 0) {
3241 unsigned long duration
=
3242 msecs_to_jiffies(le32_to_cpu(h
->duration_ms
));
3244 di
->lease_seq
= seq
;
3245 di
->time
= di
->lease_renew_from
+ duration
;
3246 di
->lease_renew_after
= di
->lease_renew_from
+
3248 di
->lease_renew_from
= 0;
3252 spin_unlock(&dentry
->d_lock
);
3259 /* let's just reuse the same message */
3260 h
->action
= CEPH_MDS_LEASE_REVOKE_ACK
;
3262 ceph_con_send(&session
->s_con
, msg
);
3266 mutex_unlock(&session
->s_mutex
);
3270 pr_err("corrupt lease message\n");
3274 void ceph_mdsc_lease_send_msg(struct ceph_mds_session
*session
,
3275 struct inode
*inode
,
3276 struct dentry
*dentry
, char action
,
3279 struct ceph_msg
*msg
;
3280 struct ceph_mds_lease
*lease
;
3281 int len
= sizeof(*lease
) + sizeof(u32
);
3284 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3285 inode
, dentry
, ceph_lease_op_name(action
), session
->s_mds
);
3286 dnamelen
= dentry
->d_name
.len
;
3289 msg
= ceph_msg_new(CEPH_MSG_CLIENT_LEASE
, len
, GFP_NOFS
, false);
3292 lease
= msg
->front
.iov_base
;
3293 lease
->action
= action
;
3294 lease
->ino
= cpu_to_le64(ceph_vino(inode
).ino
);
3295 lease
->first
= lease
->last
= cpu_to_le64(ceph_vino(inode
).snap
);
3296 lease
->seq
= cpu_to_le32(seq
);
3297 put_unaligned_le32(dnamelen
, lease
+ 1);
3298 memcpy((void *)(lease
+ 1) + 4, dentry
->d_name
.name
, dnamelen
);
3301 * if this is a preemptive lease RELEASE, no need to
3302 * flush request stream, since the actual request will
3305 msg
->more_to_follow
= (action
== CEPH_MDS_LEASE_RELEASE
);
3307 ceph_con_send(&session
->s_con
, msg
);
3311 * drop all leases (and dentry refs) in preparation for umount
3313 static void drop_leases(struct ceph_mds_client
*mdsc
)
3317 dout("drop_leases\n");
3318 mutex_lock(&mdsc
->mutex
);
3319 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3320 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3323 mutex_unlock(&mdsc
->mutex
);
3324 mutex_lock(&s
->s_mutex
);
3325 mutex_unlock(&s
->s_mutex
);
3326 ceph_put_mds_session(s
);
3327 mutex_lock(&mdsc
->mutex
);
3329 mutex_unlock(&mdsc
->mutex
);
3335 * delayed work -- periodically trim expired leases, renew caps with mds
3337 static void schedule_delayed(struct ceph_mds_client
*mdsc
)
3340 unsigned hz
= round_jiffies_relative(HZ
* delay
);
3341 schedule_delayed_work(&mdsc
->delayed_work
, hz
);
3344 static void delayed_work(struct work_struct
*work
)
3347 struct ceph_mds_client
*mdsc
=
3348 container_of(work
, struct ceph_mds_client
, delayed_work
.work
);
3352 dout("mdsc delayed_work\n");
3353 ceph_check_delayed_caps(mdsc
);
3355 mutex_lock(&mdsc
->mutex
);
3356 renew_interval
= mdsc
->mdsmap
->m_session_timeout
>> 2;
3357 renew_caps
= time_after_eq(jiffies
, HZ
*renew_interval
+
3358 mdsc
->last_renew_caps
);
3360 mdsc
->last_renew_caps
= jiffies
;
3362 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3363 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3366 if (s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3367 dout("resending session close request for mds%d\n",
3369 request_close_session(mdsc
, s
);
3370 ceph_put_mds_session(s
);
3373 if (s
->s_ttl
&& time_after(jiffies
, s
->s_ttl
)) {
3374 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
) {
3375 s
->s_state
= CEPH_MDS_SESSION_HUNG
;
3376 pr_info("mds%d hung\n", s
->s_mds
);
3379 if (s
->s_state
< CEPH_MDS_SESSION_OPEN
) {
3380 /* this mds is failed or recovering, just wait */
3381 ceph_put_mds_session(s
);
3384 mutex_unlock(&mdsc
->mutex
);
3386 mutex_lock(&s
->s_mutex
);
3388 send_renew_caps(mdsc
, s
);
3390 ceph_con_keepalive(&s
->s_con
);
3391 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3392 s
->s_state
== CEPH_MDS_SESSION_HUNG
)
3393 ceph_send_cap_releases(mdsc
, s
);
3394 mutex_unlock(&s
->s_mutex
);
3395 ceph_put_mds_session(s
);
3397 mutex_lock(&mdsc
->mutex
);
3399 mutex_unlock(&mdsc
->mutex
);
3401 schedule_delayed(mdsc
);
3404 int ceph_mdsc_init(struct ceph_fs_client
*fsc
)
3407 struct ceph_mds_client
*mdsc
;
3409 mdsc
= kzalloc(sizeof(struct ceph_mds_client
), GFP_NOFS
);
3414 mutex_init(&mdsc
->mutex
);
3415 mdsc
->mdsmap
= kzalloc(sizeof(*mdsc
->mdsmap
), GFP_NOFS
);
3416 if (mdsc
->mdsmap
== NULL
) {
3421 init_completion(&mdsc
->safe_umount_waiters
);
3422 init_waitqueue_head(&mdsc
->session_close_wq
);
3423 INIT_LIST_HEAD(&mdsc
->waiting_for_map
);
3424 mdsc
->sessions
= NULL
;
3425 atomic_set(&mdsc
->num_sessions
, 0);
3426 mdsc
->max_sessions
= 0;
3428 mdsc
->last_snap_seq
= 0;
3429 init_rwsem(&mdsc
->snap_rwsem
);
3430 mdsc
->snap_realms
= RB_ROOT
;
3431 INIT_LIST_HEAD(&mdsc
->snap_empty
);
3432 spin_lock_init(&mdsc
->snap_empty_lock
);
3434 mdsc
->oldest_tid
= 0;
3435 mdsc
->request_tree
= RB_ROOT
;
3436 INIT_DELAYED_WORK(&mdsc
->delayed_work
, delayed_work
);
3437 mdsc
->last_renew_caps
= jiffies
;
3438 INIT_LIST_HEAD(&mdsc
->cap_delay_list
);
3439 spin_lock_init(&mdsc
->cap_delay_lock
);
3440 INIT_LIST_HEAD(&mdsc
->snap_flush_list
);
3441 spin_lock_init(&mdsc
->snap_flush_lock
);
3442 mdsc
->last_cap_flush_tid
= 1;
3443 INIT_LIST_HEAD(&mdsc
->cap_flush_list
);
3444 INIT_LIST_HEAD(&mdsc
->cap_dirty
);
3445 INIT_LIST_HEAD(&mdsc
->cap_dirty_migrating
);
3446 mdsc
->num_cap_flushing
= 0;
3447 spin_lock_init(&mdsc
->cap_dirty_lock
);
3448 init_waitqueue_head(&mdsc
->cap_flushing_wq
);
3449 spin_lock_init(&mdsc
->dentry_lru_lock
);
3450 INIT_LIST_HEAD(&mdsc
->dentry_lru
);
3452 ceph_caps_init(mdsc
);
3453 ceph_adjust_min_caps(mdsc
, fsc
->min_caps
);
3455 init_rwsem(&mdsc
->pool_perm_rwsem
);
3456 mdsc
->pool_perm_tree
= RB_ROOT
;
3462 * Wait for safe replies on open mds requests. If we time out, drop
3463 * all requests from the tree to avoid dangling dentry refs.
3465 static void wait_requests(struct ceph_mds_client
*mdsc
)
3467 struct ceph_options
*opts
= mdsc
->fsc
->client
->options
;
3468 struct ceph_mds_request
*req
;
3470 mutex_lock(&mdsc
->mutex
);
3471 if (__get_oldest_req(mdsc
)) {
3472 mutex_unlock(&mdsc
->mutex
);
3474 dout("wait_requests waiting for requests\n");
3475 wait_for_completion_timeout(&mdsc
->safe_umount_waiters
,
3476 ceph_timeout_jiffies(opts
->mount_timeout
));
3478 /* tear down remaining requests */
3479 mutex_lock(&mdsc
->mutex
);
3480 while ((req
= __get_oldest_req(mdsc
))) {
3481 dout("wait_requests timed out on tid %llu\n",
3483 __unregister_request(mdsc
, req
);
3486 mutex_unlock(&mdsc
->mutex
);
3487 dout("wait_requests done\n");
3491 * called before mount is ro, and before dentries are torn down.
3492 * (hmm, does this still race with new lookups?)
3494 void ceph_mdsc_pre_umount(struct ceph_mds_client
*mdsc
)
3496 dout("pre_umount\n");
3500 ceph_flush_dirty_caps(mdsc
);
3501 wait_requests(mdsc
);
3504 * wait for reply handlers to drop their request refs and
3505 * their inode/dcache refs
3511 * wait for all write mds requests to flush.
3513 static void wait_unsafe_requests(struct ceph_mds_client
*mdsc
, u64 want_tid
)
3515 struct ceph_mds_request
*req
= NULL
, *nextreq
;
3518 mutex_lock(&mdsc
->mutex
);
3519 dout("wait_unsafe_requests want %lld\n", want_tid
);
3521 req
= __get_oldest_req(mdsc
);
3522 while (req
&& req
->r_tid
<= want_tid
) {
3523 /* find next request */
3524 n
= rb_next(&req
->r_node
);
3526 nextreq
= rb_entry(n
, struct ceph_mds_request
, r_node
);
3529 if (req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
&&
3530 (req
->r_op
& CEPH_MDS_OP_WRITE
)) {
3532 ceph_mdsc_get_request(req
);
3534 ceph_mdsc_get_request(nextreq
);
3535 mutex_unlock(&mdsc
->mutex
);
3536 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3537 req
->r_tid
, want_tid
);
3538 wait_for_completion(&req
->r_safe_completion
);
3539 mutex_lock(&mdsc
->mutex
);
3540 ceph_mdsc_put_request(req
);
3542 break; /* next dne before, so we're done! */
3543 if (RB_EMPTY_NODE(&nextreq
->r_node
)) {
3544 /* next request was removed from tree */
3545 ceph_mdsc_put_request(nextreq
);
3548 ceph_mdsc_put_request(nextreq
); /* won't go away */
3552 mutex_unlock(&mdsc
->mutex
);
3553 dout("wait_unsafe_requests done\n");
3556 void ceph_mdsc_sync(struct ceph_mds_client
*mdsc
)
3558 u64 want_tid
, want_flush
;
3560 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
3564 mutex_lock(&mdsc
->mutex
);
3565 want_tid
= mdsc
->last_tid
;
3566 mutex_unlock(&mdsc
->mutex
);
3568 ceph_flush_dirty_caps(mdsc
);
3569 spin_lock(&mdsc
->cap_dirty_lock
);
3570 want_flush
= mdsc
->last_cap_flush_tid
;
3571 if (!list_empty(&mdsc
->cap_flush_list
)) {
3572 struct ceph_cap_flush
*cf
=
3573 list_last_entry(&mdsc
->cap_flush_list
,
3574 struct ceph_cap_flush
, g_list
);
3577 spin_unlock(&mdsc
->cap_dirty_lock
);
3579 dout("sync want tid %lld flush_seq %lld\n",
3580 want_tid
, want_flush
);
3582 wait_unsafe_requests(mdsc
, want_tid
);
3583 wait_caps_flush(mdsc
, want_flush
);
3587 * true if all sessions are closed, or we force unmount
3589 static bool done_closing_sessions(struct ceph_mds_client
*mdsc
, int skipped
)
3591 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
3593 return atomic_read(&mdsc
->num_sessions
) <= skipped
;
3597 * called after sb is ro.
3599 void ceph_mdsc_close_sessions(struct ceph_mds_client
*mdsc
)
3601 struct ceph_options
*opts
= mdsc
->fsc
->client
->options
;
3602 struct ceph_mds_session
*session
;
3606 dout("close_sessions\n");
3608 /* close sessions */
3609 mutex_lock(&mdsc
->mutex
);
3610 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3611 session
= __ceph_lookup_mds_session(mdsc
, i
);
3614 mutex_unlock(&mdsc
->mutex
);
3615 mutex_lock(&session
->s_mutex
);
3616 if (__close_session(mdsc
, session
) <= 0)
3618 mutex_unlock(&session
->s_mutex
);
3619 ceph_put_mds_session(session
);
3620 mutex_lock(&mdsc
->mutex
);
3622 mutex_unlock(&mdsc
->mutex
);
3624 dout("waiting for sessions to close\n");
3625 wait_event_timeout(mdsc
->session_close_wq
,
3626 done_closing_sessions(mdsc
, skipped
),
3627 ceph_timeout_jiffies(opts
->mount_timeout
));
3629 /* tear down remaining sessions */
3630 mutex_lock(&mdsc
->mutex
);
3631 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3632 if (mdsc
->sessions
[i
]) {
3633 session
= get_session(mdsc
->sessions
[i
]);
3634 __unregister_session(mdsc
, session
);
3635 mutex_unlock(&mdsc
->mutex
);
3636 mutex_lock(&session
->s_mutex
);
3637 remove_session_caps(session
);
3638 mutex_unlock(&session
->s_mutex
);
3639 ceph_put_mds_session(session
);
3640 mutex_lock(&mdsc
->mutex
);
3643 WARN_ON(!list_empty(&mdsc
->cap_delay_list
));
3644 mutex_unlock(&mdsc
->mutex
);
3646 ceph_cleanup_empty_realms(mdsc
);
3648 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3653 void ceph_mdsc_force_umount(struct ceph_mds_client
*mdsc
)
3655 struct ceph_mds_session
*session
;
3658 dout("force umount\n");
3660 mutex_lock(&mdsc
->mutex
);
3661 for (mds
= 0; mds
< mdsc
->max_sessions
; mds
++) {
3662 session
= __ceph_lookup_mds_session(mdsc
, mds
);
3665 mutex_unlock(&mdsc
->mutex
);
3666 mutex_lock(&session
->s_mutex
);
3667 __close_session(mdsc
, session
);
3668 if (session
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3669 cleanup_session_requests(mdsc
, session
);
3670 remove_session_caps(session
);
3672 mutex_unlock(&session
->s_mutex
);
3673 ceph_put_mds_session(session
);
3674 mutex_lock(&mdsc
->mutex
);
3675 kick_requests(mdsc
, mds
);
3677 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3678 mutex_unlock(&mdsc
->mutex
);
3681 static void ceph_mdsc_stop(struct ceph_mds_client
*mdsc
)
3684 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3686 ceph_mdsmap_destroy(mdsc
->mdsmap
);
3687 kfree(mdsc
->sessions
);
3688 ceph_caps_finalize(mdsc
);
3689 ceph_pool_perm_destroy(mdsc
);
3692 void ceph_mdsc_destroy(struct ceph_fs_client
*fsc
)
3694 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
3696 dout("mdsc_destroy %p\n", mdsc
);
3697 ceph_mdsc_stop(mdsc
);
3699 /* flush out any connection work with references to us */
3704 dout("mdsc_destroy %p done\n", mdsc
);
3707 void ceph_mdsc_handle_fsmap(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3709 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3710 const char *mds_namespace
= fsc
->mount_options
->mds_namespace
;
3711 void *p
= msg
->front
.iov_base
;
3712 void *end
= p
+ msg
->front
.iov_len
;
3716 u32 mount_fscid
= (u32
)-1;
3717 u8 struct_v
, struct_cv
;
3720 ceph_decode_need(&p
, end
, sizeof(u32
), bad
);
3721 epoch
= ceph_decode_32(&p
);
3723 dout("handle_fsmap epoch %u\n", epoch
);
3725 ceph_decode_need(&p
, end
, 2 + sizeof(u32
), bad
);
3726 struct_v
= ceph_decode_8(&p
);
3727 struct_cv
= ceph_decode_8(&p
);
3728 map_len
= ceph_decode_32(&p
);
3730 ceph_decode_need(&p
, end
, sizeof(u32
) * 3, bad
);
3731 p
+= sizeof(u32
) * 2; /* skip epoch and legacy_client_fscid */
3733 num_fs
= ceph_decode_32(&p
);
3734 while (num_fs
-- > 0) {
3735 void *info_p
, *info_end
;
3740 ceph_decode_need(&p
, end
, 2 + sizeof(u32
), bad
);
3741 info_v
= ceph_decode_8(&p
);
3742 info_cv
= ceph_decode_8(&p
);
3743 info_len
= ceph_decode_32(&p
);
3744 ceph_decode_need(&p
, end
, info_len
, bad
);
3746 info_end
= p
+ info_len
;
3749 ceph_decode_need(&info_p
, info_end
, sizeof(u32
) * 2, bad
);
3750 fscid
= ceph_decode_32(&info_p
);
3751 namelen
= ceph_decode_32(&info_p
);
3752 ceph_decode_need(&info_p
, info_end
, namelen
, bad
);
3754 if (mds_namespace
&&
3755 strlen(mds_namespace
) == namelen
&&
3756 !strncmp(mds_namespace
, (char *)info_p
, namelen
)) {
3757 mount_fscid
= fscid
;
3762 ceph_monc_got_map(&fsc
->client
->monc
, CEPH_SUB_FSMAP
, epoch
);
3763 if (mount_fscid
!= (u32
)-1) {
3764 fsc
->client
->monc
.fs_cluster_id
= mount_fscid
;
3765 ceph_monc_want_map(&fsc
->client
->monc
, CEPH_SUB_MDSMAP
,
3767 ceph_monc_renew_subs(&fsc
->client
->monc
);
3774 pr_err("error decoding fsmap\n");
3776 mutex_lock(&mdsc
->mutex
);
3777 mdsc
->mdsmap_err
= -ENOENT
;
3778 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3779 mutex_unlock(&mdsc
->mutex
);
3784 * handle mds map update.
3786 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3790 void *p
= msg
->front
.iov_base
;
3791 void *end
= p
+ msg
->front
.iov_len
;
3792 struct ceph_mdsmap
*newmap
, *oldmap
;
3793 struct ceph_fsid fsid
;
3796 ceph_decode_need(&p
, end
, sizeof(fsid
)+2*sizeof(u32
), bad
);
3797 ceph_decode_copy(&p
, &fsid
, sizeof(fsid
));
3798 if (ceph_check_fsid(mdsc
->fsc
->client
, &fsid
) < 0)
3800 epoch
= ceph_decode_32(&p
);
3801 maplen
= ceph_decode_32(&p
);
3802 dout("handle_map epoch %u len %d\n", epoch
, (int)maplen
);
3804 /* do we need it? */
3805 mutex_lock(&mdsc
->mutex
);
3806 if (mdsc
->mdsmap
&& epoch
<= mdsc
->mdsmap
->m_epoch
) {
3807 dout("handle_map epoch %u <= our %u\n",
3808 epoch
, mdsc
->mdsmap
->m_epoch
);
3809 mutex_unlock(&mdsc
->mutex
);
3813 newmap
= ceph_mdsmap_decode(&p
, end
);
3814 if (IS_ERR(newmap
)) {
3815 err
= PTR_ERR(newmap
);
3819 /* swap into place */
3821 oldmap
= mdsc
->mdsmap
;
3822 mdsc
->mdsmap
= newmap
;
3823 check_new_map(mdsc
, newmap
, oldmap
);
3824 ceph_mdsmap_destroy(oldmap
);
3826 mdsc
->mdsmap
= newmap
; /* first mds map */
3828 mdsc
->fsc
->sb
->s_maxbytes
= mdsc
->mdsmap
->m_max_file_size
;
3830 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3831 ceph_monc_got_map(&mdsc
->fsc
->client
->monc
, CEPH_SUB_MDSMAP
,
3832 mdsc
->mdsmap
->m_epoch
);
3834 mutex_unlock(&mdsc
->mutex
);
3835 schedule_delayed(mdsc
);
3839 mutex_unlock(&mdsc
->mutex
);
3841 pr_err("error decoding mdsmap %d\n", err
);
3845 static struct ceph_connection
*con_get(struct ceph_connection
*con
)
3847 struct ceph_mds_session
*s
= con
->private;
3849 if (get_session(s
)) {
3850 dout("mdsc con_get %p ok (%d)\n", s
, atomic_read(&s
->s_ref
));
3853 dout("mdsc con_get %p FAIL\n", s
);
3857 static void con_put(struct ceph_connection
*con
)
3859 struct ceph_mds_session
*s
= con
->private;
3861 dout("mdsc con_put %p (%d)\n", s
, atomic_read(&s
->s_ref
) - 1);
3862 ceph_put_mds_session(s
);
3866 * if the client is unresponsive for long enough, the mds will kill
3867 * the session entirely.
3869 static void peer_reset(struct ceph_connection
*con
)
3871 struct ceph_mds_session
*s
= con
->private;
3872 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3874 pr_warn("mds%d closed our session\n", s
->s_mds
);
3875 send_mds_reconnect(mdsc
, s
);
3878 static void dispatch(struct ceph_connection
*con
, struct ceph_msg
*msg
)
3880 struct ceph_mds_session
*s
= con
->private;
3881 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3882 int type
= le16_to_cpu(msg
->hdr
.type
);
3884 mutex_lock(&mdsc
->mutex
);
3885 if (__verify_registered_session(mdsc
, s
) < 0) {
3886 mutex_unlock(&mdsc
->mutex
);
3889 mutex_unlock(&mdsc
->mutex
);
3892 case CEPH_MSG_MDS_MAP
:
3893 ceph_mdsc_handle_mdsmap(mdsc
, msg
);
3895 case CEPH_MSG_FS_MAP_USER
:
3896 ceph_mdsc_handle_fsmap(mdsc
, msg
);
3898 case CEPH_MSG_CLIENT_SESSION
:
3899 handle_session(s
, msg
);
3901 case CEPH_MSG_CLIENT_REPLY
:
3902 handle_reply(s
, msg
);
3904 case CEPH_MSG_CLIENT_REQUEST_FORWARD
:
3905 handle_forward(mdsc
, s
, msg
);
3907 case CEPH_MSG_CLIENT_CAPS
:
3908 ceph_handle_caps(s
, msg
);
3910 case CEPH_MSG_CLIENT_SNAP
:
3911 ceph_handle_snap(mdsc
, s
, msg
);
3913 case CEPH_MSG_CLIENT_LEASE
:
3914 handle_lease(mdsc
, s
, msg
);
3918 pr_err("received unknown message type %d %s\n", type
,
3919 ceph_msg_type_name(type
));
3930 * Note: returned pointer is the address of a structure that's
3931 * managed separately. Caller must *not* attempt to free it.
3933 static struct ceph_auth_handshake
*get_authorizer(struct ceph_connection
*con
,
3934 int *proto
, int force_new
)
3936 struct ceph_mds_session
*s
= con
->private;
3937 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3938 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3939 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
3941 if (force_new
&& auth
->authorizer
) {
3942 ceph_auth_destroy_authorizer(auth
->authorizer
);
3943 auth
->authorizer
= NULL
;
3945 if (!auth
->authorizer
) {
3946 int ret
= ceph_auth_create_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3949 return ERR_PTR(ret
);
3951 int ret
= ceph_auth_update_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3954 return ERR_PTR(ret
);
3956 *proto
= ac
->protocol
;
3962 static int verify_authorizer_reply(struct ceph_connection
*con
)
3964 struct ceph_mds_session
*s
= con
->private;
3965 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3966 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3968 return ceph_auth_verify_authorizer_reply(ac
, s
->s_auth
.authorizer
);
3971 static int invalidate_authorizer(struct ceph_connection
*con
)
3973 struct ceph_mds_session
*s
= con
->private;
3974 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3975 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3977 ceph_auth_invalidate_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
);
3979 return ceph_monc_validate_auth(&mdsc
->fsc
->client
->monc
);
3982 static struct ceph_msg
*mds_alloc_msg(struct ceph_connection
*con
,
3983 struct ceph_msg_header
*hdr
, int *skip
)
3985 struct ceph_msg
*msg
;
3986 int type
= (int) le16_to_cpu(hdr
->type
);
3987 int front_len
= (int) le32_to_cpu(hdr
->front_len
);
3993 msg
= ceph_msg_new(type
, front_len
, GFP_NOFS
, false);
3995 pr_err("unable to allocate msg type %d len %d\n",
4003 static int mds_sign_message(struct ceph_msg
*msg
)
4005 struct ceph_mds_session
*s
= msg
->con
->private;
4006 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
4008 return ceph_auth_sign_message(auth
, msg
);
4011 static int mds_check_message_signature(struct ceph_msg
*msg
)
4013 struct ceph_mds_session
*s
= msg
->con
->private;
4014 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
4016 return ceph_auth_check_message_signature(auth
, msg
);
4019 static const struct ceph_connection_operations mds_con_ops
= {
4022 .dispatch
= dispatch
,
4023 .get_authorizer
= get_authorizer
,
4024 .verify_authorizer_reply
= verify_authorizer_reply
,
4025 .invalidate_authorizer
= invalidate_authorizer
,
4026 .peer_reset
= peer_reset
,
4027 .alloc_msg
= mds_alloc_msg
,
4028 .sign_message
= mds_sign_message
,
4029 .check_message_signature
= mds_check_message_signature
,