1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
6 #include <linux/sched.h>
7 #include <linux/debugfs.h>
8 #include <linux/seq_file.h>
11 #include "mds_client.h"
13 #include <linux/ceph/ceph_features.h>
14 #include <linux/ceph/messenger.h>
15 #include <linux/ceph/decode.h>
16 #include <linux/ceph/pagelist.h>
17 #include <linux/ceph/auth.h>
18 #include <linux/ceph/debugfs.h>
21 * A cluster of MDS (metadata server) daemons is responsible for
22 * managing the file system namespace (the directory hierarchy and
23 * inodes) and for coordinating shared access to storage. Metadata is
24 * partitioning hierarchically across a number of servers, and that
25 * partition varies over time as the cluster adjusts the distribution
26 * in order to balance load.
28 * The MDS client is primarily responsible to managing synchronous
29 * metadata requests for operations like open, unlink, and so forth.
30 * If there is a MDS failure, we find out about it when we (possibly
31 * request and) receive a new MDS map, and can resubmit affected
34 * For the most part, though, we take advantage of a lossless
35 * communications channel to the MDS, and do not need to worry about
36 * timing out or resubmitting requests.
38 * We maintain a stateful "session" with each MDS we interact with.
39 * Within each session, we sent periodic heartbeat messages to ensure
40 * any capabilities or leases we have been issues remain valid. If
41 * the session times out and goes stale, our leases and capabilities
42 * are no longer valid.
45 struct ceph_reconnect_state
{
46 struct ceph_pagelist
*pagelist
;
50 static void __wake_requests(struct ceph_mds_client
*mdsc
,
51 struct list_head
*head
);
53 static const struct ceph_connection_operations mds_con_ops
;
61 * parse individual inode info
63 static int parse_reply_info_in(void **p
, void *end
,
64 struct ceph_mds_reply_info_in
*info
,
70 *p
+= sizeof(struct ceph_mds_reply_inode
) +
71 sizeof(*info
->in
->fragtree
.splits
) *
72 le32_to_cpu(info
->in
->fragtree
.nsplits
);
74 ceph_decode_32_safe(p
, end
, info
->symlink_len
, bad
);
75 ceph_decode_need(p
, end
, info
->symlink_len
, bad
);
77 *p
+= info
->symlink_len
;
79 if (features
& CEPH_FEATURE_DIRLAYOUTHASH
)
80 ceph_decode_copy_safe(p
, end
, &info
->dir_layout
,
81 sizeof(info
->dir_layout
), bad
);
83 memset(&info
->dir_layout
, 0, sizeof(info
->dir_layout
));
85 ceph_decode_32_safe(p
, end
, info
->xattr_len
, bad
);
86 ceph_decode_need(p
, end
, info
->xattr_len
, bad
);
87 info
->xattr_data
= *p
;
88 *p
+= info
->xattr_len
;
95 * parse a normal reply, which may contain a (dir+)dentry and/or a
98 static int parse_reply_info_trace(void **p
, void *end
,
99 struct ceph_mds_reply_info_parsed
*info
,
104 if (info
->head
->is_dentry
) {
105 err
= parse_reply_info_in(p
, end
, &info
->diri
, features
);
109 if (unlikely(*p
+ sizeof(*info
->dirfrag
) > end
))
112 *p
+= sizeof(*info
->dirfrag
) +
113 sizeof(u32
)*le32_to_cpu(info
->dirfrag
->ndist
);
114 if (unlikely(*p
> end
))
117 ceph_decode_32_safe(p
, end
, info
->dname_len
, bad
);
118 ceph_decode_need(p
, end
, info
->dname_len
, bad
);
120 *p
+= info
->dname_len
;
122 *p
+= sizeof(*info
->dlease
);
125 if (info
->head
->is_target
) {
126 err
= parse_reply_info_in(p
, end
, &info
->targeti
, features
);
131 if (unlikely(*p
!= end
))
138 pr_err("problem parsing mds trace %d\n", err
);
143 * parse readdir results
145 static int parse_reply_info_dir(void **p
, void *end
,
146 struct ceph_mds_reply_info_parsed
*info
,
153 if (*p
+ sizeof(*info
->dir_dir
) > end
)
155 *p
+= sizeof(*info
->dir_dir
) +
156 sizeof(u32
)*le32_to_cpu(info
->dir_dir
->ndist
);
160 ceph_decode_need(p
, end
, sizeof(num
) + 2, bad
);
161 num
= ceph_decode_32(p
);
162 info
->dir_end
= ceph_decode_8(p
);
163 info
->dir_complete
= ceph_decode_8(p
);
167 /* alloc large array */
169 info
->dir_in
= kcalloc(num
, sizeof(*info
->dir_in
) +
170 sizeof(*info
->dir_dname
) +
171 sizeof(*info
->dir_dname_len
) +
172 sizeof(*info
->dir_dlease
),
174 if (info
->dir_in
== NULL
) {
178 info
->dir_dname
= (void *)(info
->dir_in
+ num
);
179 info
->dir_dname_len
= (void *)(info
->dir_dname
+ num
);
180 info
->dir_dlease
= (void *)(info
->dir_dname_len
+ num
);
184 ceph_decode_need(p
, end
, sizeof(u32
)*2, bad
);
185 info
->dir_dname_len
[i
] = ceph_decode_32(p
);
186 ceph_decode_need(p
, end
, info
->dir_dname_len
[i
], bad
);
187 info
->dir_dname
[i
] = *p
;
188 *p
+= info
->dir_dname_len
[i
];
189 dout("parsed dir dname '%.*s'\n", info
->dir_dname_len
[i
],
191 info
->dir_dlease
[i
] = *p
;
192 *p
+= sizeof(struct ceph_mds_reply_lease
);
195 err
= parse_reply_info_in(p
, end
, &info
->dir_in
[i
], features
);
210 pr_err("problem parsing dir contents %d\n", err
);
215 * parse fcntl F_GETLK results
217 static int parse_reply_info_filelock(void **p
, void *end
,
218 struct ceph_mds_reply_info_parsed
*info
,
221 if (*p
+ sizeof(*info
->filelock_reply
) > end
)
224 info
->filelock_reply
= *p
;
225 *p
+= sizeof(*info
->filelock_reply
);
227 if (unlikely(*p
!= end
))
236 * parse create results
238 static int parse_reply_info_create(void **p
, void *end
,
239 struct ceph_mds_reply_info_parsed
*info
,
242 if (features
& CEPH_FEATURE_REPLY_CREATE_INODE
) {
244 info
->has_create_ino
= false;
246 info
->has_create_ino
= true;
247 info
->ino
= ceph_decode_64(p
);
251 if (unlikely(*p
!= end
))
260 * parse extra results
262 static int parse_reply_info_extra(void **p
, void *end
,
263 struct ceph_mds_reply_info_parsed
*info
,
266 if (info
->head
->op
== CEPH_MDS_OP_GETFILELOCK
)
267 return parse_reply_info_filelock(p
, end
, info
, features
);
268 else if (info
->head
->op
== CEPH_MDS_OP_READDIR
||
269 info
->head
->op
== CEPH_MDS_OP_LSSNAP
)
270 return parse_reply_info_dir(p
, end
, info
, features
);
271 else if (info
->head
->op
== CEPH_MDS_OP_CREATE
)
272 return parse_reply_info_create(p
, end
, info
, features
);
278 * parse entire mds reply
280 static int parse_reply_info(struct ceph_msg
*msg
,
281 struct ceph_mds_reply_info_parsed
*info
,
288 info
->head
= msg
->front
.iov_base
;
289 p
= msg
->front
.iov_base
+ sizeof(struct ceph_mds_reply_head
);
290 end
= p
+ msg
->front
.iov_len
- sizeof(struct ceph_mds_reply_head
);
293 ceph_decode_32_safe(&p
, end
, len
, bad
);
295 ceph_decode_need(&p
, end
, len
, bad
);
296 err
= parse_reply_info_trace(&p
, p
+len
, info
, features
);
302 ceph_decode_32_safe(&p
, end
, len
, bad
);
304 ceph_decode_need(&p
, end
, len
, bad
);
305 err
= parse_reply_info_extra(&p
, p
+len
, info
, features
);
311 ceph_decode_32_safe(&p
, end
, len
, bad
);
312 info
->snapblob_len
= len
;
323 pr_err("mds parse_reply err %d\n", err
);
327 static void destroy_reply_info(struct ceph_mds_reply_info_parsed
*info
)
336 static const char *session_state_name(int s
)
339 case CEPH_MDS_SESSION_NEW
: return "new";
340 case CEPH_MDS_SESSION_OPENING
: return "opening";
341 case CEPH_MDS_SESSION_OPEN
: return "open";
342 case CEPH_MDS_SESSION_HUNG
: return "hung";
343 case CEPH_MDS_SESSION_CLOSING
: return "closing";
344 case CEPH_MDS_SESSION_RESTARTING
: return "restarting";
345 case CEPH_MDS_SESSION_RECONNECTING
: return "reconnecting";
346 default: return "???";
350 static struct ceph_mds_session
*get_session(struct ceph_mds_session
*s
)
352 if (atomic_inc_not_zero(&s
->s_ref
)) {
353 dout("mdsc get_session %p %d -> %d\n", s
,
354 atomic_read(&s
->s_ref
)-1, atomic_read(&s
->s_ref
));
357 dout("mdsc get_session %p 0 -- FAIL", s
);
362 void ceph_put_mds_session(struct ceph_mds_session
*s
)
364 dout("mdsc put_session %p %d -> %d\n", s
,
365 atomic_read(&s
->s_ref
), atomic_read(&s
->s_ref
)-1);
366 if (atomic_dec_and_test(&s
->s_ref
)) {
367 if (s
->s_auth
.authorizer
)
368 ceph_auth_destroy_authorizer(
369 s
->s_mdsc
->fsc
->client
->monc
.auth
,
370 s
->s_auth
.authorizer
);
376 * called under mdsc->mutex
378 struct ceph_mds_session
*__ceph_lookup_mds_session(struct ceph_mds_client
*mdsc
,
381 struct ceph_mds_session
*session
;
383 if (mds
>= mdsc
->max_sessions
|| mdsc
->sessions
[mds
] == NULL
)
385 session
= mdsc
->sessions
[mds
];
386 dout("lookup_mds_session %p %d\n", session
,
387 atomic_read(&session
->s_ref
));
388 get_session(session
);
392 static bool __have_session(struct ceph_mds_client
*mdsc
, int mds
)
394 if (mds
>= mdsc
->max_sessions
)
396 return mdsc
->sessions
[mds
];
399 static int __verify_registered_session(struct ceph_mds_client
*mdsc
,
400 struct ceph_mds_session
*s
)
402 if (s
->s_mds
>= mdsc
->max_sessions
||
403 mdsc
->sessions
[s
->s_mds
] != s
)
409 * create+register a new session for given mds.
410 * called under mdsc->mutex.
412 static struct ceph_mds_session
*register_session(struct ceph_mds_client
*mdsc
,
415 struct ceph_mds_session
*s
;
417 if (mds
>= mdsc
->mdsmap
->m_max_mds
)
418 return ERR_PTR(-EINVAL
);
420 s
= kzalloc(sizeof(*s
), GFP_NOFS
);
422 return ERR_PTR(-ENOMEM
);
425 s
->s_state
= CEPH_MDS_SESSION_NEW
;
428 mutex_init(&s
->s_mutex
);
430 ceph_con_init(&s
->s_con
, s
, &mds_con_ops
, &mdsc
->fsc
->client
->msgr
);
432 spin_lock_init(&s
->s_gen_ttl_lock
);
434 s
->s_cap_ttl
= jiffies
- 1;
436 spin_lock_init(&s
->s_cap_lock
);
437 s
->s_renew_requested
= 0;
439 INIT_LIST_HEAD(&s
->s_caps
);
442 atomic_set(&s
->s_ref
, 1);
443 INIT_LIST_HEAD(&s
->s_waiting
);
444 INIT_LIST_HEAD(&s
->s_unsafe
);
445 s
->s_num_cap_releases
= 0;
446 s
->s_cap_iterator
= NULL
;
447 INIT_LIST_HEAD(&s
->s_cap_releases
);
448 INIT_LIST_HEAD(&s
->s_cap_releases_done
);
449 INIT_LIST_HEAD(&s
->s_cap_flushing
);
450 INIT_LIST_HEAD(&s
->s_cap_snaps_flushing
);
452 dout("register_session mds%d\n", mds
);
453 if (mds
>= mdsc
->max_sessions
) {
454 int newmax
= 1 << get_count_order(mds
+1);
455 struct ceph_mds_session
**sa
;
457 dout("register_session realloc to %d\n", newmax
);
458 sa
= kcalloc(newmax
, sizeof(void *), GFP_NOFS
);
461 if (mdsc
->sessions
) {
462 memcpy(sa
, mdsc
->sessions
,
463 mdsc
->max_sessions
* sizeof(void *));
464 kfree(mdsc
->sessions
);
467 mdsc
->max_sessions
= newmax
;
469 mdsc
->sessions
[mds
] = s
;
470 atomic_inc(&s
->s_ref
); /* one ref to sessions[], one to caller */
472 ceph_con_open(&s
->s_con
, CEPH_ENTITY_TYPE_MDS
, mds
,
473 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
479 return ERR_PTR(-ENOMEM
);
483 * called under mdsc->mutex
485 static void __unregister_session(struct ceph_mds_client
*mdsc
,
486 struct ceph_mds_session
*s
)
488 dout("__unregister_session mds%d %p\n", s
->s_mds
, s
);
489 BUG_ON(mdsc
->sessions
[s
->s_mds
] != s
);
490 mdsc
->sessions
[s
->s_mds
] = NULL
;
491 ceph_con_close(&s
->s_con
);
492 ceph_put_mds_session(s
);
496 * drop session refs in request.
498 * should be last request ref, or hold mdsc->mutex
500 static void put_request_session(struct ceph_mds_request
*req
)
502 if (req
->r_session
) {
503 ceph_put_mds_session(req
->r_session
);
504 req
->r_session
= NULL
;
508 void ceph_mdsc_release_request(struct kref
*kref
)
510 struct ceph_mds_request
*req
= container_of(kref
,
511 struct ceph_mds_request
,
514 ceph_msg_put(req
->r_request
);
516 ceph_msg_put(req
->r_reply
);
517 destroy_reply_info(&req
->r_reply_info
);
520 ceph_put_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
523 if (req
->r_locked_dir
)
524 ceph_put_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
525 if (req
->r_target_inode
)
526 iput(req
->r_target_inode
);
529 if (req
->r_old_dentry
) {
531 * track (and drop pins for) r_old_dentry_dir
532 * separately, since r_old_dentry's d_parent may have
533 * changed between the dir mutex being dropped and
534 * this request being freed.
536 ceph_put_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
538 dput(req
->r_old_dentry
);
539 iput(req
->r_old_dentry_dir
);
543 put_request_session(req
);
544 ceph_unreserve_caps(req
->r_mdsc
, &req
->r_caps_reservation
);
549 * lookup session, bump ref if found.
551 * called under mdsc->mutex.
553 static struct ceph_mds_request
*__lookup_request(struct ceph_mds_client
*mdsc
,
556 struct ceph_mds_request
*req
;
557 struct rb_node
*n
= mdsc
->request_tree
.rb_node
;
560 req
= rb_entry(n
, struct ceph_mds_request
, r_node
);
561 if (tid
< req
->r_tid
)
563 else if (tid
> req
->r_tid
)
566 ceph_mdsc_get_request(req
);
573 static void __insert_request(struct ceph_mds_client
*mdsc
,
574 struct ceph_mds_request
*new)
576 struct rb_node
**p
= &mdsc
->request_tree
.rb_node
;
577 struct rb_node
*parent
= NULL
;
578 struct ceph_mds_request
*req
= NULL
;
582 req
= rb_entry(parent
, struct ceph_mds_request
, r_node
);
583 if (new->r_tid
< req
->r_tid
)
585 else if (new->r_tid
> req
->r_tid
)
591 rb_link_node(&new->r_node
, parent
, p
);
592 rb_insert_color(&new->r_node
, &mdsc
->request_tree
);
596 * Register an in-flight request, and assign a tid. Link to directory
597 * are modifying (if any).
599 * Called under mdsc->mutex.
601 static void __register_request(struct ceph_mds_client
*mdsc
,
602 struct ceph_mds_request
*req
,
605 req
->r_tid
= ++mdsc
->last_tid
;
607 ceph_reserve_caps(mdsc
, &req
->r_caps_reservation
,
609 dout("__register_request %p tid %lld\n", req
, req
->r_tid
);
610 ceph_mdsc_get_request(req
);
611 __insert_request(mdsc
, req
);
613 req
->r_uid
= current_fsuid();
614 req
->r_gid
= current_fsgid();
617 struct ceph_inode_info
*ci
= ceph_inode(dir
);
620 spin_lock(&ci
->i_unsafe_lock
);
621 req
->r_unsafe_dir
= dir
;
622 list_add_tail(&req
->r_unsafe_dir_item
, &ci
->i_unsafe_dirops
);
623 spin_unlock(&ci
->i_unsafe_lock
);
627 static void __unregister_request(struct ceph_mds_client
*mdsc
,
628 struct ceph_mds_request
*req
)
630 dout("__unregister_request %p tid %lld\n", req
, req
->r_tid
);
631 rb_erase(&req
->r_node
, &mdsc
->request_tree
);
632 RB_CLEAR_NODE(&req
->r_node
);
634 if (req
->r_unsafe_dir
) {
635 struct ceph_inode_info
*ci
= ceph_inode(req
->r_unsafe_dir
);
637 spin_lock(&ci
->i_unsafe_lock
);
638 list_del_init(&req
->r_unsafe_dir_item
);
639 spin_unlock(&ci
->i_unsafe_lock
);
641 iput(req
->r_unsafe_dir
);
642 req
->r_unsafe_dir
= NULL
;
645 ceph_mdsc_put_request(req
);
649 * Choose mds to send request to next. If there is a hint set in the
650 * request (e.g., due to a prior forward hint from the mds), use that.
651 * Otherwise, consult frag tree and/or caps to identify the
652 * appropriate mds. If all else fails, choose randomly.
654 * Called under mdsc->mutex.
656 static struct dentry
*get_nonsnap_parent(struct dentry
*dentry
)
659 * we don't need to worry about protecting the d_parent access
660 * here because we never renaming inside the snapped namespace
661 * except to resplice to another snapdir, and either the old or new
662 * result is a valid result.
664 while (!IS_ROOT(dentry
) && ceph_snap(dentry
->d_inode
) != CEPH_NOSNAP
)
665 dentry
= dentry
->d_parent
;
669 static int __choose_mds(struct ceph_mds_client
*mdsc
,
670 struct ceph_mds_request
*req
)
673 struct ceph_inode_info
*ci
;
674 struct ceph_cap
*cap
;
675 int mode
= req
->r_direct_mode
;
677 u32 hash
= req
->r_direct_hash
;
678 bool is_hash
= req
->r_direct_is_hash
;
681 * is there a specific mds we should try? ignore hint if we have
682 * no session and the mds is not up (active or recovering).
684 if (req
->r_resend_mds
>= 0 &&
685 (__have_session(mdsc
, req
->r_resend_mds
) ||
686 ceph_mdsmap_get_state(mdsc
->mdsmap
, req
->r_resend_mds
) > 0)) {
687 dout("choose_mds using resend_mds mds%d\n",
689 return req
->r_resend_mds
;
692 if (mode
== USE_RANDOM_MDS
)
697 inode
= req
->r_inode
;
698 } else if (req
->r_dentry
) {
699 /* ignore race with rename; old or new d_parent is okay */
700 struct dentry
*parent
= req
->r_dentry
->d_parent
;
701 struct inode
*dir
= parent
->d_inode
;
703 if (dir
->i_sb
!= mdsc
->fsc
->sb
) {
705 inode
= req
->r_dentry
->d_inode
;
706 } else if (ceph_snap(dir
) != CEPH_NOSNAP
) {
707 /* direct snapped/virtual snapdir requests
708 * based on parent dir inode */
709 struct dentry
*dn
= get_nonsnap_parent(parent
);
711 dout("__choose_mds using nonsnap parent %p\n", inode
);
712 } else if (req
->r_dentry
->d_inode
) {
714 inode
= req
->r_dentry
->d_inode
;
718 hash
= ceph_dentry_hash(dir
, req
->r_dentry
);
723 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode
, (int)is_hash
,
727 ci
= ceph_inode(inode
);
729 if (is_hash
&& S_ISDIR(inode
->i_mode
)) {
730 struct ceph_inode_frag frag
;
733 ceph_choose_frag(ci
, hash
, &frag
, &found
);
735 if (mode
== USE_ANY_MDS
&& frag
.ndist
> 0) {
738 /* choose a random replica */
739 get_random_bytes(&r
, 1);
742 dout("choose_mds %p %llx.%llx "
743 "frag %u mds%d (%d/%d)\n",
744 inode
, ceph_vinop(inode
),
747 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
748 CEPH_MDS_STATE_ACTIVE
)
752 /* since this file/dir wasn't known to be
753 * replicated, then we want to look for the
754 * authoritative mds. */
757 /* choose auth mds */
759 dout("choose_mds %p %llx.%llx "
760 "frag %u mds%d (auth)\n",
761 inode
, ceph_vinop(inode
), frag
.frag
, mds
);
762 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
763 CEPH_MDS_STATE_ACTIVE
)
769 spin_lock(&ci
->i_ceph_lock
);
771 if (mode
== USE_AUTH_MDS
)
772 cap
= ci
->i_auth_cap
;
773 if (!cap
&& !RB_EMPTY_ROOT(&ci
->i_caps
))
774 cap
= rb_entry(rb_first(&ci
->i_caps
), struct ceph_cap
, ci_node
);
776 spin_unlock(&ci
->i_ceph_lock
);
779 mds
= cap
->session
->s_mds
;
780 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
781 inode
, ceph_vinop(inode
), mds
,
782 cap
== ci
->i_auth_cap
? "auth " : "", cap
);
783 spin_unlock(&ci
->i_ceph_lock
);
787 mds
= ceph_mdsmap_get_random_mds(mdsc
->mdsmap
);
788 dout("choose_mds chose random mds%d\n", mds
);
796 static struct ceph_msg
*create_session_msg(u32 op
, u64 seq
)
798 struct ceph_msg
*msg
;
799 struct ceph_mds_session_head
*h
;
801 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
), GFP_NOFS
,
804 pr_err("create_session_msg ENOMEM creating msg\n");
807 h
= msg
->front
.iov_base
;
808 h
->op
= cpu_to_le32(op
);
809 h
->seq
= cpu_to_le64(seq
);
814 * send session open request.
816 * called under mdsc->mutex
818 static int __open_session(struct ceph_mds_client
*mdsc
,
819 struct ceph_mds_session
*session
)
821 struct ceph_msg
*msg
;
823 int mds
= session
->s_mds
;
825 /* wait for mds to go active? */
826 mstate
= ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
);
827 dout("open_session to mds%d (%s)\n", mds
,
828 ceph_mds_state_name(mstate
));
829 session
->s_state
= CEPH_MDS_SESSION_OPENING
;
830 session
->s_renew_requested
= jiffies
;
832 /* send connect message */
833 msg
= create_session_msg(CEPH_SESSION_REQUEST_OPEN
, session
->s_seq
);
836 ceph_con_send(&session
->s_con
, msg
);
841 * open sessions for any export targets for the given mds
843 * called under mdsc->mutex
845 static void __open_export_target_sessions(struct ceph_mds_client
*mdsc
,
846 struct ceph_mds_session
*session
)
848 struct ceph_mds_info
*mi
;
849 struct ceph_mds_session
*ts
;
850 int i
, mds
= session
->s_mds
;
853 if (mds
>= mdsc
->mdsmap
->m_max_mds
)
855 mi
= &mdsc
->mdsmap
->m_info
[mds
];
856 dout("open_export_target_sessions for mds%d (%d targets)\n",
857 session
->s_mds
, mi
->num_export_targets
);
859 for (i
= 0; i
< mi
->num_export_targets
; i
++) {
860 target
= mi
->export_targets
[i
];
861 ts
= __ceph_lookup_mds_session(mdsc
, target
);
863 ts
= register_session(mdsc
, target
);
867 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
868 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
869 __open_session(mdsc
, session
);
871 dout(" mds%d target mds%d %p is %s\n", session
->s_mds
,
872 i
, ts
, session_state_name(ts
->s_state
));
873 ceph_put_mds_session(ts
);
877 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client
*mdsc
,
878 struct ceph_mds_session
*session
)
880 mutex_lock(&mdsc
->mutex
);
881 __open_export_target_sessions(mdsc
, session
);
882 mutex_unlock(&mdsc
->mutex
);
890 * Free preallocated cap messages assigned to this session
892 static void cleanup_cap_releases(struct ceph_mds_session
*session
)
894 struct ceph_msg
*msg
;
896 spin_lock(&session
->s_cap_lock
);
897 while (!list_empty(&session
->s_cap_releases
)) {
898 msg
= list_first_entry(&session
->s_cap_releases
,
899 struct ceph_msg
, list_head
);
900 list_del_init(&msg
->list_head
);
903 while (!list_empty(&session
->s_cap_releases_done
)) {
904 msg
= list_first_entry(&session
->s_cap_releases_done
,
905 struct ceph_msg
, list_head
);
906 list_del_init(&msg
->list_head
);
909 spin_unlock(&session
->s_cap_lock
);
913 * Helper to safely iterate over all caps associated with a session, with
914 * special care taken to handle a racing __ceph_remove_cap().
916 * Caller must hold session s_mutex.
918 static int iterate_session_caps(struct ceph_mds_session
*session
,
919 int (*cb
)(struct inode
*, struct ceph_cap
*,
923 struct ceph_cap
*cap
;
924 struct inode
*inode
, *last_inode
= NULL
;
925 struct ceph_cap
*old_cap
= NULL
;
928 dout("iterate_session_caps %p mds%d\n", session
, session
->s_mds
);
929 spin_lock(&session
->s_cap_lock
);
930 p
= session
->s_caps
.next
;
931 while (p
!= &session
->s_caps
) {
932 cap
= list_entry(p
, struct ceph_cap
, session_caps
);
933 inode
= igrab(&cap
->ci
->vfs_inode
);
938 session
->s_cap_iterator
= cap
;
939 spin_unlock(&session
->s_cap_lock
);
946 ceph_put_cap(session
->s_mdsc
, old_cap
);
950 ret
= cb(inode
, cap
, arg
);
953 spin_lock(&session
->s_cap_lock
);
955 if (cap
->ci
== NULL
) {
956 dout("iterate_session_caps finishing cap %p removal\n",
958 BUG_ON(cap
->session
!= session
);
959 list_del_init(&cap
->session_caps
);
960 session
->s_nr_caps
--;
962 old_cap
= cap
; /* put_cap it w/o locks held */
969 session
->s_cap_iterator
= NULL
;
970 spin_unlock(&session
->s_cap_lock
);
975 ceph_put_cap(session
->s_mdsc
, old_cap
);
980 static int remove_session_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
983 struct ceph_inode_info
*ci
= ceph_inode(inode
);
986 dout("removing cap %p, ci is %p, inode is %p\n",
987 cap
, ci
, &ci
->vfs_inode
);
988 spin_lock(&ci
->i_ceph_lock
);
989 __ceph_remove_cap(cap
);
990 if (!__ceph_is_any_real_caps(ci
)) {
991 struct ceph_mds_client
*mdsc
=
992 ceph_sb_to_client(inode
->i_sb
)->mdsc
;
994 spin_lock(&mdsc
->cap_dirty_lock
);
995 if (!list_empty(&ci
->i_dirty_item
)) {
996 pr_info(" dropping dirty %s state for %p %lld\n",
997 ceph_cap_string(ci
->i_dirty_caps
),
998 inode
, ceph_ino(inode
));
999 ci
->i_dirty_caps
= 0;
1000 list_del_init(&ci
->i_dirty_item
);
1003 if (!list_empty(&ci
->i_flushing_item
)) {
1004 pr_info(" dropping dirty+flushing %s state for %p %lld\n",
1005 ceph_cap_string(ci
->i_flushing_caps
),
1006 inode
, ceph_ino(inode
));
1007 ci
->i_flushing_caps
= 0;
1008 list_del_init(&ci
->i_flushing_item
);
1009 mdsc
->num_cap_flushing
--;
1012 if (drop
&& ci
->i_wrbuffer_ref
) {
1013 pr_info(" dropping dirty data for %p %lld\n",
1014 inode
, ceph_ino(inode
));
1015 ci
->i_wrbuffer_ref
= 0;
1016 ci
->i_wrbuffer_ref_head
= 0;
1019 spin_unlock(&mdsc
->cap_dirty_lock
);
1021 spin_unlock(&ci
->i_ceph_lock
);
1028 * caller must hold session s_mutex
1030 static void remove_session_caps(struct ceph_mds_session
*session
)
1032 dout("remove_session_caps on %p\n", session
);
1033 iterate_session_caps(session
, remove_session_caps_cb
, NULL
);
1035 spin_lock(&session
->s_cap_lock
);
1036 if (session
->s_nr_caps
> 0) {
1037 struct super_block
*sb
= session
->s_mdsc
->fsc
->sb
;
1038 struct inode
*inode
;
1039 struct ceph_cap
*cap
, *prev
= NULL
;
1040 struct ceph_vino vino
;
1042 * iterate_session_caps() skips inodes that are being
1043 * deleted, we need to wait until deletions are complete.
1044 * __wait_on_freeing_inode() is designed for the job,
1045 * but it is not exported, so use lookup inode function
1048 while (!list_empty(&session
->s_caps
)) {
1049 cap
= list_entry(session
->s_caps
.next
,
1050 struct ceph_cap
, session_caps
);
1054 vino
= cap
->ci
->i_vino
;
1055 spin_unlock(&session
->s_cap_lock
);
1057 inode
= ceph_find_inode(sb
, vino
);
1060 spin_lock(&session
->s_cap_lock
);
1063 spin_unlock(&session
->s_cap_lock
);
1065 BUG_ON(session
->s_nr_caps
> 0);
1066 BUG_ON(!list_empty(&session
->s_cap_flushing
));
1067 cleanup_cap_releases(session
);
1071 * wake up any threads waiting on this session's caps. if the cap is
1072 * old (didn't get renewed on the client reconnect), remove it now.
1074 * caller must hold s_mutex.
1076 static int wake_up_session_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1079 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1081 wake_up_all(&ci
->i_cap_wq
);
1083 spin_lock(&ci
->i_ceph_lock
);
1084 ci
->i_wanted_max_size
= 0;
1085 ci
->i_requested_max_size
= 0;
1086 spin_unlock(&ci
->i_ceph_lock
);
1091 static void wake_up_session_caps(struct ceph_mds_session
*session
,
1094 dout("wake_up_session_caps %p mds%d\n", session
, session
->s_mds
);
1095 iterate_session_caps(session
, wake_up_session_cb
,
1096 (void *)(unsigned long)reconnect
);
1100 * Send periodic message to MDS renewing all currently held caps. The
1101 * ack will reset the expiration for all caps from this session.
1103 * caller holds s_mutex
1105 static int send_renew_caps(struct ceph_mds_client
*mdsc
,
1106 struct ceph_mds_session
*session
)
1108 struct ceph_msg
*msg
;
1111 if (time_after_eq(jiffies
, session
->s_cap_ttl
) &&
1112 time_after_eq(session
->s_cap_ttl
, session
->s_renew_requested
))
1113 pr_info("mds%d caps stale\n", session
->s_mds
);
1114 session
->s_renew_requested
= jiffies
;
1116 /* do not try to renew caps until a recovering mds has reconnected
1117 * with its clients. */
1118 state
= ceph_mdsmap_get_state(mdsc
->mdsmap
, session
->s_mds
);
1119 if (state
< CEPH_MDS_STATE_RECONNECT
) {
1120 dout("send_renew_caps ignoring mds%d (%s)\n",
1121 session
->s_mds
, ceph_mds_state_name(state
));
1125 dout("send_renew_caps to mds%d (%s)\n", session
->s_mds
,
1126 ceph_mds_state_name(state
));
1127 msg
= create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS
,
1128 ++session
->s_renew_seq
);
1131 ceph_con_send(&session
->s_con
, msg
);
1136 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1138 * Called under session->s_mutex
1140 static void renewed_caps(struct ceph_mds_client
*mdsc
,
1141 struct ceph_mds_session
*session
, int is_renew
)
1146 spin_lock(&session
->s_cap_lock
);
1147 was_stale
= is_renew
&& time_after_eq(jiffies
, session
->s_cap_ttl
);
1149 session
->s_cap_ttl
= session
->s_renew_requested
+
1150 mdsc
->mdsmap
->m_session_timeout
*HZ
;
1153 if (time_before(jiffies
, session
->s_cap_ttl
)) {
1154 pr_info("mds%d caps renewed\n", session
->s_mds
);
1157 pr_info("mds%d caps still stale\n", session
->s_mds
);
1160 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1161 session
->s_mds
, session
->s_cap_ttl
, was_stale
? "stale" : "fresh",
1162 time_before(jiffies
, session
->s_cap_ttl
) ? "stale" : "fresh");
1163 spin_unlock(&session
->s_cap_lock
);
1166 wake_up_session_caps(session
, 0);
1170 * send a session close request
1172 static int request_close_session(struct ceph_mds_client
*mdsc
,
1173 struct ceph_mds_session
*session
)
1175 struct ceph_msg
*msg
;
1177 dout("request_close_session mds%d state %s seq %lld\n",
1178 session
->s_mds
, session_state_name(session
->s_state
),
1180 msg
= create_session_msg(CEPH_SESSION_REQUEST_CLOSE
, session
->s_seq
);
1183 ceph_con_send(&session
->s_con
, msg
);
1188 * Called with s_mutex held.
1190 static int __close_session(struct ceph_mds_client
*mdsc
,
1191 struct ceph_mds_session
*session
)
1193 if (session
->s_state
>= CEPH_MDS_SESSION_CLOSING
)
1195 session
->s_state
= CEPH_MDS_SESSION_CLOSING
;
1196 return request_close_session(mdsc
, session
);
1200 * Trim old(er) caps.
1202 * Because we can't cache an inode without one or more caps, we do
1203 * this indirectly: if a cap is unused, we prune its aliases, at which
1204 * point the inode will hopefully get dropped to.
1206 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1207 * memory pressure from the MDS, though, so it needn't be perfect.
1209 static int trim_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
, void *arg
)
1211 struct ceph_mds_session
*session
= arg
;
1212 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1213 int used
, oissued
, mine
;
1215 if (session
->s_trim_caps
<= 0)
1218 spin_lock(&ci
->i_ceph_lock
);
1219 mine
= cap
->issued
| cap
->implemented
;
1220 used
= __ceph_caps_used(ci
);
1221 oissued
= __ceph_caps_issued_other(ci
, cap
);
1223 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1224 inode
, cap
, ceph_cap_string(mine
), ceph_cap_string(oissued
),
1225 ceph_cap_string(used
));
1226 if (ci
->i_dirty_caps
)
1227 goto out
; /* dirty caps */
1228 if ((used
& ~oissued
) & mine
)
1229 goto out
; /* we need these caps */
1231 session
->s_trim_caps
--;
1233 /* we aren't the only cap.. just remove us */
1234 __queue_cap_release(session
, ceph_ino(inode
), cap
->cap_id
,
1235 cap
->mseq
, cap
->issue_seq
);
1236 __ceph_remove_cap(cap
);
1238 /* try to drop referring dentries */
1239 spin_unlock(&ci
->i_ceph_lock
);
1240 d_prune_aliases(inode
);
1241 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1242 inode
, cap
, atomic_read(&inode
->i_count
));
1247 spin_unlock(&ci
->i_ceph_lock
);
1252 * Trim session cap count down to some max number.
1254 static int trim_caps(struct ceph_mds_client
*mdsc
,
1255 struct ceph_mds_session
*session
,
1258 int trim_caps
= session
->s_nr_caps
- max_caps
;
1260 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1261 session
->s_mds
, session
->s_nr_caps
, max_caps
, trim_caps
);
1262 if (trim_caps
> 0) {
1263 session
->s_trim_caps
= trim_caps
;
1264 iterate_session_caps(session
, trim_caps_cb
, session
);
1265 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1266 session
->s_mds
, session
->s_nr_caps
, max_caps
,
1267 trim_caps
- session
->s_trim_caps
);
1268 session
->s_trim_caps
= 0;
1274 * Allocate cap_release messages. If there is a partially full message
1275 * in the queue, try to allocate enough to cover it's remainder, so that
1276 * we can send it immediately.
1278 * Called under s_mutex.
1280 int ceph_add_cap_releases(struct ceph_mds_client
*mdsc
,
1281 struct ceph_mds_session
*session
)
1283 struct ceph_msg
*msg
, *partial
= NULL
;
1284 struct ceph_mds_cap_release
*head
;
1286 int extra
= mdsc
->fsc
->mount_options
->cap_release_safety
;
1289 dout("add_cap_releases %p mds%d extra %d\n", session
, session
->s_mds
,
1292 spin_lock(&session
->s_cap_lock
);
1294 if (!list_empty(&session
->s_cap_releases
)) {
1295 msg
= list_first_entry(&session
->s_cap_releases
,
1298 head
= msg
->front
.iov_base
;
1299 num
= le32_to_cpu(head
->num
);
1301 dout(" partial %p with (%d/%d)\n", msg
, num
,
1302 (int)CEPH_CAPS_PER_RELEASE
);
1303 extra
+= CEPH_CAPS_PER_RELEASE
- num
;
1307 while (session
->s_num_cap_releases
< session
->s_nr_caps
+ extra
) {
1308 spin_unlock(&session
->s_cap_lock
);
1309 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE
, PAGE_CACHE_SIZE
,
1313 dout("add_cap_releases %p msg %p now %d\n", session
, msg
,
1314 (int)msg
->front
.iov_len
);
1315 head
= msg
->front
.iov_base
;
1316 head
->num
= cpu_to_le32(0);
1317 msg
->front
.iov_len
= sizeof(*head
);
1318 spin_lock(&session
->s_cap_lock
);
1319 list_add(&msg
->list_head
, &session
->s_cap_releases
);
1320 session
->s_num_cap_releases
+= CEPH_CAPS_PER_RELEASE
;
1324 head
= partial
->front
.iov_base
;
1325 num
= le32_to_cpu(head
->num
);
1326 dout(" queueing partial %p with %d/%d\n", partial
, num
,
1327 (int)CEPH_CAPS_PER_RELEASE
);
1328 list_move_tail(&partial
->list_head
,
1329 &session
->s_cap_releases_done
);
1330 session
->s_num_cap_releases
-= CEPH_CAPS_PER_RELEASE
- num
;
1333 spin_unlock(&session
->s_cap_lock
);
1339 * flush all dirty inode data to disk.
1341 * returns true if we've flushed through want_flush_seq
1343 static int check_cap_flush(struct ceph_mds_client
*mdsc
, u64 want_flush_seq
)
1347 dout("check_cap_flush want %lld\n", want_flush_seq
);
1348 mutex_lock(&mdsc
->mutex
);
1349 for (mds
= 0; ret
&& mds
< mdsc
->max_sessions
; mds
++) {
1350 struct ceph_mds_session
*session
= mdsc
->sessions
[mds
];
1354 get_session(session
);
1355 mutex_unlock(&mdsc
->mutex
);
1357 mutex_lock(&session
->s_mutex
);
1358 if (!list_empty(&session
->s_cap_flushing
)) {
1359 struct ceph_inode_info
*ci
=
1360 list_entry(session
->s_cap_flushing
.next
,
1361 struct ceph_inode_info
,
1363 struct inode
*inode
= &ci
->vfs_inode
;
1365 spin_lock(&ci
->i_ceph_lock
);
1366 if (ci
->i_cap_flush_seq
<= want_flush_seq
) {
1367 dout("check_cap_flush still flushing %p "
1368 "seq %lld <= %lld to mds%d\n", inode
,
1369 ci
->i_cap_flush_seq
, want_flush_seq
,
1373 spin_unlock(&ci
->i_ceph_lock
);
1375 mutex_unlock(&session
->s_mutex
);
1376 ceph_put_mds_session(session
);
1380 mutex_lock(&mdsc
->mutex
);
1383 mutex_unlock(&mdsc
->mutex
);
1384 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq
);
1389 * called under s_mutex
1391 void ceph_send_cap_releases(struct ceph_mds_client
*mdsc
,
1392 struct ceph_mds_session
*session
)
1394 struct ceph_msg
*msg
;
1396 dout("send_cap_releases mds%d\n", session
->s_mds
);
1397 spin_lock(&session
->s_cap_lock
);
1398 while (!list_empty(&session
->s_cap_releases_done
)) {
1399 msg
= list_first_entry(&session
->s_cap_releases_done
,
1400 struct ceph_msg
, list_head
);
1401 list_del_init(&msg
->list_head
);
1402 spin_unlock(&session
->s_cap_lock
);
1403 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1404 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1405 ceph_con_send(&session
->s_con
, msg
);
1406 spin_lock(&session
->s_cap_lock
);
1408 spin_unlock(&session
->s_cap_lock
);
1411 static void discard_cap_releases(struct ceph_mds_client
*mdsc
,
1412 struct ceph_mds_session
*session
)
1414 struct ceph_msg
*msg
;
1415 struct ceph_mds_cap_release
*head
;
1418 dout("discard_cap_releases mds%d\n", session
->s_mds
);
1419 spin_lock(&session
->s_cap_lock
);
1421 /* zero out the in-progress message */
1422 msg
= list_first_entry(&session
->s_cap_releases
,
1423 struct ceph_msg
, list_head
);
1424 head
= msg
->front
.iov_base
;
1425 num
= le32_to_cpu(head
->num
);
1426 dout("discard_cap_releases mds%d %p %u\n", session
->s_mds
, msg
, num
);
1427 head
->num
= cpu_to_le32(0);
1428 msg
->front
.iov_len
= sizeof(*head
);
1429 session
->s_num_cap_releases
+= num
;
1431 /* requeue completed messages */
1432 while (!list_empty(&session
->s_cap_releases_done
)) {
1433 msg
= list_first_entry(&session
->s_cap_releases_done
,
1434 struct ceph_msg
, list_head
);
1435 list_del_init(&msg
->list_head
);
1437 head
= msg
->front
.iov_base
;
1438 num
= le32_to_cpu(head
->num
);
1439 dout("discard_cap_releases mds%d %p %u\n", session
->s_mds
, msg
,
1441 session
->s_num_cap_releases
+= num
;
1442 head
->num
= cpu_to_le32(0);
1443 msg
->front
.iov_len
= sizeof(*head
);
1444 list_add(&msg
->list_head
, &session
->s_cap_releases
);
1447 spin_unlock(&session
->s_cap_lock
);
1455 * Create an mds request.
1457 struct ceph_mds_request
*
1458 ceph_mdsc_create_request(struct ceph_mds_client
*mdsc
, int op
, int mode
)
1460 struct ceph_mds_request
*req
= kzalloc(sizeof(*req
), GFP_NOFS
);
1463 return ERR_PTR(-ENOMEM
);
1465 mutex_init(&req
->r_fill_mutex
);
1467 req
->r_started
= jiffies
;
1468 req
->r_resend_mds
= -1;
1469 INIT_LIST_HEAD(&req
->r_unsafe_dir_item
);
1471 kref_init(&req
->r_kref
);
1472 INIT_LIST_HEAD(&req
->r_wait
);
1473 init_completion(&req
->r_completion
);
1474 init_completion(&req
->r_safe_completion
);
1475 INIT_LIST_HEAD(&req
->r_unsafe_item
);
1478 req
->r_direct_mode
= mode
;
1483 * return oldest (lowest) request, tid in request tree, 0 if none.
1485 * called under mdsc->mutex.
1487 static struct ceph_mds_request
*__get_oldest_req(struct ceph_mds_client
*mdsc
)
1489 if (RB_EMPTY_ROOT(&mdsc
->request_tree
))
1491 return rb_entry(rb_first(&mdsc
->request_tree
),
1492 struct ceph_mds_request
, r_node
);
1495 static u64
__get_oldest_tid(struct ceph_mds_client
*mdsc
)
1497 struct ceph_mds_request
*req
= __get_oldest_req(mdsc
);
1505 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1506 * on build_path_from_dentry in fs/cifs/dir.c.
1508 * If @stop_on_nosnap, generate path relative to the first non-snapped
1511 * Encode hidden .snap dirs as a double /, i.e.
1512 * foo/.snap/bar -> foo//bar
1514 char *ceph_mdsc_build_path(struct dentry
*dentry
, int *plen
, u64
*base
,
1517 struct dentry
*temp
;
1523 return ERR_PTR(-EINVAL
);
1527 seq
= read_seqbegin(&rename_lock
);
1529 for (temp
= dentry
; !IS_ROOT(temp
);) {
1530 struct inode
*inode
= temp
->d_inode
;
1531 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
)
1532 len
++; /* slash only */
1533 else if (stop_on_nosnap
&& inode
&&
1534 ceph_snap(inode
) == CEPH_NOSNAP
)
1537 len
+= 1 + temp
->d_name
.len
;
1538 temp
= temp
->d_parent
;
1542 len
--; /* no leading '/' */
1544 path
= kmalloc(len
+1, GFP_NOFS
);
1546 return ERR_PTR(-ENOMEM
);
1548 path
[pos
] = 0; /* trailing null */
1550 for (temp
= dentry
; !IS_ROOT(temp
) && pos
!= 0; ) {
1551 struct inode
*inode
;
1553 spin_lock(&temp
->d_lock
);
1554 inode
= temp
->d_inode
;
1555 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
) {
1556 dout("build_path path+%d: %p SNAPDIR\n",
1558 } else if (stop_on_nosnap
&& inode
&&
1559 ceph_snap(inode
) == CEPH_NOSNAP
) {
1560 spin_unlock(&temp
->d_lock
);
1563 pos
-= temp
->d_name
.len
;
1565 spin_unlock(&temp
->d_lock
);
1568 strncpy(path
+ pos
, temp
->d_name
.name
,
1571 spin_unlock(&temp
->d_lock
);
1574 temp
= temp
->d_parent
;
1577 if (pos
!= 0 || read_seqretry(&rename_lock
, seq
)) {
1578 pr_err("build_path did not end path lookup where "
1579 "expected, namelen is %d, pos is %d\n", len
, pos
);
1580 /* presumably this is only possible if racing with a
1581 rename of one of the parent directories (we can not
1582 lock the dentries above us to prevent this, but
1583 retrying should be harmless) */
1588 *base
= ceph_ino(temp
->d_inode
);
1590 dout("build_path on %p %d built %llx '%.*s'\n",
1591 dentry
, d_count(dentry
), *base
, len
, path
);
1595 static int build_dentry_path(struct dentry
*dentry
,
1596 const char **ppath
, int *ppathlen
, u64
*pino
,
1601 if (ceph_snap(dentry
->d_parent
->d_inode
) == CEPH_NOSNAP
) {
1602 *pino
= ceph_ino(dentry
->d_parent
->d_inode
);
1603 *ppath
= dentry
->d_name
.name
;
1604 *ppathlen
= dentry
->d_name
.len
;
1607 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1609 return PTR_ERR(path
);
1615 static int build_inode_path(struct inode
*inode
,
1616 const char **ppath
, int *ppathlen
, u64
*pino
,
1619 struct dentry
*dentry
;
1622 if (ceph_snap(inode
) == CEPH_NOSNAP
) {
1623 *pino
= ceph_ino(inode
);
1627 dentry
= d_find_alias(inode
);
1628 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1631 return PTR_ERR(path
);
1638 * request arguments may be specified via an inode *, a dentry *, or
1639 * an explicit ino+path.
1641 static int set_request_path_attr(struct inode
*rinode
, struct dentry
*rdentry
,
1642 const char *rpath
, u64 rino
,
1643 const char **ppath
, int *pathlen
,
1644 u64
*ino
, int *freepath
)
1649 r
= build_inode_path(rinode
, ppath
, pathlen
, ino
, freepath
);
1650 dout(" inode %p %llx.%llx\n", rinode
, ceph_ino(rinode
),
1652 } else if (rdentry
) {
1653 r
= build_dentry_path(rdentry
, ppath
, pathlen
, ino
, freepath
);
1654 dout(" dentry %p %llx/%.*s\n", rdentry
, *ino
, *pathlen
,
1656 } else if (rpath
|| rino
) {
1659 *pathlen
= rpath
? strlen(rpath
) : 0;
1660 dout(" path %.*s\n", *pathlen
, rpath
);
1667 * called under mdsc->mutex
1669 static struct ceph_msg
*create_request_message(struct ceph_mds_client
*mdsc
,
1670 struct ceph_mds_request
*req
,
1673 struct ceph_msg
*msg
;
1674 struct ceph_mds_request_head
*head
;
1675 const char *path1
= NULL
;
1676 const char *path2
= NULL
;
1677 u64 ino1
= 0, ino2
= 0;
1678 int pathlen1
= 0, pathlen2
= 0;
1679 int freepath1
= 0, freepath2
= 0;
1685 ret
= set_request_path_attr(req
->r_inode
, req
->r_dentry
,
1686 req
->r_path1
, req
->r_ino1
.ino
,
1687 &path1
, &pathlen1
, &ino1
, &freepath1
);
1693 ret
= set_request_path_attr(NULL
, req
->r_old_dentry
,
1694 req
->r_path2
, req
->r_ino2
.ino
,
1695 &path2
, &pathlen2
, &ino2
, &freepath2
);
1701 len
= sizeof(*head
) +
1702 pathlen1
+ pathlen2
+ 2*(1 + sizeof(u32
) + sizeof(u64
));
1704 /* calculate (max) length for cap releases */
1705 len
+= sizeof(struct ceph_mds_request_release
) *
1706 (!!req
->r_inode_drop
+ !!req
->r_dentry_drop
+
1707 !!req
->r_old_inode_drop
+ !!req
->r_old_dentry_drop
);
1708 if (req
->r_dentry_drop
)
1709 len
+= req
->r_dentry
->d_name
.len
;
1710 if (req
->r_old_dentry_drop
)
1711 len
+= req
->r_old_dentry
->d_name
.len
;
1713 msg
= ceph_msg_new(CEPH_MSG_CLIENT_REQUEST
, len
, GFP_NOFS
, false);
1715 msg
= ERR_PTR(-ENOMEM
);
1719 msg
->hdr
.tid
= cpu_to_le64(req
->r_tid
);
1721 head
= msg
->front
.iov_base
;
1722 p
= msg
->front
.iov_base
+ sizeof(*head
);
1723 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1725 head
->mdsmap_epoch
= cpu_to_le32(mdsc
->mdsmap
->m_epoch
);
1726 head
->op
= cpu_to_le32(req
->r_op
);
1727 head
->caller_uid
= cpu_to_le32(from_kuid(&init_user_ns
, req
->r_uid
));
1728 head
->caller_gid
= cpu_to_le32(from_kgid(&init_user_ns
, req
->r_gid
));
1729 head
->args
= req
->r_args
;
1731 ceph_encode_filepath(&p
, end
, ino1
, path1
);
1732 ceph_encode_filepath(&p
, end
, ino2
, path2
);
1734 /* make note of release offset, in case we need to replay */
1735 req
->r_request_release_offset
= p
- msg
->front
.iov_base
;
1739 if (req
->r_inode_drop
)
1740 releases
+= ceph_encode_inode_release(&p
,
1741 req
->r_inode
? req
->r_inode
: req
->r_dentry
->d_inode
,
1742 mds
, req
->r_inode_drop
, req
->r_inode_unless
, 0);
1743 if (req
->r_dentry_drop
)
1744 releases
+= ceph_encode_dentry_release(&p
, req
->r_dentry
,
1745 mds
, req
->r_dentry_drop
, req
->r_dentry_unless
);
1746 if (req
->r_old_dentry_drop
)
1747 releases
+= ceph_encode_dentry_release(&p
, req
->r_old_dentry
,
1748 mds
, req
->r_old_dentry_drop
, req
->r_old_dentry_unless
);
1749 if (req
->r_old_inode_drop
)
1750 releases
+= ceph_encode_inode_release(&p
,
1751 req
->r_old_dentry
->d_inode
,
1752 mds
, req
->r_old_inode_drop
, req
->r_old_inode_unless
, 0);
1753 head
->num_releases
= cpu_to_le16(releases
);
1756 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
1757 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1759 if (req
->r_data_len
) {
1760 /* outbound data set only by ceph_sync_setxattr() */
1761 BUG_ON(!req
->r_pages
);
1762 ceph_msg_data_add_pages(msg
, req
->r_pages
, req
->r_data_len
, 0);
1765 msg
->hdr
.data_len
= cpu_to_le32(req
->r_data_len
);
1766 msg
->hdr
.data_off
= cpu_to_le16(0);
1770 kfree((char *)path2
);
1773 kfree((char *)path1
);
1779 * called under mdsc->mutex if error, under no mutex if
1782 static void complete_request(struct ceph_mds_client
*mdsc
,
1783 struct ceph_mds_request
*req
)
1785 if (req
->r_callback
)
1786 req
->r_callback(mdsc
, req
);
1788 complete_all(&req
->r_completion
);
1792 * called under mdsc->mutex
1794 static int __prepare_send_request(struct ceph_mds_client
*mdsc
,
1795 struct ceph_mds_request
*req
,
1798 struct ceph_mds_request_head
*rhead
;
1799 struct ceph_msg
*msg
;
1804 struct ceph_cap
*cap
=
1805 ceph_get_cap_for_mds(ceph_inode(req
->r_inode
), mds
);
1808 req
->r_sent_on_mseq
= cap
->mseq
;
1810 req
->r_sent_on_mseq
= -1;
1812 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req
,
1813 req
->r_tid
, ceph_mds_op_name(req
->r_op
), req
->r_attempts
);
1815 if (req
->r_got_unsafe
) {
1817 * Replay. Do not regenerate message (and rebuild
1818 * paths, etc.); just use the original message.
1819 * Rebuilding paths will break for renames because
1820 * d_move mangles the src name.
1822 msg
= req
->r_request
;
1823 rhead
= msg
->front
.iov_base
;
1825 flags
= le32_to_cpu(rhead
->flags
);
1826 flags
|= CEPH_MDS_FLAG_REPLAY
;
1827 rhead
->flags
= cpu_to_le32(flags
);
1829 if (req
->r_target_inode
)
1830 rhead
->ino
= cpu_to_le64(ceph_ino(req
->r_target_inode
));
1832 rhead
->num_retry
= req
->r_attempts
- 1;
1834 /* remove cap/dentry releases from message */
1835 rhead
->num_releases
= 0;
1836 msg
->hdr
.front_len
= cpu_to_le32(req
->r_request_release_offset
);
1837 msg
->front
.iov_len
= req
->r_request_release_offset
;
1841 if (req
->r_request
) {
1842 ceph_msg_put(req
->r_request
);
1843 req
->r_request
= NULL
;
1845 msg
= create_request_message(mdsc
, req
, mds
);
1847 req
->r_err
= PTR_ERR(msg
);
1848 complete_request(mdsc
, req
);
1849 return PTR_ERR(msg
);
1851 req
->r_request
= msg
;
1853 rhead
= msg
->front
.iov_base
;
1854 rhead
->oldest_client_tid
= cpu_to_le64(__get_oldest_tid(mdsc
));
1855 if (req
->r_got_unsafe
)
1856 flags
|= CEPH_MDS_FLAG_REPLAY
;
1857 if (req
->r_locked_dir
)
1858 flags
|= CEPH_MDS_FLAG_WANT_DENTRY
;
1859 rhead
->flags
= cpu_to_le32(flags
);
1860 rhead
->num_fwd
= req
->r_num_fwd
;
1861 rhead
->num_retry
= req
->r_attempts
- 1;
1864 dout(" r_locked_dir = %p\n", req
->r_locked_dir
);
1869 * send request, or put it on the appropriate wait list.
1871 static int __do_request(struct ceph_mds_client
*mdsc
,
1872 struct ceph_mds_request
*req
)
1874 struct ceph_mds_session
*session
= NULL
;
1878 if (req
->r_err
|| req
->r_got_result
)
1881 if (req
->r_timeout
&&
1882 time_after_eq(jiffies
, req
->r_started
+ req
->r_timeout
)) {
1883 dout("do_request timed out\n");
1888 put_request_session(req
);
1890 mds
= __choose_mds(mdsc
, req
);
1892 ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) < CEPH_MDS_STATE_ACTIVE
) {
1893 dout("do_request no mds or not active, waiting for map\n");
1894 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
1898 /* get, open session */
1899 session
= __ceph_lookup_mds_session(mdsc
, mds
);
1901 session
= register_session(mdsc
, mds
);
1902 if (IS_ERR(session
)) {
1903 err
= PTR_ERR(session
);
1907 req
->r_session
= get_session(session
);
1909 dout("do_request mds%d session %p state %s\n", mds
, session
,
1910 session_state_name(session
->s_state
));
1911 if (session
->s_state
!= CEPH_MDS_SESSION_OPEN
&&
1912 session
->s_state
!= CEPH_MDS_SESSION_HUNG
) {
1913 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
1914 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
1915 __open_session(mdsc
, session
);
1916 list_add(&req
->r_wait
, &session
->s_waiting
);
1921 req
->r_resend_mds
= -1; /* forget any previous mds hint */
1923 if (req
->r_request_started
== 0) /* note request start time */
1924 req
->r_request_started
= jiffies
;
1926 err
= __prepare_send_request(mdsc
, req
, mds
);
1928 ceph_msg_get(req
->r_request
);
1929 ceph_con_send(&session
->s_con
, req
->r_request
);
1933 ceph_put_mds_session(session
);
1939 complete_request(mdsc
, req
);
1944 * called under mdsc->mutex
1946 static void __wake_requests(struct ceph_mds_client
*mdsc
,
1947 struct list_head
*head
)
1949 struct ceph_mds_request
*req
;
1950 LIST_HEAD(tmp_list
);
1952 list_splice_init(head
, &tmp_list
);
1954 while (!list_empty(&tmp_list
)) {
1955 req
= list_entry(tmp_list
.next
,
1956 struct ceph_mds_request
, r_wait
);
1957 list_del_init(&req
->r_wait
);
1958 dout(" wake request %p tid %llu\n", req
, req
->r_tid
);
1959 __do_request(mdsc
, req
);
1964 * Wake up threads with requests pending for @mds, so that they can
1965 * resubmit their requests to a possibly different mds.
1967 static void kick_requests(struct ceph_mds_client
*mdsc
, int mds
)
1969 struct ceph_mds_request
*req
;
1972 dout("kick_requests mds%d\n", mds
);
1973 for (p
= rb_first(&mdsc
->request_tree
); p
; p
= rb_next(p
)) {
1974 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
1975 if (req
->r_got_unsafe
)
1977 if (req
->r_session
&&
1978 req
->r_session
->s_mds
== mds
) {
1979 dout(" kicking tid %llu\n", req
->r_tid
);
1980 __do_request(mdsc
, req
);
1985 void ceph_mdsc_submit_request(struct ceph_mds_client
*mdsc
,
1986 struct ceph_mds_request
*req
)
1988 dout("submit_request on %p\n", req
);
1989 mutex_lock(&mdsc
->mutex
);
1990 __register_request(mdsc
, req
, NULL
);
1991 __do_request(mdsc
, req
);
1992 mutex_unlock(&mdsc
->mutex
);
1996 * Synchrously perform an mds request. Take care of all of the
1997 * session setup, forwarding, retry details.
1999 int ceph_mdsc_do_request(struct ceph_mds_client
*mdsc
,
2001 struct ceph_mds_request
*req
)
2005 dout("do_request on %p\n", req
);
2007 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2009 ceph_get_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
2010 if (req
->r_locked_dir
)
2011 ceph_get_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
2012 if (req
->r_old_dentry
)
2013 ceph_get_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
2017 mutex_lock(&mdsc
->mutex
);
2018 __register_request(mdsc
, req
, dir
);
2019 __do_request(mdsc
, req
);
2023 __unregister_request(mdsc
, req
);
2024 dout("do_request early error %d\n", err
);
2029 mutex_unlock(&mdsc
->mutex
);
2030 dout("do_request waiting\n");
2031 if (req
->r_timeout
) {
2032 err
= (long)wait_for_completion_killable_timeout(
2033 &req
->r_completion
, req
->r_timeout
);
2037 err
= wait_for_completion_killable(&req
->r_completion
);
2039 dout("do_request waited, got %d\n", err
);
2040 mutex_lock(&mdsc
->mutex
);
2042 /* only abort if we didn't race with a real reply */
2043 if (req
->r_got_result
) {
2044 err
= le32_to_cpu(req
->r_reply_info
.head
->result
);
2045 } else if (err
< 0) {
2046 dout("aborted request %lld with %d\n", req
->r_tid
, err
);
2049 * ensure we aren't running concurrently with
2050 * ceph_fill_trace or ceph_readdir_prepopulate, which
2051 * rely on locks (dir mutex) held by our caller.
2053 mutex_lock(&req
->r_fill_mutex
);
2055 req
->r_aborted
= true;
2056 mutex_unlock(&req
->r_fill_mutex
);
2058 if (req
->r_locked_dir
&&
2059 (req
->r_op
& CEPH_MDS_OP_WRITE
))
2060 ceph_invalidate_dir_request(req
);
2066 mutex_unlock(&mdsc
->mutex
);
2067 dout("do_request %p done, result %d\n", req
, err
);
2072 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2073 * namespace request.
2075 void ceph_invalidate_dir_request(struct ceph_mds_request
*req
)
2077 struct inode
*inode
= req
->r_locked_dir
;
2079 dout("invalidate_dir_request %p (complete, lease(s))\n", inode
);
2081 ceph_dir_clear_complete(inode
);
2083 ceph_invalidate_dentry_lease(req
->r_dentry
);
2084 if (req
->r_old_dentry
)
2085 ceph_invalidate_dentry_lease(req
->r_old_dentry
);
2091 * We take the session mutex and parse and process the reply immediately.
2092 * This preserves the logical ordering of replies, capabilities, etc., sent
2093 * by the MDS as they are applied to our local cache.
2095 static void handle_reply(struct ceph_mds_session
*session
, struct ceph_msg
*msg
)
2097 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2098 struct ceph_mds_request
*req
;
2099 struct ceph_mds_reply_head
*head
= msg
->front
.iov_base
;
2100 struct ceph_mds_reply_info_parsed
*rinfo
; /* parsed reply info */
2103 int mds
= session
->s_mds
;
2105 if (msg
->front
.iov_len
< sizeof(*head
)) {
2106 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2111 /* get request, session */
2112 tid
= le64_to_cpu(msg
->hdr
.tid
);
2113 mutex_lock(&mdsc
->mutex
);
2114 req
= __lookup_request(mdsc
, tid
);
2116 dout("handle_reply on unknown tid %llu\n", tid
);
2117 mutex_unlock(&mdsc
->mutex
);
2120 dout("handle_reply %p\n", req
);
2122 /* correct session? */
2123 if (req
->r_session
!= session
) {
2124 pr_err("mdsc_handle_reply got %llu on session mds%d"
2125 " not mds%d\n", tid
, session
->s_mds
,
2126 req
->r_session
? req
->r_session
->s_mds
: -1);
2127 mutex_unlock(&mdsc
->mutex
);
2132 if ((req
->r_got_unsafe
&& !head
->safe
) ||
2133 (req
->r_got_safe
&& head
->safe
)) {
2134 pr_warning("got a dup %s reply on %llu from mds%d\n",
2135 head
->safe
? "safe" : "unsafe", tid
, mds
);
2136 mutex_unlock(&mdsc
->mutex
);
2139 if (req
->r_got_safe
&& !head
->safe
) {
2140 pr_warning("got unsafe after safe on %llu from mds%d\n",
2142 mutex_unlock(&mdsc
->mutex
);
2146 result
= le32_to_cpu(head
->result
);
2150 * if we're not talking to the authority, send to them
2151 * if the authority has changed while we weren't looking,
2152 * send to new authority
2153 * Otherwise we just have to return an ESTALE
2155 if (result
== -ESTALE
) {
2156 dout("got ESTALE on request %llu", req
->r_tid
);
2157 if (!req
->r_inode
) {
2158 /* do nothing; not an authority problem */
2159 } else if (req
->r_direct_mode
!= USE_AUTH_MDS
) {
2160 dout("not using auth, setting for that now");
2161 req
->r_direct_mode
= USE_AUTH_MDS
;
2162 __do_request(mdsc
, req
);
2163 mutex_unlock(&mdsc
->mutex
);
2166 struct ceph_inode_info
*ci
= ceph_inode(req
->r_inode
);
2167 struct ceph_cap
*cap
= NULL
;
2170 cap
= ceph_get_cap_for_mds(ci
,
2171 req
->r_session
->s_mds
);
2173 dout("already using auth");
2174 if ((!cap
|| cap
!= ci
->i_auth_cap
) ||
2175 (cap
->mseq
!= req
->r_sent_on_mseq
)) {
2176 dout("but cap changed, so resending");
2177 __do_request(mdsc
, req
);
2178 mutex_unlock(&mdsc
->mutex
);
2182 dout("have to return ESTALE on request %llu", req
->r_tid
);
2187 req
->r_got_safe
= true;
2188 __unregister_request(mdsc
, req
);
2189 complete_all(&req
->r_safe_completion
);
2191 if (req
->r_got_unsafe
) {
2193 * We already handled the unsafe response, now do the
2194 * cleanup. No need to examine the response; the MDS
2195 * doesn't include any result info in the safe
2196 * response. And even if it did, there is nothing
2197 * useful we could do with a revised return value.
2199 dout("got safe reply %llu, mds%d\n", tid
, mds
);
2200 list_del_init(&req
->r_unsafe_item
);
2202 /* last unsafe request during umount? */
2203 if (mdsc
->stopping
&& !__get_oldest_req(mdsc
))
2204 complete_all(&mdsc
->safe_umount_waiters
);
2205 mutex_unlock(&mdsc
->mutex
);
2209 req
->r_got_unsafe
= true;
2210 list_add_tail(&req
->r_unsafe_item
, &req
->r_session
->s_unsafe
);
2213 dout("handle_reply tid %lld result %d\n", tid
, result
);
2214 rinfo
= &req
->r_reply_info
;
2215 err
= parse_reply_info(msg
, rinfo
, session
->s_con
.peer_features
);
2216 mutex_unlock(&mdsc
->mutex
);
2218 mutex_lock(&session
->s_mutex
);
2220 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds
, tid
);
2226 if (rinfo
->snapblob_len
) {
2227 down_write(&mdsc
->snap_rwsem
);
2228 ceph_update_snap_trace(mdsc
, rinfo
->snapblob
,
2229 rinfo
->snapblob
+ rinfo
->snapblob_len
,
2230 le32_to_cpu(head
->op
) == CEPH_MDS_OP_RMSNAP
);
2231 downgrade_write(&mdsc
->snap_rwsem
);
2233 down_read(&mdsc
->snap_rwsem
);
2236 /* insert trace into our cache */
2237 mutex_lock(&req
->r_fill_mutex
);
2238 err
= ceph_fill_trace(mdsc
->fsc
->sb
, req
, req
->r_session
);
2240 if (result
== 0 && (req
->r_op
== CEPH_MDS_OP_READDIR
||
2241 req
->r_op
== CEPH_MDS_OP_LSSNAP
) &&
2243 ceph_readdir_prepopulate(req
, req
->r_session
);
2244 ceph_unreserve_caps(mdsc
, &req
->r_caps_reservation
);
2246 mutex_unlock(&req
->r_fill_mutex
);
2248 up_read(&mdsc
->snap_rwsem
);
2250 mutex_lock(&mdsc
->mutex
);
2251 if (!req
->r_aborted
) {
2257 req
->r_got_result
= true;
2260 dout("reply arrived after request %lld was aborted\n", tid
);
2262 mutex_unlock(&mdsc
->mutex
);
2264 ceph_add_cap_releases(mdsc
, req
->r_session
);
2265 mutex_unlock(&session
->s_mutex
);
2267 /* kick calling process */
2268 complete_request(mdsc
, req
);
2270 ceph_mdsc_put_request(req
);
2277 * handle mds notification that our request has been forwarded.
2279 static void handle_forward(struct ceph_mds_client
*mdsc
,
2280 struct ceph_mds_session
*session
,
2281 struct ceph_msg
*msg
)
2283 struct ceph_mds_request
*req
;
2284 u64 tid
= le64_to_cpu(msg
->hdr
.tid
);
2288 void *p
= msg
->front
.iov_base
;
2289 void *end
= p
+ msg
->front
.iov_len
;
2291 ceph_decode_need(&p
, end
, 2*sizeof(u32
), bad
);
2292 next_mds
= ceph_decode_32(&p
);
2293 fwd_seq
= ceph_decode_32(&p
);
2295 mutex_lock(&mdsc
->mutex
);
2296 req
= __lookup_request(mdsc
, tid
);
2298 dout("forward tid %llu to mds%d - req dne\n", tid
, next_mds
);
2299 goto out
; /* dup reply? */
2302 if (req
->r_aborted
) {
2303 dout("forward tid %llu aborted, unregistering\n", tid
);
2304 __unregister_request(mdsc
, req
);
2305 } else if (fwd_seq
<= req
->r_num_fwd
) {
2306 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2307 tid
, next_mds
, req
->r_num_fwd
, fwd_seq
);
2309 /* resend. forward race not possible; mds would drop */
2310 dout("forward tid %llu to mds%d (we resend)\n", tid
, next_mds
);
2312 BUG_ON(req
->r_got_result
);
2313 req
->r_num_fwd
= fwd_seq
;
2314 req
->r_resend_mds
= next_mds
;
2315 put_request_session(req
);
2316 __do_request(mdsc
, req
);
2318 ceph_mdsc_put_request(req
);
2320 mutex_unlock(&mdsc
->mutex
);
2324 pr_err("mdsc_handle_forward decode error err=%d\n", err
);
2328 * handle a mds session control message
2330 static void handle_session(struct ceph_mds_session
*session
,
2331 struct ceph_msg
*msg
)
2333 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2336 int mds
= session
->s_mds
;
2337 struct ceph_mds_session_head
*h
= msg
->front
.iov_base
;
2341 if (msg
->front
.iov_len
!= sizeof(*h
))
2343 op
= le32_to_cpu(h
->op
);
2344 seq
= le64_to_cpu(h
->seq
);
2346 mutex_lock(&mdsc
->mutex
);
2347 if (op
== CEPH_SESSION_CLOSE
)
2348 __unregister_session(mdsc
, session
);
2349 /* FIXME: this ttl calculation is generous */
2350 session
->s_ttl
= jiffies
+ HZ
*mdsc
->mdsmap
->m_session_autoclose
;
2351 mutex_unlock(&mdsc
->mutex
);
2353 mutex_lock(&session
->s_mutex
);
2355 dout("handle_session mds%d %s %p state %s seq %llu\n",
2356 mds
, ceph_session_op_name(op
), session
,
2357 session_state_name(session
->s_state
), seq
);
2359 if (session
->s_state
== CEPH_MDS_SESSION_HUNG
) {
2360 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2361 pr_info("mds%d came back\n", session
->s_mds
);
2365 case CEPH_SESSION_OPEN
:
2366 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2367 pr_info("mds%d reconnect success\n", session
->s_mds
);
2368 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2369 renewed_caps(mdsc
, session
, 0);
2372 __close_session(mdsc
, session
);
2375 case CEPH_SESSION_RENEWCAPS
:
2376 if (session
->s_renew_seq
== seq
)
2377 renewed_caps(mdsc
, session
, 1);
2380 case CEPH_SESSION_CLOSE
:
2381 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2382 pr_info("mds%d reconnect denied\n", session
->s_mds
);
2383 remove_session_caps(session
);
2384 wake
= 1; /* for good measure */
2385 wake_up_all(&mdsc
->session_close_wq
);
2386 kick_requests(mdsc
, mds
);
2389 case CEPH_SESSION_STALE
:
2390 pr_info("mds%d caps went stale, renewing\n",
2392 spin_lock(&session
->s_gen_ttl_lock
);
2393 session
->s_cap_gen
++;
2394 session
->s_cap_ttl
= jiffies
- 1;
2395 spin_unlock(&session
->s_gen_ttl_lock
);
2396 send_renew_caps(mdsc
, session
);
2399 case CEPH_SESSION_RECALL_STATE
:
2400 trim_caps(mdsc
, session
, le32_to_cpu(h
->max_caps
));
2404 pr_err("mdsc_handle_session bad op %d mds%d\n", op
, mds
);
2408 mutex_unlock(&session
->s_mutex
);
2410 mutex_lock(&mdsc
->mutex
);
2411 __wake_requests(mdsc
, &session
->s_waiting
);
2412 mutex_unlock(&mdsc
->mutex
);
2417 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds
,
2418 (int)msg
->front
.iov_len
);
2425 * called under session->mutex.
2427 static void replay_unsafe_requests(struct ceph_mds_client
*mdsc
,
2428 struct ceph_mds_session
*session
)
2430 struct ceph_mds_request
*req
, *nreq
;
2433 dout("replay_unsafe_requests mds%d\n", session
->s_mds
);
2435 mutex_lock(&mdsc
->mutex
);
2436 list_for_each_entry_safe(req
, nreq
, &session
->s_unsafe
, r_unsafe_item
) {
2437 err
= __prepare_send_request(mdsc
, req
, session
->s_mds
);
2439 ceph_msg_get(req
->r_request
);
2440 ceph_con_send(&session
->s_con
, req
->r_request
);
2443 mutex_unlock(&mdsc
->mutex
);
2447 * Encode information about a cap for a reconnect with the MDS.
2449 static int encode_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
2453 struct ceph_mds_cap_reconnect v2
;
2454 struct ceph_mds_cap_reconnect_v1 v1
;
2457 struct ceph_inode_info
*ci
;
2458 struct ceph_reconnect_state
*recon_state
= arg
;
2459 struct ceph_pagelist
*pagelist
= recon_state
->pagelist
;
2463 struct dentry
*dentry
;
2467 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2468 inode
, ceph_vinop(inode
), cap
, cap
->cap_id
,
2469 ceph_cap_string(cap
->issued
));
2470 err
= ceph_pagelist_encode_64(pagelist
, ceph_ino(inode
));
2474 dentry
= d_find_alias(inode
);
2476 path
= ceph_mdsc_build_path(dentry
, &pathlen
, &pathbase
, 0);
2478 err
= PTR_ERR(path
);
2485 err
= ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
2489 spin_lock(&ci
->i_ceph_lock
);
2490 cap
->seq
= 0; /* reset cap seq */
2491 cap
->issue_seq
= 0; /* and issue_seq */
2492 cap
->mseq
= 0; /* and migrate_seq */
2494 if (recon_state
->flock
) {
2495 rec
.v2
.cap_id
= cpu_to_le64(cap
->cap_id
);
2496 rec
.v2
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2497 rec
.v2
.issued
= cpu_to_le32(cap
->issued
);
2498 rec
.v2
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2499 rec
.v2
.pathbase
= cpu_to_le64(pathbase
);
2500 rec
.v2
.flock_len
= 0;
2501 reclen
= sizeof(rec
.v2
);
2503 rec
.v1
.cap_id
= cpu_to_le64(cap
->cap_id
);
2504 rec
.v1
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2505 rec
.v1
.issued
= cpu_to_le32(cap
->issued
);
2506 rec
.v1
.size
= cpu_to_le64(inode
->i_size
);
2507 ceph_encode_timespec(&rec
.v1
.mtime
, &inode
->i_mtime
);
2508 ceph_encode_timespec(&rec
.v1
.atime
, &inode
->i_atime
);
2509 rec
.v1
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2510 rec
.v1
.pathbase
= cpu_to_le64(pathbase
);
2511 reclen
= sizeof(rec
.v1
);
2513 spin_unlock(&ci
->i_ceph_lock
);
2515 if (recon_state
->flock
) {
2516 int num_fcntl_locks
, num_flock_locks
;
2517 struct ceph_filelock
*flocks
;
2520 spin_lock(&inode
->i_lock
);
2521 ceph_count_locks(inode
, &num_fcntl_locks
, &num_flock_locks
);
2522 spin_unlock(&inode
->i_lock
);
2523 flocks
= kmalloc((num_fcntl_locks
+num_flock_locks
) *
2524 sizeof(struct ceph_filelock
), GFP_NOFS
);
2529 spin_lock(&inode
->i_lock
);
2530 err
= ceph_encode_locks_to_buffer(inode
, flocks
,
2533 spin_unlock(&inode
->i_lock
);
2541 * number of encoded locks is stable, so copy to pagelist
2543 rec
.v2
.flock_len
= cpu_to_le32(2*sizeof(u32
) +
2544 (num_fcntl_locks
+num_flock_locks
) *
2545 sizeof(struct ceph_filelock
));
2546 err
= ceph_pagelist_append(pagelist
, &rec
, reclen
);
2548 err
= ceph_locks_to_pagelist(flocks
, pagelist
,
2553 err
= ceph_pagelist_append(pagelist
, &rec
, reclen
);
2564 * If an MDS fails and recovers, clients need to reconnect in order to
2565 * reestablish shared state. This includes all caps issued through
2566 * this session _and_ the snap_realm hierarchy. Because it's not
2567 * clear which snap realms the mds cares about, we send everything we
2568 * know about.. that ensures we'll then get any new info the
2569 * recovering MDS might have.
2571 * This is a relatively heavyweight operation, but it's rare.
2573 * called with mdsc->mutex held.
2575 static void send_mds_reconnect(struct ceph_mds_client
*mdsc
,
2576 struct ceph_mds_session
*session
)
2578 struct ceph_msg
*reply
;
2580 int mds
= session
->s_mds
;
2582 struct ceph_pagelist
*pagelist
;
2583 struct ceph_reconnect_state recon_state
;
2585 pr_info("mds%d reconnect start\n", mds
);
2587 pagelist
= kmalloc(sizeof(*pagelist
), GFP_NOFS
);
2589 goto fail_nopagelist
;
2590 ceph_pagelist_init(pagelist
);
2592 reply
= ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT
, 0, GFP_NOFS
, false);
2596 mutex_lock(&session
->s_mutex
);
2597 session
->s_state
= CEPH_MDS_SESSION_RECONNECTING
;
2600 ceph_con_close(&session
->s_con
);
2601 ceph_con_open(&session
->s_con
,
2602 CEPH_ENTITY_TYPE_MDS
, mds
,
2603 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
2605 /* replay unsafe requests */
2606 replay_unsafe_requests(mdsc
, session
);
2608 down_read(&mdsc
->snap_rwsem
);
2610 dout("session %p state %s\n", session
,
2611 session_state_name(session
->s_state
));
2613 /* drop old cap expires; we're about to reestablish that state */
2614 discard_cap_releases(mdsc
, session
);
2616 /* traverse this session's caps */
2617 err
= ceph_pagelist_encode_32(pagelist
, session
->s_nr_caps
);
2621 recon_state
.pagelist
= pagelist
;
2622 recon_state
.flock
= session
->s_con
.peer_features
& CEPH_FEATURE_FLOCK
;
2623 err
= iterate_session_caps(session
, encode_caps_cb
, &recon_state
);
2628 * snaprealms. we provide mds with the ino, seq (version), and
2629 * parent for all of our realms. If the mds has any newer info,
2632 for (p
= rb_first(&mdsc
->snap_realms
); p
; p
= rb_next(p
)) {
2633 struct ceph_snap_realm
*realm
=
2634 rb_entry(p
, struct ceph_snap_realm
, node
);
2635 struct ceph_mds_snaprealm_reconnect sr_rec
;
2637 dout(" adding snap realm %llx seq %lld parent %llx\n",
2638 realm
->ino
, realm
->seq
, realm
->parent_ino
);
2639 sr_rec
.ino
= cpu_to_le64(realm
->ino
);
2640 sr_rec
.seq
= cpu_to_le64(realm
->seq
);
2641 sr_rec
.parent
= cpu_to_le64(realm
->parent_ino
);
2642 err
= ceph_pagelist_append(pagelist
, &sr_rec
, sizeof(sr_rec
));
2647 if (recon_state
.flock
)
2648 reply
->hdr
.version
= cpu_to_le16(2);
2649 if (pagelist
->length
) {
2650 /* set up outbound data if we have any */
2651 reply
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
2652 ceph_msg_data_add_pagelist(reply
, pagelist
);
2654 ceph_con_send(&session
->s_con
, reply
);
2656 mutex_unlock(&session
->s_mutex
);
2658 mutex_lock(&mdsc
->mutex
);
2659 __wake_requests(mdsc
, &session
->s_waiting
);
2660 mutex_unlock(&mdsc
->mutex
);
2662 up_read(&mdsc
->snap_rwsem
);
2666 ceph_msg_put(reply
);
2667 up_read(&mdsc
->snap_rwsem
);
2668 mutex_unlock(&session
->s_mutex
);
2670 ceph_pagelist_release(pagelist
);
2673 pr_err("error %d preparing reconnect for mds%d\n", err
, mds
);
2679 * compare old and new mdsmaps, kicking requests
2680 * and closing out old connections as necessary
2682 * called under mdsc->mutex.
2684 static void check_new_map(struct ceph_mds_client
*mdsc
,
2685 struct ceph_mdsmap
*newmap
,
2686 struct ceph_mdsmap
*oldmap
)
2689 int oldstate
, newstate
;
2690 struct ceph_mds_session
*s
;
2692 dout("check_new_map new %u old %u\n",
2693 newmap
->m_epoch
, oldmap
->m_epoch
);
2695 for (i
= 0; i
< oldmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
2696 if (mdsc
->sessions
[i
] == NULL
)
2698 s
= mdsc
->sessions
[i
];
2699 oldstate
= ceph_mdsmap_get_state(oldmap
, i
);
2700 newstate
= ceph_mdsmap_get_state(newmap
, i
);
2702 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2703 i
, ceph_mds_state_name(oldstate
),
2704 ceph_mdsmap_is_laggy(oldmap
, i
) ? " (laggy)" : "",
2705 ceph_mds_state_name(newstate
),
2706 ceph_mdsmap_is_laggy(newmap
, i
) ? " (laggy)" : "",
2707 session_state_name(s
->s_state
));
2709 if (i
>= newmap
->m_max_mds
||
2710 memcmp(ceph_mdsmap_get_addr(oldmap
, i
),
2711 ceph_mdsmap_get_addr(newmap
, i
),
2712 sizeof(struct ceph_entity_addr
))) {
2713 if (s
->s_state
== CEPH_MDS_SESSION_OPENING
) {
2714 /* the session never opened, just close it
2716 __wake_requests(mdsc
, &s
->s_waiting
);
2717 __unregister_session(mdsc
, s
);
2720 mutex_unlock(&mdsc
->mutex
);
2721 mutex_lock(&s
->s_mutex
);
2722 mutex_lock(&mdsc
->mutex
);
2723 ceph_con_close(&s
->s_con
);
2724 mutex_unlock(&s
->s_mutex
);
2725 s
->s_state
= CEPH_MDS_SESSION_RESTARTING
;
2728 /* kick any requests waiting on the recovering mds */
2729 kick_requests(mdsc
, i
);
2730 } else if (oldstate
== newstate
) {
2731 continue; /* nothing new with this mds */
2737 if (s
->s_state
== CEPH_MDS_SESSION_RESTARTING
&&
2738 newstate
>= CEPH_MDS_STATE_RECONNECT
) {
2739 mutex_unlock(&mdsc
->mutex
);
2740 send_mds_reconnect(mdsc
, s
);
2741 mutex_lock(&mdsc
->mutex
);
2745 * kick request on any mds that has gone active.
2747 if (oldstate
< CEPH_MDS_STATE_ACTIVE
&&
2748 newstate
>= CEPH_MDS_STATE_ACTIVE
) {
2749 if (oldstate
!= CEPH_MDS_STATE_CREATING
&&
2750 oldstate
!= CEPH_MDS_STATE_STARTING
)
2751 pr_info("mds%d recovery completed\n", s
->s_mds
);
2752 kick_requests(mdsc
, i
);
2753 ceph_kick_flushing_caps(mdsc
, s
);
2754 wake_up_session_caps(s
, 1);
2758 for (i
= 0; i
< newmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
2759 s
= mdsc
->sessions
[i
];
2762 if (!ceph_mdsmap_is_laggy(newmap
, i
))
2764 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
2765 s
->s_state
== CEPH_MDS_SESSION_HUNG
||
2766 s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
2767 dout(" connecting to export targets of laggy mds%d\n",
2769 __open_export_target_sessions(mdsc
, s
);
2781 * caller must hold session s_mutex, dentry->d_lock
2783 void __ceph_mdsc_drop_dentry_lease(struct dentry
*dentry
)
2785 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
2787 ceph_put_mds_session(di
->lease_session
);
2788 di
->lease_session
= NULL
;
2791 static void handle_lease(struct ceph_mds_client
*mdsc
,
2792 struct ceph_mds_session
*session
,
2793 struct ceph_msg
*msg
)
2795 struct super_block
*sb
= mdsc
->fsc
->sb
;
2796 struct inode
*inode
;
2797 struct dentry
*parent
, *dentry
;
2798 struct ceph_dentry_info
*di
;
2799 int mds
= session
->s_mds
;
2800 struct ceph_mds_lease
*h
= msg
->front
.iov_base
;
2802 struct ceph_vino vino
;
2806 dout("handle_lease from mds%d\n", mds
);
2809 if (msg
->front
.iov_len
< sizeof(*h
) + sizeof(u32
))
2811 vino
.ino
= le64_to_cpu(h
->ino
);
2812 vino
.snap
= CEPH_NOSNAP
;
2813 seq
= le32_to_cpu(h
->seq
);
2814 dname
.name
= (void *)h
+ sizeof(*h
) + sizeof(u32
);
2815 dname
.len
= msg
->front
.iov_len
- sizeof(*h
) - sizeof(u32
);
2816 if (dname
.len
!= get_unaligned_le32(h
+1))
2819 mutex_lock(&session
->s_mutex
);
2823 inode
= ceph_find_inode(sb
, vino
);
2824 dout("handle_lease %s, ino %llx %p %.*s\n",
2825 ceph_lease_op_name(h
->action
), vino
.ino
, inode
,
2826 dname
.len
, dname
.name
);
2827 if (inode
== NULL
) {
2828 dout("handle_lease no inode %llx\n", vino
.ino
);
2833 parent
= d_find_alias(inode
);
2835 dout("no parent dentry on inode %p\n", inode
);
2837 goto release
; /* hrm... */
2839 dname
.hash
= full_name_hash(dname
.name
, dname
.len
);
2840 dentry
= d_lookup(parent
, &dname
);
2845 spin_lock(&dentry
->d_lock
);
2846 di
= ceph_dentry(dentry
);
2847 switch (h
->action
) {
2848 case CEPH_MDS_LEASE_REVOKE
:
2849 if (di
->lease_session
== session
) {
2850 if (ceph_seq_cmp(di
->lease_seq
, seq
) > 0)
2851 h
->seq
= cpu_to_le32(di
->lease_seq
);
2852 __ceph_mdsc_drop_dentry_lease(dentry
);
2857 case CEPH_MDS_LEASE_RENEW
:
2858 if (di
->lease_session
== session
&&
2859 di
->lease_gen
== session
->s_cap_gen
&&
2860 di
->lease_renew_from
&&
2861 di
->lease_renew_after
== 0) {
2862 unsigned long duration
=
2863 le32_to_cpu(h
->duration_ms
) * HZ
/ 1000;
2865 di
->lease_seq
= seq
;
2866 dentry
->d_time
= di
->lease_renew_from
+ duration
;
2867 di
->lease_renew_after
= di
->lease_renew_from
+
2869 di
->lease_renew_from
= 0;
2873 spin_unlock(&dentry
->d_lock
);
2880 /* let's just reuse the same message */
2881 h
->action
= CEPH_MDS_LEASE_REVOKE_ACK
;
2883 ceph_con_send(&session
->s_con
, msg
);
2887 mutex_unlock(&session
->s_mutex
);
2891 pr_err("corrupt lease message\n");
2895 void ceph_mdsc_lease_send_msg(struct ceph_mds_session
*session
,
2896 struct inode
*inode
,
2897 struct dentry
*dentry
, char action
,
2900 struct ceph_msg
*msg
;
2901 struct ceph_mds_lease
*lease
;
2902 int len
= sizeof(*lease
) + sizeof(u32
);
2905 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2906 inode
, dentry
, ceph_lease_op_name(action
), session
->s_mds
);
2907 dnamelen
= dentry
->d_name
.len
;
2910 msg
= ceph_msg_new(CEPH_MSG_CLIENT_LEASE
, len
, GFP_NOFS
, false);
2913 lease
= msg
->front
.iov_base
;
2914 lease
->action
= action
;
2915 lease
->ino
= cpu_to_le64(ceph_vino(inode
).ino
);
2916 lease
->first
= lease
->last
= cpu_to_le64(ceph_vino(inode
).snap
);
2917 lease
->seq
= cpu_to_le32(seq
);
2918 put_unaligned_le32(dnamelen
, lease
+ 1);
2919 memcpy((void *)(lease
+ 1) + 4, dentry
->d_name
.name
, dnamelen
);
2922 * if this is a preemptive lease RELEASE, no need to
2923 * flush request stream, since the actual request will
2926 msg
->more_to_follow
= (action
== CEPH_MDS_LEASE_RELEASE
);
2928 ceph_con_send(&session
->s_con
, msg
);
2932 * Preemptively release a lease we expect to invalidate anyway.
2933 * Pass @inode always, @dentry is optional.
2935 void ceph_mdsc_lease_release(struct ceph_mds_client
*mdsc
, struct inode
*inode
,
2936 struct dentry
*dentry
)
2938 struct ceph_dentry_info
*di
;
2939 struct ceph_mds_session
*session
;
2942 BUG_ON(inode
== NULL
);
2943 BUG_ON(dentry
== NULL
);
2945 /* is dentry lease valid? */
2946 spin_lock(&dentry
->d_lock
);
2947 di
= ceph_dentry(dentry
);
2948 if (!di
|| !di
->lease_session
||
2949 di
->lease_session
->s_mds
< 0 ||
2950 di
->lease_gen
!= di
->lease_session
->s_cap_gen
||
2951 !time_before(jiffies
, dentry
->d_time
)) {
2952 dout("lease_release inode %p dentry %p -- "
2955 spin_unlock(&dentry
->d_lock
);
2959 /* we do have a lease on this dentry; note mds and seq */
2960 session
= ceph_get_mds_session(di
->lease_session
);
2961 seq
= di
->lease_seq
;
2962 __ceph_mdsc_drop_dentry_lease(dentry
);
2963 spin_unlock(&dentry
->d_lock
);
2965 dout("lease_release inode %p dentry %p to mds%d\n",
2966 inode
, dentry
, session
->s_mds
);
2967 ceph_mdsc_lease_send_msg(session
, inode
, dentry
,
2968 CEPH_MDS_LEASE_RELEASE
, seq
);
2969 ceph_put_mds_session(session
);
2973 * drop all leases (and dentry refs) in preparation for umount
2975 static void drop_leases(struct ceph_mds_client
*mdsc
)
2979 dout("drop_leases\n");
2980 mutex_lock(&mdsc
->mutex
);
2981 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
2982 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
2985 mutex_unlock(&mdsc
->mutex
);
2986 mutex_lock(&s
->s_mutex
);
2987 mutex_unlock(&s
->s_mutex
);
2988 ceph_put_mds_session(s
);
2989 mutex_lock(&mdsc
->mutex
);
2991 mutex_unlock(&mdsc
->mutex
);
2997 * delayed work -- periodically trim expired leases, renew caps with mds
2999 static void schedule_delayed(struct ceph_mds_client
*mdsc
)
3002 unsigned hz
= round_jiffies_relative(HZ
* delay
);
3003 schedule_delayed_work(&mdsc
->delayed_work
, hz
);
3006 static void delayed_work(struct work_struct
*work
)
3009 struct ceph_mds_client
*mdsc
=
3010 container_of(work
, struct ceph_mds_client
, delayed_work
.work
);
3014 dout("mdsc delayed_work\n");
3015 ceph_check_delayed_caps(mdsc
);
3017 mutex_lock(&mdsc
->mutex
);
3018 renew_interval
= mdsc
->mdsmap
->m_session_timeout
>> 2;
3019 renew_caps
= time_after_eq(jiffies
, HZ
*renew_interval
+
3020 mdsc
->last_renew_caps
);
3022 mdsc
->last_renew_caps
= jiffies
;
3024 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3025 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3028 if (s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3029 dout("resending session close request for mds%d\n",
3031 request_close_session(mdsc
, s
);
3032 ceph_put_mds_session(s
);
3035 if (s
->s_ttl
&& time_after(jiffies
, s
->s_ttl
)) {
3036 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
) {
3037 s
->s_state
= CEPH_MDS_SESSION_HUNG
;
3038 pr_info("mds%d hung\n", s
->s_mds
);
3041 if (s
->s_state
< CEPH_MDS_SESSION_OPEN
) {
3042 /* this mds is failed or recovering, just wait */
3043 ceph_put_mds_session(s
);
3046 mutex_unlock(&mdsc
->mutex
);
3048 mutex_lock(&s
->s_mutex
);
3050 send_renew_caps(mdsc
, s
);
3052 ceph_con_keepalive(&s
->s_con
);
3053 ceph_add_cap_releases(mdsc
, s
);
3054 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3055 s
->s_state
== CEPH_MDS_SESSION_HUNG
)
3056 ceph_send_cap_releases(mdsc
, s
);
3057 mutex_unlock(&s
->s_mutex
);
3058 ceph_put_mds_session(s
);
3060 mutex_lock(&mdsc
->mutex
);
3062 mutex_unlock(&mdsc
->mutex
);
3064 schedule_delayed(mdsc
);
3067 int ceph_mdsc_init(struct ceph_fs_client
*fsc
)
3070 struct ceph_mds_client
*mdsc
;
3072 mdsc
= kzalloc(sizeof(struct ceph_mds_client
), GFP_NOFS
);
3077 mutex_init(&mdsc
->mutex
);
3078 mdsc
->mdsmap
= kzalloc(sizeof(*mdsc
->mdsmap
), GFP_NOFS
);
3079 if (mdsc
->mdsmap
== NULL
) {
3084 init_completion(&mdsc
->safe_umount_waiters
);
3085 init_waitqueue_head(&mdsc
->session_close_wq
);
3086 INIT_LIST_HEAD(&mdsc
->waiting_for_map
);
3087 mdsc
->sessions
= NULL
;
3088 mdsc
->max_sessions
= 0;
3090 init_rwsem(&mdsc
->snap_rwsem
);
3091 mdsc
->snap_realms
= RB_ROOT
;
3092 INIT_LIST_HEAD(&mdsc
->snap_empty
);
3093 spin_lock_init(&mdsc
->snap_empty_lock
);
3095 mdsc
->request_tree
= RB_ROOT
;
3096 INIT_DELAYED_WORK(&mdsc
->delayed_work
, delayed_work
);
3097 mdsc
->last_renew_caps
= jiffies
;
3098 INIT_LIST_HEAD(&mdsc
->cap_delay_list
);
3099 spin_lock_init(&mdsc
->cap_delay_lock
);
3100 INIT_LIST_HEAD(&mdsc
->snap_flush_list
);
3101 spin_lock_init(&mdsc
->snap_flush_lock
);
3102 mdsc
->cap_flush_seq
= 0;
3103 INIT_LIST_HEAD(&mdsc
->cap_dirty
);
3104 INIT_LIST_HEAD(&mdsc
->cap_dirty_migrating
);
3105 mdsc
->num_cap_flushing
= 0;
3106 spin_lock_init(&mdsc
->cap_dirty_lock
);
3107 init_waitqueue_head(&mdsc
->cap_flushing_wq
);
3108 spin_lock_init(&mdsc
->dentry_lru_lock
);
3109 INIT_LIST_HEAD(&mdsc
->dentry_lru
);
3111 ceph_caps_init(mdsc
);
3112 ceph_adjust_min_caps(mdsc
, fsc
->min_caps
);
3118 * Wait for safe replies on open mds requests. If we time out, drop
3119 * all requests from the tree to avoid dangling dentry refs.
3121 static void wait_requests(struct ceph_mds_client
*mdsc
)
3123 struct ceph_mds_request
*req
;
3124 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3126 mutex_lock(&mdsc
->mutex
);
3127 if (__get_oldest_req(mdsc
)) {
3128 mutex_unlock(&mdsc
->mutex
);
3130 dout("wait_requests waiting for requests\n");
3131 wait_for_completion_timeout(&mdsc
->safe_umount_waiters
,
3132 fsc
->client
->options
->mount_timeout
* HZ
);
3134 /* tear down remaining requests */
3135 mutex_lock(&mdsc
->mutex
);
3136 while ((req
= __get_oldest_req(mdsc
))) {
3137 dout("wait_requests timed out on tid %llu\n",
3139 __unregister_request(mdsc
, req
);
3142 mutex_unlock(&mdsc
->mutex
);
3143 dout("wait_requests done\n");
3147 * called before mount is ro, and before dentries are torn down.
3148 * (hmm, does this still race with new lookups?)
3150 void ceph_mdsc_pre_umount(struct ceph_mds_client
*mdsc
)
3152 dout("pre_umount\n");
3156 ceph_flush_dirty_caps(mdsc
);
3157 wait_requests(mdsc
);
3160 * wait for reply handlers to drop their request refs and
3161 * their inode/dcache refs
3167 * wait for all write mds requests to flush.
3169 static void wait_unsafe_requests(struct ceph_mds_client
*mdsc
, u64 want_tid
)
3171 struct ceph_mds_request
*req
= NULL
, *nextreq
;
3174 mutex_lock(&mdsc
->mutex
);
3175 dout("wait_unsafe_requests want %lld\n", want_tid
);
3177 req
= __get_oldest_req(mdsc
);
3178 while (req
&& req
->r_tid
<= want_tid
) {
3179 /* find next request */
3180 n
= rb_next(&req
->r_node
);
3182 nextreq
= rb_entry(n
, struct ceph_mds_request
, r_node
);
3185 if ((req
->r_op
& CEPH_MDS_OP_WRITE
)) {
3187 ceph_mdsc_get_request(req
);
3189 ceph_mdsc_get_request(nextreq
);
3190 mutex_unlock(&mdsc
->mutex
);
3191 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3192 req
->r_tid
, want_tid
);
3193 wait_for_completion(&req
->r_safe_completion
);
3194 mutex_lock(&mdsc
->mutex
);
3195 ceph_mdsc_put_request(req
);
3197 break; /* next dne before, so we're done! */
3198 if (RB_EMPTY_NODE(&nextreq
->r_node
)) {
3199 /* next request was removed from tree */
3200 ceph_mdsc_put_request(nextreq
);
3203 ceph_mdsc_put_request(nextreq
); /* won't go away */
3207 mutex_unlock(&mdsc
->mutex
);
3208 dout("wait_unsafe_requests done\n");
3211 void ceph_mdsc_sync(struct ceph_mds_client
*mdsc
)
3213 u64 want_tid
, want_flush
;
3215 if (mdsc
->fsc
->mount_state
== CEPH_MOUNT_SHUTDOWN
)
3219 mutex_lock(&mdsc
->mutex
);
3220 want_tid
= mdsc
->last_tid
;
3221 want_flush
= mdsc
->cap_flush_seq
;
3222 mutex_unlock(&mdsc
->mutex
);
3223 dout("sync want tid %lld flush_seq %lld\n", want_tid
, want_flush
);
3225 ceph_flush_dirty_caps(mdsc
);
3227 wait_unsafe_requests(mdsc
, want_tid
);
3228 wait_event(mdsc
->cap_flushing_wq
, check_cap_flush(mdsc
, want_flush
));
3232 * true if all sessions are closed, or we force unmount
3234 static bool done_closing_sessions(struct ceph_mds_client
*mdsc
)
3238 if (mdsc
->fsc
->mount_state
== CEPH_MOUNT_SHUTDOWN
)
3241 mutex_lock(&mdsc
->mutex
);
3242 for (i
= 0; i
< mdsc
->max_sessions
; i
++)
3243 if (mdsc
->sessions
[i
])
3245 mutex_unlock(&mdsc
->mutex
);
3250 * called after sb is ro.
3252 void ceph_mdsc_close_sessions(struct ceph_mds_client
*mdsc
)
3254 struct ceph_mds_session
*session
;
3256 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3257 unsigned long timeout
= fsc
->client
->options
->mount_timeout
* HZ
;
3259 dout("close_sessions\n");
3261 /* close sessions */
3262 mutex_lock(&mdsc
->mutex
);
3263 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3264 session
= __ceph_lookup_mds_session(mdsc
, i
);
3267 mutex_unlock(&mdsc
->mutex
);
3268 mutex_lock(&session
->s_mutex
);
3269 __close_session(mdsc
, session
);
3270 mutex_unlock(&session
->s_mutex
);
3271 ceph_put_mds_session(session
);
3272 mutex_lock(&mdsc
->mutex
);
3274 mutex_unlock(&mdsc
->mutex
);
3276 dout("waiting for sessions to close\n");
3277 wait_event_timeout(mdsc
->session_close_wq
, done_closing_sessions(mdsc
),
3280 /* tear down remaining sessions */
3281 mutex_lock(&mdsc
->mutex
);
3282 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3283 if (mdsc
->sessions
[i
]) {
3284 session
= get_session(mdsc
->sessions
[i
]);
3285 __unregister_session(mdsc
, session
);
3286 mutex_unlock(&mdsc
->mutex
);
3287 mutex_lock(&session
->s_mutex
);
3288 remove_session_caps(session
);
3289 mutex_unlock(&session
->s_mutex
);
3290 ceph_put_mds_session(session
);
3291 mutex_lock(&mdsc
->mutex
);
3294 WARN_ON(!list_empty(&mdsc
->cap_delay_list
));
3295 mutex_unlock(&mdsc
->mutex
);
3297 ceph_cleanup_empty_realms(mdsc
);
3299 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3304 static void ceph_mdsc_stop(struct ceph_mds_client
*mdsc
)
3307 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3309 ceph_mdsmap_destroy(mdsc
->mdsmap
);
3310 kfree(mdsc
->sessions
);
3311 ceph_caps_finalize(mdsc
);
3314 void ceph_mdsc_destroy(struct ceph_fs_client
*fsc
)
3316 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
3318 dout("mdsc_destroy %p\n", mdsc
);
3319 ceph_mdsc_stop(mdsc
);
3321 /* flush out any connection work with references to us */
3326 dout("mdsc_destroy %p done\n", mdsc
);
3331 * handle mds map update.
3333 void ceph_mdsc_handle_map(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3337 void *p
= msg
->front
.iov_base
;
3338 void *end
= p
+ msg
->front
.iov_len
;
3339 struct ceph_mdsmap
*newmap
, *oldmap
;
3340 struct ceph_fsid fsid
;
3343 ceph_decode_need(&p
, end
, sizeof(fsid
)+2*sizeof(u32
), bad
);
3344 ceph_decode_copy(&p
, &fsid
, sizeof(fsid
));
3345 if (ceph_check_fsid(mdsc
->fsc
->client
, &fsid
) < 0)
3347 epoch
= ceph_decode_32(&p
);
3348 maplen
= ceph_decode_32(&p
);
3349 dout("handle_map epoch %u len %d\n", epoch
, (int)maplen
);
3351 /* do we need it? */
3352 ceph_monc_got_mdsmap(&mdsc
->fsc
->client
->monc
, epoch
);
3353 mutex_lock(&mdsc
->mutex
);
3354 if (mdsc
->mdsmap
&& epoch
<= mdsc
->mdsmap
->m_epoch
) {
3355 dout("handle_map epoch %u <= our %u\n",
3356 epoch
, mdsc
->mdsmap
->m_epoch
);
3357 mutex_unlock(&mdsc
->mutex
);
3361 newmap
= ceph_mdsmap_decode(&p
, end
);
3362 if (IS_ERR(newmap
)) {
3363 err
= PTR_ERR(newmap
);
3367 /* swap into place */
3369 oldmap
= mdsc
->mdsmap
;
3370 mdsc
->mdsmap
= newmap
;
3371 check_new_map(mdsc
, newmap
, oldmap
);
3372 ceph_mdsmap_destroy(oldmap
);
3374 mdsc
->mdsmap
= newmap
; /* first mds map */
3376 mdsc
->fsc
->sb
->s_maxbytes
= mdsc
->mdsmap
->m_max_file_size
;
3378 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3380 mutex_unlock(&mdsc
->mutex
);
3381 schedule_delayed(mdsc
);
3385 mutex_unlock(&mdsc
->mutex
);
3387 pr_err("error decoding mdsmap %d\n", err
);
3391 static struct ceph_connection
*con_get(struct ceph_connection
*con
)
3393 struct ceph_mds_session
*s
= con
->private;
3395 if (get_session(s
)) {
3396 dout("mdsc con_get %p ok (%d)\n", s
, atomic_read(&s
->s_ref
));
3399 dout("mdsc con_get %p FAIL\n", s
);
3403 static void con_put(struct ceph_connection
*con
)
3405 struct ceph_mds_session
*s
= con
->private;
3407 dout("mdsc con_put %p (%d)\n", s
, atomic_read(&s
->s_ref
) - 1);
3408 ceph_put_mds_session(s
);
3412 * if the client is unresponsive for long enough, the mds will kill
3413 * the session entirely.
3415 static void peer_reset(struct ceph_connection
*con
)
3417 struct ceph_mds_session
*s
= con
->private;
3418 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3420 pr_warning("mds%d closed our session\n", s
->s_mds
);
3421 send_mds_reconnect(mdsc
, s
);
3424 static void dispatch(struct ceph_connection
*con
, struct ceph_msg
*msg
)
3426 struct ceph_mds_session
*s
= con
->private;
3427 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3428 int type
= le16_to_cpu(msg
->hdr
.type
);
3430 mutex_lock(&mdsc
->mutex
);
3431 if (__verify_registered_session(mdsc
, s
) < 0) {
3432 mutex_unlock(&mdsc
->mutex
);
3435 mutex_unlock(&mdsc
->mutex
);
3438 case CEPH_MSG_MDS_MAP
:
3439 ceph_mdsc_handle_map(mdsc
, msg
);
3441 case CEPH_MSG_CLIENT_SESSION
:
3442 handle_session(s
, msg
);
3444 case CEPH_MSG_CLIENT_REPLY
:
3445 handle_reply(s
, msg
);
3447 case CEPH_MSG_CLIENT_REQUEST_FORWARD
:
3448 handle_forward(mdsc
, s
, msg
);
3450 case CEPH_MSG_CLIENT_CAPS
:
3451 ceph_handle_caps(s
, msg
);
3453 case CEPH_MSG_CLIENT_SNAP
:
3454 ceph_handle_snap(mdsc
, s
, msg
);
3456 case CEPH_MSG_CLIENT_LEASE
:
3457 handle_lease(mdsc
, s
, msg
);
3461 pr_err("received unknown message type %d %s\n", type
,
3462 ceph_msg_type_name(type
));
3473 * Note: returned pointer is the address of a structure that's
3474 * managed separately. Caller must *not* attempt to free it.
3476 static struct ceph_auth_handshake
*get_authorizer(struct ceph_connection
*con
,
3477 int *proto
, int force_new
)
3479 struct ceph_mds_session
*s
= con
->private;
3480 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3481 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3482 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
3484 if (force_new
&& auth
->authorizer
) {
3485 ceph_auth_destroy_authorizer(ac
, auth
->authorizer
);
3486 auth
->authorizer
= NULL
;
3488 if (!auth
->authorizer
) {
3489 int ret
= ceph_auth_create_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3492 return ERR_PTR(ret
);
3494 int ret
= ceph_auth_update_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3497 return ERR_PTR(ret
);
3499 *proto
= ac
->protocol
;
3505 static int verify_authorizer_reply(struct ceph_connection
*con
, int len
)
3507 struct ceph_mds_session
*s
= con
->private;
3508 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3509 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3511 return ceph_auth_verify_authorizer_reply(ac
, s
->s_auth
.authorizer
, len
);
3514 static int invalidate_authorizer(struct ceph_connection
*con
)
3516 struct ceph_mds_session
*s
= con
->private;
3517 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3518 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3520 ceph_auth_invalidate_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
);
3522 return ceph_monc_validate_auth(&mdsc
->fsc
->client
->monc
);
3525 static struct ceph_msg
*mds_alloc_msg(struct ceph_connection
*con
,
3526 struct ceph_msg_header
*hdr
, int *skip
)
3528 struct ceph_msg
*msg
;
3529 int type
= (int) le16_to_cpu(hdr
->type
);
3530 int front_len
= (int) le32_to_cpu(hdr
->front_len
);
3536 msg
= ceph_msg_new(type
, front_len
, GFP_NOFS
, false);
3538 pr_err("unable to allocate msg type %d len %d\n",
3546 static const struct ceph_connection_operations mds_con_ops
= {
3549 .dispatch
= dispatch
,
3550 .get_authorizer
= get_authorizer
,
3551 .verify_authorizer_reply
= verify_authorizer_reply
,
3552 .invalidate_authorizer
= invalidate_authorizer
,
3553 .peer_reset
= peer_reset
,
3554 .alloc_msg
= mds_alloc_msg
,