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42 * are no longer valid.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 return parse_reply_info_dir(p
, end
, info
, features
);
270 else if (info
->head
->op
== CEPH_MDS_OP_CREATE
)
271 return parse_reply_info_create(p
, end
, info
, features
);
277 * parse entire mds reply
279 static int parse_reply_info(struct ceph_msg
*msg
,
280 struct ceph_mds_reply_info_parsed
*info
,
287 info
->head
= msg
->front
.iov_base
;
288 p
= msg
->front
.iov_base
+ sizeof(struct ceph_mds_reply_head
);
289 end
= p
+ msg
->front
.iov_len
- sizeof(struct ceph_mds_reply_head
);
292 ceph_decode_32_safe(&p
, end
, len
, bad
);
294 ceph_decode_need(&p
, end
, len
, bad
);
295 err
= parse_reply_info_trace(&p
, p
+len
, info
, features
);
301 ceph_decode_32_safe(&p
, end
, len
, bad
);
303 ceph_decode_need(&p
, end
, len
, bad
);
304 err
= parse_reply_info_extra(&p
, p
+len
, info
, features
);
310 ceph_decode_32_safe(&p
, end
, len
, bad
);
311 info
->snapblob_len
= len
;
322 pr_err("mds parse_reply err %d\n", err
);
326 static void destroy_reply_info(struct ceph_mds_reply_info_parsed
*info
)
335 static const char *session_state_name(int s
)
338 case CEPH_MDS_SESSION_NEW
: return "new";
339 case CEPH_MDS_SESSION_OPENING
: return "opening";
340 case CEPH_MDS_SESSION_OPEN
: return "open";
341 case CEPH_MDS_SESSION_HUNG
: return "hung";
342 case CEPH_MDS_SESSION_CLOSING
: return "closing";
343 case CEPH_MDS_SESSION_RESTARTING
: return "restarting";
344 case CEPH_MDS_SESSION_RECONNECTING
: return "reconnecting";
345 default: return "???";
349 static struct ceph_mds_session
*get_session(struct ceph_mds_session
*s
)
351 if (atomic_inc_not_zero(&s
->s_ref
)) {
352 dout("mdsc get_session %p %d -> %d\n", s
,
353 atomic_read(&s
->s_ref
)-1, atomic_read(&s
->s_ref
));
356 dout("mdsc get_session %p 0 -- FAIL", s
);
361 void ceph_put_mds_session(struct ceph_mds_session
*s
)
363 dout("mdsc put_session %p %d -> %d\n", s
,
364 atomic_read(&s
->s_ref
), atomic_read(&s
->s_ref
)-1);
365 if (atomic_dec_and_test(&s
->s_ref
)) {
366 if (s
->s_auth
.authorizer
)
367 s
->s_mdsc
->fsc
->client
->monc
.auth
->ops
->destroy_authorizer(
368 s
->s_mdsc
->fsc
->client
->monc
.auth
,
369 s
->s_auth
.authorizer
);
375 * called under mdsc->mutex
377 struct ceph_mds_session
*__ceph_lookup_mds_session(struct ceph_mds_client
*mdsc
,
380 struct ceph_mds_session
*session
;
382 if (mds
>= mdsc
->max_sessions
|| mdsc
->sessions
[mds
] == NULL
)
384 session
= mdsc
->sessions
[mds
];
385 dout("lookup_mds_session %p %d\n", session
,
386 atomic_read(&session
->s_ref
));
387 get_session(session
);
391 static bool __have_session(struct ceph_mds_client
*mdsc
, int mds
)
393 if (mds
>= mdsc
->max_sessions
)
395 return mdsc
->sessions
[mds
];
398 static int __verify_registered_session(struct ceph_mds_client
*mdsc
,
399 struct ceph_mds_session
*s
)
401 if (s
->s_mds
>= mdsc
->max_sessions
||
402 mdsc
->sessions
[s
->s_mds
] != s
)
408 * create+register a new session for given mds.
409 * called under mdsc->mutex.
411 static struct ceph_mds_session
*register_session(struct ceph_mds_client
*mdsc
,
414 struct ceph_mds_session
*s
;
416 s
= kzalloc(sizeof(*s
), GFP_NOFS
);
418 return ERR_PTR(-ENOMEM
);
421 s
->s_state
= CEPH_MDS_SESSION_NEW
;
424 mutex_init(&s
->s_mutex
);
426 ceph_con_init(&s
->s_con
, s
, &mds_con_ops
, &mdsc
->fsc
->client
->msgr
);
428 spin_lock_init(&s
->s_gen_ttl_lock
);
430 s
->s_cap_ttl
= jiffies
- 1;
432 spin_lock_init(&s
->s_cap_lock
);
433 s
->s_renew_requested
= 0;
435 INIT_LIST_HEAD(&s
->s_caps
);
438 atomic_set(&s
->s_ref
, 1);
439 INIT_LIST_HEAD(&s
->s_waiting
);
440 INIT_LIST_HEAD(&s
->s_unsafe
);
441 s
->s_num_cap_releases
= 0;
442 s
->s_cap_iterator
= NULL
;
443 INIT_LIST_HEAD(&s
->s_cap_releases
);
444 INIT_LIST_HEAD(&s
->s_cap_releases_done
);
445 INIT_LIST_HEAD(&s
->s_cap_flushing
);
446 INIT_LIST_HEAD(&s
->s_cap_snaps_flushing
);
448 dout("register_session mds%d\n", mds
);
449 if (mds
>= mdsc
->max_sessions
) {
450 int newmax
= 1 << get_count_order(mds
+1);
451 struct ceph_mds_session
**sa
;
453 dout("register_session realloc to %d\n", newmax
);
454 sa
= kcalloc(newmax
, sizeof(void *), GFP_NOFS
);
457 if (mdsc
->sessions
) {
458 memcpy(sa
, mdsc
->sessions
,
459 mdsc
->max_sessions
* sizeof(void *));
460 kfree(mdsc
->sessions
);
463 mdsc
->max_sessions
= newmax
;
465 mdsc
->sessions
[mds
] = s
;
466 atomic_inc(&s
->s_ref
); /* one ref to sessions[], one to caller */
468 ceph_con_open(&s
->s_con
, CEPH_ENTITY_TYPE_MDS
, mds
,
469 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
475 return ERR_PTR(-ENOMEM
);
479 * called under mdsc->mutex
481 static void __unregister_session(struct ceph_mds_client
*mdsc
,
482 struct ceph_mds_session
*s
)
484 dout("__unregister_session mds%d %p\n", s
->s_mds
, s
);
485 BUG_ON(mdsc
->sessions
[s
->s_mds
] != s
);
486 mdsc
->sessions
[s
->s_mds
] = NULL
;
487 ceph_con_close(&s
->s_con
);
488 ceph_put_mds_session(s
);
492 * drop session refs in request.
494 * should be last request ref, or hold mdsc->mutex
496 static void put_request_session(struct ceph_mds_request
*req
)
498 if (req
->r_session
) {
499 ceph_put_mds_session(req
->r_session
);
500 req
->r_session
= NULL
;
504 void ceph_mdsc_release_request(struct kref
*kref
)
506 struct ceph_mds_request
*req
= container_of(kref
,
507 struct ceph_mds_request
,
510 ceph_msg_put(req
->r_request
);
512 ceph_msg_put(req
->r_reply
);
513 destroy_reply_info(&req
->r_reply_info
);
516 ceph_put_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
519 if (req
->r_locked_dir
)
520 ceph_put_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
521 if (req
->r_target_inode
)
522 iput(req
->r_target_inode
);
525 if (req
->r_old_dentry
) {
527 * track (and drop pins for) r_old_dentry_dir
528 * separately, since r_old_dentry's d_parent may have
529 * changed between the dir mutex being dropped and
530 * this request being freed.
532 ceph_put_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
534 dput(req
->r_old_dentry
);
535 iput(req
->r_old_dentry_dir
);
539 put_request_session(req
);
540 ceph_unreserve_caps(req
->r_mdsc
, &req
->r_caps_reservation
);
545 * lookup session, bump ref if found.
547 * called under mdsc->mutex.
549 static struct ceph_mds_request
*__lookup_request(struct ceph_mds_client
*mdsc
,
552 struct ceph_mds_request
*req
;
553 struct rb_node
*n
= mdsc
->request_tree
.rb_node
;
556 req
= rb_entry(n
, struct ceph_mds_request
, r_node
);
557 if (tid
< req
->r_tid
)
559 else if (tid
> req
->r_tid
)
562 ceph_mdsc_get_request(req
);
569 static void __insert_request(struct ceph_mds_client
*mdsc
,
570 struct ceph_mds_request
*new)
572 struct rb_node
**p
= &mdsc
->request_tree
.rb_node
;
573 struct rb_node
*parent
= NULL
;
574 struct ceph_mds_request
*req
= NULL
;
578 req
= rb_entry(parent
, struct ceph_mds_request
, r_node
);
579 if (new->r_tid
< req
->r_tid
)
581 else if (new->r_tid
> req
->r_tid
)
587 rb_link_node(&new->r_node
, parent
, p
);
588 rb_insert_color(&new->r_node
, &mdsc
->request_tree
);
592 * Register an in-flight request, and assign a tid. Link to directory
593 * are modifying (if any).
595 * Called under mdsc->mutex.
597 static void __register_request(struct ceph_mds_client
*mdsc
,
598 struct ceph_mds_request
*req
,
601 req
->r_tid
= ++mdsc
->last_tid
;
603 ceph_reserve_caps(mdsc
, &req
->r_caps_reservation
,
605 dout("__register_request %p tid %lld\n", req
, req
->r_tid
);
606 ceph_mdsc_get_request(req
);
607 __insert_request(mdsc
, req
);
609 req
->r_uid
= current_fsuid();
610 req
->r_gid
= current_fsgid();
613 struct ceph_inode_info
*ci
= ceph_inode(dir
);
616 spin_lock(&ci
->i_unsafe_lock
);
617 req
->r_unsafe_dir
= dir
;
618 list_add_tail(&req
->r_unsafe_dir_item
, &ci
->i_unsafe_dirops
);
619 spin_unlock(&ci
->i_unsafe_lock
);
623 static void __unregister_request(struct ceph_mds_client
*mdsc
,
624 struct ceph_mds_request
*req
)
626 dout("__unregister_request %p tid %lld\n", req
, req
->r_tid
);
627 rb_erase(&req
->r_node
, &mdsc
->request_tree
);
628 RB_CLEAR_NODE(&req
->r_node
);
630 if (req
->r_unsafe_dir
) {
631 struct ceph_inode_info
*ci
= ceph_inode(req
->r_unsafe_dir
);
633 spin_lock(&ci
->i_unsafe_lock
);
634 list_del_init(&req
->r_unsafe_dir_item
);
635 spin_unlock(&ci
->i_unsafe_lock
);
637 iput(req
->r_unsafe_dir
);
638 req
->r_unsafe_dir
= NULL
;
641 ceph_mdsc_put_request(req
);
645 * Choose mds to send request to next. If there is a hint set in the
646 * request (e.g., due to a prior forward hint from the mds), use that.
647 * Otherwise, consult frag tree and/or caps to identify the
648 * appropriate mds. If all else fails, choose randomly.
650 * Called under mdsc->mutex.
652 static struct dentry
*get_nonsnap_parent(struct dentry
*dentry
)
655 * we don't need to worry about protecting the d_parent access
656 * here because we never renaming inside the snapped namespace
657 * except to resplice to another snapdir, and either the old or new
658 * result is a valid result.
660 while (!IS_ROOT(dentry
) && ceph_snap(dentry
->d_inode
) != CEPH_NOSNAP
)
661 dentry
= dentry
->d_parent
;
665 static int __choose_mds(struct ceph_mds_client
*mdsc
,
666 struct ceph_mds_request
*req
)
669 struct ceph_inode_info
*ci
;
670 struct ceph_cap
*cap
;
671 int mode
= req
->r_direct_mode
;
673 u32 hash
= req
->r_direct_hash
;
674 bool is_hash
= req
->r_direct_is_hash
;
677 * is there a specific mds we should try? ignore hint if we have
678 * no session and the mds is not up (active or recovering).
680 if (req
->r_resend_mds
>= 0 &&
681 (__have_session(mdsc
, req
->r_resend_mds
) ||
682 ceph_mdsmap_get_state(mdsc
->mdsmap
, req
->r_resend_mds
) > 0)) {
683 dout("choose_mds using resend_mds mds%d\n",
685 return req
->r_resend_mds
;
688 if (mode
== USE_RANDOM_MDS
)
693 inode
= req
->r_inode
;
694 } else if (req
->r_dentry
) {
695 /* ignore race with rename; old or new d_parent is okay */
696 struct dentry
*parent
= req
->r_dentry
->d_parent
;
697 struct inode
*dir
= parent
->d_inode
;
699 if (dir
->i_sb
!= mdsc
->fsc
->sb
) {
701 inode
= req
->r_dentry
->d_inode
;
702 } else if (ceph_snap(dir
) != CEPH_NOSNAP
) {
703 /* direct snapped/virtual snapdir requests
704 * based on parent dir inode */
705 struct dentry
*dn
= get_nonsnap_parent(parent
);
707 dout("__choose_mds using nonsnap parent %p\n", inode
);
708 } else if (req
->r_dentry
->d_inode
) {
710 inode
= req
->r_dentry
->d_inode
;
714 hash
= ceph_dentry_hash(dir
, req
->r_dentry
);
719 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode
, (int)is_hash
,
723 ci
= ceph_inode(inode
);
725 if (is_hash
&& S_ISDIR(inode
->i_mode
)) {
726 struct ceph_inode_frag frag
;
729 ceph_choose_frag(ci
, hash
, &frag
, &found
);
731 if (mode
== USE_ANY_MDS
&& frag
.ndist
> 0) {
734 /* choose a random replica */
735 get_random_bytes(&r
, 1);
738 dout("choose_mds %p %llx.%llx "
739 "frag %u mds%d (%d/%d)\n",
740 inode
, ceph_vinop(inode
),
743 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
744 CEPH_MDS_STATE_ACTIVE
)
748 /* since this file/dir wasn't known to be
749 * replicated, then we want to look for the
750 * authoritative mds. */
753 /* choose auth mds */
755 dout("choose_mds %p %llx.%llx "
756 "frag %u mds%d (auth)\n",
757 inode
, ceph_vinop(inode
), frag
.frag
, mds
);
758 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
759 CEPH_MDS_STATE_ACTIVE
)
765 spin_lock(&ci
->i_ceph_lock
);
767 if (mode
== USE_AUTH_MDS
)
768 cap
= ci
->i_auth_cap
;
769 if (!cap
&& !RB_EMPTY_ROOT(&ci
->i_caps
))
770 cap
= rb_entry(rb_first(&ci
->i_caps
), struct ceph_cap
, ci_node
);
772 spin_unlock(&ci
->i_ceph_lock
);
775 mds
= cap
->session
->s_mds
;
776 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
777 inode
, ceph_vinop(inode
), mds
,
778 cap
== ci
->i_auth_cap
? "auth " : "", cap
);
779 spin_unlock(&ci
->i_ceph_lock
);
783 mds
= ceph_mdsmap_get_random_mds(mdsc
->mdsmap
);
784 dout("choose_mds chose random mds%d\n", mds
);
792 static struct ceph_msg
*create_session_msg(u32 op
, u64 seq
)
794 struct ceph_msg
*msg
;
795 struct ceph_mds_session_head
*h
;
797 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
), GFP_NOFS
,
800 pr_err("create_session_msg ENOMEM creating msg\n");
803 h
= msg
->front
.iov_base
;
804 h
->op
= cpu_to_le32(op
);
805 h
->seq
= cpu_to_le64(seq
);
810 * send session open request.
812 * called under mdsc->mutex
814 static int __open_session(struct ceph_mds_client
*mdsc
,
815 struct ceph_mds_session
*session
)
817 struct ceph_msg
*msg
;
819 int mds
= session
->s_mds
;
821 /* wait for mds to go active? */
822 mstate
= ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
);
823 dout("open_session to mds%d (%s)\n", mds
,
824 ceph_mds_state_name(mstate
));
825 session
->s_state
= CEPH_MDS_SESSION_OPENING
;
826 session
->s_renew_requested
= jiffies
;
828 /* send connect message */
829 msg
= create_session_msg(CEPH_SESSION_REQUEST_OPEN
, session
->s_seq
);
832 ceph_con_send(&session
->s_con
, msg
);
837 * open sessions for any export targets for the given mds
839 * called under mdsc->mutex
841 static void __open_export_target_sessions(struct ceph_mds_client
*mdsc
,
842 struct ceph_mds_session
*session
)
844 struct ceph_mds_info
*mi
;
845 struct ceph_mds_session
*ts
;
846 int i
, mds
= session
->s_mds
;
849 if (mds
>= mdsc
->mdsmap
->m_max_mds
)
851 mi
= &mdsc
->mdsmap
->m_info
[mds
];
852 dout("open_export_target_sessions for mds%d (%d targets)\n",
853 session
->s_mds
, mi
->num_export_targets
);
855 for (i
= 0; i
< mi
->num_export_targets
; i
++) {
856 target
= mi
->export_targets
[i
];
857 ts
= __ceph_lookup_mds_session(mdsc
, target
);
859 ts
= register_session(mdsc
, target
);
863 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
864 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
865 __open_session(mdsc
, session
);
867 dout(" mds%d target mds%d %p is %s\n", session
->s_mds
,
868 i
, ts
, session_state_name(ts
->s_state
));
869 ceph_put_mds_session(ts
);
873 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client
*mdsc
,
874 struct ceph_mds_session
*session
)
876 mutex_lock(&mdsc
->mutex
);
877 __open_export_target_sessions(mdsc
, session
);
878 mutex_unlock(&mdsc
->mutex
);
886 * Free preallocated cap messages assigned to this session
888 static void cleanup_cap_releases(struct ceph_mds_session
*session
)
890 struct ceph_msg
*msg
;
892 spin_lock(&session
->s_cap_lock
);
893 while (!list_empty(&session
->s_cap_releases
)) {
894 msg
= list_first_entry(&session
->s_cap_releases
,
895 struct ceph_msg
, list_head
);
896 list_del_init(&msg
->list_head
);
899 while (!list_empty(&session
->s_cap_releases_done
)) {
900 msg
= list_first_entry(&session
->s_cap_releases_done
,
901 struct ceph_msg
, list_head
);
902 list_del_init(&msg
->list_head
);
905 spin_unlock(&session
->s_cap_lock
);
909 * Helper to safely iterate over all caps associated with a session, with
910 * special care taken to handle a racing __ceph_remove_cap().
912 * Caller must hold session s_mutex.
914 static int iterate_session_caps(struct ceph_mds_session
*session
,
915 int (*cb
)(struct inode
*, struct ceph_cap
*,
919 struct ceph_cap
*cap
;
920 struct inode
*inode
, *last_inode
= NULL
;
921 struct ceph_cap
*old_cap
= NULL
;
924 dout("iterate_session_caps %p mds%d\n", session
, session
->s_mds
);
925 spin_lock(&session
->s_cap_lock
);
926 p
= session
->s_caps
.next
;
927 while (p
!= &session
->s_caps
) {
928 cap
= list_entry(p
, struct ceph_cap
, session_caps
);
929 inode
= igrab(&cap
->ci
->vfs_inode
);
934 session
->s_cap_iterator
= cap
;
935 spin_unlock(&session
->s_cap_lock
);
942 ceph_put_cap(session
->s_mdsc
, old_cap
);
946 ret
= cb(inode
, cap
, arg
);
949 spin_lock(&session
->s_cap_lock
);
951 if (cap
->ci
== NULL
) {
952 dout("iterate_session_caps finishing cap %p removal\n",
954 BUG_ON(cap
->session
!= session
);
955 list_del_init(&cap
->session_caps
);
956 session
->s_nr_caps
--;
958 old_cap
= cap
; /* put_cap it w/o locks held */
965 session
->s_cap_iterator
= NULL
;
966 spin_unlock(&session
->s_cap_lock
);
971 ceph_put_cap(session
->s_mdsc
, old_cap
);
976 static int remove_session_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
979 struct ceph_inode_info
*ci
= ceph_inode(inode
);
982 dout("removing cap %p, ci is %p, inode is %p\n",
983 cap
, ci
, &ci
->vfs_inode
);
984 spin_lock(&ci
->i_ceph_lock
);
985 __ceph_remove_cap(cap
);
986 if (!__ceph_is_any_real_caps(ci
)) {
987 struct ceph_mds_client
*mdsc
=
988 ceph_sb_to_client(inode
->i_sb
)->mdsc
;
990 spin_lock(&mdsc
->cap_dirty_lock
);
991 if (!list_empty(&ci
->i_dirty_item
)) {
992 pr_info(" dropping dirty %s state for %p %lld\n",
993 ceph_cap_string(ci
->i_dirty_caps
),
994 inode
, ceph_ino(inode
));
995 ci
->i_dirty_caps
= 0;
996 list_del_init(&ci
->i_dirty_item
);
999 if (!list_empty(&ci
->i_flushing_item
)) {
1000 pr_info(" dropping dirty+flushing %s state for %p %lld\n",
1001 ceph_cap_string(ci
->i_flushing_caps
),
1002 inode
, ceph_ino(inode
));
1003 ci
->i_flushing_caps
= 0;
1004 list_del_init(&ci
->i_flushing_item
);
1005 mdsc
->num_cap_flushing
--;
1008 if (drop
&& ci
->i_wrbuffer_ref
) {
1009 pr_info(" dropping dirty data for %p %lld\n",
1010 inode
, ceph_ino(inode
));
1011 ci
->i_wrbuffer_ref
= 0;
1012 ci
->i_wrbuffer_ref_head
= 0;
1015 spin_unlock(&mdsc
->cap_dirty_lock
);
1017 spin_unlock(&ci
->i_ceph_lock
);
1024 * caller must hold session s_mutex
1026 static void remove_session_caps(struct ceph_mds_session
*session
)
1028 dout("remove_session_caps on %p\n", session
);
1029 iterate_session_caps(session
, remove_session_caps_cb
, NULL
);
1030 BUG_ON(session
->s_nr_caps
> 0);
1031 BUG_ON(!list_empty(&session
->s_cap_flushing
));
1032 cleanup_cap_releases(session
);
1036 * wake up any threads waiting on this session's caps. if the cap is
1037 * old (didn't get renewed on the client reconnect), remove it now.
1039 * caller must hold s_mutex.
1041 static int wake_up_session_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1044 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1046 wake_up_all(&ci
->i_cap_wq
);
1048 spin_lock(&ci
->i_ceph_lock
);
1049 ci
->i_wanted_max_size
= 0;
1050 ci
->i_requested_max_size
= 0;
1051 spin_unlock(&ci
->i_ceph_lock
);
1056 static void wake_up_session_caps(struct ceph_mds_session
*session
,
1059 dout("wake_up_session_caps %p mds%d\n", session
, session
->s_mds
);
1060 iterate_session_caps(session
, wake_up_session_cb
,
1061 (void *)(unsigned long)reconnect
);
1065 * Send periodic message to MDS renewing all currently held caps. The
1066 * ack will reset the expiration for all caps from this session.
1068 * caller holds s_mutex
1070 static int send_renew_caps(struct ceph_mds_client
*mdsc
,
1071 struct ceph_mds_session
*session
)
1073 struct ceph_msg
*msg
;
1076 if (time_after_eq(jiffies
, session
->s_cap_ttl
) &&
1077 time_after_eq(session
->s_cap_ttl
, session
->s_renew_requested
))
1078 pr_info("mds%d caps stale\n", session
->s_mds
);
1079 session
->s_renew_requested
= jiffies
;
1081 /* do not try to renew caps until a recovering mds has reconnected
1082 * with its clients. */
1083 state
= ceph_mdsmap_get_state(mdsc
->mdsmap
, session
->s_mds
);
1084 if (state
< CEPH_MDS_STATE_RECONNECT
) {
1085 dout("send_renew_caps ignoring mds%d (%s)\n",
1086 session
->s_mds
, ceph_mds_state_name(state
));
1090 dout("send_renew_caps to mds%d (%s)\n", session
->s_mds
,
1091 ceph_mds_state_name(state
));
1092 msg
= create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS
,
1093 ++session
->s_renew_seq
);
1096 ceph_con_send(&session
->s_con
, msg
);
1101 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1103 * Called under session->s_mutex
1105 static void renewed_caps(struct ceph_mds_client
*mdsc
,
1106 struct ceph_mds_session
*session
, int is_renew
)
1111 spin_lock(&session
->s_cap_lock
);
1112 was_stale
= is_renew
&& time_after_eq(jiffies
, session
->s_cap_ttl
);
1114 session
->s_cap_ttl
= session
->s_renew_requested
+
1115 mdsc
->mdsmap
->m_session_timeout
*HZ
;
1118 if (time_before(jiffies
, session
->s_cap_ttl
)) {
1119 pr_info("mds%d caps renewed\n", session
->s_mds
);
1122 pr_info("mds%d caps still stale\n", session
->s_mds
);
1125 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1126 session
->s_mds
, session
->s_cap_ttl
, was_stale
? "stale" : "fresh",
1127 time_before(jiffies
, session
->s_cap_ttl
) ? "stale" : "fresh");
1128 spin_unlock(&session
->s_cap_lock
);
1131 wake_up_session_caps(session
, 0);
1135 * send a session close request
1137 static int request_close_session(struct ceph_mds_client
*mdsc
,
1138 struct ceph_mds_session
*session
)
1140 struct ceph_msg
*msg
;
1142 dout("request_close_session mds%d state %s seq %lld\n",
1143 session
->s_mds
, session_state_name(session
->s_state
),
1145 msg
= create_session_msg(CEPH_SESSION_REQUEST_CLOSE
, session
->s_seq
);
1148 ceph_con_send(&session
->s_con
, msg
);
1153 * Called with s_mutex held.
1155 static int __close_session(struct ceph_mds_client
*mdsc
,
1156 struct ceph_mds_session
*session
)
1158 if (session
->s_state
>= CEPH_MDS_SESSION_CLOSING
)
1160 session
->s_state
= CEPH_MDS_SESSION_CLOSING
;
1161 return request_close_session(mdsc
, session
);
1165 * Trim old(er) caps.
1167 * Because we can't cache an inode without one or more caps, we do
1168 * this indirectly: if a cap is unused, we prune its aliases, at which
1169 * point the inode will hopefully get dropped to.
1171 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1172 * memory pressure from the MDS, though, so it needn't be perfect.
1174 static int trim_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
, void *arg
)
1176 struct ceph_mds_session
*session
= arg
;
1177 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1178 int used
, oissued
, mine
;
1180 if (session
->s_trim_caps
<= 0)
1183 spin_lock(&ci
->i_ceph_lock
);
1184 mine
= cap
->issued
| cap
->implemented
;
1185 used
= __ceph_caps_used(ci
);
1186 oissued
= __ceph_caps_issued_other(ci
, cap
);
1188 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1189 inode
, cap
, ceph_cap_string(mine
), ceph_cap_string(oissued
),
1190 ceph_cap_string(used
));
1191 if (ci
->i_dirty_caps
)
1192 goto out
; /* dirty caps */
1193 if ((used
& ~oissued
) & mine
)
1194 goto out
; /* we need these caps */
1196 session
->s_trim_caps
--;
1198 /* we aren't the only cap.. just remove us */
1199 __ceph_remove_cap(cap
);
1201 /* try to drop referring dentries */
1202 spin_unlock(&ci
->i_ceph_lock
);
1203 d_prune_aliases(inode
);
1204 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1205 inode
, cap
, atomic_read(&inode
->i_count
));
1210 spin_unlock(&ci
->i_ceph_lock
);
1215 * Trim session cap count down to some max number.
1217 static int trim_caps(struct ceph_mds_client
*mdsc
,
1218 struct ceph_mds_session
*session
,
1221 int trim_caps
= session
->s_nr_caps
- max_caps
;
1223 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1224 session
->s_mds
, session
->s_nr_caps
, max_caps
, trim_caps
);
1225 if (trim_caps
> 0) {
1226 session
->s_trim_caps
= trim_caps
;
1227 iterate_session_caps(session
, trim_caps_cb
, session
);
1228 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1229 session
->s_mds
, session
->s_nr_caps
, max_caps
,
1230 trim_caps
- session
->s_trim_caps
);
1231 session
->s_trim_caps
= 0;
1237 * Allocate cap_release messages. If there is a partially full message
1238 * in the queue, try to allocate enough to cover it's remainder, so that
1239 * we can send it immediately.
1241 * Called under s_mutex.
1243 int ceph_add_cap_releases(struct ceph_mds_client
*mdsc
,
1244 struct ceph_mds_session
*session
)
1246 struct ceph_msg
*msg
, *partial
= NULL
;
1247 struct ceph_mds_cap_release
*head
;
1249 int extra
= mdsc
->fsc
->mount_options
->cap_release_safety
;
1252 dout("add_cap_releases %p mds%d extra %d\n", session
, session
->s_mds
,
1255 spin_lock(&session
->s_cap_lock
);
1257 if (!list_empty(&session
->s_cap_releases
)) {
1258 msg
= list_first_entry(&session
->s_cap_releases
,
1261 head
= msg
->front
.iov_base
;
1262 num
= le32_to_cpu(head
->num
);
1264 dout(" partial %p with (%d/%d)\n", msg
, num
,
1265 (int)CEPH_CAPS_PER_RELEASE
);
1266 extra
+= CEPH_CAPS_PER_RELEASE
- num
;
1270 while (session
->s_num_cap_releases
< session
->s_nr_caps
+ extra
) {
1271 spin_unlock(&session
->s_cap_lock
);
1272 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE
, PAGE_CACHE_SIZE
,
1276 dout("add_cap_releases %p msg %p now %d\n", session
, msg
,
1277 (int)msg
->front
.iov_len
);
1278 head
= msg
->front
.iov_base
;
1279 head
->num
= cpu_to_le32(0);
1280 msg
->front
.iov_len
= sizeof(*head
);
1281 spin_lock(&session
->s_cap_lock
);
1282 list_add(&msg
->list_head
, &session
->s_cap_releases
);
1283 session
->s_num_cap_releases
+= CEPH_CAPS_PER_RELEASE
;
1287 head
= partial
->front
.iov_base
;
1288 num
= le32_to_cpu(head
->num
);
1289 dout(" queueing partial %p with %d/%d\n", partial
, num
,
1290 (int)CEPH_CAPS_PER_RELEASE
);
1291 list_move_tail(&partial
->list_head
,
1292 &session
->s_cap_releases_done
);
1293 session
->s_num_cap_releases
-= CEPH_CAPS_PER_RELEASE
- num
;
1296 spin_unlock(&session
->s_cap_lock
);
1302 * flush all dirty inode data to disk.
1304 * returns true if we've flushed through want_flush_seq
1306 static int check_cap_flush(struct ceph_mds_client
*mdsc
, u64 want_flush_seq
)
1310 dout("check_cap_flush want %lld\n", want_flush_seq
);
1311 mutex_lock(&mdsc
->mutex
);
1312 for (mds
= 0; ret
&& mds
< mdsc
->max_sessions
; mds
++) {
1313 struct ceph_mds_session
*session
= mdsc
->sessions
[mds
];
1317 get_session(session
);
1318 mutex_unlock(&mdsc
->mutex
);
1320 mutex_lock(&session
->s_mutex
);
1321 if (!list_empty(&session
->s_cap_flushing
)) {
1322 struct ceph_inode_info
*ci
=
1323 list_entry(session
->s_cap_flushing
.next
,
1324 struct ceph_inode_info
,
1326 struct inode
*inode
= &ci
->vfs_inode
;
1328 spin_lock(&ci
->i_ceph_lock
);
1329 if (ci
->i_cap_flush_seq
<= want_flush_seq
) {
1330 dout("check_cap_flush still flushing %p "
1331 "seq %lld <= %lld to mds%d\n", inode
,
1332 ci
->i_cap_flush_seq
, want_flush_seq
,
1336 spin_unlock(&ci
->i_ceph_lock
);
1338 mutex_unlock(&session
->s_mutex
);
1339 ceph_put_mds_session(session
);
1343 mutex_lock(&mdsc
->mutex
);
1346 mutex_unlock(&mdsc
->mutex
);
1347 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq
);
1352 * called under s_mutex
1354 void ceph_send_cap_releases(struct ceph_mds_client
*mdsc
,
1355 struct ceph_mds_session
*session
)
1357 struct ceph_msg
*msg
;
1359 dout("send_cap_releases mds%d\n", session
->s_mds
);
1360 spin_lock(&session
->s_cap_lock
);
1361 while (!list_empty(&session
->s_cap_releases_done
)) {
1362 msg
= list_first_entry(&session
->s_cap_releases_done
,
1363 struct ceph_msg
, list_head
);
1364 list_del_init(&msg
->list_head
);
1365 spin_unlock(&session
->s_cap_lock
);
1366 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1367 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1368 ceph_con_send(&session
->s_con
, msg
);
1369 spin_lock(&session
->s_cap_lock
);
1371 spin_unlock(&session
->s_cap_lock
);
1374 static void discard_cap_releases(struct ceph_mds_client
*mdsc
,
1375 struct ceph_mds_session
*session
)
1377 struct ceph_msg
*msg
;
1378 struct ceph_mds_cap_release
*head
;
1381 dout("discard_cap_releases mds%d\n", session
->s_mds
);
1382 spin_lock(&session
->s_cap_lock
);
1384 /* zero out the in-progress message */
1385 msg
= list_first_entry(&session
->s_cap_releases
,
1386 struct ceph_msg
, list_head
);
1387 head
= msg
->front
.iov_base
;
1388 num
= le32_to_cpu(head
->num
);
1389 dout("discard_cap_releases mds%d %p %u\n", session
->s_mds
, msg
, num
);
1390 head
->num
= cpu_to_le32(0);
1391 session
->s_num_cap_releases
+= num
;
1393 /* requeue completed messages */
1394 while (!list_empty(&session
->s_cap_releases_done
)) {
1395 msg
= list_first_entry(&session
->s_cap_releases_done
,
1396 struct ceph_msg
, list_head
);
1397 list_del_init(&msg
->list_head
);
1399 head
= msg
->front
.iov_base
;
1400 num
= le32_to_cpu(head
->num
);
1401 dout("discard_cap_releases mds%d %p %u\n", session
->s_mds
, msg
,
1403 session
->s_num_cap_releases
+= num
;
1404 head
->num
= cpu_to_le32(0);
1405 msg
->front
.iov_len
= sizeof(*head
);
1406 list_add(&msg
->list_head
, &session
->s_cap_releases
);
1409 spin_unlock(&session
->s_cap_lock
);
1417 * Create an mds request.
1419 struct ceph_mds_request
*
1420 ceph_mdsc_create_request(struct ceph_mds_client
*mdsc
, int op
, int mode
)
1422 struct ceph_mds_request
*req
= kzalloc(sizeof(*req
), GFP_NOFS
);
1425 return ERR_PTR(-ENOMEM
);
1427 mutex_init(&req
->r_fill_mutex
);
1429 req
->r_started
= jiffies
;
1430 req
->r_resend_mds
= -1;
1431 INIT_LIST_HEAD(&req
->r_unsafe_dir_item
);
1433 kref_init(&req
->r_kref
);
1434 INIT_LIST_HEAD(&req
->r_wait
);
1435 init_completion(&req
->r_completion
);
1436 init_completion(&req
->r_safe_completion
);
1437 INIT_LIST_HEAD(&req
->r_unsafe_item
);
1440 req
->r_direct_mode
= mode
;
1445 * return oldest (lowest) request, tid in request tree, 0 if none.
1447 * called under mdsc->mutex.
1449 static struct ceph_mds_request
*__get_oldest_req(struct ceph_mds_client
*mdsc
)
1451 if (RB_EMPTY_ROOT(&mdsc
->request_tree
))
1453 return rb_entry(rb_first(&mdsc
->request_tree
),
1454 struct ceph_mds_request
, r_node
);
1457 static u64
__get_oldest_tid(struct ceph_mds_client
*mdsc
)
1459 struct ceph_mds_request
*req
= __get_oldest_req(mdsc
);
1467 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1468 * on build_path_from_dentry in fs/cifs/dir.c.
1470 * If @stop_on_nosnap, generate path relative to the first non-snapped
1473 * Encode hidden .snap dirs as a double /, i.e.
1474 * foo/.snap/bar -> foo//bar
1476 char *ceph_mdsc_build_path(struct dentry
*dentry
, int *plen
, u64
*base
,
1479 struct dentry
*temp
;
1485 return ERR_PTR(-EINVAL
);
1489 seq
= read_seqbegin(&rename_lock
);
1491 for (temp
= dentry
; !IS_ROOT(temp
);) {
1492 struct inode
*inode
= temp
->d_inode
;
1493 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
)
1494 len
++; /* slash only */
1495 else if (stop_on_nosnap
&& inode
&&
1496 ceph_snap(inode
) == CEPH_NOSNAP
)
1499 len
+= 1 + temp
->d_name
.len
;
1500 temp
= temp
->d_parent
;
1504 len
--; /* no leading '/' */
1506 path
= kmalloc(len
+1, GFP_NOFS
);
1508 return ERR_PTR(-ENOMEM
);
1510 path
[pos
] = 0; /* trailing null */
1512 for (temp
= dentry
; !IS_ROOT(temp
) && pos
!= 0; ) {
1513 struct inode
*inode
;
1515 spin_lock(&temp
->d_lock
);
1516 inode
= temp
->d_inode
;
1517 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
) {
1518 dout("build_path path+%d: %p SNAPDIR\n",
1520 } else if (stop_on_nosnap
&& inode
&&
1521 ceph_snap(inode
) == CEPH_NOSNAP
) {
1522 spin_unlock(&temp
->d_lock
);
1525 pos
-= temp
->d_name
.len
;
1527 spin_unlock(&temp
->d_lock
);
1530 strncpy(path
+ pos
, temp
->d_name
.name
,
1533 spin_unlock(&temp
->d_lock
);
1536 temp
= temp
->d_parent
;
1539 if (pos
!= 0 || read_seqretry(&rename_lock
, seq
)) {
1540 pr_err("build_path did not end path lookup where "
1541 "expected, namelen is %d, pos is %d\n", len
, pos
);
1542 /* presumably this is only possible if racing with a
1543 rename of one of the parent directories (we can not
1544 lock the dentries above us to prevent this, but
1545 retrying should be harmless) */
1550 *base
= ceph_ino(temp
->d_inode
);
1552 dout("build_path on %p %d built %llx '%.*s'\n",
1553 dentry
, dentry
->d_count
, *base
, len
, path
);
1557 static int build_dentry_path(struct dentry
*dentry
,
1558 const char **ppath
, int *ppathlen
, u64
*pino
,
1563 if (ceph_snap(dentry
->d_parent
->d_inode
) == CEPH_NOSNAP
) {
1564 *pino
= ceph_ino(dentry
->d_parent
->d_inode
);
1565 *ppath
= dentry
->d_name
.name
;
1566 *ppathlen
= dentry
->d_name
.len
;
1569 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1571 return PTR_ERR(path
);
1577 static int build_inode_path(struct inode
*inode
,
1578 const char **ppath
, int *ppathlen
, u64
*pino
,
1581 struct dentry
*dentry
;
1584 if (ceph_snap(inode
) == CEPH_NOSNAP
) {
1585 *pino
= ceph_ino(inode
);
1589 dentry
= d_find_alias(inode
);
1590 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1593 return PTR_ERR(path
);
1600 * request arguments may be specified via an inode *, a dentry *, or
1601 * an explicit ino+path.
1603 static int set_request_path_attr(struct inode
*rinode
, struct dentry
*rdentry
,
1604 const char *rpath
, u64 rino
,
1605 const char **ppath
, int *pathlen
,
1606 u64
*ino
, int *freepath
)
1611 r
= build_inode_path(rinode
, ppath
, pathlen
, ino
, freepath
);
1612 dout(" inode %p %llx.%llx\n", rinode
, ceph_ino(rinode
),
1614 } else if (rdentry
) {
1615 r
= build_dentry_path(rdentry
, ppath
, pathlen
, ino
, freepath
);
1616 dout(" dentry %p %llx/%.*s\n", rdentry
, *ino
, *pathlen
,
1618 } else if (rpath
|| rino
) {
1621 *pathlen
= rpath
? strlen(rpath
) : 0;
1622 dout(" path %.*s\n", *pathlen
, rpath
);
1629 * called under mdsc->mutex
1631 static struct ceph_msg
*create_request_message(struct ceph_mds_client
*mdsc
,
1632 struct ceph_mds_request
*req
,
1635 struct ceph_msg
*msg
;
1636 struct ceph_mds_request_head
*head
;
1637 const char *path1
= NULL
;
1638 const char *path2
= NULL
;
1639 u64 ino1
= 0, ino2
= 0;
1640 int pathlen1
= 0, pathlen2
= 0;
1641 int freepath1
= 0, freepath2
= 0;
1647 ret
= set_request_path_attr(req
->r_inode
, req
->r_dentry
,
1648 req
->r_path1
, req
->r_ino1
.ino
,
1649 &path1
, &pathlen1
, &ino1
, &freepath1
);
1655 ret
= set_request_path_attr(NULL
, req
->r_old_dentry
,
1656 req
->r_path2
, req
->r_ino2
.ino
,
1657 &path2
, &pathlen2
, &ino2
, &freepath2
);
1663 len
= sizeof(*head
) +
1664 pathlen1
+ pathlen2
+ 2*(1 + sizeof(u32
) + sizeof(u64
));
1666 /* calculate (max) length for cap releases */
1667 len
+= sizeof(struct ceph_mds_request_release
) *
1668 (!!req
->r_inode_drop
+ !!req
->r_dentry_drop
+
1669 !!req
->r_old_inode_drop
+ !!req
->r_old_dentry_drop
);
1670 if (req
->r_dentry_drop
)
1671 len
+= req
->r_dentry
->d_name
.len
;
1672 if (req
->r_old_dentry_drop
)
1673 len
+= req
->r_old_dentry
->d_name
.len
;
1675 msg
= ceph_msg_new(CEPH_MSG_CLIENT_REQUEST
, len
, GFP_NOFS
, false);
1677 msg
= ERR_PTR(-ENOMEM
);
1681 msg
->hdr
.tid
= cpu_to_le64(req
->r_tid
);
1683 head
= msg
->front
.iov_base
;
1684 p
= msg
->front
.iov_base
+ sizeof(*head
);
1685 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1687 head
->mdsmap_epoch
= cpu_to_le32(mdsc
->mdsmap
->m_epoch
);
1688 head
->op
= cpu_to_le32(req
->r_op
);
1689 head
->caller_uid
= cpu_to_le32(from_kuid(&init_user_ns
, req
->r_uid
));
1690 head
->caller_gid
= cpu_to_le32(from_kgid(&init_user_ns
, req
->r_gid
));
1691 head
->args
= req
->r_args
;
1693 ceph_encode_filepath(&p
, end
, ino1
, path1
);
1694 ceph_encode_filepath(&p
, end
, ino2
, path2
);
1696 /* make note of release offset, in case we need to replay */
1697 req
->r_request_release_offset
= p
- msg
->front
.iov_base
;
1701 if (req
->r_inode_drop
)
1702 releases
+= ceph_encode_inode_release(&p
,
1703 req
->r_inode
? req
->r_inode
: req
->r_dentry
->d_inode
,
1704 mds
, req
->r_inode_drop
, req
->r_inode_unless
, 0);
1705 if (req
->r_dentry_drop
)
1706 releases
+= ceph_encode_dentry_release(&p
, req
->r_dentry
,
1707 mds
, req
->r_dentry_drop
, req
->r_dentry_unless
);
1708 if (req
->r_old_dentry_drop
)
1709 releases
+= ceph_encode_dentry_release(&p
, req
->r_old_dentry
,
1710 mds
, req
->r_old_dentry_drop
, req
->r_old_dentry_unless
);
1711 if (req
->r_old_inode_drop
)
1712 releases
+= ceph_encode_inode_release(&p
,
1713 req
->r_old_dentry
->d_inode
,
1714 mds
, req
->r_old_inode_drop
, req
->r_old_inode_unless
, 0);
1715 head
->num_releases
= cpu_to_le16(releases
);
1718 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
1719 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1721 msg
->pages
= req
->r_pages
;
1722 msg
->nr_pages
= req
->r_num_pages
;
1723 msg
->hdr
.data_len
= cpu_to_le32(req
->r_data_len
);
1724 msg
->hdr
.data_off
= cpu_to_le16(0);
1728 kfree((char *)path2
);
1731 kfree((char *)path1
);
1737 * called under mdsc->mutex if error, under no mutex if
1740 static void complete_request(struct ceph_mds_client
*mdsc
,
1741 struct ceph_mds_request
*req
)
1743 if (req
->r_callback
)
1744 req
->r_callback(mdsc
, req
);
1746 complete_all(&req
->r_completion
);
1750 * called under mdsc->mutex
1752 static int __prepare_send_request(struct ceph_mds_client
*mdsc
,
1753 struct ceph_mds_request
*req
,
1756 struct ceph_mds_request_head
*rhead
;
1757 struct ceph_msg
*msg
;
1762 struct ceph_cap
*cap
=
1763 ceph_get_cap_for_mds(ceph_inode(req
->r_inode
), mds
);
1766 req
->r_sent_on_mseq
= cap
->mseq
;
1768 req
->r_sent_on_mseq
= -1;
1770 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req
,
1771 req
->r_tid
, ceph_mds_op_name(req
->r_op
), req
->r_attempts
);
1773 if (req
->r_got_unsafe
) {
1775 * Replay. Do not regenerate message (and rebuild
1776 * paths, etc.); just use the original message.
1777 * Rebuilding paths will break for renames because
1778 * d_move mangles the src name.
1780 msg
= req
->r_request
;
1781 rhead
= msg
->front
.iov_base
;
1783 flags
= le32_to_cpu(rhead
->flags
);
1784 flags
|= CEPH_MDS_FLAG_REPLAY
;
1785 rhead
->flags
= cpu_to_le32(flags
);
1787 if (req
->r_target_inode
)
1788 rhead
->ino
= cpu_to_le64(ceph_ino(req
->r_target_inode
));
1790 rhead
->num_retry
= req
->r_attempts
- 1;
1792 /* remove cap/dentry releases from message */
1793 rhead
->num_releases
= 0;
1794 msg
->hdr
.front_len
= cpu_to_le32(req
->r_request_release_offset
);
1795 msg
->front
.iov_len
= req
->r_request_release_offset
;
1799 if (req
->r_request
) {
1800 ceph_msg_put(req
->r_request
);
1801 req
->r_request
= NULL
;
1803 msg
= create_request_message(mdsc
, req
, mds
);
1805 req
->r_err
= PTR_ERR(msg
);
1806 complete_request(mdsc
, req
);
1807 return PTR_ERR(msg
);
1809 req
->r_request
= msg
;
1811 rhead
= msg
->front
.iov_base
;
1812 rhead
->oldest_client_tid
= cpu_to_le64(__get_oldest_tid(mdsc
));
1813 if (req
->r_got_unsafe
)
1814 flags
|= CEPH_MDS_FLAG_REPLAY
;
1815 if (req
->r_locked_dir
)
1816 flags
|= CEPH_MDS_FLAG_WANT_DENTRY
;
1817 rhead
->flags
= cpu_to_le32(flags
);
1818 rhead
->num_fwd
= req
->r_num_fwd
;
1819 rhead
->num_retry
= req
->r_attempts
- 1;
1822 dout(" r_locked_dir = %p\n", req
->r_locked_dir
);
1827 * send request, or put it on the appropriate wait list.
1829 static int __do_request(struct ceph_mds_client
*mdsc
,
1830 struct ceph_mds_request
*req
)
1832 struct ceph_mds_session
*session
= NULL
;
1836 if (req
->r_err
|| req
->r_got_result
)
1839 if (req
->r_timeout
&&
1840 time_after_eq(jiffies
, req
->r_started
+ req
->r_timeout
)) {
1841 dout("do_request timed out\n");
1846 put_request_session(req
);
1848 mds
= __choose_mds(mdsc
, req
);
1850 ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) < CEPH_MDS_STATE_ACTIVE
) {
1851 dout("do_request no mds or not active, waiting for map\n");
1852 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
1856 /* get, open session */
1857 session
= __ceph_lookup_mds_session(mdsc
, mds
);
1859 session
= register_session(mdsc
, mds
);
1860 if (IS_ERR(session
)) {
1861 err
= PTR_ERR(session
);
1865 req
->r_session
= get_session(session
);
1867 dout("do_request mds%d session %p state %s\n", mds
, session
,
1868 session_state_name(session
->s_state
));
1869 if (session
->s_state
!= CEPH_MDS_SESSION_OPEN
&&
1870 session
->s_state
!= CEPH_MDS_SESSION_HUNG
) {
1871 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
1872 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
1873 __open_session(mdsc
, session
);
1874 list_add(&req
->r_wait
, &session
->s_waiting
);
1879 req
->r_resend_mds
= -1; /* forget any previous mds hint */
1881 if (req
->r_request_started
== 0) /* note request start time */
1882 req
->r_request_started
= jiffies
;
1884 err
= __prepare_send_request(mdsc
, req
, mds
);
1886 ceph_msg_get(req
->r_request
);
1887 ceph_con_send(&session
->s_con
, req
->r_request
);
1891 ceph_put_mds_session(session
);
1897 complete_request(mdsc
, req
);
1902 * called under mdsc->mutex
1904 static void __wake_requests(struct ceph_mds_client
*mdsc
,
1905 struct list_head
*head
)
1907 struct ceph_mds_request
*req
;
1908 LIST_HEAD(tmp_list
);
1910 list_splice_init(head
, &tmp_list
);
1912 while (!list_empty(&tmp_list
)) {
1913 req
= list_entry(tmp_list
.next
,
1914 struct ceph_mds_request
, r_wait
);
1915 list_del_init(&req
->r_wait
);
1916 __do_request(mdsc
, req
);
1921 * Wake up threads with requests pending for @mds, so that they can
1922 * resubmit their requests to a possibly different mds.
1924 static void kick_requests(struct ceph_mds_client
*mdsc
, int mds
)
1926 struct ceph_mds_request
*req
;
1929 dout("kick_requests mds%d\n", mds
);
1930 for (p
= rb_first(&mdsc
->request_tree
); p
; p
= rb_next(p
)) {
1931 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
1932 if (req
->r_got_unsafe
)
1934 if (req
->r_session
&&
1935 req
->r_session
->s_mds
== mds
) {
1936 dout(" kicking tid %llu\n", req
->r_tid
);
1937 __do_request(mdsc
, req
);
1942 void ceph_mdsc_submit_request(struct ceph_mds_client
*mdsc
,
1943 struct ceph_mds_request
*req
)
1945 dout("submit_request on %p\n", req
);
1946 mutex_lock(&mdsc
->mutex
);
1947 __register_request(mdsc
, req
, NULL
);
1948 __do_request(mdsc
, req
);
1949 mutex_unlock(&mdsc
->mutex
);
1953 * Synchrously perform an mds request. Take care of all of the
1954 * session setup, forwarding, retry details.
1956 int ceph_mdsc_do_request(struct ceph_mds_client
*mdsc
,
1958 struct ceph_mds_request
*req
)
1962 dout("do_request on %p\n", req
);
1964 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1966 ceph_get_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
1967 if (req
->r_locked_dir
)
1968 ceph_get_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
1969 if (req
->r_old_dentry
)
1970 ceph_get_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
1974 mutex_lock(&mdsc
->mutex
);
1975 __register_request(mdsc
, req
, dir
);
1976 __do_request(mdsc
, req
);
1980 __unregister_request(mdsc
, req
);
1981 dout("do_request early error %d\n", err
);
1986 mutex_unlock(&mdsc
->mutex
);
1987 dout("do_request waiting\n");
1988 if (req
->r_timeout
) {
1989 err
= (long)wait_for_completion_killable_timeout(
1990 &req
->r_completion
, req
->r_timeout
);
1994 err
= wait_for_completion_killable(&req
->r_completion
);
1996 dout("do_request waited, got %d\n", err
);
1997 mutex_lock(&mdsc
->mutex
);
1999 /* only abort if we didn't race with a real reply */
2000 if (req
->r_got_result
) {
2001 err
= le32_to_cpu(req
->r_reply_info
.head
->result
);
2002 } else if (err
< 0) {
2003 dout("aborted request %lld with %d\n", req
->r_tid
, err
);
2006 * ensure we aren't running concurrently with
2007 * ceph_fill_trace or ceph_readdir_prepopulate, which
2008 * rely on locks (dir mutex) held by our caller.
2010 mutex_lock(&req
->r_fill_mutex
);
2012 req
->r_aborted
= true;
2013 mutex_unlock(&req
->r_fill_mutex
);
2015 if (req
->r_locked_dir
&&
2016 (req
->r_op
& CEPH_MDS_OP_WRITE
))
2017 ceph_invalidate_dir_request(req
);
2023 mutex_unlock(&mdsc
->mutex
);
2024 dout("do_request %p done, result %d\n", req
, err
);
2029 * Invalidate dir D_COMPLETE, dentry lease state on an aborted MDS
2030 * namespace request.
2032 void ceph_invalidate_dir_request(struct ceph_mds_request
*req
)
2034 struct inode
*inode
= req
->r_locked_dir
;
2035 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2037 dout("invalidate_dir_request %p (D_COMPLETE, lease(s))\n", inode
);
2038 spin_lock(&ci
->i_ceph_lock
);
2039 ceph_dir_clear_complete(inode
);
2040 ci
->i_release_count
++;
2041 spin_unlock(&ci
->i_ceph_lock
);
2044 ceph_invalidate_dentry_lease(req
->r_dentry
);
2045 if (req
->r_old_dentry
)
2046 ceph_invalidate_dentry_lease(req
->r_old_dentry
);
2052 * We take the session mutex and parse and process the reply immediately.
2053 * This preserves the logical ordering of replies, capabilities, etc., sent
2054 * by the MDS as they are applied to our local cache.
2056 static void handle_reply(struct ceph_mds_session
*session
, struct ceph_msg
*msg
)
2058 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2059 struct ceph_mds_request
*req
;
2060 struct ceph_mds_reply_head
*head
= msg
->front
.iov_base
;
2061 struct ceph_mds_reply_info_parsed
*rinfo
; /* parsed reply info */
2064 int mds
= session
->s_mds
;
2066 if (msg
->front
.iov_len
< sizeof(*head
)) {
2067 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2072 /* get request, session */
2073 tid
= le64_to_cpu(msg
->hdr
.tid
);
2074 mutex_lock(&mdsc
->mutex
);
2075 req
= __lookup_request(mdsc
, tid
);
2077 dout("handle_reply on unknown tid %llu\n", tid
);
2078 mutex_unlock(&mdsc
->mutex
);
2081 dout("handle_reply %p\n", req
);
2083 /* correct session? */
2084 if (req
->r_session
!= session
) {
2085 pr_err("mdsc_handle_reply got %llu on session mds%d"
2086 " not mds%d\n", tid
, session
->s_mds
,
2087 req
->r_session
? req
->r_session
->s_mds
: -1);
2088 mutex_unlock(&mdsc
->mutex
);
2093 if ((req
->r_got_unsafe
&& !head
->safe
) ||
2094 (req
->r_got_safe
&& head
->safe
)) {
2095 pr_warning("got a dup %s reply on %llu from mds%d\n",
2096 head
->safe
? "safe" : "unsafe", tid
, mds
);
2097 mutex_unlock(&mdsc
->mutex
);
2100 if (req
->r_got_safe
&& !head
->safe
) {
2101 pr_warning("got unsafe after safe on %llu from mds%d\n",
2103 mutex_unlock(&mdsc
->mutex
);
2107 result
= le32_to_cpu(head
->result
);
2111 * if we're not talking to the authority, send to them
2112 * if the authority has changed while we weren't looking,
2113 * send to new authority
2114 * Otherwise we just have to return an ESTALE
2116 if (result
== -ESTALE
) {
2117 dout("got ESTALE on request %llu", req
->r_tid
);
2118 if (!req
->r_inode
) {
2119 /* do nothing; not an authority problem */
2120 } else if (req
->r_direct_mode
!= USE_AUTH_MDS
) {
2121 dout("not using auth, setting for that now");
2122 req
->r_direct_mode
= USE_AUTH_MDS
;
2123 __do_request(mdsc
, req
);
2124 mutex_unlock(&mdsc
->mutex
);
2127 struct ceph_inode_info
*ci
= ceph_inode(req
->r_inode
);
2128 struct ceph_cap
*cap
= NULL
;
2131 cap
= ceph_get_cap_for_mds(ci
,
2132 req
->r_session
->s_mds
);
2134 dout("already using auth");
2135 if ((!cap
|| cap
!= ci
->i_auth_cap
) ||
2136 (cap
->mseq
!= req
->r_sent_on_mseq
)) {
2137 dout("but cap changed, so resending");
2138 __do_request(mdsc
, req
);
2139 mutex_unlock(&mdsc
->mutex
);
2143 dout("have to return ESTALE on request %llu", req
->r_tid
);
2148 req
->r_got_safe
= true;
2149 __unregister_request(mdsc
, req
);
2150 complete_all(&req
->r_safe_completion
);
2152 if (req
->r_got_unsafe
) {
2154 * We already handled the unsafe response, now do the
2155 * cleanup. No need to examine the response; the MDS
2156 * doesn't include any result info in the safe
2157 * response. And even if it did, there is nothing
2158 * useful we could do with a revised return value.
2160 dout("got safe reply %llu, mds%d\n", tid
, mds
);
2161 list_del_init(&req
->r_unsafe_item
);
2163 /* last unsafe request during umount? */
2164 if (mdsc
->stopping
&& !__get_oldest_req(mdsc
))
2165 complete_all(&mdsc
->safe_umount_waiters
);
2166 mutex_unlock(&mdsc
->mutex
);
2170 req
->r_got_unsafe
= true;
2171 list_add_tail(&req
->r_unsafe_item
, &req
->r_session
->s_unsafe
);
2174 dout("handle_reply tid %lld result %d\n", tid
, result
);
2175 rinfo
= &req
->r_reply_info
;
2176 err
= parse_reply_info(msg
, rinfo
, session
->s_con
.peer_features
);
2177 mutex_unlock(&mdsc
->mutex
);
2179 mutex_lock(&session
->s_mutex
);
2181 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds
, tid
);
2187 if (rinfo
->snapblob_len
) {
2188 down_write(&mdsc
->snap_rwsem
);
2189 ceph_update_snap_trace(mdsc
, rinfo
->snapblob
,
2190 rinfo
->snapblob
+ rinfo
->snapblob_len
,
2191 le32_to_cpu(head
->op
) == CEPH_MDS_OP_RMSNAP
);
2192 downgrade_write(&mdsc
->snap_rwsem
);
2194 down_read(&mdsc
->snap_rwsem
);
2197 /* insert trace into our cache */
2198 mutex_lock(&req
->r_fill_mutex
);
2199 err
= ceph_fill_trace(mdsc
->fsc
->sb
, req
, req
->r_session
);
2201 if (result
== 0 && (req
->r_op
== CEPH_MDS_OP_READDIR
||
2202 req
->r_op
== CEPH_MDS_OP_LSSNAP
) &&
2204 ceph_readdir_prepopulate(req
, req
->r_session
);
2205 ceph_unreserve_caps(mdsc
, &req
->r_caps_reservation
);
2207 mutex_unlock(&req
->r_fill_mutex
);
2209 up_read(&mdsc
->snap_rwsem
);
2211 mutex_lock(&mdsc
->mutex
);
2212 if (!req
->r_aborted
) {
2218 req
->r_got_result
= true;
2221 dout("reply arrived after request %lld was aborted\n", tid
);
2223 mutex_unlock(&mdsc
->mutex
);
2225 ceph_add_cap_releases(mdsc
, req
->r_session
);
2226 mutex_unlock(&session
->s_mutex
);
2228 /* kick calling process */
2229 complete_request(mdsc
, req
);
2231 ceph_mdsc_put_request(req
);
2238 * handle mds notification that our request has been forwarded.
2240 static void handle_forward(struct ceph_mds_client
*mdsc
,
2241 struct ceph_mds_session
*session
,
2242 struct ceph_msg
*msg
)
2244 struct ceph_mds_request
*req
;
2245 u64 tid
= le64_to_cpu(msg
->hdr
.tid
);
2249 void *p
= msg
->front
.iov_base
;
2250 void *end
= p
+ msg
->front
.iov_len
;
2252 ceph_decode_need(&p
, end
, 2*sizeof(u32
), bad
);
2253 next_mds
= ceph_decode_32(&p
);
2254 fwd_seq
= ceph_decode_32(&p
);
2256 mutex_lock(&mdsc
->mutex
);
2257 req
= __lookup_request(mdsc
, tid
);
2259 dout("forward tid %llu to mds%d - req dne\n", tid
, next_mds
);
2260 goto out
; /* dup reply? */
2263 if (req
->r_aborted
) {
2264 dout("forward tid %llu aborted, unregistering\n", tid
);
2265 __unregister_request(mdsc
, req
);
2266 } else if (fwd_seq
<= req
->r_num_fwd
) {
2267 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2268 tid
, next_mds
, req
->r_num_fwd
, fwd_seq
);
2270 /* resend. forward race not possible; mds would drop */
2271 dout("forward tid %llu to mds%d (we resend)\n", tid
, next_mds
);
2273 BUG_ON(req
->r_got_result
);
2274 req
->r_num_fwd
= fwd_seq
;
2275 req
->r_resend_mds
= next_mds
;
2276 put_request_session(req
);
2277 __do_request(mdsc
, req
);
2279 ceph_mdsc_put_request(req
);
2281 mutex_unlock(&mdsc
->mutex
);
2285 pr_err("mdsc_handle_forward decode error err=%d\n", err
);
2289 * handle a mds session control message
2291 static void handle_session(struct ceph_mds_session
*session
,
2292 struct ceph_msg
*msg
)
2294 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2297 int mds
= session
->s_mds
;
2298 struct ceph_mds_session_head
*h
= msg
->front
.iov_base
;
2302 if (msg
->front
.iov_len
!= sizeof(*h
))
2304 op
= le32_to_cpu(h
->op
);
2305 seq
= le64_to_cpu(h
->seq
);
2307 mutex_lock(&mdsc
->mutex
);
2308 if (op
== CEPH_SESSION_CLOSE
)
2309 __unregister_session(mdsc
, session
);
2310 /* FIXME: this ttl calculation is generous */
2311 session
->s_ttl
= jiffies
+ HZ
*mdsc
->mdsmap
->m_session_autoclose
;
2312 mutex_unlock(&mdsc
->mutex
);
2314 mutex_lock(&session
->s_mutex
);
2316 dout("handle_session mds%d %s %p state %s seq %llu\n",
2317 mds
, ceph_session_op_name(op
), session
,
2318 session_state_name(session
->s_state
), seq
);
2320 if (session
->s_state
== CEPH_MDS_SESSION_HUNG
) {
2321 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2322 pr_info("mds%d came back\n", session
->s_mds
);
2326 case CEPH_SESSION_OPEN
:
2327 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2328 pr_info("mds%d reconnect success\n", session
->s_mds
);
2329 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2330 renewed_caps(mdsc
, session
, 0);
2333 __close_session(mdsc
, session
);
2336 case CEPH_SESSION_RENEWCAPS
:
2337 if (session
->s_renew_seq
== seq
)
2338 renewed_caps(mdsc
, session
, 1);
2341 case CEPH_SESSION_CLOSE
:
2342 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2343 pr_info("mds%d reconnect denied\n", session
->s_mds
);
2344 remove_session_caps(session
);
2345 wake
= 1; /* for good measure */
2346 wake_up_all(&mdsc
->session_close_wq
);
2347 kick_requests(mdsc
, mds
);
2350 case CEPH_SESSION_STALE
:
2351 pr_info("mds%d caps went stale, renewing\n",
2353 spin_lock(&session
->s_gen_ttl_lock
);
2354 session
->s_cap_gen
++;
2355 session
->s_cap_ttl
= jiffies
- 1;
2356 spin_unlock(&session
->s_gen_ttl_lock
);
2357 send_renew_caps(mdsc
, session
);
2360 case CEPH_SESSION_RECALL_STATE
:
2361 trim_caps(mdsc
, session
, le32_to_cpu(h
->max_caps
));
2365 pr_err("mdsc_handle_session bad op %d mds%d\n", op
, mds
);
2369 mutex_unlock(&session
->s_mutex
);
2371 mutex_lock(&mdsc
->mutex
);
2372 __wake_requests(mdsc
, &session
->s_waiting
);
2373 mutex_unlock(&mdsc
->mutex
);
2378 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds
,
2379 (int)msg
->front
.iov_len
);
2386 * called under session->mutex.
2388 static void replay_unsafe_requests(struct ceph_mds_client
*mdsc
,
2389 struct ceph_mds_session
*session
)
2391 struct ceph_mds_request
*req
, *nreq
;
2394 dout("replay_unsafe_requests mds%d\n", session
->s_mds
);
2396 mutex_lock(&mdsc
->mutex
);
2397 list_for_each_entry_safe(req
, nreq
, &session
->s_unsafe
, r_unsafe_item
) {
2398 err
= __prepare_send_request(mdsc
, req
, session
->s_mds
);
2400 ceph_msg_get(req
->r_request
);
2401 ceph_con_send(&session
->s_con
, req
->r_request
);
2404 mutex_unlock(&mdsc
->mutex
);
2408 * Encode information about a cap for a reconnect with the MDS.
2410 static int encode_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
2414 struct ceph_mds_cap_reconnect v2
;
2415 struct ceph_mds_cap_reconnect_v1 v1
;
2418 struct ceph_inode_info
*ci
;
2419 struct ceph_reconnect_state
*recon_state
= arg
;
2420 struct ceph_pagelist
*pagelist
= recon_state
->pagelist
;
2424 struct dentry
*dentry
;
2428 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2429 inode
, ceph_vinop(inode
), cap
, cap
->cap_id
,
2430 ceph_cap_string(cap
->issued
));
2431 err
= ceph_pagelist_encode_64(pagelist
, ceph_ino(inode
));
2435 dentry
= d_find_alias(inode
);
2437 path
= ceph_mdsc_build_path(dentry
, &pathlen
, &pathbase
, 0);
2439 err
= PTR_ERR(path
);
2446 err
= ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
2450 spin_lock(&ci
->i_ceph_lock
);
2451 cap
->seq
= 0; /* reset cap seq */
2452 cap
->issue_seq
= 0; /* and issue_seq */
2454 if (recon_state
->flock
) {
2455 rec
.v2
.cap_id
= cpu_to_le64(cap
->cap_id
);
2456 rec
.v2
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2457 rec
.v2
.issued
= cpu_to_le32(cap
->issued
);
2458 rec
.v2
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2459 rec
.v2
.pathbase
= cpu_to_le64(pathbase
);
2460 rec
.v2
.flock_len
= 0;
2461 reclen
= sizeof(rec
.v2
);
2463 rec
.v1
.cap_id
= cpu_to_le64(cap
->cap_id
);
2464 rec
.v1
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2465 rec
.v1
.issued
= cpu_to_le32(cap
->issued
);
2466 rec
.v1
.size
= cpu_to_le64(inode
->i_size
);
2467 ceph_encode_timespec(&rec
.v1
.mtime
, &inode
->i_mtime
);
2468 ceph_encode_timespec(&rec
.v1
.atime
, &inode
->i_atime
);
2469 rec
.v1
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2470 rec
.v1
.pathbase
= cpu_to_le64(pathbase
);
2471 reclen
= sizeof(rec
.v1
);
2473 spin_unlock(&ci
->i_ceph_lock
);
2475 if (recon_state
->flock
) {
2476 int num_fcntl_locks
, num_flock_locks
;
2477 struct ceph_pagelist_cursor trunc_point
;
2479 ceph_pagelist_set_cursor(pagelist
, &trunc_point
);
2482 ceph_count_locks(inode
, &num_fcntl_locks
,
2484 rec
.v2
.flock_len
= (2*sizeof(u32
) +
2485 (num_fcntl_locks
+num_flock_locks
) *
2486 sizeof(struct ceph_filelock
));
2489 /* pre-alloc pagelist */
2490 ceph_pagelist_truncate(pagelist
, &trunc_point
);
2491 err
= ceph_pagelist_append(pagelist
, &rec
, reclen
);
2493 err
= ceph_pagelist_reserve(pagelist
,
2499 err
= ceph_encode_locks(inode
,
2505 } while (err
== -ENOSPC
);
2507 err
= ceph_pagelist_append(pagelist
, &rec
, reclen
);
2519 * If an MDS fails and recovers, clients need to reconnect in order to
2520 * reestablish shared state. This includes all caps issued through
2521 * this session _and_ the snap_realm hierarchy. Because it's not
2522 * clear which snap realms the mds cares about, we send everything we
2523 * know about.. that ensures we'll then get any new info the
2524 * recovering MDS might have.
2526 * This is a relatively heavyweight operation, but it's rare.
2528 * called with mdsc->mutex held.
2530 static void send_mds_reconnect(struct ceph_mds_client
*mdsc
,
2531 struct ceph_mds_session
*session
)
2533 struct ceph_msg
*reply
;
2535 int mds
= session
->s_mds
;
2537 struct ceph_pagelist
*pagelist
;
2538 struct ceph_reconnect_state recon_state
;
2540 pr_info("mds%d reconnect start\n", mds
);
2542 pagelist
= kmalloc(sizeof(*pagelist
), GFP_NOFS
);
2544 goto fail_nopagelist
;
2545 ceph_pagelist_init(pagelist
);
2547 reply
= ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT
, 0, GFP_NOFS
, false);
2551 mutex_lock(&session
->s_mutex
);
2552 session
->s_state
= CEPH_MDS_SESSION_RECONNECTING
;
2555 ceph_con_close(&session
->s_con
);
2556 ceph_con_open(&session
->s_con
,
2557 CEPH_ENTITY_TYPE_MDS
, mds
,
2558 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
2560 /* replay unsafe requests */
2561 replay_unsafe_requests(mdsc
, session
);
2563 down_read(&mdsc
->snap_rwsem
);
2565 dout("session %p state %s\n", session
,
2566 session_state_name(session
->s_state
));
2568 /* drop old cap expires; we're about to reestablish that state */
2569 discard_cap_releases(mdsc
, session
);
2571 /* traverse this session's caps */
2572 err
= ceph_pagelist_encode_32(pagelist
, session
->s_nr_caps
);
2576 recon_state
.pagelist
= pagelist
;
2577 recon_state
.flock
= session
->s_con
.peer_features
& CEPH_FEATURE_FLOCK
;
2578 err
= iterate_session_caps(session
, encode_caps_cb
, &recon_state
);
2583 * snaprealms. we provide mds with the ino, seq (version), and
2584 * parent for all of our realms. If the mds has any newer info,
2587 for (p
= rb_first(&mdsc
->snap_realms
); p
; p
= rb_next(p
)) {
2588 struct ceph_snap_realm
*realm
=
2589 rb_entry(p
, struct ceph_snap_realm
, node
);
2590 struct ceph_mds_snaprealm_reconnect sr_rec
;
2592 dout(" adding snap realm %llx seq %lld parent %llx\n",
2593 realm
->ino
, realm
->seq
, realm
->parent_ino
);
2594 sr_rec
.ino
= cpu_to_le64(realm
->ino
);
2595 sr_rec
.seq
= cpu_to_le64(realm
->seq
);
2596 sr_rec
.parent
= cpu_to_le64(realm
->parent_ino
);
2597 err
= ceph_pagelist_append(pagelist
, &sr_rec
, sizeof(sr_rec
));
2602 reply
->pagelist
= pagelist
;
2603 if (recon_state
.flock
)
2604 reply
->hdr
.version
= cpu_to_le16(2);
2605 reply
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
2606 reply
->nr_pages
= calc_pages_for(0, pagelist
->length
);
2607 ceph_con_send(&session
->s_con
, reply
);
2609 mutex_unlock(&session
->s_mutex
);
2611 mutex_lock(&mdsc
->mutex
);
2612 __wake_requests(mdsc
, &session
->s_waiting
);
2613 mutex_unlock(&mdsc
->mutex
);
2615 up_read(&mdsc
->snap_rwsem
);
2619 ceph_msg_put(reply
);
2620 up_read(&mdsc
->snap_rwsem
);
2621 mutex_unlock(&session
->s_mutex
);
2623 ceph_pagelist_release(pagelist
);
2626 pr_err("error %d preparing reconnect for mds%d\n", err
, mds
);
2632 * compare old and new mdsmaps, kicking requests
2633 * and closing out old connections as necessary
2635 * called under mdsc->mutex.
2637 static void check_new_map(struct ceph_mds_client
*mdsc
,
2638 struct ceph_mdsmap
*newmap
,
2639 struct ceph_mdsmap
*oldmap
)
2642 int oldstate
, newstate
;
2643 struct ceph_mds_session
*s
;
2645 dout("check_new_map new %u old %u\n",
2646 newmap
->m_epoch
, oldmap
->m_epoch
);
2648 for (i
= 0; i
< oldmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
2649 if (mdsc
->sessions
[i
] == NULL
)
2651 s
= mdsc
->sessions
[i
];
2652 oldstate
= ceph_mdsmap_get_state(oldmap
, i
);
2653 newstate
= ceph_mdsmap_get_state(newmap
, i
);
2655 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2656 i
, ceph_mds_state_name(oldstate
),
2657 ceph_mdsmap_is_laggy(oldmap
, i
) ? " (laggy)" : "",
2658 ceph_mds_state_name(newstate
),
2659 ceph_mdsmap_is_laggy(newmap
, i
) ? " (laggy)" : "",
2660 session_state_name(s
->s_state
));
2662 if (i
>= newmap
->m_max_mds
||
2663 memcmp(ceph_mdsmap_get_addr(oldmap
, i
),
2664 ceph_mdsmap_get_addr(newmap
, i
),
2665 sizeof(struct ceph_entity_addr
))) {
2666 if (s
->s_state
== CEPH_MDS_SESSION_OPENING
) {
2667 /* the session never opened, just close it
2669 __wake_requests(mdsc
, &s
->s_waiting
);
2670 __unregister_session(mdsc
, s
);
2673 mutex_unlock(&mdsc
->mutex
);
2674 mutex_lock(&s
->s_mutex
);
2675 mutex_lock(&mdsc
->mutex
);
2676 ceph_con_close(&s
->s_con
);
2677 mutex_unlock(&s
->s_mutex
);
2678 s
->s_state
= CEPH_MDS_SESSION_RESTARTING
;
2681 /* kick any requests waiting on the recovering mds */
2682 kick_requests(mdsc
, i
);
2683 } else if (oldstate
== newstate
) {
2684 continue; /* nothing new with this mds */
2690 if (s
->s_state
== CEPH_MDS_SESSION_RESTARTING
&&
2691 newstate
>= CEPH_MDS_STATE_RECONNECT
) {
2692 mutex_unlock(&mdsc
->mutex
);
2693 send_mds_reconnect(mdsc
, s
);
2694 mutex_lock(&mdsc
->mutex
);
2698 * kick request on any mds that has gone active.
2700 if (oldstate
< CEPH_MDS_STATE_ACTIVE
&&
2701 newstate
>= CEPH_MDS_STATE_ACTIVE
) {
2702 if (oldstate
!= CEPH_MDS_STATE_CREATING
&&
2703 oldstate
!= CEPH_MDS_STATE_STARTING
)
2704 pr_info("mds%d recovery completed\n", s
->s_mds
);
2705 kick_requests(mdsc
, i
);
2706 ceph_kick_flushing_caps(mdsc
, s
);
2707 wake_up_session_caps(s
, 1);
2711 for (i
= 0; i
< newmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
2712 s
= mdsc
->sessions
[i
];
2715 if (!ceph_mdsmap_is_laggy(newmap
, i
))
2717 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
2718 s
->s_state
== CEPH_MDS_SESSION_HUNG
||
2719 s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
2720 dout(" connecting to export targets of laggy mds%d\n",
2722 __open_export_target_sessions(mdsc
, s
);
2734 * caller must hold session s_mutex, dentry->d_lock
2736 void __ceph_mdsc_drop_dentry_lease(struct dentry
*dentry
)
2738 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
2740 ceph_put_mds_session(di
->lease_session
);
2741 di
->lease_session
= NULL
;
2744 static void handle_lease(struct ceph_mds_client
*mdsc
,
2745 struct ceph_mds_session
*session
,
2746 struct ceph_msg
*msg
)
2748 struct super_block
*sb
= mdsc
->fsc
->sb
;
2749 struct inode
*inode
;
2750 struct dentry
*parent
, *dentry
;
2751 struct ceph_dentry_info
*di
;
2752 int mds
= session
->s_mds
;
2753 struct ceph_mds_lease
*h
= msg
->front
.iov_base
;
2755 struct ceph_vino vino
;
2759 dout("handle_lease from mds%d\n", mds
);
2762 if (msg
->front
.iov_len
< sizeof(*h
) + sizeof(u32
))
2764 vino
.ino
= le64_to_cpu(h
->ino
);
2765 vino
.snap
= CEPH_NOSNAP
;
2766 seq
= le32_to_cpu(h
->seq
);
2767 dname
.name
= (void *)h
+ sizeof(*h
) + sizeof(u32
);
2768 dname
.len
= msg
->front
.iov_len
- sizeof(*h
) - sizeof(u32
);
2769 if (dname
.len
!= get_unaligned_le32(h
+1))
2772 mutex_lock(&session
->s_mutex
);
2776 inode
= ceph_find_inode(sb
, vino
);
2777 dout("handle_lease %s, ino %llx %p %.*s\n",
2778 ceph_lease_op_name(h
->action
), vino
.ino
, inode
,
2779 dname
.len
, dname
.name
);
2780 if (inode
== NULL
) {
2781 dout("handle_lease no inode %llx\n", vino
.ino
);
2786 parent
= d_find_alias(inode
);
2788 dout("no parent dentry on inode %p\n", inode
);
2790 goto release
; /* hrm... */
2792 dname
.hash
= full_name_hash(dname
.name
, dname
.len
);
2793 dentry
= d_lookup(parent
, &dname
);
2798 spin_lock(&dentry
->d_lock
);
2799 di
= ceph_dentry(dentry
);
2800 switch (h
->action
) {
2801 case CEPH_MDS_LEASE_REVOKE
:
2802 if (di
->lease_session
== session
) {
2803 if (ceph_seq_cmp(di
->lease_seq
, seq
) > 0)
2804 h
->seq
= cpu_to_le32(di
->lease_seq
);
2805 __ceph_mdsc_drop_dentry_lease(dentry
);
2810 case CEPH_MDS_LEASE_RENEW
:
2811 if (di
->lease_session
== session
&&
2812 di
->lease_gen
== session
->s_cap_gen
&&
2813 di
->lease_renew_from
&&
2814 di
->lease_renew_after
== 0) {
2815 unsigned long duration
=
2816 le32_to_cpu(h
->duration_ms
) * HZ
/ 1000;
2818 di
->lease_seq
= seq
;
2819 dentry
->d_time
= di
->lease_renew_from
+ duration
;
2820 di
->lease_renew_after
= di
->lease_renew_from
+
2822 di
->lease_renew_from
= 0;
2826 spin_unlock(&dentry
->d_lock
);
2833 /* let's just reuse the same message */
2834 h
->action
= CEPH_MDS_LEASE_REVOKE_ACK
;
2836 ceph_con_send(&session
->s_con
, msg
);
2840 mutex_unlock(&session
->s_mutex
);
2844 pr_err("corrupt lease message\n");
2848 void ceph_mdsc_lease_send_msg(struct ceph_mds_session
*session
,
2849 struct inode
*inode
,
2850 struct dentry
*dentry
, char action
,
2853 struct ceph_msg
*msg
;
2854 struct ceph_mds_lease
*lease
;
2855 int len
= sizeof(*lease
) + sizeof(u32
);
2858 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2859 inode
, dentry
, ceph_lease_op_name(action
), session
->s_mds
);
2860 dnamelen
= dentry
->d_name
.len
;
2863 msg
= ceph_msg_new(CEPH_MSG_CLIENT_LEASE
, len
, GFP_NOFS
, false);
2866 lease
= msg
->front
.iov_base
;
2867 lease
->action
= action
;
2868 lease
->ino
= cpu_to_le64(ceph_vino(inode
).ino
);
2869 lease
->first
= lease
->last
= cpu_to_le64(ceph_vino(inode
).snap
);
2870 lease
->seq
= cpu_to_le32(seq
);
2871 put_unaligned_le32(dnamelen
, lease
+ 1);
2872 memcpy((void *)(lease
+ 1) + 4, dentry
->d_name
.name
, dnamelen
);
2875 * if this is a preemptive lease RELEASE, no need to
2876 * flush request stream, since the actual request will
2879 msg
->more_to_follow
= (action
== CEPH_MDS_LEASE_RELEASE
);
2881 ceph_con_send(&session
->s_con
, msg
);
2885 * Preemptively release a lease we expect to invalidate anyway.
2886 * Pass @inode always, @dentry is optional.
2888 void ceph_mdsc_lease_release(struct ceph_mds_client
*mdsc
, struct inode
*inode
,
2889 struct dentry
*dentry
)
2891 struct ceph_dentry_info
*di
;
2892 struct ceph_mds_session
*session
;
2895 BUG_ON(inode
== NULL
);
2896 BUG_ON(dentry
== NULL
);
2898 /* is dentry lease valid? */
2899 spin_lock(&dentry
->d_lock
);
2900 di
= ceph_dentry(dentry
);
2901 if (!di
|| !di
->lease_session
||
2902 di
->lease_session
->s_mds
< 0 ||
2903 di
->lease_gen
!= di
->lease_session
->s_cap_gen
||
2904 !time_before(jiffies
, dentry
->d_time
)) {
2905 dout("lease_release inode %p dentry %p -- "
2908 spin_unlock(&dentry
->d_lock
);
2912 /* we do have a lease on this dentry; note mds and seq */
2913 session
= ceph_get_mds_session(di
->lease_session
);
2914 seq
= di
->lease_seq
;
2915 __ceph_mdsc_drop_dentry_lease(dentry
);
2916 spin_unlock(&dentry
->d_lock
);
2918 dout("lease_release inode %p dentry %p to mds%d\n",
2919 inode
, dentry
, session
->s_mds
);
2920 ceph_mdsc_lease_send_msg(session
, inode
, dentry
,
2921 CEPH_MDS_LEASE_RELEASE
, seq
);
2922 ceph_put_mds_session(session
);
2926 * drop all leases (and dentry refs) in preparation for umount
2928 static void drop_leases(struct ceph_mds_client
*mdsc
)
2932 dout("drop_leases\n");
2933 mutex_lock(&mdsc
->mutex
);
2934 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
2935 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
2938 mutex_unlock(&mdsc
->mutex
);
2939 mutex_lock(&s
->s_mutex
);
2940 mutex_unlock(&s
->s_mutex
);
2941 ceph_put_mds_session(s
);
2942 mutex_lock(&mdsc
->mutex
);
2944 mutex_unlock(&mdsc
->mutex
);
2950 * delayed work -- periodically trim expired leases, renew caps with mds
2952 static void schedule_delayed(struct ceph_mds_client
*mdsc
)
2955 unsigned hz
= round_jiffies_relative(HZ
* delay
);
2956 schedule_delayed_work(&mdsc
->delayed_work
, hz
);
2959 static void delayed_work(struct work_struct
*work
)
2962 struct ceph_mds_client
*mdsc
=
2963 container_of(work
, struct ceph_mds_client
, delayed_work
.work
);
2967 dout("mdsc delayed_work\n");
2968 ceph_check_delayed_caps(mdsc
);
2970 mutex_lock(&mdsc
->mutex
);
2971 renew_interval
= mdsc
->mdsmap
->m_session_timeout
>> 2;
2972 renew_caps
= time_after_eq(jiffies
, HZ
*renew_interval
+
2973 mdsc
->last_renew_caps
);
2975 mdsc
->last_renew_caps
= jiffies
;
2977 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
2978 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
2981 if (s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
2982 dout("resending session close request for mds%d\n",
2984 request_close_session(mdsc
, s
);
2985 ceph_put_mds_session(s
);
2988 if (s
->s_ttl
&& time_after(jiffies
, s
->s_ttl
)) {
2989 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
) {
2990 s
->s_state
= CEPH_MDS_SESSION_HUNG
;
2991 pr_info("mds%d hung\n", s
->s_mds
);
2994 if (s
->s_state
< CEPH_MDS_SESSION_OPEN
) {
2995 /* this mds is failed or recovering, just wait */
2996 ceph_put_mds_session(s
);
2999 mutex_unlock(&mdsc
->mutex
);
3001 mutex_lock(&s
->s_mutex
);
3003 send_renew_caps(mdsc
, s
);
3005 ceph_con_keepalive(&s
->s_con
);
3006 ceph_add_cap_releases(mdsc
, s
);
3007 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3008 s
->s_state
== CEPH_MDS_SESSION_HUNG
)
3009 ceph_send_cap_releases(mdsc
, s
);
3010 mutex_unlock(&s
->s_mutex
);
3011 ceph_put_mds_session(s
);
3013 mutex_lock(&mdsc
->mutex
);
3015 mutex_unlock(&mdsc
->mutex
);
3017 schedule_delayed(mdsc
);
3020 int ceph_mdsc_init(struct ceph_fs_client
*fsc
)
3023 struct ceph_mds_client
*mdsc
;
3025 mdsc
= kzalloc(sizeof(struct ceph_mds_client
), GFP_NOFS
);
3030 mutex_init(&mdsc
->mutex
);
3031 mdsc
->mdsmap
= kzalloc(sizeof(*mdsc
->mdsmap
), GFP_NOFS
);
3032 if (mdsc
->mdsmap
== NULL
)
3035 init_completion(&mdsc
->safe_umount_waiters
);
3036 init_waitqueue_head(&mdsc
->session_close_wq
);
3037 INIT_LIST_HEAD(&mdsc
->waiting_for_map
);
3038 mdsc
->sessions
= NULL
;
3039 mdsc
->max_sessions
= 0;
3041 init_rwsem(&mdsc
->snap_rwsem
);
3042 mdsc
->snap_realms
= RB_ROOT
;
3043 INIT_LIST_HEAD(&mdsc
->snap_empty
);
3044 spin_lock_init(&mdsc
->snap_empty_lock
);
3046 mdsc
->request_tree
= RB_ROOT
;
3047 INIT_DELAYED_WORK(&mdsc
->delayed_work
, delayed_work
);
3048 mdsc
->last_renew_caps
= jiffies
;
3049 INIT_LIST_HEAD(&mdsc
->cap_delay_list
);
3050 spin_lock_init(&mdsc
->cap_delay_lock
);
3051 INIT_LIST_HEAD(&mdsc
->snap_flush_list
);
3052 spin_lock_init(&mdsc
->snap_flush_lock
);
3053 mdsc
->cap_flush_seq
= 0;
3054 INIT_LIST_HEAD(&mdsc
->cap_dirty
);
3055 INIT_LIST_HEAD(&mdsc
->cap_dirty_migrating
);
3056 mdsc
->num_cap_flushing
= 0;
3057 spin_lock_init(&mdsc
->cap_dirty_lock
);
3058 init_waitqueue_head(&mdsc
->cap_flushing_wq
);
3059 spin_lock_init(&mdsc
->dentry_lru_lock
);
3060 INIT_LIST_HEAD(&mdsc
->dentry_lru
);
3062 ceph_caps_init(mdsc
);
3063 ceph_adjust_min_caps(mdsc
, fsc
->min_caps
);
3069 * Wait for safe replies on open mds requests. If we time out, drop
3070 * all requests from the tree to avoid dangling dentry refs.
3072 static void wait_requests(struct ceph_mds_client
*mdsc
)
3074 struct ceph_mds_request
*req
;
3075 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3077 mutex_lock(&mdsc
->mutex
);
3078 if (__get_oldest_req(mdsc
)) {
3079 mutex_unlock(&mdsc
->mutex
);
3081 dout("wait_requests waiting for requests\n");
3082 wait_for_completion_timeout(&mdsc
->safe_umount_waiters
,
3083 fsc
->client
->options
->mount_timeout
* HZ
);
3085 /* tear down remaining requests */
3086 mutex_lock(&mdsc
->mutex
);
3087 while ((req
= __get_oldest_req(mdsc
))) {
3088 dout("wait_requests timed out on tid %llu\n",
3090 __unregister_request(mdsc
, req
);
3093 mutex_unlock(&mdsc
->mutex
);
3094 dout("wait_requests done\n");
3098 * called before mount is ro, and before dentries are torn down.
3099 * (hmm, does this still race with new lookups?)
3101 void ceph_mdsc_pre_umount(struct ceph_mds_client
*mdsc
)
3103 dout("pre_umount\n");
3107 ceph_flush_dirty_caps(mdsc
);
3108 wait_requests(mdsc
);
3111 * wait for reply handlers to drop their request refs and
3112 * their inode/dcache refs
3118 * wait for all write mds requests to flush.
3120 static void wait_unsafe_requests(struct ceph_mds_client
*mdsc
, u64 want_tid
)
3122 struct ceph_mds_request
*req
= NULL
, *nextreq
;
3125 mutex_lock(&mdsc
->mutex
);
3126 dout("wait_unsafe_requests want %lld\n", want_tid
);
3128 req
= __get_oldest_req(mdsc
);
3129 while (req
&& req
->r_tid
<= want_tid
) {
3130 /* find next request */
3131 n
= rb_next(&req
->r_node
);
3133 nextreq
= rb_entry(n
, struct ceph_mds_request
, r_node
);
3136 if ((req
->r_op
& CEPH_MDS_OP_WRITE
)) {
3138 ceph_mdsc_get_request(req
);
3140 ceph_mdsc_get_request(nextreq
);
3141 mutex_unlock(&mdsc
->mutex
);
3142 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3143 req
->r_tid
, want_tid
);
3144 wait_for_completion(&req
->r_safe_completion
);
3145 mutex_lock(&mdsc
->mutex
);
3146 ceph_mdsc_put_request(req
);
3148 break; /* next dne before, so we're done! */
3149 if (RB_EMPTY_NODE(&nextreq
->r_node
)) {
3150 /* next request was removed from tree */
3151 ceph_mdsc_put_request(nextreq
);
3154 ceph_mdsc_put_request(nextreq
); /* won't go away */
3158 mutex_unlock(&mdsc
->mutex
);
3159 dout("wait_unsafe_requests done\n");
3162 void ceph_mdsc_sync(struct ceph_mds_client
*mdsc
)
3164 u64 want_tid
, want_flush
;
3166 if (mdsc
->fsc
->mount_state
== CEPH_MOUNT_SHUTDOWN
)
3170 mutex_lock(&mdsc
->mutex
);
3171 want_tid
= mdsc
->last_tid
;
3172 want_flush
= mdsc
->cap_flush_seq
;
3173 mutex_unlock(&mdsc
->mutex
);
3174 dout("sync want tid %lld flush_seq %lld\n", want_tid
, want_flush
);
3176 ceph_flush_dirty_caps(mdsc
);
3178 wait_unsafe_requests(mdsc
, want_tid
);
3179 wait_event(mdsc
->cap_flushing_wq
, check_cap_flush(mdsc
, want_flush
));
3183 * true if all sessions are closed, or we force unmount
3185 static bool done_closing_sessions(struct ceph_mds_client
*mdsc
)
3189 if (mdsc
->fsc
->mount_state
== CEPH_MOUNT_SHUTDOWN
)
3192 mutex_lock(&mdsc
->mutex
);
3193 for (i
= 0; i
< mdsc
->max_sessions
; i
++)
3194 if (mdsc
->sessions
[i
])
3196 mutex_unlock(&mdsc
->mutex
);
3201 * called after sb is ro.
3203 void ceph_mdsc_close_sessions(struct ceph_mds_client
*mdsc
)
3205 struct ceph_mds_session
*session
;
3207 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3208 unsigned long timeout
= fsc
->client
->options
->mount_timeout
* HZ
;
3210 dout("close_sessions\n");
3212 /* close sessions */
3213 mutex_lock(&mdsc
->mutex
);
3214 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3215 session
= __ceph_lookup_mds_session(mdsc
, i
);
3218 mutex_unlock(&mdsc
->mutex
);
3219 mutex_lock(&session
->s_mutex
);
3220 __close_session(mdsc
, session
);
3221 mutex_unlock(&session
->s_mutex
);
3222 ceph_put_mds_session(session
);
3223 mutex_lock(&mdsc
->mutex
);
3225 mutex_unlock(&mdsc
->mutex
);
3227 dout("waiting for sessions to close\n");
3228 wait_event_timeout(mdsc
->session_close_wq
, done_closing_sessions(mdsc
),
3231 /* tear down remaining sessions */
3232 mutex_lock(&mdsc
->mutex
);
3233 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3234 if (mdsc
->sessions
[i
]) {
3235 session
= get_session(mdsc
->sessions
[i
]);
3236 __unregister_session(mdsc
, session
);
3237 mutex_unlock(&mdsc
->mutex
);
3238 mutex_lock(&session
->s_mutex
);
3239 remove_session_caps(session
);
3240 mutex_unlock(&session
->s_mutex
);
3241 ceph_put_mds_session(session
);
3242 mutex_lock(&mdsc
->mutex
);
3245 WARN_ON(!list_empty(&mdsc
->cap_delay_list
));
3246 mutex_unlock(&mdsc
->mutex
);
3248 ceph_cleanup_empty_realms(mdsc
);
3250 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3255 static void ceph_mdsc_stop(struct ceph_mds_client
*mdsc
)
3258 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3260 ceph_mdsmap_destroy(mdsc
->mdsmap
);
3261 kfree(mdsc
->sessions
);
3262 ceph_caps_finalize(mdsc
);
3265 void ceph_mdsc_destroy(struct ceph_fs_client
*fsc
)
3267 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
3269 dout("mdsc_destroy %p\n", mdsc
);
3270 ceph_mdsc_stop(mdsc
);
3272 /* flush out any connection work with references to us */
3277 dout("mdsc_destroy %p done\n", mdsc
);
3282 * handle mds map update.
3284 void ceph_mdsc_handle_map(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3288 void *p
= msg
->front
.iov_base
;
3289 void *end
= p
+ msg
->front
.iov_len
;
3290 struct ceph_mdsmap
*newmap
, *oldmap
;
3291 struct ceph_fsid fsid
;
3294 ceph_decode_need(&p
, end
, sizeof(fsid
)+2*sizeof(u32
), bad
);
3295 ceph_decode_copy(&p
, &fsid
, sizeof(fsid
));
3296 if (ceph_check_fsid(mdsc
->fsc
->client
, &fsid
) < 0)
3298 epoch
= ceph_decode_32(&p
);
3299 maplen
= ceph_decode_32(&p
);
3300 dout("handle_map epoch %u len %d\n", epoch
, (int)maplen
);
3302 /* do we need it? */
3303 ceph_monc_got_mdsmap(&mdsc
->fsc
->client
->monc
, epoch
);
3304 mutex_lock(&mdsc
->mutex
);
3305 if (mdsc
->mdsmap
&& epoch
<= mdsc
->mdsmap
->m_epoch
) {
3306 dout("handle_map epoch %u <= our %u\n",
3307 epoch
, mdsc
->mdsmap
->m_epoch
);
3308 mutex_unlock(&mdsc
->mutex
);
3312 newmap
= ceph_mdsmap_decode(&p
, end
);
3313 if (IS_ERR(newmap
)) {
3314 err
= PTR_ERR(newmap
);
3318 /* swap into place */
3320 oldmap
= mdsc
->mdsmap
;
3321 mdsc
->mdsmap
= newmap
;
3322 check_new_map(mdsc
, newmap
, oldmap
);
3323 ceph_mdsmap_destroy(oldmap
);
3325 mdsc
->mdsmap
= newmap
; /* first mds map */
3327 mdsc
->fsc
->sb
->s_maxbytes
= mdsc
->mdsmap
->m_max_file_size
;
3329 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3331 mutex_unlock(&mdsc
->mutex
);
3332 schedule_delayed(mdsc
);
3336 mutex_unlock(&mdsc
->mutex
);
3338 pr_err("error decoding mdsmap %d\n", err
);
3342 static struct ceph_connection
*con_get(struct ceph_connection
*con
)
3344 struct ceph_mds_session
*s
= con
->private;
3346 if (get_session(s
)) {
3347 dout("mdsc con_get %p ok (%d)\n", s
, atomic_read(&s
->s_ref
));
3350 dout("mdsc con_get %p FAIL\n", s
);
3354 static void con_put(struct ceph_connection
*con
)
3356 struct ceph_mds_session
*s
= con
->private;
3358 dout("mdsc con_put %p (%d)\n", s
, atomic_read(&s
->s_ref
) - 1);
3359 ceph_put_mds_session(s
);
3363 * if the client is unresponsive for long enough, the mds will kill
3364 * the session entirely.
3366 static void peer_reset(struct ceph_connection
*con
)
3368 struct ceph_mds_session
*s
= con
->private;
3369 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3371 pr_warning("mds%d closed our session\n", s
->s_mds
);
3372 send_mds_reconnect(mdsc
, s
);
3375 static void dispatch(struct ceph_connection
*con
, struct ceph_msg
*msg
)
3377 struct ceph_mds_session
*s
= con
->private;
3378 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3379 int type
= le16_to_cpu(msg
->hdr
.type
);
3381 mutex_lock(&mdsc
->mutex
);
3382 if (__verify_registered_session(mdsc
, s
) < 0) {
3383 mutex_unlock(&mdsc
->mutex
);
3386 mutex_unlock(&mdsc
->mutex
);
3389 case CEPH_MSG_MDS_MAP
:
3390 ceph_mdsc_handle_map(mdsc
, msg
);
3392 case CEPH_MSG_CLIENT_SESSION
:
3393 handle_session(s
, msg
);
3395 case CEPH_MSG_CLIENT_REPLY
:
3396 handle_reply(s
, msg
);
3398 case CEPH_MSG_CLIENT_REQUEST_FORWARD
:
3399 handle_forward(mdsc
, s
, msg
);
3401 case CEPH_MSG_CLIENT_CAPS
:
3402 ceph_handle_caps(s
, msg
);
3404 case CEPH_MSG_CLIENT_SNAP
:
3405 ceph_handle_snap(mdsc
, s
, msg
);
3407 case CEPH_MSG_CLIENT_LEASE
:
3408 handle_lease(mdsc
, s
, msg
);
3412 pr_err("received unknown message type %d %s\n", type
,
3413 ceph_msg_type_name(type
));
3424 * Note: returned pointer is the address of a structure that's
3425 * managed separately. Caller must *not* attempt to free it.
3427 static struct ceph_auth_handshake
*get_authorizer(struct ceph_connection
*con
,
3428 int *proto
, int force_new
)
3430 struct ceph_mds_session
*s
= con
->private;
3431 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3432 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3433 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
3435 if (force_new
&& auth
->authorizer
) {
3436 if (ac
->ops
&& ac
->ops
->destroy_authorizer
)
3437 ac
->ops
->destroy_authorizer(ac
, auth
->authorizer
);
3438 auth
->authorizer
= NULL
;
3440 if (!auth
->authorizer
&& ac
->ops
&& ac
->ops
->create_authorizer
) {
3441 int ret
= ac
->ops
->create_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3444 return ERR_PTR(ret
);
3446 *proto
= ac
->protocol
;
3452 static int verify_authorizer_reply(struct ceph_connection
*con
, int len
)
3454 struct ceph_mds_session
*s
= con
->private;
3455 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3456 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3458 return ac
->ops
->verify_authorizer_reply(ac
, s
->s_auth
.authorizer
, len
);
3461 static int invalidate_authorizer(struct ceph_connection
*con
)
3463 struct ceph_mds_session
*s
= con
->private;
3464 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3465 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3467 if (ac
->ops
->invalidate_authorizer
)
3468 ac
->ops
->invalidate_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
);
3470 return ceph_monc_validate_auth(&mdsc
->fsc
->client
->monc
);
3473 static const struct ceph_connection_operations mds_con_ops
= {
3476 .dispatch
= dispatch
,
3477 .get_authorizer
= get_authorizer
,
3478 .verify_authorizer_reply
= verify_authorizer_reply
,
3479 .invalidate_authorizer
= invalidate_authorizer
,
3480 .peer_reset
= peer_reset
,