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42 * are no longer valid.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 s
->s_mdsc
->fsc
->client
->monc
.auth
->ops
->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 s
= kzalloc(sizeof(*s
), GFP_NOFS
);
419 return ERR_PTR(-ENOMEM
);
422 s
->s_state
= CEPH_MDS_SESSION_NEW
;
425 mutex_init(&s
->s_mutex
);
427 ceph_con_init(&s
->s_con
, s
, &mds_con_ops
, &mdsc
->fsc
->client
->msgr
);
429 spin_lock_init(&s
->s_gen_ttl_lock
);
431 s
->s_cap_ttl
= jiffies
- 1;
433 spin_lock_init(&s
->s_cap_lock
);
434 s
->s_renew_requested
= 0;
436 INIT_LIST_HEAD(&s
->s_caps
);
439 atomic_set(&s
->s_ref
, 1);
440 INIT_LIST_HEAD(&s
->s_waiting
);
441 INIT_LIST_HEAD(&s
->s_unsafe
);
442 s
->s_num_cap_releases
= 0;
443 s
->s_cap_iterator
= NULL
;
444 INIT_LIST_HEAD(&s
->s_cap_releases
);
445 INIT_LIST_HEAD(&s
->s_cap_releases_done
);
446 INIT_LIST_HEAD(&s
->s_cap_flushing
);
447 INIT_LIST_HEAD(&s
->s_cap_snaps_flushing
);
449 dout("register_session mds%d\n", mds
);
450 if (mds
>= mdsc
->max_sessions
) {
451 int newmax
= 1 << get_count_order(mds
+1);
452 struct ceph_mds_session
**sa
;
454 dout("register_session realloc to %d\n", newmax
);
455 sa
= kcalloc(newmax
, sizeof(void *), GFP_NOFS
);
458 if (mdsc
->sessions
) {
459 memcpy(sa
, mdsc
->sessions
,
460 mdsc
->max_sessions
* sizeof(void *));
461 kfree(mdsc
->sessions
);
464 mdsc
->max_sessions
= newmax
;
466 mdsc
->sessions
[mds
] = s
;
467 atomic_inc(&s
->s_ref
); /* one ref to sessions[], one to caller */
469 ceph_con_open(&s
->s_con
, CEPH_ENTITY_TYPE_MDS
, mds
,
470 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
476 return ERR_PTR(-ENOMEM
);
480 * called under mdsc->mutex
482 static void __unregister_session(struct ceph_mds_client
*mdsc
,
483 struct ceph_mds_session
*s
)
485 dout("__unregister_session mds%d %p\n", s
->s_mds
, s
);
486 BUG_ON(mdsc
->sessions
[s
->s_mds
] != s
);
487 mdsc
->sessions
[s
->s_mds
] = NULL
;
488 ceph_con_close(&s
->s_con
);
489 ceph_put_mds_session(s
);
493 * drop session refs in request.
495 * should be last request ref, or hold mdsc->mutex
497 static void put_request_session(struct ceph_mds_request
*req
)
499 if (req
->r_session
) {
500 ceph_put_mds_session(req
->r_session
);
501 req
->r_session
= NULL
;
505 void ceph_mdsc_release_request(struct kref
*kref
)
507 struct ceph_mds_request
*req
= container_of(kref
,
508 struct ceph_mds_request
,
511 ceph_msg_put(req
->r_request
);
513 ceph_msg_put(req
->r_reply
);
514 destroy_reply_info(&req
->r_reply_info
);
517 ceph_put_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
520 if (req
->r_locked_dir
)
521 ceph_put_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
522 if (req
->r_target_inode
)
523 iput(req
->r_target_inode
);
526 if (req
->r_old_dentry
) {
528 * track (and drop pins for) r_old_dentry_dir
529 * separately, since r_old_dentry's d_parent may have
530 * changed between the dir mutex being dropped and
531 * this request being freed.
533 ceph_put_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
535 dput(req
->r_old_dentry
);
536 iput(req
->r_old_dentry_dir
);
540 put_request_session(req
);
541 ceph_unreserve_caps(req
->r_mdsc
, &req
->r_caps_reservation
);
546 * lookup session, bump ref if found.
548 * called under mdsc->mutex.
550 static struct ceph_mds_request
*__lookup_request(struct ceph_mds_client
*mdsc
,
553 struct ceph_mds_request
*req
;
554 struct rb_node
*n
= mdsc
->request_tree
.rb_node
;
557 req
= rb_entry(n
, struct ceph_mds_request
, r_node
);
558 if (tid
< req
->r_tid
)
560 else if (tid
> req
->r_tid
)
563 ceph_mdsc_get_request(req
);
570 static void __insert_request(struct ceph_mds_client
*mdsc
,
571 struct ceph_mds_request
*new)
573 struct rb_node
**p
= &mdsc
->request_tree
.rb_node
;
574 struct rb_node
*parent
= NULL
;
575 struct ceph_mds_request
*req
= NULL
;
579 req
= rb_entry(parent
, struct ceph_mds_request
, r_node
);
580 if (new->r_tid
< req
->r_tid
)
582 else if (new->r_tid
> req
->r_tid
)
588 rb_link_node(&new->r_node
, parent
, p
);
589 rb_insert_color(&new->r_node
, &mdsc
->request_tree
);
593 * Register an in-flight request, and assign a tid. Link to directory
594 * are modifying (if any).
596 * Called under mdsc->mutex.
598 static void __register_request(struct ceph_mds_client
*mdsc
,
599 struct ceph_mds_request
*req
,
602 req
->r_tid
= ++mdsc
->last_tid
;
604 ceph_reserve_caps(mdsc
, &req
->r_caps_reservation
,
606 dout("__register_request %p tid %lld\n", req
, req
->r_tid
);
607 ceph_mdsc_get_request(req
);
608 __insert_request(mdsc
, req
);
610 req
->r_uid
= current_fsuid();
611 req
->r_gid
= current_fsgid();
614 struct ceph_inode_info
*ci
= ceph_inode(dir
);
617 spin_lock(&ci
->i_unsafe_lock
);
618 req
->r_unsafe_dir
= dir
;
619 list_add_tail(&req
->r_unsafe_dir_item
, &ci
->i_unsafe_dirops
);
620 spin_unlock(&ci
->i_unsafe_lock
);
624 static void __unregister_request(struct ceph_mds_client
*mdsc
,
625 struct ceph_mds_request
*req
)
627 dout("__unregister_request %p tid %lld\n", req
, req
->r_tid
);
628 rb_erase(&req
->r_node
, &mdsc
->request_tree
);
629 RB_CLEAR_NODE(&req
->r_node
);
631 if (req
->r_unsafe_dir
) {
632 struct ceph_inode_info
*ci
= ceph_inode(req
->r_unsafe_dir
);
634 spin_lock(&ci
->i_unsafe_lock
);
635 list_del_init(&req
->r_unsafe_dir_item
);
636 spin_unlock(&ci
->i_unsafe_lock
);
638 iput(req
->r_unsafe_dir
);
639 req
->r_unsafe_dir
= NULL
;
642 ceph_mdsc_put_request(req
);
646 * Choose mds to send request to next. If there is a hint set in the
647 * request (e.g., due to a prior forward hint from the mds), use that.
648 * Otherwise, consult frag tree and/or caps to identify the
649 * appropriate mds. If all else fails, choose randomly.
651 * Called under mdsc->mutex.
653 static struct dentry
*get_nonsnap_parent(struct dentry
*dentry
)
656 * we don't need to worry about protecting the d_parent access
657 * here because we never renaming inside the snapped namespace
658 * except to resplice to another snapdir, and either the old or new
659 * result is a valid result.
661 while (!IS_ROOT(dentry
) && ceph_snap(dentry
->d_inode
) != CEPH_NOSNAP
)
662 dentry
= dentry
->d_parent
;
666 static int __choose_mds(struct ceph_mds_client
*mdsc
,
667 struct ceph_mds_request
*req
)
670 struct ceph_inode_info
*ci
;
671 struct ceph_cap
*cap
;
672 int mode
= req
->r_direct_mode
;
674 u32 hash
= req
->r_direct_hash
;
675 bool is_hash
= req
->r_direct_is_hash
;
678 * is there a specific mds we should try? ignore hint if we have
679 * no session and the mds is not up (active or recovering).
681 if (req
->r_resend_mds
>= 0 &&
682 (__have_session(mdsc
, req
->r_resend_mds
) ||
683 ceph_mdsmap_get_state(mdsc
->mdsmap
, req
->r_resend_mds
) > 0)) {
684 dout("choose_mds using resend_mds mds%d\n",
686 return req
->r_resend_mds
;
689 if (mode
== USE_RANDOM_MDS
)
694 inode
= req
->r_inode
;
695 } else if (req
->r_dentry
) {
696 /* ignore race with rename; old or new d_parent is okay */
697 struct dentry
*parent
= req
->r_dentry
->d_parent
;
698 struct inode
*dir
= parent
->d_inode
;
700 if (dir
->i_sb
!= mdsc
->fsc
->sb
) {
702 inode
= req
->r_dentry
->d_inode
;
703 } else if (ceph_snap(dir
) != CEPH_NOSNAP
) {
704 /* direct snapped/virtual snapdir requests
705 * based on parent dir inode */
706 struct dentry
*dn
= get_nonsnap_parent(parent
);
708 dout("__choose_mds using nonsnap parent %p\n", inode
);
709 } else if (req
->r_dentry
->d_inode
) {
711 inode
= req
->r_dentry
->d_inode
;
715 hash
= ceph_dentry_hash(dir
, req
->r_dentry
);
720 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode
, (int)is_hash
,
724 ci
= ceph_inode(inode
);
726 if (is_hash
&& S_ISDIR(inode
->i_mode
)) {
727 struct ceph_inode_frag frag
;
730 ceph_choose_frag(ci
, hash
, &frag
, &found
);
732 if (mode
== USE_ANY_MDS
&& frag
.ndist
> 0) {
735 /* choose a random replica */
736 get_random_bytes(&r
, 1);
739 dout("choose_mds %p %llx.%llx "
740 "frag %u mds%d (%d/%d)\n",
741 inode
, ceph_vinop(inode
),
744 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
745 CEPH_MDS_STATE_ACTIVE
)
749 /* since this file/dir wasn't known to be
750 * replicated, then we want to look for the
751 * authoritative mds. */
754 /* choose auth mds */
756 dout("choose_mds %p %llx.%llx "
757 "frag %u mds%d (auth)\n",
758 inode
, ceph_vinop(inode
), frag
.frag
, mds
);
759 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
760 CEPH_MDS_STATE_ACTIVE
)
766 spin_lock(&ci
->i_ceph_lock
);
768 if (mode
== USE_AUTH_MDS
)
769 cap
= ci
->i_auth_cap
;
770 if (!cap
&& !RB_EMPTY_ROOT(&ci
->i_caps
))
771 cap
= rb_entry(rb_first(&ci
->i_caps
), struct ceph_cap
, ci_node
);
773 spin_unlock(&ci
->i_ceph_lock
);
776 mds
= cap
->session
->s_mds
;
777 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
778 inode
, ceph_vinop(inode
), mds
,
779 cap
== ci
->i_auth_cap
? "auth " : "", cap
);
780 spin_unlock(&ci
->i_ceph_lock
);
784 mds
= ceph_mdsmap_get_random_mds(mdsc
->mdsmap
);
785 dout("choose_mds chose random mds%d\n", mds
);
793 static struct ceph_msg
*create_session_msg(u32 op
, u64 seq
)
795 struct ceph_msg
*msg
;
796 struct ceph_mds_session_head
*h
;
798 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
), GFP_NOFS
,
801 pr_err("create_session_msg ENOMEM creating msg\n");
804 h
= msg
->front
.iov_base
;
805 h
->op
= cpu_to_le32(op
);
806 h
->seq
= cpu_to_le64(seq
);
811 * send session open request.
813 * called under mdsc->mutex
815 static int __open_session(struct ceph_mds_client
*mdsc
,
816 struct ceph_mds_session
*session
)
818 struct ceph_msg
*msg
;
820 int mds
= session
->s_mds
;
822 /* wait for mds to go active? */
823 mstate
= ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
);
824 dout("open_session to mds%d (%s)\n", mds
,
825 ceph_mds_state_name(mstate
));
826 session
->s_state
= CEPH_MDS_SESSION_OPENING
;
827 session
->s_renew_requested
= jiffies
;
829 /* send connect message */
830 msg
= create_session_msg(CEPH_SESSION_REQUEST_OPEN
, session
->s_seq
);
833 ceph_con_send(&session
->s_con
, msg
);
838 * open sessions for any export targets for the given mds
840 * called under mdsc->mutex
842 static void __open_export_target_sessions(struct ceph_mds_client
*mdsc
,
843 struct ceph_mds_session
*session
)
845 struct ceph_mds_info
*mi
;
846 struct ceph_mds_session
*ts
;
847 int i
, mds
= session
->s_mds
;
850 if (mds
>= mdsc
->mdsmap
->m_max_mds
)
852 mi
= &mdsc
->mdsmap
->m_info
[mds
];
853 dout("open_export_target_sessions for mds%d (%d targets)\n",
854 session
->s_mds
, mi
->num_export_targets
);
856 for (i
= 0; i
< mi
->num_export_targets
; i
++) {
857 target
= mi
->export_targets
[i
];
858 ts
= __ceph_lookup_mds_session(mdsc
, target
);
860 ts
= register_session(mdsc
, target
);
864 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
865 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
866 __open_session(mdsc
, session
);
868 dout(" mds%d target mds%d %p is %s\n", session
->s_mds
,
869 i
, ts
, session_state_name(ts
->s_state
));
870 ceph_put_mds_session(ts
);
874 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client
*mdsc
,
875 struct ceph_mds_session
*session
)
877 mutex_lock(&mdsc
->mutex
);
878 __open_export_target_sessions(mdsc
, session
);
879 mutex_unlock(&mdsc
->mutex
);
887 * Free preallocated cap messages assigned to this session
889 static void cleanup_cap_releases(struct ceph_mds_session
*session
)
891 struct ceph_msg
*msg
;
893 spin_lock(&session
->s_cap_lock
);
894 while (!list_empty(&session
->s_cap_releases
)) {
895 msg
= list_first_entry(&session
->s_cap_releases
,
896 struct ceph_msg
, list_head
);
897 list_del_init(&msg
->list_head
);
900 while (!list_empty(&session
->s_cap_releases_done
)) {
901 msg
= list_first_entry(&session
->s_cap_releases_done
,
902 struct ceph_msg
, list_head
);
903 list_del_init(&msg
->list_head
);
906 spin_unlock(&session
->s_cap_lock
);
910 * Helper to safely iterate over all caps associated with a session, with
911 * special care taken to handle a racing __ceph_remove_cap().
913 * Caller must hold session s_mutex.
915 static int iterate_session_caps(struct ceph_mds_session
*session
,
916 int (*cb
)(struct inode
*, struct ceph_cap
*,
920 struct ceph_cap
*cap
;
921 struct inode
*inode
, *last_inode
= NULL
;
922 struct ceph_cap
*old_cap
= NULL
;
925 dout("iterate_session_caps %p mds%d\n", session
, session
->s_mds
);
926 spin_lock(&session
->s_cap_lock
);
927 p
= session
->s_caps
.next
;
928 while (p
!= &session
->s_caps
) {
929 cap
= list_entry(p
, struct ceph_cap
, session_caps
);
930 inode
= igrab(&cap
->ci
->vfs_inode
);
935 session
->s_cap_iterator
= cap
;
936 spin_unlock(&session
->s_cap_lock
);
943 ceph_put_cap(session
->s_mdsc
, old_cap
);
947 ret
= cb(inode
, cap
, arg
);
950 spin_lock(&session
->s_cap_lock
);
952 if (cap
->ci
== NULL
) {
953 dout("iterate_session_caps finishing cap %p removal\n",
955 BUG_ON(cap
->session
!= session
);
956 list_del_init(&cap
->session_caps
);
957 session
->s_nr_caps
--;
959 old_cap
= cap
; /* put_cap it w/o locks held */
966 session
->s_cap_iterator
= NULL
;
967 spin_unlock(&session
->s_cap_lock
);
972 ceph_put_cap(session
->s_mdsc
, old_cap
);
977 static int remove_session_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
980 struct ceph_inode_info
*ci
= ceph_inode(inode
);
983 dout("removing cap %p, ci is %p, inode is %p\n",
984 cap
, ci
, &ci
->vfs_inode
);
985 spin_lock(&ci
->i_ceph_lock
);
986 __ceph_remove_cap(cap
);
987 if (!__ceph_is_any_real_caps(ci
)) {
988 struct ceph_mds_client
*mdsc
=
989 ceph_sb_to_client(inode
->i_sb
)->mdsc
;
991 spin_lock(&mdsc
->cap_dirty_lock
);
992 if (!list_empty(&ci
->i_dirty_item
)) {
993 pr_info(" dropping dirty %s state for %p %lld\n",
994 ceph_cap_string(ci
->i_dirty_caps
),
995 inode
, ceph_ino(inode
));
996 ci
->i_dirty_caps
= 0;
997 list_del_init(&ci
->i_dirty_item
);
1000 if (!list_empty(&ci
->i_flushing_item
)) {
1001 pr_info(" dropping dirty+flushing %s state for %p %lld\n",
1002 ceph_cap_string(ci
->i_flushing_caps
),
1003 inode
, ceph_ino(inode
));
1004 ci
->i_flushing_caps
= 0;
1005 list_del_init(&ci
->i_flushing_item
);
1006 mdsc
->num_cap_flushing
--;
1009 if (drop
&& ci
->i_wrbuffer_ref
) {
1010 pr_info(" dropping dirty data for %p %lld\n",
1011 inode
, ceph_ino(inode
));
1012 ci
->i_wrbuffer_ref
= 0;
1013 ci
->i_wrbuffer_ref_head
= 0;
1016 spin_unlock(&mdsc
->cap_dirty_lock
);
1018 spin_unlock(&ci
->i_ceph_lock
);
1025 * caller must hold session s_mutex
1027 static void remove_session_caps(struct ceph_mds_session
*session
)
1029 dout("remove_session_caps on %p\n", session
);
1030 iterate_session_caps(session
, remove_session_caps_cb
, NULL
);
1031 BUG_ON(session
->s_nr_caps
> 0);
1032 BUG_ON(!list_empty(&session
->s_cap_flushing
));
1033 cleanup_cap_releases(session
);
1037 * wake up any threads waiting on this session's caps. if the cap is
1038 * old (didn't get renewed on the client reconnect), remove it now.
1040 * caller must hold s_mutex.
1042 static int wake_up_session_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1045 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1047 wake_up_all(&ci
->i_cap_wq
);
1049 spin_lock(&ci
->i_ceph_lock
);
1050 ci
->i_wanted_max_size
= 0;
1051 ci
->i_requested_max_size
= 0;
1052 spin_unlock(&ci
->i_ceph_lock
);
1057 static void wake_up_session_caps(struct ceph_mds_session
*session
,
1060 dout("wake_up_session_caps %p mds%d\n", session
, session
->s_mds
);
1061 iterate_session_caps(session
, wake_up_session_cb
,
1062 (void *)(unsigned long)reconnect
);
1066 * Send periodic message to MDS renewing all currently held caps. The
1067 * ack will reset the expiration for all caps from this session.
1069 * caller holds s_mutex
1071 static int send_renew_caps(struct ceph_mds_client
*mdsc
,
1072 struct ceph_mds_session
*session
)
1074 struct ceph_msg
*msg
;
1077 if (time_after_eq(jiffies
, session
->s_cap_ttl
) &&
1078 time_after_eq(session
->s_cap_ttl
, session
->s_renew_requested
))
1079 pr_info("mds%d caps stale\n", session
->s_mds
);
1080 session
->s_renew_requested
= jiffies
;
1082 /* do not try to renew caps until a recovering mds has reconnected
1083 * with its clients. */
1084 state
= ceph_mdsmap_get_state(mdsc
->mdsmap
, session
->s_mds
);
1085 if (state
< CEPH_MDS_STATE_RECONNECT
) {
1086 dout("send_renew_caps ignoring mds%d (%s)\n",
1087 session
->s_mds
, ceph_mds_state_name(state
));
1091 dout("send_renew_caps to mds%d (%s)\n", session
->s_mds
,
1092 ceph_mds_state_name(state
));
1093 msg
= create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS
,
1094 ++session
->s_renew_seq
);
1097 ceph_con_send(&session
->s_con
, msg
);
1102 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1104 * Called under session->s_mutex
1106 static void renewed_caps(struct ceph_mds_client
*mdsc
,
1107 struct ceph_mds_session
*session
, int is_renew
)
1112 spin_lock(&session
->s_cap_lock
);
1113 was_stale
= is_renew
&& time_after_eq(jiffies
, session
->s_cap_ttl
);
1115 session
->s_cap_ttl
= session
->s_renew_requested
+
1116 mdsc
->mdsmap
->m_session_timeout
*HZ
;
1119 if (time_before(jiffies
, session
->s_cap_ttl
)) {
1120 pr_info("mds%d caps renewed\n", session
->s_mds
);
1123 pr_info("mds%d caps still stale\n", session
->s_mds
);
1126 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1127 session
->s_mds
, session
->s_cap_ttl
, was_stale
? "stale" : "fresh",
1128 time_before(jiffies
, session
->s_cap_ttl
) ? "stale" : "fresh");
1129 spin_unlock(&session
->s_cap_lock
);
1132 wake_up_session_caps(session
, 0);
1136 * send a session close request
1138 static int request_close_session(struct ceph_mds_client
*mdsc
,
1139 struct ceph_mds_session
*session
)
1141 struct ceph_msg
*msg
;
1143 dout("request_close_session mds%d state %s seq %lld\n",
1144 session
->s_mds
, session_state_name(session
->s_state
),
1146 msg
= create_session_msg(CEPH_SESSION_REQUEST_CLOSE
, session
->s_seq
);
1149 ceph_con_send(&session
->s_con
, msg
);
1154 * Called with s_mutex held.
1156 static int __close_session(struct ceph_mds_client
*mdsc
,
1157 struct ceph_mds_session
*session
)
1159 if (session
->s_state
>= CEPH_MDS_SESSION_CLOSING
)
1161 session
->s_state
= CEPH_MDS_SESSION_CLOSING
;
1162 return request_close_session(mdsc
, session
);
1166 * Trim old(er) caps.
1168 * Because we can't cache an inode without one or more caps, we do
1169 * this indirectly: if a cap is unused, we prune its aliases, at which
1170 * point the inode will hopefully get dropped to.
1172 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1173 * memory pressure from the MDS, though, so it needn't be perfect.
1175 static int trim_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
, void *arg
)
1177 struct ceph_mds_session
*session
= arg
;
1178 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1179 int used
, oissued
, mine
;
1181 if (session
->s_trim_caps
<= 0)
1184 spin_lock(&ci
->i_ceph_lock
);
1185 mine
= cap
->issued
| cap
->implemented
;
1186 used
= __ceph_caps_used(ci
);
1187 oissued
= __ceph_caps_issued_other(ci
, cap
);
1189 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1190 inode
, cap
, ceph_cap_string(mine
), ceph_cap_string(oissued
),
1191 ceph_cap_string(used
));
1192 if (ci
->i_dirty_caps
)
1193 goto out
; /* dirty caps */
1194 if ((used
& ~oissued
) & mine
)
1195 goto out
; /* we need these caps */
1197 session
->s_trim_caps
--;
1199 /* we aren't the only cap.. just remove us */
1200 __queue_cap_release(session
, ceph_ino(inode
), cap
->cap_id
,
1201 cap
->mseq
, cap
->issue_seq
);
1202 __ceph_remove_cap(cap
);
1204 /* try to drop referring dentries */
1205 spin_unlock(&ci
->i_ceph_lock
);
1206 d_prune_aliases(inode
);
1207 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1208 inode
, cap
, atomic_read(&inode
->i_count
));
1213 spin_unlock(&ci
->i_ceph_lock
);
1218 * Trim session cap count down to some max number.
1220 static int trim_caps(struct ceph_mds_client
*mdsc
,
1221 struct ceph_mds_session
*session
,
1224 int trim_caps
= session
->s_nr_caps
- max_caps
;
1226 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1227 session
->s_mds
, session
->s_nr_caps
, max_caps
, trim_caps
);
1228 if (trim_caps
> 0) {
1229 session
->s_trim_caps
= trim_caps
;
1230 iterate_session_caps(session
, trim_caps_cb
, session
);
1231 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1232 session
->s_mds
, session
->s_nr_caps
, max_caps
,
1233 trim_caps
- session
->s_trim_caps
);
1234 session
->s_trim_caps
= 0;
1240 * Allocate cap_release messages. If there is a partially full message
1241 * in the queue, try to allocate enough to cover it's remainder, so that
1242 * we can send it immediately.
1244 * Called under s_mutex.
1246 int ceph_add_cap_releases(struct ceph_mds_client
*mdsc
,
1247 struct ceph_mds_session
*session
)
1249 struct ceph_msg
*msg
, *partial
= NULL
;
1250 struct ceph_mds_cap_release
*head
;
1252 int extra
= mdsc
->fsc
->mount_options
->cap_release_safety
;
1255 dout("add_cap_releases %p mds%d extra %d\n", session
, session
->s_mds
,
1258 spin_lock(&session
->s_cap_lock
);
1260 if (!list_empty(&session
->s_cap_releases
)) {
1261 msg
= list_first_entry(&session
->s_cap_releases
,
1264 head
= msg
->front
.iov_base
;
1265 num
= le32_to_cpu(head
->num
);
1267 dout(" partial %p with (%d/%d)\n", msg
, num
,
1268 (int)CEPH_CAPS_PER_RELEASE
);
1269 extra
+= CEPH_CAPS_PER_RELEASE
- num
;
1273 while (session
->s_num_cap_releases
< session
->s_nr_caps
+ extra
) {
1274 spin_unlock(&session
->s_cap_lock
);
1275 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE
, PAGE_CACHE_SIZE
,
1279 dout("add_cap_releases %p msg %p now %d\n", session
, msg
,
1280 (int)msg
->front
.iov_len
);
1281 head
= msg
->front
.iov_base
;
1282 head
->num
= cpu_to_le32(0);
1283 msg
->front
.iov_len
= sizeof(*head
);
1284 spin_lock(&session
->s_cap_lock
);
1285 list_add(&msg
->list_head
, &session
->s_cap_releases
);
1286 session
->s_num_cap_releases
+= CEPH_CAPS_PER_RELEASE
;
1290 head
= partial
->front
.iov_base
;
1291 num
= le32_to_cpu(head
->num
);
1292 dout(" queueing partial %p with %d/%d\n", partial
, num
,
1293 (int)CEPH_CAPS_PER_RELEASE
);
1294 list_move_tail(&partial
->list_head
,
1295 &session
->s_cap_releases_done
);
1296 session
->s_num_cap_releases
-= CEPH_CAPS_PER_RELEASE
- num
;
1299 spin_unlock(&session
->s_cap_lock
);
1305 * flush all dirty inode data to disk.
1307 * returns true if we've flushed through want_flush_seq
1309 static int check_cap_flush(struct ceph_mds_client
*mdsc
, u64 want_flush_seq
)
1313 dout("check_cap_flush want %lld\n", want_flush_seq
);
1314 mutex_lock(&mdsc
->mutex
);
1315 for (mds
= 0; ret
&& mds
< mdsc
->max_sessions
; mds
++) {
1316 struct ceph_mds_session
*session
= mdsc
->sessions
[mds
];
1320 get_session(session
);
1321 mutex_unlock(&mdsc
->mutex
);
1323 mutex_lock(&session
->s_mutex
);
1324 if (!list_empty(&session
->s_cap_flushing
)) {
1325 struct ceph_inode_info
*ci
=
1326 list_entry(session
->s_cap_flushing
.next
,
1327 struct ceph_inode_info
,
1329 struct inode
*inode
= &ci
->vfs_inode
;
1331 spin_lock(&ci
->i_ceph_lock
);
1332 if (ci
->i_cap_flush_seq
<= want_flush_seq
) {
1333 dout("check_cap_flush still flushing %p "
1334 "seq %lld <= %lld to mds%d\n", inode
,
1335 ci
->i_cap_flush_seq
, want_flush_seq
,
1339 spin_unlock(&ci
->i_ceph_lock
);
1341 mutex_unlock(&session
->s_mutex
);
1342 ceph_put_mds_session(session
);
1346 mutex_lock(&mdsc
->mutex
);
1349 mutex_unlock(&mdsc
->mutex
);
1350 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq
);
1355 * called under s_mutex
1357 void ceph_send_cap_releases(struct ceph_mds_client
*mdsc
,
1358 struct ceph_mds_session
*session
)
1360 struct ceph_msg
*msg
;
1362 dout("send_cap_releases mds%d\n", session
->s_mds
);
1363 spin_lock(&session
->s_cap_lock
);
1364 while (!list_empty(&session
->s_cap_releases_done
)) {
1365 msg
= list_first_entry(&session
->s_cap_releases_done
,
1366 struct ceph_msg
, list_head
);
1367 list_del_init(&msg
->list_head
);
1368 spin_unlock(&session
->s_cap_lock
);
1369 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1370 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1371 ceph_con_send(&session
->s_con
, msg
);
1372 spin_lock(&session
->s_cap_lock
);
1374 spin_unlock(&session
->s_cap_lock
);
1377 static void discard_cap_releases(struct ceph_mds_client
*mdsc
,
1378 struct ceph_mds_session
*session
)
1380 struct ceph_msg
*msg
;
1381 struct ceph_mds_cap_release
*head
;
1384 dout("discard_cap_releases mds%d\n", session
->s_mds
);
1385 spin_lock(&session
->s_cap_lock
);
1387 /* zero out the in-progress message */
1388 msg
= list_first_entry(&session
->s_cap_releases
,
1389 struct ceph_msg
, list_head
);
1390 head
= msg
->front
.iov_base
;
1391 num
= le32_to_cpu(head
->num
);
1392 dout("discard_cap_releases mds%d %p %u\n", session
->s_mds
, msg
, num
);
1393 head
->num
= cpu_to_le32(0);
1394 session
->s_num_cap_releases
+= num
;
1396 /* requeue completed messages */
1397 while (!list_empty(&session
->s_cap_releases_done
)) {
1398 msg
= list_first_entry(&session
->s_cap_releases_done
,
1399 struct ceph_msg
, list_head
);
1400 list_del_init(&msg
->list_head
);
1402 head
= msg
->front
.iov_base
;
1403 num
= le32_to_cpu(head
->num
);
1404 dout("discard_cap_releases mds%d %p %u\n", session
->s_mds
, msg
,
1406 session
->s_num_cap_releases
+= num
;
1407 head
->num
= cpu_to_le32(0);
1408 msg
->front
.iov_len
= sizeof(*head
);
1409 list_add(&msg
->list_head
, &session
->s_cap_releases
);
1412 spin_unlock(&session
->s_cap_lock
);
1420 * Create an mds request.
1422 struct ceph_mds_request
*
1423 ceph_mdsc_create_request(struct ceph_mds_client
*mdsc
, int op
, int mode
)
1425 struct ceph_mds_request
*req
= kzalloc(sizeof(*req
), GFP_NOFS
);
1428 return ERR_PTR(-ENOMEM
);
1430 mutex_init(&req
->r_fill_mutex
);
1432 req
->r_started
= jiffies
;
1433 req
->r_resend_mds
= -1;
1434 INIT_LIST_HEAD(&req
->r_unsafe_dir_item
);
1436 kref_init(&req
->r_kref
);
1437 INIT_LIST_HEAD(&req
->r_wait
);
1438 init_completion(&req
->r_completion
);
1439 init_completion(&req
->r_safe_completion
);
1440 INIT_LIST_HEAD(&req
->r_unsafe_item
);
1443 req
->r_direct_mode
= mode
;
1448 * return oldest (lowest) request, tid in request tree, 0 if none.
1450 * called under mdsc->mutex.
1452 static struct ceph_mds_request
*__get_oldest_req(struct ceph_mds_client
*mdsc
)
1454 if (RB_EMPTY_ROOT(&mdsc
->request_tree
))
1456 return rb_entry(rb_first(&mdsc
->request_tree
),
1457 struct ceph_mds_request
, r_node
);
1460 static u64
__get_oldest_tid(struct ceph_mds_client
*mdsc
)
1462 struct ceph_mds_request
*req
= __get_oldest_req(mdsc
);
1470 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1471 * on build_path_from_dentry in fs/cifs/dir.c.
1473 * If @stop_on_nosnap, generate path relative to the first non-snapped
1476 * Encode hidden .snap dirs as a double /, i.e.
1477 * foo/.snap/bar -> foo//bar
1479 char *ceph_mdsc_build_path(struct dentry
*dentry
, int *plen
, u64
*base
,
1482 struct dentry
*temp
;
1488 return ERR_PTR(-EINVAL
);
1492 seq
= read_seqbegin(&rename_lock
);
1494 for (temp
= dentry
; !IS_ROOT(temp
);) {
1495 struct inode
*inode
= temp
->d_inode
;
1496 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
)
1497 len
++; /* slash only */
1498 else if (stop_on_nosnap
&& inode
&&
1499 ceph_snap(inode
) == CEPH_NOSNAP
)
1502 len
+= 1 + temp
->d_name
.len
;
1503 temp
= temp
->d_parent
;
1507 len
--; /* no leading '/' */
1509 path
= kmalloc(len
+1, GFP_NOFS
);
1511 return ERR_PTR(-ENOMEM
);
1513 path
[pos
] = 0; /* trailing null */
1515 for (temp
= dentry
; !IS_ROOT(temp
) && pos
!= 0; ) {
1516 struct inode
*inode
;
1518 spin_lock(&temp
->d_lock
);
1519 inode
= temp
->d_inode
;
1520 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
) {
1521 dout("build_path path+%d: %p SNAPDIR\n",
1523 } else if (stop_on_nosnap
&& inode
&&
1524 ceph_snap(inode
) == CEPH_NOSNAP
) {
1525 spin_unlock(&temp
->d_lock
);
1528 pos
-= temp
->d_name
.len
;
1530 spin_unlock(&temp
->d_lock
);
1533 strncpy(path
+ pos
, temp
->d_name
.name
,
1536 spin_unlock(&temp
->d_lock
);
1539 temp
= temp
->d_parent
;
1542 if (pos
!= 0 || read_seqretry(&rename_lock
, seq
)) {
1543 pr_err("build_path did not end path lookup where "
1544 "expected, namelen is %d, pos is %d\n", len
, pos
);
1545 /* presumably this is only possible if racing with a
1546 rename of one of the parent directories (we can not
1547 lock the dentries above us to prevent this, but
1548 retrying should be harmless) */
1553 *base
= ceph_ino(temp
->d_inode
);
1555 dout("build_path on %p %d built %llx '%.*s'\n",
1556 dentry
, dentry
->d_count
, *base
, len
, path
);
1560 static int build_dentry_path(struct dentry
*dentry
,
1561 const char **ppath
, int *ppathlen
, u64
*pino
,
1566 if (ceph_snap(dentry
->d_parent
->d_inode
) == CEPH_NOSNAP
) {
1567 *pino
= ceph_ino(dentry
->d_parent
->d_inode
);
1568 *ppath
= dentry
->d_name
.name
;
1569 *ppathlen
= dentry
->d_name
.len
;
1572 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1574 return PTR_ERR(path
);
1580 static int build_inode_path(struct inode
*inode
,
1581 const char **ppath
, int *ppathlen
, u64
*pino
,
1584 struct dentry
*dentry
;
1587 if (ceph_snap(inode
) == CEPH_NOSNAP
) {
1588 *pino
= ceph_ino(inode
);
1592 dentry
= d_find_alias(inode
);
1593 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1596 return PTR_ERR(path
);
1603 * request arguments may be specified via an inode *, a dentry *, or
1604 * an explicit ino+path.
1606 static int set_request_path_attr(struct inode
*rinode
, struct dentry
*rdentry
,
1607 const char *rpath
, u64 rino
,
1608 const char **ppath
, int *pathlen
,
1609 u64
*ino
, int *freepath
)
1614 r
= build_inode_path(rinode
, ppath
, pathlen
, ino
, freepath
);
1615 dout(" inode %p %llx.%llx\n", rinode
, ceph_ino(rinode
),
1617 } else if (rdentry
) {
1618 r
= build_dentry_path(rdentry
, ppath
, pathlen
, ino
, freepath
);
1619 dout(" dentry %p %llx/%.*s\n", rdentry
, *ino
, *pathlen
,
1621 } else if (rpath
|| rino
) {
1624 *pathlen
= rpath
? strlen(rpath
) : 0;
1625 dout(" path %.*s\n", *pathlen
, rpath
);
1632 * called under mdsc->mutex
1634 static struct ceph_msg
*create_request_message(struct ceph_mds_client
*mdsc
,
1635 struct ceph_mds_request
*req
,
1638 struct ceph_msg
*msg
;
1639 struct ceph_mds_request_head
*head
;
1640 const char *path1
= NULL
;
1641 const char *path2
= NULL
;
1642 u64 ino1
= 0, ino2
= 0;
1643 int pathlen1
= 0, pathlen2
= 0;
1644 int freepath1
= 0, freepath2
= 0;
1650 ret
= set_request_path_attr(req
->r_inode
, req
->r_dentry
,
1651 req
->r_path1
, req
->r_ino1
.ino
,
1652 &path1
, &pathlen1
, &ino1
, &freepath1
);
1658 ret
= set_request_path_attr(NULL
, req
->r_old_dentry
,
1659 req
->r_path2
, req
->r_ino2
.ino
,
1660 &path2
, &pathlen2
, &ino2
, &freepath2
);
1666 len
= sizeof(*head
) +
1667 pathlen1
+ pathlen2
+ 2*(1 + sizeof(u32
) + sizeof(u64
));
1669 /* calculate (max) length for cap releases */
1670 len
+= sizeof(struct ceph_mds_request_release
) *
1671 (!!req
->r_inode_drop
+ !!req
->r_dentry_drop
+
1672 !!req
->r_old_inode_drop
+ !!req
->r_old_dentry_drop
);
1673 if (req
->r_dentry_drop
)
1674 len
+= req
->r_dentry
->d_name
.len
;
1675 if (req
->r_old_dentry_drop
)
1676 len
+= req
->r_old_dentry
->d_name
.len
;
1678 msg
= ceph_msg_new(CEPH_MSG_CLIENT_REQUEST
, len
, GFP_NOFS
, false);
1680 msg
= ERR_PTR(-ENOMEM
);
1684 msg
->hdr
.tid
= cpu_to_le64(req
->r_tid
);
1686 head
= msg
->front
.iov_base
;
1687 p
= msg
->front
.iov_base
+ sizeof(*head
);
1688 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1690 head
->mdsmap_epoch
= cpu_to_le32(mdsc
->mdsmap
->m_epoch
);
1691 head
->op
= cpu_to_le32(req
->r_op
);
1692 head
->caller_uid
= cpu_to_le32(from_kuid(&init_user_ns
, req
->r_uid
));
1693 head
->caller_gid
= cpu_to_le32(from_kgid(&init_user_ns
, req
->r_gid
));
1694 head
->args
= req
->r_args
;
1696 ceph_encode_filepath(&p
, end
, ino1
, path1
);
1697 ceph_encode_filepath(&p
, end
, ino2
, path2
);
1699 /* make note of release offset, in case we need to replay */
1700 req
->r_request_release_offset
= p
- msg
->front
.iov_base
;
1704 if (req
->r_inode_drop
)
1705 releases
+= ceph_encode_inode_release(&p
,
1706 req
->r_inode
? req
->r_inode
: req
->r_dentry
->d_inode
,
1707 mds
, req
->r_inode_drop
, req
->r_inode_unless
, 0);
1708 if (req
->r_dentry_drop
)
1709 releases
+= ceph_encode_dentry_release(&p
, req
->r_dentry
,
1710 mds
, req
->r_dentry_drop
, req
->r_dentry_unless
);
1711 if (req
->r_old_dentry_drop
)
1712 releases
+= ceph_encode_dentry_release(&p
, req
->r_old_dentry
,
1713 mds
, req
->r_old_dentry_drop
, req
->r_old_dentry_unless
);
1714 if (req
->r_old_inode_drop
)
1715 releases
+= ceph_encode_inode_release(&p
,
1716 req
->r_old_dentry
->d_inode
,
1717 mds
, req
->r_old_inode_drop
, req
->r_old_inode_unless
, 0);
1718 head
->num_releases
= cpu_to_le16(releases
);
1721 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
1722 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1724 msg
->pages
= req
->r_pages
;
1725 msg
->page_count
= req
->r_num_pages
;
1726 msg
->hdr
.data_len
= cpu_to_le32(req
->r_data_len
);
1727 msg
->hdr
.data_off
= cpu_to_le16(0);
1731 kfree((char *)path2
);
1734 kfree((char *)path1
);
1740 * called under mdsc->mutex if error, under no mutex if
1743 static void complete_request(struct ceph_mds_client
*mdsc
,
1744 struct ceph_mds_request
*req
)
1746 if (req
->r_callback
)
1747 req
->r_callback(mdsc
, req
);
1749 complete_all(&req
->r_completion
);
1753 * called under mdsc->mutex
1755 static int __prepare_send_request(struct ceph_mds_client
*mdsc
,
1756 struct ceph_mds_request
*req
,
1759 struct ceph_mds_request_head
*rhead
;
1760 struct ceph_msg
*msg
;
1765 struct ceph_cap
*cap
=
1766 ceph_get_cap_for_mds(ceph_inode(req
->r_inode
), mds
);
1769 req
->r_sent_on_mseq
= cap
->mseq
;
1771 req
->r_sent_on_mseq
= -1;
1773 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req
,
1774 req
->r_tid
, ceph_mds_op_name(req
->r_op
), req
->r_attempts
);
1776 if (req
->r_got_unsafe
) {
1778 * Replay. Do not regenerate message (and rebuild
1779 * paths, etc.); just use the original message.
1780 * Rebuilding paths will break for renames because
1781 * d_move mangles the src name.
1783 msg
= req
->r_request
;
1784 rhead
= msg
->front
.iov_base
;
1786 flags
= le32_to_cpu(rhead
->flags
);
1787 flags
|= CEPH_MDS_FLAG_REPLAY
;
1788 rhead
->flags
= cpu_to_le32(flags
);
1790 if (req
->r_target_inode
)
1791 rhead
->ino
= cpu_to_le64(ceph_ino(req
->r_target_inode
));
1793 rhead
->num_retry
= req
->r_attempts
- 1;
1795 /* remove cap/dentry releases from message */
1796 rhead
->num_releases
= 0;
1797 msg
->hdr
.front_len
= cpu_to_le32(req
->r_request_release_offset
);
1798 msg
->front
.iov_len
= req
->r_request_release_offset
;
1802 if (req
->r_request
) {
1803 ceph_msg_put(req
->r_request
);
1804 req
->r_request
= NULL
;
1806 msg
= create_request_message(mdsc
, req
, mds
);
1808 req
->r_err
= PTR_ERR(msg
);
1809 complete_request(mdsc
, req
);
1810 return PTR_ERR(msg
);
1812 req
->r_request
= msg
;
1814 rhead
= msg
->front
.iov_base
;
1815 rhead
->oldest_client_tid
= cpu_to_le64(__get_oldest_tid(mdsc
));
1816 if (req
->r_got_unsafe
)
1817 flags
|= CEPH_MDS_FLAG_REPLAY
;
1818 if (req
->r_locked_dir
)
1819 flags
|= CEPH_MDS_FLAG_WANT_DENTRY
;
1820 rhead
->flags
= cpu_to_le32(flags
);
1821 rhead
->num_fwd
= req
->r_num_fwd
;
1822 rhead
->num_retry
= req
->r_attempts
- 1;
1825 dout(" r_locked_dir = %p\n", req
->r_locked_dir
);
1830 * send request, or put it on the appropriate wait list.
1832 static int __do_request(struct ceph_mds_client
*mdsc
,
1833 struct ceph_mds_request
*req
)
1835 struct ceph_mds_session
*session
= NULL
;
1839 if (req
->r_err
|| req
->r_got_result
)
1842 if (req
->r_timeout
&&
1843 time_after_eq(jiffies
, req
->r_started
+ req
->r_timeout
)) {
1844 dout("do_request timed out\n");
1849 put_request_session(req
);
1851 mds
= __choose_mds(mdsc
, req
);
1853 ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) < CEPH_MDS_STATE_ACTIVE
) {
1854 dout("do_request no mds or not active, waiting for map\n");
1855 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
1859 /* get, open session */
1860 session
= __ceph_lookup_mds_session(mdsc
, mds
);
1862 session
= register_session(mdsc
, mds
);
1863 if (IS_ERR(session
)) {
1864 err
= PTR_ERR(session
);
1868 req
->r_session
= get_session(session
);
1870 dout("do_request mds%d session %p state %s\n", mds
, session
,
1871 session_state_name(session
->s_state
));
1872 if (session
->s_state
!= CEPH_MDS_SESSION_OPEN
&&
1873 session
->s_state
!= CEPH_MDS_SESSION_HUNG
) {
1874 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
1875 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
1876 __open_session(mdsc
, session
);
1877 list_add(&req
->r_wait
, &session
->s_waiting
);
1882 req
->r_resend_mds
= -1; /* forget any previous mds hint */
1884 if (req
->r_request_started
== 0) /* note request start time */
1885 req
->r_request_started
= jiffies
;
1887 err
= __prepare_send_request(mdsc
, req
, mds
);
1889 ceph_msg_get(req
->r_request
);
1890 ceph_con_send(&session
->s_con
, req
->r_request
);
1894 ceph_put_mds_session(session
);
1900 complete_request(mdsc
, req
);
1905 * called under mdsc->mutex
1907 static void __wake_requests(struct ceph_mds_client
*mdsc
,
1908 struct list_head
*head
)
1910 struct ceph_mds_request
*req
;
1911 LIST_HEAD(tmp_list
);
1913 list_splice_init(head
, &tmp_list
);
1915 while (!list_empty(&tmp_list
)) {
1916 req
= list_entry(tmp_list
.next
,
1917 struct ceph_mds_request
, r_wait
);
1918 list_del_init(&req
->r_wait
);
1919 dout(" wake request %p tid %llu\n", req
, req
->r_tid
);
1920 __do_request(mdsc
, req
);
1925 * Wake up threads with requests pending for @mds, so that they can
1926 * resubmit their requests to a possibly different mds.
1928 static void kick_requests(struct ceph_mds_client
*mdsc
, int mds
)
1930 struct ceph_mds_request
*req
;
1933 dout("kick_requests mds%d\n", mds
);
1934 for (p
= rb_first(&mdsc
->request_tree
); p
; p
= rb_next(p
)) {
1935 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
1936 if (req
->r_got_unsafe
)
1938 if (req
->r_session
&&
1939 req
->r_session
->s_mds
== mds
) {
1940 dout(" kicking tid %llu\n", req
->r_tid
);
1941 __do_request(mdsc
, req
);
1946 void ceph_mdsc_submit_request(struct ceph_mds_client
*mdsc
,
1947 struct ceph_mds_request
*req
)
1949 dout("submit_request on %p\n", req
);
1950 mutex_lock(&mdsc
->mutex
);
1951 __register_request(mdsc
, req
, NULL
);
1952 __do_request(mdsc
, req
);
1953 mutex_unlock(&mdsc
->mutex
);
1957 * Synchrously perform an mds request. Take care of all of the
1958 * session setup, forwarding, retry details.
1960 int ceph_mdsc_do_request(struct ceph_mds_client
*mdsc
,
1962 struct ceph_mds_request
*req
)
1966 dout("do_request on %p\n", req
);
1968 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1970 ceph_get_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
1971 if (req
->r_locked_dir
)
1972 ceph_get_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
1973 if (req
->r_old_dentry
)
1974 ceph_get_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
1978 mutex_lock(&mdsc
->mutex
);
1979 __register_request(mdsc
, req
, dir
);
1980 __do_request(mdsc
, req
);
1984 __unregister_request(mdsc
, req
);
1985 dout("do_request early error %d\n", err
);
1990 mutex_unlock(&mdsc
->mutex
);
1991 dout("do_request waiting\n");
1992 if (req
->r_timeout
) {
1993 err
= (long)wait_for_completion_killable_timeout(
1994 &req
->r_completion
, req
->r_timeout
);
1998 err
= wait_for_completion_killable(&req
->r_completion
);
2000 dout("do_request waited, got %d\n", err
);
2001 mutex_lock(&mdsc
->mutex
);
2003 /* only abort if we didn't race with a real reply */
2004 if (req
->r_got_result
) {
2005 err
= le32_to_cpu(req
->r_reply_info
.head
->result
);
2006 } else if (err
< 0) {
2007 dout("aborted request %lld with %d\n", req
->r_tid
, err
);
2010 * ensure we aren't running concurrently with
2011 * ceph_fill_trace or ceph_readdir_prepopulate, which
2012 * rely on locks (dir mutex) held by our caller.
2014 mutex_lock(&req
->r_fill_mutex
);
2016 req
->r_aborted
= true;
2017 mutex_unlock(&req
->r_fill_mutex
);
2019 if (req
->r_locked_dir
&&
2020 (req
->r_op
& CEPH_MDS_OP_WRITE
))
2021 ceph_invalidate_dir_request(req
);
2027 mutex_unlock(&mdsc
->mutex
);
2028 dout("do_request %p done, result %d\n", req
, err
);
2033 * Invalidate dir I_COMPLETE, dentry lease state on an aborted MDS
2034 * namespace request.
2036 void ceph_invalidate_dir_request(struct ceph_mds_request
*req
)
2038 struct inode
*inode
= req
->r_locked_dir
;
2039 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2041 dout("invalidate_dir_request %p (I_COMPLETE, lease(s))\n", inode
);
2042 spin_lock(&ci
->i_ceph_lock
);
2043 ci
->i_ceph_flags
&= ~CEPH_I_COMPLETE
;
2044 ci
->i_release_count
++;
2045 spin_unlock(&ci
->i_ceph_lock
);
2048 ceph_invalidate_dentry_lease(req
->r_dentry
);
2049 if (req
->r_old_dentry
)
2050 ceph_invalidate_dentry_lease(req
->r_old_dentry
);
2056 * We take the session mutex and parse and process the reply immediately.
2057 * This preserves the logical ordering of replies, capabilities, etc., sent
2058 * by the MDS as they are applied to our local cache.
2060 static void handle_reply(struct ceph_mds_session
*session
, struct ceph_msg
*msg
)
2062 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2063 struct ceph_mds_request
*req
;
2064 struct ceph_mds_reply_head
*head
= msg
->front
.iov_base
;
2065 struct ceph_mds_reply_info_parsed
*rinfo
; /* parsed reply info */
2068 int mds
= session
->s_mds
;
2070 if (msg
->front
.iov_len
< sizeof(*head
)) {
2071 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2076 /* get request, session */
2077 tid
= le64_to_cpu(msg
->hdr
.tid
);
2078 mutex_lock(&mdsc
->mutex
);
2079 req
= __lookup_request(mdsc
, tid
);
2081 dout("handle_reply on unknown tid %llu\n", tid
);
2082 mutex_unlock(&mdsc
->mutex
);
2085 dout("handle_reply %p\n", req
);
2087 /* correct session? */
2088 if (req
->r_session
!= session
) {
2089 pr_err("mdsc_handle_reply got %llu on session mds%d"
2090 " not mds%d\n", tid
, session
->s_mds
,
2091 req
->r_session
? req
->r_session
->s_mds
: -1);
2092 mutex_unlock(&mdsc
->mutex
);
2097 if ((req
->r_got_unsafe
&& !head
->safe
) ||
2098 (req
->r_got_safe
&& head
->safe
)) {
2099 pr_warning("got a dup %s reply on %llu from mds%d\n",
2100 head
->safe
? "safe" : "unsafe", tid
, mds
);
2101 mutex_unlock(&mdsc
->mutex
);
2104 if (req
->r_got_safe
&& !head
->safe
) {
2105 pr_warning("got unsafe after safe on %llu from mds%d\n",
2107 mutex_unlock(&mdsc
->mutex
);
2111 result
= le32_to_cpu(head
->result
);
2115 * if we're not talking to the authority, send to them
2116 * if the authority has changed while we weren't looking,
2117 * send to new authority
2118 * Otherwise we just have to return an ESTALE
2120 if (result
== -ESTALE
) {
2121 dout("got ESTALE on request %llu", req
->r_tid
);
2122 if (!req
->r_inode
) {
2123 /* do nothing; not an authority problem */
2124 } else if (req
->r_direct_mode
!= USE_AUTH_MDS
) {
2125 dout("not using auth, setting for that now");
2126 req
->r_direct_mode
= USE_AUTH_MDS
;
2127 __do_request(mdsc
, req
);
2128 mutex_unlock(&mdsc
->mutex
);
2131 struct ceph_inode_info
*ci
= ceph_inode(req
->r_inode
);
2132 struct ceph_cap
*cap
= NULL
;
2135 cap
= ceph_get_cap_for_mds(ci
,
2136 req
->r_session
->s_mds
);
2138 dout("already using auth");
2139 if ((!cap
|| cap
!= ci
->i_auth_cap
) ||
2140 (cap
->mseq
!= req
->r_sent_on_mseq
)) {
2141 dout("but cap changed, so resending");
2142 __do_request(mdsc
, req
);
2143 mutex_unlock(&mdsc
->mutex
);
2147 dout("have to return ESTALE on request %llu", req
->r_tid
);
2152 req
->r_got_safe
= true;
2153 __unregister_request(mdsc
, req
);
2154 complete_all(&req
->r_safe_completion
);
2156 if (req
->r_got_unsafe
) {
2158 * We already handled the unsafe response, now do the
2159 * cleanup. No need to examine the response; the MDS
2160 * doesn't include any result info in the safe
2161 * response. And even if it did, there is nothing
2162 * useful we could do with a revised return value.
2164 dout("got safe reply %llu, mds%d\n", tid
, mds
);
2165 list_del_init(&req
->r_unsafe_item
);
2167 /* last unsafe request during umount? */
2168 if (mdsc
->stopping
&& !__get_oldest_req(mdsc
))
2169 complete_all(&mdsc
->safe_umount_waiters
);
2170 mutex_unlock(&mdsc
->mutex
);
2174 req
->r_got_unsafe
= true;
2175 list_add_tail(&req
->r_unsafe_item
, &req
->r_session
->s_unsafe
);
2178 dout("handle_reply tid %lld result %d\n", tid
, result
);
2179 rinfo
= &req
->r_reply_info
;
2180 err
= parse_reply_info(msg
, rinfo
, session
->s_con
.peer_features
);
2181 mutex_unlock(&mdsc
->mutex
);
2183 mutex_lock(&session
->s_mutex
);
2185 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds
, tid
);
2191 if (rinfo
->snapblob_len
) {
2192 down_write(&mdsc
->snap_rwsem
);
2193 ceph_update_snap_trace(mdsc
, rinfo
->snapblob
,
2194 rinfo
->snapblob
+ rinfo
->snapblob_len
,
2195 le32_to_cpu(head
->op
) == CEPH_MDS_OP_RMSNAP
);
2196 downgrade_write(&mdsc
->snap_rwsem
);
2198 down_read(&mdsc
->snap_rwsem
);
2201 /* insert trace into our cache */
2202 mutex_lock(&req
->r_fill_mutex
);
2203 err
= ceph_fill_trace(mdsc
->fsc
->sb
, req
, req
->r_session
);
2205 if (result
== 0 && (req
->r_op
== CEPH_MDS_OP_READDIR
||
2206 req
->r_op
== CEPH_MDS_OP_LSSNAP
) &&
2208 ceph_readdir_prepopulate(req
, req
->r_session
);
2209 ceph_unreserve_caps(mdsc
, &req
->r_caps_reservation
);
2211 mutex_unlock(&req
->r_fill_mutex
);
2213 up_read(&mdsc
->snap_rwsem
);
2215 mutex_lock(&mdsc
->mutex
);
2216 if (!req
->r_aborted
) {
2222 req
->r_got_result
= true;
2225 dout("reply arrived after request %lld was aborted\n", tid
);
2227 mutex_unlock(&mdsc
->mutex
);
2229 ceph_add_cap_releases(mdsc
, req
->r_session
);
2230 mutex_unlock(&session
->s_mutex
);
2232 /* kick calling process */
2233 complete_request(mdsc
, req
);
2235 ceph_mdsc_put_request(req
);
2242 * handle mds notification that our request has been forwarded.
2244 static void handle_forward(struct ceph_mds_client
*mdsc
,
2245 struct ceph_mds_session
*session
,
2246 struct ceph_msg
*msg
)
2248 struct ceph_mds_request
*req
;
2249 u64 tid
= le64_to_cpu(msg
->hdr
.tid
);
2253 void *p
= msg
->front
.iov_base
;
2254 void *end
= p
+ msg
->front
.iov_len
;
2256 ceph_decode_need(&p
, end
, 2*sizeof(u32
), bad
);
2257 next_mds
= ceph_decode_32(&p
);
2258 fwd_seq
= ceph_decode_32(&p
);
2260 mutex_lock(&mdsc
->mutex
);
2261 req
= __lookup_request(mdsc
, tid
);
2263 dout("forward tid %llu to mds%d - req dne\n", tid
, next_mds
);
2264 goto out
; /* dup reply? */
2267 if (req
->r_aborted
) {
2268 dout("forward tid %llu aborted, unregistering\n", tid
);
2269 __unregister_request(mdsc
, req
);
2270 } else if (fwd_seq
<= req
->r_num_fwd
) {
2271 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2272 tid
, next_mds
, req
->r_num_fwd
, fwd_seq
);
2274 /* resend. forward race not possible; mds would drop */
2275 dout("forward tid %llu to mds%d (we resend)\n", tid
, next_mds
);
2277 BUG_ON(req
->r_got_result
);
2278 req
->r_num_fwd
= fwd_seq
;
2279 req
->r_resend_mds
= next_mds
;
2280 put_request_session(req
);
2281 __do_request(mdsc
, req
);
2283 ceph_mdsc_put_request(req
);
2285 mutex_unlock(&mdsc
->mutex
);
2289 pr_err("mdsc_handle_forward decode error err=%d\n", err
);
2293 * handle a mds session control message
2295 static void handle_session(struct ceph_mds_session
*session
,
2296 struct ceph_msg
*msg
)
2298 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2301 int mds
= session
->s_mds
;
2302 struct ceph_mds_session_head
*h
= msg
->front
.iov_base
;
2306 if (msg
->front
.iov_len
!= sizeof(*h
))
2308 op
= le32_to_cpu(h
->op
);
2309 seq
= le64_to_cpu(h
->seq
);
2311 mutex_lock(&mdsc
->mutex
);
2312 if (op
== CEPH_SESSION_CLOSE
)
2313 __unregister_session(mdsc
, session
);
2314 /* FIXME: this ttl calculation is generous */
2315 session
->s_ttl
= jiffies
+ HZ
*mdsc
->mdsmap
->m_session_autoclose
;
2316 mutex_unlock(&mdsc
->mutex
);
2318 mutex_lock(&session
->s_mutex
);
2320 dout("handle_session mds%d %s %p state %s seq %llu\n",
2321 mds
, ceph_session_op_name(op
), session
,
2322 session_state_name(session
->s_state
), seq
);
2324 if (session
->s_state
== CEPH_MDS_SESSION_HUNG
) {
2325 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2326 pr_info("mds%d came back\n", session
->s_mds
);
2330 case CEPH_SESSION_OPEN
:
2331 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2332 pr_info("mds%d reconnect success\n", session
->s_mds
);
2333 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2334 renewed_caps(mdsc
, session
, 0);
2337 __close_session(mdsc
, session
);
2340 case CEPH_SESSION_RENEWCAPS
:
2341 if (session
->s_renew_seq
== seq
)
2342 renewed_caps(mdsc
, session
, 1);
2345 case CEPH_SESSION_CLOSE
:
2346 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2347 pr_info("mds%d reconnect denied\n", session
->s_mds
);
2348 remove_session_caps(session
);
2349 wake
= 1; /* for good measure */
2350 wake_up_all(&mdsc
->session_close_wq
);
2351 kick_requests(mdsc
, mds
);
2354 case CEPH_SESSION_STALE
:
2355 pr_info("mds%d caps went stale, renewing\n",
2357 spin_lock(&session
->s_gen_ttl_lock
);
2358 session
->s_cap_gen
++;
2359 session
->s_cap_ttl
= jiffies
- 1;
2360 spin_unlock(&session
->s_gen_ttl_lock
);
2361 send_renew_caps(mdsc
, session
);
2364 case CEPH_SESSION_RECALL_STATE
:
2365 trim_caps(mdsc
, session
, le32_to_cpu(h
->max_caps
));
2369 pr_err("mdsc_handle_session bad op %d mds%d\n", op
, mds
);
2373 mutex_unlock(&session
->s_mutex
);
2375 mutex_lock(&mdsc
->mutex
);
2376 __wake_requests(mdsc
, &session
->s_waiting
);
2377 mutex_unlock(&mdsc
->mutex
);
2382 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds
,
2383 (int)msg
->front
.iov_len
);
2390 * called under session->mutex.
2392 static void replay_unsafe_requests(struct ceph_mds_client
*mdsc
,
2393 struct ceph_mds_session
*session
)
2395 struct ceph_mds_request
*req
, *nreq
;
2398 dout("replay_unsafe_requests mds%d\n", session
->s_mds
);
2400 mutex_lock(&mdsc
->mutex
);
2401 list_for_each_entry_safe(req
, nreq
, &session
->s_unsafe
, r_unsafe_item
) {
2402 err
= __prepare_send_request(mdsc
, req
, session
->s_mds
);
2404 ceph_msg_get(req
->r_request
);
2405 ceph_con_send(&session
->s_con
, req
->r_request
);
2408 mutex_unlock(&mdsc
->mutex
);
2412 * Encode information about a cap for a reconnect with the MDS.
2414 static int encode_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
2418 struct ceph_mds_cap_reconnect v2
;
2419 struct ceph_mds_cap_reconnect_v1 v1
;
2422 struct ceph_inode_info
*ci
;
2423 struct ceph_reconnect_state
*recon_state
= arg
;
2424 struct ceph_pagelist
*pagelist
= recon_state
->pagelist
;
2428 struct dentry
*dentry
;
2432 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2433 inode
, ceph_vinop(inode
), cap
, cap
->cap_id
,
2434 ceph_cap_string(cap
->issued
));
2435 err
= ceph_pagelist_encode_64(pagelist
, ceph_ino(inode
));
2439 dentry
= d_find_alias(inode
);
2441 path
= ceph_mdsc_build_path(dentry
, &pathlen
, &pathbase
, 0);
2443 err
= PTR_ERR(path
);
2450 err
= ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
2454 spin_lock(&ci
->i_ceph_lock
);
2455 cap
->seq
= 0; /* reset cap seq */
2456 cap
->issue_seq
= 0; /* and issue_seq */
2458 if (recon_state
->flock
) {
2459 rec
.v2
.cap_id
= cpu_to_le64(cap
->cap_id
);
2460 rec
.v2
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2461 rec
.v2
.issued
= cpu_to_le32(cap
->issued
);
2462 rec
.v2
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2463 rec
.v2
.pathbase
= cpu_to_le64(pathbase
);
2464 rec
.v2
.flock_len
= 0;
2465 reclen
= sizeof(rec
.v2
);
2467 rec
.v1
.cap_id
= cpu_to_le64(cap
->cap_id
);
2468 rec
.v1
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2469 rec
.v1
.issued
= cpu_to_le32(cap
->issued
);
2470 rec
.v1
.size
= cpu_to_le64(inode
->i_size
);
2471 ceph_encode_timespec(&rec
.v1
.mtime
, &inode
->i_mtime
);
2472 ceph_encode_timespec(&rec
.v1
.atime
, &inode
->i_atime
);
2473 rec
.v1
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2474 rec
.v1
.pathbase
= cpu_to_le64(pathbase
);
2475 reclen
= sizeof(rec
.v1
);
2477 spin_unlock(&ci
->i_ceph_lock
);
2479 if (recon_state
->flock
) {
2480 int num_fcntl_locks
, num_flock_locks
;
2481 struct ceph_pagelist_cursor trunc_point
;
2483 ceph_pagelist_set_cursor(pagelist
, &trunc_point
);
2486 ceph_count_locks(inode
, &num_fcntl_locks
,
2488 rec
.v2
.flock_len
= (2*sizeof(u32
) +
2489 (num_fcntl_locks
+num_flock_locks
) *
2490 sizeof(struct ceph_filelock
));
2493 /* pre-alloc pagelist */
2494 ceph_pagelist_truncate(pagelist
, &trunc_point
);
2495 err
= ceph_pagelist_append(pagelist
, &rec
, reclen
);
2497 err
= ceph_pagelist_reserve(pagelist
,
2503 err
= ceph_encode_locks(inode
,
2509 } while (err
== -ENOSPC
);
2511 err
= ceph_pagelist_append(pagelist
, &rec
, reclen
);
2523 * If an MDS fails and recovers, clients need to reconnect in order to
2524 * reestablish shared state. This includes all caps issued through
2525 * this session _and_ the snap_realm hierarchy. Because it's not
2526 * clear which snap realms the mds cares about, we send everything we
2527 * know about.. that ensures we'll then get any new info the
2528 * recovering MDS might have.
2530 * This is a relatively heavyweight operation, but it's rare.
2532 * called with mdsc->mutex held.
2534 static void send_mds_reconnect(struct ceph_mds_client
*mdsc
,
2535 struct ceph_mds_session
*session
)
2537 struct ceph_msg
*reply
;
2539 int mds
= session
->s_mds
;
2541 struct ceph_pagelist
*pagelist
;
2542 struct ceph_reconnect_state recon_state
;
2544 pr_info("mds%d reconnect start\n", mds
);
2546 pagelist
= kmalloc(sizeof(*pagelist
), GFP_NOFS
);
2548 goto fail_nopagelist
;
2549 ceph_pagelist_init(pagelist
);
2551 reply
= ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT
, 0, GFP_NOFS
, false);
2555 mutex_lock(&session
->s_mutex
);
2556 session
->s_state
= CEPH_MDS_SESSION_RECONNECTING
;
2559 ceph_con_close(&session
->s_con
);
2560 ceph_con_open(&session
->s_con
,
2561 CEPH_ENTITY_TYPE_MDS
, mds
,
2562 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
2564 /* replay unsafe requests */
2565 replay_unsafe_requests(mdsc
, session
);
2567 down_read(&mdsc
->snap_rwsem
);
2569 dout("session %p state %s\n", session
,
2570 session_state_name(session
->s_state
));
2572 /* drop old cap expires; we're about to reestablish that state */
2573 discard_cap_releases(mdsc
, session
);
2575 /* traverse this session's caps */
2576 err
= ceph_pagelist_encode_32(pagelist
, session
->s_nr_caps
);
2580 recon_state
.pagelist
= pagelist
;
2581 recon_state
.flock
= session
->s_con
.peer_features
& CEPH_FEATURE_FLOCK
;
2582 err
= iterate_session_caps(session
, encode_caps_cb
, &recon_state
);
2587 * snaprealms. we provide mds with the ino, seq (version), and
2588 * parent for all of our realms. If the mds has any newer info,
2591 for (p
= rb_first(&mdsc
->snap_realms
); p
; p
= rb_next(p
)) {
2592 struct ceph_snap_realm
*realm
=
2593 rb_entry(p
, struct ceph_snap_realm
, node
);
2594 struct ceph_mds_snaprealm_reconnect sr_rec
;
2596 dout(" adding snap realm %llx seq %lld parent %llx\n",
2597 realm
->ino
, realm
->seq
, realm
->parent_ino
);
2598 sr_rec
.ino
= cpu_to_le64(realm
->ino
);
2599 sr_rec
.seq
= cpu_to_le64(realm
->seq
);
2600 sr_rec
.parent
= cpu_to_le64(realm
->parent_ino
);
2601 err
= ceph_pagelist_append(pagelist
, &sr_rec
, sizeof(sr_rec
));
2606 reply
->pagelist
= pagelist
;
2607 if (recon_state
.flock
)
2608 reply
->hdr
.version
= cpu_to_le16(2);
2609 reply
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
2610 ceph_con_send(&session
->s_con
, reply
);
2612 mutex_unlock(&session
->s_mutex
);
2614 mutex_lock(&mdsc
->mutex
);
2615 __wake_requests(mdsc
, &session
->s_waiting
);
2616 mutex_unlock(&mdsc
->mutex
);
2618 up_read(&mdsc
->snap_rwsem
);
2622 ceph_msg_put(reply
);
2623 up_read(&mdsc
->snap_rwsem
);
2624 mutex_unlock(&session
->s_mutex
);
2626 ceph_pagelist_release(pagelist
);
2629 pr_err("error %d preparing reconnect for mds%d\n", err
, mds
);
2635 * compare old and new mdsmaps, kicking requests
2636 * and closing out old connections as necessary
2638 * called under mdsc->mutex.
2640 static void check_new_map(struct ceph_mds_client
*mdsc
,
2641 struct ceph_mdsmap
*newmap
,
2642 struct ceph_mdsmap
*oldmap
)
2645 int oldstate
, newstate
;
2646 struct ceph_mds_session
*s
;
2648 dout("check_new_map new %u old %u\n",
2649 newmap
->m_epoch
, oldmap
->m_epoch
);
2651 for (i
= 0; i
< oldmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
2652 if (mdsc
->sessions
[i
] == NULL
)
2654 s
= mdsc
->sessions
[i
];
2655 oldstate
= ceph_mdsmap_get_state(oldmap
, i
);
2656 newstate
= ceph_mdsmap_get_state(newmap
, i
);
2658 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2659 i
, ceph_mds_state_name(oldstate
),
2660 ceph_mdsmap_is_laggy(oldmap
, i
) ? " (laggy)" : "",
2661 ceph_mds_state_name(newstate
),
2662 ceph_mdsmap_is_laggy(newmap
, i
) ? " (laggy)" : "",
2663 session_state_name(s
->s_state
));
2665 if (i
>= newmap
->m_max_mds
||
2666 memcmp(ceph_mdsmap_get_addr(oldmap
, i
),
2667 ceph_mdsmap_get_addr(newmap
, i
),
2668 sizeof(struct ceph_entity_addr
))) {
2669 if (s
->s_state
== CEPH_MDS_SESSION_OPENING
) {
2670 /* the session never opened, just close it
2672 __wake_requests(mdsc
, &s
->s_waiting
);
2673 __unregister_session(mdsc
, s
);
2676 mutex_unlock(&mdsc
->mutex
);
2677 mutex_lock(&s
->s_mutex
);
2678 mutex_lock(&mdsc
->mutex
);
2679 ceph_con_close(&s
->s_con
);
2680 mutex_unlock(&s
->s_mutex
);
2681 s
->s_state
= CEPH_MDS_SESSION_RESTARTING
;
2684 /* kick any requests waiting on the recovering mds */
2685 kick_requests(mdsc
, i
);
2686 } else if (oldstate
== newstate
) {
2687 continue; /* nothing new with this mds */
2693 if (s
->s_state
== CEPH_MDS_SESSION_RESTARTING
&&
2694 newstate
>= CEPH_MDS_STATE_RECONNECT
) {
2695 mutex_unlock(&mdsc
->mutex
);
2696 send_mds_reconnect(mdsc
, s
);
2697 mutex_lock(&mdsc
->mutex
);
2701 * kick request on any mds that has gone active.
2703 if (oldstate
< CEPH_MDS_STATE_ACTIVE
&&
2704 newstate
>= CEPH_MDS_STATE_ACTIVE
) {
2705 if (oldstate
!= CEPH_MDS_STATE_CREATING
&&
2706 oldstate
!= CEPH_MDS_STATE_STARTING
)
2707 pr_info("mds%d recovery completed\n", s
->s_mds
);
2708 kick_requests(mdsc
, i
);
2709 ceph_kick_flushing_caps(mdsc
, s
);
2710 wake_up_session_caps(s
, 1);
2714 for (i
= 0; i
< newmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
2715 s
= mdsc
->sessions
[i
];
2718 if (!ceph_mdsmap_is_laggy(newmap
, i
))
2720 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
2721 s
->s_state
== CEPH_MDS_SESSION_HUNG
||
2722 s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
2723 dout(" connecting to export targets of laggy mds%d\n",
2725 __open_export_target_sessions(mdsc
, s
);
2737 * caller must hold session s_mutex, dentry->d_lock
2739 void __ceph_mdsc_drop_dentry_lease(struct dentry
*dentry
)
2741 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
2743 ceph_put_mds_session(di
->lease_session
);
2744 di
->lease_session
= NULL
;
2747 static void handle_lease(struct ceph_mds_client
*mdsc
,
2748 struct ceph_mds_session
*session
,
2749 struct ceph_msg
*msg
)
2751 struct super_block
*sb
= mdsc
->fsc
->sb
;
2752 struct inode
*inode
;
2753 struct dentry
*parent
, *dentry
;
2754 struct ceph_dentry_info
*di
;
2755 int mds
= session
->s_mds
;
2756 struct ceph_mds_lease
*h
= msg
->front
.iov_base
;
2758 struct ceph_vino vino
;
2762 dout("handle_lease from mds%d\n", mds
);
2765 if (msg
->front
.iov_len
< sizeof(*h
) + sizeof(u32
))
2767 vino
.ino
= le64_to_cpu(h
->ino
);
2768 vino
.snap
= CEPH_NOSNAP
;
2769 seq
= le32_to_cpu(h
->seq
);
2770 dname
.name
= (void *)h
+ sizeof(*h
) + sizeof(u32
);
2771 dname
.len
= msg
->front
.iov_len
- sizeof(*h
) - sizeof(u32
);
2772 if (dname
.len
!= get_unaligned_le32(h
+1))
2775 mutex_lock(&session
->s_mutex
);
2779 inode
= ceph_find_inode(sb
, vino
);
2780 dout("handle_lease %s, ino %llx %p %.*s\n",
2781 ceph_lease_op_name(h
->action
), vino
.ino
, inode
,
2782 dname
.len
, dname
.name
);
2783 if (inode
== NULL
) {
2784 dout("handle_lease no inode %llx\n", vino
.ino
);
2789 parent
= d_find_alias(inode
);
2791 dout("no parent dentry on inode %p\n", inode
);
2793 goto release
; /* hrm... */
2795 dname
.hash
= full_name_hash(dname
.name
, dname
.len
);
2796 dentry
= d_lookup(parent
, &dname
);
2801 spin_lock(&dentry
->d_lock
);
2802 di
= ceph_dentry(dentry
);
2803 switch (h
->action
) {
2804 case CEPH_MDS_LEASE_REVOKE
:
2805 if (di
->lease_session
== session
) {
2806 if (ceph_seq_cmp(di
->lease_seq
, seq
) > 0)
2807 h
->seq
= cpu_to_le32(di
->lease_seq
);
2808 __ceph_mdsc_drop_dentry_lease(dentry
);
2813 case CEPH_MDS_LEASE_RENEW
:
2814 if (di
->lease_session
== session
&&
2815 di
->lease_gen
== session
->s_cap_gen
&&
2816 di
->lease_renew_from
&&
2817 di
->lease_renew_after
== 0) {
2818 unsigned long duration
=
2819 le32_to_cpu(h
->duration_ms
) * HZ
/ 1000;
2821 di
->lease_seq
= seq
;
2822 dentry
->d_time
= di
->lease_renew_from
+ duration
;
2823 di
->lease_renew_after
= di
->lease_renew_from
+
2825 di
->lease_renew_from
= 0;
2829 spin_unlock(&dentry
->d_lock
);
2836 /* let's just reuse the same message */
2837 h
->action
= CEPH_MDS_LEASE_REVOKE_ACK
;
2839 ceph_con_send(&session
->s_con
, msg
);
2843 mutex_unlock(&session
->s_mutex
);
2847 pr_err("corrupt lease message\n");
2851 void ceph_mdsc_lease_send_msg(struct ceph_mds_session
*session
,
2852 struct inode
*inode
,
2853 struct dentry
*dentry
, char action
,
2856 struct ceph_msg
*msg
;
2857 struct ceph_mds_lease
*lease
;
2858 int len
= sizeof(*lease
) + sizeof(u32
);
2861 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2862 inode
, dentry
, ceph_lease_op_name(action
), session
->s_mds
);
2863 dnamelen
= dentry
->d_name
.len
;
2866 msg
= ceph_msg_new(CEPH_MSG_CLIENT_LEASE
, len
, GFP_NOFS
, false);
2869 lease
= msg
->front
.iov_base
;
2870 lease
->action
= action
;
2871 lease
->ino
= cpu_to_le64(ceph_vino(inode
).ino
);
2872 lease
->first
= lease
->last
= cpu_to_le64(ceph_vino(inode
).snap
);
2873 lease
->seq
= cpu_to_le32(seq
);
2874 put_unaligned_le32(dnamelen
, lease
+ 1);
2875 memcpy((void *)(lease
+ 1) + 4, dentry
->d_name
.name
, dnamelen
);
2878 * if this is a preemptive lease RELEASE, no need to
2879 * flush request stream, since the actual request will
2882 msg
->more_to_follow
= (action
== CEPH_MDS_LEASE_RELEASE
);
2884 ceph_con_send(&session
->s_con
, msg
);
2888 * Preemptively release a lease we expect to invalidate anyway.
2889 * Pass @inode always, @dentry is optional.
2891 void ceph_mdsc_lease_release(struct ceph_mds_client
*mdsc
, struct inode
*inode
,
2892 struct dentry
*dentry
)
2894 struct ceph_dentry_info
*di
;
2895 struct ceph_mds_session
*session
;
2898 BUG_ON(inode
== NULL
);
2899 BUG_ON(dentry
== NULL
);
2901 /* is dentry lease valid? */
2902 spin_lock(&dentry
->d_lock
);
2903 di
= ceph_dentry(dentry
);
2904 if (!di
|| !di
->lease_session
||
2905 di
->lease_session
->s_mds
< 0 ||
2906 di
->lease_gen
!= di
->lease_session
->s_cap_gen
||
2907 !time_before(jiffies
, dentry
->d_time
)) {
2908 dout("lease_release inode %p dentry %p -- "
2911 spin_unlock(&dentry
->d_lock
);
2915 /* we do have a lease on this dentry; note mds and seq */
2916 session
= ceph_get_mds_session(di
->lease_session
);
2917 seq
= di
->lease_seq
;
2918 __ceph_mdsc_drop_dentry_lease(dentry
);
2919 spin_unlock(&dentry
->d_lock
);
2921 dout("lease_release inode %p dentry %p to mds%d\n",
2922 inode
, dentry
, session
->s_mds
);
2923 ceph_mdsc_lease_send_msg(session
, inode
, dentry
,
2924 CEPH_MDS_LEASE_RELEASE
, seq
);
2925 ceph_put_mds_session(session
);
2929 * drop all leases (and dentry refs) in preparation for umount
2931 static void drop_leases(struct ceph_mds_client
*mdsc
)
2935 dout("drop_leases\n");
2936 mutex_lock(&mdsc
->mutex
);
2937 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
2938 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
2941 mutex_unlock(&mdsc
->mutex
);
2942 mutex_lock(&s
->s_mutex
);
2943 mutex_unlock(&s
->s_mutex
);
2944 ceph_put_mds_session(s
);
2945 mutex_lock(&mdsc
->mutex
);
2947 mutex_unlock(&mdsc
->mutex
);
2953 * delayed work -- periodically trim expired leases, renew caps with mds
2955 static void schedule_delayed(struct ceph_mds_client
*mdsc
)
2958 unsigned hz
= round_jiffies_relative(HZ
* delay
);
2959 schedule_delayed_work(&mdsc
->delayed_work
, hz
);
2962 static void delayed_work(struct work_struct
*work
)
2965 struct ceph_mds_client
*mdsc
=
2966 container_of(work
, struct ceph_mds_client
, delayed_work
.work
);
2970 dout("mdsc delayed_work\n");
2971 ceph_check_delayed_caps(mdsc
);
2973 mutex_lock(&mdsc
->mutex
);
2974 renew_interval
= mdsc
->mdsmap
->m_session_timeout
>> 2;
2975 renew_caps
= time_after_eq(jiffies
, HZ
*renew_interval
+
2976 mdsc
->last_renew_caps
);
2978 mdsc
->last_renew_caps
= jiffies
;
2980 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
2981 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
2984 if (s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
2985 dout("resending session close request for mds%d\n",
2987 request_close_session(mdsc
, s
);
2988 ceph_put_mds_session(s
);
2991 if (s
->s_ttl
&& time_after(jiffies
, s
->s_ttl
)) {
2992 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
) {
2993 s
->s_state
= CEPH_MDS_SESSION_HUNG
;
2994 pr_info("mds%d hung\n", s
->s_mds
);
2997 if (s
->s_state
< CEPH_MDS_SESSION_OPEN
) {
2998 /* this mds is failed or recovering, just wait */
2999 ceph_put_mds_session(s
);
3002 mutex_unlock(&mdsc
->mutex
);
3004 mutex_lock(&s
->s_mutex
);
3006 send_renew_caps(mdsc
, s
);
3008 ceph_con_keepalive(&s
->s_con
);
3009 ceph_add_cap_releases(mdsc
, s
);
3010 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3011 s
->s_state
== CEPH_MDS_SESSION_HUNG
)
3012 ceph_send_cap_releases(mdsc
, s
);
3013 mutex_unlock(&s
->s_mutex
);
3014 ceph_put_mds_session(s
);
3016 mutex_lock(&mdsc
->mutex
);
3018 mutex_unlock(&mdsc
->mutex
);
3020 schedule_delayed(mdsc
);
3023 int ceph_mdsc_init(struct ceph_fs_client
*fsc
)
3026 struct ceph_mds_client
*mdsc
;
3028 mdsc
= kzalloc(sizeof(struct ceph_mds_client
), GFP_NOFS
);
3033 mutex_init(&mdsc
->mutex
);
3034 mdsc
->mdsmap
= kzalloc(sizeof(*mdsc
->mdsmap
), GFP_NOFS
);
3035 if (mdsc
->mdsmap
== NULL
)
3038 init_completion(&mdsc
->safe_umount_waiters
);
3039 init_waitqueue_head(&mdsc
->session_close_wq
);
3040 INIT_LIST_HEAD(&mdsc
->waiting_for_map
);
3041 mdsc
->sessions
= NULL
;
3042 mdsc
->max_sessions
= 0;
3044 init_rwsem(&mdsc
->snap_rwsem
);
3045 mdsc
->snap_realms
= RB_ROOT
;
3046 INIT_LIST_HEAD(&mdsc
->snap_empty
);
3047 spin_lock_init(&mdsc
->snap_empty_lock
);
3049 mdsc
->request_tree
= RB_ROOT
;
3050 INIT_DELAYED_WORK(&mdsc
->delayed_work
, delayed_work
);
3051 mdsc
->last_renew_caps
= jiffies
;
3052 INIT_LIST_HEAD(&mdsc
->cap_delay_list
);
3053 spin_lock_init(&mdsc
->cap_delay_lock
);
3054 INIT_LIST_HEAD(&mdsc
->snap_flush_list
);
3055 spin_lock_init(&mdsc
->snap_flush_lock
);
3056 mdsc
->cap_flush_seq
= 0;
3057 INIT_LIST_HEAD(&mdsc
->cap_dirty
);
3058 INIT_LIST_HEAD(&mdsc
->cap_dirty_migrating
);
3059 mdsc
->num_cap_flushing
= 0;
3060 spin_lock_init(&mdsc
->cap_dirty_lock
);
3061 init_waitqueue_head(&mdsc
->cap_flushing_wq
);
3062 spin_lock_init(&mdsc
->dentry_lru_lock
);
3063 INIT_LIST_HEAD(&mdsc
->dentry_lru
);
3065 ceph_caps_init(mdsc
);
3066 ceph_adjust_min_caps(mdsc
, fsc
->min_caps
);
3072 * Wait for safe replies on open mds requests. If we time out, drop
3073 * all requests from the tree to avoid dangling dentry refs.
3075 static void wait_requests(struct ceph_mds_client
*mdsc
)
3077 struct ceph_mds_request
*req
;
3078 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3080 mutex_lock(&mdsc
->mutex
);
3081 if (__get_oldest_req(mdsc
)) {
3082 mutex_unlock(&mdsc
->mutex
);
3084 dout("wait_requests waiting for requests\n");
3085 wait_for_completion_timeout(&mdsc
->safe_umount_waiters
,
3086 fsc
->client
->options
->mount_timeout
* HZ
);
3088 /* tear down remaining requests */
3089 mutex_lock(&mdsc
->mutex
);
3090 while ((req
= __get_oldest_req(mdsc
))) {
3091 dout("wait_requests timed out on tid %llu\n",
3093 __unregister_request(mdsc
, req
);
3096 mutex_unlock(&mdsc
->mutex
);
3097 dout("wait_requests done\n");
3101 * called before mount is ro, and before dentries are torn down.
3102 * (hmm, does this still race with new lookups?)
3104 void ceph_mdsc_pre_umount(struct ceph_mds_client
*mdsc
)
3106 dout("pre_umount\n");
3110 ceph_flush_dirty_caps(mdsc
);
3111 wait_requests(mdsc
);
3114 * wait for reply handlers to drop their request refs and
3115 * their inode/dcache refs
3121 * wait for all write mds requests to flush.
3123 static void wait_unsafe_requests(struct ceph_mds_client
*mdsc
, u64 want_tid
)
3125 struct ceph_mds_request
*req
= NULL
, *nextreq
;
3128 mutex_lock(&mdsc
->mutex
);
3129 dout("wait_unsafe_requests want %lld\n", want_tid
);
3131 req
= __get_oldest_req(mdsc
);
3132 while (req
&& req
->r_tid
<= want_tid
) {
3133 /* find next request */
3134 n
= rb_next(&req
->r_node
);
3136 nextreq
= rb_entry(n
, struct ceph_mds_request
, r_node
);
3139 if ((req
->r_op
& CEPH_MDS_OP_WRITE
)) {
3141 ceph_mdsc_get_request(req
);
3143 ceph_mdsc_get_request(nextreq
);
3144 mutex_unlock(&mdsc
->mutex
);
3145 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3146 req
->r_tid
, want_tid
);
3147 wait_for_completion(&req
->r_safe_completion
);
3148 mutex_lock(&mdsc
->mutex
);
3149 ceph_mdsc_put_request(req
);
3151 break; /* next dne before, so we're done! */
3152 if (RB_EMPTY_NODE(&nextreq
->r_node
)) {
3153 /* next request was removed from tree */
3154 ceph_mdsc_put_request(nextreq
);
3157 ceph_mdsc_put_request(nextreq
); /* won't go away */
3161 mutex_unlock(&mdsc
->mutex
);
3162 dout("wait_unsafe_requests done\n");
3165 void ceph_mdsc_sync(struct ceph_mds_client
*mdsc
)
3167 u64 want_tid
, want_flush
;
3169 if (mdsc
->fsc
->mount_state
== CEPH_MOUNT_SHUTDOWN
)
3173 mutex_lock(&mdsc
->mutex
);
3174 want_tid
= mdsc
->last_tid
;
3175 want_flush
= mdsc
->cap_flush_seq
;
3176 mutex_unlock(&mdsc
->mutex
);
3177 dout("sync want tid %lld flush_seq %lld\n", want_tid
, want_flush
);
3179 ceph_flush_dirty_caps(mdsc
);
3181 wait_unsafe_requests(mdsc
, want_tid
);
3182 wait_event(mdsc
->cap_flushing_wq
, check_cap_flush(mdsc
, want_flush
));
3186 * true if all sessions are closed, or we force unmount
3188 static bool done_closing_sessions(struct ceph_mds_client
*mdsc
)
3192 if (mdsc
->fsc
->mount_state
== CEPH_MOUNT_SHUTDOWN
)
3195 mutex_lock(&mdsc
->mutex
);
3196 for (i
= 0; i
< mdsc
->max_sessions
; i
++)
3197 if (mdsc
->sessions
[i
])
3199 mutex_unlock(&mdsc
->mutex
);
3204 * called after sb is ro.
3206 void ceph_mdsc_close_sessions(struct ceph_mds_client
*mdsc
)
3208 struct ceph_mds_session
*session
;
3210 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3211 unsigned long timeout
= fsc
->client
->options
->mount_timeout
* HZ
;
3213 dout("close_sessions\n");
3215 /* close sessions */
3216 mutex_lock(&mdsc
->mutex
);
3217 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3218 session
= __ceph_lookup_mds_session(mdsc
, i
);
3221 mutex_unlock(&mdsc
->mutex
);
3222 mutex_lock(&session
->s_mutex
);
3223 __close_session(mdsc
, session
);
3224 mutex_unlock(&session
->s_mutex
);
3225 ceph_put_mds_session(session
);
3226 mutex_lock(&mdsc
->mutex
);
3228 mutex_unlock(&mdsc
->mutex
);
3230 dout("waiting for sessions to close\n");
3231 wait_event_timeout(mdsc
->session_close_wq
, done_closing_sessions(mdsc
),
3234 /* tear down remaining sessions */
3235 mutex_lock(&mdsc
->mutex
);
3236 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3237 if (mdsc
->sessions
[i
]) {
3238 session
= get_session(mdsc
->sessions
[i
]);
3239 __unregister_session(mdsc
, session
);
3240 mutex_unlock(&mdsc
->mutex
);
3241 mutex_lock(&session
->s_mutex
);
3242 remove_session_caps(session
);
3243 mutex_unlock(&session
->s_mutex
);
3244 ceph_put_mds_session(session
);
3245 mutex_lock(&mdsc
->mutex
);
3248 WARN_ON(!list_empty(&mdsc
->cap_delay_list
));
3249 mutex_unlock(&mdsc
->mutex
);
3251 ceph_cleanup_empty_realms(mdsc
);
3253 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3258 static void ceph_mdsc_stop(struct ceph_mds_client
*mdsc
)
3261 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3263 ceph_mdsmap_destroy(mdsc
->mdsmap
);
3264 kfree(mdsc
->sessions
);
3265 ceph_caps_finalize(mdsc
);
3268 void ceph_mdsc_destroy(struct ceph_fs_client
*fsc
)
3270 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
3272 dout("mdsc_destroy %p\n", mdsc
);
3273 ceph_mdsc_stop(mdsc
);
3275 /* flush out any connection work with references to us */
3280 dout("mdsc_destroy %p done\n", mdsc
);
3285 * handle mds map update.
3287 void ceph_mdsc_handle_map(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3291 void *p
= msg
->front
.iov_base
;
3292 void *end
= p
+ msg
->front
.iov_len
;
3293 struct ceph_mdsmap
*newmap
, *oldmap
;
3294 struct ceph_fsid fsid
;
3297 ceph_decode_need(&p
, end
, sizeof(fsid
)+2*sizeof(u32
), bad
);
3298 ceph_decode_copy(&p
, &fsid
, sizeof(fsid
));
3299 if (ceph_check_fsid(mdsc
->fsc
->client
, &fsid
) < 0)
3301 epoch
= ceph_decode_32(&p
);
3302 maplen
= ceph_decode_32(&p
);
3303 dout("handle_map epoch %u len %d\n", epoch
, (int)maplen
);
3305 /* do we need it? */
3306 ceph_monc_got_mdsmap(&mdsc
->fsc
->client
->monc
, epoch
);
3307 mutex_lock(&mdsc
->mutex
);
3308 if (mdsc
->mdsmap
&& epoch
<= mdsc
->mdsmap
->m_epoch
) {
3309 dout("handle_map epoch %u <= our %u\n",
3310 epoch
, mdsc
->mdsmap
->m_epoch
);
3311 mutex_unlock(&mdsc
->mutex
);
3315 newmap
= ceph_mdsmap_decode(&p
, end
);
3316 if (IS_ERR(newmap
)) {
3317 err
= PTR_ERR(newmap
);
3321 /* swap into place */
3323 oldmap
= mdsc
->mdsmap
;
3324 mdsc
->mdsmap
= newmap
;
3325 check_new_map(mdsc
, newmap
, oldmap
);
3326 ceph_mdsmap_destroy(oldmap
);
3328 mdsc
->mdsmap
= newmap
; /* first mds map */
3330 mdsc
->fsc
->sb
->s_maxbytes
= mdsc
->mdsmap
->m_max_file_size
;
3332 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3334 mutex_unlock(&mdsc
->mutex
);
3335 schedule_delayed(mdsc
);
3339 mutex_unlock(&mdsc
->mutex
);
3341 pr_err("error decoding mdsmap %d\n", err
);
3345 static struct ceph_connection
*con_get(struct ceph_connection
*con
)
3347 struct ceph_mds_session
*s
= con
->private;
3349 if (get_session(s
)) {
3350 dout("mdsc con_get %p ok (%d)\n", s
, atomic_read(&s
->s_ref
));
3353 dout("mdsc con_get %p FAIL\n", s
);
3357 static void con_put(struct ceph_connection
*con
)
3359 struct ceph_mds_session
*s
= con
->private;
3361 dout("mdsc con_put %p (%d)\n", s
, atomic_read(&s
->s_ref
) - 1);
3362 ceph_put_mds_session(s
);
3366 * if the client is unresponsive for long enough, the mds will kill
3367 * the session entirely.
3369 static void peer_reset(struct ceph_connection
*con
)
3371 struct ceph_mds_session
*s
= con
->private;
3372 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3374 pr_warning("mds%d closed our session\n", s
->s_mds
);
3375 send_mds_reconnect(mdsc
, s
);
3378 static void dispatch(struct ceph_connection
*con
, struct ceph_msg
*msg
)
3380 struct ceph_mds_session
*s
= con
->private;
3381 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3382 int type
= le16_to_cpu(msg
->hdr
.type
);
3384 mutex_lock(&mdsc
->mutex
);
3385 if (__verify_registered_session(mdsc
, s
) < 0) {
3386 mutex_unlock(&mdsc
->mutex
);
3389 mutex_unlock(&mdsc
->mutex
);
3392 case CEPH_MSG_MDS_MAP
:
3393 ceph_mdsc_handle_map(mdsc
, msg
);
3395 case CEPH_MSG_CLIENT_SESSION
:
3396 handle_session(s
, msg
);
3398 case CEPH_MSG_CLIENT_REPLY
:
3399 handle_reply(s
, msg
);
3401 case CEPH_MSG_CLIENT_REQUEST_FORWARD
:
3402 handle_forward(mdsc
, s
, msg
);
3404 case CEPH_MSG_CLIENT_CAPS
:
3405 ceph_handle_caps(s
, msg
);
3407 case CEPH_MSG_CLIENT_SNAP
:
3408 ceph_handle_snap(mdsc
, s
, msg
);
3410 case CEPH_MSG_CLIENT_LEASE
:
3411 handle_lease(mdsc
, s
, msg
);
3415 pr_err("received unknown message type %d %s\n", type
,
3416 ceph_msg_type_name(type
));
3427 * Note: returned pointer is the address of a structure that's
3428 * managed separately. Caller must *not* attempt to free it.
3430 static struct ceph_auth_handshake
*get_authorizer(struct ceph_connection
*con
,
3431 int *proto
, int force_new
)
3433 struct ceph_mds_session
*s
= con
->private;
3434 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3435 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3436 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
3438 if (force_new
&& auth
->authorizer
) {
3439 if (ac
->ops
&& ac
->ops
->destroy_authorizer
)
3440 ac
->ops
->destroy_authorizer(ac
, auth
->authorizer
);
3441 auth
->authorizer
= NULL
;
3443 if (!auth
->authorizer
&& ac
->ops
&& ac
->ops
->create_authorizer
) {
3444 int ret
= ac
->ops
->create_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3447 return ERR_PTR(ret
);
3449 *proto
= ac
->protocol
;
3455 static int verify_authorizer_reply(struct ceph_connection
*con
, int len
)
3457 struct ceph_mds_session
*s
= con
->private;
3458 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3459 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3461 return ac
->ops
->verify_authorizer_reply(ac
, s
->s_auth
.authorizer
, len
);
3464 static int invalidate_authorizer(struct ceph_connection
*con
)
3466 struct ceph_mds_session
*s
= con
->private;
3467 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3468 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3470 if (ac
->ops
->invalidate_authorizer
)
3471 ac
->ops
->invalidate_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
);
3473 return ceph_monc_validate_auth(&mdsc
->fsc
->client
->monc
);
3476 static const struct ceph_connection_operations mds_con_ops
= {
3479 .dispatch
= dispatch
,
3480 .get_authorizer
= get_authorizer
,
3481 .verify_authorizer_reply
= verify_authorizer_reply
,
3482 .invalidate_authorizer
= invalidate_authorizer
,
3483 .peer_reset
= peer_reset
,