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/messenger.h>
14 #include <linux/ceph/decode.h>
15 #include <linux/ceph/pagelist.h>
16 #include <linux/ceph/auth.h>
17 #include <linux/ceph/debugfs.h>
20 * A cluster of MDS (metadata server) daemons is responsible for
21 * managing the file system namespace (the directory hierarchy and
22 * inodes) and for coordinating shared access to storage. Metadata is
23 * partitioning hierarchically across a number of servers, and that
24 * partition varies over time as the cluster adjusts the distribution
25 * in order to balance load.
27 * The MDS client is primarily responsible to managing synchronous
28 * metadata requests for operations like open, unlink, and so forth.
29 * If there is a MDS failure, we find out about it when we (possibly
30 * request and) receive a new MDS map, and can resubmit affected
33 * For the most part, though, we take advantage of a lossless
34 * communications channel to the MDS, and do not need to worry about
35 * timing out or resubmitting requests.
37 * We maintain a stateful "session" with each MDS we interact with.
38 * Within each session, we sent periodic heartbeat messages to ensure
39 * any capabilities or leases we have been issues remain valid. If
40 * the session times out and goes stale, our leases and capabilities
41 * are no longer valid.
44 struct ceph_reconnect_state
{
45 struct ceph_pagelist
*pagelist
;
49 static void __wake_requests(struct ceph_mds_client
*mdsc
,
50 struct list_head
*head
);
52 static const struct ceph_connection_operations mds_con_ops
;
60 * parse individual inode info
62 static int parse_reply_info_in(void **p
, void *end
,
63 struct ceph_mds_reply_info_in
*info
,
69 *p
+= sizeof(struct ceph_mds_reply_inode
) +
70 sizeof(*info
->in
->fragtree
.splits
) *
71 le32_to_cpu(info
->in
->fragtree
.nsplits
);
73 ceph_decode_32_safe(p
, end
, info
->symlink_len
, bad
);
74 ceph_decode_need(p
, end
, info
->symlink_len
, bad
);
76 *p
+= info
->symlink_len
;
78 if (features
& CEPH_FEATURE_DIRLAYOUTHASH
)
79 ceph_decode_copy_safe(p
, end
, &info
->dir_layout
,
80 sizeof(info
->dir_layout
), bad
);
82 memset(&info
->dir_layout
, 0, sizeof(info
->dir_layout
));
84 ceph_decode_32_safe(p
, end
, info
->xattr_len
, bad
);
85 ceph_decode_need(p
, end
, info
->xattr_len
, bad
);
86 info
->xattr_data
= *p
;
87 *p
+= info
->xattr_len
;
94 * parse a normal reply, which may contain a (dir+)dentry and/or a
97 static int parse_reply_info_trace(void **p
, void *end
,
98 struct ceph_mds_reply_info_parsed
*info
,
103 if (info
->head
->is_dentry
) {
104 err
= parse_reply_info_in(p
, end
, &info
->diri
, features
);
108 if (unlikely(*p
+ sizeof(*info
->dirfrag
) > end
))
111 *p
+= sizeof(*info
->dirfrag
) +
112 sizeof(u32
)*le32_to_cpu(info
->dirfrag
->ndist
);
113 if (unlikely(*p
> end
))
116 ceph_decode_32_safe(p
, end
, info
->dname_len
, bad
);
117 ceph_decode_need(p
, end
, info
->dname_len
, bad
);
119 *p
+= info
->dname_len
;
121 *p
+= sizeof(*info
->dlease
);
124 if (info
->head
->is_target
) {
125 err
= parse_reply_info_in(p
, end
, &info
->targeti
, features
);
130 if (unlikely(*p
!= end
))
137 pr_err("problem parsing mds trace %d\n", err
);
142 * parse readdir results
144 static int parse_reply_info_dir(void **p
, void *end
,
145 struct ceph_mds_reply_info_parsed
*info
,
152 if (*p
+ sizeof(*info
->dir_dir
) > end
)
154 *p
+= sizeof(*info
->dir_dir
) +
155 sizeof(u32
)*le32_to_cpu(info
->dir_dir
->ndist
);
159 ceph_decode_need(p
, end
, sizeof(num
) + 2, bad
);
160 num
= ceph_decode_32(p
);
161 info
->dir_end
= ceph_decode_8(p
);
162 info
->dir_complete
= ceph_decode_8(p
);
166 /* alloc large array */
168 info
->dir_in
= kcalloc(num
, sizeof(*info
->dir_in
) +
169 sizeof(*info
->dir_dname
) +
170 sizeof(*info
->dir_dname_len
) +
171 sizeof(*info
->dir_dlease
),
173 if (info
->dir_in
== NULL
) {
177 info
->dir_dname
= (void *)(info
->dir_in
+ num
);
178 info
->dir_dname_len
= (void *)(info
->dir_dname
+ num
);
179 info
->dir_dlease
= (void *)(info
->dir_dname_len
+ num
);
183 ceph_decode_need(p
, end
, sizeof(u32
)*2, bad
);
184 info
->dir_dname_len
[i
] = ceph_decode_32(p
);
185 ceph_decode_need(p
, end
, info
->dir_dname_len
[i
], bad
);
186 info
->dir_dname
[i
] = *p
;
187 *p
+= info
->dir_dname_len
[i
];
188 dout("parsed dir dname '%.*s'\n", info
->dir_dname_len
[i
],
190 info
->dir_dlease
[i
] = *p
;
191 *p
+= sizeof(struct ceph_mds_reply_lease
);
194 err
= parse_reply_info_in(p
, end
, &info
->dir_in
[i
], features
);
209 pr_err("problem parsing dir contents %d\n", err
);
214 * parse fcntl F_GETLK results
216 static int parse_reply_info_filelock(void **p
, void *end
,
217 struct ceph_mds_reply_info_parsed
*info
,
220 if (*p
+ sizeof(*info
->filelock_reply
) > end
)
223 info
->filelock_reply
= *p
;
224 *p
+= sizeof(*info
->filelock_reply
);
226 if (unlikely(*p
!= end
))
235 * parse extra results
237 static int parse_reply_info_extra(void **p
, void *end
,
238 struct ceph_mds_reply_info_parsed
*info
,
241 if (info
->head
->op
== CEPH_MDS_OP_GETFILELOCK
)
242 return parse_reply_info_filelock(p
, end
, info
, features
);
244 return parse_reply_info_dir(p
, end
, info
, features
);
248 * parse entire mds reply
250 static int parse_reply_info(struct ceph_msg
*msg
,
251 struct ceph_mds_reply_info_parsed
*info
,
258 info
->head
= msg
->front
.iov_base
;
259 p
= msg
->front
.iov_base
+ sizeof(struct ceph_mds_reply_head
);
260 end
= p
+ msg
->front
.iov_len
- sizeof(struct ceph_mds_reply_head
);
263 ceph_decode_32_safe(&p
, end
, len
, bad
);
265 ceph_decode_need(&p
, end
, len
, bad
);
266 err
= parse_reply_info_trace(&p
, p
+len
, info
, features
);
272 ceph_decode_32_safe(&p
, end
, len
, bad
);
274 ceph_decode_need(&p
, end
, len
, bad
);
275 err
= parse_reply_info_extra(&p
, p
+len
, info
, features
);
281 ceph_decode_32_safe(&p
, end
, len
, bad
);
282 info
->snapblob_len
= len
;
293 pr_err("mds parse_reply err %d\n", err
);
297 static void destroy_reply_info(struct ceph_mds_reply_info_parsed
*info
)
306 static const char *session_state_name(int s
)
309 case CEPH_MDS_SESSION_NEW
: return "new";
310 case CEPH_MDS_SESSION_OPENING
: return "opening";
311 case CEPH_MDS_SESSION_OPEN
: return "open";
312 case CEPH_MDS_SESSION_HUNG
: return "hung";
313 case CEPH_MDS_SESSION_CLOSING
: return "closing";
314 case CEPH_MDS_SESSION_RESTARTING
: return "restarting";
315 case CEPH_MDS_SESSION_RECONNECTING
: return "reconnecting";
316 default: return "???";
320 static struct ceph_mds_session
*get_session(struct ceph_mds_session
*s
)
322 if (atomic_inc_not_zero(&s
->s_ref
)) {
323 dout("mdsc get_session %p %d -> %d\n", s
,
324 atomic_read(&s
->s_ref
)-1, atomic_read(&s
->s_ref
));
327 dout("mdsc get_session %p 0 -- FAIL", s
);
332 void ceph_put_mds_session(struct ceph_mds_session
*s
)
334 dout("mdsc put_session %p %d -> %d\n", s
,
335 atomic_read(&s
->s_ref
), atomic_read(&s
->s_ref
)-1);
336 if (atomic_dec_and_test(&s
->s_ref
)) {
337 if (s
->s_auth
.authorizer
)
338 s
->s_mdsc
->fsc
->client
->monc
.auth
->ops
->destroy_authorizer(
339 s
->s_mdsc
->fsc
->client
->monc
.auth
,
340 s
->s_auth
.authorizer
);
346 * called under mdsc->mutex
348 struct ceph_mds_session
*__ceph_lookup_mds_session(struct ceph_mds_client
*mdsc
,
351 struct ceph_mds_session
*session
;
353 if (mds
>= mdsc
->max_sessions
|| mdsc
->sessions
[mds
] == NULL
)
355 session
= mdsc
->sessions
[mds
];
356 dout("lookup_mds_session %p %d\n", session
,
357 atomic_read(&session
->s_ref
));
358 get_session(session
);
362 static bool __have_session(struct ceph_mds_client
*mdsc
, int mds
)
364 if (mds
>= mdsc
->max_sessions
)
366 return mdsc
->sessions
[mds
];
369 static int __verify_registered_session(struct ceph_mds_client
*mdsc
,
370 struct ceph_mds_session
*s
)
372 if (s
->s_mds
>= mdsc
->max_sessions
||
373 mdsc
->sessions
[s
->s_mds
] != s
)
379 * create+register a new session for given mds.
380 * called under mdsc->mutex.
382 static struct ceph_mds_session
*register_session(struct ceph_mds_client
*mdsc
,
385 struct ceph_mds_session
*s
;
387 s
= kzalloc(sizeof(*s
), GFP_NOFS
);
389 return ERR_PTR(-ENOMEM
);
392 s
->s_state
= CEPH_MDS_SESSION_NEW
;
395 mutex_init(&s
->s_mutex
);
397 ceph_con_init(&s
->s_con
, s
, &mds_con_ops
, &mdsc
->fsc
->client
->msgr
);
399 spin_lock_init(&s
->s_gen_ttl_lock
);
401 s
->s_cap_ttl
= jiffies
- 1;
403 spin_lock_init(&s
->s_cap_lock
);
404 s
->s_renew_requested
= 0;
406 INIT_LIST_HEAD(&s
->s_caps
);
409 atomic_set(&s
->s_ref
, 1);
410 INIT_LIST_HEAD(&s
->s_waiting
);
411 INIT_LIST_HEAD(&s
->s_unsafe
);
412 s
->s_num_cap_releases
= 0;
413 s
->s_cap_iterator
= NULL
;
414 INIT_LIST_HEAD(&s
->s_cap_releases
);
415 INIT_LIST_HEAD(&s
->s_cap_releases_done
);
416 INIT_LIST_HEAD(&s
->s_cap_flushing
);
417 INIT_LIST_HEAD(&s
->s_cap_snaps_flushing
);
419 dout("register_session mds%d\n", mds
);
420 if (mds
>= mdsc
->max_sessions
) {
421 int newmax
= 1 << get_count_order(mds
+1);
422 struct ceph_mds_session
**sa
;
424 dout("register_session realloc to %d\n", newmax
);
425 sa
= kcalloc(newmax
, sizeof(void *), GFP_NOFS
);
428 if (mdsc
->sessions
) {
429 memcpy(sa
, mdsc
->sessions
,
430 mdsc
->max_sessions
* sizeof(void *));
431 kfree(mdsc
->sessions
);
434 mdsc
->max_sessions
= newmax
;
436 mdsc
->sessions
[mds
] = s
;
437 atomic_inc(&s
->s_ref
); /* one ref to sessions[], one to caller */
439 ceph_con_open(&s
->s_con
, CEPH_ENTITY_TYPE_MDS
, mds
,
440 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
446 return ERR_PTR(-ENOMEM
);
450 * called under mdsc->mutex
452 static void __unregister_session(struct ceph_mds_client
*mdsc
,
453 struct ceph_mds_session
*s
)
455 dout("__unregister_session mds%d %p\n", s
->s_mds
, s
);
456 BUG_ON(mdsc
->sessions
[s
->s_mds
] != s
);
457 mdsc
->sessions
[s
->s_mds
] = NULL
;
458 ceph_con_close(&s
->s_con
);
459 ceph_put_mds_session(s
);
463 * drop session refs in request.
465 * should be last request ref, or hold mdsc->mutex
467 static void put_request_session(struct ceph_mds_request
*req
)
469 if (req
->r_session
) {
470 ceph_put_mds_session(req
->r_session
);
471 req
->r_session
= NULL
;
475 void ceph_mdsc_release_request(struct kref
*kref
)
477 struct ceph_mds_request
*req
= container_of(kref
,
478 struct ceph_mds_request
,
481 ceph_msg_put(req
->r_request
);
483 ceph_msg_put(req
->r_reply
);
484 destroy_reply_info(&req
->r_reply_info
);
487 ceph_put_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
490 if (req
->r_locked_dir
)
491 ceph_put_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
492 if (req
->r_target_inode
)
493 iput(req
->r_target_inode
);
496 if (req
->r_old_dentry
) {
498 * track (and drop pins for) r_old_dentry_dir
499 * separately, since r_old_dentry's d_parent may have
500 * changed between the dir mutex being dropped and
501 * this request being freed.
503 ceph_put_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
505 dput(req
->r_old_dentry
);
506 iput(req
->r_old_dentry_dir
);
510 put_request_session(req
);
511 ceph_unreserve_caps(req
->r_mdsc
, &req
->r_caps_reservation
);
516 * lookup session, bump ref if found.
518 * called under mdsc->mutex.
520 static struct ceph_mds_request
*__lookup_request(struct ceph_mds_client
*mdsc
,
523 struct ceph_mds_request
*req
;
524 struct rb_node
*n
= mdsc
->request_tree
.rb_node
;
527 req
= rb_entry(n
, struct ceph_mds_request
, r_node
);
528 if (tid
< req
->r_tid
)
530 else if (tid
> req
->r_tid
)
533 ceph_mdsc_get_request(req
);
540 static void __insert_request(struct ceph_mds_client
*mdsc
,
541 struct ceph_mds_request
*new)
543 struct rb_node
**p
= &mdsc
->request_tree
.rb_node
;
544 struct rb_node
*parent
= NULL
;
545 struct ceph_mds_request
*req
= NULL
;
549 req
= rb_entry(parent
, struct ceph_mds_request
, r_node
);
550 if (new->r_tid
< req
->r_tid
)
552 else if (new->r_tid
> req
->r_tid
)
558 rb_link_node(&new->r_node
, parent
, p
);
559 rb_insert_color(&new->r_node
, &mdsc
->request_tree
);
563 * Register an in-flight request, and assign a tid. Link to directory
564 * are modifying (if any).
566 * Called under mdsc->mutex.
568 static void __register_request(struct ceph_mds_client
*mdsc
,
569 struct ceph_mds_request
*req
,
572 req
->r_tid
= ++mdsc
->last_tid
;
574 ceph_reserve_caps(mdsc
, &req
->r_caps_reservation
,
576 dout("__register_request %p tid %lld\n", req
, req
->r_tid
);
577 ceph_mdsc_get_request(req
);
578 __insert_request(mdsc
, req
);
580 req
->r_uid
= current_fsuid();
581 req
->r_gid
= current_fsgid();
584 struct ceph_inode_info
*ci
= ceph_inode(dir
);
587 spin_lock(&ci
->i_unsafe_lock
);
588 req
->r_unsafe_dir
= dir
;
589 list_add_tail(&req
->r_unsafe_dir_item
, &ci
->i_unsafe_dirops
);
590 spin_unlock(&ci
->i_unsafe_lock
);
594 static void __unregister_request(struct ceph_mds_client
*mdsc
,
595 struct ceph_mds_request
*req
)
597 dout("__unregister_request %p tid %lld\n", req
, req
->r_tid
);
598 rb_erase(&req
->r_node
, &mdsc
->request_tree
);
599 RB_CLEAR_NODE(&req
->r_node
);
601 if (req
->r_unsafe_dir
) {
602 struct ceph_inode_info
*ci
= ceph_inode(req
->r_unsafe_dir
);
604 spin_lock(&ci
->i_unsafe_lock
);
605 list_del_init(&req
->r_unsafe_dir_item
);
606 spin_unlock(&ci
->i_unsafe_lock
);
608 iput(req
->r_unsafe_dir
);
609 req
->r_unsafe_dir
= NULL
;
612 ceph_mdsc_put_request(req
);
616 * Choose mds to send request to next. If there is a hint set in the
617 * request (e.g., due to a prior forward hint from the mds), use that.
618 * Otherwise, consult frag tree and/or caps to identify the
619 * appropriate mds. If all else fails, choose randomly.
621 * Called under mdsc->mutex.
623 static struct dentry
*get_nonsnap_parent(struct dentry
*dentry
)
626 * we don't need to worry about protecting the d_parent access
627 * here because we never renaming inside the snapped namespace
628 * except to resplice to another snapdir, and either the old or new
629 * result is a valid result.
631 while (!IS_ROOT(dentry
) && ceph_snap(dentry
->d_inode
) != CEPH_NOSNAP
)
632 dentry
= dentry
->d_parent
;
636 static int __choose_mds(struct ceph_mds_client
*mdsc
,
637 struct ceph_mds_request
*req
)
640 struct ceph_inode_info
*ci
;
641 struct ceph_cap
*cap
;
642 int mode
= req
->r_direct_mode
;
644 u32 hash
= req
->r_direct_hash
;
645 bool is_hash
= req
->r_direct_is_hash
;
648 * is there a specific mds we should try? ignore hint if we have
649 * no session and the mds is not up (active or recovering).
651 if (req
->r_resend_mds
>= 0 &&
652 (__have_session(mdsc
, req
->r_resend_mds
) ||
653 ceph_mdsmap_get_state(mdsc
->mdsmap
, req
->r_resend_mds
) > 0)) {
654 dout("choose_mds using resend_mds mds%d\n",
656 return req
->r_resend_mds
;
659 if (mode
== USE_RANDOM_MDS
)
664 inode
= req
->r_inode
;
665 } else if (req
->r_dentry
) {
666 /* ignore race with rename; old or new d_parent is okay */
667 struct dentry
*parent
= req
->r_dentry
->d_parent
;
668 struct inode
*dir
= parent
->d_inode
;
670 if (dir
->i_sb
!= mdsc
->fsc
->sb
) {
672 inode
= req
->r_dentry
->d_inode
;
673 } else if (ceph_snap(dir
) != CEPH_NOSNAP
) {
674 /* direct snapped/virtual snapdir requests
675 * based on parent dir inode */
676 struct dentry
*dn
= get_nonsnap_parent(parent
);
678 dout("__choose_mds using nonsnap parent %p\n", inode
);
679 } else if (req
->r_dentry
->d_inode
) {
681 inode
= req
->r_dentry
->d_inode
;
685 hash
= ceph_dentry_hash(dir
, req
->r_dentry
);
690 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode
, (int)is_hash
,
694 ci
= ceph_inode(inode
);
696 if (is_hash
&& S_ISDIR(inode
->i_mode
)) {
697 struct ceph_inode_frag frag
;
700 ceph_choose_frag(ci
, hash
, &frag
, &found
);
702 if (mode
== USE_ANY_MDS
&& frag
.ndist
> 0) {
705 /* choose a random replica */
706 get_random_bytes(&r
, 1);
709 dout("choose_mds %p %llx.%llx "
710 "frag %u mds%d (%d/%d)\n",
711 inode
, ceph_vinop(inode
),
714 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
715 CEPH_MDS_STATE_ACTIVE
)
719 /* since this file/dir wasn't known to be
720 * replicated, then we want to look for the
721 * authoritative mds. */
724 /* choose auth mds */
726 dout("choose_mds %p %llx.%llx "
727 "frag %u mds%d (auth)\n",
728 inode
, ceph_vinop(inode
), frag
.frag
, mds
);
729 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
730 CEPH_MDS_STATE_ACTIVE
)
736 spin_lock(&ci
->i_ceph_lock
);
738 if (mode
== USE_AUTH_MDS
)
739 cap
= ci
->i_auth_cap
;
740 if (!cap
&& !RB_EMPTY_ROOT(&ci
->i_caps
))
741 cap
= rb_entry(rb_first(&ci
->i_caps
), struct ceph_cap
, ci_node
);
743 spin_unlock(&ci
->i_ceph_lock
);
746 mds
= cap
->session
->s_mds
;
747 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
748 inode
, ceph_vinop(inode
), mds
,
749 cap
== ci
->i_auth_cap
? "auth " : "", cap
);
750 spin_unlock(&ci
->i_ceph_lock
);
754 mds
= ceph_mdsmap_get_random_mds(mdsc
->mdsmap
);
755 dout("choose_mds chose random mds%d\n", mds
);
763 static struct ceph_msg
*create_session_msg(u32 op
, u64 seq
)
765 struct ceph_msg
*msg
;
766 struct ceph_mds_session_head
*h
;
768 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
), GFP_NOFS
,
771 pr_err("create_session_msg ENOMEM creating msg\n");
774 h
= msg
->front
.iov_base
;
775 h
->op
= cpu_to_le32(op
);
776 h
->seq
= cpu_to_le64(seq
);
781 * send session open request.
783 * called under mdsc->mutex
785 static int __open_session(struct ceph_mds_client
*mdsc
,
786 struct ceph_mds_session
*session
)
788 struct ceph_msg
*msg
;
790 int mds
= session
->s_mds
;
792 /* wait for mds to go active? */
793 mstate
= ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
);
794 dout("open_session to mds%d (%s)\n", mds
,
795 ceph_mds_state_name(mstate
));
796 session
->s_state
= CEPH_MDS_SESSION_OPENING
;
797 session
->s_renew_requested
= jiffies
;
799 /* send connect message */
800 msg
= create_session_msg(CEPH_SESSION_REQUEST_OPEN
, session
->s_seq
);
803 ceph_con_send(&session
->s_con
, msg
);
808 * open sessions for any export targets for the given mds
810 * called under mdsc->mutex
812 static void __open_export_target_sessions(struct ceph_mds_client
*mdsc
,
813 struct ceph_mds_session
*session
)
815 struct ceph_mds_info
*mi
;
816 struct ceph_mds_session
*ts
;
817 int i
, mds
= session
->s_mds
;
820 if (mds
>= mdsc
->mdsmap
->m_max_mds
)
822 mi
= &mdsc
->mdsmap
->m_info
[mds
];
823 dout("open_export_target_sessions for mds%d (%d targets)\n",
824 session
->s_mds
, mi
->num_export_targets
);
826 for (i
= 0; i
< mi
->num_export_targets
; i
++) {
827 target
= mi
->export_targets
[i
];
828 ts
= __ceph_lookup_mds_session(mdsc
, target
);
830 ts
= register_session(mdsc
, target
);
834 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
835 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
836 __open_session(mdsc
, session
);
838 dout(" mds%d target mds%d %p is %s\n", session
->s_mds
,
839 i
, ts
, session_state_name(ts
->s_state
));
840 ceph_put_mds_session(ts
);
844 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client
*mdsc
,
845 struct ceph_mds_session
*session
)
847 mutex_lock(&mdsc
->mutex
);
848 __open_export_target_sessions(mdsc
, session
);
849 mutex_unlock(&mdsc
->mutex
);
857 * Free preallocated cap messages assigned to this session
859 static void cleanup_cap_releases(struct ceph_mds_session
*session
)
861 struct ceph_msg
*msg
;
863 spin_lock(&session
->s_cap_lock
);
864 while (!list_empty(&session
->s_cap_releases
)) {
865 msg
= list_first_entry(&session
->s_cap_releases
,
866 struct ceph_msg
, list_head
);
867 list_del_init(&msg
->list_head
);
870 while (!list_empty(&session
->s_cap_releases_done
)) {
871 msg
= list_first_entry(&session
->s_cap_releases_done
,
872 struct ceph_msg
, list_head
);
873 list_del_init(&msg
->list_head
);
876 spin_unlock(&session
->s_cap_lock
);
880 * Helper to safely iterate over all caps associated with a session, with
881 * special care taken to handle a racing __ceph_remove_cap().
883 * Caller must hold session s_mutex.
885 static int iterate_session_caps(struct ceph_mds_session
*session
,
886 int (*cb
)(struct inode
*, struct ceph_cap
*,
890 struct ceph_cap
*cap
;
891 struct inode
*inode
, *last_inode
= NULL
;
892 struct ceph_cap
*old_cap
= NULL
;
895 dout("iterate_session_caps %p mds%d\n", session
, session
->s_mds
);
896 spin_lock(&session
->s_cap_lock
);
897 p
= session
->s_caps
.next
;
898 while (p
!= &session
->s_caps
) {
899 cap
= list_entry(p
, struct ceph_cap
, session_caps
);
900 inode
= igrab(&cap
->ci
->vfs_inode
);
905 session
->s_cap_iterator
= cap
;
906 spin_unlock(&session
->s_cap_lock
);
913 ceph_put_cap(session
->s_mdsc
, old_cap
);
917 ret
= cb(inode
, cap
, arg
);
920 spin_lock(&session
->s_cap_lock
);
922 if (cap
->ci
== NULL
) {
923 dout("iterate_session_caps finishing cap %p removal\n",
925 BUG_ON(cap
->session
!= session
);
926 list_del_init(&cap
->session_caps
);
927 session
->s_nr_caps
--;
929 old_cap
= cap
; /* put_cap it w/o locks held */
936 session
->s_cap_iterator
= NULL
;
937 spin_unlock(&session
->s_cap_lock
);
942 ceph_put_cap(session
->s_mdsc
, old_cap
);
947 static int remove_session_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
950 struct ceph_inode_info
*ci
= ceph_inode(inode
);
953 dout("removing cap %p, ci is %p, inode is %p\n",
954 cap
, ci
, &ci
->vfs_inode
);
955 spin_lock(&ci
->i_ceph_lock
);
956 __ceph_remove_cap(cap
);
957 if (!__ceph_is_any_real_caps(ci
)) {
958 struct ceph_mds_client
*mdsc
=
959 ceph_sb_to_client(inode
->i_sb
)->mdsc
;
961 spin_lock(&mdsc
->cap_dirty_lock
);
962 if (!list_empty(&ci
->i_dirty_item
)) {
963 pr_info(" dropping dirty %s state for %p %lld\n",
964 ceph_cap_string(ci
->i_dirty_caps
),
965 inode
, ceph_ino(inode
));
966 ci
->i_dirty_caps
= 0;
967 list_del_init(&ci
->i_dirty_item
);
970 if (!list_empty(&ci
->i_flushing_item
)) {
971 pr_info(" dropping dirty+flushing %s state for %p %lld\n",
972 ceph_cap_string(ci
->i_flushing_caps
),
973 inode
, ceph_ino(inode
));
974 ci
->i_flushing_caps
= 0;
975 list_del_init(&ci
->i_flushing_item
);
976 mdsc
->num_cap_flushing
--;
979 if (drop
&& ci
->i_wrbuffer_ref
) {
980 pr_info(" dropping dirty data for %p %lld\n",
981 inode
, ceph_ino(inode
));
982 ci
->i_wrbuffer_ref
= 0;
983 ci
->i_wrbuffer_ref_head
= 0;
986 spin_unlock(&mdsc
->cap_dirty_lock
);
988 spin_unlock(&ci
->i_ceph_lock
);
995 * caller must hold session s_mutex
997 static void remove_session_caps(struct ceph_mds_session
*session
)
999 dout("remove_session_caps on %p\n", session
);
1000 iterate_session_caps(session
, remove_session_caps_cb
, NULL
);
1001 BUG_ON(session
->s_nr_caps
> 0);
1002 BUG_ON(!list_empty(&session
->s_cap_flushing
));
1003 cleanup_cap_releases(session
);
1007 * wake up any threads waiting on this session's caps. if the cap is
1008 * old (didn't get renewed on the client reconnect), remove it now.
1010 * caller must hold s_mutex.
1012 static int wake_up_session_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1015 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1017 wake_up_all(&ci
->i_cap_wq
);
1019 spin_lock(&ci
->i_ceph_lock
);
1020 ci
->i_wanted_max_size
= 0;
1021 ci
->i_requested_max_size
= 0;
1022 spin_unlock(&ci
->i_ceph_lock
);
1027 static void wake_up_session_caps(struct ceph_mds_session
*session
,
1030 dout("wake_up_session_caps %p mds%d\n", session
, session
->s_mds
);
1031 iterate_session_caps(session
, wake_up_session_cb
,
1032 (void *)(unsigned long)reconnect
);
1036 * Send periodic message to MDS renewing all currently held caps. The
1037 * ack will reset the expiration for all caps from this session.
1039 * caller holds s_mutex
1041 static int send_renew_caps(struct ceph_mds_client
*mdsc
,
1042 struct ceph_mds_session
*session
)
1044 struct ceph_msg
*msg
;
1047 if (time_after_eq(jiffies
, session
->s_cap_ttl
) &&
1048 time_after_eq(session
->s_cap_ttl
, session
->s_renew_requested
))
1049 pr_info("mds%d caps stale\n", session
->s_mds
);
1050 session
->s_renew_requested
= jiffies
;
1052 /* do not try to renew caps until a recovering mds has reconnected
1053 * with its clients. */
1054 state
= ceph_mdsmap_get_state(mdsc
->mdsmap
, session
->s_mds
);
1055 if (state
< CEPH_MDS_STATE_RECONNECT
) {
1056 dout("send_renew_caps ignoring mds%d (%s)\n",
1057 session
->s_mds
, ceph_mds_state_name(state
));
1061 dout("send_renew_caps to mds%d (%s)\n", session
->s_mds
,
1062 ceph_mds_state_name(state
));
1063 msg
= create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS
,
1064 ++session
->s_renew_seq
);
1067 ceph_con_send(&session
->s_con
, msg
);
1072 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1074 * Called under session->s_mutex
1076 static void renewed_caps(struct ceph_mds_client
*mdsc
,
1077 struct ceph_mds_session
*session
, int is_renew
)
1082 spin_lock(&session
->s_cap_lock
);
1083 was_stale
= is_renew
&& time_after_eq(jiffies
, session
->s_cap_ttl
);
1085 session
->s_cap_ttl
= session
->s_renew_requested
+
1086 mdsc
->mdsmap
->m_session_timeout
*HZ
;
1089 if (time_before(jiffies
, session
->s_cap_ttl
)) {
1090 pr_info("mds%d caps renewed\n", session
->s_mds
);
1093 pr_info("mds%d caps still stale\n", session
->s_mds
);
1096 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1097 session
->s_mds
, session
->s_cap_ttl
, was_stale
? "stale" : "fresh",
1098 time_before(jiffies
, session
->s_cap_ttl
) ? "stale" : "fresh");
1099 spin_unlock(&session
->s_cap_lock
);
1102 wake_up_session_caps(session
, 0);
1106 * send a session close request
1108 static int request_close_session(struct ceph_mds_client
*mdsc
,
1109 struct ceph_mds_session
*session
)
1111 struct ceph_msg
*msg
;
1113 dout("request_close_session mds%d state %s seq %lld\n",
1114 session
->s_mds
, session_state_name(session
->s_state
),
1116 msg
= create_session_msg(CEPH_SESSION_REQUEST_CLOSE
, session
->s_seq
);
1119 ceph_con_send(&session
->s_con
, msg
);
1124 * Called with s_mutex held.
1126 static int __close_session(struct ceph_mds_client
*mdsc
,
1127 struct ceph_mds_session
*session
)
1129 if (session
->s_state
>= CEPH_MDS_SESSION_CLOSING
)
1131 session
->s_state
= CEPH_MDS_SESSION_CLOSING
;
1132 return request_close_session(mdsc
, session
);
1136 * Trim old(er) caps.
1138 * Because we can't cache an inode without one or more caps, we do
1139 * this indirectly: if a cap is unused, we prune its aliases, at which
1140 * point the inode will hopefully get dropped to.
1142 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1143 * memory pressure from the MDS, though, so it needn't be perfect.
1145 static int trim_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
, void *arg
)
1147 struct ceph_mds_session
*session
= arg
;
1148 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1149 int used
, oissued
, mine
;
1151 if (session
->s_trim_caps
<= 0)
1154 spin_lock(&ci
->i_ceph_lock
);
1155 mine
= cap
->issued
| cap
->implemented
;
1156 used
= __ceph_caps_used(ci
);
1157 oissued
= __ceph_caps_issued_other(ci
, cap
);
1159 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1160 inode
, cap
, ceph_cap_string(mine
), ceph_cap_string(oissued
),
1161 ceph_cap_string(used
));
1162 if (ci
->i_dirty_caps
)
1163 goto out
; /* dirty caps */
1164 if ((used
& ~oissued
) & mine
)
1165 goto out
; /* we need these caps */
1167 session
->s_trim_caps
--;
1169 /* we aren't the only cap.. just remove us */
1170 __ceph_remove_cap(cap
);
1172 /* try to drop referring dentries */
1173 spin_unlock(&ci
->i_ceph_lock
);
1174 d_prune_aliases(inode
);
1175 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1176 inode
, cap
, atomic_read(&inode
->i_count
));
1181 spin_unlock(&ci
->i_ceph_lock
);
1186 * Trim session cap count down to some max number.
1188 static int trim_caps(struct ceph_mds_client
*mdsc
,
1189 struct ceph_mds_session
*session
,
1192 int trim_caps
= session
->s_nr_caps
- max_caps
;
1194 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1195 session
->s_mds
, session
->s_nr_caps
, max_caps
, trim_caps
);
1196 if (trim_caps
> 0) {
1197 session
->s_trim_caps
= trim_caps
;
1198 iterate_session_caps(session
, trim_caps_cb
, session
);
1199 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1200 session
->s_mds
, session
->s_nr_caps
, max_caps
,
1201 trim_caps
- session
->s_trim_caps
);
1202 session
->s_trim_caps
= 0;
1208 * Allocate cap_release messages. If there is a partially full message
1209 * in the queue, try to allocate enough to cover it's remainder, so that
1210 * we can send it immediately.
1212 * Called under s_mutex.
1214 int ceph_add_cap_releases(struct ceph_mds_client
*mdsc
,
1215 struct ceph_mds_session
*session
)
1217 struct ceph_msg
*msg
, *partial
= NULL
;
1218 struct ceph_mds_cap_release
*head
;
1220 int extra
= mdsc
->fsc
->mount_options
->cap_release_safety
;
1223 dout("add_cap_releases %p mds%d extra %d\n", session
, session
->s_mds
,
1226 spin_lock(&session
->s_cap_lock
);
1228 if (!list_empty(&session
->s_cap_releases
)) {
1229 msg
= list_first_entry(&session
->s_cap_releases
,
1232 head
= msg
->front
.iov_base
;
1233 num
= le32_to_cpu(head
->num
);
1235 dout(" partial %p with (%d/%d)\n", msg
, num
,
1236 (int)CEPH_CAPS_PER_RELEASE
);
1237 extra
+= CEPH_CAPS_PER_RELEASE
- num
;
1241 while (session
->s_num_cap_releases
< session
->s_nr_caps
+ extra
) {
1242 spin_unlock(&session
->s_cap_lock
);
1243 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE
, PAGE_CACHE_SIZE
,
1247 dout("add_cap_releases %p msg %p now %d\n", session
, msg
,
1248 (int)msg
->front
.iov_len
);
1249 head
= msg
->front
.iov_base
;
1250 head
->num
= cpu_to_le32(0);
1251 msg
->front
.iov_len
= sizeof(*head
);
1252 spin_lock(&session
->s_cap_lock
);
1253 list_add(&msg
->list_head
, &session
->s_cap_releases
);
1254 session
->s_num_cap_releases
+= CEPH_CAPS_PER_RELEASE
;
1258 head
= partial
->front
.iov_base
;
1259 num
= le32_to_cpu(head
->num
);
1260 dout(" queueing partial %p with %d/%d\n", partial
, num
,
1261 (int)CEPH_CAPS_PER_RELEASE
);
1262 list_move_tail(&partial
->list_head
,
1263 &session
->s_cap_releases_done
);
1264 session
->s_num_cap_releases
-= CEPH_CAPS_PER_RELEASE
- num
;
1267 spin_unlock(&session
->s_cap_lock
);
1273 * flush all dirty inode data to disk.
1275 * returns true if we've flushed through want_flush_seq
1277 static int check_cap_flush(struct ceph_mds_client
*mdsc
, u64 want_flush_seq
)
1281 dout("check_cap_flush want %lld\n", want_flush_seq
);
1282 mutex_lock(&mdsc
->mutex
);
1283 for (mds
= 0; ret
&& mds
< mdsc
->max_sessions
; mds
++) {
1284 struct ceph_mds_session
*session
= mdsc
->sessions
[mds
];
1288 get_session(session
);
1289 mutex_unlock(&mdsc
->mutex
);
1291 mutex_lock(&session
->s_mutex
);
1292 if (!list_empty(&session
->s_cap_flushing
)) {
1293 struct ceph_inode_info
*ci
=
1294 list_entry(session
->s_cap_flushing
.next
,
1295 struct ceph_inode_info
,
1297 struct inode
*inode
= &ci
->vfs_inode
;
1299 spin_lock(&ci
->i_ceph_lock
);
1300 if (ci
->i_cap_flush_seq
<= want_flush_seq
) {
1301 dout("check_cap_flush still flushing %p "
1302 "seq %lld <= %lld to mds%d\n", inode
,
1303 ci
->i_cap_flush_seq
, want_flush_seq
,
1307 spin_unlock(&ci
->i_ceph_lock
);
1309 mutex_unlock(&session
->s_mutex
);
1310 ceph_put_mds_session(session
);
1314 mutex_lock(&mdsc
->mutex
);
1317 mutex_unlock(&mdsc
->mutex
);
1318 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq
);
1323 * called under s_mutex
1325 void ceph_send_cap_releases(struct ceph_mds_client
*mdsc
,
1326 struct ceph_mds_session
*session
)
1328 struct ceph_msg
*msg
;
1330 dout("send_cap_releases mds%d\n", session
->s_mds
);
1331 spin_lock(&session
->s_cap_lock
);
1332 while (!list_empty(&session
->s_cap_releases_done
)) {
1333 msg
= list_first_entry(&session
->s_cap_releases_done
,
1334 struct ceph_msg
, list_head
);
1335 list_del_init(&msg
->list_head
);
1336 spin_unlock(&session
->s_cap_lock
);
1337 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1338 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1339 ceph_con_send(&session
->s_con
, msg
);
1340 spin_lock(&session
->s_cap_lock
);
1342 spin_unlock(&session
->s_cap_lock
);
1345 static void discard_cap_releases(struct ceph_mds_client
*mdsc
,
1346 struct ceph_mds_session
*session
)
1348 struct ceph_msg
*msg
;
1349 struct ceph_mds_cap_release
*head
;
1352 dout("discard_cap_releases mds%d\n", session
->s_mds
);
1353 spin_lock(&session
->s_cap_lock
);
1355 /* zero out the in-progress message */
1356 msg
= list_first_entry(&session
->s_cap_releases
,
1357 struct ceph_msg
, list_head
);
1358 head
= msg
->front
.iov_base
;
1359 num
= le32_to_cpu(head
->num
);
1360 dout("discard_cap_releases mds%d %p %u\n", session
->s_mds
, msg
, num
);
1361 head
->num
= cpu_to_le32(0);
1362 session
->s_num_cap_releases
+= num
;
1364 /* requeue completed messages */
1365 while (!list_empty(&session
->s_cap_releases_done
)) {
1366 msg
= list_first_entry(&session
->s_cap_releases_done
,
1367 struct ceph_msg
, list_head
);
1368 list_del_init(&msg
->list_head
);
1370 head
= msg
->front
.iov_base
;
1371 num
= le32_to_cpu(head
->num
);
1372 dout("discard_cap_releases mds%d %p %u\n", session
->s_mds
, msg
,
1374 session
->s_num_cap_releases
+= num
;
1375 head
->num
= cpu_to_le32(0);
1376 msg
->front
.iov_len
= sizeof(*head
);
1377 list_add(&msg
->list_head
, &session
->s_cap_releases
);
1380 spin_unlock(&session
->s_cap_lock
);
1388 * Create an mds request.
1390 struct ceph_mds_request
*
1391 ceph_mdsc_create_request(struct ceph_mds_client
*mdsc
, int op
, int mode
)
1393 struct ceph_mds_request
*req
= kzalloc(sizeof(*req
), GFP_NOFS
);
1396 return ERR_PTR(-ENOMEM
);
1398 mutex_init(&req
->r_fill_mutex
);
1400 req
->r_started
= jiffies
;
1401 req
->r_resend_mds
= -1;
1402 INIT_LIST_HEAD(&req
->r_unsafe_dir_item
);
1404 kref_init(&req
->r_kref
);
1405 INIT_LIST_HEAD(&req
->r_wait
);
1406 init_completion(&req
->r_completion
);
1407 init_completion(&req
->r_safe_completion
);
1408 INIT_LIST_HEAD(&req
->r_unsafe_item
);
1411 req
->r_direct_mode
= mode
;
1416 * return oldest (lowest) request, tid in request tree, 0 if none.
1418 * called under mdsc->mutex.
1420 static struct ceph_mds_request
*__get_oldest_req(struct ceph_mds_client
*mdsc
)
1422 if (RB_EMPTY_ROOT(&mdsc
->request_tree
))
1424 return rb_entry(rb_first(&mdsc
->request_tree
),
1425 struct ceph_mds_request
, r_node
);
1428 static u64
__get_oldest_tid(struct ceph_mds_client
*mdsc
)
1430 struct ceph_mds_request
*req
= __get_oldest_req(mdsc
);
1438 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1439 * on build_path_from_dentry in fs/cifs/dir.c.
1441 * If @stop_on_nosnap, generate path relative to the first non-snapped
1444 * Encode hidden .snap dirs as a double /, i.e.
1445 * foo/.snap/bar -> foo//bar
1447 char *ceph_mdsc_build_path(struct dentry
*dentry
, int *plen
, u64
*base
,
1450 struct dentry
*temp
;
1456 return ERR_PTR(-EINVAL
);
1460 seq
= read_seqbegin(&rename_lock
);
1462 for (temp
= dentry
; !IS_ROOT(temp
);) {
1463 struct inode
*inode
= temp
->d_inode
;
1464 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
)
1465 len
++; /* slash only */
1466 else if (stop_on_nosnap
&& inode
&&
1467 ceph_snap(inode
) == CEPH_NOSNAP
)
1470 len
+= 1 + temp
->d_name
.len
;
1471 temp
= temp
->d_parent
;
1474 pr_err("build_path corrupt dentry %p\n", dentry
);
1475 return ERR_PTR(-EINVAL
);
1480 len
--; /* no leading '/' */
1482 path
= kmalloc(len
+1, GFP_NOFS
);
1484 return ERR_PTR(-ENOMEM
);
1486 path
[pos
] = 0; /* trailing null */
1488 for (temp
= dentry
; !IS_ROOT(temp
) && pos
!= 0; ) {
1489 struct inode
*inode
;
1491 spin_lock(&temp
->d_lock
);
1492 inode
= temp
->d_inode
;
1493 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
) {
1494 dout("build_path path+%d: %p SNAPDIR\n",
1496 } else if (stop_on_nosnap
&& inode
&&
1497 ceph_snap(inode
) == CEPH_NOSNAP
) {
1498 spin_unlock(&temp
->d_lock
);
1501 pos
-= temp
->d_name
.len
;
1503 spin_unlock(&temp
->d_lock
);
1506 strncpy(path
+ pos
, temp
->d_name
.name
,
1509 spin_unlock(&temp
->d_lock
);
1512 temp
= temp
->d_parent
;
1515 pr_err("build_path corrupt dentry\n");
1517 return ERR_PTR(-EINVAL
);
1521 if (pos
!= 0 || read_seqretry(&rename_lock
, seq
)) {
1522 pr_err("build_path did not end path lookup where "
1523 "expected, namelen is %d, pos is %d\n", len
, pos
);
1524 /* presumably this is only possible if racing with a
1525 rename of one of the parent directories (we can not
1526 lock the dentries above us to prevent this, but
1527 retrying should be harmless) */
1532 *base
= ceph_ino(temp
->d_inode
);
1534 dout("build_path on %p %d built %llx '%.*s'\n",
1535 dentry
, dentry
->d_count
, *base
, len
, path
);
1539 static int build_dentry_path(struct dentry
*dentry
,
1540 const char **ppath
, int *ppathlen
, u64
*pino
,
1545 if (ceph_snap(dentry
->d_parent
->d_inode
) == CEPH_NOSNAP
) {
1546 *pino
= ceph_ino(dentry
->d_parent
->d_inode
);
1547 *ppath
= dentry
->d_name
.name
;
1548 *ppathlen
= dentry
->d_name
.len
;
1551 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1553 return PTR_ERR(path
);
1559 static int build_inode_path(struct inode
*inode
,
1560 const char **ppath
, int *ppathlen
, u64
*pino
,
1563 struct dentry
*dentry
;
1566 if (ceph_snap(inode
) == CEPH_NOSNAP
) {
1567 *pino
= ceph_ino(inode
);
1571 dentry
= d_find_alias(inode
);
1572 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1575 return PTR_ERR(path
);
1582 * request arguments may be specified via an inode *, a dentry *, or
1583 * an explicit ino+path.
1585 static int set_request_path_attr(struct inode
*rinode
, struct dentry
*rdentry
,
1586 const char *rpath
, u64 rino
,
1587 const char **ppath
, int *pathlen
,
1588 u64
*ino
, int *freepath
)
1593 r
= build_inode_path(rinode
, ppath
, pathlen
, ino
, freepath
);
1594 dout(" inode %p %llx.%llx\n", rinode
, ceph_ino(rinode
),
1596 } else if (rdentry
) {
1597 r
= build_dentry_path(rdentry
, ppath
, pathlen
, ino
, freepath
);
1598 dout(" dentry %p %llx/%.*s\n", rdentry
, *ino
, *pathlen
,
1600 } else if (rpath
|| rino
) {
1603 *pathlen
= strlen(rpath
);
1604 dout(" path %.*s\n", *pathlen
, rpath
);
1611 * called under mdsc->mutex
1613 static struct ceph_msg
*create_request_message(struct ceph_mds_client
*mdsc
,
1614 struct ceph_mds_request
*req
,
1617 struct ceph_msg
*msg
;
1618 struct ceph_mds_request_head
*head
;
1619 const char *path1
= NULL
;
1620 const char *path2
= NULL
;
1621 u64 ino1
= 0, ino2
= 0;
1622 int pathlen1
= 0, pathlen2
= 0;
1623 int freepath1
= 0, freepath2
= 0;
1629 ret
= set_request_path_attr(req
->r_inode
, req
->r_dentry
,
1630 req
->r_path1
, req
->r_ino1
.ino
,
1631 &path1
, &pathlen1
, &ino1
, &freepath1
);
1637 ret
= set_request_path_attr(NULL
, req
->r_old_dentry
,
1638 req
->r_path2
, req
->r_ino2
.ino
,
1639 &path2
, &pathlen2
, &ino2
, &freepath2
);
1645 len
= sizeof(*head
) +
1646 pathlen1
+ pathlen2
+ 2*(1 + sizeof(u32
) + sizeof(u64
));
1648 /* calculate (max) length for cap releases */
1649 len
+= sizeof(struct ceph_mds_request_release
) *
1650 (!!req
->r_inode_drop
+ !!req
->r_dentry_drop
+
1651 !!req
->r_old_inode_drop
+ !!req
->r_old_dentry_drop
);
1652 if (req
->r_dentry_drop
)
1653 len
+= req
->r_dentry
->d_name
.len
;
1654 if (req
->r_old_dentry_drop
)
1655 len
+= req
->r_old_dentry
->d_name
.len
;
1657 msg
= ceph_msg_new(CEPH_MSG_CLIENT_REQUEST
, len
, GFP_NOFS
, false);
1659 msg
= ERR_PTR(-ENOMEM
);
1663 msg
->hdr
.tid
= cpu_to_le64(req
->r_tid
);
1665 head
= msg
->front
.iov_base
;
1666 p
= msg
->front
.iov_base
+ sizeof(*head
);
1667 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1669 head
->mdsmap_epoch
= cpu_to_le32(mdsc
->mdsmap
->m_epoch
);
1670 head
->op
= cpu_to_le32(req
->r_op
);
1671 head
->caller_uid
= cpu_to_le32(req
->r_uid
);
1672 head
->caller_gid
= cpu_to_le32(req
->r_gid
);
1673 head
->args
= req
->r_args
;
1675 ceph_encode_filepath(&p
, end
, ino1
, path1
);
1676 ceph_encode_filepath(&p
, end
, ino2
, path2
);
1678 /* make note of release offset, in case we need to replay */
1679 req
->r_request_release_offset
= p
- msg
->front
.iov_base
;
1683 if (req
->r_inode_drop
)
1684 releases
+= ceph_encode_inode_release(&p
,
1685 req
->r_inode
? req
->r_inode
: req
->r_dentry
->d_inode
,
1686 mds
, req
->r_inode_drop
, req
->r_inode_unless
, 0);
1687 if (req
->r_dentry_drop
)
1688 releases
+= ceph_encode_dentry_release(&p
, req
->r_dentry
,
1689 mds
, req
->r_dentry_drop
, req
->r_dentry_unless
);
1690 if (req
->r_old_dentry_drop
)
1691 releases
+= ceph_encode_dentry_release(&p
, req
->r_old_dentry
,
1692 mds
, req
->r_old_dentry_drop
, req
->r_old_dentry_unless
);
1693 if (req
->r_old_inode_drop
)
1694 releases
+= ceph_encode_inode_release(&p
,
1695 req
->r_old_dentry
->d_inode
,
1696 mds
, req
->r_old_inode_drop
, req
->r_old_inode_unless
, 0);
1697 head
->num_releases
= cpu_to_le16(releases
);
1700 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
1701 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1703 msg
->pages
= req
->r_pages
;
1704 msg
->nr_pages
= req
->r_num_pages
;
1705 msg
->hdr
.data_len
= cpu_to_le32(req
->r_data_len
);
1706 msg
->hdr
.data_off
= cpu_to_le16(0);
1710 kfree((char *)path2
);
1713 kfree((char *)path1
);
1719 * called under mdsc->mutex if error, under no mutex if
1722 static void complete_request(struct ceph_mds_client
*mdsc
,
1723 struct ceph_mds_request
*req
)
1725 if (req
->r_callback
)
1726 req
->r_callback(mdsc
, req
);
1728 complete_all(&req
->r_completion
);
1732 * called under mdsc->mutex
1734 static int __prepare_send_request(struct ceph_mds_client
*mdsc
,
1735 struct ceph_mds_request
*req
,
1738 struct ceph_mds_request_head
*rhead
;
1739 struct ceph_msg
*msg
;
1744 struct ceph_cap
*cap
=
1745 ceph_get_cap_for_mds(ceph_inode(req
->r_inode
), mds
);
1748 req
->r_sent_on_mseq
= cap
->mseq
;
1750 req
->r_sent_on_mseq
= -1;
1752 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req
,
1753 req
->r_tid
, ceph_mds_op_name(req
->r_op
), req
->r_attempts
);
1755 if (req
->r_got_unsafe
) {
1757 * Replay. Do not regenerate message (and rebuild
1758 * paths, etc.); just use the original message.
1759 * Rebuilding paths will break for renames because
1760 * d_move mangles the src name.
1762 msg
= req
->r_request
;
1763 rhead
= msg
->front
.iov_base
;
1765 flags
= le32_to_cpu(rhead
->flags
);
1766 flags
|= CEPH_MDS_FLAG_REPLAY
;
1767 rhead
->flags
= cpu_to_le32(flags
);
1769 if (req
->r_target_inode
)
1770 rhead
->ino
= cpu_to_le64(ceph_ino(req
->r_target_inode
));
1772 rhead
->num_retry
= req
->r_attempts
- 1;
1774 /* remove cap/dentry releases from message */
1775 rhead
->num_releases
= 0;
1776 msg
->hdr
.front_len
= cpu_to_le32(req
->r_request_release_offset
);
1777 msg
->front
.iov_len
= req
->r_request_release_offset
;
1781 if (req
->r_request
) {
1782 ceph_msg_put(req
->r_request
);
1783 req
->r_request
= NULL
;
1785 msg
= create_request_message(mdsc
, req
, mds
);
1787 req
->r_err
= PTR_ERR(msg
);
1788 complete_request(mdsc
, req
);
1789 return PTR_ERR(msg
);
1791 req
->r_request
= msg
;
1793 rhead
= msg
->front
.iov_base
;
1794 rhead
->oldest_client_tid
= cpu_to_le64(__get_oldest_tid(mdsc
));
1795 if (req
->r_got_unsafe
)
1796 flags
|= CEPH_MDS_FLAG_REPLAY
;
1797 if (req
->r_locked_dir
)
1798 flags
|= CEPH_MDS_FLAG_WANT_DENTRY
;
1799 rhead
->flags
= cpu_to_le32(flags
);
1800 rhead
->num_fwd
= req
->r_num_fwd
;
1801 rhead
->num_retry
= req
->r_attempts
- 1;
1804 dout(" r_locked_dir = %p\n", req
->r_locked_dir
);
1809 * send request, or put it on the appropriate wait list.
1811 static int __do_request(struct ceph_mds_client
*mdsc
,
1812 struct ceph_mds_request
*req
)
1814 struct ceph_mds_session
*session
= NULL
;
1818 if (req
->r_err
|| req
->r_got_result
)
1821 if (req
->r_timeout
&&
1822 time_after_eq(jiffies
, req
->r_started
+ req
->r_timeout
)) {
1823 dout("do_request timed out\n");
1828 put_request_session(req
);
1830 mds
= __choose_mds(mdsc
, req
);
1832 ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) < CEPH_MDS_STATE_ACTIVE
) {
1833 dout("do_request no mds or not active, waiting for map\n");
1834 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
1838 /* get, open session */
1839 session
= __ceph_lookup_mds_session(mdsc
, mds
);
1841 session
= register_session(mdsc
, mds
);
1842 if (IS_ERR(session
)) {
1843 err
= PTR_ERR(session
);
1847 req
->r_session
= get_session(session
);
1849 dout("do_request mds%d session %p state %s\n", mds
, session
,
1850 session_state_name(session
->s_state
));
1851 if (session
->s_state
!= CEPH_MDS_SESSION_OPEN
&&
1852 session
->s_state
!= CEPH_MDS_SESSION_HUNG
) {
1853 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
1854 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
1855 __open_session(mdsc
, session
);
1856 list_add(&req
->r_wait
, &session
->s_waiting
);
1861 req
->r_resend_mds
= -1; /* forget any previous mds hint */
1863 if (req
->r_request_started
== 0) /* note request start time */
1864 req
->r_request_started
= jiffies
;
1866 err
= __prepare_send_request(mdsc
, req
, mds
);
1868 ceph_msg_get(req
->r_request
);
1869 ceph_con_send(&session
->s_con
, req
->r_request
);
1873 ceph_put_mds_session(session
);
1879 complete_request(mdsc
, req
);
1884 * called under mdsc->mutex
1886 static void __wake_requests(struct ceph_mds_client
*mdsc
,
1887 struct list_head
*head
)
1889 struct ceph_mds_request
*req
, *nreq
;
1891 list_for_each_entry_safe(req
, nreq
, head
, r_wait
) {
1892 list_del_init(&req
->r_wait
);
1893 __do_request(mdsc
, req
);
1898 * Wake up threads with requests pending for @mds, so that they can
1899 * resubmit their requests to a possibly different mds.
1901 static void kick_requests(struct ceph_mds_client
*mdsc
, int mds
)
1903 struct ceph_mds_request
*req
;
1906 dout("kick_requests mds%d\n", mds
);
1907 for (p
= rb_first(&mdsc
->request_tree
); p
; p
= rb_next(p
)) {
1908 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
1909 if (req
->r_got_unsafe
)
1911 if (req
->r_session
&&
1912 req
->r_session
->s_mds
== mds
) {
1913 dout(" kicking tid %llu\n", req
->r_tid
);
1914 __do_request(mdsc
, req
);
1919 void ceph_mdsc_submit_request(struct ceph_mds_client
*mdsc
,
1920 struct ceph_mds_request
*req
)
1922 dout("submit_request on %p\n", req
);
1923 mutex_lock(&mdsc
->mutex
);
1924 __register_request(mdsc
, req
, NULL
);
1925 __do_request(mdsc
, req
);
1926 mutex_unlock(&mdsc
->mutex
);
1930 * Synchrously perform an mds request. Take care of all of the
1931 * session setup, forwarding, retry details.
1933 int ceph_mdsc_do_request(struct ceph_mds_client
*mdsc
,
1935 struct ceph_mds_request
*req
)
1939 dout("do_request on %p\n", req
);
1941 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1943 ceph_get_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
1944 if (req
->r_locked_dir
)
1945 ceph_get_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
1946 if (req
->r_old_dentry
)
1947 ceph_get_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
1951 mutex_lock(&mdsc
->mutex
);
1952 __register_request(mdsc
, req
, dir
);
1953 __do_request(mdsc
, req
);
1957 __unregister_request(mdsc
, req
);
1958 dout("do_request early error %d\n", err
);
1963 mutex_unlock(&mdsc
->mutex
);
1964 dout("do_request waiting\n");
1965 if (req
->r_timeout
) {
1966 err
= (long)wait_for_completion_killable_timeout(
1967 &req
->r_completion
, req
->r_timeout
);
1971 err
= wait_for_completion_killable(&req
->r_completion
);
1973 dout("do_request waited, got %d\n", err
);
1974 mutex_lock(&mdsc
->mutex
);
1976 /* only abort if we didn't race with a real reply */
1977 if (req
->r_got_result
) {
1978 err
= le32_to_cpu(req
->r_reply_info
.head
->result
);
1979 } else if (err
< 0) {
1980 dout("aborted request %lld with %d\n", req
->r_tid
, err
);
1983 * ensure we aren't running concurrently with
1984 * ceph_fill_trace or ceph_readdir_prepopulate, which
1985 * rely on locks (dir mutex) held by our caller.
1987 mutex_lock(&req
->r_fill_mutex
);
1989 req
->r_aborted
= true;
1990 mutex_unlock(&req
->r_fill_mutex
);
1992 if (req
->r_locked_dir
&&
1993 (req
->r_op
& CEPH_MDS_OP_WRITE
))
1994 ceph_invalidate_dir_request(req
);
2000 mutex_unlock(&mdsc
->mutex
);
2001 dout("do_request %p done, result %d\n", req
, err
);
2006 * Invalidate dir D_COMPLETE, dentry lease state on an aborted MDS
2007 * namespace request.
2009 void ceph_invalidate_dir_request(struct ceph_mds_request
*req
)
2011 struct inode
*inode
= req
->r_locked_dir
;
2012 struct ceph_inode_info
*ci
= ceph_inode(inode
);
2014 dout("invalidate_dir_request %p (D_COMPLETE, lease(s))\n", inode
);
2015 spin_lock(&ci
->i_ceph_lock
);
2016 ceph_dir_clear_complete(inode
);
2017 ci
->i_release_count
++;
2018 spin_unlock(&ci
->i_ceph_lock
);
2021 ceph_invalidate_dentry_lease(req
->r_dentry
);
2022 if (req
->r_old_dentry
)
2023 ceph_invalidate_dentry_lease(req
->r_old_dentry
);
2029 * We take the session mutex and parse and process the reply immediately.
2030 * This preserves the logical ordering of replies, capabilities, etc., sent
2031 * by the MDS as they are applied to our local cache.
2033 static void handle_reply(struct ceph_mds_session
*session
, struct ceph_msg
*msg
)
2035 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2036 struct ceph_mds_request
*req
;
2037 struct ceph_mds_reply_head
*head
= msg
->front
.iov_base
;
2038 struct ceph_mds_reply_info_parsed
*rinfo
; /* parsed reply info */
2041 int mds
= session
->s_mds
;
2043 if (msg
->front
.iov_len
< sizeof(*head
)) {
2044 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2049 /* get request, session */
2050 tid
= le64_to_cpu(msg
->hdr
.tid
);
2051 mutex_lock(&mdsc
->mutex
);
2052 req
= __lookup_request(mdsc
, tid
);
2054 dout("handle_reply on unknown tid %llu\n", tid
);
2055 mutex_unlock(&mdsc
->mutex
);
2058 dout("handle_reply %p\n", req
);
2060 /* correct session? */
2061 if (req
->r_session
!= session
) {
2062 pr_err("mdsc_handle_reply got %llu on session mds%d"
2063 " not mds%d\n", tid
, session
->s_mds
,
2064 req
->r_session
? req
->r_session
->s_mds
: -1);
2065 mutex_unlock(&mdsc
->mutex
);
2070 if ((req
->r_got_unsafe
&& !head
->safe
) ||
2071 (req
->r_got_safe
&& head
->safe
)) {
2072 pr_warning("got a dup %s reply on %llu from mds%d\n",
2073 head
->safe
? "safe" : "unsafe", tid
, mds
);
2074 mutex_unlock(&mdsc
->mutex
);
2077 if (req
->r_got_safe
&& !head
->safe
) {
2078 pr_warning("got unsafe after safe on %llu from mds%d\n",
2080 mutex_unlock(&mdsc
->mutex
);
2084 result
= le32_to_cpu(head
->result
);
2088 * if we're not talking to the authority, send to them
2089 * if the authority has changed while we weren't looking,
2090 * send to new authority
2091 * Otherwise we just have to return an ESTALE
2093 if (result
== -ESTALE
) {
2094 dout("got ESTALE on request %llu", req
->r_tid
);
2095 if (!req
->r_inode
) {
2096 /* do nothing; not an authority problem */
2097 } else if (req
->r_direct_mode
!= USE_AUTH_MDS
) {
2098 dout("not using auth, setting for that now");
2099 req
->r_direct_mode
= USE_AUTH_MDS
;
2100 __do_request(mdsc
, req
);
2101 mutex_unlock(&mdsc
->mutex
);
2104 struct ceph_inode_info
*ci
= ceph_inode(req
->r_inode
);
2105 struct ceph_cap
*cap
= NULL
;
2108 cap
= ceph_get_cap_for_mds(ci
,
2109 req
->r_session
->s_mds
);
2111 dout("already using auth");
2112 if ((!cap
|| cap
!= ci
->i_auth_cap
) ||
2113 (cap
->mseq
!= req
->r_sent_on_mseq
)) {
2114 dout("but cap changed, so resending");
2115 __do_request(mdsc
, req
);
2116 mutex_unlock(&mdsc
->mutex
);
2120 dout("have to return ESTALE on request %llu", req
->r_tid
);
2125 req
->r_got_safe
= true;
2126 __unregister_request(mdsc
, req
);
2127 complete_all(&req
->r_safe_completion
);
2129 if (req
->r_got_unsafe
) {
2131 * We already handled the unsafe response, now do the
2132 * cleanup. No need to examine the response; the MDS
2133 * doesn't include any result info in the safe
2134 * response. And even if it did, there is nothing
2135 * useful we could do with a revised return value.
2137 dout("got safe reply %llu, mds%d\n", tid
, mds
);
2138 list_del_init(&req
->r_unsafe_item
);
2140 /* last unsafe request during umount? */
2141 if (mdsc
->stopping
&& !__get_oldest_req(mdsc
))
2142 complete_all(&mdsc
->safe_umount_waiters
);
2143 mutex_unlock(&mdsc
->mutex
);
2147 req
->r_got_unsafe
= true;
2148 list_add_tail(&req
->r_unsafe_item
, &req
->r_session
->s_unsafe
);
2151 dout("handle_reply tid %lld result %d\n", tid
, result
);
2152 rinfo
= &req
->r_reply_info
;
2153 err
= parse_reply_info(msg
, rinfo
, session
->s_con
.peer_features
);
2154 mutex_unlock(&mdsc
->mutex
);
2156 mutex_lock(&session
->s_mutex
);
2158 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds
, tid
);
2164 if (rinfo
->snapblob_len
) {
2165 down_write(&mdsc
->snap_rwsem
);
2166 ceph_update_snap_trace(mdsc
, rinfo
->snapblob
,
2167 rinfo
->snapblob
+ rinfo
->snapblob_len
,
2168 le32_to_cpu(head
->op
) == CEPH_MDS_OP_RMSNAP
);
2169 downgrade_write(&mdsc
->snap_rwsem
);
2171 down_read(&mdsc
->snap_rwsem
);
2174 /* insert trace into our cache */
2175 mutex_lock(&req
->r_fill_mutex
);
2176 err
= ceph_fill_trace(mdsc
->fsc
->sb
, req
, req
->r_session
);
2178 if (result
== 0 && req
->r_op
!= CEPH_MDS_OP_GETFILELOCK
&&
2180 ceph_readdir_prepopulate(req
, req
->r_session
);
2181 ceph_unreserve_caps(mdsc
, &req
->r_caps_reservation
);
2183 mutex_unlock(&req
->r_fill_mutex
);
2185 up_read(&mdsc
->snap_rwsem
);
2187 mutex_lock(&mdsc
->mutex
);
2188 if (!req
->r_aborted
) {
2194 req
->r_got_result
= true;
2197 dout("reply arrived after request %lld was aborted\n", tid
);
2199 mutex_unlock(&mdsc
->mutex
);
2201 ceph_add_cap_releases(mdsc
, req
->r_session
);
2202 mutex_unlock(&session
->s_mutex
);
2204 /* kick calling process */
2205 complete_request(mdsc
, req
);
2207 ceph_mdsc_put_request(req
);
2214 * handle mds notification that our request has been forwarded.
2216 static void handle_forward(struct ceph_mds_client
*mdsc
,
2217 struct ceph_mds_session
*session
,
2218 struct ceph_msg
*msg
)
2220 struct ceph_mds_request
*req
;
2221 u64 tid
= le64_to_cpu(msg
->hdr
.tid
);
2225 void *p
= msg
->front
.iov_base
;
2226 void *end
= p
+ msg
->front
.iov_len
;
2228 ceph_decode_need(&p
, end
, 2*sizeof(u32
), bad
);
2229 next_mds
= ceph_decode_32(&p
);
2230 fwd_seq
= ceph_decode_32(&p
);
2232 mutex_lock(&mdsc
->mutex
);
2233 req
= __lookup_request(mdsc
, tid
);
2235 dout("forward tid %llu to mds%d - req dne\n", tid
, next_mds
);
2236 goto out
; /* dup reply? */
2239 if (req
->r_aborted
) {
2240 dout("forward tid %llu aborted, unregistering\n", tid
);
2241 __unregister_request(mdsc
, req
);
2242 } else if (fwd_seq
<= req
->r_num_fwd
) {
2243 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2244 tid
, next_mds
, req
->r_num_fwd
, fwd_seq
);
2246 /* resend. forward race not possible; mds would drop */
2247 dout("forward tid %llu to mds%d (we resend)\n", tid
, next_mds
);
2249 BUG_ON(req
->r_got_result
);
2250 req
->r_num_fwd
= fwd_seq
;
2251 req
->r_resend_mds
= next_mds
;
2252 put_request_session(req
);
2253 __do_request(mdsc
, req
);
2255 ceph_mdsc_put_request(req
);
2257 mutex_unlock(&mdsc
->mutex
);
2261 pr_err("mdsc_handle_forward decode error err=%d\n", err
);
2265 * handle a mds session control message
2267 static void handle_session(struct ceph_mds_session
*session
,
2268 struct ceph_msg
*msg
)
2270 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2273 int mds
= session
->s_mds
;
2274 struct ceph_mds_session_head
*h
= msg
->front
.iov_base
;
2278 if (msg
->front
.iov_len
!= sizeof(*h
))
2280 op
= le32_to_cpu(h
->op
);
2281 seq
= le64_to_cpu(h
->seq
);
2283 mutex_lock(&mdsc
->mutex
);
2284 if (op
== CEPH_SESSION_CLOSE
)
2285 __unregister_session(mdsc
, session
);
2286 /* FIXME: this ttl calculation is generous */
2287 session
->s_ttl
= jiffies
+ HZ
*mdsc
->mdsmap
->m_session_autoclose
;
2288 mutex_unlock(&mdsc
->mutex
);
2290 mutex_lock(&session
->s_mutex
);
2292 dout("handle_session mds%d %s %p state %s seq %llu\n",
2293 mds
, ceph_session_op_name(op
), session
,
2294 session_state_name(session
->s_state
), seq
);
2296 if (session
->s_state
== CEPH_MDS_SESSION_HUNG
) {
2297 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2298 pr_info("mds%d came back\n", session
->s_mds
);
2302 case CEPH_SESSION_OPEN
:
2303 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2304 pr_info("mds%d reconnect success\n", session
->s_mds
);
2305 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2306 renewed_caps(mdsc
, session
, 0);
2309 __close_session(mdsc
, session
);
2312 case CEPH_SESSION_RENEWCAPS
:
2313 if (session
->s_renew_seq
== seq
)
2314 renewed_caps(mdsc
, session
, 1);
2317 case CEPH_SESSION_CLOSE
:
2318 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2319 pr_info("mds%d reconnect denied\n", session
->s_mds
);
2320 remove_session_caps(session
);
2321 wake
= 1; /* for good measure */
2322 wake_up_all(&mdsc
->session_close_wq
);
2323 kick_requests(mdsc
, mds
);
2326 case CEPH_SESSION_STALE
:
2327 pr_info("mds%d caps went stale, renewing\n",
2329 spin_lock(&session
->s_gen_ttl_lock
);
2330 session
->s_cap_gen
++;
2331 session
->s_cap_ttl
= jiffies
- 1;
2332 spin_unlock(&session
->s_gen_ttl_lock
);
2333 send_renew_caps(mdsc
, session
);
2336 case CEPH_SESSION_RECALL_STATE
:
2337 trim_caps(mdsc
, session
, le32_to_cpu(h
->max_caps
));
2341 pr_err("mdsc_handle_session bad op %d mds%d\n", op
, mds
);
2345 mutex_unlock(&session
->s_mutex
);
2347 mutex_lock(&mdsc
->mutex
);
2348 __wake_requests(mdsc
, &session
->s_waiting
);
2349 mutex_unlock(&mdsc
->mutex
);
2354 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds
,
2355 (int)msg
->front
.iov_len
);
2362 * called under session->mutex.
2364 static void replay_unsafe_requests(struct ceph_mds_client
*mdsc
,
2365 struct ceph_mds_session
*session
)
2367 struct ceph_mds_request
*req
, *nreq
;
2370 dout("replay_unsafe_requests mds%d\n", session
->s_mds
);
2372 mutex_lock(&mdsc
->mutex
);
2373 list_for_each_entry_safe(req
, nreq
, &session
->s_unsafe
, r_unsafe_item
) {
2374 err
= __prepare_send_request(mdsc
, req
, session
->s_mds
);
2376 ceph_msg_get(req
->r_request
);
2377 ceph_con_send(&session
->s_con
, req
->r_request
);
2380 mutex_unlock(&mdsc
->mutex
);
2384 * Encode information about a cap for a reconnect with the MDS.
2386 static int encode_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
2390 struct ceph_mds_cap_reconnect v2
;
2391 struct ceph_mds_cap_reconnect_v1 v1
;
2394 struct ceph_inode_info
*ci
;
2395 struct ceph_reconnect_state
*recon_state
= arg
;
2396 struct ceph_pagelist
*pagelist
= recon_state
->pagelist
;
2400 struct dentry
*dentry
;
2404 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2405 inode
, ceph_vinop(inode
), cap
, cap
->cap_id
,
2406 ceph_cap_string(cap
->issued
));
2407 err
= ceph_pagelist_encode_64(pagelist
, ceph_ino(inode
));
2411 dentry
= d_find_alias(inode
);
2413 path
= ceph_mdsc_build_path(dentry
, &pathlen
, &pathbase
, 0);
2415 err
= PTR_ERR(path
);
2422 err
= ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
2426 spin_lock(&ci
->i_ceph_lock
);
2427 cap
->seq
= 0; /* reset cap seq */
2428 cap
->issue_seq
= 0; /* and issue_seq */
2430 if (recon_state
->flock
) {
2431 rec
.v2
.cap_id
= cpu_to_le64(cap
->cap_id
);
2432 rec
.v2
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2433 rec
.v2
.issued
= cpu_to_le32(cap
->issued
);
2434 rec
.v2
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2435 rec
.v2
.pathbase
= cpu_to_le64(pathbase
);
2436 rec
.v2
.flock_len
= 0;
2437 reclen
= sizeof(rec
.v2
);
2439 rec
.v1
.cap_id
= cpu_to_le64(cap
->cap_id
);
2440 rec
.v1
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2441 rec
.v1
.issued
= cpu_to_le32(cap
->issued
);
2442 rec
.v1
.size
= cpu_to_le64(inode
->i_size
);
2443 ceph_encode_timespec(&rec
.v1
.mtime
, &inode
->i_mtime
);
2444 ceph_encode_timespec(&rec
.v1
.atime
, &inode
->i_atime
);
2445 rec
.v1
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2446 rec
.v1
.pathbase
= cpu_to_le64(pathbase
);
2447 reclen
= sizeof(rec
.v1
);
2449 spin_unlock(&ci
->i_ceph_lock
);
2451 if (recon_state
->flock
) {
2452 int num_fcntl_locks
, num_flock_locks
;
2453 struct ceph_pagelist_cursor trunc_point
;
2455 ceph_pagelist_set_cursor(pagelist
, &trunc_point
);
2458 ceph_count_locks(inode
, &num_fcntl_locks
,
2460 rec
.v2
.flock_len
= (2*sizeof(u32
) +
2461 (num_fcntl_locks
+num_flock_locks
) *
2462 sizeof(struct ceph_filelock
));
2465 /* pre-alloc pagelist */
2466 ceph_pagelist_truncate(pagelist
, &trunc_point
);
2467 err
= ceph_pagelist_append(pagelist
, &rec
, reclen
);
2469 err
= ceph_pagelist_reserve(pagelist
,
2475 err
= ceph_encode_locks(inode
,
2481 } while (err
== -ENOSPC
);
2483 err
= ceph_pagelist_append(pagelist
, &rec
, reclen
);
2495 * If an MDS fails and recovers, clients need to reconnect in order to
2496 * reestablish shared state. This includes all caps issued through
2497 * this session _and_ the snap_realm hierarchy. Because it's not
2498 * clear which snap realms the mds cares about, we send everything we
2499 * know about.. that ensures we'll then get any new info the
2500 * recovering MDS might have.
2502 * This is a relatively heavyweight operation, but it's rare.
2504 * called with mdsc->mutex held.
2506 static void send_mds_reconnect(struct ceph_mds_client
*mdsc
,
2507 struct ceph_mds_session
*session
)
2509 struct ceph_msg
*reply
;
2511 int mds
= session
->s_mds
;
2513 struct ceph_pagelist
*pagelist
;
2514 struct ceph_reconnect_state recon_state
;
2516 pr_info("mds%d reconnect start\n", mds
);
2518 pagelist
= kmalloc(sizeof(*pagelist
), GFP_NOFS
);
2520 goto fail_nopagelist
;
2521 ceph_pagelist_init(pagelist
);
2523 reply
= ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT
, 0, GFP_NOFS
, false);
2527 mutex_lock(&session
->s_mutex
);
2528 session
->s_state
= CEPH_MDS_SESSION_RECONNECTING
;
2531 ceph_con_open(&session
->s_con
,
2532 CEPH_ENTITY_TYPE_MDS
, mds
,
2533 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
2535 /* replay unsafe requests */
2536 replay_unsafe_requests(mdsc
, session
);
2538 down_read(&mdsc
->snap_rwsem
);
2540 dout("session %p state %s\n", session
,
2541 session_state_name(session
->s_state
));
2543 /* drop old cap expires; we're about to reestablish that state */
2544 discard_cap_releases(mdsc
, session
);
2546 /* traverse this session's caps */
2547 err
= ceph_pagelist_encode_32(pagelist
, session
->s_nr_caps
);
2551 recon_state
.pagelist
= pagelist
;
2552 recon_state
.flock
= session
->s_con
.peer_features
& CEPH_FEATURE_FLOCK
;
2553 err
= iterate_session_caps(session
, encode_caps_cb
, &recon_state
);
2558 * snaprealms. we provide mds with the ino, seq (version), and
2559 * parent for all of our realms. If the mds has any newer info,
2562 for (p
= rb_first(&mdsc
->snap_realms
); p
; p
= rb_next(p
)) {
2563 struct ceph_snap_realm
*realm
=
2564 rb_entry(p
, struct ceph_snap_realm
, node
);
2565 struct ceph_mds_snaprealm_reconnect sr_rec
;
2567 dout(" adding snap realm %llx seq %lld parent %llx\n",
2568 realm
->ino
, realm
->seq
, realm
->parent_ino
);
2569 sr_rec
.ino
= cpu_to_le64(realm
->ino
);
2570 sr_rec
.seq
= cpu_to_le64(realm
->seq
);
2571 sr_rec
.parent
= cpu_to_le64(realm
->parent_ino
);
2572 err
= ceph_pagelist_append(pagelist
, &sr_rec
, sizeof(sr_rec
));
2577 reply
->pagelist
= pagelist
;
2578 if (recon_state
.flock
)
2579 reply
->hdr
.version
= cpu_to_le16(2);
2580 reply
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
2581 reply
->nr_pages
= calc_pages_for(0, pagelist
->length
);
2582 ceph_con_send(&session
->s_con
, reply
);
2584 mutex_unlock(&session
->s_mutex
);
2586 mutex_lock(&mdsc
->mutex
);
2587 __wake_requests(mdsc
, &session
->s_waiting
);
2588 mutex_unlock(&mdsc
->mutex
);
2590 up_read(&mdsc
->snap_rwsem
);
2594 ceph_msg_put(reply
);
2595 up_read(&mdsc
->snap_rwsem
);
2596 mutex_unlock(&session
->s_mutex
);
2598 ceph_pagelist_release(pagelist
);
2601 pr_err("error %d preparing reconnect for mds%d\n", err
, mds
);
2607 * compare old and new mdsmaps, kicking requests
2608 * and closing out old connections as necessary
2610 * called under mdsc->mutex.
2612 static void check_new_map(struct ceph_mds_client
*mdsc
,
2613 struct ceph_mdsmap
*newmap
,
2614 struct ceph_mdsmap
*oldmap
)
2617 int oldstate
, newstate
;
2618 struct ceph_mds_session
*s
;
2620 dout("check_new_map new %u old %u\n",
2621 newmap
->m_epoch
, oldmap
->m_epoch
);
2623 for (i
= 0; i
< oldmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
2624 if (mdsc
->sessions
[i
] == NULL
)
2626 s
= mdsc
->sessions
[i
];
2627 oldstate
= ceph_mdsmap_get_state(oldmap
, i
);
2628 newstate
= ceph_mdsmap_get_state(newmap
, i
);
2630 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2631 i
, ceph_mds_state_name(oldstate
),
2632 ceph_mdsmap_is_laggy(oldmap
, i
) ? " (laggy)" : "",
2633 ceph_mds_state_name(newstate
),
2634 ceph_mdsmap_is_laggy(newmap
, i
) ? " (laggy)" : "",
2635 session_state_name(s
->s_state
));
2637 if (memcmp(ceph_mdsmap_get_addr(oldmap
, i
),
2638 ceph_mdsmap_get_addr(newmap
, i
),
2639 sizeof(struct ceph_entity_addr
))) {
2640 if (s
->s_state
== CEPH_MDS_SESSION_OPENING
) {
2641 /* the session never opened, just close it
2643 __wake_requests(mdsc
, &s
->s_waiting
);
2644 __unregister_session(mdsc
, s
);
2647 mutex_unlock(&mdsc
->mutex
);
2648 mutex_lock(&s
->s_mutex
);
2649 mutex_lock(&mdsc
->mutex
);
2650 ceph_con_close(&s
->s_con
);
2651 mutex_unlock(&s
->s_mutex
);
2652 s
->s_state
= CEPH_MDS_SESSION_RESTARTING
;
2655 /* kick any requests waiting on the recovering mds */
2656 kick_requests(mdsc
, i
);
2657 } else if (oldstate
== newstate
) {
2658 continue; /* nothing new with this mds */
2664 if (s
->s_state
== CEPH_MDS_SESSION_RESTARTING
&&
2665 newstate
>= CEPH_MDS_STATE_RECONNECT
) {
2666 mutex_unlock(&mdsc
->mutex
);
2667 send_mds_reconnect(mdsc
, s
);
2668 mutex_lock(&mdsc
->mutex
);
2672 * kick request on any mds that has gone active.
2674 if (oldstate
< CEPH_MDS_STATE_ACTIVE
&&
2675 newstate
>= CEPH_MDS_STATE_ACTIVE
) {
2676 if (oldstate
!= CEPH_MDS_STATE_CREATING
&&
2677 oldstate
!= CEPH_MDS_STATE_STARTING
)
2678 pr_info("mds%d recovery completed\n", s
->s_mds
);
2679 kick_requests(mdsc
, i
);
2680 ceph_kick_flushing_caps(mdsc
, s
);
2681 wake_up_session_caps(s
, 1);
2685 for (i
= 0; i
< newmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
2686 s
= mdsc
->sessions
[i
];
2689 if (!ceph_mdsmap_is_laggy(newmap
, i
))
2691 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
2692 s
->s_state
== CEPH_MDS_SESSION_HUNG
||
2693 s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
2694 dout(" connecting to export targets of laggy mds%d\n",
2696 __open_export_target_sessions(mdsc
, s
);
2708 * caller must hold session s_mutex, dentry->d_lock
2710 void __ceph_mdsc_drop_dentry_lease(struct dentry
*dentry
)
2712 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
2714 ceph_put_mds_session(di
->lease_session
);
2715 di
->lease_session
= NULL
;
2718 static void handle_lease(struct ceph_mds_client
*mdsc
,
2719 struct ceph_mds_session
*session
,
2720 struct ceph_msg
*msg
)
2722 struct super_block
*sb
= mdsc
->fsc
->sb
;
2723 struct inode
*inode
;
2724 struct dentry
*parent
, *dentry
;
2725 struct ceph_dentry_info
*di
;
2726 int mds
= session
->s_mds
;
2727 struct ceph_mds_lease
*h
= msg
->front
.iov_base
;
2729 struct ceph_vino vino
;
2733 dout("handle_lease from mds%d\n", mds
);
2736 if (msg
->front
.iov_len
< sizeof(*h
) + sizeof(u32
))
2738 vino
.ino
= le64_to_cpu(h
->ino
);
2739 vino
.snap
= CEPH_NOSNAP
;
2740 seq
= le32_to_cpu(h
->seq
);
2741 dname
.name
= (void *)h
+ sizeof(*h
) + sizeof(u32
);
2742 dname
.len
= msg
->front
.iov_len
- sizeof(*h
) - sizeof(u32
);
2743 if (dname
.len
!= get_unaligned_le32(h
+1))
2746 mutex_lock(&session
->s_mutex
);
2750 inode
= ceph_find_inode(sb
, vino
);
2751 dout("handle_lease %s, ino %llx %p %.*s\n",
2752 ceph_lease_op_name(h
->action
), vino
.ino
, inode
,
2753 dname
.len
, dname
.name
);
2754 if (inode
== NULL
) {
2755 dout("handle_lease no inode %llx\n", vino
.ino
);
2760 parent
= d_find_alias(inode
);
2762 dout("no parent dentry on inode %p\n", inode
);
2764 goto release
; /* hrm... */
2766 dname
.hash
= full_name_hash(dname
.name
, dname
.len
);
2767 dentry
= d_lookup(parent
, &dname
);
2772 spin_lock(&dentry
->d_lock
);
2773 di
= ceph_dentry(dentry
);
2774 switch (h
->action
) {
2775 case CEPH_MDS_LEASE_REVOKE
:
2776 if (di
->lease_session
== session
) {
2777 if (ceph_seq_cmp(di
->lease_seq
, seq
) > 0)
2778 h
->seq
= cpu_to_le32(di
->lease_seq
);
2779 __ceph_mdsc_drop_dentry_lease(dentry
);
2784 case CEPH_MDS_LEASE_RENEW
:
2785 if (di
->lease_session
== session
&&
2786 di
->lease_gen
== session
->s_cap_gen
&&
2787 di
->lease_renew_from
&&
2788 di
->lease_renew_after
== 0) {
2789 unsigned long duration
=
2790 le32_to_cpu(h
->duration_ms
) * HZ
/ 1000;
2792 di
->lease_seq
= seq
;
2793 dentry
->d_time
= di
->lease_renew_from
+ duration
;
2794 di
->lease_renew_after
= di
->lease_renew_from
+
2796 di
->lease_renew_from
= 0;
2800 spin_unlock(&dentry
->d_lock
);
2807 /* let's just reuse the same message */
2808 h
->action
= CEPH_MDS_LEASE_REVOKE_ACK
;
2810 ceph_con_send(&session
->s_con
, msg
);
2814 mutex_unlock(&session
->s_mutex
);
2818 pr_err("corrupt lease message\n");
2822 void ceph_mdsc_lease_send_msg(struct ceph_mds_session
*session
,
2823 struct inode
*inode
,
2824 struct dentry
*dentry
, char action
,
2827 struct ceph_msg
*msg
;
2828 struct ceph_mds_lease
*lease
;
2829 int len
= sizeof(*lease
) + sizeof(u32
);
2832 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2833 inode
, dentry
, ceph_lease_op_name(action
), session
->s_mds
);
2834 dnamelen
= dentry
->d_name
.len
;
2837 msg
= ceph_msg_new(CEPH_MSG_CLIENT_LEASE
, len
, GFP_NOFS
, false);
2840 lease
= msg
->front
.iov_base
;
2841 lease
->action
= action
;
2842 lease
->ino
= cpu_to_le64(ceph_vino(inode
).ino
);
2843 lease
->first
= lease
->last
= cpu_to_le64(ceph_vino(inode
).snap
);
2844 lease
->seq
= cpu_to_le32(seq
);
2845 put_unaligned_le32(dnamelen
, lease
+ 1);
2846 memcpy((void *)(lease
+ 1) + 4, dentry
->d_name
.name
, dnamelen
);
2849 * if this is a preemptive lease RELEASE, no need to
2850 * flush request stream, since the actual request will
2853 msg
->more_to_follow
= (action
== CEPH_MDS_LEASE_RELEASE
);
2855 ceph_con_send(&session
->s_con
, msg
);
2859 * Preemptively release a lease we expect to invalidate anyway.
2860 * Pass @inode always, @dentry is optional.
2862 void ceph_mdsc_lease_release(struct ceph_mds_client
*mdsc
, struct inode
*inode
,
2863 struct dentry
*dentry
)
2865 struct ceph_dentry_info
*di
;
2866 struct ceph_mds_session
*session
;
2869 BUG_ON(inode
== NULL
);
2870 BUG_ON(dentry
== NULL
);
2872 /* is dentry lease valid? */
2873 spin_lock(&dentry
->d_lock
);
2874 di
= ceph_dentry(dentry
);
2875 if (!di
|| !di
->lease_session
||
2876 di
->lease_session
->s_mds
< 0 ||
2877 di
->lease_gen
!= di
->lease_session
->s_cap_gen
||
2878 !time_before(jiffies
, dentry
->d_time
)) {
2879 dout("lease_release inode %p dentry %p -- "
2882 spin_unlock(&dentry
->d_lock
);
2886 /* we do have a lease on this dentry; note mds and seq */
2887 session
= ceph_get_mds_session(di
->lease_session
);
2888 seq
= di
->lease_seq
;
2889 __ceph_mdsc_drop_dentry_lease(dentry
);
2890 spin_unlock(&dentry
->d_lock
);
2892 dout("lease_release inode %p dentry %p to mds%d\n",
2893 inode
, dentry
, session
->s_mds
);
2894 ceph_mdsc_lease_send_msg(session
, inode
, dentry
,
2895 CEPH_MDS_LEASE_RELEASE
, seq
);
2896 ceph_put_mds_session(session
);
2900 * drop all leases (and dentry refs) in preparation for umount
2902 static void drop_leases(struct ceph_mds_client
*mdsc
)
2906 dout("drop_leases\n");
2907 mutex_lock(&mdsc
->mutex
);
2908 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
2909 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
2912 mutex_unlock(&mdsc
->mutex
);
2913 mutex_lock(&s
->s_mutex
);
2914 mutex_unlock(&s
->s_mutex
);
2915 ceph_put_mds_session(s
);
2916 mutex_lock(&mdsc
->mutex
);
2918 mutex_unlock(&mdsc
->mutex
);
2924 * delayed work -- periodically trim expired leases, renew caps with mds
2926 static void schedule_delayed(struct ceph_mds_client
*mdsc
)
2929 unsigned hz
= round_jiffies_relative(HZ
* delay
);
2930 schedule_delayed_work(&mdsc
->delayed_work
, hz
);
2933 static void delayed_work(struct work_struct
*work
)
2936 struct ceph_mds_client
*mdsc
=
2937 container_of(work
, struct ceph_mds_client
, delayed_work
.work
);
2941 dout("mdsc delayed_work\n");
2942 ceph_check_delayed_caps(mdsc
);
2944 mutex_lock(&mdsc
->mutex
);
2945 renew_interval
= mdsc
->mdsmap
->m_session_timeout
>> 2;
2946 renew_caps
= time_after_eq(jiffies
, HZ
*renew_interval
+
2947 mdsc
->last_renew_caps
);
2949 mdsc
->last_renew_caps
= jiffies
;
2951 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
2952 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
2955 if (s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
2956 dout("resending session close request for mds%d\n",
2958 request_close_session(mdsc
, s
);
2959 ceph_put_mds_session(s
);
2962 if (s
->s_ttl
&& time_after(jiffies
, s
->s_ttl
)) {
2963 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
) {
2964 s
->s_state
= CEPH_MDS_SESSION_HUNG
;
2965 pr_info("mds%d hung\n", s
->s_mds
);
2968 if (s
->s_state
< CEPH_MDS_SESSION_OPEN
) {
2969 /* this mds is failed or recovering, just wait */
2970 ceph_put_mds_session(s
);
2973 mutex_unlock(&mdsc
->mutex
);
2975 mutex_lock(&s
->s_mutex
);
2977 send_renew_caps(mdsc
, s
);
2979 ceph_con_keepalive(&s
->s_con
);
2980 ceph_add_cap_releases(mdsc
, s
);
2981 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
2982 s
->s_state
== CEPH_MDS_SESSION_HUNG
)
2983 ceph_send_cap_releases(mdsc
, s
);
2984 mutex_unlock(&s
->s_mutex
);
2985 ceph_put_mds_session(s
);
2987 mutex_lock(&mdsc
->mutex
);
2989 mutex_unlock(&mdsc
->mutex
);
2991 schedule_delayed(mdsc
);
2994 int ceph_mdsc_init(struct ceph_fs_client
*fsc
)
2997 struct ceph_mds_client
*mdsc
;
2999 mdsc
= kzalloc(sizeof(struct ceph_mds_client
), GFP_NOFS
);
3004 mutex_init(&mdsc
->mutex
);
3005 mdsc
->mdsmap
= kzalloc(sizeof(*mdsc
->mdsmap
), GFP_NOFS
);
3006 if (mdsc
->mdsmap
== NULL
)
3009 init_completion(&mdsc
->safe_umount_waiters
);
3010 init_waitqueue_head(&mdsc
->session_close_wq
);
3011 INIT_LIST_HEAD(&mdsc
->waiting_for_map
);
3012 mdsc
->sessions
= NULL
;
3013 mdsc
->max_sessions
= 0;
3015 init_rwsem(&mdsc
->snap_rwsem
);
3016 mdsc
->snap_realms
= RB_ROOT
;
3017 INIT_LIST_HEAD(&mdsc
->snap_empty
);
3018 spin_lock_init(&mdsc
->snap_empty_lock
);
3020 mdsc
->request_tree
= RB_ROOT
;
3021 INIT_DELAYED_WORK(&mdsc
->delayed_work
, delayed_work
);
3022 mdsc
->last_renew_caps
= jiffies
;
3023 INIT_LIST_HEAD(&mdsc
->cap_delay_list
);
3024 spin_lock_init(&mdsc
->cap_delay_lock
);
3025 INIT_LIST_HEAD(&mdsc
->snap_flush_list
);
3026 spin_lock_init(&mdsc
->snap_flush_lock
);
3027 mdsc
->cap_flush_seq
= 0;
3028 INIT_LIST_HEAD(&mdsc
->cap_dirty
);
3029 INIT_LIST_HEAD(&mdsc
->cap_dirty_migrating
);
3030 mdsc
->num_cap_flushing
= 0;
3031 spin_lock_init(&mdsc
->cap_dirty_lock
);
3032 init_waitqueue_head(&mdsc
->cap_flushing_wq
);
3033 spin_lock_init(&mdsc
->dentry_lru_lock
);
3034 INIT_LIST_HEAD(&mdsc
->dentry_lru
);
3036 ceph_caps_init(mdsc
);
3037 ceph_adjust_min_caps(mdsc
, fsc
->min_caps
);
3043 * Wait for safe replies on open mds requests. If we time out, drop
3044 * all requests from the tree to avoid dangling dentry refs.
3046 static void wait_requests(struct ceph_mds_client
*mdsc
)
3048 struct ceph_mds_request
*req
;
3049 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3051 mutex_lock(&mdsc
->mutex
);
3052 if (__get_oldest_req(mdsc
)) {
3053 mutex_unlock(&mdsc
->mutex
);
3055 dout("wait_requests waiting for requests\n");
3056 wait_for_completion_timeout(&mdsc
->safe_umount_waiters
,
3057 fsc
->client
->options
->mount_timeout
* HZ
);
3059 /* tear down remaining requests */
3060 mutex_lock(&mdsc
->mutex
);
3061 while ((req
= __get_oldest_req(mdsc
))) {
3062 dout("wait_requests timed out on tid %llu\n",
3064 __unregister_request(mdsc
, req
);
3067 mutex_unlock(&mdsc
->mutex
);
3068 dout("wait_requests done\n");
3072 * called before mount is ro, and before dentries are torn down.
3073 * (hmm, does this still race with new lookups?)
3075 void ceph_mdsc_pre_umount(struct ceph_mds_client
*mdsc
)
3077 dout("pre_umount\n");
3081 ceph_flush_dirty_caps(mdsc
);
3082 wait_requests(mdsc
);
3085 * wait for reply handlers to drop their request refs and
3086 * their inode/dcache refs
3092 * wait for all write mds requests to flush.
3094 static void wait_unsafe_requests(struct ceph_mds_client
*mdsc
, u64 want_tid
)
3096 struct ceph_mds_request
*req
= NULL
, *nextreq
;
3099 mutex_lock(&mdsc
->mutex
);
3100 dout("wait_unsafe_requests want %lld\n", want_tid
);
3102 req
= __get_oldest_req(mdsc
);
3103 while (req
&& req
->r_tid
<= want_tid
) {
3104 /* find next request */
3105 n
= rb_next(&req
->r_node
);
3107 nextreq
= rb_entry(n
, struct ceph_mds_request
, r_node
);
3110 if ((req
->r_op
& CEPH_MDS_OP_WRITE
)) {
3112 ceph_mdsc_get_request(req
);
3114 ceph_mdsc_get_request(nextreq
);
3115 mutex_unlock(&mdsc
->mutex
);
3116 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3117 req
->r_tid
, want_tid
);
3118 wait_for_completion(&req
->r_safe_completion
);
3119 mutex_lock(&mdsc
->mutex
);
3120 ceph_mdsc_put_request(req
);
3122 break; /* next dne before, so we're done! */
3123 if (RB_EMPTY_NODE(&nextreq
->r_node
)) {
3124 /* next request was removed from tree */
3125 ceph_mdsc_put_request(nextreq
);
3128 ceph_mdsc_put_request(nextreq
); /* won't go away */
3132 mutex_unlock(&mdsc
->mutex
);
3133 dout("wait_unsafe_requests done\n");
3136 void ceph_mdsc_sync(struct ceph_mds_client
*mdsc
)
3138 u64 want_tid
, want_flush
;
3140 if (mdsc
->fsc
->mount_state
== CEPH_MOUNT_SHUTDOWN
)
3144 mutex_lock(&mdsc
->mutex
);
3145 want_tid
= mdsc
->last_tid
;
3146 want_flush
= mdsc
->cap_flush_seq
;
3147 mutex_unlock(&mdsc
->mutex
);
3148 dout("sync want tid %lld flush_seq %lld\n", want_tid
, want_flush
);
3150 ceph_flush_dirty_caps(mdsc
);
3152 wait_unsafe_requests(mdsc
, want_tid
);
3153 wait_event(mdsc
->cap_flushing_wq
, check_cap_flush(mdsc
, want_flush
));
3157 * true if all sessions are closed, or we force unmount
3159 static bool done_closing_sessions(struct ceph_mds_client
*mdsc
)
3163 if (mdsc
->fsc
->mount_state
== CEPH_MOUNT_SHUTDOWN
)
3166 mutex_lock(&mdsc
->mutex
);
3167 for (i
= 0; i
< mdsc
->max_sessions
; i
++)
3168 if (mdsc
->sessions
[i
])
3170 mutex_unlock(&mdsc
->mutex
);
3175 * called after sb is ro.
3177 void ceph_mdsc_close_sessions(struct ceph_mds_client
*mdsc
)
3179 struct ceph_mds_session
*session
;
3181 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3182 unsigned long timeout
= fsc
->client
->options
->mount_timeout
* HZ
;
3184 dout("close_sessions\n");
3186 /* close sessions */
3187 mutex_lock(&mdsc
->mutex
);
3188 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3189 session
= __ceph_lookup_mds_session(mdsc
, i
);
3192 mutex_unlock(&mdsc
->mutex
);
3193 mutex_lock(&session
->s_mutex
);
3194 __close_session(mdsc
, session
);
3195 mutex_unlock(&session
->s_mutex
);
3196 ceph_put_mds_session(session
);
3197 mutex_lock(&mdsc
->mutex
);
3199 mutex_unlock(&mdsc
->mutex
);
3201 dout("waiting for sessions to close\n");
3202 wait_event_timeout(mdsc
->session_close_wq
, done_closing_sessions(mdsc
),
3205 /* tear down remaining sessions */
3206 mutex_lock(&mdsc
->mutex
);
3207 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3208 if (mdsc
->sessions
[i
]) {
3209 session
= get_session(mdsc
->sessions
[i
]);
3210 __unregister_session(mdsc
, session
);
3211 mutex_unlock(&mdsc
->mutex
);
3212 mutex_lock(&session
->s_mutex
);
3213 remove_session_caps(session
);
3214 mutex_unlock(&session
->s_mutex
);
3215 ceph_put_mds_session(session
);
3216 mutex_lock(&mdsc
->mutex
);
3219 WARN_ON(!list_empty(&mdsc
->cap_delay_list
));
3220 mutex_unlock(&mdsc
->mutex
);
3222 ceph_cleanup_empty_realms(mdsc
);
3224 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3229 static void ceph_mdsc_stop(struct ceph_mds_client
*mdsc
)
3232 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3234 ceph_mdsmap_destroy(mdsc
->mdsmap
);
3235 kfree(mdsc
->sessions
);
3236 ceph_caps_finalize(mdsc
);
3239 void ceph_mdsc_destroy(struct ceph_fs_client
*fsc
)
3241 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
3243 dout("mdsc_destroy %p\n", mdsc
);
3244 ceph_mdsc_stop(mdsc
);
3246 /* flush out any connection work with references to us */
3251 dout("mdsc_destroy %p done\n", mdsc
);
3256 * handle mds map update.
3258 void ceph_mdsc_handle_map(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3262 void *p
= msg
->front
.iov_base
;
3263 void *end
= p
+ msg
->front
.iov_len
;
3264 struct ceph_mdsmap
*newmap
, *oldmap
;
3265 struct ceph_fsid fsid
;
3268 ceph_decode_need(&p
, end
, sizeof(fsid
)+2*sizeof(u32
), bad
);
3269 ceph_decode_copy(&p
, &fsid
, sizeof(fsid
));
3270 if (ceph_check_fsid(mdsc
->fsc
->client
, &fsid
) < 0)
3272 epoch
= ceph_decode_32(&p
);
3273 maplen
= ceph_decode_32(&p
);
3274 dout("handle_map epoch %u len %d\n", epoch
, (int)maplen
);
3276 /* do we need it? */
3277 ceph_monc_got_mdsmap(&mdsc
->fsc
->client
->monc
, epoch
);
3278 mutex_lock(&mdsc
->mutex
);
3279 if (mdsc
->mdsmap
&& epoch
<= mdsc
->mdsmap
->m_epoch
) {
3280 dout("handle_map epoch %u <= our %u\n",
3281 epoch
, mdsc
->mdsmap
->m_epoch
);
3282 mutex_unlock(&mdsc
->mutex
);
3286 newmap
= ceph_mdsmap_decode(&p
, end
);
3287 if (IS_ERR(newmap
)) {
3288 err
= PTR_ERR(newmap
);
3292 /* swap into place */
3294 oldmap
= mdsc
->mdsmap
;
3295 mdsc
->mdsmap
= newmap
;
3296 check_new_map(mdsc
, newmap
, oldmap
);
3297 ceph_mdsmap_destroy(oldmap
);
3299 mdsc
->mdsmap
= newmap
; /* first mds map */
3301 mdsc
->fsc
->sb
->s_maxbytes
= mdsc
->mdsmap
->m_max_file_size
;
3303 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3305 mutex_unlock(&mdsc
->mutex
);
3306 schedule_delayed(mdsc
);
3310 mutex_unlock(&mdsc
->mutex
);
3312 pr_err("error decoding mdsmap %d\n", err
);
3316 static struct ceph_connection
*con_get(struct ceph_connection
*con
)
3318 struct ceph_mds_session
*s
= con
->private;
3320 if (get_session(s
)) {
3321 dout("mdsc con_get %p ok (%d)\n", s
, atomic_read(&s
->s_ref
));
3324 dout("mdsc con_get %p FAIL\n", s
);
3328 static void con_put(struct ceph_connection
*con
)
3330 struct ceph_mds_session
*s
= con
->private;
3332 dout("mdsc con_put %p (%d)\n", s
, atomic_read(&s
->s_ref
) - 1);
3333 ceph_put_mds_session(s
);
3337 * if the client is unresponsive for long enough, the mds will kill
3338 * the session entirely.
3340 static void peer_reset(struct ceph_connection
*con
)
3342 struct ceph_mds_session
*s
= con
->private;
3343 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3345 pr_warning("mds%d closed our session\n", s
->s_mds
);
3346 send_mds_reconnect(mdsc
, s
);
3349 static void dispatch(struct ceph_connection
*con
, struct ceph_msg
*msg
)
3351 struct ceph_mds_session
*s
= con
->private;
3352 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3353 int type
= le16_to_cpu(msg
->hdr
.type
);
3355 mutex_lock(&mdsc
->mutex
);
3356 if (__verify_registered_session(mdsc
, s
) < 0) {
3357 mutex_unlock(&mdsc
->mutex
);
3360 mutex_unlock(&mdsc
->mutex
);
3363 case CEPH_MSG_MDS_MAP
:
3364 ceph_mdsc_handle_map(mdsc
, msg
);
3366 case CEPH_MSG_CLIENT_SESSION
:
3367 handle_session(s
, msg
);
3369 case CEPH_MSG_CLIENT_REPLY
:
3370 handle_reply(s
, msg
);
3372 case CEPH_MSG_CLIENT_REQUEST_FORWARD
:
3373 handle_forward(mdsc
, s
, msg
);
3375 case CEPH_MSG_CLIENT_CAPS
:
3376 ceph_handle_caps(s
, msg
);
3378 case CEPH_MSG_CLIENT_SNAP
:
3379 ceph_handle_snap(mdsc
, s
, msg
);
3381 case CEPH_MSG_CLIENT_LEASE
:
3382 handle_lease(mdsc
, s
, msg
);
3386 pr_err("received unknown message type %d %s\n", type
,
3387 ceph_msg_type_name(type
));
3398 * Note: returned pointer is the address of a structure that's
3399 * managed separately. Caller must *not* attempt to free it.
3401 static struct ceph_auth_handshake
*get_authorizer(struct ceph_connection
*con
,
3402 int *proto
, int force_new
)
3404 struct ceph_mds_session
*s
= con
->private;
3405 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3406 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3407 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
3409 if (force_new
&& auth
->authorizer
) {
3410 if (ac
->ops
&& ac
->ops
->destroy_authorizer
)
3411 ac
->ops
->destroy_authorizer(ac
, auth
->authorizer
);
3412 auth
->authorizer
= NULL
;
3414 if (!auth
->authorizer
&& ac
->ops
&& ac
->ops
->create_authorizer
) {
3415 int ret
= ac
->ops
->create_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3418 return ERR_PTR(ret
);
3420 *proto
= ac
->protocol
;
3426 static int verify_authorizer_reply(struct ceph_connection
*con
, int len
)
3428 struct ceph_mds_session
*s
= con
->private;
3429 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3430 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3432 return ac
->ops
->verify_authorizer_reply(ac
, s
->s_auth
.authorizer
, len
);
3435 static int invalidate_authorizer(struct ceph_connection
*con
)
3437 struct ceph_mds_session
*s
= con
->private;
3438 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3439 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3441 if (ac
->ops
->invalidate_authorizer
)
3442 ac
->ops
->invalidate_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
);
3444 return ceph_monc_validate_auth(&mdsc
->fsc
->client
->monc
);
3447 static const struct ceph_connection_operations mds_con_ops
= {
3450 .dispatch
= dispatch
,
3451 .get_authorizer
= get_authorizer
,
3452 .verify_authorizer_reply
= verify_authorizer_reply
,
3453 .invalidate_authorizer
= invalidate_authorizer
,
3454 .peer_reset
= peer_reset
,