1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
7 #include <linux/sched.h>
8 #include <linux/debugfs.h>
9 #include <linux/seq_file.h>
10 #include <linux/utsname.h>
11 #include <linux/ratelimit.h>
14 #include "mds_client.h"
16 #include <linux/ceph/ceph_features.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/pagelist.h>
20 #include <linux/ceph/auth.h>
21 #include <linux/ceph/debugfs.h>
24 * A cluster of MDS (metadata server) daemons is responsible for
25 * managing the file system namespace (the directory hierarchy and
26 * inodes) and for coordinating shared access to storage. Metadata is
27 * partitioning hierarchically across a number of servers, and that
28 * partition varies over time as the cluster adjusts the distribution
29 * in order to balance load.
31 * The MDS client is primarily responsible to managing synchronous
32 * metadata requests for operations like open, unlink, and so forth.
33 * If there is a MDS failure, we find out about it when we (possibly
34 * request and) receive a new MDS map, and can resubmit affected
37 * For the most part, though, we take advantage of a lossless
38 * communications channel to the MDS, and do not need to worry about
39 * timing out or resubmitting requests.
41 * We maintain a stateful "session" with each MDS we interact with.
42 * Within each session, we sent periodic heartbeat messages to ensure
43 * any capabilities or leases we have been issues remain valid. If
44 * the session times out and goes stale, our leases and capabilities
45 * are no longer valid.
48 struct ceph_reconnect_state
{
50 struct ceph_pagelist
*pagelist
;
54 static void __wake_requests(struct ceph_mds_client
*mdsc
,
55 struct list_head
*head
);
57 static const struct ceph_connection_operations mds_con_ops
;
65 * parse individual inode info
67 static int parse_reply_info_in(void **p
, void *end
,
68 struct ceph_mds_reply_info_in
*info
,
74 *p
+= sizeof(struct ceph_mds_reply_inode
) +
75 sizeof(*info
->in
->fragtree
.splits
) *
76 le32_to_cpu(info
->in
->fragtree
.nsplits
);
78 ceph_decode_32_safe(p
, end
, info
->symlink_len
, bad
);
79 ceph_decode_need(p
, end
, info
->symlink_len
, bad
);
81 *p
+= info
->symlink_len
;
83 if (features
& CEPH_FEATURE_DIRLAYOUTHASH
)
84 ceph_decode_copy_safe(p
, end
, &info
->dir_layout
,
85 sizeof(info
->dir_layout
), bad
);
87 memset(&info
->dir_layout
, 0, sizeof(info
->dir_layout
));
89 ceph_decode_32_safe(p
, end
, info
->xattr_len
, bad
);
90 ceph_decode_need(p
, end
, info
->xattr_len
, bad
);
91 info
->xattr_data
= *p
;
92 *p
+= info
->xattr_len
;
94 if (features
& CEPH_FEATURE_MDS_INLINE_DATA
) {
95 ceph_decode_64_safe(p
, end
, info
->inline_version
, bad
);
96 ceph_decode_32_safe(p
, end
, info
->inline_len
, bad
);
97 ceph_decode_need(p
, end
, info
->inline_len
, bad
);
98 info
->inline_data
= *p
;
99 *p
+= info
->inline_len
;
101 info
->inline_version
= CEPH_INLINE_NONE
;
103 if (features
& CEPH_FEATURE_FS_FILE_LAYOUT_V2
) {
104 ceph_decode_32_safe(p
, end
, info
->pool_ns_len
, bad
);
105 ceph_decode_need(p
, end
, info
->pool_ns_len
, bad
);
106 *p
+= info
->pool_ns_len
;
108 info
->pool_ns_len
= 0;
117 * parse a normal reply, which may contain a (dir+)dentry and/or a
120 static int parse_reply_info_trace(void **p
, void *end
,
121 struct ceph_mds_reply_info_parsed
*info
,
126 if (info
->head
->is_dentry
) {
127 err
= parse_reply_info_in(p
, end
, &info
->diri
, features
);
131 if (unlikely(*p
+ sizeof(*info
->dirfrag
) > end
))
134 *p
+= sizeof(*info
->dirfrag
) +
135 sizeof(u32
)*le32_to_cpu(info
->dirfrag
->ndist
);
136 if (unlikely(*p
> end
))
139 ceph_decode_32_safe(p
, end
, info
->dname_len
, bad
);
140 ceph_decode_need(p
, end
, info
->dname_len
, bad
);
142 *p
+= info
->dname_len
;
144 *p
+= sizeof(*info
->dlease
);
147 if (info
->head
->is_target
) {
148 err
= parse_reply_info_in(p
, end
, &info
->targeti
, features
);
153 if (unlikely(*p
!= end
))
160 pr_err("problem parsing mds trace %d\n", err
);
165 * parse readdir results
167 static int parse_reply_info_dir(void **p
, void *end
,
168 struct ceph_mds_reply_info_parsed
*info
,
175 if (*p
+ sizeof(*info
->dir_dir
) > end
)
177 *p
+= sizeof(*info
->dir_dir
) +
178 sizeof(u32
)*le32_to_cpu(info
->dir_dir
->ndist
);
182 ceph_decode_need(p
, end
, sizeof(num
) + 2, bad
);
183 num
= ceph_decode_32(p
);
185 u16 flags
= ceph_decode_16(p
);
186 info
->dir_end
= !!(flags
& CEPH_READDIR_FRAG_END
);
187 info
->dir_complete
= !!(flags
& CEPH_READDIR_FRAG_COMPLETE
);
192 BUG_ON(!info
->dir_entries
);
193 if ((unsigned long)(info
->dir_entries
+ num
) >
194 (unsigned long)info
->dir_entries
+ info
->dir_buf_size
) {
195 pr_err("dir contents are larger than expected\n");
202 struct ceph_mds_reply_dir_entry
*rde
= info
->dir_entries
+ i
;
204 ceph_decode_need(p
, end
, sizeof(u32
)*2, bad
);
205 rde
->name_len
= ceph_decode_32(p
);
206 ceph_decode_need(p
, end
, rde
->name_len
, bad
);
209 dout("parsed dir dname '%.*s'\n", rde
->name_len
, rde
->name
);
211 *p
+= sizeof(struct ceph_mds_reply_lease
);
214 err
= parse_reply_info_in(p
, end
, &rde
->inode
, features
);
229 pr_err("problem parsing dir contents %d\n", err
);
234 * parse fcntl F_GETLK results
236 static int parse_reply_info_filelock(void **p
, void *end
,
237 struct ceph_mds_reply_info_parsed
*info
,
240 if (*p
+ sizeof(*info
->filelock_reply
) > end
)
243 info
->filelock_reply
= *p
;
244 *p
+= sizeof(*info
->filelock_reply
);
246 if (unlikely(*p
!= end
))
255 * parse create results
257 static int parse_reply_info_create(void **p
, void *end
,
258 struct ceph_mds_reply_info_parsed
*info
,
261 if (features
& CEPH_FEATURE_REPLY_CREATE_INODE
) {
263 info
->has_create_ino
= false;
265 info
->has_create_ino
= true;
266 info
->ino
= ceph_decode_64(p
);
270 if (unlikely(*p
!= end
))
279 * parse extra results
281 static int parse_reply_info_extra(void **p
, void *end
,
282 struct ceph_mds_reply_info_parsed
*info
,
285 if (info
->head
->op
== CEPH_MDS_OP_GETFILELOCK
)
286 return parse_reply_info_filelock(p
, end
, info
, features
);
287 else if (info
->head
->op
== CEPH_MDS_OP_READDIR
||
288 info
->head
->op
== CEPH_MDS_OP_LSSNAP
)
289 return parse_reply_info_dir(p
, end
, info
, features
);
290 else if (info
->head
->op
== CEPH_MDS_OP_CREATE
)
291 return parse_reply_info_create(p
, end
, info
, features
);
297 * parse entire mds reply
299 static int parse_reply_info(struct ceph_msg
*msg
,
300 struct ceph_mds_reply_info_parsed
*info
,
307 info
->head
= msg
->front
.iov_base
;
308 p
= msg
->front
.iov_base
+ sizeof(struct ceph_mds_reply_head
);
309 end
= p
+ msg
->front
.iov_len
- sizeof(struct ceph_mds_reply_head
);
312 ceph_decode_32_safe(&p
, end
, len
, bad
);
314 ceph_decode_need(&p
, end
, len
, bad
);
315 err
= parse_reply_info_trace(&p
, p
+len
, info
, features
);
321 ceph_decode_32_safe(&p
, end
, len
, bad
);
323 ceph_decode_need(&p
, end
, len
, bad
);
324 err
= parse_reply_info_extra(&p
, p
+len
, info
, features
);
330 ceph_decode_32_safe(&p
, end
, len
, bad
);
331 info
->snapblob_len
= len
;
342 pr_err("mds parse_reply err %d\n", err
);
346 static void destroy_reply_info(struct ceph_mds_reply_info_parsed
*info
)
348 if (!info
->dir_entries
)
350 free_pages((unsigned long)info
->dir_entries
, get_order(info
->dir_buf_size
));
357 const char *ceph_session_state_name(int s
)
360 case CEPH_MDS_SESSION_NEW
: return "new";
361 case CEPH_MDS_SESSION_OPENING
: return "opening";
362 case CEPH_MDS_SESSION_OPEN
: return "open";
363 case CEPH_MDS_SESSION_HUNG
: return "hung";
364 case CEPH_MDS_SESSION_CLOSING
: return "closing";
365 case CEPH_MDS_SESSION_RESTARTING
: return "restarting";
366 case CEPH_MDS_SESSION_RECONNECTING
: return "reconnecting";
367 default: return "???";
371 static struct ceph_mds_session
*get_session(struct ceph_mds_session
*s
)
373 if (atomic_inc_not_zero(&s
->s_ref
)) {
374 dout("mdsc get_session %p %d -> %d\n", s
,
375 atomic_read(&s
->s_ref
)-1, atomic_read(&s
->s_ref
));
378 dout("mdsc get_session %p 0 -- FAIL", s
);
383 void ceph_put_mds_session(struct ceph_mds_session
*s
)
385 dout("mdsc put_session %p %d -> %d\n", s
,
386 atomic_read(&s
->s_ref
), atomic_read(&s
->s_ref
)-1);
387 if (atomic_dec_and_test(&s
->s_ref
)) {
388 if (s
->s_auth
.authorizer
)
389 ceph_auth_destroy_authorizer(s
->s_auth
.authorizer
);
395 * called under mdsc->mutex
397 struct ceph_mds_session
*__ceph_lookup_mds_session(struct ceph_mds_client
*mdsc
,
400 struct ceph_mds_session
*session
;
402 if (mds
>= mdsc
->max_sessions
|| mdsc
->sessions
[mds
] == NULL
)
404 session
= mdsc
->sessions
[mds
];
405 dout("lookup_mds_session %p %d\n", session
,
406 atomic_read(&session
->s_ref
));
407 get_session(session
);
411 static bool __have_session(struct ceph_mds_client
*mdsc
, int mds
)
413 if (mds
>= mdsc
->max_sessions
)
415 return mdsc
->sessions
[mds
];
418 static int __verify_registered_session(struct ceph_mds_client
*mdsc
,
419 struct ceph_mds_session
*s
)
421 if (s
->s_mds
>= mdsc
->max_sessions
||
422 mdsc
->sessions
[s
->s_mds
] != s
)
428 * create+register a new session for given mds.
429 * called under mdsc->mutex.
431 static struct ceph_mds_session
*register_session(struct ceph_mds_client
*mdsc
,
434 struct ceph_mds_session
*s
;
436 if (mds
>= mdsc
->mdsmap
->m_max_mds
)
437 return ERR_PTR(-EINVAL
);
439 s
= kzalloc(sizeof(*s
), GFP_NOFS
);
441 return ERR_PTR(-ENOMEM
);
444 s
->s_state
= CEPH_MDS_SESSION_NEW
;
447 mutex_init(&s
->s_mutex
);
449 ceph_con_init(&s
->s_con
, s
, &mds_con_ops
, &mdsc
->fsc
->client
->msgr
);
451 spin_lock_init(&s
->s_gen_ttl_lock
);
453 s
->s_cap_ttl
= jiffies
- 1;
455 spin_lock_init(&s
->s_cap_lock
);
456 s
->s_renew_requested
= 0;
458 INIT_LIST_HEAD(&s
->s_caps
);
461 atomic_set(&s
->s_ref
, 1);
462 INIT_LIST_HEAD(&s
->s_waiting
);
463 INIT_LIST_HEAD(&s
->s_unsafe
);
464 s
->s_num_cap_releases
= 0;
465 s
->s_cap_reconnect
= 0;
466 s
->s_cap_iterator
= NULL
;
467 INIT_LIST_HEAD(&s
->s_cap_releases
);
468 INIT_LIST_HEAD(&s
->s_cap_flushing
);
469 INIT_LIST_HEAD(&s
->s_cap_snaps_flushing
);
471 dout("register_session mds%d\n", mds
);
472 if (mds
>= mdsc
->max_sessions
) {
473 int newmax
= 1 << get_count_order(mds
+1);
474 struct ceph_mds_session
**sa
;
476 dout("register_session realloc to %d\n", newmax
);
477 sa
= kcalloc(newmax
, sizeof(void *), GFP_NOFS
);
480 if (mdsc
->sessions
) {
481 memcpy(sa
, mdsc
->sessions
,
482 mdsc
->max_sessions
* sizeof(void *));
483 kfree(mdsc
->sessions
);
486 mdsc
->max_sessions
= newmax
;
488 mdsc
->sessions
[mds
] = s
;
489 atomic_inc(&mdsc
->num_sessions
);
490 atomic_inc(&s
->s_ref
); /* one ref to sessions[], one to caller */
492 ceph_con_open(&s
->s_con
, CEPH_ENTITY_TYPE_MDS
, mds
,
493 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
499 return ERR_PTR(-ENOMEM
);
503 * called under mdsc->mutex
505 static void __unregister_session(struct ceph_mds_client
*mdsc
,
506 struct ceph_mds_session
*s
)
508 dout("__unregister_session mds%d %p\n", s
->s_mds
, s
);
509 BUG_ON(mdsc
->sessions
[s
->s_mds
] != s
);
510 mdsc
->sessions
[s
->s_mds
] = NULL
;
511 ceph_con_close(&s
->s_con
);
512 ceph_put_mds_session(s
);
513 atomic_dec(&mdsc
->num_sessions
);
517 * drop session refs in request.
519 * should be last request ref, or hold mdsc->mutex
521 static void put_request_session(struct ceph_mds_request
*req
)
523 if (req
->r_session
) {
524 ceph_put_mds_session(req
->r_session
);
525 req
->r_session
= NULL
;
529 void ceph_mdsc_release_request(struct kref
*kref
)
531 struct ceph_mds_request
*req
= container_of(kref
,
532 struct ceph_mds_request
,
534 destroy_reply_info(&req
->r_reply_info
);
536 ceph_msg_put(req
->r_request
);
538 ceph_msg_put(req
->r_reply
);
540 ceph_put_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
543 if (req
->r_locked_dir
)
544 ceph_put_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
545 iput(req
->r_target_inode
);
548 if (req
->r_old_dentry
)
549 dput(req
->r_old_dentry
);
550 if (req
->r_old_dentry_dir
) {
552 * track (and drop pins for) r_old_dentry_dir
553 * separately, since r_old_dentry's d_parent may have
554 * changed between the dir mutex being dropped and
555 * this request being freed.
557 ceph_put_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
559 iput(req
->r_old_dentry_dir
);
564 ceph_pagelist_release(req
->r_pagelist
);
565 put_request_session(req
);
566 ceph_unreserve_caps(req
->r_mdsc
, &req
->r_caps_reservation
);
570 DEFINE_RB_FUNCS(request
, struct ceph_mds_request
, r_tid
, r_node
)
573 * lookup session, bump ref if found.
575 * called under mdsc->mutex.
577 static struct ceph_mds_request
*
578 lookup_get_request(struct ceph_mds_client
*mdsc
, u64 tid
)
580 struct ceph_mds_request
*req
;
582 req
= lookup_request(&mdsc
->request_tree
, tid
);
584 ceph_mdsc_get_request(req
);
590 * Register an in-flight request, and assign a tid. Link to directory
591 * are modifying (if any).
593 * Called under mdsc->mutex.
595 static void __register_request(struct ceph_mds_client
*mdsc
,
596 struct ceph_mds_request
*req
,
599 req
->r_tid
= ++mdsc
->last_tid
;
601 ceph_reserve_caps(mdsc
, &req
->r_caps_reservation
,
603 dout("__register_request %p tid %lld\n", req
, req
->r_tid
);
604 ceph_mdsc_get_request(req
);
605 insert_request(&mdsc
->request_tree
, req
);
607 req
->r_uid
= current_fsuid();
608 req
->r_gid
= current_fsgid();
610 if (mdsc
->oldest_tid
== 0 && req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
)
611 mdsc
->oldest_tid
= req
->r_tid
;
615 req
->r_unsafe_dir
= dir
;
619 static void __unregister_request(struct ceph_mds_client
*mdsc
,
620 struct ceph_mds_request
*req
)
622 dout("__unregister_request %p tid %lld\n", req
, req
->r_tid
);
624 if (req
->r_tid
== mdsc
->oldest_tid
) {
625 struct rb_node
*p
= rb_next(&req
->r_node
);
626 mdsc
->oldest_tid
= 0;
628 struct ceph_mds_request
*next_req
=
629 rb_entry(p
, struct ceph_mds_request
, r_node
);
630 if (next_req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
) {
631 mdsc
->oldest_tid
= next_req
->r_tid
;
638 erase_request(&mdsc
->request_tree
, req
);
640 if (req
->r_unsafe_dir
&& req
->r_got_unsafe
) {
641 struct ceph_inode_info
*ci
= ceph_inode(req
->r_unsafe_dir
);
642 spin_lock(&ci
->i_unsafe_lock
);
643 list_del_init(&req
->r_unsafe_dir_item
);
644 spin_unlock(&ci
->i_unsafe_lock
);
646 if (req
->r_target_inode
&& req
->r_got_unsafe
) {
647 struct ceph_inode_info
*ci
= ceph_inode(req
->r_target_inode
);
648 spin_lock(&ci
->i_unsafe_lock
);
649 list_del_init(&req
->r_unsafe_target_item
);
650 spin_unlock(&ci
->i_unsafe_lock
);
653 if (req
->r_unsafe_dir
) {
654 iput(req
->r_unsafe_dir
);
655 req
->r_unsafe_dir
= NULL
;
658 complete_all(&req
->r_safe_completion
);
660 ceph_mdsc_put_request(req
);
664 * Choose mds to send request to next. If there is a hint set in the
665 * request (e.g., due to a prior forward hint from the mds), use that.
666 * Otherwise, consult frag tree and/or caps to identify the
667 * appropriate mds. If all else fails, choose randomly.
669 * Called under mdsc->mutex.
671 static struct dentry
*get_nonsnap_parent(struct dentry
*dentry
)
674 * we don't need to worry about protecting the d_parent access
675 * here because we never renaming inside the snapped namespace
676 * except to resplice to another snapdir, and either the old or new
677 * result is a valid result.
679 while (!IS_ROOT(dentry
) && ceph_snap(d_inode(dentry
)) != CEPH_NOSNAP
)
680 dentry
= dentry
->d_parent
;
684 static int __choose_mds(struct ceph_mds_client
*mdsc
,
685 struct ceph_mds_request
*req
)
688 struct ceph_inode_info
*ci
;
689 struct ceph_cap
*cap
;
690 int mode
= req
->r_direct_mode
;
692 u32 hash
= req
->r_direct_hash
;
693 bool is_hash
= req
->r_direct_is_hash
;
696 * is there a specific mds we should try? ignore hint if we have
697 * no session and the mds is not up (active or recovering).
699 if (req
->r_resend_mds
>= 0 &&
700 (__have_session(mdsc
, req
->r_resend_mds
) ||
701 ceph_mdsmap_get_state(mdsc
->mdsmap
, req
->r_resend_mds
) > 0)) {
702 dout("choose_mds using resend_mds mds%d\n",
704 return req
->r_resend_mds
;
707 if (mode
== USE_RANDOM_MDS
)
712 inode
= req
->r_inode
;
713 } else if (req
->r_dentry
) {
714 /* ignore race with rename; old or new d_parent is okay */
715 struct dentry
*parent
= req
->r_dentry
->d_parent
;
716 struct inode
*dir
= d_inode(parent
);
718 if (dir
->i_sb
!= mdsc
->fsc
->sb
) {
720 inode
= d_inode(req
->r_dentry
);
721 } else if (ceph_snap(dir
) != CEPH_NOSNAP
) {
722 /* direct snapped/virtual snapdir requests
723 * based on parent dir inode */
724 struct dentry
*dn
= get_nonsnap_parent(parent
);
726 dout("__choose_mds using nonsnap parent %p\n", inode
);
729 inode
= d_inode(req
->r_dentry
);
730 if (!inode
|| mode
== USE_AUTH_MDS
) {
733 hash
= ceph_dentry_hash(dir
, req
->r_dentry
);
739 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode
, (int)is_hash
,
743 ci
= ceph_inode(inode
);
745 if (is_hash
&& S_ISDIR(inode
->i_mode
)) {
746 struct ceph_inode_frag frag
;
749 ceph_choose_frag(ci
, hash
, &frag
, &found
);
751 if (mode
== USE_ANY_MDS
&& frag
.ndist
> 0) {
754 /* choose a random replica */
755 get_random_bytes(&r
, 1);
758 dout("choose_mds %p %llx.%llx "
759 "frag %u mds%d (%d/%d)\n",
760 inode
, ceph_vinop(inode
),
763 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
764 CEPH_MDS_STATE_ACTIVE
)
768 /* since this file/dir wasn't known to be
769 * replicated, then we want to look for the
770 * authoritative mds. */
773 /* choose auth mds */
775 dout("choose_mds %p %llx.%llx "
776 "frag %u mds%d (auth)\n",
777 inode
, ceph_vinop(inode
), frag
.frag
, mds
);
778 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
779 CEPH_MDS_STATE_ACTIVE
)
785 spin_lock(&ci
->i_ceph_lock
);
787 if (mode
== USE_AUTH_MDS
)
788 cap
= ci
->i_auth_cap
;
789 if (!cap
&& !RB_EMPTY_ROOT(&ci
->i_caps
))
790 cap
= rb_entry(rb_first(&ci
->i_caps
), struct ceph_cap
, ci_node
);
792 spin_unlock(&ci
->i_ceph_lock
);
795 mds
= cap
->session
->s_mds
;
796 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
797 inode
, ceph_vinop(inode
), mds
,
798 cap
== ci
->i_auth_cap
? "auth " : "", cap
);
799 spin_unlock(&ci
->i_ceph_lock
);
803 mds
= ceph_mdsmap_get_random_mds(mdsc
->mdsmap
);
804 dout("choose_mds chose random mds%d\n", mds
);
812 static struct ceph_msg
*create_session_msg(u32 op
, u64 seq
)
814 struct ceph_msg
*msg
;
815 struct ceph_mds_session_head
*h
;
817 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
), GFP_NOFS
,
820 pr_err("create_session_msg ENOMEM creating msg\n");
823 h
= msg
->front
.iov_base
;
824 h
->op
= cpu_to_le32(op
);
825 h
->seq
= cpu_to_le64(seq
);
831 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
832 * to include additional client metadata fields.
834 static struct ceph_msg
*create_session_open_msg(struct ceph_mds_client
*mdsc
, u64 seq
)
836 struct ceph_msg
*msg
;
837 struct ceph_mds_session_head
*h
;
839 int metadata_bytes
= 0;
840 int metadata_key_count
= 0;
841 struct ceph_options
*opt
= mdsc
->fsc
->client
->options
;
842 struct ceph_mount_options
*fsopt
= mdsc
->fsc
->mount_options
;
845 const char* metadata
[][2] = {
846 {"hostname", utsname()->nodename
},
847 {"kernel_version", utsname()->release
},
848 {"entity_id", opt
->name
? : ""},
849 {"root", fsopt
->server_path
? : "/"},
853 /* Calculate serialized length of metadata */
854 metadata_bytes
= 4; /* map length */
855 for (i
= 0; metadata
[i
][0] != NULL
; ++i
) {
856 metadata_bytes
+= 8 + strlen(metadata
[i
][0]) +
857 strlen(metadata
[i
][1]);
858 metadata_key_count
++;
861 /* Allocate the message */
862 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
) + metadata_bytes
,
865 pr_err("create_session_msg ENOMEM creating msg\n");
868 h
= msg
->front
.iov_base
;
869 h
->op
= cpu_to_le32(CEPH_SESSION_REQUEST_OPEN
);
870 h
->seq
= cpu_to_le64(seq
);
873 * Serialize client metadata into waiting buffer space, using
874 * the format that userspace expects for map<string, string>
876 * ClientSession messages with metadata are v2
878 msg
->hdr
.version
= cpu_to_le16(2);
879 msg
->hdr
.compat_version
= cpu_to_le16(1);
881 /* The write pointer, following the session_head structure */
882 p
= msg
->front
.iov_base
+ sizeof(*h
);
884 /* Number of entries in the map */
885 ceph_encode_32(&p
, metadata_key_count
);
887 /* Two length-prefixed strings for each entry in the map */
888 for (i
= 0; metadata
[i
][0] != NULL
; ++i
) {
889 size_t const key_len
= strlen(metadata
[i
][0]);
890 size_t const val_len
= strlen(metadata
[i
][1]);
892 ceph_encode_32(&p
, key_len
);
893 memcpy(p
, metadata
[i
][0], key_len
);
895 ceph_encode_32(&p
, val_len
);
896 memcpy(p
, metadata
[i
][1], val_len
);
904 * send session open request.
906 * called under mdsc->mutex
908 static int __open_session(struct ceph_mds_client
*mdsc
,
909 struct ceph_mds_session
*session
)
911 struct ceph_msg
*msg
;
913 int mds
= session
->s_mds
;
915 /* wait for mds to go active? */
916 mstate
= ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
);
917 dout("open_session to mds%d (%s)\n", mds
,
918 ceph_mds_state_name(mstate
));
919 session
->s_state
= CEPH_MDS_SESSION_OPENING
;
920 session
->s_renew_requested
= jiffies
;
922 /* send connect message */
923 msg
= create_session_open_msg(mdsc
, session
->s_seq
);
926 ceph_con_send(&session
->s_con
, msg
);
931 * open sessions for any export targets for the given mds
933 * called under mdsc->mutex
935 static struct ceph_mds_session
*
936 __open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
938 struct ceph_mds_session
*session
;
940 session
= __ceph_lookup_mds_session(mdsc
, target
);
942 session
= register_session(mdsc
, target
);
946 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
947 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
948 __open_session(mdsc
, session
);
953 struct ceph_mds_session
*
954 ceph_mdsc_open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
956 struct ceph_mds_session
*session
;
958 dout("open_export_target_session to mds%d\n", target
);
960 mutex_lock(&mdsc
->mutex
);
961 session
= __open_export_target_session(mdsc
, target
);
962 mutex_unlock(&mdsc
->mutex
);
967 static void __open_export_target_sessions(struct ceph_mds_client
*mdsc
,
968 struct ceph_mds_session
*session
)
970 struct ceph_mds_info
*mi
;
971 struct ceph_mds_session
*ts
;
972 int i
, mds
= session
->s_mds
;
974 if (mds
>= mdsc
->mdsmap
->m_max_mds
)
977 mi
= &mdsc
->mdsmap
->m_info
[mds
];
978 dout("open_export_target_sessions for mds%d (%d targets)\n",
979 session
->s_mds
, mi
->num_export_targets
);
981 for (i
= 0; i
< mi
->num_export_targets
; i
++) {
982 ts
= __open_export_target_session(mdsc
, mi
->export_targets
[i
]);
984 ceph_put_mds_session(ts
);
988 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client
*mdsc
,
989 struct ceph_mds_session
*session
)
991 mutex_lock(&mdsc
->mutex
);
992 __open_export_target_sessions(mdsc
, session
);
993 mutex_unlock(&mdsc
->mutex
);
1000 /* caller holds s_cap_lock, we drop it */
1001 static void cleanup_cap_releases(struct ceph_mds_client
*mdsc
,
1002 struct ceph_mds_session
*session
)
1003 __releases(session
->s_cap_lock
)
1005 LIST_HEAD(tmp_list
);
1006 list_splice_init(&session
->s_cap_releases
, &tmp_list
);
1007 session
->s_num_cap_releases
= 0;
1008 spin_unlock(&session
->s_cap_lock
);
1010 dout("cleanup_cap_releases mds%d\n", session
->s_mds
);
1011 while (!list_empty(&tmp_list
)) {
1012 struct ceph_cap
*cap
;
1013 /* zero out the in-progress message */
1014 cap
= list_first_entry(&tmp_list
,
1015 struct ceph_cap
, session_caps
);
1016 list_del(&cap
->session_caps
);
1017 ceph_put_cap(mdsc
, cap
);
1021 static void cleanup_session_requests(struct ceph_mds_client
*mdsc
,
1022 struct ceph_mds_session
*session
)
1024 struct ceph_mds_request
*req
;
1027 dout("cleanup_session_requests mds%d\n", session
->s_mds
);
1028 mutex_lock(&mdsc
->mutex
);
1029 while (!list_empty(&session
->s_unsafe
)) {
1030 req
= list_first_entry(&session
->s_unsafe
,
1031 struct ceph_mds_request
, r_unsafe_item
);
1032 list_del_init(&req
->r_unsafe_item
);
1033 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1035 __unregister_request(mdsc
, req
);
1037 /* zero r_attempts, so kick_requests() will re-send requests */
1038 p
= rb_first(&mdsc
->request_tree
);
1040 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
1042 if (req
->r_session
&&
1043 req
->r_session
->s_mds
== session
->s_mds
)
1044 req
->r_attempts
= 0;
1046 mutex_unlock(&mdsc
->mutex
);
1050 * Helper to safely iterate over all caps associated with a session, with
1051 * special care taken to handle a racing __ceph_remove_cap().
1053 * Caller must hold session s_mutex.
1055 static int iterate_session_caps(struct ceph_mds_session
*session
,
1056 int (*cb
)(struct inode
*, struct ceph_cap
*,
1059 struct list_head
*p
;
1060 struct ceph_cap
*cap
;
1061 struct inode
*inode
, *last_inode
= NULL
;
1062 struct ceph_cap
*old_cap
= NULL
;
1065 dout("iterate_session_caps %p mds%d\n", session
, session
->s_mds
);
1066 spin_lock(&session
->s_cap_lock
);
1067 p
= session
->s_caps
.next
;
1068 while (p
!= &session
->s_caps
) {
1069 cap
= list_entry(p
, struct ceph_cap
, session_caps
);
1070 inode
= igrab(&cap
->ci
->vfs_inode
);
1075 session
->s_cap_iterator
= cap
;
1076 spin_unlock(&session
->s_cap_lock
);
1083 ceph_put_cap(session
->s_mdsc
, old_cap
);
1087 ret
= cb(inode
, cap
, arg
);
1090 spin_lock(&session
->s_cap_lock
);
1092 if (cap
->ci
== NULL
) {
1093 dout("iterate_session_caps finishing cap %p removal\n",
1095 BUG_ON(cap
->session
!= session
);
1096 cap
->session
= NULL
;
1097 list_del_init(&cap
->session_caps
);
1098 session
->s_nr_caps
--;
1099 if (cap
->queue_release
) {
1100 list_add_tail(&cap
->session_caps
,
1101 &session
->s_cap_releases
);
1102 session
->s_num_cap_releases
++;
1104 old_cap
= cap
; /* put_cap it w/o locks held */
1112 session
->s_cap_iterator
= NULL
;
1113 spin_unlock(&session
->s_cap_lock
);
1117 ceph_put_cap(session
->s_mdsc
, old_cap
);
1122 static int remove_session_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1125 struct ceph_fs_client
*fsc
= (struct ceph_fs_client
*)arg
;
1126 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1127 LIST_HEAD(to_remove
);
1129 bool invalidate
= false;
1131 dout("removing cap %p, ci is %p, inode is %p\n",
1132 cap
, ci
, &ci
->vfs_inode
);
1133 spin_lock(&ci
->i_ceph_lock
);
1134 __ceph_remove_cap(cap
, false);
1135 if (!ci
->i_auth_cap
) {
1136 struct ceph_cap_flush
*cf
;
1137 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
1139 ci
->i_ceph_flags
|= CEPH_I_CAP_DROPPED
;
1141 if (ci
->i_wrbuffer_ref
> 0 &&
1142 ACCESS_ONCE(fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
1146 struct rb_node
*n
= rb_first(&ci
->i_cap_flush_tree
);
1149 cf
= rb_entry(n
, struct ceph_cap_flush
, i_node
);
1150 rb_erase(&cf
->i_node
, &ci
->i_cap_flush_tree
);
1151 list_add(&cf
->list
, &to_remove
);
1154 spin_lock(&mdsc
->cap_dirty_lock
);
1156 list_for_each_entry(cf
, &to_remove
, list
)
1157 rb_erase(&cf
->g_node
, &mdsc
->cap_flush_tree
);
1159 if (!list_empty(&ci
->i_dirty_item
)) {
1160 pr_warn_ratelimited(
1161 " dropping dirty %s state for %p %lld\n",
1162 ceph_cap_string(ci
->i_dirty_caps
),
1163 inode
, ceph_ino(inode
));
1164 ci
->i_dirty_caps
= 0;
1165 list_del_init(&ci
->i_dirty_item
);
1168 if (!list_empty(&ci
->i_flushing_item
)) {
1169 pr_warn_ratelimited(
1170 " dropping dirty+flushing %s state for %p %lld\n",
1171 ceph_cap_string(ci
->i_flushing_caps
),
1172 inode
, ceph_ino(inode
));
1173 ci
->i_flushing_caps
= 0;
1174 list_del_init(&ci
->i_flushing_item
);
1175 mdsc
->num_cap_flushing
--;
1178 spin_unlock(&mdsc
->cap_dirty_lock
);
1180 if (!ci
->i_dirty_caps
&& ci
->i_prealloc_cap_flush
) {
1181 list_add(&ci
->i_prealloc_cap_flush
->list
, &to_remove
);
1182 ci
->i_prealloc_cap_flush
= NULL
;
1185 spin_unlock(&ci
->i_ceph_lock
);
1186 while (!list_empty(&to_remove
)) {
1187 struct ceph_cap_flush
*cf
;
1188 cf
= list_first_entry(&to_remove
,
1189 struct ceph_cap_flush
, list
);
1190 list_del(&cf
->list
);
1191 ceph_free_cap_flush(cf
);
1194 wake_up_all(&ci
->i_cap_wq
);
1196 ceph_queue_invalidate(inode
);
1203 * caller must hold session s_mutex
1205 static void remove_session_caps(struct ceph_mds_session
*session
)
1207 struct ceph_fs_client
*fsc
= session
->s_mdsc
->fsc
;
1208 struct super_block
*sb
= fsc
->sb
;
1209 dout("remove_session_caps on %p\n", session
);
1210 iterate_session_caps(session
, remove_session_caps_cb
, fsc
);
1212 spin_lock(&session
->s_cap_lock
);
1213 if (session
->s_nr_caps
> 0) {
1214 struct inode
*inode
;
1215 struct ceph_cap
*cap
, *prev
= NULL
;
1216 struct ceph_vino vino
;
1218 * iterate_session_caps() skips inodes that are being
1219 * deleted, we need to wait until deletions are complete.
1220 * __wait_on_freeing_inode() is designed for the job,
1221 * but it is not exported, so use lookup inode function
1224 while (!list_empty(&session
->s_caps
)) {
1225 cap
= list_entry(session
->s_caps
.next
,
1226 struct ceph_cap
, session_caps
);
1230 vino
= cap
->ci
->i_vino
;
1231 spin_unlock(&session
->s_cap_lock
);
1233 inode
= ceph_find_inode(sb
, vino
);
1236 spin_lock(&session
->s_cap_lock
);
1240 // drop cap expires and unlock s_cap_lock
1241 cleanup_cap_releases(session
->s_mdsc
, session
);
1243 BUG_ON(session
->s_nr_caps
> 0);
1244 BUG_ON(!list_empty(&session
->s_cap_flushing
));
1248 * wake up any threads waiting on this session's caps. if the cap is
1249 * old (didn't get renewed on the client reconnect), remove it now.
1251 * caller must hold s_mutex.
1253 static int wake_up_session_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1256 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1258 wake_up_all(&ci
->i_cap_wq
);
1260 spin_lock(&ci
->i_ceph_lock
);
1261 ci
->i_wanted_max_size
= 0;
1262 ci
->i_requested_max_size
= 0;
1263 spin_unlock(&ci
->i_ceph_lock
);
1268 static void wake_up_session_caps(struct ceph_mds_session
*session
,
1271 dout("wake_up_session_caps %p mds%d\n", session
, session
->s_mds
);
1272 iterate_session_caps(session
, wake_up_session_cb
,
1273 (void *)(unsigned long)reconnect
);
1277 * Send periodic message to MDS renewing all currently held caps. The
1278 * ack will reset the expiration for all caps from this session.
1280 * caller holds s_mutex
1282 static int send_renew_caps(struct ceph_mds_client
*mdsc
,
1283 struct ceph_mds_session
*session
)
1285 struct ceph_msg
*msg
;
1288 if (time_after_eq(jiffies
, session
->s_cap_ttl
) &&
1289 time_after_eq(session
->s_cap_ttl
, session
->s_renew_requested
))
1290 pr_info("mds%d caps stale\n", session
->s_mds
);
1291 session
->s_renew_requested
= jiffies
;
1293 /* do not try to renew caps until a recovering mds has reconnected
1294 * with its clients. */
1295 state
= ceph_mdsmap_get_state(mdsc
->mdsmap
, session
->s_mds
);
1296 if (state
< CEPH_MDS_STATE_RECONNECT
) {
1297 dout("send_renew_caps ignoring mds%d (%s)\n",
1298 session
->s_mds
, ceph_mds_state_name(state
));
1302 dout("send_renew_caps to mds%d (%s)\n", session
->s_mds
,
1303 ceph_mds_state_name(state
));
1304 msg
= create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS
,
1305 ++session
->s_renew_seq
);
1308 ceph_con_send(&session
->s_con
, msg
);
1312 static int send_flushmsg_ack(struct ceph_mds_client
*mdsc
,
1313 struct ceph_mds_session
*session
, u64 seq
)
1315 struct ceph_msg
*msg
;
1317 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1318 session
->s_mds
, ceph_session_state_name(session
->s_state
), seq
);
1319 msg
= create_session_msg(CEPH_SESSION_FLUSHMSG_ACK
, seq
);
1322 ceph_con_send(&session
->s_con
, msg
);
1328 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1330 * Called under session->s_mutex
1332 static void renewed_caps(struct ceph_mds_client
*mdsc
,
1333 struct ceph_mds_session
*session
, int is_renew
)
1338 spin_lock(&session
->s_cap_lock
);
1339 was_stale
= is_renew
&& time_after_eq(jiffies
, session
->s_cap_ttl
);
1341 session
->s_cap_ttl
= session
->s_renew_requested
+
1342 mdsc
->mdsmap
->m_session_timeout
*HZ
;
1345 if (time_before(jiffies
, session
->s_cap_ttl
)) {
1346 pr_info("mds%d caps renewed\n", session
->s_mds
);
1349 pr_info("mds%d caps still stale\n", session
->s_mds
);
1352 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1353 session
->s_mds
, session
->s_cap_ttl
, was_stale
? "stale" : "fresh",
1354 time_before(jiffies
, session
->s_cap_ttl
) ? "stale" : "fresh");
1355 spin_unlock(&session
->s_cap_lock
);
1358 wake_up_session_caps(session
, 0);
1362 * send a session close request
1364 static int request_close_session(struct ceph_mds_client
*mdsc
,
1365 struct ceph_mds_session
*session
)
1367 struct ceph_msg
*msg
;
1369 dout("request_close_session mds%d state %s seq %lld\n",
1370 session
->s_mds
, ceph_session_state_name(session
->s_state
),
1372 msg
= create_session_msg(CEPH_SESSION_REQUEST_CLOSE
, session
->s_seq
);
1375 ceph_con_send(&session
->s_con
, msg
);
1380 * Called with s_mutex held.
1382 static int __close_session(struct ceph_mds_client
*mdsc
,
1383 struct ceph_mds_session
*session
)
1385 if (session
->s_state
>= CEPH_MDS_SESSION_CLOSING
)
1387 session
->s_state
= CEPH_MDS_SESSION_CLOSING
;
1388 return request_close_session(mdsc
, session
);
1392 * Trim old(er) caps.
1394 * Because we can't cache an inode without one or more caps, we do
1395 * this indirectly: if a cap is unused, we prune its aliases, at which
1396 * point the inode will hopefully get dropped to.
1398 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1399 * memory pressure from the MDS, though, so it needn't be perfect.
1401 static int trim_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
, void *arg
)
1403 struct ceph_mds_session
*session
= arg
;
1404 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1405 int used
, wanted
, oissued
, mine
;
1407 if (session
->s_trim_caps
<= 0)
1410 spin_lock(&ci
->i_ceph_lock
);
1411 mine
= cap
->issued
| cap
->implemented
;
1412 used
= __ceph_caps_used(ci
);
1413 wanted
= __ceph_caps_file_wanted(ci
);
1414 oissued
= __ceph_caps_issued_other(ci
, cap
);
1416 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1417 inode
, cap
, ceph_cap_string(mine
), ceph_cap_string(oissued
),
1418 ceph_cap_string(used
), ceph_cap_string(wanted
));
1419 if (cap
== ci
->i_auth_cap
) {
1420 if (ci
->i_dirty_caps
|| ci
->i_flushing_caps
||
1421 !list_empty(&ci
->i_cap_snaps
))
1423 if ((used
| wanted
) & CEPH_CAP_ANY_WR
)
1426 /* The inode has cached pages, but it's no longer used.
1427 * we can safely drop it */
1428 if (wanted
== 0 && used
== CEPH_CAP_FILE_CACHE
&&
1429 !(oissued
& CEPH_CAP_FILE_CACHE
)) {
1433 if ((used
| wanted
) & ~oissued
& mine
)
1434 goto out
; /* we need these caps */
1436 session
->s_trim_caps
--;
1438 /* we aren't the only cap.. just remove us */
1439 __ceph_remove_cap(cap
, true);
1441 /* try dropping referring dentries */
1442 spin_unlock(&ci
->i_ceph_lock
);
1443 d_prune_aliases(inode
);
1444 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1445 inode
, cap
, atomic_read(&inode
->i_count
));
1450 spin_unlock(&ci
->i_ceph_lock
);
1455 * Trim session cap count down to some max number.
1457 static int trim_caps(struct ceph_mds_client
*mdsc
,
1458 struct ceph_mds_session
*session
,
1461 int trim_caps
= session
->s_nr_caps
- max_caps
;
1463 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1464 session
->s_mds
, session
->s_nr_caps
, max_caps
, trim_caps
);
1465 if (trim_caps
> 0) {
1466 session
->s_trim_caps
= trim_caps
;
1467 iterate_session_caps(session
, trim_caps_cb
, session
);
1468 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1469 session
->s_mds
, session
->s_nr_caps
, max_caps
,
1470 trim_caps
- session
->s_trim_caps
);
1471 session
->s_trim_caps
= 0;
1474 ceph_send_cap_releases(mdsc
, session
);
1478 static int check_capsnap_flush(struct ceph_inode_info
*ci
,
1482 spin_lock(&ci
->i_ceph_lock
);
1483 if (want_snap_seq
> 0 && !list_empty(&ci
->i_cap_snaps
)) {
1484 struct ceph_cap_snap
*capsnap
=
1485 list_first_entry(&ci
->i_cap_snaps
,
1486 struct ceph_cap_snap
, ci_item
);
1487 ret
= capsnap
->follows
>= want_snap_seq
;
1489 spin_unlock(&ci
->i_ceph_lock
);
1493 static int check_caps_flush(struct ceph_mds_client
*mdsc
,
1497 struct ceph_cap_flush
*cf
;
1500 spin_lock(&mdsc
->cap_dirty_lock
);
1501 n
= rb_first(&mdsc
->cap_flush_tree
);
1502 cf
= n
? rb_entry(n
, struct ceph_cap_flush
, g_node
) : NULL
;
1503 if (cf
&& cf
->tid
<= want_flush_tid
) {
1504 dout("check_caps_flush still flushing tid %llu <= %llu\n",
1505 cf
->tid
, want_flush_tid
);
1508 spin_unlock(&mdsc
->cap_dirty_lock
);
1513 * flush all dirty inode data to disk.
1515 * returns true if we've flushed through want_flush_tid
1517 static void wait_caps_flush(struct ceph_mds_client
*mdsc
,
1518 u64 want_flush_tid
, u64 want_snap_seq
)
1522 dout("check_caps_flush want %llu snap want %llu\n",
1523 want_flush_tid
, want_snap_seq
);
1524 mutex_lock(&mdsc
->mutex
);
1525 for (mds
= 0; mds
< mdsc
->max_sessions
; ) {
1526 struct ceph_mds_session
*session
= mdsc
->sessions
[mds
];
1527 struct inode
*inode
= NULL
;
1533 get_session(session
);
1534 mutex_unlock(&mdsc
->mutex
);
1536 mutex_lock(&session
->s_mutex
);
1537 if (!list_empty(&session
->s_cap_snaps_flushing
)) {
1538 struct ceph_cap_snap
*capsnap
=
1539 list_first_entry(&session
->s_cap_snaps_flushing
,
1540 struct ceph_cap_snap
,
1542 struct ceph_inode_info
*ci
= capsnap
->ci
;
1543 if (!check_capsnap_flush(ci
, want_snap_seq
)) {
1544 dout("check_cap_flush still flushing snap %p "
1545 "follows %lld <= %lld to mds%d\n",
1546 &ci
->vfs_inode
, capsnap
->follows
,
1547 want_snap_seq
, mds
);
1548 inode
= igrab(&ci
->vfs_inode
);
1551 mutex_unlock(&session
->s_mutex
);
1552 ceph_put_mds_session(session
);
1555 wait_event(mdsc
->cap_flushing_wq
,
1556 check_capsnap_flush(ceph_inode(inode
),
1563 mutex_lock(&mdsc
->mutex
);
1565 mutex_unlock(&mdsc
->mutex
);
1567 wait_event(mdsc
->cap_flushing_wq
,
1568 check_caps_flush(mdsc
, want_flush_tid
));
1570 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid
);
1574 * called under s_mutex
1576 void ceph_send_cap_releases(struct ceph_mds_client
*mdsc
,
1577 struct ceph_mds_session
*session
)
1579 struct ceph_msg
*msg
= NULL
;
1580 struct ceph_mds_cap_release
*head
;
1581 struct ceph_mds_cap_item
*item
;
1582 struct ceph_cap
*cap
;
1583 LIST_HEAD(tmp_list
);
1584 int num_cap_releases
;
1586 spin_lock(&session
->s_cap_lock
);
1588 list_splice_init(&session
->s_cap_releases
, &tmp_list
);
1589 num_cap_releases
= session
->s_num_cap_releases
;
1590 session
->s_num_cap_releases
= 0;
1591 spin_unlock(&session
->s_cap_lock
);
1593 while (!list_empty(&tmp_list
)) {
1595 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE
,
1596 PAGE_SIZE
, GFP_NOFS
, false);
1599 head
= msg
->front
.iov_base
;
1600 head
->num
= cpu_to_le32(0);
1601 msg
->front
.iov_len
= sizeof(*head
);
1603 cap
= list_first_entry(&tmp_list
, struct ceph_cap
,
1605 list_del(&cap
->session_caps
);
1608 head
= msg
->front
.iov_base
;
1609 le32_add_cpu(&head
->num
, 1);
1610 item
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1611 item
->ino
= cpu_to_le64(cap
->cap_ino
);
1612 item
->cap_id
= cpu_to_le64(cap
->cap_id
);
1613 item
->migrate_seq
= cpu_to_le32(cap
->mseq
);
1614 item
->seq
= cpu_to_le32(cap
->issue_seq
);
1615 msg
->front
.iov_len
+= sizeof(*item
);
1617 ceph_put_cap(mdsc
, cap
);
1619 if (le32_to_cpu(head
->num
) == CEPH_CAPS_PER_RELEASE
) {
1620 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1621 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1622 ceph_con_send(&session
->s_con
, msg
);
1627 BUG_ON(num_cap_releases
!= 0);
1629 spin_lock(&session
->s_cap_lock
);
1630 if (!list_empty(&session
->s_cap_releases
))
1632 spin_unlock(&session
->s_cap_lock
);
1635 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1636 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1637 ceph_con_send(&session
->s_con
, msg
);
1641 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1643 spin_lock(&session
->s_cap_lock
);
1644 list_splice(&tmp_list
, &session
->s_cap_releases
);
1645 session
->s_num_cap_releases
+= num_cap_releases
;
1646 spin_unlock(&session
->s_cap_lock
);
1653 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request
*req
,
1656 struct ceph_inode_info
*ci
= ceph_inode(dir
);
1657 struct ceph_mds_reply_info_parsed
*rinfo
= &req
->r_reply_info
;
1658 struct ceph_mount_options
*opt
= req
->r_mdsc
->fsc
->mount_options
;
1659 size_t size
= sizeof(struct ceph_mds_reply_dir_entry
);
1660 int order
, num_entries
;
1662 spin_lock(&ci
->i_ceph_lock
);
1663 num_entries
= ci
->i_files
+ ci
->i_subdirs
;
1664 spin_unlock(&ci
->i_ceph_lock
);
1665 num_entries
= max(num_entries
, 1);
1666 num_entries
= min(num_entries
, opt
->max_readdir
);
1668 order
= get_order(size
* num_entries
);
1669 while (order
>= 0) {
1670 rinfo
->dir_entries
= (void*)__get_free_pages(GFP_KERNEL
|
1673 if (rinfo
->dir_entries
)
1677 if (!rinfo
->dir_entries
)
1680 num_entries
= (PAGE_SIZE
<< order
) / size
;
1681 num_entries
= min(num_entries
, opt
->max_readdir
);
1683 rinfo
->dir_buf_size
= PAGE_SIZE
<< order
;
1684 req
->r_num_caps
= num_entries
+ 1;
1685 req
->r_args
.readdir
.max_entries
= cpu_to_le32(num_entries
);
1686 req
->r_args
.readdir
.max_bytes
= cpu_to_le32(opt
->max_readdir_bytes
);
1691 * Create an mds request.
1693 struct ceph_mds_request
*
1694 ceph_mdsc_create_request(struct ceph_mds_client
*mdsc
, int op
, int mode
)
1696 struct ceph_mds_request
*req
= kzalloc(sizeof(*req
), GFP_NOFS
);
1699 return ERR_PTR(-ENOMEM
);
1701 mutex_init(&req
->r_fill_mutex
);
1703 req
->r_started
= jiffies
;
1704 req
->r_resend_mds
= -1;
1705 INIT_LIST_HEAD(&req
->r_unsafe_dir_item
);
1706 INIT_LIST_HEAD(&req
->r_unsafe_target_item
);
1708 kref_init(&req
->r_kref
);
1709 RB_CLEAR_NODE(&req
->r_node
);
1710 INIT_LIST_HEAD(&req
->r_wait
);
1711 init_completion(&req
->r_completion
);
1712 init_completion(&req
->r_safe_completion
);
1713 INIT_LIST_HEAD(&req
->r_unsafe_item
);
1715 req
->r_stamp
= current_fs_time(mdsc
->fsc
->sb
);
1718 req
->r_direct_mode
= mode
;
1723 * return oldest (lowest) request, tid in request tree, 0 if none.
1725 * called under mdsc->mutex.
1727 static struct ceph_mds_request
*__get_oldest_req(struct ceph_mds_client
*mdsc
)
1729 if (RB_EMPTY_ROOT(&mdsc
->request_tree
))
1731 return rb_entry(rb_first(&mdsc
->request_tree
),
1732 struct ceph_mds_request
, r_node
);
1735 static inline u64
__get_oldest_tid(struct ceph_mds_client
*mdsc
)
1737 return mdsc
->oldest_tid
;
1741 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1742 * on build_path_from_dentry in fs/cifs/dir.c.
1744 * If @stop_on_nosnap, generate path relative to the first non-snapped
1747 * Encode hidden .snap dirs as a double /, i.e.
1748 * foo/.snap/bar -> foo//bar
1750 char *ceph_mdsc_build_path(struct dentry
*dentry
, int *plen
, u64
*base
,
1753 struct dentry
*temp
;
1759 return ERR_PTR(-EINVAL
);
1763 seq
= read_seqbegin(&rename_lock
);
1765 for (temp
= dentry
; !IS_ROOT(temp
);) {
1766 struct inode
*inode
= d_inode(temp
);
1767 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
)
1768 len
++; /* slash only */
1769 else if (stop_on_nosnap
&& inode
&&
1770 ceph_snap(inode
) == CEPH_NOSNAP
)
1773 len
+= 1 + temp
->d_name
.len
;
1774 temp
= temp
->d_parent
;
1778 len
--; /* no leading '/' */
1780 path
= kmalloc(len
+1, GFP_NOFS
);
1782 return ERR_PTR(-ENOMEM
);
1784 path
[pos
] = 0; /* trailing null */
1786 for (temp
= dentry
; !IS_ROOT(temp
) && pos
!= 0; ) {
1787 struct inode
*inode
;
1789 spin_lock(&temp
->d_lock
);
1790 inode
= d_inode(temp
);
1791 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
) {
1792 dout("build_path path+%d: %p SNAPDIR\n",
1794 } else if (stop_on_nosnap
&& inode
&&
1795 ceph_snap(inode
) == CEPH_NOSNAP
) {
1796 spin_unlock(&temp
->d_lock
);
1799 pos
-= temp
->d_name
.len
;
1801 spin_unlock(&temp
->d_lock
);
1804 strncpy(path
+ pos
, temp
->d_name
.name
,
1807 spin_unlock(&temp
->d_lock
);
1810 temp
= temp
->d_parent
;
1813 if (pos
!= 0 || read_seqretry(&rename_lock
, seq
)) {
1814 pr_err("build_path did not end path lookup where "
1815 "expected, namelen is %d, pos is %d\n", len
, pos
);
1816 /* presumably this is only possible if racing with a
1817 rename of one of the parent directories (we can not
1818 lock the dentries above us to prevent this, but
1819 retrying should be harmless) */
1824 *base
= ceph_ino(d_inode(temp
));
1826 dout("build_path on %p %d built %llx '%.*s'\n",
1827 dentry
, d_count(dentry
), *base
, len
, path
);
1831 static int build_dentry_path(struct dentry
*dentry
,
1832 const char **ppath
, int *ppathlen
, u64
*pino
,
1837 if (ceph_snap(d_inode(dentry
->d_parent
)) == CEPH_NOSNAP
) {
1838 *pino
= ceph_ino(d_inode(dentry
->d_parent
));
1839 *ppath
= dentry
->d_name
.name
;
1840 *ppathlen
= dentry
->d_name
.len
;
1843 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1845 return PTR_ERR(path
);
1851 static int build_inode_path(struct inode
*inode
,
1852 const char **ppath
, int *ppathlen
, u64
*pino
,
1855 struct dentry
*dentry
;
1858 if (ceph_snap(inode
) == CEPH_NOSNAP
) {
1859 *pino
= ceph_ino(inode
);
1863 dentry
= d_find_alias(inode
);
1864 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1867 return PTR_ERR(path
);
1874 * request arguments may be specified via an inode *, a dentry *, or
1875 * an explicit ino+path.
1877 static int set_request_path_attr(struct inode
*rinode
, struct dentry
*rdentry
,
1878 const char *rpath
, u64 rino
,
1879 const char **ppath
, int *pathlen
,
1880 u64
*ino
, int *freepath
)
1885 r
= build_inode_path(rinode
, ppath
, pathlen
, ino
, freepath
);
1886 dout(" inode %p %llx.%llx\n", rinode
, ceph_ino(rinode
),
1888 } else if (rdentry
) {
1889 r
= build_dentry_path(rdentry
, ppath
, pathlen
, ino
, freepath
);
1890 dout(" dentry %p %llx/%.*s\n", rdentry
, *ino
, *pathlen
,
1892 } else if (rpath
|| rino
) {
1895 *pathlen
= rpath
? strlen(rpath
) : 0;
1896 dout(" path %.*s\n", *pathlen
, rpath
);
1903 * called under mdsc->mutex
1905 static struct ceph_msg
*create_request_message(struct ceph_mds_client
*mdsc
,
1906 struct ceph_mds_request
*req
,
1907 int mds
, bool drop_cap_releases
)
1909 struct ceph_msg
*msg
;
1910 struct ceph_mds_request_head
*head
;
1911 const char *path1
= NULL
;
1912 const char *path2
= NULL
;
1913 u64 ino1
= 0, ino2
= 0;
1914 int pathlen1
= 0, pathlen2
= 0;
1915 int freepath1
= 0, freepath2
= 0;
1921 ret
= set_request_path_attr(req
->r_inode
, req
->r_dentry
,
1922 req
->r_path1
, req
->r_ino1
.ino
,
1923 &path1
, &pathlen1
, &ino1
, &freepath1
);
1929 ret
= set_request_path_attr(NULL
, req
->r_old_dentry
,
1930 req
->r_path2
, req
->r_ino2
.ino
,
1931 &path2
, &pathlen2
, &ino2
, &freepath2
);
1937 len
= sizeof(*head
) +
1938 pathlen1
+ pathlen2
+ 2*(1 + sizeof(u32
) + sizeof(u64
)) +
1939 sizeof(struct ceph_timespec
);
1941 /* calculate (max) length for cap releases */
1942 len
+= sizeof(struct ceph_mds_request_release
) *
1943 (!!req
->r_inode_drop
+ !!req
->r_dentry_drop
+
1944 !!req
->r_old_inode_drop
+ !!req
->r_old_dentry_drop
);
1945 if (req
->r_dentry_drop
)
1946 len
+= req
->r_dentry
->d_name
.len
;
1947 if (req
->r_old_dentry_drop
)
1948 len
+= req
->r_old_dentry
->d_name
.len
;
1950 msg
= ceph_msg_new(CEPH_MSG_CLIENT_REQUEST
, len
, GFP_NOFS
, false);
1952 msg
= ERR_PTR(-ENOMEM
);
1956 msg
->hdr
.version
= cpu_to_le16(2);
1957 msg
->hdr
.tid
= cpu_to_le64(req
->r_tid
);
1959 head
= msg
->front
.iov_base
;
1960 p
= msg
->front
.iov_base
+ sizeof(*head
);
1961 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1963 head
->mdsmap_epoch
= cpu_to_le32(mdsc
->mdsmap
->m_epoch
);
1964 head
->op
= cpu_to_le32(req
->r_op
);
1965 head
->caller_uid
= cpu_to_le32(from_kuid(&init_user_ns
, req
->r_uid
));
1966 head
->caller_gid
= cpu_to_le32(from_kgid(&init_user_ns
, req
->r_gid
));
1967 head
->args
= req
->r_args
;
1969 ceph_encode_filepath(&p
, end
, ino1
, path1
);
1970 ceph_encode_filepath(&p
, end
, ino2
, path2
);
1972 /* make note of release offset, in case we need to replay */
1973 req
->r_request_release_offset
= p
- msg
->front
.iov_base
;
1977 if (req
->r_inode_drop
)
1978 releases
+= ceph_encode_inode_release(&p
,
1979 req
->r_inode
? req
->r_inode
: d_inode(req
->r_dentry
),
1980 mds
, req
->r_inode_drop
, req
->r_inode_unless
, 0);
1981 if (req
->r_dentry_drop
)
1982 releases
+= ceph_encode_dentry_release(&p
, req
->r_dentry
,
1983 mds
, req
->r_dentry_drop
, req
->r_dentry_unless
);
1984 if (req
->r_old_dentry_drop
)
1985 releases
+= ceph_encode_dentry_release(&p
, req
->r_old_dentry
,
1986 mds
, req
->r_old_dentry_drop
, req
->r_old_dentry_unless
);
1987 if (req
->r_old_inode_drop
)
1988 releases
+= ceph_encode_inode_release(&p
,
1989 d_inode(req
->r_old_dentry
),
1990 mds
, req
->r_old_inode_drop
, req
->r_old_inode_unless
, 0);
1992 if (drop_cap_releases
) {
1994 p
= msg
->front
.iov_base
+ req
->r_request_release_offset
;
1997 head
->num_releases
= cpu_to_le16(releases
);
2001 struct ceph_timespec ts
;
2002 ceph_encode_timespec(&ts
, &req
->r_stamp
);
2003 ceph_encode_copy(&p
, &ts
, sizeof(ts
));
2007 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
2008 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
2010 if (req
->r_pagelist
) {
2011 struct ceph_pagelist
*pagelist
= req
->r_pagelist
;
2012 atomic_inc(&pagelist
->refcnt
);
2013 ceph_msg_data_add_pagelist(msg
, pagelist
);
2014 msg
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
2016 msg
->hdr
.data_len
= 0;
2019 msg
->hdr
.data_off
= cpu_to_le16(0);
2023 kfree((char *)path2
);
2026 kfree((char *)path1
);
2032 * called under mdsc->mutex if error, under no mutex if
2035 static void complete_request(struct ceph_mds_client
*mdsc
,
2036 struct ceph_mds_request
*req
)
2038 if (req
->r_callback
)
2039 req
->r_callback(mdsc
, req
);
2041 complete_all(&req
->r_completion
);
2045 * called under mdsc->mutex
2047 static int __prepare_send_request(struct ceph_mds_client
*mdsc
,
2048 struct ceph_mds_request
*req
,
2049 int mds
, bool drop_cap_releases
)
2051 struct ceph_mds_request_head
*rhead
;
2052 struct ceph_msg
*msg
;
2057 struct ceph_cap
*cap
=
2058 ceph_get_cap_for_mds(ceph_inode(req
->r_inode
), mds
);
2061 req
->r_sent_on_mseq
= cap
->mseq
;
2063 req
->r_sent_on_mseq
= -1;
2065 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req
,
2066 req
->r_tid
, ceph_mds_op_name(req
->r_op
), req
->r_attempts
);
2068 if (req
->r_got_unsafe
) {
2071 * Replay. Do not regenerate message (and rebuild
2072 * paths, etc.); just use the original message.
2073 * Rebuilding paths will break for renames because
2074 * d_move mangles the src name.
2076 msg
= req
->r_request
;
2077 rhead
= msg
->front
.iov_base
;
2079 flags
= le32_to_cpu(rhead
->flags
);
2080 flags
|= CEPH_MDS_FLAG_REPLAY
;
2081 rhead
->flags
= cpu_to_le32(flags
);
2083 if (req
->r_target_inode
)
2084 rhead
->ino
= cpu_to_le64(ceph_ino(req
->r_target_inode
));
2086 rhead
->num_retry
= req
->r_attempts
- 1;
2088 /* remove cap/dentry releases from message */
2089 rhead
->num_releases
= 0;
2092 p
= msg
->front
.iov_base
+ req
->r_request_release_offset
;
2094 struct ceph_timespec ts
;
2095 ceph_encode_timespec(&ts
, &req
->r_stamp
);
2096 ceph_encode_copy(&p
, &ts
, sizeof(ts
));
2099 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
2100 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
2104 if (req
->r_request
) {
2105 ceph_msg_put(req
->r_request
);
2106 req
->r_request
= NULL
;
2108 msg
= create_request_message(mdsc
, req
, mds
, drop_cap_releases
);
2110 req
->r_err
= PTR_ERR(msg
);
2111 return PTR_ERR(msg
);
2113 req
->r_request
= msg
;
2115 rhead
= msg
->front
.iov_base
;
2116 rhead
->oldest_client_tid
= cpu_to_le64(__get_oldest_tid(mdsc
));
2117 if (req
->r_got_unsafe
)
2118 flags
|= CEPH_MDS_FLAG_REPLAY
;
2119 if (req
->r_locked_dir
)
2120 flags
|= CEPH_MDS_FLAG_WANT_DENTRY
;
2121 rhead
->flags
= cpu_to_le32(flags
);
2122 rhead
->num_fwd
= req
->r_num_fwd
;
2123 rhead
->num_retry
= req
->r_attempts
- 1;
2126 dout(" r_locked_dir = %p\n", req
->r_locked_dir
);
2131 * send request, or put it on the appropriate wait list.
2133 static int __do_request(struct ceph_mds_client
*mdsc
,
2134 struct ceph_mds_request
*req
)
2136 struct ceph_mds_session
*session
= NULL
;
2140 if (req
->r_err
|| req
->r_got_result
) {
2142 __unregister_request(mdsc
, req
);
2146 if (req
->r_timeout
&&
2147 time_after_eq(jiffies
, req
->r_started
+ req
->r_timeout
)) {
2148 dout("do_request timed out\n");
2152 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
) {
2153 dout("do_request forced umount\n");
2158 put_request_session(req
);
2160 mds
= __choose_mds(mdsc
, req
);
2162 ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) < CEPH_MDS_STATE_ACTIVE
) {
2163 dout("do_request no mds or not active, waiting for map\n");
2164 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
2168 /* get, open session */
2169 session
= __ceph_lookup_mds_session(mdsc
, mds
);
2171 session
= register_session(mdsc
, mds
);
2172 if (IS_ERR(session
)) {
2173 err
= PTR_ERR(session
);
2177 req
->r_session
= get_session(session
);
2179 dout("do_request mds%d session %p state %s\n", mds
, session
,
2180 ceph_session_state_name(session
->s_state
));
2181 if (session
->s_state
!= CEPH_MDS_SESSION_OPEN
&&
2182 session
->s_state
!= CEPH_MDS_SESSION_HUNG
) {
2183 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
2184 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
2185 __open_session(mdsc
, session
);
2186 list_add(&req
->r_wait
, &session
->s_waiting
);
2191 req
->r_resend_mds
= -1; /* forget any previous mds hint */
2193 if (req
->r_request_started
== 0) /* note request start time */
2194 req
->r_request_started
= jiffies
;
2196 err
= __prepare_send_request(mdsc
, req
, mds
, false);
2198 ceph_msg_get(req
->r_request
);
2199 ceph_con_send(&session
->s_con
, req
->r_request
);
2203 ceph_put_mds_session(session
);
2206 dout("__do_request early error %d\n", err
);
2208 complete_request(mdsc
, req
);
2209 __unregister_request(mdsc
, req
);
2216 * called under mdsc->mutex
2218 static void __wake_requests(struct ceph_mds_client
*mdsc
,
2219 struct list_head
*head
)
2221 struct ceph_mds_request
*req
;
2222 LIST_HEAD(tmp_list
);
2224 list_splice_init(head
, &tmp_list
);
2226 while (!list_empty(&tmp_list
)) {
2227 req
= list_entry(tmp_list
.next
,
2228 struct ceph_mds_request
, r_wait
);
2229 list_del_init(&req
->r_wait
);
2230 dout(" wake request %p tid %llu\n", req
, req
->r_tid
);
2231 __do_request(mdsc
, req
);
2236 * Wake up threads with requests pending for @mds, so that they can
2237 * resubmit their requests to a possibly different mds.
2239 static void kick_requests(struct ceph_mds_client
*mdsc
, int mds
)
2241 struct ceph_mds_request
*req
;
2242 struct rb_node
*p
= rb_first(&mdsc
->request_tree
);
2244 dout("kick_requests mds%d\n", mds
);
2246 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2248 if (req
->r_got_unsafe
)
2250 if (req
->r_attempts
> 0)
2251 continue; /* only new requests */
2252 if (req
->r_session
&&
2253 req
->r_session
->s_mds
== mds
) {
2254 dout(" kicking tid %llu\n", req
->r_tid
);
2255 list_del_init(&req
->r_wait
);
2256 __do_request(mdsc
, req
);
2261 void ceph_mdsc_submit_request(struct ceph_mds_client
*mdsc
,
2262 struct ceph_mds_request
*req
)
2264 dout("submit_request on %p\n", req
);
2265 mutex_lock(&mdsc
->mutex
);
2266 __register_request(mdsc
, req
, NULL
);
2267 __do_request(mdsc
, req
);
2268 mutex_unlock(&mdsc
->mutex
);
2272 * Synchrously perform an mds request. Take care of all of the
2273 * session setup, forwarding, retry details.
2275 int ceph_mdsc_do_request(struct ceph_mds_client
*mdsc
,
2277 struct ceph_mds_request
*req
)
2281 dout("do_request on %p\n", req
);
2283 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2285 ceph_get_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
2286 if (req
->r_locked_dir
)
2287 ceph_get_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
2288 if (req
->r_old_dentry_dir
)
2289 ceph_get_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
2292 /* deny access to directories with pool_ns layouts */
2293 if (req
->r_inode
&& S_ISDIR(req
->r_inode
->i_mode
) &&
2294 ceph_inode(req
->r_inode
)->i_pool_ns_len
)
2296 if (req
->r_locked_dir
&&
2297 ceph_inode(req
->r_locked_dir
)->i_pool_ns_len
)
2301 mutex_lock(&mdsc
->mutex
);
2302 __register_request(mdsc
, req
, dir
);
2303 __do_request(mdsc
, req
);
2311 mutex_unlock(&mdsc
->mutex
);
2312 dout("do_request waiting\n");
2313 if (!req
->r_timeout
&& req
->r_wait_for_completion
) {
2314 err
= req
->r_wait_for_completion(mdsc
, req
);
2316 long timeleft
= wait_for_completion_killable_timeout(
2318 ceph_timeout_jiffies(req
->r_timeout
));
2322 err
= -EIO
; /* timed out */
2324 err
= timeleft
; /* killed */
2326 dout("do_request waited, got %d\n", err
);
2327 mutex_lock(&mdsc
->mutex
);
2329 /* only abort if we didn't race with a real reply */
2330 if (req
->r_got_result
) {
2331 err
= le32_to_cpu(req
->r_reply_info
.head
->result
);
2332 } else if (err
< 0) {
2333 dout("aborted request %lld with %d\n", req
->r_tid
, err
);
2336 * ensure we aren't running concurrently with
2337 * ceph_fill_trace or ceph_readdir_prepopulate, which
2338 * rely on locks (dir mutex) held by our caller.
2340 mutex_lock(&req
->r_fill_mutex
);
2342 req
->r_aborted
= true;
2343 mutex_unlock(&req
->r_fill_mutex
);
2345 if (req
->r_locked_dir
&&
2346 (req
->r_op
& CEPH_MDS_OP_WRITE
))
2347 ceph_invalidate_dir_request(req
);
2353 mutex_unlock(&mdsc
->mutex
);
2354 dout("do_request %p done, result %d\n", req
, err
);
2359 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2360 * namespace request.
2362 void ceph_invalidate_dir_request(struct ceph_mds_request
*req
)
2364 struct inode
*inode
= req
->r_locked_dir
;
2366 dout("invalidate_dir_request %p (complete, lease(s))\n", inode
);
2368 ceph_dir_clear_complete(inode
);
2370 ceph_invalidate_dentry_lease(req
->r_dentry
);
2371 if (req
->r_old_dentry
)
2372 ceph_invalidate_dentry_lease(req
->r_old_dentry
);
2378 * We take the session mutex and parse and process the reply immediately.
2379 * This preserves the logical ordering of replies, capabilities, etc., sent
2380 * by the MDS as they are applied to our local cache.
2382 static void handle_reply(struct ceph_mds_session
*session
, struct ceph_msg
*msg
)
2384 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2385 struct ceph_mds_request
*req
;
2386 struct ceph_mds_reply_head
*head
= msg
->front
.iov_base
;
2387 struct ceph_mds_reply_info_parsed
*rinfo
; /* parsed reply info */
2388 struct ceph_snap_realm
*realm
;
2391 int mds
= session
->s_mds
;
2393 if (msg
->front
.iov_len
< sizeof(*head
)) {
2394 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2399 /* get request, session */
2400 tid
= le64_to_cpu(msg
->hdr
.tid
);
2401 mutex_lock(&mdsc
->mutex
);
2402 req
= lookup_get_request(mdsc
, tid
);
2404 dout("handle_reply on unknown tid %llu\n", tid
);
2405 mutex_unlock(&mdsc
->mutex
);
2408 dout("handle_reply %p\n", req
);
2410 /* correct session? */
2411 if (req
->r_session
!= session
) {
2412 pr_err("mdsc_handle_reply got %llu on session mds%d"
2413 " not mds%d\n", tid
, session
->s_mds
,
2414 req
->r_session
? req
->r_session
->s_mds
: -1);
2415 mutex_unlock(&mdsc
->mutex
);
2420 if ((req
->r_got_unsafe
&& !head
->safe
) ||
2421 (req
->r_got_safe
&& head
->safe
)) {
2422 pr_warn("got a dup %s reply on %llu from mds%d\n",
2423 head
->safe
? "safe" : "unsafe", tid
, mds
);
2424 mutex_unlock(&mdsc
->mutex
);
2427 if (req
->r_got_safe
) {
2428 pr_warn("got unsafe after safe on %llu from mds%d\n",
2430 mutex_unlock(&mdsc
->mutex
);
2434 result
= le32_to_cpu(head
->result
);
2438 * if we're not talking to the authority, send to them
2439 * if the authority has changed while we weren't looking,
2440 * send to new authority
2441 * Otherwise we just have to return an ESTALE
2443 if (result
== -ESTALE
) {
2444 dout("got ESTALE on request %llu", req
->r_tid
);
2445 req
->r_resend_mds
= -1;
2446 if (req
->r_direct_mode
!= USE_AUTH_MDS
) {
2447 dout("not using auth, setting for that now");
2448 req
->r_direct_mode
= USE_AUTH_MDS
;
2449 __do_request(mdsc
, req
);
2450 mutex_unlock(&mdsc
->mutex
);
2453 int mds
= __choose_mds(mdsc
, req
);
2454 if (mds
>= 0 && mds
!= req
->r_session
->s_mds
) {
2455 dout("but auth changed, so resending");
2456 __do_request(mdsc
, req
);
2457 mutex_unlock(&mdsc
->mutex
);
2461 dout("have to return ESTALE on request %llu", req
->r_tid
);
2466 req
->r_got_safe
= true;
2467 __unregister_request(mdsc
, req
);
2469 if (req
->r_got_unsafe
) {
2471 * We already handled the unsafe response, now do the
2472 * cleanup. No need to examine the response; the MDS
2473 * doesn't include any result info in the safe
2474 * response. And even if it did, there is nothing
2475 * useful we could do with a revised return value.
2477 dout("got safe reply %llu, mds%d\n", tid
, mds
);
2478 list_del_init(&req
->r_unsafe_item
);
2480 /* last unsafe request during umount? */
2481 if (mdsc
->stopping
&& !__get_oldest_req(mdsc
))
2482 complete_all(&mdsc
->safe_umount_waiters
);
2483 mutex_unlock(&mdsc
->mutex
);
2487 req
->r_got_unsafe
= true;
2488 list_add_tail(&req
->r_unsafe_item
, &req
->r_session
->s_unsafe
);
2489 if (req
->r_unsafe_dir
) {
2490 struct ceph_inode_info
*ci
=
2491 ceph_inode(req
->r_unsafe_dir
);
2492 spin_lock(&ci
->i_unsafe_lock
);
2493 list_add_tail(&req
->r_unsafe_dir_item
,
2494 &ci
->i_unsafe_dirops
);
2495 spin_unlock(&ci
->i_unsafe_lock
);
2499 dout("handle_reply tid %lld result %d\n", tid
, result
);
2500 rinfo
= &req
->r_reply_info
;
2501 err
= parse_reply_info(msg
, rinfo
, session
->s_con
.peer_features
);
2502 mutex_unlock(&mdsc
->mutex
);
2504 mutex_lock(&session
->s_mutex
);
2506 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds
, tid
);
2513 if (rinfo
->snapblob_len
) {
2514 down_write(&mdsc
->snap_rwsem
);
2515 ceph_update_snap_trace(mdsc
, rinfo
->snapblob
,
2516 rinfo
->snapblob
+ rinfo
->snapblob_len
,
2517 le32_to_cpu(head
->op
) == CEPH_MDS_OP_RMSNAP
,
2519 downgrade_write(&mdsc
->snap_rwsem
);
2521 down_read(&mdsc
->snap_rwsem
);
2524 /* insert trace into our cache */
2525 mutex_lock(&req
->r_fill_mutex
);
2526 current
->journal_info
= req
;
2527 err
= ceph_fill_trace(mdsc
->fsc
->sb
, req
, req
->r_session
);
2529 if (result
== 0 && (req
->r_op
== CEPH_MDS_OP_READDIR
||
2530 req
->r_op
== CEPH_MDS_OP_LSSNAP
))
2531 ceph_readdir_prepopulate(req
, req
->r_session
);
2532 ceph_unreserve_caps(mdsc
, &req
->r_caps_reservation
);
2534 current
->journal_info
= NULL
;
2535 mutex_unlock(&req
->r_fill_mutex
);
2537 up_read(&mdsc
->snap_rwsem
);
2539 ceph_put_snap_realm(mdsc
, realm
);
2541 if (err
== 0 && req
->r_got_unsafe
&& req
->r_target_inode
) {
2542 struct ceph_inode_info
*ci
= ceph_inode(req
->r_target_inode
);
2543 spin_lock(&ci
->i_unsafe_lock
);
2544 list_add_tail(&req
->r_unsafe_target_item
, &ci
->i_unsafe_iops
);
2545 spin_unlock(&ci
->i_unsafe_lock
);
2548 mutex_lock(&mdsc
->mutex
);
2549 if (!req
->r_aborted
) {
2553 req
->r_reply
= ceph_msg_get(msg
);
2554 req
->r_got_result
= true;
2557 dout("reply arrived after request %lld was aborted\n", tid
);
2559 mutex_unlock(&mdsc
->mutex
);
2561 mutex_unlock(&session
->s_mutex
);
2563 /* kick calling process */
2564 complete_request(mdsc
, req
);
2566 ceph_mdsc_put_request(req
);
2573 * handle mds notification that our request has been forwarded.
2575 static void handle_forward(struct ceph_mds_client
*mdsc
,
2576 struct ceph_mds_session
*session
,
2577 struct ceph_msg
*msg
)
2579 struct ceph_mds_request
*req
;
2580 u64 tid
= le64_to_cpu(msg
->hdr
.tid
);
2584 void *p
= msg
->front
.iov_base
;
2585 void *end
= p
+ msg
->front
.iov_len
;
2587 ceph_decode_need(&p
, end
, 2*sizeof(u32
), bad
);
2588 next_mds
= ceph_decode_32(&p
);
2589 fwd_seq
= ceph_decode_32(&p
);
2591 mutex_lock(&mdsc
->mutex
);
2592 req
= lookup_get_request(mdsc
, tid
);
2594 dout("forward tid %llu to mds%d - req dne\n", tid
, next_mds
);
2595 goto out
; /* dup reply? */
2598 if (req
->r_aborted
) {
2599 dout("forward tid %llu aborted, unregistering\n", tid
);
2600 __unregister_request(mdsc
, req
);
2601 } else if (fwd_seq
<= req
->r_num_fwd
) {
2602 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2603 tid
, next_mds
, req
->r_num_fwd
, fwd_seq
);
2605 /* resend. forward race not possible; mds would drop */
2606 dout("forward tid %llu to mds%d (we resend)\n", tid
, next_mds
);
2608 BUG_ON(req
->r_got_result
);
2609 req
->r_attempts
= 0;
2610 req
->r_num_fwd
= fwd_seq
;
2611 req
->r_resend_mds
= next_mds
;
2612 put_request_session(req
);
2613 __do_request(mdsc
, req
);
2615 ceph_mdsc_put_request(req
);
2617 mutex_unlock(&mdsc
->mutex
);
2621 pr_err("mdsc_handle_forward decode error err=%d\n", err
);
2625 * handle a mds session control message
2627 static void handle_session(struct ceph_mds_session
*session
,
2628 struct ceph_msg
*msg
)
2630 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2633 int mds
= session
->s_mds
;
2634 struct ceph_mds_session_head
*h
= msg
->front
.iov_base
;
2638 if (msg
->front
.iov_len
!= sizeof(*h
))
2640 op
= le32_to_cpu(h
->op
);
2641 seq
= le64_to_cpu(h
->seq
);
2643 mutex_lock(&mdsc
->mutex
);
2644 if (op
== CEPH_SESSION_CLOSE
)
2645 __unregister_session(mdsc
, session
);
2646 /* FIXME: this ttl calculation is generous */
2647 session
->s_ttl
= jiffies
+ HZ
*mdsc
->mdsmap
->m_session_autoclose
;
2648 mutex_unlock(&mdsc
->mutex
);
2650 mutex_lock(&session
->s_mutex
);
2652 dout("handle_session mds%d %s %p state %s seq %llu\n",
2653 mds
, ceph_session_op_name(op
), session
,
2654 ceph_session_state_name(session
->s_state
), seq
);
2656 if (session
->s_state
== CEPH_MDS_SESSION_HUNG
) {
2657 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2658 pr_info("mds%d came back\n", session
->s_mds
);
2662 case CEPH_SESSION_OPEN
:
2663 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2664 pr_info("mds%d reconnect success\n", session
->s_mds
);
2665 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2666 renewed_caps(mdsc
, session
, 0);
2669 __close_session(mdsc
, session
);
2672 case CEPH_SESSION_RENEWCAPS
:
2673 if (session
->s_renew_seq
== seq
)
2674 renewed_caps(mdsc
, session
, 1);
2677 case CEPH_SESSION_CLOSE
:
2678 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2679 pr_info("mds%d reconnect denied\n", session
->s_mds
);
2680 cleanup_session_requests(mdsc
, session
);
2681 remove_session_caps(session
);
2682 wake
= 2; /* for good measure */
2683 wake_up_all(&mdsc
->session_close_wq
);
2686 case CEPH_SESSION_STALE
:
2687 pr_info("mds%d caps went stale, renewing\n",
2689 spin_lock(&session
->s_gen_ttl_lock
);
2690 session
->s_cap_gen
++;
2691 session
->s_cap_ttl
= jiffies
- 1;
2692 spin_unlock(&session
->s_gen_ttl_lock
);
2693 send_renew_caps(mdsc
, session
);
2696 case CEPH_SESSION_RECALL_STATE
:
2697 trim_caps(mdsc
, session
, le32_to_cpu(h
->max_caps
));
2700 case CEPH_SESSION_FLUSHMSG
:
2701 send_flushmsg_ack(mdsc
, session
, seq
);
2704 case CEPH_SESSION_FORCE_RO
:
2705 dout("force_session_readonly %p\n", session
);
2706 spin_lock(&session
->s_cap_lock
);
2707 session
->s_readonly
= true;
2708 spin_unlock(&session
->s_cap_lock
);
2709 wake_up_session_caps(session
, 0);
2713 pr_err("mdsc_handle_session bad op %d mds%d\n", op
, mds
);
2717 mutex_unlock(&session
->s_mutex
);
2719 mutex_lock(&mdsc
->mutex
);
2720 __wake_requests(mdsc
, &session
->s_waiting
);
2722 kick_requests(mdsc
, mds
);
2723 mutex_unlock(&mdsc
->mutex
);
2728 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds
,
2729 (int)msg
->front
.iov_len
);
2736 * called under session->mutex.
2738 static void replay_unsafe_requests(struct ceph_mds_client
*mdsc
,
2739 struct ceph_mds_session
*session
)
2741 struct ceph_mds_request
*req
, *nreq
;
2745 dout("replay_unsafe_requests mds%d\n", session
->s_mds
);
2747 mutex_lock(&mdsc
->mutex
);
2748 list_for_each_entry_safe(req
, nreq
, &session
->s_unsafe
, r_unsafe_item
) {
2749 err
= __prepare_send_request(mdsc
, req
, session
->s_mds
, true);
2751 ceph_msg_get(req
->r_request
);
2752 ceph_con_send(&session
->s_con
, req
->r_request
);
2757 * also re-send old requests when MDS enters reconnect stage. So that MDS
2758 * can process completed request in clientreplay stage.
2760 p
= rb_first(&mdsc
->request_tree
);
2762 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2764 if (req
->r_got_unsafe
)
2766 if (req
->r_attempts
== 0)
2767 continue; /* only old requests */
2768 if (req
->r_session
&&
2769 req
->r_session
->s_mds
== session
->s_mds
) {
2770 err
= __prepare_send_request(mdsc
, req
,
2771 session
->s_mds
, true);
2773 ceph_msg_get(req
->r_request
);
2774 ceph_con_send(&session
->s_con
, req
->r_request
);
2778 mutex_unlock(&mdsc
->mutex
);
2782 * Encode information about a cap for a reconnect with the MDS.
2784 static int encode_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
2788 struct ceph_mds_cap_reconnect v2
;
2789 struct ceph_mds_cap_reconnect_v1 v1
;
2792 struct ceph_inode_info
*ci
;
2793 struct ceph_reconnect_state
*recon_state
= arg
;
2794 struct ceph_pagelist
*pagelist
= recon_state
->pagelist
;
2798 struct dentry
*dentry
;
2802 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2803 inode
, ceph_vinop(inode
), cap
, cap
->cap_id
,
2804 ceph_cap_string(cap
->issued
));
2805 err
= ceph_pagelist_encode_64(pagelist
, ceph_ino(inode
));
2809 dentry
= d_find_alias(inode
);
2811 path
= ceph_mdsc_build_path(dentry
, &pathlen
, &pathbase
, 0);
2813 err
= PTR_ERR(path
);
2820 err
= ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
2824 spin_lock(&ci
->i_ceph_lock
);
2825 cap
->seq
= 0; /* reset cap seq */
2826 cap
->issue_seq
= 0; /* and issue_seq */
2827 cap
->mseq
= 0; /* and migrate_seq */
2828 cap
->cap_gen
= cap
->session
->s_cap_gen
;
2830 if (recon_state
->flock
) {
2831 rec
.v2
.cap_id
= cpu_to_le64(cap
->cap_id
);
2832 rec
.v2
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2833 rec
.v2
.issued
= cpu_to_le32(cap
->issued
);
2834 rec
.v2
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2835 rec
.v2
.pathbase
= cpu_to_le64(pathbase
);
2836 rec
.v2
.flock_len
= 0;
2837 reclen
= sizeof(rec
.v2
);
2839 rec
.v1
.cap_id
= cpu_to_le64(cap
->cap_id
);
2840 rec
.v1
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2841 rec
.v1
.issued
= cpu_to_le32(cap
->issued
);
2842 rec
.v1
.size
= cpu_to_le64(inode
->i_size
);
2843 ceph_encode_timespec(&rec
.v1
.mtime
, &inode
->i_mtime
);
2844 ceph_encode_timespec(&rec
.v1
.atime
, &inode
->i_atime
);
2845 rec
.v1
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2846 rec
.v1
.pathbase
= cpu_to_le64(pathbase
);
2847 reclen
= sizeof(rec
.v1
);
2849 spin_unlock(&ci
->i_ceph_lock
);
2851 if (recon_state
->flock
) {
2852 int num_fcntl_locks
, num_flock_locks
;
2853 struct ceph_filelock
*flocks
;
2856 ceph_count_locks(inode
, &num_fcntl_locks
, &num_flock_locks
);
2857 flocks
= kmalloc((num_fcntl_locks
+num_flock_locks
) *
2858 sizeof(struct ceph_filelock
), GFP_NOFS
);
2863 err
= ceph_encode_locks_to_buffer(inode
, flocks
,
2873 * number of encoded locks is stable, so copy to pagelist
2875 rec
.v2
.flock_len
= cpu_to_le32(2*sizeof(u32
) +
2876 (num_fcntl_locks
+num_flock_locks
) *
2877 sizeof(struct ceph_filelock
));
2878 err
= ceph_pagelist_append(pagelist
, &rec
, reclen
);
2880 err
= ceph_locks_to_pagelist(flocks
, pagelist
,
2885 err
= ceph_pagelist_append(pagelist
, &rec
, reclen
);
2888 recon_state
->nr_caps
++;
2898 * If an MDS fails and recovers, clients need to reconnect in order to
2899 * reestablish shared state. This includes all caps issued through
2900 * this session _and_ the snap_realm hierarchy. Because it's not
2901 * clear which snap realms the mds cares about, we send everything we
2902 * know about.. that ensures we'll then get any new info the
2903 * recovering MDS might have.
2905 * This is a relatively heavyweight operation, but it's rare.
2907 * called with mdsc->mutex held.
2909 static void send_mds_reconnect(struct ceph_mds_client
*mdsc
,
2910 struct ceph_mds_session
*session
)
2912 struct ceph_msg
*reply
;
2914 int mds
= session
->s_mds
;
2917 struct ceph_pagelist
*pagelist
;
2918 struct ceph_reconnect_state recon_state
;
2920 pr_info("mds%d reconnect start\n", mds
);
2922 pagelist
= kmalloc(sizeof(*pagelist
), GFP_NOFS
);
2924 goto fail_nopagelist
;
2925 ceph_pagelist_init(pagelist
);
2927 reply
= ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT
, 0, GFP_NOFS
, false);
2931 mutex_lock(&session
->s_mutex
);
2932 session
->s_state
= CEPH_MDS_SESSION_RECONNECTING
;
2935 dout("session %p state %s\n", session
,
2936 ceph_session_state_name(session
->s_state
));
2938 spin_lock(&session
->s_gen_ttl_lock
);
2939 session
->s_cap_gen
++;
2940 spin_unlock(&session
->s_gen_ttl_lock
);
2942 spin_lock(&session
->s_cap_lock
);
2943 /* don't know if session is readonly */
2944 session
->s_readonly
= 0;
2946 * notify __ceph_remove_cap() that we are composing cap reconnect.
2947 * If a cap get released before being added to the cap reconnect,
2948 * __ceph_remove_cap() should skip queuing cap release.
2950 session
->s_cap_reconnect
= 1;
2951 /* drop old cap expires; we're about to reestablish that state */
2952 cleanup_cap_releases(mdsc
, session
);
2954 /* trim unused caps to reduce MDS's cache rejoin time */
2955 if (mdsc
->fsc
->sb
->s_root
)
2956 shrink_dcache_parent(mdsc
->fsc
->sb
->s_root
);
2958 ceph_con_close(&session
->s_con
);
2959 ceph_con_open(&session
->s_con
,
2960 CEPH_ENTITY_TYPE_MDS
, mds
,
2961 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
2963 /* replay unsafe requests */
2964 replay_unsafe_requests(mdsc
, session
);
2966 down_read(&mdsc
->snap_rwsem
);
2968 /* traverse this session's caps */
2969 s_nr_caps
= session
->s_nr_caps
;
2970 err
= ceph_pagelist_encode_32(pagelist
, s_nr_caps
);
2974 recon_state
.nr_caps
= 0;
2975 recon_state
.pagelist
= pagelist
;
2976 recon_state
.flock
= session
->s_con
.peer_features
& CEPH_FEATURE_FLOCK
;
2977 err
= iterate_session_caps(session
, encode_caps_cb
, &recon_state
);
2981 spin_lock(&session
->s_cap_lock
);
2982 session
->s_cap_reconnect
= 0;
2983 spin_unlock(&session
->s_cap_lock
);
2986 * snaprealms. we provide mds with the ino, seq (version), and
2987 * parent for all of our realms. If the mds has any newer info,
2990 for (p
= rb_first(&mdsc
->snap_realms
); p
; p
= rb_next(p
)) {
2991 struct ceph_snap_realm
*realm
=
2992 rb_entry(p
, struct ceph_snap_realm
, node
);
2993 struct ceph_mds_snaprealm_reconnect sr_rec
;
2995 dout(" adding snap realm %llx seq %lld parent %llx\n",
2996 realm
->ino
, realm
->seq
, realm
->parent_ino
);
2997 sr_rec
.ino
= cpu_to_le64(realm
->ino
);
2998 sr_rec
.seq
= cpu_to_le64(realm
->seq
);
2999 sr_rec
.parent
= cpu_to_le64(realm
->parent_ino
);
3000 err
= ceph_pagelist_append(pagelist
, &sr_rec
, sizeof(sr_rec
));
3005 if (recon_state
.flock
)
3006 reply
->hdr
.version
= cpu_to_le16(2);
3008 /* raced with cap release? */
3009 if (s_nr_caps
!= recon_state
.nr_caps
) {
3010 struct page
*page
= list_first_entry(&pagelist
->head
,
3012 __le32
*addr
= kmap_atomic(page
);
3013 *addr
= cpu_to_le32(recon_state
.nr_caps
);
3014 kunmap_atomic(addr
);
3017 reply
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
3018 ceph_msg_data_add_pagelist(reply
, pagelist
);
3020 ceph_early_kick_flushing_caps(mdsc
, session
);
3022 ceph_con_send(&session
->s_con
, reply
);
3024 mutex_unlock(&session
->s_mutex
);
3026 mutex_lock(&mdsc
->mutex
);
3027 __wake_requests(mdsc
, &session
->s_waiting
);
3028 mutex_unlock(&mdsc
->mutex
);
3030 up_read(&mdsc
->snap_rwsem
);
3034 ceph_msg_put(reply
);
3035 up_read(&mdsc
->snap_rwsem
);
3036 mutex_unlock(&session
->s_mutex
);
3038 ceph_pagelist_release(pagelist
);
3040 pr_err("error %d preparing reconnect for mds%d\n", err
, mds
);
3046 * compare old and new mdsmaps, kicking requests
3047 * and closing out old connections as necessary
3049 * called under mdsc->mutex.
3051 static void check_new_map(struct ceph_mds_client
*mdsc
,
3052 struct ceph_mdsmap
*newmap
,
3053 struct ceph_mdsmap
*oldmap
)
3056 int oldstate
, newstate
;
3057 struct ceph_mds_session
*s
;
3059 dout("check_new_map new %u old %u\n",
3060 newmap
->m_epoch
, oldmap
->m_epoch
);
3062 for (i
= 0; i
< oldmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
3063 if (mdsc
->sessions
[i
] == NULL
)
3065 s
= mdsc
->sessions
[i
];
3066 oldstate
= ceph_mdsmap_get_state(oldmap
, i
);
3067 newstate
= ceph_mdsmap_get_state(newmap
, i
);
3069 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3070 i
, ceph_mds_state_name(oldstate
),
3071 ceph_mdsmap_is_laggy(oldmap
, i
) ? " (laggy)" : "",
3072 ceph_mds_state_name(newstate
),
3073 ceph_mdsmap_is_laggy(newmap
, i
) ? " (laggy)" : "",
3074 ceph_session_state_name(s
->s_state
));
3076 if (i
>= newmap
->m_max_mds
||
3077 memcmp(ceph_mdsmap_get_addr(oldmap
, i
),
3078 ceph_mdsmap_get_addr(newmap
, i
),
3079 sizeof(struct ceph_entity_addr
))) {
3080 if (s
->s_state
== CEPH_MDS_SESSION_OPENING
) {
3081 /* the session never opened, just close it
3083 __wake_requests(mdsc
, &s
->s_waiting
);
3084 __unregister_session(mdsc
, s
);
3087 mutex_unlock(&mdsc
->mutex
);
3088 mutex_lock(&s
->s_mutex
);
3089 mutex_lock(&mdsc
->mutex
);
3090 ceph_con_close(&s
->s_con
);
3091 mutex_unlock(&s
->s_mutex
);
3092 s
->s_state
= CEPH_MDS_SESSION_RESTARTING
;
3094 } else if (oldstate
== newstate
) {
3095 continue; /* nothing new with this mds */
3101 if (s
->s_state
== CEPH_MDS_SESSION_RESTARTING
&&
3102 newstate
>= CEPH_MDS_STATE_RECONNECT
) {
3103 mutex_unlock(&mdsc
->mutex
);
3104 send_mds_reconnect(mdsc
, s
);
3105 mutex_lock(&mdsc
->mutex
);
3109 * kick request on any mds that has gone active.
3111 if (oldstate
< CEPH_MDS_STATE_ACTIVE
&&
3112 newstate
>= CEPH_MDS_STATE_ACTIVE
) {
3113 if (oldstate
!= CEPH_MDS_STATE_CREATING
&&
3114 oldstate
!= CEPH_MDS_STATE_STARTING
)
3115 pr_info("mds%d recovery completed\n", s
->s_mds
);
3116 kick_requests(mdsc
, i
);
3117 ceph_kick_flushing_caps(mdsc
, s
);
3118 wake_up_session_caps(s
, 1);
3122 for (i
= 0; i
< newmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
3123 s
= mdsc
->sessions
[i
];
3126 if (!ceph_mdsmap_is_laggy(newmap
, i
))
3128 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3129 s
->s_state
== CEPH_MDS_SESSION_HUNG
||
3130 s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3131 dout(" connecting to export targets of laggy mds%d\n",
3133 __open_export_target_sessions(mdsc
, s
);
3145 * caller must hold session s_mutex, dentry->d_lock
3147 void __ceph_mdsc_drop_dentry_lease(struct dentry
*dentry
)
3149 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
3151 ceph_put_mds_session(di
->lease_session
);
3152 di
->lease_session
= NULL
;
3155 static void handle_lease(struct ceph_mds_client
*mdsc
,
3156 struct ceph_mds_session
*session
,
3157 struct ceph_msg
*msg
)
3159 struct super_block
*sb
= mdsc
->fsc
->sb
;
3160 struct inode
*inode
;
3161 struct dentry
*parent
, *dentry
;
3162 struct ceph_dentry_info
*di
;
3163 int mds
= session
->s_mds
;
3164 struct ceph_mds_lease
*h
= msg
->front
.iov_base
;
3166 struct ceph_vino vino
;
3170 dout("handle_lease from mds%d\n", mds
);
3173 if (msg
->front
.iov_len
< sizeof(*h
) + sizeof(u32
))
3175 vino
.ino
= le64_to_cpu(h
->ino
);
3176 vino
.snap
= CEPH_NOSNAP
;
3177 seq
= le32_to_cpu(h
->seq
);
3178 dname
.name
= (void *)h
+ sizeof(*h
) + sizeof(u32
);
3179 dname
.len
= msg
->front
.iov_len
- sizeof(*h
) - sizeof(u32
);
3180 if (dname
.len
!= get_unaligned_le32(h
+1))
3184 inode
= ceph_find_inode(sb
, vino
);
3185 dout("handle_lease %s, ino %llx %p %.*s\n",
3186 ceph_lease_op_name(h
->action
), vino
.ino
, inode
,
3187 dname
.len
, dname
.name
);
3189 mutex_lock(&session
->s_mutex
);
3192 if (inode
== NULL
) {
3193 dout("handle_lease no inode %llx\n", vino
.ino
);
3198 parent
= d_find_alias(inode
);
3200 dout("no parent dentry on inode %p\n", inode
);
3202 goto release
; /* hrm... */
3204 dname
.hash
= full_name_hash(dname
.name
, dname
.len
);
3205 dentry
= d_lookup(parent
, &dname
);
3210 spin_lock(&dentry
->d_lock
);
3211 di
= ceph_dentry(dentry
);
3212 switch (h
->action
) {
3213 case CEPH_MDS_LEASE_REVOKE
:
3214 if (di
->lease_session
== session
) {
3215 if (ceph_seq_cmp(di
->lease_seq
, seq
) > 0)
3216 h
->seq
= cpu_to_le32(di
->lease_seq
);
3217 __ceph_mdsc_drop_dentry_lease(dentry
);
3222 case CEPH_MDS_LEASE_RENEW
:
3223 if (di
->lease_session
== session
&&
3224 di
->lease_gen
== session
->s_cap_gen
&&
3225 di
->lease_renew_from
&&
3226 di
->lease_renew_after
== 0) {
3227 unsigned long duration
=
3228 msecs_to_jiffies(le32_to_cpu(h
->duration_ms
));
3230 di
->lease_seq
= seq
;
3231 dentry
->d_time
= di
->lease_renew_from
+ duration
;
3232 di
->lease_renew_after
= di
->lease_renew_from
+
3234 di
->lease_renew_from
= 0;
3238 spin_unlock(&dentry
->d_lock
);
3245 /* let's just reuse the same message */
3246 h
->action
= CEPH_MDS_LEASE_REVOKE_ACK
;
3248 ceph_con_send(&session
->s_con
, msg
);
3252 mutex_unlock(&session
->s_mutex
);
3256 pr_err("corrupt lease message\n");
3260 void ceph_mdsc_lease_send_msg(struct ceph_mds_session
*session
,
3261 struct inode
*inode
,
3262 struct dentry
*dentry
, char action
,
3265 struct ceph_msg
*msg
;
3266 struct ceph_mds_lease
*lease
;
3267 int len
= sizeof(*lease
) + sizeof(u32
);
3270 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3271 inode
, dentry
, ceph_lease_op_name(action
), session
->s_mds
);
3272 dnamelen
= dentry
->d_name
.len
;
3275 msg
= ceph_msg_new(CEPH_MSG_CLIENT_LEASE
, len
, GFP_NOFS
, false);
3278 lease
= msg
->front
.iov_base
;
3279 lease
->action
= action
;
3280 lease
->ino
= cpu_to_le64(ceph_vino(inode
).ino
);
3281 lease
->first
= lease
->last
= cpu_to_le64(ceph_vino(inode
).snap
);
3282 lease
->seq
= cpu_to_le32(seq
);
3283 put_unaligned_le32(dnamelen
, lease
+ 1);
3284 memcpy((void *)(lease
+ 1) + 4, dentry
->d_name
.name
, dnamelen
);
3287 * if this is a preemptive lease RELEASE, no need to
3288 * flush request stream, since the actual request will
3291 msg
->more_to_follow
= (action
== CEPH_MDS_LEASE_RELEASE
);
3293 ceph_con_send(&session
->s_con
, msg
);
3297 * Preemptively release a lease we expect to invalidate anyway.
3298 * Pass @inode always, @dentry is optional.
3300 void ceph_mdsc_lease_release(struct ceph_mds_client
*mdsc
, struct inode
*inode
,
3301 struct dentry
*dentry
)
3303 struct ceph_dentry_info
*di
;
3304 struct ceph_mds_session
*session
;
3307 BUG_ON(inode
== NULL
);
3308 BUG_ON(dentry
== NULL
);
3310 /* is dentry lease valid? */
3311 spin_lock(&dentry
->d_lock
);
3312 di
= ceph_dentry(dentry
);
3313 if (!di
|| !di
->lease_session
||
3314 di
->lease_session
->s_mds
< 0 ||
3315 di
->lease_gen
!= di
->lease_session
->s_cap_gen
||
3316 !time_before(jiffies
, dentry
->d_time
)) {
3317 dout("lease_release inode %p dentry %p -- "
3320 spin_unlock(&dentry
->d_lock
);
3324 /* we do have a lease on this dentry; note mds and seq */
3325 session
= ceph_get_mds_session(di
->lease_session
);
3326 seq
= di
->lease_seq
;
3327 __ceph_mdsc_drop_dentry_lease(dentry
);
3328 spin_unlock(&dentry
->d_lock
);
3330 dout("lease_release inode %p dentry %p to mds%d\n",
3331 inode
, dentry
, session
->s_mds
);
3332 ceph_mdsc_lease_send_msg(session
, inode
, dentry
,
3333 CEPH_MDS_LEASE_RELEASE
, seq
);
3334 ceph_put_mds_session(session
);
3338 * drop all leases (and dentry refs) in preparation for umount
3340 static void drop_leases(struct ceph_mds_client
*mdsc
)
3344 dout("drop_leases\n");
3345 mutex_lock(&mdsc
->mutex
);
3346 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3347 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3350 mutex_unlock(&mdsc
->mutex
);
3351 mutex_lock(&s
->s_mutex
);
3352 mutex_unlock(&s
->s_mutex
);
3353 ceph_put_mds_session(s
);
3354 mutex_lock(&mdsc
->mutex
);
3356 mutex_unlock(&mdsc
->mutex
);
3362 * delayed work -- periodically trim expired leases, renew caps with mds
3364 static void schedule_delayed(struct ceph_mds_client
*mdsc
)
3367 unsigned hz
= round_jiffies_relative(HZ
* delay
);
3368 schedule_delayed_work(&mdsc
->delayed_work
, hz
);
3371 static void delayed_work(struct work_struct
*work
)
3374 struct ceph_mds_client
*mdsc
=
3375 container_of(work
, struct ceph_mds_client
, delayed_work
.work
);
3379 dout("mdsc delayed_work\n");
3380 ceph_check_delayed_caps(mdsc
);
3382 mutex_lock(&mdsc
->mutex
);
3383 renew_interval
= mdsc
->mdsmap
->m_session_timeout
>> 2;
3384 renew_caps
= time_after_eq(jiffies
, HZ
*renew_interval
+
3385 mdsc
->last_renew_caps
);
3387 mdsc
->last_renew_caps
= jiffies
;
3389 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3390 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3393 if (s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3394 dout("resending session close request for mds%d\n",
3396 request_close_session(mdsc
, s
);
3397 ceph_put_mds_session(s
);
3400 if (s
->s_ttl
&& time_after(jiffies
, s
->s_ttl
)) {
3401 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
) {
3402 s
->s_state
= CEPH_MDS_SESSION_HUNG
;
3403 pr_info("mds%d hung\n", s
->s_mds
);
3406 if (s
->s_state
< CEPH_MDS_SESSION_OPEN
) {
3407 /* this mds is failed or recovering, just wait */
3408 ceph_put_mds_session(s
);
3411 mutex_unlock(&mdsc
->mutex
);
3413 mutex_lock(&s
->s_mutex
);
3415 send_renew_caps(mdsc
, s
);
3417 ceph_con_keepalive(&s
->s_con
);
3418 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3419 s
->s_state
== CEPH_MDS_SESSION_HUNG
)
3420 ceph_send_cap_releases(mdsc
, s
);
3421 mutex_unlock(&s
->s_mutex
);
3422 ceph_put_mds_session(s
);
3424 mutex_lock(&mdsc
->mutex
);
3426 mutex_unlock(&mdsc
->mutex
);
3428 schedule_delayed(mdsc
);
3431 int ceph_mdsc_init(struct ceph_fs_client
*fsc
)
3434 struct ceph_mds_client
*mdsc
;
3436 mdsc
= kzalloc(sizeof(struct ceph_mds_client
), GFP_NOFS
);
3441 mutex_init(&mdsc
->mutex
);
3442 mdsc
->mdsmap
= kzalloc(sizeof(*mdsc
->mdsmap
), GFP_NOFS
);
3443 if (mdsc
->mdsmap
== NULL
) {
3448 init_completion(&mdsc
->safe_umount_waiters
);
3449 init_waitqueue_head(&mdsc
->session_close_wq
);
3450 INIT_LIST_HEAD(&mdsc
->waiting_for_map
);
3451 mdsc
->sessions
= NULL
;
3452 atomic_set(&mdsc
->num_sessions
, 0);
3453 mdsc
->max_sessions
= 0;
3455 mdsc
->last_snap_seq
= 0;
3456 init_rwsem(&mdsc
->snap_rwsem
);
3457 mdsc
->snap_realms
= RB_ROOT
;
3458 INIT_LIST_HEAD(&mdsc
->snap_empty
);
3459 spin_lock_init(&mdsc
->snap_empty_lock
);
3461 mdsc
->oldest_tid
= 0;
3462 mdsc
->request_tree
= RB_ROOT
;
3463 INIT_DELAYED_WORK(&mdsc
->delayed_work
, delayed_work
);
3464 mdsc
->last_renew_caps
= jiffies
;
3465 INIT_LIST_HEAD(&mdsc
->cap_delay_list
);
3466 spin_lock_init(&mdsc
->cap_delay_lock
);
3467 INIT_LIST_HEAD(&mdsc
->snap_flush_list
);
3468 spin_lock_init(&mdsc
->snap_flush_lock
);
3469 mdsc
->last_cap_flush_tid
= 1;
3470 mdsc
->cap_flush_tree
= RB_ROOT
;
3471 INIT_LIST_HEAD(&mdsc
->cap_dirty
);
3472 INIT_LIST_HEAD(&mdsc
->cap_dirty_migrating
);
3473 mdsc
->num_cap_flushing
= 0;
3474 spin_lock_init(&mdsc
->cap_dirty_lock
);
3475 init_waitqueue_head(&mdsc
->cap_flushing_wq
);
3476 spin_lock_init(&mdsc
->dentry_lru_lock
);
3477 INIT_LIST_HEAD(&mdsc
->dentry_lru
);
3479 ceph_caps_init(mdsc
);
3480 ceph_adjust_min_caps(mdsc
, fsc
->min_caps
);
3482 init_rwsem(&mdsc
->pool_perm_rwsem
);
3483 mdsc
->pool_perm_tree
= RB_ROOT
;
3489 * Wait for safe replies on open mds requests. If we time out, drop
3490 * all requests from the tree to avoid dangling dentry refs.
3492 static void wait_requests(struct ceph_mds_client
*mdsc
)
3494 struct ceph_options
*opts
= mdsc
->fsc
->client
->options
;
3495 struct ceph_mds_request
*req
;
3497 mutex_lock(&mdsc
->mutex
);
3498 if (__get_oldest_req(mdsc
)) {
3499 mutex_unlock(&mdsc
->mutex
);
3501 dout("wait_requests waiting for requests\n");
3502 wait_for_completion_timeout(&mdsc
->safe_umount_waiters
,
3503 ceph_timeout_jiffies(opts
->mount_timeout
));
3505 /* tear down remaining requests */
3506 mutex_lock(&mdsc
->mutex
);
3507 while ((req
= __get_oldest_req(mdsc
))) {
3508 dout("wait_requests timed out on tid %llu\n",
3510 __unregister_request(mdsc
, req
);
3513 mutex_unlock(&mdsc
->mutex
);
3514 dout("wait_requests done\n");
3518 * called before mount is ro, and before dentries are torn down.
3519 * (hmm, does this still race with new lookups?)
3521 void ceph_mdsc_pre_umount(struct ceph_mds_client
*mdsc
)
3523 dout("pre_umount\n");
3527 ceph_flush_dirty_caps(mdsc
);
3528 wait_requests(mdsc
);
3531 * wait for reply handlers to drop their request refs and
3532 * their inode/dcache refs
3538 * wait for all write mds requests to flush.
3540 static void wait_unsafe_requests(struct ceph_mds_client
*mdsc
, u64 want_tid
)
3542 struct ceph_mds_request
*req
= NULL
, *nextreq
;
3545 mutex_lock(&mdsc
->mutex
);
3546 dout("wait_unsafe_requests want %lld\n", want_tid
);
3548 req
= __get_oldest_req(mdsc
);
3549 while (req
&& req
->r_tid
<= want_tid
) {
3550 /* find next request */
3551 n
= rb_next(&req
->r_node
);
3553 nextreq
= rb_entry(n
, struct ceph_mds_request
, r_node
);
3556 if (req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
&&
3557 (req
->r_op
& CEPH_MDS_OP_WRITE
)) {
3559 ceph_mdsc_get_request(req
);
3561 ceph_mdsc_get_request(nextreq
);
3562 mutex_unlock(&mdsc
->mutex
);
3563 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3564 req
->r_tid
, want_tid
);
3565 wait_for_completion(&req
->r_safe_completion
);
3566 mutex_lock(&mdsc
->mutex
);
3567 ceph_mdsc_put_request(req
);
3569 break; /* next dne before, so we're done! */
3570 if (RB_EMPTY_NODE(&nextreq
->r_node
)) {
3571 /* next request was removed from tree */
3572 ceph_mdsc_put_request(nextreq
);
3575 ceph_mdsc_put_request(nextreq
); /* won't go away */
3579 mutex_unlock(&mdsc
->mutex
);
3580 dout("wait_unsafe_requests done\n");
3583 void ceph_mdsc_sync(struct ceph_mds_client
*mdsc
)
3585 u64 want_tid
, want_flush
, want_snap
;
3587 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
3591 mutex_lock(&mdsc
->mutex
);
3592 want_tid
= mdsc
->last_tid
;
3593 mutex_unlock(&mdsc
->mutex
);
3595 ceph_flush_dirty_caps(mdsc
);
3596 spin_lock(&mdsc
->cap_dirty_lock
);
3597 want_flush
= mdsc
->last_cap_flush_tid
;
3598 spin_unlock(&mdsc
->cap_dirty_lock
);
3600 down_read(&mdsc
->snap_rwsem
);
3601 want_snap
= mdsc
->last_snap_seq
;
3602 up_read(&mdsc
->snap_rwsem
);
3604 dout("sync want tid %lld flush_seq %lld snap_seq %lld\n",
3605 want_tid
, want_flush
, want_snap
);
3607 wait_unsafe_requests(mdsc
, want_tid
);
3608 wait_caps_flush(mdsc
, want_flush
, want_snap
);
3612 * true if all sessions are closed, or we force unmount
3614 static bool done_closing_sessions(struct ceph_mds_client
*mdsc
)
3616 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
3618 return atomic_read(&mdsc
->num_sessions
) == 0;
3622 * called after sb is ro.
3624 void ceph_mdsc_close_sessions(struct ceph_mds_client
*mdsc
)
3626 struct ceph_options
*opts
= mdsc
->fsc
->client
->options
;
3627 struct ceph_mds_session
*session
;
3630 dout("close_sessions\n");
3632 /* close sessions */
3633 mutex_lock(&mdsc
->mutex
);
3634 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3635 session
= __ceph_lookup_mds_session(mdsc
, i
);
3638 mutex_unlock(&mdsc
->mutex
);
3639 mutex_lock(&session
->s_mutex
);
3640 __close_session(mdsc
, session
);
3641 mutex_unlock(&session
->s_mutex
);
3642 ceph_put_mds_session(session
);
3643 mutex_lock(&mdsc
->mutex
);
3645 mutex_unlock(&mdsc
->mutex
);
3647 dout("waiting for sessions to close\n");
3648 wait_event_timeout(mdsc
->session_close_wq
, done_closing_sessions(mdsc
),
3649 ceph_timeout_jiffies(opts
->mount_timeout
));
3651 /* tear down remaining sessions */
3652 mutex_lock(&mdsc
->mutex
);
3653 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3654 if (mdsc
->sessions
[i
]) {
3655 session
= get_session(mdsc
->sessions
[i
]);
3656 __unregister_session(mdsc
, session
);
3657 mutex_unlock(&mdsc
->mutex
);
3658 mutex_lock(&session
->s_mutex
);
3659 remove_session_caps(session
);
3660 mutex_unlock(&session
->s_mutex
);
3661 ceph_put_mds_session(session
);
3662 mutex_lock(&mdsc
->mutex
);
3665 WARN_ON(!list_empty(&mdsc
->cap_delay_list
));
3666 mutex_unlock(&mdsc
->mutex
);
3668 ceph_cleanup_empty_realms(mdsc
);
3670 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3675 void ceph_mdsc_force_umount(struct ceph_mds_client
*mdsc
)
3677 struct ceph_mds_session
*session
;
3680 dout("force umount\n");
3682 mutex_lock(&mdsc
->mutex
);
3683 for (mds
= 0; mds
< mdsc
->max_sessions
; mds
++) {
3684 session
= __ceph_lookup_mds_session(mdsc
, mds
);
3687 mutex_unlock(&mdsc
->mutex
);
3688 mutex_lock(&session
->s_mutex
);
3689 __close_session(mdsc
, session
);
3690 if (session
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3691 cleanup_session_requests(mdsc
, session
);
3692 remove_session_caps(session
);
3694 mutex_unlock(&session
->s_mutex
);
3695 ceph_put_mds_session(session
);
3696 mutex_lock(&mdsc
->mutex
);
3697 kick_requests(mdsc
, mds
);
3699 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3700 mutex_unlock(&mdsc
->mutex
);
3703 static void ceph_mdsc_stop(struct ceph_mds_client
*mdsc
)
3706 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3708 ceph_mdsmap_destroy(mdsc
->mdsmap
);
3709 kfree(mdsc
->sessions
);
3710 ceph_caps_finalize(mdsc
);
3711 ceph_pool_perm_destroy(mdsc
);
3714 void ceph_mdsc_destroy(struct ceph_fs_client
*fsc
)
3716 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
3718 dout("mdsc_destroy %p\n", mdsc
);
3719 ceph_mdsc_stop(mdsc
);
3721 /* flush out any connection work with references to us */
3726 dout("mdsc_destroy %p done\n", mdsc
);
3731 * handle mds map update.
3733 void ceph_mdsc_handle_map(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3737 void *p
= msg
->front
.iov_base
;
3738 void *end
= p
+ msg
->front
.iov_len
;
3739 struct ceph_mdsmap
*newmap
, *oldmap
;
3740 struct ceph_fsid fsid
;
3743 ceph_decode_need(&p
, end
, sizeof(fsid
)+2*sizeof(u32
), bad
);
3744 ceph_decode_copy(&p
, &fsid
, sizeof(fsid
));
3745 if (ceph_check_fsid(mdsc
->fsc
->client
, &fsid
) < 0)
3747 epoch
= ceph_decode_32(&p
);
3748 maplen
= ceph_decode_32(&p
);
3749 dout("handle_map epoch %u len %d\n", epoch
, (int)maplen
);
3751 /* do we need it? */
3752 mutex_lock(&mdsc
->mutex
);
3753 if (mdsc
->mdsmap
&& epoch
<= mdsc
->mdsmap
->m_epoch
) {
3754 dout("handle_map epoch %u <= our %u\n",
3755 epoch
, mdsc
->mdsmap
->m_epoch
);
3756 mutex_unlock(&mdsc
->mutex
);
3760 newmap
= ceph_mdsmap_decode(&p
, end
);
3761 if (IS_ERR(newmap
)) {
3762 err
= PTR_ERR(newmap
);
3766 /* swap into place */
3768 oldmap
= mdsc
->mdsmap
;
3769 mdsc
->mdsmap
= newmap
;
3770 check_new_map(mdsc
, newmap
, oldmap
);
3771 ceph_mdsmap_destroy(oldmap
);
3773 mdsc
->mdsmap
= newmap
; /* first mds map */
3775 mdsc
->fsc
->sb
->s_maxbytes
= mdsc
->mdsmap
->m_max_file_size
;
3777 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3778 ceph_monc_got_map(&mdsc
->fsc
->client
->monc
, CEPH_SUB_MDSMAP
,
3779 mdsc
->mdsmap
->m_epoch
);
3781 mutex_unlock(&mdsc
->mutex
);
3782 schedule_delayed(mdsc
);
3786 mutex_unlock(&mdsc
->mutex
);
3788 pr_err("error decoding mdsmap %d\n", err
);
3792 static struct ceph_connection
*con_get(struct ceph_connection
*con
)
3794 struct ceph_mds_session
*s
= con
->private;
3796 if (get_session(s
)) {
3797 dout("mdsc con_get %p ok (%d)\n", s
, atomic_read(&s
->s_ref
));
3800 dout("mdsc con_get %p FAIL\n", s
);
3804 static void con_put(struct ceph_connection
*con
)
3806 struct ceph_mds_session
*s
= con
->private;
3808 dout("mdsc con_put %p (%d)\n", s
, atomic_read(&s
->s_ref
) - 1);
3809 ceph_put_mds_session(s
);
3813 * if the client is unresponsive for long enough, the mds will kill
3814 * the session entirely.
3816 static void peer_reset(struct ceph_connection
*con
)
3818 struct ceph_mds_session
*s
= con
->private;
3819 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3821 pr_warn("mds%d closed our session\n", s
->s_mds
);
3822 send_mds_reconnect(mdsc
, s
);
3825 static void dispatch(struct ceph_connection
*con
, struct ceph_msg
*msg
)
3827 struct ceph_mds_session
*s
= con
->private;
3828 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3829 int type
= le16_to_cpu(msg
->hdr
.type
);
3831 mutex_lock(&mdsc
->mutex
);
3832 if (__verify_registered_session(mdsc
, s
) < 0) {
3833 mutex_unlock(&mdsc
->mutex
);
3836 mutex_unlock(&mdsc
->mutex
);
3839 case CEPH_MSG_MDS_MAP
:
3840 ceph_mdsc_handle_map(mdsc
, msg
);
3842 case CEPH_MSG_CLIENT_SESSION
:
3843 handle_session(s
, msg
);
3845 case CEPH_MSG_CLIENT_REPLY
:
3846 handle_reply(s
, msg
);
3848 case CEPH_MSG_CLIENT_REQUEST_FORWARD
:
3849 handle_forward(mdsc
, s
, msg
);
3851 case CEPH_MSG_CLIENT_CAPS
:
3852 ceph_handle_caps(s
, msg
);
3854 case CEPH_MSG_CLIENT_SNAP
:
3855 ceph_handle_snap(mdsc
, s
, msg
);
3857 case CEPH_MSG_CLIENT_LEASE
:
3858 handle_lease(mdsc
, s
, msg
);
3862 pr_err("received unknown message type %d %s\n", type
,
3863 ceph_msg_type_name(type
));
3874 * Note: returned pointer is the address of a structure that's
3875 * managed separately. Caller must *not* attempt to free it.
3877 static struct ceph_auth_handshake
*get_authorizer(struct ceph_connection
*con
,
3878 int *proto
, int force_new
)
3880 struct ceph_mds_session
*s
= con
->private;
3881 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3882 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3883 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
3885 if (force_new
&& auth
->authorizer
) {
3886 ceph_auth_destroy_authorizer(auth
->authorizer
);
3887 auth
->authorizer
= NULL
;
3889 if (!auth
->authorizer
) {
3890 int ret
= ceph_auth_create_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3893 return ERR_PTR(ret
);
3895 int ret
= ceph_auth_update_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3898 return ERR_PTR(ret
);
3900 *proto
= ac
->protocol
;
3906 static int verify_authorizer_reply(struct ceph_connection
*con
, int len
)
3908 struct ceph_mds_session
*s
= con
->private;
3909 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3910 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3912 return ceph_auth_verify_authorizer_reply(ac
, s
->s_auth
.authorizer
, len
);
3915 static int invalidate_authorizer(struct ceph_connection
*con
)
3917 struct ceph_mds_session
*s
= con
->private;
3918 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3919 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3921 ceph_auth_invalidate_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
);
3923 return ceph_monc_validate_auth(&mdsc
->fsc
->client
->monc
);
3926 static struct ceph_msg
*mds_alloc_msg(struct ceph_connection
*con
,
3927 struct ceph_msg_header
*hdr
, int *skip
)
3929 struct ceph_msg
*msg
;
3930 int type
= (int) le16_to_cpu(hdr
->type
);
3931 int front_len
= (int) le32_to_cpu(hdr
->front_len
);
3937 msg
= ceph_msg_new(type
, front_len
, GFP_NOFS
, false);
3939 pr_err("unable to allocate msg type %d len %d\n",
3947 static int mds_sign_message(struct ceph_msg
*msg
)
3949 struct ceph_mds_session
*s
= msg
->con
->private;
3950 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
3952 return ceph_auth_sign_message(auth
, msg
);
3955 static int mds_check_message_signature(struct ceph_msg
*msg
)
3957 struct ceph_mds_session
*s
= msg
->con
->private;
3958 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
3960 return ceph_auth_check_message_signature(auth
, msg
);
3963 static const struct ceph_connection_operations mds_con_ops
= {
3966 .dispatch
= dispatch
,
3967 .get_authorizer
= get_authorizer
,
3968 .verify_authorizer_reply
= verify_authorizer_reply
,
3969 .invalidate_authorizer
= invalidate_authorizer
,
3970 .peer_reset
= peer_reset
,
3971 .alloc_msg
= mds_alloc_msg
,
3972 .sign_message
= mds_sign_message
,
3973 .check_message_signature
= mds_check_message_signature
,