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45 * are no longer valid.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
);
184 info
->dir_end
= ceph_decode_8(p
);
185 info
->dir_complete
= ceph_decode_8(p
);
189 BUG_ON(!info
->dir_in
);
190 info
->dir_dname
= (void *)(info
->dir_in
+ num
);
191 info
->dir_dname_len
= (void *)(info
->dir_dname
+ num
);
192 info
->dir_dlease
= (void *)(info
->dir_dname_len
+ num
);
193 if ((unsigned long)(info
->dir_dlease
+ num
) >
194 (unsigned long)info
->dir_in
+ info
->dir_buf_size
) {
195 pr_err("dir contents are larger than expected\n");
203 ceph_decode_need(p
, end
, sizeof(u32
)*2, bad
);
204 info
->dir_dname_len
[i
] = ceph_decode_32(p
);
205 ceph_decode_need(p
, end
, info
->dir_dname_len
[i
], bad
);
206 info
->dir_dname
[i
] = *p
;
207 *p
+= info
->dir_dname_len
[i
];
208 dout("parsed dir dname '%.*s'\n", info
->dir_dname_len
[i
],
210 info
->dir_dlease
[i
] = *p
;
211 *p
+= sizeof(struct ceph_mds_reply_lease
);
214 err
= parse_reply_info_in(p
, end
, &info
->dir_in
[i
], 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
)
350 free_pages((unsigned long)info
->dir_in
, 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
;
844 const char* metadata
[][2] = {
845 {"hostname", utsname()->nodename
},
846 {"kernel_version", utsname()->release
},
847 {"entity_id", opt
->name
? opt
->name
: ""},
851 /* Calculate serialized length of metadata */
852 metadata_bytes
= 4; /* map length */
853 for (i
= 0; metadata
[i
][0] != NULL
; ++i
) {
854 metadata_bytes
+= 8 + strlen(metadata
[i
][0]) +
855 strlen(metadata
[i
][1]);
856 metadata_key_count
++;
859 /* Allocate the message */
860 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
) + metadata_bytes
,
863 pr_err("create_session_msg ENOMEM creating msg\n");
866 h
= msg
->front
.iov_base
;
867 h
->op
= cpu_to_le32(CEPH_SESSION_REQUEST_OPEN
);
868 h
->seq
= cpu_to_le64(seq
);
871 * Serialize client metadata into waiting buffer space, using
872 * the format that userspace expects for map<string, string>
874 * ClientSession messages with metadata are v2
876 msg
->hdr
.version
= cpu_to_le16(2);
877 msg
->hdr
.compat_version
= cpu_to_le16(1);
879 /* The write pointer, following the session_head structure */
880 p
= msg
->front
.iov_base
+ sizeof(*h
);
882 /* Number of entries in the map */
883 ceph_encode_32(&p
, metadata_key_count
);
885 /* Two length-prefixed strings for each entry in the map */
886 for (i
= 0; metadata
[i
][0] != NULL
; ++i
) {
887 size_t const key_len
= strlen(metadata
[i
][0]);
888 size_t const val_len
= strlen(metadata
[i
][1]);
890 ceph_encode_32(&p
, key_len
);
891 memcpy(p
, metadata
[i
][0], key_len
);
893 ceph_encode_32(&p
, val_len
);
894 memcpy(p
, metadata
[i
][1], val_len
);
902 * send session open request.
904 * called under mdsc->mutex
906 static int __open_session(struct ceph_mds_client
*mdsc
,
907 struct ceph_mds_session
*session
)
909 struct ceph_msg
*msg
;
911 int mds
= session
->s_mds
;
913 /* wait for mds to go active? */
914 mstate
= ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
);
915 dout("open_session to mds%d (%s)\n", mds
,
916 ceph_mds_state_name(mstate
));
917 session
->s_state
= CEPH_MDS_SESSION_OPENING
;
918 session
->s_renew_requested
= jiffies
;
920 /* send connect message */
921 msg
= create_session_open_msg(mdsc
, session
->s_seq
);
924 ceph_con_send(&session
->s_con
, msg
);
929 * open sessions for any export targets for the given mds
931 * called under mdsc->mutex
933 static struct ceph_mds_session
*
934 __open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
936 struct ceph_mds_session
*session
;
938 session
= __ceph_lookup_mds_session(mdsc
, target
);
940 session
= register_session(mdsc
, target
);
944 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
945 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
946 __open_session(mdsc
, session
);
951 struct ceph_mds_session
*
952 ceph_mdsc_open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
954 struct ceph_mds_session
*session
;
956 dout("open_export_target_session to mds%d\n", target
);
958 mutex_lock(&mdsc
->mutex
);
959 session
= __open_export_target_session(mdsc
, target
);
960 mutex_unlock(&mdsc
->mutex
);
965 static void __open_export_target_sessions(struct ceph_mds_client
*mdsc
,
966 struct ceph_mds_session
*session
)
968 struct ceph_mds_info
*mi
;
969 struct ceph_mds_session
*ts
;
970 int i
, mds
= session
->s_mds
;
972 if (mds
>= mdsc
->mdsmap
->m_max_mds
)
975 mi
= &mdsc
->mdsmap
->m_info
[mds
];
976 dout("open_export_target_sessions for mds%d (%d targets)\n",
977 session
->s_mds
, mi
->num_export_targets
);
979 for (i
= 0; i
< mi
->num_export_targets
; i
++) {
980 ts
= __open_export_target_session(mdsc
, mi
->export_targets
[i
]);
982 ceph_put_mds_session(ts
);
986 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client
*mdsc
,
987 struct ceph_mds_session
*session
)
989 mutex_lock(&mdsc
->mutex
);
990 __open_export_target_sessions(mdsc
, session
);
991 mutex_unlock(&mdsc
->mutex
);
998 /* caller holds s_cap_lock, we drop it */
999 static void cleanup_cap_releases(struct ceph_mds_client
*mdsc
,
1000 struct ceph_mds_session
*session
)
1001 __releases(session
->s_cap_lock
)
1003 LIST_HEAD(tmp_list
);
1004 list_splice_init(&session
->s_cap_releases
, &tmp_list
);
1005 session
->s_num_cap_releases
= 0;
1006 spin_unlock(&session
->s_cap_lock
);
1008 dout("cleanup_cap_releases mds%d\n", session
->s_mds
);
1009 while (!list_empty(&tmp_list
)) {
1010 struct ceph_cap
*cap
;
1011 /* zero out the in-progress message */
1012 cap
= list_first_entry(&tmp_list
,
1013 struct ceph_cap
, session_caps
);
1014 list_del(&cap
->session_caps
);
1015 ceph_put_cap(mdsc
, cap
);
1019 static void cleanup_session_requests(struct ceph_mds_client
*mdsc
,
1020 struct ceph_mds_session
*session
)
1022 struct ceph_mds_request
*req
;
1025 dout("cleanup_session_requests mds%d\n", session
->s_mds
);
1026 mutex_lock(&mdsc
->mutex
);
1027 while (!list_empty(&session
->s_unsafe
)) {
1028 req
= list_first_entry(&session
->s_unsafe
,
1029 struct ceph_mds_request
, r_unsafe_item
);
1030 list_del_init(&req
->r_unsafe_item
);
1031 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1033 __unregister_request(mdsc
, req
);
1035 /* zero r_attempts, so kick_requests() will re-send requests */
1036 p
= rb_first(&mdsc
->request_tree
);
1038 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
1040 if (req
->r_session
&&
1041 req
->r_session
->s_mds
== session
->s_mds
)
1042 req
->r_attempts
= 0;
1044 mutex_unlock(&mdsc
->mutex
);
1048 * Helper to safely iterate over all caps associated with a session, with
1049 * special care taken to handle a racing __ceph_remove_cap().
1051 * Caller must hold session s_mutex.
1053 static int iterate_session_caps(struct ceph_mds_session
*session
,
1054 int (*cb
)(struct inode
*, struct ceph_cap
*,
1057 struct list_head
*p
;
1058 struct ceph_cap
*cap
;
1059 struct inode
*inode
, *last_inode
= NULL
;
1060 struct ceph_cap
*old_cap
= NULL
;
1063 dout("iterate_session_caps %p mds%d\n", session
, session
->s_mds
);
1064 spin_lock(&session
->s_cap_lock
);
1065 p
= session
->s_caps
.next
;
1066 while (p
!= &session
->s_caps
) {
1067 cap
= list_entry(p
, struct ceph_cap
, session_caps
);
1068 inode
= igrab(&cap
->ci
->vfs_inode
);
1073 session
->s_cap_iterator
= cap
;
1074 spin_unlock(&session
->s_cap_lock
);
1081 ceph_put_cap(session
->s_mdsc
, old_cap
);
1085 ret
= cb(inode
, cap
, arg
);
1088 spin_lock(&session
->s_cap_lock
);
1090 if (cap
->ci
== NULL
) {
1091 dout("iterate_session_caps finishing cap %p removal\n",
1093 BUG_ON(cap
->session
!= session
);
1094 cap
->session
= NULL
;
1095 list_del_init(&cap
->session_caps
);
1096 session
->s_nr_caps
--;
1097 if (cap
->queue_release
) {
1098 list_add_tail(&cap
->session_caps
,
1099 &session
->s_cap_releases
);
1100 session
->s_num_cap_releases
++;
1102 old_cap
= cap
; /* put_cap it w/o locks held */
1110 session
->s_cap_iterator
= NULL
;
1111 spin_unlock(&session
->s_cap_lock
);
1115 ceph_put_cap(session
->s_mdsc
, old_cap
);
1120 static int remove_session_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1123 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1124 LIST_HEAD(to_remove
);
1127 dout("removing cap %p, ci is %p, inode is %p\n",
1128 cap
, ci
, &ci
->vfs_inode
);
1129 spin_lock(&ci
->i_ceph_lock
);
1130 __ceph_remove_cap(cap
, false);
1131 if (!ci
->i_auth_cap
) {
1132 struct ceph_cap_flush
*cf
;
1133 struct ceph_mds_client
*mdsc
=
1134 ceph_sb_to_client(inode
->i_sb
)->mdsc
;
1136 ci
->i_ceph_flags
|= CEPH_I_CAP_DROPPED
;
1139 struct rb_node
*n
= rb_first(&ci
->i_cap_flush_tree
);
1142 cf
= rb_entry(n
, struct ceph_cap_flush
, i_node
);
1143 rb_erase(&cf
->i_node
, &ci
->i_cap_flush_tree
);
1144 list_add(&cf
->list
, &to_remove
);
1147 spin_lock(&mdsc
->cap_dirty_lock
);
1149 list_for_each_entry(cf
, &to_remove
, list
)
1150 rb_erase(&cf
->g_node
, &mdsc
->cap_flush_tree
);
1152 if (!list_empty(&ci
->i_dirty_item
)) {
1153 pr_warn_ratelimited(
1154 " dropping dirty %s state for %p %lld\n",
1155 ceph_cap_string(ci
->i_dirty_caps
),
1156 inode
, ceph_ino(inode
));
1157 ci
->i_dirty_caps
= 0;
1158 list_del_init(&ci
->i_dirty_item
);
1161 if (!list_empty(&ci
->i_flushing_item
)) {
1162 pr_warn_ratelimited(
1163 " dropping dirty+flushing %s state for %p %lld\n",
1164 ceph_cap_string(ci
->i_flushing_caps
),
1165 inode
, ceph_ino(inode
));
1166 ci
->i_flushing_caps
= 0;
1167 list_del_init(&ci
->i_flushing_item
);
1168 mdsc
->num_cap_flushing
--;
1171 spin_unlock(&mdsc
->cap_dirty_lock
);
1173 if (!ci
->i_dirty_caps
&& ci
->i_prealloc_cap_flush
) {
1174 list_add(&ci
->i_prealloc_cap_flush
->list
, &to_remove
);
1175 ci
->i_prealloc_cap_flush
= NULL
;
1178 spin_unlock(&ci
->i_ceph_lock
);
1179 while (!list_empty(&to_remove
)) {
1180 struct ceph_cap_flush
*cf
;
1181 cf
= list_first_entry(&to_remove
,
1182 struct ceph_cap_flush
, list
);
1183 list_del(&cf
->list
);
1184 ceph_free_cap_flush(cf
);
1187 wake_up_all(&ci
->i_cap_wq
);
1194 * caller must hold session s_mutex
1196 static void remove_session_caps(struct ceph_mds_session
*session
)
1198 dout("remove_session_caps on %p\n", session
);
1199 iterate_session_caps(session
, remove_session_caps_cb
, NULL
);
1201 spin_lock(&session
->s_cap_lock
);
1202 if (session
->s_nr_caps
> 0) {
1203 struct super_block
*sb
= session
->s_mdsc
->fsc
->sb
;
1204 struct inode
*inode
;
1205 struct ceph_cap
*cap
, *prev
= NULL
;
1206 struct ceph_vino vino
;
1208 * iterate_session_caps() skips inodes that are being
1209 * deleted, we need to wait until deletions are complete.
1210 * __wait_on_freeing_inode() is designed for the job,
1211 * but it is not exported, so use lookup inode function
1214 while (!list_empty(&session
->s_caps
)) {
1215 cap
= list_entry(session
->s_caps
.next
,
1216 struct ceph_cap
, session_caps
);
1220 vino
= cap
->ci
->i_vino
;
1221 spin_unlock(&session
->s_cap_lock
);
1223 inode
= ceph_find_inode(sb
, vino
);
1226 spin_lock(&session
->s_cap_lock
);
1230 // drop cap expires and unlock s_cap_lock
1231 cleanup_cap_releases(session
->s_mdsc
, session
);
1233 BUG_ON(session
->s_nr_caps
> 0);
1234 BUG_ON(!list_empty(&session
->s_cap_flushing
));
1238 * wake up any threads waiting on this session's caps. if the cap is
1239 * old (didn't get renewed on the client reconnect), remove it now.
1241 * caller must hold s_mutex.
1243 static int wake_up_session_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1246 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1248 wake_up_all(&ci
->i_cap_wq
);
1250 spin_lock(&ci
->i_ceph_lock
);
1251 ci
->i_wanted_max_size
= 0;
1252 ci
->i_requested_max_size
= 0;
1253 spin_unlock(&ci
->i_ceph_lock
);
1258 static void wake_up_session_caps(struct ceph_mds_session
*session
,
1261 dout("wake_up_session_caps %p mds%d\n", session
, session
->s_mds
);
1262 iterate_session_caps(session
, wake_up_session_cb
,
1263 (void *)(unsigned long)reconnect
);
1267 * Send periodic message to MDS renewing all currently held caps. The
1268 * ack will reset the expiration for all caps from this session.
1270 * caller holds s_mutex
1272 static int send_renew_caps(struct ceph_mds_client
*mdsc
,
1273 struct ceph_mds_session
*session
)
1275 struct ceph_msg
*msg
;
1278 if (time_after_eq(jiffies
, session
->s_cap_ttl
) &&
1279 time_after_eq(session
->s_cap_ttl
, session
->s_renew_requested
))
1280 pr_info("mds%d caps stale\n", session
->s_mds
);
1281 session
->s_renew_requested
= jiffies
;
1283 /* do not try to renew caps until a recovering mds has reconnected
1284 * with its clients. */
1285 state
= ceph_mdsmap_get_state(mdsc
->mdsmap
, session
->s_mds
);
1286 if (state
< CEPH_MDS_STATE_RECONNECT
) {
1287 dout("send_renew_caps ignoring mds%d (%s)\n",
1288 session
->s_mds
, ceph_mds_state_name(state
));
1292 dout("send_renew_caps to mds%d (%s)\n", session
->s_mds
,
1293 ceph_mds_state_name(state
));
1294 msg
= create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS
,
1295 ++session
->s_renew_seq
);
1298 ceph_con_send(&session
->s_con
, msg
);
1302 static int send_flushmsg_ack(struct ceph_mds_client
*mdsc
,
1303 struct ceph_mds_session
*session
, u64 seq
)
1305 struct ceph_msg
*msg
;
1307 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1308 session
->s_mds
, ceph_session_state_name(session
->s_state
), seq
);
1309 msg
= create_session_msg(CEPH_SESSION_FLUSHMSG_ACK
, seq
);
1312 ceph_con_send(&session
->s_con
, msg
);
1318 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1320 * Called under session->s_mutex
1322 static void renewed_caps(struct ceph_mds_client
*mdsc
,
1323 struct ceph_mds_session
*session
, int is_renew
)
1328 spin_lock(&session
->s_cap_lock
);
1329 was_stale
= is_renew
&& time_after_eq(jiffies
, session
->s_cap_ttl
);
1331 session
->s_cap_ttl
= session
->s_renew_requested
+
1332 mdsc
->mdsmap
->m_session_timeout
*HZ
;
1335 if (time_before(jiffies
, session
->s_cap_ttl
)) {
1336 pr_info("mds%d caps renewed\n", session
->s_mds
);
1339 pr_info("mds%d caps still stale\n", session
->s_mds
);
1342 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1343 session
->s_mds
, session
->s_cap_ttl
, was_stale
? "stale" : "fresh",
1344 time_before(jiffies
, session
->s_cap_ttl
) ? "stale" : "fresh");
1345 spin_unlock(&session
->s_cap_lock
);
1348 wake_up_session_caps(session
, 0);
1352 * send a session close request
1354 static int request_close_session(struct ceph_mds_client
*mdsc
,
1355 struct ceph_mds_session
*session
)
1357 struct ceph_msg
*msg
;
1359 dout("request_close_session mds%d state %s seq %lld\n",
1360 session
->s_mds
, ceph_session_state_name(session
->s_state
),
1362 msg
= create_session_msg(CEPH_SESSION_REQUEST_CLOSE
, session
->s_seq
);
1365 ceph_con_send(&session
->s_con
, msg
);
1370 * Called with s_mutex held.
1372 static int __close_session(struct ceph_mds_client
*mdsc
,
1373 struct ceph_mds_session
*session
)
1375 if (session
->s_state
>= CEPH_MDS_SESSION_CLOSING
)
1377 session
->s_state
= CEPH_MDS_SESSION_CLOSING
;
1378 return request_close_session(mdsc
, session
);
1382 * Trim old(er) caps.
1384 * Because we can't cache an inode without one or more caps, we do
1385 * this indirectly: if a cap is unused, we prune its aliases, at which
1386 * point the inode will hopefully get dropped to.
1388 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1389 * memory pressure from the MDS, though, so it needn't be perfect.
1391 static int trim_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
, void *arg
)
1393 struct ceph_mds_session
*session
= arg
;
1394 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1395 int used
, wanted
, oissued
, mine
;
1397 if (session
->s_trim_caps
<= 0)
1400 spin_lock(&ci
->i_ceph_lock
);
1401 mine
= cap
->issued
| cap
->implemented
;
1402 used
= __ceph_caps_used(ci
);
1403 wanted
= __ceph_caps_file_wanted(ci
);
1404 oissued
= __ceph_caps_issued_other(ci
, cap
);
1406 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1407 inode
, cap
, ceph_cap_string(mine
), ceph_cap_string(oissued
),
1408 ceph_cap_string(used
), ceph_cap_string(wanted
));
1409 if (cap
== ci
->i_auth_cap
) {
1410 if (ci
->i_dirty_caps
|| ci
->i_flushing_caps
||
1411 !list_empty(&ci
->i_cap_snaps
))
1413 if ((used
| wanted
) & CEPH_CAP_ANY_WR
)
1416 /* The inode has cached pages, but it's no longer used.
1417 * we can safely drop it */
1418 if (wanted
== 0 && used
== CEPH_CAP_FILE_CACHE
&&
1419 !(oissued
& CEPH_CAP_FILE_CACHE
)) {
1423 if ((used
| wanted
) & ~oissued
& mine
)
1424 goto out
; /* we need these caps */
1426 session
->s_trim_caps
--;
1428 /* we aren't the only cap.. just remove us */
1429 __ceph_remove_cap(cap
, true);
1431 /* try dropping referring dentries */
1432 spin_unlock(&ci
->i_ceph_lock
);
1433 d_prune_aliases(inode
);
1434 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1435 inode
, cap
, atomic_read(&inode
->i_count
));
1440 spin_unlock(&ci
->i_ceph_lock
);
1445 * Trim session cap count down to some max number.
1447 static int trim_caps(struct ceph_mds_client
*mdsc
,
1448 struct ceph_mds_session
*session
,
1451 int trim_caps
= session
->s_nr_caps
- max_caps
;
1453 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1454 session
->s_mds
, session
->s_nr_caps
, max_caps
, trim_caps
);
1455 if (trim_caps
> 0) {
1456 session
->s_trim_caps
= trim_caps
;
1457 iterate_session_caps(session
, trim_caps_cb
, session
);
1458 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1459 session
->s_mds
, session
->s_nr_caps
, max_caps
,
1460 trim_caps
- session
->s_trim_caps
);
1461 session
->s_trim_caps
= 0;
1464 ceph_send_cap_releases(mdsc
, session
);
1468 static int check_capsnap_flush(struct ceph_inode_info
*ci
,
1472 spin_lock(&ci
->i_ceph_lock
);
1473 if (want_snap_seq
> 0 && !list_empty(&ci
->i_cap_snaps
)) {
1474 struct ceph_cap_snap
*capsnap
=
1475 list_first_entry(&ci
->i_cap_snaps
,
1476 struct ceph_cap_snap
, ci_item
);
1477 ret
= capsnap
->follows
>= want_snap_seq
;
1479 spin_unlock(&ci
->i_ceph_lock
);
1483 static int check_caps_flush(struct ceph_mds_client
*mdsc
,
1487 struct ceph_cap_flush
*cf
;
1490 spin_lock(&mdsc
->cap_dirty_lock
);
1491 n
= rb_first(&mdsc
->cap_flush_tree
);
1492 cf
= n
? rb_entry(n
, struct ceph_cap_flush
, g_node
) : NULL
;
1493 if (cf
&& cf
->tid
<= want_flush_tid
) {
1494 dout("check_caps_flush still flushing tid %llu <= %llu\n",
1495 cf
->tid
, want_flush_tid
);
1498 spin_unlock(&mdsc
->cap_dirty_lock
);
1503 * flush all dirty inode data to disk.
1505 * returns true if we've flushed through want_flush_tid
1507 static void wait_caps_flush(struct ceph_mds_client
*mdsc
,
1508 u64 want_flush_tid
, u64 want_snap_seq
)
1512 dout("check_caps_flush want %llu snap want %llu\n",
1513 want_flush_tid
, want_snap_seq
);
1514 mutex_lock(&mdsc
->mutex
);
1515 for (mds
= 0; mds
< mdsc
->max_sessions
; ) {
1516 struct ceph_mds_session
*session
= mdsc
->sessions
[mds
];
1517 struct inode
*inode
= NULL
;
1523 get_session(session
);
1524 mutex_unlock(&mdsc
->mutex
);
1526 mutex_lock(&session
->s_mutex
);
1527 if (!list_empty(&session
->s_cap_snaps_flushing
)) {
1528 struct ceph_cap_snap
*capsnap
=
1529 list_first_entry(&session
->s_cap_snaps_flushing
,
1530 struct ceph_cap_snap
,
1532 struct ceph_inode_info
*ci
= capsnap
->ci
;
1533 if (!check_capsnap_flush(ci
, want_snap_seq
)) {
1534 dout("check_cap_flush still flushing snap %p "
1535 "follows %lld <= %lld to mds%d\n",
1536 &ci
->vfs_inode
, capsnap
->follows
,
1537 want_snap_seq
, mds
);
1538 inode
= igrab(&ci
->vfs_inode
);
1541 mutex_unlock(&session
->s_mutex
);
1542 ceph_put_mds_session(session
);
1545 wait_event(mdsc
->cap_flushing_wq
,
1546 check_capsnap_flush(ceph_inode(inode
),
1553 mutex_lock(&mdsc
->mutex
);
1555 mutex_unlock(&mdsc
->mutex
);
1557 wait_event(mdsc
->cap_flushing_wq
,
1558 check_caps_flush(mdsc
, want_flush_tid
));
1560 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid
);
1564 * called under s_mutex
1566 void ceph_send_cap_releases(struct ceph_mds_client
*mdsc
,
1567 struct ceph_mds_session
*session
)
1569 struct ceph_msg
*msg
= NULL
;
1570 struct ceph_mds_cap_release
*head
;
1571 struct ceph_mds_cap_item
*item
;
1572 struct ceph_cap
*cap
;
1573 LIST_HEAD(tmp_list
);
1574 int num_cap_releases
;
1576 spin_lock(&session
->s_cap_lock
);
1578 list_splice_init(&session
->s_cap_releases
, &tmp_list
);
1579 num_cap_releases
= session
->s_num_cap_releases
;
1580 session
->s_num_cap_releases
= 0;
1581 spin_unlock(&session
->s_cap_lock
);
1583 while (!list_empty(&tmp_list
)) {
1585 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE
,
1586 PAGE_SIZE
, GFP_NOFS
, false);
1589 head
= msg
->front
.iov_base
;
1590 head
->num
= cpu_to_le32(0);
1591 msg
->front
.iov_len
= sizeof(*head
);
1593 cap
= list_first_entry(&tmp_list
, struct ceph_cap
,
1595 list_del(&cap
->session_caps
);
1598 head
= msg
->front
.iov_base
;
1599 le32_add_cpu(&head
->num
, 1);
1600 item
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1601 item
->ino
= cpu_to_le64(cap
->cap_ino
);
1602 item
->cap_id
= cpu_to_le64(cap
->cap_id
);
1603 item
->migrate_seq
= cpu_to_le32(cap
->mseq
);
1604 item
->seq
= cpu_to_le32(cap
->issue_seq
);
1605 msg
->front
.iov_len
+= sizeof(*item
);
1607 ceph_put_cap(mdsc
, cap
);
1609 if (le32_to_cpu(head
->num
) == CEPH_CAPS_PER_RELEASE
) {
1610 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1611 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1612 ceph_con_send(&session
->s_con
, msg
);
1617 BUG_ON(num_cap_releases
!= 0);
1619 spin_lock(&session
->s_cap_lock
);
1620 if (!list_empty(&session
->s_cap_releases
))
1622 spin_unlock(&session
->s_cap_lock
);
1625 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1626 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1627 ceph_con_send(&session
->s_con
, msg
);
1631 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1633 spin_lock(&session
->s_cap_lock
);
1634 list_splice(&tmp_list
, &session
->s_cap_releases
);
1635 session
->s_num_cap_releases
+= num_cap_releases
;
1636 spin_unlock(&session
->s_cap_lock
);
1643 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request
*req
,
1646 struct ceph_inode_info
*ci
= ceph_inode(dir
);
1647 struct ceph_mds_reply_info_parsed
*rinfo
= &req
->r_reply_info
;
1648 struct ceph_mount_options
*opt
= req
->r_mdsc
->fsc
->mount_options
;
1649 size_t size
= sizeof(*rinfo
->dir_in
) + sizeof(*rinfo
->dir_dname_len
) +
1650 sizeof(*rinfo
->dir_dname
) + sizeof(*rinfo
->dir_dlease
);
1651 int order
, num_entries
;
1653 spin_lock(&ci
->i_ceph_lock
);
1654 num_entries
= ci
->i_files
+ ci
->i_subdirs
;
1655 spin_unlock(&ci
->i_ceph_lock
);
1656 num_entries
= max(num_entries
, 1);
1657 num_entries
= min(num_entries
, opt
->max_readdir
);
1659 order
= get_order(size
* num_entries
);
1660 while (order
>= 0) {
1661 rinfo
->dir_in
= (void*)__get_free_pages(GFP_KERNEL
|
1671 num_entries
= (PAGE_SIZE
<< order
) / size
;
1672 num_entries
= min(num_entries
, opt
->max_readdir
);
1674 rinfo
->dir_buf_size
= PAGE_SIZE
<< order
;
1675 req
->r_num_caps
= num_entries
+ 1;
1676 req
->r_args
.readdir
.max_entries
= cpu_to_le32(num_entries
);
1677 req
->r_args
.readdir
.max_bytes
= cpu_to_le32(opt
->max_readdir_bytes
);
1682 * Create an mds request.
1684 struct ceph_mds_request
*
1685 ceph_mdsc_create_request(struct ceph_mds_client
*mdsc
, int op
, int mode
)
1687 struct ceph_mds_request
*req
= kzalloc(sizeof(*req
), GFP_NOFS
);
1690 return ERR_PTR(-ENOMEM
);
1692 mutex_init(&req
->r_fill_mutex
);
1694 req
->r_started
= jiffies
;
1695 req
->r_resend_mds
= -1;
1696 INIT_LIST_HEAD(&req
->r_unsafe_dir_item
);
1697 INIT_LIST_HEAD(&req
->r_unsafe_target_item
);
1699 kref_init(&req
->r_kref
);
1700 RB_CLEAR_NODE(&req
->r_node
);
1701 INIT_LIST_HEAD(&req
->r_wait
);
1702 init_completion(&req
->r_completion
);
1703 init_completion(&req
->r_safe_completion
);
1704 INIT_LIST_HEAD(&req
->r_unsafe_item
);
1706 req
->r_stamp
= current_fs_time(mdsc
->fsc
->sb
);
1709 req
->r_direct_mode
= mode
;
1714 * return oldest (lowest) request, tid in request tree, 0 if none.
1716 * called under mdsc->mutex.
1718 static struct ceph_mds_request
*__get_oldest_req(struct ceph_mds_client
*mdsc
)
1720 if (RB_EMPTY_ROOT(&mdsc
->request_tree
))
1722 return rb_entry(rb_first(&mdsc
->request_tree
),
1723 struct ceph_mds_request
, r_node
);
1726 static inline u64
__get_oldest_tid(struct ceph_mds_client
*mdsc
)
1728 return mdsc
->oldest_tid
;
1732 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1733 * on build_path_from_dentry in fs/cifs/dir.c.
1735 * If @stop_on_nosnap, generate path relative to the first non-snapped
1738 * Encode hidden .snap dirs as a double /, i.e.
1739 * foo/.snap/bar -> foo//bar
1741 char *ceph_mdsc_build_path(struct dentry
*dentry
, int *plen
, u64
*base
,
1744 struct dentry
*temp
;
1750 return ERR_PTR(-EINVAL
);
1754 seq
= read_seqbegin(&rename_lock
);
1756 for (temp
= dentry
; !IS_ROOT(temp
);) {
1757 struct inode
*inode
= d_inode(temp
);
1758 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
)
1759 len
++; /* slash only */
1760 else if (stop_on_nosnap
&& inode
&&
1761 ceph_snap(inode
) == CEPH_NOSNAP
)
1764 len
+= 1 + temp
->d_name
.len
;
1765 temp
= temp
->d_parent
;
1769 len
--; /* no leading '/' */
1771 path
= kmalloc(len
+1, GFP_NOFS
);
1773 return ERR_PTR(-ENOMEM
);
1775 path
[pos
] = 0; /* trailing null */
1777 for (temp
= dentry
; !IS_ROOT(temp
) && pos
!= 0; ) {
1778 struct inode
*inode
;
1780 spin_lock(&temp
->d_lock
);
1781 inode
= d_inode(temp
);
1782 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
) {
1783 dout("build_path path+%d: %p SNAPDIR\n",
1785 } else if (stop_on_nosnap
&& inode
&&
1786 ceph_snap(inode
) == CEPH_NOSNAP
) {
1787 spin_unlock(&temp
->d_lock
);
1790 pos
-= temp
->d_name
.len
;
1792 spin_unlock(&temp
->d_lock
);
1795 strncpy(path
+ pos
, temp
->d_name
.name
,
1798 spin_unlock(&temp
->d_lock
);
1801 temp
= temp
->d_parent
;
1804 if (pos
!= 0 || read_seqretry(&rename_lock
, seq
)) {
1805 pr_err("build_path did not end path lookup where "
1806 "expected, namelen is %d, pos is %d\n", len
, pos
);
1807 /* presumably this is only possible if racing with a
1808 rename of one of the parent directories (we can not
1809 lock the dentries above us to prevent this, but
1810 retrying should be harmless) */
1815 *base
= ceph_ino(d_inode(temp
));
1817 dout("build_path on %p %d built %llx '%.*s'\n",
1818 dentry
, d_count(dentry
), *base
, len
, path
);
1822 static int build_dentry_path(struct dentry
*dentry
,
1823 const char **ppath
, int *ppathlen
, u64
*pino
,
1828 if (ceph_snap(d_inode(dentry
->d_parent
)) == CEPH_NOSNAP
) {
1829 *pino
= ceph_ino(d_inode(dentry
->d_parent
));
1830 *ppath
= dentry
->d_name
.name
;
1831 *ppathlen
= dentry
->d_name
.len
;
1834 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1836 return PTR_ERR(path
);
1842 static int build_inode_path(struct inode
*inode
,
1843 const char **ppath
, int *ppathlen
, u64
*pino
,
1846 struct dentry
*dentry
;
1849 if (ceph_snap(inode
) == CEPH_NOSNAP
) {
1850 *pino
= ceph_ino(inode
);
1854 dentry
= d_find_alias(inode
);
1855 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1858 return PTR_ERR(path
);
1865 * request arguments may be specified via an inode *, a dentry *, or
1866 * an explicit ino+path.
1868 static int set_request_path_attr(struct inode
*rinode
, struct dentry
*rdentry
,
1869 const char *rpath
, u64 rino
,
1870 const char **ppath
, int *pathlen
,
1871 u64
*ino
, int *freepath
)
1876 r
= build_inode_path(rinode
, ppath
, pathlen
, ino
, freepath
);
1877 dout(" inode %p %llx.%llx\n", rinode
, ceph_ino(rinode
),
1879 } else if (rdentry
) {
1880 r
= build_dentry_path(rdentry
, ppath
, pathlen
, ino
, freepath
);
1881 dout(" dentry %p %llx/%.*s\n", rdentry
, *ino
, *pathlen
,
1883 } else if (rpath
|| rino
) {
1886 *pathlen
= rpath
? strlen(rpath
) : 0;
1887 dout(" path %.*s\n", *pathlen
, rpath
);
1894 * called under mdsc->mutex
1896 static struct ceph_msg
*create_request_message(struct ceph_mds_client
*mdsc
,
1897 struct ceph_mds_request
*req
,
1898 int mds
, bool drop_cap_releases
)
1900 struct ceph_msg
*msg
;
1901 struct ceph_mds_request_head
*head
;
1902 const char *path1
= NULL
;
1903 const char *path2
= NULL
;
1904 u64 ino1
= 0, ino2
= 0;
1905 int pathlen1
= 0, pathlen2
= 0;
1906 int freepath1
= 0, freepath2
= 0;
1912 ret
= set_request_path_attr(req
->r_inode
, req
->r_dentry
,
1913 req
->r_path1
, req
->r_ino1
.ino
,
1914 &path1
, &pathlen1
, &ino1
, &freepath1
);
1920 ret
= set_request_path_attr(NULL
, req
->r_old_dentry
,
1921 req
->r_path2
, req
->r_ino2
.ino
,
1922 &path2
, &pathlen2
, &ino2
, &freepath2
);
1928 len
= sizeof(*head
) +
1929 pathlen1
+ pathlen2
+ 2*(1 + sizeof(u32
) + sizeof(u64
)) +
1930 sizeof(struct ceph_timespec
);
1932 /* calculate (max) length for cap releases */
1933 len
+= sizeof(struct ceph_mds_request_release
) *
1934 (!!req
->r_inode_drop
+ !!req
->r_dentry_drop
+
1935 !!req
->r_old_inode_drop
+ !!req
->r_old_dentry_drop
);
1936 if (req
->r_dentry_drop
)
1937 len
+= req
->r_dentry
->d_name
.len
;
1938 if (req
->r_old_dentry_drop
)
1939 len
+= req
->r_old_dentry
->d_name
.len
;
1941 msg
= ceph_msg_new(CEPH_MSG_CLIENT_REQUEST
, len
, GFP_NOFS
, false);
1943 msg
= ERR_PTR(-ENOMEM
);
1947 msg
->hdr
.version
= cpu_to_le16(2);
1948 msg
->hdr
.tid
= cpu_to_le64(req
->r_tid
);
1950 head
= msg
->front
.iov_base
;
1951 p
= msg
->front
.iov_base
+ sizeof(*head
);
1952 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1954 head
->mdsmap_epoch
= cpu_to_le32(mdsc
->mdsmap
->m_epoch
);
1955 head
->op
= cpu_to_le32(req
->r_op
);
1956 head
->caller_uid
= cpu_to_le32(from_kuid(&init_user_ns
, req
->r_uid
));
1957 head
->caller_gid
= cpu_to_le32(from_kgid(&init_user_ns
, req
->r_gid
));
1958 head
->args
= req
->r_args
;
1960 ceph_encode_filepath(&p
, end
, ino1
, path1
);
1961 ceph_encode_filepath(&p
, end
, ino2
, path2
);
1963 /* make note of release offset, in case we need to replay */
1964 req
->r_request_release_offset
= p
- msg
->front
.iov_base
;
1968 if (req
->r_inode_drop
)
1969 releases
+= ceph_encode_inode_release(&p
,
1970 req
->r_inode
? req
->r_inode
: d_inode(req
->r_dentry
),
1971 mds
, req
->r_inode_drop
, req
->r_inode_unless
, 0);
1972 if (req
->r_dentry_drop
)
1973 releases
+= ceph_encode_dentry_release(&p
, req
->r_dentry
,
1974 mds
, req
->r_dentry_drop
, req
->r_dentry_unless
);
1975 if (req
->r_old_dentry_drop
)
1976 releases
+= ceph_encode_dentry_release(&p
, req
->r_old_dentry
,
1977 mds
, req
->r_old_dentry_drop
, req
->r_old_dentry_unless
);
1978 if (req
->r_old_inode_drop
)
1979 releases
+= ceph_encode_inode_release(&p
,
1980 d_inode(req
->r_old_dentry
),
1981 mds
, req
->r_old_inode_drop
, req
->r_old_inode_unless
, 0);
1983 if (drop_cap_releases
) {
1985 p
= msg
->front
.iov_base
+ req
->r_request_release_offset
;
1988 head
->num_releases
= cpu_to_le16(releases
);
1992 struct ceph_timespec ts
;
1993 ceph_encode_timespec(&ts
, &req
->r_stamp
);
1994 ceph_encode_copy(&p
, &ts
, sizeof(ts
));
1998 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
1999 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
2001 if (req
->r_pagelist
) {
2002 struct ceph_pagelist
*pagelist
= req
->r_pagelist
;
2003 atomic_inc(&pagelist
->refcnt
);
2004 ceph_msg_data_add_pagelist(msg
, pagelist
);
2005 msg
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
2007 msg
->hdr
.data_len
= 0;
2010 msg
->hdr
.data_off
= cpu_to_le16(0);
2014 kfree((char *)path2
);
2017 kfree((char *)path1
);
2023 * called under mdsc->mutex if error, under no mutex if
2026 static void complete_request(struct ceph_mds_client
*mdsc
,
2027 struct ceph_mds_request
*req
)
2029 if (req
->r_callback
)
2030 req
->r_callback(mdsc
, req
);
2032 complete_all(&req
->r_completion
);
2036 * called under mdsc->mutex
2038 static int __prepare_send_request(struct ceph_mds_client
*mdsc
,
2039 struct ceph_mds_request
*req
,
2040 int mds
, bool drop_cap_releases
)
2042 struct ceph_mds_request_head
*rhead
;
2043 struct ceph_msg
*msg
;
2048 struct ceph_cap
*cap
=
2049 ceph_get_cap_for_mds(ceph_inode(req
->r_inode
), mds
);
2052 req
->r_sent_on_mseq
= cap
->mseq
;
2054 req
->r_sent_on_mseq
= -1;
2056 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req
,
2057 req
->r_tid
, ceph_mds_op_name(req
->r_op
), req
->r_attempts
);
2059 if (req
->r_got_unsafe
) {
2062 * Replay. Do not regenerate message (and rebuild
2063 * paths, etc.); just use the original message.
2064 * Rebuilding paths will break for renames because
2065 * d_move mangles the src name.
2067 msg
= req
->r_request
;
2068 rhead
= msg
->front
.iov_base
;
2070 flags
= le32_to_cpu(rhead
->flags
);
2071 flags
|= CEPH_MDS_FLAG_REPLAY
;
2072 rhead
->flags
= cpu_to_le32(flags
);
2074 if (req
->r_target_inode
)
2075 rhead
->ino
= cpu_to_le64(ceph_ino(req
->r_target_inode
));
2077 rhead
->num_retry
= req
->r_attempts
- 1;
2079 /* remove cap/dentry releases from message */
2080 rhead
->num_releases
= 0;
2083 p
= msg
->front
.iov_base
+ req
->r_request_release_offset
;
2085 struct ceph_timespec ts
;
2086 ceph_encode_timespec(&ts
, &req
->r_stamp
);
2087 ceph_encode_copy(&p
, &ts
, sizeof(ts
));
2090 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
2091 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
2095 if (req
->r_request
) {
2096 ceph_msg_put(req
->r_request
);
2097 req
->r_request
= NULL
;
2099 msg
= create_request_message(mdsc
, req
, mds
, drop_cap_releases
);
2101 req
->r_err
= PTR_ERR(msg
);
2102 return PTR_ERR(msg
);
2104 req
->r_request
= msg
;
2106 rhead
= msg
->front
.iov_base
;
2107 rhead
->oldest_client_tid
= cpu_to_le64(__get_oldest_tid(mdsc
));
2108 if (req
->r_got_unsafe
)
2109 flags
|= CEPH_MDS_FLAG_REPLAY
;
2110 if (req
->r_locked_dir
)
2111 flags
|= CEPH_MDS_FLAG_WANT_DENTRY
;
2112 rhead
->flags
= cpu_to_le32(flags
);
2113 rhead
->num_fwd
= req
->r_num_fwd
;
2114 rhead
->num_retry
= req
->r_attempts
- 1;
2117 dout(" r_locked_dir = %p\n", req
->r_locked_dir
);
2122 * send request, or put it on the appropriate wait list.
2124 static int __do_request(struct ceph_mds_client
*mdsc
,
2125 struct ceph_mds_request
*req
)
2127 struct ceph_mds_session
*session
= NULL
;
2131 if (req
->r_err
|| req
->r_got_result
) {
2133 __unregister_request(mdsc
, req
);
2137 if (req
->r_timeout
&&
2138 time_after_eq(jiffies
, req
->r_started
+ req
->r_timeout
)) {
2139 dout("do_request timed out\n");
2143 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
) {
2144 dout("do_request forced umount\n");
2149 put_request_session(req
);
2151 mds
= __choose_mds(mdsc
, req
);
2153 ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) < CEPH_MDS_STATE_ACTIVE
) {
2154 dout("do_request no mds or not active, waiting for map\n");
2155 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
2159 /* get, open session */
2160 session
= __ceph_lookup_mds_session(mdsc
, mds
);
2162 session
= register_session(mdsc
, mds
);
2163 if (IS_ERR(session
)) {
2164 err
= PTR_ERR(session
);
2168 req
->r_session
= get_session(session
);
2170 dout("do_request mds%d session %p state %s\n", mds
, session
,
2171 ceph_session_state_name(session
->s_state
));
2172 if (session
->s_state
!= CEPH_MDS_SESSION_OPEN
&&
2173 session
->s_state
!= CEPH_MDS_SESSION_HUNG
) {
2174 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
2175 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
2176 __open_session(mdsc
, session
);
2177 list_add(&req
->r_wait
, &session
->s_waiting
);
2182 req
->r_resend_mds
= -1; /* forget any previous mds hint */
2184 if (req
->r_request_started
== 0) /* note request start time */
2185 req
->r_request_started
= jiffies
;
2187 err
= __prepare_send_request(mdsc
, req
, mds
, false);
2189 ceph_msg_get(req
->r_request
);
2190 ceph_con_send(&session
->s_con
, req
->r_request
);
2194 ceph_put_mds_session(session
);
2197 dout("__do_request early error %d\n", err
);
2199 complete_request(mdsc
, req
);
2200 __unregister_request(mdsc
, req
);
2207 * called under mdsc->mutex
2209 static void __wake_requests(struct ceph_mds_client
*mdsc
,
2210 struct list_head
*head
)
2212 struct ceph_mds_request
*req
;
2213 LIST_HEAD(tmp_list
);
2215 list_splice_init(head
, &tmp_list
);
2217 while (!list_empty(&tmp_list
)) {
2218 req
= list_entry(tmp_list
.next
,
2219 struct ceph_mds_request
, r_wait
);
2220 list_del_init(&req
->r_wait
);
2221 dout(" wake request %p tid %llu\n", req
, req
->r_tid
);
2222 __do_request(mdsc
, req
);
2227 * Wake up threads with requests pending for @mds, so that they can
2228 * resubmit their requests to a possibly different mds.
2230 static void kick_requests(struct ceph_mds_client
*mdsc
, int mds
)
2232 struct ceph_mds_request
*req
;
2233 struct rb_node
*p
= rb_first(&mdsc
->request_tree
);
2235 dout("kick_requests mds%d\n", mds
);
2237 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2239 if (req
->r_got_unsafe
)
2241 if (req
->r_attempts
> 0)
2242 continue; /* only new requests */
2243 if (req
->r_session
&&
2244 req
->r_session
->s_mds
== mds
) {
2245 dout(" kicking tid %llu\n", req
->r_tid
);
2246 list_del_init(&req
->r_wait
);
2247 __do_request(mdsc
, req
);
2252 void ceph_mdsc_submit_request(struct ceph_mds_client
*mdsc
,
2253 struct ceph_mds_request
*req
)
2255 dout("submit_request on %p\n", req
);
2256 mutex_lock(&mdsc
->mutex
);
2257 __register_request(mdsc
, req
, NULL
);
2258 __do_request(mdsc
, req
);
2259 mutex_unlock(&mdsc
->mutex
);
2263 * Synchrously perform an mds request. Take care of all of the
2264 * session setup, forwarding, retry details.
2266 int ceph_mdsc_do_request(struct ceph_mds_client
*mdsc
,
2268 struct ceph_mds_request
*req
)
2272 dout("do_request on %p\n", req
);
2274 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2276 ceph_get_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
2277 if (req
->r_locked_dir
)
2278 ceph_get_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
2279 if (req
->r_old_dentry_dir
)
2280 ceph_get_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
2283 /* deny access to directories with pool_ns layouts */
2284 if (req
->r_inode
&& S_ISDIR(req
->r_inode
->i_mode
) &&
2285 ceph_inode(req
->r_inode
)->i_pool_ns_len
)
2287 if (req
->r_locked_dir
&&
2288 ceph_inode(req
->r_locked_dir
)->i_pool_ns_len
)
2292 mutex_lock(&mdsc
->mutex
);
2293 __register_request(mdsc
, req
, dir
);
2294 __do_request(mdsc
, req
);
2302 mutex_unlock(&mdsc
->mutex
);
2303 dout("do_request waiting\n");
2304 if (!req
->r_timeout
&& req
->r_wait_for_completion
) {
2305 err
= req
->r_wait_for_completion(mdsc
, req
);
2307 long timeleft
= wait_for_completion_killable_timeout(
2309 ceph_timeout_jiffies(req
->r_timeout
));
2313 err
= -EIO
; /* timed out */
2315 err
= timeleft
; /* killed */
2317 dout("do_request waited, got %d\n", err
);
2318 mutex_lock(&mdsc
->mutex
);
2320 /* only abort if we didn't race with a real reply */
2321 if (req
->r_got_result
) {
2322 err
= le32_to_cpu(req
->r_reply_info
.head
->result
);
2323 } else if (err
< 0) {
2324 dout("aborted request %lld with %d\n", req
->r_tid
, err
);
2327 * ensure we aren't running concurrently with
2328 * ceph_fill_trace or ceph_readdir_prepopulate, which
2329 * rely on locks (dir mutex) held by our caller.
2331 mutex_lock(&req
->r_fill_mutex
);
2333 req
->r_aborted
= true;
2334 mutex_unlock(&req
->r_fill_mutex
);
2336 if (req
->r_locked_dir
&&
2337 (req
->r_op
& CEPH_MDS_OP_WRITE
))
2338 ceph_invalidate_dir_request(req
);
2344 mutex_unlock(&mdsc
->mutex
);
2345 dout("do_request %p done, result %d\n", req
, err
);
2350 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2351 * namespace request.
2353 void ceph_invalidate_dir_request(struct ceph_mds_request
*req
)
2355 struct inode
*inode
= req
->r_locked_dir
;
2357 dout("invalidate_dir_request %p (complete, lease(s))\n", inode
);
2359 ceph_dir_clear_complete(inode
);
2361 ceph_invalidate_dentry_lease(req
->r_dentry
);
2362 if (req
->r_old_dentry
)
2363 ceph_invalidate_dentry_lease(req
->r_old_dentry
);
2369 * We take the session mutex and parse and process the reply immediately.
2370 * This preserves the logical ordering of replies, capabilities, etc., sent
2371 * by the MDS as they are applied to our local cache.
2373 static void handle_reply(struct ceph_mds_session
*session
, struct ceph_msg
*msg
)
2375 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2376 struct ceph_mds_request
*req
;
2377 struct ceph_mds_reply_head
*head
= msg
->front
.iov_base
;
2378 struct ceph_mds_reply_info_parsed
*rinfo
; /* parsed reply info */
2379 struct ceph_snap_realm
*realm
;
2382 int mds
= session
->s_mds
;
2384 if (msg
->front
.iov_len
< sizeof(*head
)) {
2385 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2390 /* get request, session */
2391 tid
= le64_to_cpu(msg
->hdr
.tid
);
2392 mutex_lock(&mdsc
->mutex
);
2393 req
= lookup_get_request(mdsc
, tid
);
2395 dout("handle_reply on unknown tid %llu\n", tid
);
2396 mutex_unlock(&mdsc
->mutex
);
2399 dout("handle_reply %p\n", req
);
2401 /* correct session? */
2402 if (req
->r_session
!= session
) {
2403 pr_err("mdsc_handle_reply got %llu on session mds%d"
2404 " not mds%d\n", tid
, session
->s_mds
,
2405 req
->r_session
? req
->r_session
->s_mds
: -1);
2406 mutex_unlock(&mdsc
->mutex
);
2411 if ((req
->r_got_unsafe
&& !head
->safe
) ||
2412 (req
->r_got_safe
&& head
->safe
)) {
2413 pr_warn("got a dup %s reply on %llu from mds%d\n",
2414 head
->safe
? "safe" : "unsafe", tid
, mds
);
2415 mutex_unlock(&mdsc
->mutex
);
2418 if (req
->r_got_safe
) {
2419 pr_warn("got unsafe after safe on %llu from mds%d\n",
2421 mutex_unlock(&mdsc
->mutex
);
2425 result
= le32_to_cpu(head
->result
);
2429 * if we're not talking to the authority, send to them
2430 * if the authority has changed while we weren't looking,
2431 * send to new authority
2432 * Otherwise we just have to return an ESTALE
2434 if (result
== -ESTALE
) {
2435 dout("got ESTALE on request %llu", req
->r_tid
);
2436 req
->r_resend_mds
= -1;
2437 if (req
->r_direct_mode
!= USE_AUTH_MDS
) {
2438 dout("not using auth, setting for that now");
2439 req
->r_direct_mode
= USE_AUTH_MDS
;
2440 __do_request(mdsc
, req
);
2441 mutex_unlock(&mdsc
->mutex
);
2444 int mds
= __choose_mds(mdsc
, req
);
2445 if (mds
>= 0 && mds
!= req
->r_session
->s_mds
) {
2446 dout("but auth changed, so resending");
2447 __do_request(mdsc
, req
);
2448 mutex_unlock(&mdsc
->mutex
);
2452 dout("have to return ESTALE on request %llu", req
->r_tid
);
2457 req
->r_got_safe
= true;
2458 __unregister_request(mdsc
, req
);
2460 if (req
->r_got_unsafe
) {
2462 * We already handled the unsafe response, now do the
2463 * cleanup. No need to examine the response; the MDS
2464 * doesn't include any result info in the safe
2465 * response. And even if it did, there is nothing
2466 * useful we could do with a revised return value.
2468 dout("got safe reply %llu, mds%d\n", tid
, mds
);
2469 list_del_init(&req
->r_unsafe_item
);
2471 /* last unsafe request during umount? */
2472 if (mdsc
->stopping
&& !__get_oldest_req(mdsc
))
2473 complete_all(&mdsc
->safe_umount_waiters
);
2474 mutex_unlock(&mdsc
->mutex
);
2478 req
->r_got_unsafe
= true;
2479 list_add_tail(&req
->r_unsafe_item
, &req
->r_session
->s_unsafe
);
2480 if (req
->r_unsafe_dir
) {
2481 struct ceph_inode_info
*ci
=
2482 ceph_inode(req
->r_unsafe_dir
);
2483 spin_lock(&ci
->i_unsafe_lock
);
2484 list_add_tail(&req
->r_unsafe_dir_item
,
2485 &ci
->i_unsafe_dirops
);
2486 spin_unlock(&ci
->i_unsafe_lock
);
2490 dout("handle_reply tid %lld result %d\n", tid
, result
);
2491 rinfo
= &req
->r_reply_info
;
2492 err
= parse_reply_info(msg
, rinfo
, session
->s_con
.peer_features
);
2493 mutex_unlock(&mdsc
->mutex
);
2495 mutex_lock(&session
->s_mutex
);
2497 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds
, tid
);
2504 if (rinfo
->snapblob_len
) {
2505 down_write(&mdsc
->snap_rwsem
);
2506 ceph_update_snap_trace(mdsc
, rinfo
->snapblob
,
2507 rinfo
->snapblob
+ rinfo
->snapblob_len
,
2508 le32_to_cpu(head
->op
) == CEPH_MDS_OP_RMSNAP
,
2510 downgrade_write(&mdsc
->snap_rwsem
);
2512 down_read(&mdsc
->snap_rwsem
);
2515 /* insert trace into our cache */
2516 mutex_lock(&req
->r_fill_mutex
);
2517 current
->journal_info
= req
;
2518 err
= ceph_fill_trace(mdsc
->fsc
->sb
, req
, req
->r_session
);
2520 if (result
== 0 && (req
->r_op
== CEPH_MDS_OP_READDIR
||
2521 req
->r_op
== CEPH_MDS_OP_LSSNAP
))
2522 ceph_readdir_prepopulate(req
, req
->r_session
);
2523 ceph_unreserve_caps(mdsc
, &req
->r_caps_reservation
);
2525 current
->journal_info
= NULL
;
2526 mutex_unlock(&req
->r_fill_mutex
);
2528 up_read(&mdsc
->snap_rwsem
);
2530 ceph_put_snap_realm(mdsc
, realm
);
2532 if (err
== 0 && req
->r_got_unsafe
&& req
->r_target_inode
) {
2533 struct ceph_inode_info
*ci
= ceph_inode(req
->r_target_inode
);
2534 spin_lock(&ci
->i_unsafe_lock
);
2535 list_add_tail(&req
->r_unsafe_target_item
, &ci
->i_unsafe_iops
);
2536 spin_unlock(&ci
->i_unsafe_lock
);
2539 mutex_lock(&mdsc
->mutex
);
2540 if (!req
->r_aborted
) {
2544 req
->r_reply
= ceph_msg_get(msg
);
2545 req
->r_got_result
= true;
2548 dout("reply arrived after request %lld was aborted\n", tid
);
2550 mutex_unlock(&mdsc
->mutex
);
2552 mutex_unlock(&session
->s_mutex
);
2554 /* kick calling process */
2555 complete_request(mdsc
, req
);
2557 ceph_mdsc_put_request(req
);
2564 * handle mds notification that our request has been forwarded.
2566 static void handle_forward(struct ceph_mds_client
*mdsc
,
2567 struct ceph_mds_session
*session
,
2568 struct ceph_msg
*msg
)
2570 struct ceph_mds_request
*req
;
2571 u64 tid
= le64_to_cpu(msg
->hdr
.tid
);
2575 void *p
= msg
->front
.iov_base
;
2576 void *end
= p
+ msg
->front
.iov_len
;
2578 ceph_decode_need(&p
, end
, 2*sizeof(u32
), bad
);
2579 next_mds
= ceph_decode_32(&p
);
2580 fwd_seq
= ceph_decode_32(&p
);
2582 mutex_lock(&mdsc
->mutex
);
2583 req
= lookup_get_request(mdsc
, tid
);
2585 dout("forward tid %llu to mds%d - req dne\n", tid
, next_mds
);
2586 goto out
; /* dup reply? */
2589 if (req
->r_aborted
) {
2590 dout("forward tid %llu aborted, unregistering\n", tid
);
2591 __unregister_request(mdsc
, req
);
2592 } else if (fwd_seq
<= req
->r_num_fwd
) {
2593 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2594 tid
, next_mds
, req
->r_num_fwd
, fwd_seq
);
2596 /* resend. forward race not possible; mds would drop */
2597 dout("forward tid %llu to mds%d (we resend)\n", tid
, next_mds
);
2599 BUG_ON(req
->r_got_result
);
2600 req
->r_attempts
= 0;
2601 req
->r_num_fwd
= fwd_seq
;
2602 req
->r_resend_mds
= next_mds
;
2603 put_request_session(req
);
2604 __do_request(mdsc
, req
);
2606 ceph_mdsc_put_request(req
);
2608 mutex_unlock(&mdsc
->mutex
);
2612 pr_err("mdsc_handle_forward decode error err=%d\n", err
);
2616 * handle a mds session control message
2618 static void handle_session(struct ceph_mds_session
*session
,
2619 struct ceph_msg
*msg
)
2621 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2624 int mds
= session
->s_mds
;
2625 struct ceph_mds_session_head
*h
= msg
->front
.iov_base
;
2629 if (msg
->front
.iov_len
!= sizeof(*h
))
2631 op
= le32_to_cpu(h
->op
);
2632 seq
= le64_to_cpu(h
->seq
);
2634 mutex_lock(&mdsc
->mutex
);
2635 if (op
== CEPH_SESSION_CLOSE
)
2636 __unregister_session(mdsc
, session
);
2637 /* FIXME: this ttl calculation is generous */
2638 session
->s_ttl
= jiffies
+ HZ
*mdsc
->mdsmap
->m_session_autoclose
;
2639 mutex_unlock(&mdsc
->mutex
);
2641 mutex_lock(&session
->s_mutex
);
2643 dout("handle_session mds%d %s %p state %s seq %llu\n",
2644 mds
, ceph_session_op_name(op
), session
,
2645 ceph_session_state_name(session
->s_state
), seq
);
2647 if (session
->s_state
== CEPH_MDS_SESSION_HUNG
) {
2648 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2649 pr_info("mds%d came back\n", session
->s_mds
);
2653 case CEPH_SESSION_OPEN
:
2654 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2655 pr_info("mds%d reconnect success\n", session
->s_mds
);
2656 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2657 renewed_caps(mdsc
, session
, 0);
2660 __close_session(mdsc
, session
);
2663 case CEPH_SESSION_RENEWCAPS
:
2664 if (session
->s_renew_seq
== seq
)
2665 renewed_caps(mdsc
, session
, 1);
2668 case CEPH_SESSION_CLOSE
:
2669 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2670 pr_info("mds%d reconnect denied\n", session
->s_mds
);
2671 cleanup_session_requests(mdsc
, session
);
2672 remove_session_caps(session
);
2673 wake
= 2; /* for good measure */
2674 wake_up_all(&mdsc
->session_close_wq
);
2677 case CEPH_SESSION_STALE
:
2678 pr_info("mds%d caps went stale, renewing\n",
2680 spin_lock(&session
->s_gen_ttl_lock
);
2681 session
->s_cap_gen
++;
2682 session
->s_cap_ttl
= jiffies
- 1;
2683 spin_unlock(&session
->s_gen_ttl_lock
);
2684 send_renew_caps(mdsc
, session
);
2687 case CEPH_SESSION_RECALL_STATE
:
2688 trim_caps(mdsc
, session
, le32_to_cpu(h
->max_caps
));
2691 case CEPH_SESSION_FLUSHMSG
:
2692 send_flushmsg_ack(mdsc
, session
, seq
);
2695 case CEPH_SESSION_FORCE_RO
:
2696 dout("force_session_readonly %p\n", session
);
2697 spin_lock(&session
->s_cap_lock
);
2698 session
->s_readonly
= true;
2699 spin_unlock(&session
->s_cap_lock
);
2700 wake_up_session_caps(session
, 0);
2704 pr_err("mdsc_handle_session bad op %d mds%d\n", op
, mds
);
2708 mutex_unlock(&session
->s_mutex
);
2710 mutex_lock(&mdsc
->mutex
);
2711 __wake_requests(mdsc
, &session
->s_waiting
);
2713 kick_requests(mdsc
, mds
);
2714 mutex_unlock(&mdsc
->mutex
);
2719 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds
,
2720 (int)msg
->front
.iov_len
);
2727 * called under session->mutex.
2729 static void replay_unsafe_requests(struct ceph_mds_client
*mdsc
,
2730 struct ceph_mds_session
*session
)
2732 struct ceph_mds_request
*req
, *nreq
;
2736 dout("replay_unsafe_requests mds%d\n", session
->s_mds
);
2738 mutex_lock(&mdsc
->mutex
);
2739 list_for_each_entry_safe(req
, nreq
, &session
->s_unsafe
, r_unsafe_item
) {
2740 err
= __prepare_send_request(mdsc
, req
, session
->s_mds
, true);
2742 ceph_msg_get(req
->r_request
);
2743 ceph_con_send(&session
->s_con
, req
->r_request
);
2748 * also re-send old requests when MDS enters reconnect stage. So that MDS
2749 * can process completed request in clientreplay stage.
2751 p
= rb_first(&mdsc
->request_tree
);
2753 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2755 if (req
->r_got_unsafe
)
2757 if (req
->r_attempts
== 0)
2758 continue; /* only old requests */
2759 if (req
->r_session
&&
2760 req
->r_session
->s_mds
== session
->s_mds
) {
2761 err
= __prepare_send_request(mdsc
, req
,
2762 session
->s_mds
, true);
2764 ceph_msg_get(req
->r_request
);
2765 ceph_con_send(&session
->s_con
, req
->r_request
);
2769 mutex_unlock(&mdsc
->mutex
);
2773 * Encode information about a cap for a reconnect with the MDS.
2775 static int encode_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
2779 struct ceph_mds_cap_reconnect v2
;
2780 struct ceph_mds_cap_reconnect_v1 v1
;
2783 struct ceph_inode_info
*ci
;
2784 struct ceph_reconnect_state
*recon_state
= arg
;
2785 struct ceph_pagelist
*pagelist
= recon_state
->pagelist
;
2789 struct dentry
*dentry
;
2793 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2794 inode
, ceph_vinop(inode
), cap
, cap
->cap_id
,
2795 ceph_cap_string(cap
->issued
));
2796 err
= ceph_pagelist_encode_64(pagelist
, ceph_ino(inode
));
2800 dentry
= d_find_alias(inode
);
2802 path
= ceph_mdsc_build_path(dentry
, &pathlen
, &pathbase
, 0);
2804 err
= PTR_ERR(path
);
2811 err
= ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
2815 spin_lock(&ci
->i_ceph_lock
);
2816 cap
->seq
= 0; /* reset cap seq */
2817 cap
->issue_seq
= 0; /* and issue_seq */
2818 cap
->mseq
= 0; /* and migrate_seq */
2819 cap
->cap_gen
= cap
->session
->s_cap_gen
;
2821 if (recon_state
->flock
) {
2822 rec
.v2
.cap_id
= cpu_to_le64(cap
->cap_id
);
2823 rec
.v2
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2824 rec
.v2
.issued
= cpu_to_le32(cap
->issued
);
2825 rec
.v2
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2826 rec
.v2
.pathbase
= cpu_to_le64(pathbase
);
2827 rec
.v2
.flock_len
= 0;
2828 reclen
= sizeof(rec
.v2
);
2830 rec
.v1
.cap_id
= cpu_to_le64(cap
->cap_id
);
2831 rec
.v1
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2832 rec
.v1
.issued
= cpu_to_le32(cap
->issued
);
2833 rec
.v1
.size
= cpu_to_le64(inode
->i_size
);
2834 ceph_encode_timespec(&rec
.v1
.mtime
, &inode
->i_mtime
);
2835 ceph_encode_timespec(&rec
.v1
.atime
, &inode
->i_atime
);
2836 rec
.v1
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2837 rec
.v1
.pathbase
= cpu_to_le64(pathbase
);
2838 reclen
= sizeof(rec
.v1
);
2840 spin_unlock(&ci
->i_ceph_lock
);
2842 if (recon_state
->flock
) {
2843 int num_fcntl_locks
, num_flock_locks
;
2844 struct ceph_filelock
*flocks
;
2847 ceph_count_locks(inode
, &num_fcntl_locks
, &num_flock_locks
);
2848 flocks
= kmalloc((num_fcntl_locks
+num_flock_locks
) *
2849 sizeof(struct ceph_filelock
), GFP_NOFS
);
2854 err
= ceph_encode_locks_to_buffer(inode
, flocks
,
2864 * number of encoded locks is stable, so copy to pagelist
2866 rec
.v2
.flock_len
= cpu_to_le32(2*sizeof(u32
) +
2867 (num_fcntl_locks
+num_flock_locks
) *
2868 sizeof(struct ceph_filelock
));
2869 err
= ceph_pagelist_append(pagelist
, &rec
, reclen
);
2871 err
= ceph_locks_to_pagelist(flocks
, pagelist
,
2876 err
= ceph_pagelist_append(pagelist
, &rec
, reclen
);
2879 recon_state
->nr_caps
++;
2889 * If an MDS fails and recovers, clients need to reconnect in order to
2890 * reestablish shared state. This includes all caps issued through
2891 * this session _and_ the snap_realm hierarchy. Because it's not
2892 * clear which snap realms the mds cares about, we send everything we
2893 * know about.. that ensures we'll then get any new info the
2894 * recovering MDS might have.
2896 * This is a relatively heavyweight operation, but it's rare.
2898 * called with mdsc->mutex held.
2900 static void send_mds_reconnect(struct ceph_mds_client
*mdsc
,
2901 struct ceph_mds_session
*session
)
2903 struct ceph_msg
*reply
;
2905 int mds
= session
->s_mds
;
2908 struct ceph_pagelist
*pagelist
;
2909 struct ceph_reconnect_state recon_state
;
2911 pr_info("mds%d reconnect start\n", mds
);
2913 pagelist
= kmalloc(sizeof(*pagelist
), GFP_NOFS
);
2915 goto fail_nopagelist
;
2916 ceph_pagelist_init(pagelist
);
2918 reply
= ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT
, 0, GFP_NOFS
, false);
2922 mutex_lock(&session
->s_mutex
);
2923 session
->s_state
= CEPH_MDS_SESSION_RECONNECTING
;
2926 dout("session %p state %s\n", session
,
2927 ceph_session_state_name(session
->s_state
));
2929 spin_lock(&session
->s_gen_ttl_lock
);
2930 session
->s_cap_gen
++;
2931 spin_unlock(&session
->s_gen_ttl_lock
);
2933 spin_lock(&session
->s_cap_lock
);
2934 /* don't know if session is readonly */
2935 session
->s_readonly
= 0;
2937 * notify __ceph_remove_cap() that we are composing cap reconnect.
2938 * If a cap get released before being added to the cap reconnect,
2939 * __ceph_remove_cap() should skip queuing cap release.
2941 session
->s_cap_reconnect
= 1;
2942 /* drop old cap expires; we're about to reestablish that state */
2943 cleanup_cap_releases(mdsc
, session
);
2945 /* trim unused caps to reduce MDS's cache rejoin time */
2946 if (mdsc
->fsc
->sb
->s_root
)
2947 shrink_dcache_parent(mdsc
->fsc
->sb
->s_root
);
2949 ceph_con_close(&session
->s_con
);
2950 ceph_con_open(&session
->s_con
,
2951 CEPH_ENTITY_TYPE_MDS
, mds
,
2952 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
2954 /* replay unsafe requests */
2955 replay_unsafe_requests(mdsc
, session
);
2957 down_read(&mdsc
->snap_rwsem
);
2959 /* traverse this session's caps */
2960 s_nr_caps
= session
->s_nr_caps
;
2961 err
= ceph_pagelist_encode_32(pagelist
, s_nr_caps
);
2965 recon_state
.nr_caps
= 0;
2966 recon_state
.pagelist
= pagelist
;
2967 recon_state
.flock
= session
->s_con
.peer_features
& CEPH_FEATURE_FLOCK
;
2968 err
= iterate_session_caps(session
, encode_caps_cb
, &recon_state
);
2972 spin_lock(&session
->s_cap_lock
);
2973 session
->s_cap_reconnect
= 0;
2974 spin_unlock(&session
->s_cap_lock
);
2977 * snaprealms. we provide mds with the ino, seq (version), and
2978 * parent for all of our realms. If the mds has any newer info,
2981 for (p
= rb_first(&mdsc
->snap_realms
); p
; p
= rb_next(p
)) {
2982 struct ceph_snap_realm
*realm
=
2983 rb_entry(p
, struct ceph_snap_realm
, node
);
2984 struct ceph_mds_snaprealm_reconnect sr_rec
;
2986 dout(" adding snap realm %llx seq %lld parent %llx\n",
2987 realm
->ino
, realm
->seq
, realm
->parent_ino
);
2988 sr_rec
.ino
= cpu_to_le64(realm
->ino
);
2989 sr_rec
.seq
= cpu_to_le64(realm
->seq
);
2990 sr_rec
.parent
= cpu_to_le64(realm
->parent_ino
);
2991 err
= ceph_pagelist_append(pagelist
, &sr_rec
, sizeof(sr_rec
));
2996 if (recon_state
.flock
)
2997 reply
->hdr
.version
= cpu_to_le16(2);
2999 /* raced with cap release? */
3000 if (s_nr_caps
!= recon_state
.nr_caps
) {
3001 struct page
*page
= list_first_entry(&pagelist
->head
,
3003 __le32
*addr
= kmap_atomic(page
);
3004 *addr
= cpu_to_le32(recon_state
.nr_caps
);
3005 kunmap_atomic(addr
);
3008 reply
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
3009 ceph_msg_data_add_pagelist(reply
, pagelist
);
3011 ceph_early_kick_flushing_caps(mdsc
, session
);
3013 ceph_con_send(&session
->s_con
, reply
);
3015 mutex_unlock(&session
->s_mutex
);
3017 mutex_lock(&mdsc
->mutex
);
3018 __wake_requests(mdsc
, &session
->s_waiting
);
3019 mutex_unlock(&mdsc
->mutex
);
3021 up_read(&mdsc
->snap_rwsem
);
3025 ceph_msg_put(reply
);
3026 up_read(&mdsc
->snap_rwsem
);
3027 mutex_unlock(&session
->s_mutex
);
3029 ceph_pagelist_release(pagelist
);
3031 pr_err("error %d preparing reconnect for mds%d\n", err
, mds
);
3037 * compare old and new mdsmaps, kicking requests
3038 * and closing out old connections as necessary
3040 * called under mdsc->mutex.
3042 static void check_new_map(struct ceph_mds_client
*mdsc
,
3043 struct ceph_mdsmap
*newmap
,
3044 struct ceph_mdsmap
*oldmap
)
3047 int oldstate
, newstate
;
3048 struct ceph_mds_session
*s
;
3050 dout("check_new_map new %u old %u\n",
3051 newmap
->m_epoch
, oldmap
->m_epoch
);
3053 for (i
= 0; i
< oldmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
3054 if (mdsc
->sessions
[i
] == NULL
)
3056 s
= mdsc
->sessions
[i
];
3057 oldstate
= ceph_mdsmap_get_state(oldmap
, i
);
3058 newstate
= ceph_mdsmap_get_state(newmap
, i
);
3060 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3061 i
, ceph_mds_state_name(oldstate
),
3062 ceph_mdsmap_is_laggy(oldmap
, i
) ? " (laggy)" : "",
3063 ceph_mds_state_name(newstate
),
3064 ceph_mdsmap_is_laggy(newmap
, i
) ? " (laggy)" : "",
3065 ceph_session_state_name(s
->s_state
));
3067 if (i
>= newmap
->m_max_mds
||
3068 memcmp(ceph_mdsmap_get_addr(oldmap
, i
),
3069 ceph_mdsmap_get_addr(newmap
, i
),
3070 sizeof(struct ceph_entity_addr
))) {
3071 if (s
->s_state
== CEPH_MDS_SESSION_OPENING
) {
3072 /* the session never opened, just close it
3074 __wake_requests(mdsc
, &s
->s_waiting
);
3075 __unregister_session(mdsc
, s
);
3078 mutex_unlock(&mdsc
->mutex
);
3079 mutex_lock(&s
->s_mutex
);
3080 mutex_lock(&mdsc
->mutex
);
3081 ceph_con_close(&s
->s_con
);
3082 mutex_unlock(&s
->s_mutex
);
3083 s
->s_state
= CEPH_MDS_SESSION_RESTARTING
;
3085 } else if (oldstate
== newstate
) {
3086 continue; /* nothing new with this mds */
3092 if (s
->s_state
== CEPH_MDS_SESSION_RESTARTING
&&
3093 newstate
>= CEPH_MDS_STATE_RECONNECT
) {
3094 mutex_unlock(&mdsc
->mutex
);
3095 send_mds_reconnect(mdsc
, s
);
3096 mutex_lock(&mdsc
->mutex
);
3100 * kick request on any mds that has gone active.
3102 if (oldstate
< CEPH_MDS_STATE_ACTIVE
&&
3103 newstate
>= CEPH_MDS_STATE_ACTIVE
) {
3104 if (oldstate
!= CEPH_MDS_STATE_CREATING
&&
3105 oldstate
!= CEPH_MDS_STATE_STARTING
)
3106 pr_info("mds%d recovery completed\n", s
->s_mds
);
3107 kick_requests(mdsc
, i
);
3108 ceph_kick_flushing_caps(mdsc
, s
);
3109 wake_up_session_caps(s
, 1);
3113 for (i
= 0; i
< newmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
3114 s
= mdsc
->sessions
[i
];
3117 if (!ceph_mdsmap_is_laggy(newmap
, i
))
3119 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3120 s
->s_state
== CEPH_MDS_SESSION_HUNG
||
3121 s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3122 dout(" connecting to export targets of laggy mds%d\n",
3124 __open_export_target_sessions(mdsc
, s
);
3136 * caller must hold session s_mutex, dentry->d_lock
3138 void __ceph_mdsc_drop_dentry_lease(struct dentry
*dentry
)
3140 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
3142 ceph_put_mds_session(di
->lease_session
);
3143 di
->lease_session
= NULL
;
3146 static void handle_lease(struct ceph_mds_client
*mdsc
,
3147 struct ceph_mds_session
*session
,
3148 struct ceph_msg
*msg
)
3150 struct super_block
*sb
= mdsc
->fsc
->sb
;
3151 struct inode
*inode
;
3152 struct dentry
*parent
, *dentry
;
3153 struct ceph_dentry_info
*di
;
3154 int mds
= session
->s_mds
;
3155 struct ceph_mds_lease
*h
= msg
->front
.iov_base
;
3157 struct ceph_vino vino
;
3161 dout("handle_lease from mds%d\n", mds
);
3164 if (msg
->front
.iov_len
< sizeof(*h
) + sizeof(u32
))
3166 vino
.ino
= le64_to_cpu(h
->ino
);
3167 vino
.snap
= CEPH_NOSNAP
;
3168 seq
= le32_to_cpu(h
->seq
);
3169 dname
.name
= (void *)h
+ sizeof(*h
) + sizeof(u32
);
3170 dname
.len
= msg
->front
.iov_len
- sizeof(*h
) - sizeof(u32
);
3171 if (dname
.len
!= get_unaligned_le32(h
+1))
3175 inode
= ceph_find_inode(sb
, vino
);
3176 dout("handle_lease %s, ino %llx %p %.*s\n",
3177 ceph_lease_op_name(h
->action
), vino
.ino
, inode
,
3178 dname
.len
, dname
.name
);
3180 mutex_lock(&session
->s_mutex
);
3183 if (inode
== NULL
) {
3184 dout("handle_lease no inode %llx\n", vino
.ino
);
3189 parent
= d_find_alias(inode
);
3191 dout("no parent dentry on inode %p\n", inode
);
3193 goto release
; /* hrm... */
3195 dname
.hash
= full_name_hash(dname
.name
, dname
.len
);
3196 dentry
= d_lookup(parent
, &dname
);
3201 spin_lock(&dentry
->d_lock
);
3202 di
= ceph_dentry(dentry
);
3203 switch (h
->action
) {
3204 case CEPH_MDS_LEASE_REVOKE
:
3205 if (di
->lease_session
== session
) {
3206 if (ceph_seq_cmp(di
->lease_seq
, seq
) > 0)
3207 h
->seq
= cpu_to_le32(di
->lease_seq
);
3208 __ceph_mdsc_drop_dentry_lease(dentry
);
3213 case CEPH_MDS_LEASE_RENEW
:
3214 if (di
->lease_session
== session
&&
3215 di
->lease_gen
== session
->s_cap_gen
&&
3216 di
->lease_renew_from
&&
3217 di
->lease_renew_after
== 0) {
3218 unsigned long duration
=
3219 msecs_to_jiffies(le32_to_cpu(h
->duration_ms
));
3221 di
->lease_seq
= seq
;
3222 dentry
->d_time
= di
->lease_renew_from
+ duration
;
3223 di
->lease_renew_after
= di
->lease_renew_from
+
3225 di
->lease_renew_from
= 0;
3229 spin_unlock(&dentry
->d_lock
);
3236 /* let's just reuse the same message */
3237 h
->action
= CEPH_MDS_LEASE_REVOKE_ACK
;
3239 ceph_con_send(&session
->s_con
, msg
);
3243 mutex_unlock(&session
->s_mutex
);
3247 pr_err("corrupt lease message\n");
3251 void ceph_mdsc_lease_send_msg(struct ceph_mds_session
*session
,
3252 struct inode
*inode
,
3253 struct dentry
*dentry
, char action
,
3256 struct ceph_msg
*msg
;
3257 struct ceph_mds_lease
*lease
;
3258 int len
= sizeof(*lease
) + sizeof(u32
);
3261 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3262 inode
, dentry
, ceph_lease_op_name(action
), session
->s_mds
);
3263 dnamelen
= dentry
->d_name
.len
;
3266 msg
= ceph_msg_new(CEPH_MSG_CLIENT_LEASE
, len
, GFP_NOFS
, false);
3269 lease
= msg
->front
.iov_base
;
3270 lease
->action
= action
;
3271 lease
->ino
= cpu_to_le64(ceph_vino(inode
).ino
);
3272 lease
->first
= lease
->last
= cpu_to_le64(ceph_vino(inode
).snap
);
3273 lease
->seq
= cpu_to_le32(seq
);
3274 put_unaligned_le32(dnamelen
, lease
+ 1);
3275 memcpy((void *)(lease
+ 1) + 4, dentry
->d_name
.name
, dnamelen
);
3278 * if this is a preemptive lease RELEASE, no need to
3279 * flush request stream, since the actual request will
3282 msg
->more_to_follow
= (action
== CEPH_MDS_LEASE_RELEASE
);
3284 ceph_con_send(&session
->s_con
, msg
);
3288 * Preemptively release a lease we expect to invalidate anyway.
3289 * Pass @inode always, @dentry is optional.
3291 void ceph_mdsc_lease_release(struct ceph_mds_client
*mdsc
, struct inode
*inode
,
3292 struct dentry
*dentry
)
3294 struct ceph_dentry_info
*di
;
3295 struct ceph_mds_session
*session
;
3298 BUG_ON(inode
== NULL
);
3299 BUG_ON(dentry
== NULL
);
3301 /* is dentry lease valid? */
3302 spin_lock(&dentry
->d_lock
);
3303 di
= ceph_dentry(dentry
);
3304 if (!di
|| !di
->lease_session
||
3305 di
->lease_session
->s_mds
< 0 ||
3306 di
->lease_gen
!= di
->lease_session
->s_cap_gen
||
3307 !time_before(jiffies
, dentry
->d_time
)) {
3308 dout("lease_release inode %p dentry %p -- "
3311 spin_unlock(&dentry
->d_lock
);
3315 /* we do have a lease on this dentry; note mds and seq */
3316 session
= ceph_get_mds_session(di
->lease_session
);
3317 seq
= di
->lease_seq
;
3318 __ceph_mdsc_drop_dentry_lease(dentry
);
3319 spin_unlock(&dentry
->d_lock
);
3321 dout("lease_release inode %p dentry %p to mds%d\n",
3322 inode
, dentry
, session
->s_mds
);
3323 ceph_mdsc_lease_send_msg(session
, inode
, dentry
,
3324 CEPH_MDS_LEASE_RELEASE
, seq
);
3325 ceph_put_mds_session(session
);
3329 * drop all leases (and dentry refs) in preparation for umount
3331 static void drop_leases(struct ceph_mds_client
*mdsc
)
3335 dout("drop_leases\n");
3336 mutex_lock(&mdsc
->mutex
);
3337 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3338 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3341 mutex_unlock(&mdsc
->mutex
);
3342 mutex_lock(&s
->s_mutex
);
3343 mutex_unlock(&s
->s_mutex
);
3344 ceph_put_mds_session(s
);
3345 mutex_lock(&mdsc
->mutex
);
3347 mutex_unlock(&mdsc
->mutex
);
3353 * delayed work -- periodically trim expired leases, renew caps with mds
3355 static void schedule_delayed(struct ceph_mds_client
*mdsc
)
3358 unsigned hz
= round_jiffies_relative(HZ
* delay
);
3359 schedule_delayed_work(&mdsc
->delayed_work
, hz
);
3362 static void delayed_work(struct work_struct
*work
)
3365 struct ceph_mds_client
*mdsc
=
3366 container_of(work
, struct ceph_mds_client
, delayed_work
.work
);
3370 dout("mdsc delayed_work\n");
3371 ceph_check_delayed_caps(mdsc
);
3373 mutex_lock(&mdsc
->mutex
);
3374 renew_interval
= mdsc
->mdsmap
->m_session_timeout
>> 2;
3375 renew_caps
= time_after_eq(jiffies
, HZ
*renew_interval
+
3376 mdsc
->last_renew_caps
);
3378 mdsc
->last_renew_caps
= jiffies
;
3380 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3381 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3384 if (s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3385 dout("resending session close request for mds%d\n",
3387 request_close_session(mdsc
, s
);
3388 ceph_put_mds_session(s
);
3391 if (s
->s_ttl
&& time_after(jiffies
, s
->s_ttl
)) {
3392 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
) {
3393 s
->s_state
= CEPH_MDS_SESSION_HUNG
;
3394 pr_info("mds%d hung\n", s
->s_mds
);
3397 if (s
->s_state
< CEPH_MDS_SESSION_OPEN
) {
3398 /* this mds is failed or recovering, just wait */
3399 ceph_put_mds_session(s
);
3402 mutex_unlock(&mdsc
->mutex
);
3404 mutex_lock(&s
->s_mutex
);
3406 send_renew_caps(mdsc
, s
);
3408 ceph_con_keepalive(&s
->s_con
);
3409 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3410 s
->s_state
== CEPH_MDS_SESSION_HUNG
)
3411 ceph_send_cap_releases(mdsc
, s
);
3412 mutex_unlock(&s
->s_mutex
);
3413 ceph_put_mds_session(s
);
3415 mutex_lock(&mdsc
->mutex
);
3417 mutex_unlock(&mdsc
->mutex
);
3419 schedule_delayed(mdsc
);
3422 int ceph_mdsc_init(struct ceph_fs_client
*fsc
)
3425 struct ceph_mds_client
*mdsc
;
3427 mdsc
= kzalloc(sizeof(struct ceph_mds_client
), GFP_NOFS
);
3432 mutex_init(&mdsc
->mutex
);
3433 mdsc
->mdsmap
= kzalloc(sizeof(*mdsc
->mdsmap
), GFP_NOFS
);
3434 if (mdsc
->mdsmap
== NULL
) {
3439 init_completion(&mdsc
->safe_umount_waiters
);
3440 init_waitqueue_head(&mdsc
->session_close_wq
);
3441 INIT_LIST_HEAD(&mdsc
->waiting_for_map
);
3442 mdsc
->sessions
= NULL
;
3443 atomic_set(&mdsc
->num_sessions
, 0);
3444 mdsc
->max_sessions
= 0;
3446 mdsc
->last_snap_seq
= 0;
3447 init_rwsem(&mdsc
->snap_rwsem
);
3448 mdsc
->snap_realms
= RB_ROOT
;
3449 INIT_LIST_HEAD(&mdsc
->snap_empty
);
3450 spin_lock_init(&mdsc
->snap_empty_lock
);
3452 mdsc
->oldest_tid
= 0;
3453 mdsc
->request_tree
= RB_ROOT
;
3454 INIT_DELAYED_WORK(&mdsc
->delayed_work
, delayed_work
);
3455 mdsc
->last_renew_caps
= jiffies
;
3456 INIT_LIST_HEAD(&mdsc
->cap_delay_list
);
3457 spin_lock_init(&mdsc
->cap_delay_lock
);
3458 INIT_LIST_HEAD(&mdsc
->snap_flush_list
);
3459 spin_lock_init(&mdsc
->snap_flush_lock
);
3460 mdsc
->last_cap_flush_tid
= 1;
3461 mdsc
->cap_flush_tree
= RB_ROOT
;
3462 INIT_LIST_HEAD(&mdsc
->cap_dirty
);
3463 INIT_LIST_HEAD(&mdsc
->cap_dirty_migrating
);
3464 mdsc
->num_cap_flushing
= 0;
3465 spin_lock_init(&mdsc
->cap_dirty_lock
);
3466 init_waitqueue_head(&mdsc
->cap_flushing_wq
);
3467 spin_lock_init(&mdsc
->dentry_lru_lock
);
3468 INIT_LIST_HEAD(&mdsc
->dentry_lru
);
3470 ceph_caps_init(mdsc
);
3471 ceph_adjust_min_caps(mdsc
, fsc
->min_caps
);
3473 init_rwsem(&mdsc
->pool_perm_rwsem
);
3474 mdsc
->pool_perm_tree
= RB_ROOT
;
3480 * Wait for safe replies on open mds requests. If we time out, drop
3481 * all requests from the tree to avoid dangling dentry refs.
3483 static void wait_requests(struct ceph_mds_client
*mdsc
)
3485 struct ceph_options
*opts
= mdsc
->fsc
->client
->options
;
3486 struct ceph_mds_request
*req
;
3488 mutex_lock(&mdsc
->mutex
);
3489 if (__get_oldest_req(mdsc
)) {
3490 mutex_unlock(&mdsc
->mutex
);
3492 dout("wait_requests waiting for requests\n");
3493 wait_for_completion_timeout(&mdsc
->safe_umount_waiters
,
3494 ceph_timeout_jiffies(opts
->mount_timeout
));
3496 /* tear down remaining requests */
3497 mutex_lock(&mdsc
->mutex
);
3498 while ((req
= __get_oldest_req(mdsc
))) {
3499 dout("wait_requests timed out on tid %llu\n",
3501 __unregister_request(mdsc
, req
);
3504 mutex_unlock(&mdsc
->mutex
);
3505 dout("wait_requests done\n");
3509 * called before mount is ro, and before dentries are torn down.
3510 * (hmm, does this still race with new lookups?)
3512 void ceph_mdsc_pre_umount(struct ceph_mds_client
*mdsc
)
3514 dout("pre_umount\n");
3518 ceph_flush_dirty_caps(mdsc
);
3519 wait_requests(mdsc
);
3522 * wait for reply handlers to drop their request refs and
3523 * their inode/dcache refs
3529 * wait for all write mds requests to flush.
3531 static void wait_unsafe_requests(struct ceph_mds_client
*mdsc
, u64 want_tid
)
3533 struct ceph_mds_request
*req
= NULL
, *nextreq
;
3536 mutex_lock(&mdsc
->mutex
);
3537 dout("wait_unsafe_requests want %lld\n", want_tid
);
3539 req
= __get_oldest_req(mdsc
);
3540 while (req
&& req
->r_tid
<= want_tid
) {
3541 /* find next request */
3542 n
= rb_next(&req
->r_node
);
3544 nextreq
= rb_entry(n
, struct ceph_mds_request
, r_node
);
3547 if (req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
&&
3548 (req
->r_op
& CEPH_MDS_OP_WRITE
)) {
3550 ceph_mdsc_get_request(req
);
3552 ceph_mdsc_get_request(nextreq
);
3553 mutex_unlock(&mdsc
->mutex
);
3554 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3555 req
->r_tid
, want_tid
);
3556 wait_for_completion(&req
->r_safe_completion
);
3557 mutex_lock(&mdsc
->mutex
);
3558 ceph_mdsc_put_request(req
);
3560 break; /* next dne before, so we're done! */
3561 if (RB_EMPTY_NODE(&nextreq
->r_node
)) {
3562 /* next request was removed from tree */
3563 ceph_mdsc_put_request(nextreq
);
3566 ceph_mdsc_put_request(nextreq
); /* won't go away */
3570 mutex_unlock(&mdsc
->mutex
);
3571 dout("wait_unsafe_requests done\n");
3574 void ceph_mdsc_sync(struct ceph_mds_client
*mdsc
)
3576 u64 want_tid
, want_flush
, want_snap
;
3578 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
3582 mutex_lock(&mdsc
->mutex
);
3583 want_tid
= mdsc
->last_tid
;
3584 mutex_unlock(&mdsc
->mutex
);
3586 ceph_flush_dirty_caps(mdsc
);
3587 spin_lock(&mdsc
->cap_dirty_lock
);
3588 want_flush
= mdsc
->last_cap_flush_tid
;
3589 spin_unlock(&mdsc
->cap_dirty_lock
);
3591 down_read(&mdsc
->snap_rwsem
);
3592 want_snap
= mdsc
->last_snap_seq
;
3593 up_read(&mdsc
->snap_rwsem
);
3595 dout("sync want tid %lld flush_seq %lld snap_seq %lld\n",
3596 want_tid
, want_flush
, want_snap
);
3598 wait_unsafe_requests(mdsc
, want_tid
);
3599 wait_caps_flush(mdsc
, want_flush
, want_snap
);
3603 * true if all sessions are closed, or we force unmount
3605 static bool done_closing_sessions(struct ceph_mds_client
*mdsc
)
3607 if (ACCESS_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
3609 return atomic_read(&mdsc
->num_sessions
) == 0;
3613 * called after sb is ro.
3615 void ceph_mdsc_close_sessions(struct ceph_mds_client
*mdsc
)
3617 struct ceph_options
*opts
= mdsc
->fsc
->client
->options
;
3618 struct ceph_mds_session
*session
;
3621 dout("close_sessions\n");
3623 /* close sessions */
3624 mutex_lock(&mdsc
->mutex
);
3625 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3626 session
= __ceph_lookup_mds_session(mdsc
, i
);
3629 mutex_unlock(&mdsc
->mutex
);
3630 mutex_lock(&session
->s_mutex
);
3631 __close_session(mdsc
, session
);
3632 mutex_unlock(&session
->s_mutex
);
3633 ceph_put_mds_session(session
);
3634 mutex_lock(&mdsc
->mutex
);
3636 mutex_unlock(&mdsc
->mutex
);
3638 dout("waiting for sessions to close\n");
3639 wait_event_timeout(mdsc
->session_close_wq
, done_closing_sessions(mdsc
),
3640 ceph_timeout_jiffies(opts
->mount_timeout
));
3642 /* tear down remaining sessions */
3643 mutex_lock(&mdsc
->mutex
);
3644 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3645 if (mdsc
->sessions
[i
]) {
3646 session
= get_session(mdsc
->sessions
[i
]);
3647 __unregister_session(mdsc
, session
);
3648 mutex_unlock(&mdsc
->mutex
);
3649 mutex_lock(&session
->s_mutex
);
3650 remove_session_caps(session
);
3651 mutex_unlock(&session
->s_mutex
);
3652 ceph_put_mds_session(session
);
3653 mutex_lock(&mdsc
->mutex
);
3656 WARN_ON(!list_empty(&mdsc
->cap_delay_list
));
3657 mutex_unlock(&mdsc
->mutex
);
3659 ceph_cleanup_empty_realms(mdsc
);
3661 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3666 void ceph_mdsc_force_umount(struct ceph_mds_client
*mdsc
)
3668 struct ceph_mds_session
*session
;
3671 dout("force umount\n");
3673 mutex_lock(&mdsc
->mutex
);
3674 for (mds
= 0; mds
< mdsc
->max_sessions
; mds
++) {
3675 session
= __ceph_lookup_mds_session(mdsc
, mds
);
3678 mutex_unlock(&mdsc
->mutex
);
3679 mutex_lock(&session
->s_mutex
);
3680 __close_session(mdsc
, session
);
3681 if (session
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3682 cleanup_session_requests(mdsc
, session
);
3683 remove_session_caps(session
);
3685 mutex_unlock(&session
->s_mutex
);
3686 ceph_put_mds_session(session
);
3687 mutex_lock(&mdsc
->mutex
);
3688 kick_requests(mdsc
, mds
);
3690 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3691 mutex_unlock(&mdsc
->mutex
);
3694 static void ceph_mdsc_stop(struct ceph_mds_client
*mdsc
)
3697 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3699 ceph_mdsmap_destroy(mdsc
->mdsmap
);
3700 kfree(mdsc
->sessions
);
3701 ceph_caps_finalize(mdsc
);
3702 ceph_pool_perm_destroy(mdsc
);
3705 void ceph_mdsc_destroy(struct ceph_fs_client
*fsc
)
3707 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
3709 dout("mdsc_destroy %p\n", mdsc
);
3710 ceph_mdsc_stop(mdsc
);
3712 /* flush out any connection work with references to us */
3717 dout("mdsc_destroy %p done\n", mdsc
);
3722 * handle mds map update.
3724 void ceph_mdsc_handle_map(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3728 void *p
= msg
->front
.iov_base
;
3729 void *end
= p
+ msg
->front
.iov_len
;
3730 struct ceph_mdsmap
*newmap
, *oldmap
;
3731 struct ceph_fsid fsid
;
3734 ceph_decode_need(&p
, end
, sizeof(fsid
)+2*sizeof(u32
), bad
);
3735 ceph_decode_copy(&p
, &fsid
, sizeof(fsid
));
3736 if (ceph_check_fsid(mdsc
->fsc
->client
, &fsid
) < 0)
3738 epoch
= ceph_decode_32(&p
);
3739 maplen
= ceph_decode_32(&p
);
3740 dout("handle_map epoch %u len %d\n", epoch
, (int)maplen
);
3742 /* do we need it? */
3743 mutex_lock(&mdsc
->mutex
);
3744 if (mdsc
->mdsmap
&& epoch
<= mdsc
->mdsmap
->m_epoch
) {
3745 dout("handle_map epoch %u <= our %u\n",
3746 epoch
, mdsc
->mdsmap
->m_epoch
);
3747 mutex_unlock(&mdsc
->mutex
);
3751 newmap
= ceph_mdsmap_decode(&p
, end
);
3752 if (IS_ERR(newmap
)) {
3753 err
= PTR_ERR(newmap
);
3757 /* swap into place */
3759 oldmap
= mdsc
->mdsmap
;
3760 mdsc
->mdsmap
= newmap
;
3761 check_new_map(mdsc
, newmap
, oldmap
);
3762 ceph_mdsmap_destroy(oldmap
);
3764 mdsc
->mdsmap
= newmap
; /* first mds map */
3766 mdsc
->fsc
->sb
->s_maxbytes
= mdsc
->mdsmap
->m_max_file_size
;
3768 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3769 ceph_monc_got_map(&mdsc
->fsc
->client
->monc
, CEPH_SUB_MDSMAP
,
3770 mdsc
->mdsmap
->m_epoch
);
3772 mutex_unlock(&mdsc
->mutex
);
3773 schedule_delayed(mdsc
);
3777 mutex_unlock(&mdsc
->mutex
);
3779 pr_err("error decoding mdsmap %d\n", err
);
3783 static struct ceph_connection
*con_get(struct ceph_connection
*con
)
3785 struct ceph_mds_session
*s
= con
->private;
3787 if (get_session(s
)) {
3788 dout("mdsc con_get %p ok (%d)\n", s
, atomic_read(&s
->s_ref
));
3791 dout("mdsc con_get %p FAIL\n", s
);
3795 static void con_put(struct ceph_connection
*con
)
3797 struct ceph_mds_session
*s
= con
->private;
3799 dout("mdsc con_put %p (%d)\n", s
, atomic_read(&s
->s_ref
) - 1);
3800 ceph_put_mds_session(s
);
3804 * if the client is unresponsive for long enough, the mds will kill
3805 * the session entirely.
3807 static void peer_reset(struct ceph_connection
*con
)
3809 struct ceph_mds_session
*s
= con
->private;
3810 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3812 pr_warn("mds%d closed our session\n", s
->s_mds
);
3813 send_mds_reconnect(mdsc
, s
);
3816 static void dispatch(struct ceph_connection
*con
, struct ceph_msg
*msg
)
3818 struct ceph_mds_session
*s
= con
->private;
3819 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3820 int type
= le16_to_cpu(msg
->hdr
.type
);
3822 mutex_lock(&mdsc
->mutex
);
3823 if (__verify_registered_session(mdsc
, s
) < 0) {
3824 mutex_unlock(&mdsc
->mutex
);
3827 mutex_unlock(&mdsc
->mutex
);
3830 case CEPH_MSG_MDS_MAP
:
3831 ceph_mdsc_handle_map(mdsc
, msg
);
3833 case CEPH_MSG_CLIENT_SESSION
:
3834 handle_session(s
, msg
);
3836 case CEPH_MSG_CLIENT_REPLY
:
3837 handle_reply(s
, msg
);
3839 case CEPH_MSG_CLIENT_REQUEST_FORWARD
:
3840 handle_forward(mdsc
, s
, msg
);
3842 case CEPH_MSG_CLIENT_CAPS
:
3843 ceph_handle_caps(s
, msg
);
3845 case CEPH_MSG_CLIENT_SNAP
:
3846 ceph_handle_snap(mdsc
, s
, msg
);
3848 case CEPH_MSG_CLIENT_LEASE
:
3849 handle_lease(mdsc
, s
, msg
);
3853 pr_err("received unknown message type %d %s\n", type
,
3854 ceph_msg_type_name(type
));
3865 * Note: returned pointer is the address of a structure that's
3866 * managed separately. Caller must *not* attempt to free it.
3868 static struct ceph_auth_handshake
*get_authorizer(struct ceph_connection
*con
,
3869 int *proto
, int force_new
)
3871 struct ceph_mds_session
*s
= con
->private;
3872 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3873 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3874 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
3876 if (force_new
&& auth
->authorizer
) {
3877 ceph_auth_destroy_authorizer(auth
->authorizer
);
3878 auth
->authorizer
= NULL
;
3880 if (!auth
->authorizer
) {
3881 int ret
= ceph_auth_create_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3884 return ERR_PTR(ret
);
3886 int ret
= ceph_auth_update_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3889 return ERR_PTR(ret
);
3891 *proto
= ac
->protocol
;
3897 static int verify_authorizer_reply(struct ceph_connection
*con
, int len
)
3899 struct ceph_mds_session
*s
= con
->private;
3900 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3901 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3903 return ceph_auth_verify_authorizer_reply(ac
, s
->s_auth
.authorizer
, len
);
3906 static int invalidate_authorizer(struct ceph_connection
*con
)
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 ceph_auth_invalidate_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
);
3914 return ceph_monc_validate_auth(&mdsc
->fsc
->client
->monc
);
3917 static struct ceph_msg
*mds_alloc_msg(struct ceph_connection
*con
,
3918 struct ceph_msg_header
*hdr
, int *skip
)
3920 struct ceph_msg
*msg
;
3921 int type
= (int) le16_to_cpu(hdr
->type
);
3922 int front_len
= (int) le32_to_cpu(hdr
->front_len
);
3928 msg
= ceph_msg_new(type
, front_len
, GFP_NOFS
, false);
3930 pr_err("unable to allocate msg type %d len %d\n",
3938 static int mds_sign_message(struct ceph_msg
*msg
)
3940 struct ceph_mds_session
*s
= msg
->con
->private;
3941 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
3943 return ceph_auth_sign_message(auth
, msg
);
3946 static int mds_check_message_signature(struct ceph_msg
*msg
)
3948 struct ceph_mds_session
*s
= msg
->con
->private;
3949 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
3951 return ceph_auth_check_message_signature(auth
, msg
);
3954 static const struct ceph_connection_operations mds_con_ops
= {
3957 .dispatch
= dispatch
,
3958 .get_authorizer
= get_authorizer
,
3959 .verify_authorizer_reply
= verify_authorizer_reply
,
3960 .invalidate_authorizer
= invalidate_authorizer
,
3961 .peer_reset
= peer_reset
,
3962 .alloc_msg
= mds_alloc_msg
,
3963 .sign_message
= mds_sign_message
,
3964 .check_message_signature
= mds_check_message_signature
,