1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
5 #include <linux/wait.h>
6 #include <linux/slab.h>
8 #include <linux/sched.h>
9 #include <linux/debugfs.h>
10 #include <linux/seq_file.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45 * are no longer valid.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_MDS_QUOTA
) {
104 u8 struct_v
, struct_compat
;
108 * both struct_v and struct_compat are expected to be >= 1
110 ceph_decode_8_safe(p
, end
, struct_v
, bad
);
111 ceph_decode_8_safe(p
, end
, struct_compat
, bad
);
112 if (!struct_v
|| !struct_compat
)
114 ceph_decode_32_safe(p
, end
, struct_len
, bad
);
115 ceph_decode_need(p
, end
, struct_len
, bad
);
116 ceph_decode_64_safe(p
, end
, info
->max_bytes
, bad
);
117 ceph_decode_64_safe(p
, end
, info
->max_files
, bad
);
123 info
->pool_ns_len
= 0;
124 info
->pool_ns_data
= NULL
;
125 if (features
& CEPH_FEATURE_FS_FILE_LAYOUT_V2
) {
126 ceph_decode_32_safe(p
, end
, info
->pool_ns_len
, bad
);
127 if (info
->pool_ns_len
> 0) {
128 ceph_decode_need(p
, end
, info
->pool_ns_len
, bad
);
129 info
->pool_ns_data
= *p
;
130 *p
+= info
->pool_ns_len
;
140 * parse a normal reply, which may contain a (dir+)dentry and/or a
143 static int parse_reply_info_trace(void **p
, void *end
,
144 struct ceph_mds_reply_info_parsed
*info
,
149 if (info
->head
->is_dentry
) {
150 err
= parse_reply_info_in(p
, end
, &info
->diri
, features
);
154 if (unlikely(*p
+ sizeof(*info
->dirfrag
) > end
))
157 *p
+= sizeof(*info
->dirfrag
) +
158 sizeof(u32
)*le32_to_cpu(info
->dirfrag
->ndist
);
159 if (unlikely(*p
> end
))
162 ceph_decode_32_safe(p
, end
, info
->dname_len
, bad
);
163 ceph_decode_need(p
, end
, info
->dname_len
, bad
);
165 *p
+= info
->dname_len
;
167 *p
+= sizeof(*info
->dlease
);
170 if (info
->head
->is_target
) {
171 err
= parse_reply_info_in(p
, end
, &info
->targeti
, features
);
176 if (unlikely(*p
!= end
))
183 pr_err("problem parsing mds trace %d\n", err
);
188 * parse readdir results
190 static int parse_reply_info_dir(void **p
, void *end
,
191 struct ceph_mds_reply_info_parsed
*info
,
198 if (*p
+ sizeof(*info
->dir_dir
) > end
)
200 *p
+= sizeof(*info
->dir_dir
) +
201 sizeof(u32
)*le32_to_cpu(info
->dir_dir
->ndist
);
205 ceph_decode_need(p
, end
, sizeof(num
) + 2, bad
);
206 num
= ceph_decode_32(p
);
208 u16 flags
= ceph_decode_16(p
);
209 info
->dir_end
= !!(flags
& CEPH_READDIR_FRAG_END
);
210 info
->dir_complete
= !!(flags
& CEPH_READDIR_FRAG_COMPLETE
);
211 info
->hash_order
= !!(flags
& CEPH_READDIR_HASH_ORDER
);
212 info
->offset_hash
= !!(flags
& CEPH_READDIR_OFFSET_HASH
);
217 BUG_ON(!info
->dir_entries
);
218 if ((unsigned long)(info
->dir_entries
+ num
) >
219 (unsigned long)info
->dir_entries
+ info
->dir_buf_size
) {
220 pr_err("dir contents are larger than expected\n");
227 struct ceph_mds_reply_dir_entry
*rde
= info
->dir_entries
+ i
;
229 ceph_decode_need(p
, end
, sizeof(u32
)*2, bad
);
230 rde
->name_len
= ceph_decode_32(p
);
231 ceph_decode_need(p
, end
, rde
->name_len
, bad
);
234 dout("parsed dir dname '%.*s'\n", rde
->name_len
, rde
->name
);
236 *p
+= sizeof(struct ceph_mds_reply_lease
);
239 err
= parse_reply_info_in(p
, end
, &rde
->inode
, features
);
242 /* ceph_readdir_prepopulate() will update it */
256 pr_err("problem parsing dir contents %d\n", err
);
261 * parse fcntl F_GETLK results
263 static int parse_reply_info_filelock(void **p
, void *end
,
264 struct ceph_mds_reply_info_parsed
*info
,
267 if (*p
+ sizeof(*info
->filelock_reply
) > end
)
270 info
->filelock_reply
= *p
;
271 *p
+= sizeof(*info
->filelock_reply
);
273 if (unlikely(*p
!= end
))
282 * parse create results
284 static int parse_reply_info_create(void **p
, void *end
,
285 struct ceph_mds_reply_info_parsed
*info
,
288 if (features
& CEPH_FEATURE_REPLY_CREATE_INODE
) {
290 info
->has_create_ino
= false;
292 info
->has_create_ino
= true;
293 info
->ino
= ceph_decode_64(p
);
297 if (unlikely(*p
!= end
))
306 * parse extra results
308 static int parse_reply_info_extra(void **p
, void *end
,
309 struct ceph_mds_reply_info_parsed
*info
,
312 u32 op
= le32_to_cpu(info
->head
->op
);
314 if (op
== CEPH_MDS_OP_GETFILELOCK
)
315 return parse_reply_info_filelock(p
, end
, info
, features
);
316 else if (op
== CEPH_MDS_OP_READDIR
|| op
== CEPH_MDS_OP_LSSNAP
)
317 return parse_reply_info_dir(p
, end
, info
, features
);
318 else if (op
== CEPH_MDS_OP_CREATE
)
319 return parse_reply_info_create(p
, end
, info
, features
);
325 * parse entire mds reply
327 static int parse_reply_info(struct ceph_msg
*msg
,
328 struct ceph_mds_reply_info_parsed
*info
,
335 info
->head
= msg
->front
.iov_base
;
336 p
= msg
->front
.iov_base
+ sizeof(struct ceph_mds_reply_head
);
337 end
= p
+ msg
->front
.iov_len
- sizeof(struct ceph_mds_reply_head
);
340 ceph_decode_32_safe(&p
, end
, len
, bad
);
342 ceph_decode_need(&p
, end
, len
, bad
);
343 err
= parse_reply_info_trace(&p
, p
+len
, info
, features
);
349 ceph_decode_32_safe(&p
, end
, len
, bad
);
351 ceph_decode_need(&p
, end
, len
, bad
);
352 err
= parse_reply_info_extra(&p
, p
+len
, info
, features
);
358 ceph_decode_32_safe(&p
, end
, len
, bad
);
359 info
->snapblob_len
= len
;
370 pr_err("mds parse_reply err %d\n", err
);
374 static void destroy_reply_info(struct ceph_mds_reply_info_parsed
*info
)
376 if (!info
->dir_entries
)
378 free_pages((unsigned long)info
->dir_entries
, get_order(info
->dir_buf_size
));
385 const char *ceph_session_state_name(int s
)
388 case CEPH_MDS_SESSION_NEW
: return "new";
389 case CEPH_MDS_SESSION_OPENING
: return "opening";
390 case CEPH_MDS_SESSION_OPEN
: return "open";
391 case CEPH_MDS_SESSION_HUNG
: return "hung";
392 case CEPH_MDS_SESSION_CLOSING
: return "closing";
393 case CEPH_MDS_SESSION_RESTARTING
: return "restarting";
394 case CEPH_MDS_SESSION_RECONNECTING
: return "reconnecting";
395 case CEPH_MDS_SESSION_REJECTED
: return "rejected";
396 default: return "???";
400 static struct ceph_mds_session
*get_session(struct ceph_mds_session
*s
)
402 if (refcount_inc_not_zero(&s
->s_ref
)) {
403 dout("mdsc get_session %p %d -> %d\n", s
,
404 refcount_read(&s
->s_ref
)-1, refcount_read(&s
->s_ref
));
407 dout("mdsc get_session %p 0 -- FAIL\n", s
);
412 void ceph_put_mds_session(struct ceph_mds_session
*s
)
414 dout("mdsc put_session %p %d -> %d\n", s
,
415 refcount_read(&s
->s_ref
), refcount_read(&s
->s_ref
)-1);
416 if (refcount_dec_and_test(&s
->s_ref
)) {
417 if (s
->s_auth
.authorizer
)
418 ceph_auth_destroy_authorizer(s
->s_auth
.authorizer
);
424 * called under mdsc->mutex
426 struct ceph_mds_session
*__ceph_lookup_mds_session(struct ceph_mds_client
*mdsc
,
429 struct ceph_mds_session
*session
;
431 if (mds
>= mdsc
->max_sessions
|| !mdsc
->sessions
[mds
])
433 session
= mdsc
->sessions
[mds
];
434 dout("lookup_mds_session %p %d\n", session
,
435 refcount_read(&session
->s_ref
));
436 get_session(session
);
440 static bool __have_session(struct ceph_mds_client
*mdsc
, int mds
)
442 if (mds
>= mdsc
->max_sessions
|| !mdsc
->sessions
[mds
])
448 static int __verify_registered_session(struct ceph_mds_client
*mdsc
,
449 struct ceph_mds_session
*s
)
451 if (s
->s_mds
>= mdsc
->max_sessions
||
452 mdsc
->sessions
[s
->s_mds
] != s
)
458 * create+register a new session for given mds.
459 * called under mdsc->mutex.
461 static struct ceph_mds_session
*register_session(struct ceph_mds_client
*mdsc
,
464 struct ceph_mds_session
*s
;
466 if (mds
>= mdsc
->mdsmap
->m_num_mds
)
467 return ERR_PTR(-EINVAL
);
469 s
= kzalloc(sizeof(*s
), GFP_NOFS
);
471 return ERR_PTR(-ENOMEM
);
473 if (mds
>= mdsc
->max_sessions
) {
474 int newmax
= 1 << get_count_order(mds
+ 1);
475 struct ceph_mds_session
**sa
;
477 dout("%s: realloc to %d\n", __func__
, newmax
);
478 sa
= kcalloc(newmax
, sizeof(void *), GFP_NOFS
);
481 if (mdsc
->sessions
) {
482 memcpy(sa
, mdsc
->sessions
,
483 mdsc
->max_sessions
* sizeof(void *));
484 kfree(mdsc
->sessions
);
487 mdsc
->max_sessions
= newmax
;
490 dout("%s: mds%d\n", __func__
, mds
);
493 s
->s_state
= CEPH_MDS_SESSION_NEW
;
496 mutex_init(&s
->s_mutex
);
498 ceph_con_init(&s
->s_con
, s
, &mds_con_ops
, &mdsc
->fsc
->client
->msgr
);
500 spin_lock_init(&s
->s_gen_ttl_lock
);
502 s
->s_cap_ttl
= jiffies
- 1;
504 spin_lock_init(&s
->s_cap_lock
);
505 s
->s_renew_requested
= 0;
507 INIT_LIST_HEAD(&s
->s_caps
);
510 refcount_set(&s
->s_ref
, 1);
511 INIT_LIST_HEAD(&s
->s_waiting
);
512 INIT_LIST_HEAD(&s
->s_unsafe
);
513 s
->s_num_cap_releases
= 0;
514 s
->s_cap_reconnect
= 0;
515 s
->s_cap_iterator
= NULL
;
516 INIT_LIST_HEAD(&s
->s_cap_releases
);
517 INIT_LIST_HEAD(&s
->s_cap_flushing
);
519 mdsc
->sessions
[mds
] = s
;
520 atomic_inc(&mdsc
->num_sessions
);
521 refcount_inc(&s
->s_ref
); /* one ref to sessions[], one to caller */
523 ceph_con_open(&s
->s_con
, CEPH_ENTITY_TYPE_MDS
, mds
,
524 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
530 return ERR_PTR(-ENOMEM
);
534 * called under mdsc->mutex
536 static void __unregister_session(struct ceph_mds_client
*mdsc
,
537 struct ceph_mds_session
*s
)
539 dout("__unregister_session mds%d %p\n", s
->s_mds
, s
);
540 BUG_ON(mdsc
->sessions
[s
->s_mds
] != s
);
541 mdsc
->sessions
[s
->s_mds
] = NULL
;
542 ceph_con_close(&s
->s_con
);
543 ceph_put_mds_session(s
);
544 atomic_dec(&mdsc
->num_sessions
);
548 * drop session refs in request.
550 * should be last request ref, or hold mdsc->mutex
552 static void put_request_session(struct ceph_mds_request
*req
)
554 if (req
->r_session
) {
555 ceph_put_mds_session(req
->r_session
);
556 req
->r_session
= NULL
;
560 void ceph_mdsc_release_request(struct kref
*kref
)
562 struct ceph_mds_request
*req
= container_of(kref
,
563 struct ceph_mds_request
,
565 destroy_reply_info(&req
->r_reply_info
);
567 ceph_msg_put(req
->r_request
);
569 ceph_msg_put(req
->r_reply
);
571 ceph_put_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
575 ceph_put_cap_refs(ceph_inode(req
->r_parent
), CEPH_CAP_PIN
);
576 iput(req
->r_target_inode
);
579 if (req
->r_old_dentry
)
580 dput(req
->r_old_dentry
);
581 if (req
->r_old_dentry_dir
) {
583 * track (and drop pins for) r_old_dentry_dir
584 * separately, since r_old_dentry's d_parent may have
585 * changed between the dir mutex being dropped and
586 * this request being freed.
588 ceph_put_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
590 iput(req
->r_old_dentry_dir
);
595 ceph_pagelist_release(req
->r_pagelist
);
596 put_request_session(req
);
597 ceph_unreserve_caps(req
->r_mdsc
, &req
->r_caps_reservation
);
601 DEFINE_RB_FUNCS(request
, struct ceph_mds_request
, r_tid
, r_node
)
604 * lookup session, bump ref if found.
606 * called under mdsc->mutex.
608 static struct ceph_mds_request
*
609 lookup_get_request(struct ceph_mds_client
*mdsc
, u64 tid
)
611 struct ceph_mds_request
*req
;
613 req
= lookup_request(&mdsc
->request_tree
, tid
);
615 ceph_mdsc_get_request(req
);
621 * Register an in-flight request, and assign a tid. Link to directory
622 * are modifying (if any).
624 * Called under mdsc->mutex.
626 static void __register_request(struct ceph_mds_client
*mdsc
,
627 struct ceph_mds_request
*req
,
632 req
->r_tid
= ++mdsc
->last_tid
;
633 if (req
->r_num_caps
) {
634 ret
= ceph_reserve_caps(mdsc
, &req
->r_caps_reservation
,
637 pr_err("__register_request %p "
638 "failed to reserve caps: %d\n", req
, ret
);
639 /* set req->r_err to fail early from __do_request */
644 dout("__register_request %p tid %lld\n", req
, req
->r_tid
);
645 ceph_mdsc_get_request(req
);
646 insert_request(&mdsc
->request_tree
, req
);
648 req
->r_uid
= current_fsuid();
649 req
->r_gid
= current_fsgid();
651 if (mdsc
->oldest_tid
== 0 && req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
)
652 mdsc
->oldest_tid
= req
->r_tid
;
656 req
->r_unsafe_dir
= dir
;
660 static void __unregister_request(struct ceph_mds_client
*mdsc
,
661 struct ceph_mds_request
*req
)
663 dout("__unregister_request %p tid %lld\n", req
, req
->r_tid
);
665 /* Never leave an unregistered request on an unsafe list! */
666 list_del_init(&req
->r_unsafe_item
);
668 if (req
->r_tid
== mdsc
->oldest_tid
) {
669 struct rb_node
*p
= rb_next(&req
->r_node
);
670 mdsc
->oldest_tid
= 0;
672 struct ceph_mds_request
*next_req
=
673 rb_entry(p
, struct ceph_mds_request
, r_node
);
674 if (next_req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
) {
675 mdsc
->oldest_tid
= next_req
->r_tid
;
682 erase_request(&mdsc
->request_tree
, req
);
684 if (req
->r_unsafe_dir
&&
685 test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
)) {
686 struct ceph_inode_info
*ci
= ceph_inode(req
->r_unsafe_dir
);
687 spin_lock(&ci
->i_unsafe_lock
);
688 list_del_init(&req
->r_unsafe_dir_item
);
689 spin_unlock(&ci
->i_unsafe_lock
);
691 if (req
->r_target_inode
&&
692 test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
)) {
693 struct ceph_inode_info
*ci
= ceph_inode(req
->r_target_inode
);
694 spin_lock(&ci
->i_unsafe_lock
);
695 list_del_init(&req
->r_unsafe_target_item
);
696 spin_unlock(&ci
->i_unsafe_lock
);
699 if (req
->r_unsafe_dir
) {
700 iput(req
->r_unsafe_dir
);
701 req
->r_unsafe_dir
= NULL
;
704 complete_all(&req
->r_safe_completion
);
706 ceph_mdsc_put_request(req
);
710 * Walk back up the dentry tree until we hit a dentry representing a
711 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
712 * when calling this) to ensure that the objects won't disappear while we're
713 * working with them. Once we hit a candidate dentry, we attempt to take a
714 * reference to it, and return that as the result.
716 static struct inode
*get_nonsnap_parent(struct dentry
*dentry
)
718 struct inode
*inode
= NULL
;
720 while (dentry
&& !IS_ROOT(dentry
)) {
721 inode
= d_inode_rcu(dentry
);
722 if (!inode
|| ceph_snap(inode
) == CEPH_NOSNAP
)
724 dentry
= dentry
->d_parent
;
727 inode
= igrab(inode
);
732 * Choose mds to send request to next. If there is a hint set in the
733 * request (e.g., due to a prior forward hint from the mds), use that.
734 * Otherwise, consult frag tree and/or caps to identify the
735 * appropriate mds. If all else fails, choose randomly.
737 * Called under mdsc->mutex.
739 static int __choose_mds(struct ceph_mds_client
*mdsc
,
740 struct ceph_mds_request
*req
)
743 struct ceph_inode_info
*ci
;
744 struct ceph_cap
*cap
;
745 int mode
= req
->r_direct_mode
;
747 u32 hash
= req
->r_direct_hash
;
748 bool is_hash
= test_bit(CEPH_MDS_R_DIRECT_IS_HASH
, &req
->r_req_flags
);
751 * is there a specific mds we should try? ignore hint if we have
752 * no session and the mds is not up (active or recovering).
754 if (req
->r_resend_mds
>= 0 &&
755 (__have_session(mdsc
, req
->r_resend_mds
) ||
756 ceph_mdsmap_get_state(mdsc
->mdsmap
, req
->r_resend_mds
) > 0)) {
757 dout("choose_mds using resend_mds mds%d\n",
759 return req
->r_resend_mds
;
762 if (mode
== USE_RANDOM_MDS
)
767 if (ceph_snap(req
->r_inode
) != CEPH_SNAPDIR
) {
768 inode
= req
->r_inode
;
771 /* req->r_dentry is non-null for LSSNAP request */
773 inode
= get_nonsnap_parent(req
->r_dentry
);
775 dout("__choose_mds using snapdir's parent %p\n", inode
);
777 } else if (req
->r_dentry
) {
778 /* ignore race with rename; old or new d_parent is okay */
779 struct dentry
*parent
;
783 parent
= req
->r_dentry
->d_parent
;
784 dir
= req
->r_parent
? : d_inode_rcu(parent
);
786 if (!dir
|| dir
->i_sb
!= mdsc
->fsc
->sb
) {
787 /* not this fs or parent went negative */
788 inode
= d_inode(req
->r_dentry
);
791 } else if (ceph_snap(dir
) != CEPH_NOSNAP
) {
792 /* direct snapped/virtual snapdir requests
793 * based on parent dir inode */
794 inode
= get_nonsnap_parent(parent
);
795 dout("__choose_mds using nonsnap parent %p\n", inode
);
798 inode
= d_inode(req
->r_dentry
);
799 if (!inode
|| mode
== USE_AUTH_MDS
) {
802 hash
= ceph_dentry_hash(dir
, req
->r_dentry
);
811 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode
, (int)is_hash
,
815 ci
= ceph_inode(inode
);
817 if (is_hash
&& S_ISDIR(inode
->i_mode
)) {
818 struct ceph_inode_frag frag
;
821 ceph_choose_frag(ci
, hash
, &frag
, &found
);
823 if (mode
== USE_ANY_MDS
&& frag
.ndist
> 0) {
826 /* choose a random replica */
827 get_random_bytes(&r
, 1);
830 dout("choose_mds %p %llx.%llx "
831 "frag %u mds%d (%d/%d)\n",
832 inode
, ceph_vinop(inode
),
835 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
836 CEPH_MDS_STATE_ACTIVE
)
840 /* since this file/dir wasn't known to be
841 * replicated, then we want to look for the
842 * authoritative mds. */
845 /* choose auth mds */
847 dout("choose_mds %p %llx.%llx "
848 "frag %u mds%d (auth)\n",
849 inode
, ceph_vinop(inode
), frag
.frag
, mds
);
850 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
851 CEPH_MDS_STATE_ACTIVE
)
857 spin_lock(&ci
->i_ceph_lock
);
859 if (mode
== USE_AUTH_MDS
)
860 cap
= ci
->i_auth_cap
;
861 if (!cap
&& !RB_EMPTY_ROOT(&ci
->i_caps
))
862 cap
= rb_entry(rb_first(&ci
->i_caps
), struct ceph_cap
, ci_node
);
864 spin_unlock(&ci
->i_ceph_lock
);
868 mds
= cap
->session
->s_mds
;
869 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
870 inode
, ceph_vinop(inode
), mds
,
871 cap
== ci
->i_auth_cap
? "auth " : "", cap
);
872 spin_unlock(&ci
->i_ceph_lock
);
878 mds
= ceph_mdsmap_get_random_mds(mdsc
->mdsmap
);
879 dout("choose_mds chose random mds%d\n", mds
);
887 static struct ceph_msg
*create_session_msg(u32 op
, u64 seq
)
889 struct ceph_msg
*msg
;
890 struct ceph_mds_session_head
*h
;
892 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
), GFP_NOFS
,
895 pr_err("create_session_msg ENOMEM creating msg\n");
898 h
= msg
->front
.iov_base
;
899 h
->op
= cpu_to_le32(op
);
900 h
->seq
= cpu_to_le64(seq
);
905 static void encode_supported_features(void **p
, void *end
)
907 static const unsigned char bits
[] = CEPHFS_FEATURES_CLIENT_SUPPORTED
;
908 static const size_t count
= ARRAY_SIZE(bits
);
912 size_t size
= ((size_t)bits
[count
- 1] + 64) / 64 * 8;
914 BUG_ON(*p
+ 4 + size
> end
);
915 ceph_encode_32(p
, size
);
917 for (i
= 0; i
< count
; i
++)
918 ((unsigned char*)(*p
))[i
/ 8] |= 1 << (bits
[i
] % 8);
921 BUG_ON(*p
+ 4 > end
);
922 ceph_encode_32(p
, 0);
927 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
928 * to include additional client metadata fields.
930 static struct ceph_msg
*create_session_open_msg(struct ceph_mds_client
*mdsc
, u64 seq
)
932 struct ceph_msg
*msg
;
933 struct ceph_mds_session_head
*h
;
936 int metadata_key_count
= 0;
937 struct ceph_options
*opt
= mdsc
->fsc
->client
->options
;
938 struct ceph_mount_options
*fsopt
= mdsc
->fsc
->mount_options
;
941 const char* metadata
[][2] = {
942 {"hostname", mdsc
->nodename
},
943 {"kernel_version", init_utsname()->release
},
944 {"entity_id", opt
->name
? : ""},
945 {"root", fsopt
->server_path
? : "/"},
949 /* Calculate serialized length of metadata */
950 extra_bytes
= 4; /* map length */
951 for (i
= 0; metadata
[i
][0]; ++i
) {
952 extra_bytes
+= 8 + strlen(metadata
[i
][0]) +
953 strlen(metadata
[i
][1]);
954 metadata_key_count
++;
956 /* supported feature */
957 extra_bytes
+= 4 + 8;
959 /* Allocate the message */
960 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
) + extra_bytes
,
963 pr_err("create_session_msg ENOMEM creating msg\n");
966 p
= msg
->front
.iov_base
;
967 end
= p
+ msg
->front
.iov_len
;
970 h
->op
= cpu_to_le32(CEPH_SESSION_REQUEST_OPEN
);
971 h
->seq
= cpu_to_le64(seq
);
974 * Serialize client metadata into waiting buffer space, using
975 * the format that userspace expects for map<string, string>
977 * ClientSession messages with metadata are v2
979 msg
->hdr
.version
= cpu_to_le16(3);
980 msg
->hdr
.compat_version
= cpu_to_le16(1);
982 /* The write pointer, following the session_head structure */
985 /* Number of entries in the map */
986 ceph_encode_32(&p
, metadata_key_count
);
988 /* Two length-prefixed strings for each entry in the map */
989 for (i
= 0; metadata
[i
][0]; ++i
) {
990 size_t const key_len
= strlen(metadata
[i
][0]);
991 size_t const val_len
= strlen(metadata
[i
][1]);
993 ceph_encode_32(&p
, key_len
);
994 memcpy(p
, metadata
[i
][0], key_len
);
996 ceph_encode_32(&p
, val_len
);
997 memcpy(p
, metadata
[i
][1], val_len
);
1001 encode_supported_features(&p
, end
);
1002 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
1003 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1009 * send session open request.
1011 * called under mdsc->mutex
1013 static int __open_session(struct ceph_mds_client
*mdsc
,
1014 struct ceph_mds_session
*session
)
1016 struct ceph_msg
*msg
;
1018 int mds
= session
->s_mds
;
1020 /* wait for mds to go active? */
1021 mstate
= ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
);
1022 dout("open_session to mds%d (%s)\n", mds
,
1023 ceph_mds_state_name(mstate
));
1024 session
->s_state
= CEPH_MDS_SESSION_OPENING
;
1025 session
->s_renew_requested
= jiffies
;
1027 /* send connect message */
1028 msg
= create_session_open_msg(mdsc
, session
->s_seq
);
1031 ceph_con_send(&session
->s_con
, msg
);
1036 * open sessions for any export targets for the given mds
1038 * called under mdsc->mutex
1040 static struct ceph_mds_session
*
1041 __open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
1043 struct ceph_mds_session
*session
;
1045 session
= __ceph_lookup_mds_session(mdsc
, target
);
1047 session
= register_session(mdsc
, target
);
1048 if (IS_ERR(session
))
1051 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
1052 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
1053 __open_session(mdsc
, session
);
1058 struct ceph_mds_session
*
1059 ceph_mdsc_open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
1061 struct ceph_mds_session
*session
;
1063 dout("open_export_target_session to mds%d\n", target
);
1065 mutex_lock(&mdsc
->mutex
);
1066 session
= __open_export_target_session(mdsc
, target
);
1067 mutex_unlock(&mdsc
->mutex
);
1072 static void __open_export_target_sessions(struct ceph_mds_client
*mdsc
,
1073 struct ceph_mds_session
*session
)
1075 struct ceph_mds_info
*mi
;
1076 struct ceph_mds_session
*ts
;
1077 int i
, mds
= session
->s_mds
;
1079 if (mds
>= mdsc
->mdsmap
->m_num_mds
)
1082 mi
= &mdsc
->mdsmap
->m_info
[mds
];
1083 dout("open_export_target_sessions for mds%d (%d targets)\n",
1084 session
->s_mds
, mi
->num_export_targets
);
1086 for (i
= 0; i
< mi
->num_export_targets
; i
++) {
1087 ts
= __open_export_target_session(mdsc
, mi
->export_targets
[i
]);
1089 ceph_put_mds_session(ts
);
1093 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client
*mdsc
,
1094 struct ceph_mds_session
*session
)
1096 mutex_lock(&mdsc
->mutex
);
1097 __open_export_target_sessions(mdsc
, session
);
1098 mutex_unlock(&mdsc
->mutex
);
1105 static void detach_cap_releases(struct ceph_mds_session
*session
,
1106 struct list_head
*target
)
1108 lockdep_assert_held(&session
->s_cap_lock
);
1110 list_splice_init(&session
->s_cap_releases
, target
);
1111 session
->s_num_cap_releases
= 0;
1112 dout("dispose_cap_releases mds%d\n", session
->s_mds
);
1115 static void dispose_cap_releases(struct ceph_mds_client
*mdsc
,
1116 struct list_head
*dispose
)
1118 while (!list_empty(dispose
)) {
1119 struct ceph_cap
*cap
;
1120 /* zero out the in-progress message */
1121 cap
= list_first_entry(dispose
, struct ceph_cap
, session_caps
);
1122 list_del(&cap
->session_caps
);
1123 ceph_put_cap(mdsc
, cap
);
1127 static void cleanup_session_requests(struct ceph_mds_client
*mdsc
,
1128 struct ceph_mds_session
*session
)
1130 struct ceph_mds_request
*req
;
1133 dout("cleanup_session_requests mds%d\n", session
->s_mds
);
1134 mutex_lock(&mdsc
->mutex
);
1135 while (!list_empty(&session
->s_unsafe
)) {
1136 req
= list_first_entry(&session
->s_unsafe
,
1137 struct ceph_mds_request
, r_unsafe_item
);
1138 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1140 __unregister_request(mdsc
, req
);
1142 /* zero r_attempts, so kick_requests() will re-send requests */
1143 p
= rb_first(&mdsc
->request_tree
);
1145 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
1147 if (req
->r_session
&&
1148 req
->r_session
->s_mds
== session
->s_mds
)
1149 req
->r_attempts
= 0;
1151 mutex_unlock(&mdsc
->mutex
);
1155 * Helper to safely iterate over all caps associated with a session, with
1156 * special care taken to handle a racing __ceph_remove_cap().
1158 * Caller must hold session s_mutex.
1160 static int iterate_session_caps(struct ceph_mds_session
*session
,
1161 int (*cb
)(struct inode
*, struct ceph_cap
*,
1164 struct list_head
*p
;
1165 struct ceph_cap
*cap
;
1166 struct inode
*inode
, *last_inode
= NULL
;
1167 struct ceph_cap
*old_cap
= NULL
;
1170 dout("iterate_session_caps %p mds%d\n", session
, session
->s_mds
);
1171 spin_lock(&session
->s_cap_lock
);
1172 p
= session
->s_caps
.next
;
1173 while (p
!= &session
->s_caps
) {
1174 cap
= list_entry(p
, struct ceph_cap
, session_caps
);
1175 inode
= igrab(&cap
->ci
->vfs_inode
);
1180 session
->s_cap_iterator
= cap
;
1181 spin_unlock(&session
->s_cap_lock
);
1188 ceph_put_cap(session
->s_mdsc
, old_cap
);
1192 ret
= cb(inode
, cap
, arg
);
1195 spin_lock(&session
->s_cap_lock
);
1198 dout("iterate_session_caps finishing cap %p removal\n",
1200 BUG_ON(cap
->session
!= session
);
1201 cap
->session
= NULL
;
1202 list_del_init(&cap
->session_caps
);
1203 session
->s_nr_caps
--;
1204 if (cap
->queue_release
) {
1205 list_add_tail(&cap
->session_caps
,
1206 &session
->s_cap_releases
);
1207 session
->s_num_cap_releases
++;
1209 old_cap
= cap
; /* put_cap it w/o locks held */
1217 session
->s_cap_iterator
= NULL
;
1218 spin_unlock(&session
->s_cap_lock
);
1222 ceph_put_cap(session
->s_mdsc
, old_cap
);
1227 static int remove_session_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1230 struct ceph_fs_client
*fsc
= (struct ceph_fs_client
*)arg
;
1231 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1232 LIST_HEAD(to_remove
);
1234 bool invalidate
= false;
1236 dout("removing cap %p, ci is %p, inode is %p\n",
1237 cap
, ci
, &ci
->vfs_inode
);
1238 spin_lock(&ci
->i_ceph_lock
);
1239 __ceph_remove_cap(cap
, false);
1240 if (!ci
->i_auth_cap
) {
1241 struct ceph_cap_flush
*cf
;
1242 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
1244 ci
->i_ceph_flags
|= CEPH_I_CAP_DROPPED
;
1246 if (ci
->i_wrbuffer_ref
> 0 &&
1247 READ_ONCE(fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
1250 while (!list_empty(&ci
->i_cap_flush_list
)) {
1251 cf
= list_first_entry(&ci
->i_cap_flush_list
,
1252 struct ceph_cap_flush
, i_list
);
1253 list_move(&cf
->i_list
, &to_remove
);
1256 spin_lock(&mdsc
->cap_dirty_lock
);
1258 list_for_each_entry(cf
, &to_remove
, i_list
)
1259 list_del(&cf
->g_list
);
1261 if (!list_empty(&ci
->i_dirty_item
)) {
1262 pr_warn_ratelimited(
1263 " dropping dirty %s state for %p %lld\n",
1264 ceph_cap_string(ci
->i_dirty_caps
),
1265 inode
, ceph_ino(inode
));
1266 ci
->i_dirty_caps
= 0;
1267 list_del_init(&ci
->i_dirty_item
);
1270 if (!list_empty(&ci
->i_flushing_item
)) {
1271 pr_warn_ratelimited(
1272 " dropping dirty+flushing %s state for %p %lld\n",
1273 ceph_cap_string(ci
->i_flushing_caps
),
1274 inode
, ceph_ino(inode
));
1275 ci
->i_flushing_caps
= 0;
1276 list_del_init(&ci
->i_flushing_item
);
1277 mdsc
->num_cap_flushing
--;
1280 spin_unlock(&mdsc
->cap_dirty_lock
);
1282 if (atomic_read(&ci
->i_filelock_ref
) > 0) {
1283 /* make further file lock syscall return -EIO */
1284 ci
->i_ceph_flags
|= CEPH_I_ERROR_FILELOCK
;
1285 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1286 inode
, ceph_ino(inode
));
1289 if (!ci
->i_dirty_caps
&& ci
->i_prealloc_cap_flush
) {
1290 list_add(&ci
->i_prealloc_cap_flush
->i_list
, &to_remove
);
1291 ci
->i_prealloc_cap_flush
= NULL
;
1294 spin_unlock(&ci
->i_ceph_lock
);
1295 while (!list_empty(&to_remove
)) {
1296 struct ceph_cap_flush
*cf
;
1297 cf
= list_first_entry(&to_remove
,
1298 struct ceph_cap_flush
, i_list
);
1299 list_del(&cf
->i_list
);
1300 ceph_free_cap_flush(cf
);
1303 wake_up_all(&ci
->i_cap_wq
);
1305 ceph_queue_invalidate(inode
);
1312 * caller must hold session s_mutex
1314 static void remove_session_caps(struct ceph_mds_session
*session
)
1316 struct ceph_fs_client
*fsc
= session
->s_mdsc
->fsc
;
1317 struct super_block
*sb
= fsc
->sb
;
1320 dout("remove_session_caps on %p\n", session
);
1321 iterate_session_caps(session
, remove_session_caps_cb
, fsc
);
1323 wake_up_all(&fsc
->mdsc
->cap_flushing_wq
);
1325 spin_lock(&session
->s_cap_lock
);
1326 if (session
->s_nr_caps
> 0) {
1327 struct inode
*inode
;
1328 struct ceph_cap
*cap
, *prev
= NULL
;
1329 struct ceph_vino vino
;
1331 * iterate_session_caps() skips inodes that are being
1332 * deleted, we need to wait until deletions are complete.
1333 * __wait_on_freeing_inode() is designed for the job,
1334 * but it is not exported, so use lookup inode function
1337 while (!list_empty(&session
->s_caps
)) {
1338 cap
= list_entry(session
->s_caps
.next
,
1339 struct ceph_cap
, session_caps
);
1343 vino
= cap
->ci
->i_vino
;
1344 spin_unlock(&session
->s_cap_lock
);
1346 inode
= ceph_find_inode(sb
, vino
);
1349 spin_lock(&session
->s_cap_lock
);
1353 // drop cap expires and unlock s_cap_lock
1354 detach_cap_releases(session
, &dispose
);
1356 BUG_ON(session
->s_nr_caps
> 0);
1357 BUG_ON(!list_empty(&session
->s_cap_flushing
));
1358 spin_unlock(&session
->s_cap_lock
);
1359 dispose_cap_releases(session
->s_mdsc
, &dispose
);
1363 * wake up any threads waiting on this session's caps. if the cap is
1364 * old (didn't get renewed on the client reconnect), remove it now.
1366 * caller must hold s_mutex.
1368 static int wake_up_session_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1371 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1374 spin_lock(&ci
->i_ceph_lock
);
1375 ci
->i_wanted_max_size
= 0;
1376 ci
->i_requested_max_size
= 0;
1377 spin_unlock(&ci
->i_ceph_lock
);
1379 wake_up_all(&ci
->i_cap_wq
);
1383 static void wake_up_session_caps(struct ceph_mds_session
*session
,
1386 dout("wake_up_session_caps %p mds%d\n", session
, session
->s_mds
);
1387 iterate_session_caps(session
, wake_up_session_cb
,
1388 (void *)(unsigned long)reconnect
);
1392 * Send periodic message to MDS renewing all currently held caps. The
1393 * ack will reset the expiration for all caps from this session.
1395 * caller holds s_mutex
1397 static int send_renew_caps(struct ceph_mds_client
*mdsc
,
1398 struct ceph_mds_session
*session
)
1400 struct ceph_msg
*msg
;
1403 if (time_after_eq(jiffies
, session
->s_cap_ttl
) &&
1404 time_after_eq(session
->s_cap_ttl
, session
->s_renew_requested
))
1405 pr_info("mds%d caps stale\n", session
->s_mds
);
1406 session
->s_renew_requested
= jiffies
;
1408 /* do not try to renew caps until a recovering mds has reconnected
1409 * with its clients. */
1410 state
= ceph_mdsmap_get_state(mdsc
->mdsmap
, session
->s_mds
);
1411 if (state
< CEPH_MDS_STATE_RECONNECT
) {
1412 dout("send_renew_caps ignoring mds%d (%s)\n",
1413 session
->s_mds
, ceph_mds_state_name(state
));
1417 dout("send_renew_caps to mds%d (%s)\n", session
->s_mds
,
1418 ceph_mds_state_name(state
));
1419 msg
= create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS
,
1420 ++session
->s_renew_seq
);
1423 ceph_con_send(&session
->s_con
, msg
);
1427 static int send_flushmsg_ack(struct ceph_mds_client
*mdsc
,
1428 struct ceph_mds_session
*session
, u64 seq
)
1430 struct ceph_msg
*msg
;
1432 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1433 session
->s_mds
, ceph_session_state_name(session
->s_state
), seq
);
1434 msg
= create_session_msg(CEPH_SESSION_FLUSHMSG_ACK
, seq
);
1437 ceph_con_send(&session
->s_con
, msg
);
1443 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1445 * Called under session->s_mutex
1447 static void renewed_caps(struct ceph_mds_client
*mdsc
,
1448 struct ceph_mds_session
*session
, int is_renew
)
1453 spin_lock(&session
->s_cap_lock
);
1454 was_stale
= is_renew
&& time_after_eq(jiffies
, session
->s_cap_ttl
);
1456 session
->s_cap_ttl
= session
->s_renew_requested
+
1457 mdsc
->mdsmap
->m_session_timeout
*HZ
;
1460 if (time_before(jiffies
, session
->s_cap_ttl
)) {
1461 pr_info("mds%d caps renewed\n", session
->s_mds
);
1464 pr_info("mds%d caps still stale\n", session
->s_mds
);
1467 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1468 session
->s_mds
, session
->s_cap_ttl
, was_stale
? "stale" : "fresh",
1469 time_before(jiffies
, session
->s_cap_ttl
) ? "stale" : "fresh");
1470 spin_unlock(&session
->s_cap_lock
);
1473 wake_up_session_caps(session
, 0);
1477 * send a session close request
1479 static int request_close_session(struct ceph_mds_client
*mdsc
,
1480 struct ceph_mds_session
*session
)
1482 struct ceph_msg
*msg
;
1484 dout("request_close_session mds%d state %s seq %lld\n",
1485 session
->s_mds
, ceph_session_state_name(session
->s_state
),
1487 msg
= create_session_msg(CEPH_SESSION_REQUEST_CLOSE
, session
->s_seq
);
1490 ceph_con_send(&session
->s_con
, msg
);
1495 * Called with s_mutex held.
1497 static int __close_session(struct ceph_mds_client
*mdsc
,
1498 struct ceph_mds_session
*session
)
1500 if (session
->s_state
>= CEPH_MDS_SESSION_CLOSING
)
1502 session
->s_state
= CEPH_MDS_SESSION_CLOSING
;
1503 return request_close_session(mdsc
, session
);
1506 static bool drop_negative_children(struct dentry
*dentry
)
1508 struct dentry
*child
;
1509 bool all_negative
= true;
1511 if (!d_is_dir(dentry
))
1514 spin_lock(&dentry
->d_lock
);
1515 list_for_each_entry(child
, &dentry
->d_subdirs
, d_child
) {
1516 if (d_really_is_positive(child
)) {
1517 all_negative
= false;
1521 spin_unlock(&dentry
->d_lock
);
1524 shrink_dcache_parent(dentry
);
1526 return all_negative
;
1530 * Trim old(er) caps.
1532 * Because we can't cache an inode without one or more caps, we do
1533 * this indirectly: if a cap is unused, we prune its aliases, at which
1534 * point the inode will hopefully get dropped to.
1536 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1537 * memory pressure from the MDS, though, so it needn't be perfect.
1539 static int trim_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
, void *arg
)
1541 struct ceph_mds_session
*session
= arg
;
1542 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1543 int used
, wanted
, oissued
, mine
;
1545 if (session
->s_trim_caps
<= 0)
1548 spin_lock(&ci
->i_ceph_lock
);
1549 mine
= cap
->issued
| cap
->implemented
;
1550 used
= __ceph_caps_used(ci
);
1551 wanted
= __ceph_caps_file_wanted(ci
);
1552 oissued
= __ceph_caps_issued_other(ci
, cap
);
1554 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1555 inode
, cap
, ceph_cap_string(mine
), ceph_cap_string(oissued
),
1556 ceph_cap_string(used
), ceph_cap_string(wanted
));
1557 if (cap
== ci
->i_auth_cap
) {
1558 if (ci
->i_dirty_caps
|| ci
->i_flushing_caps
||
1559 !list_empty(&ci
->i_cap_snaps
))
1561 if ((used
| wanted
) & CEPH_CAP_ANY_WR
)
1563 /* Note: it's possible that i_filelock_ref becomes non-zero
1564 * after dropping auth caps. It doesn't hurt because reply
1565 * of lock mds request will re-add auth caps. */
1566 if (atomic_read(&ci
->i_filelock_ref
) > 0)
1569 /* The inode has cached pages, but it's no longer used.
1570 * we can safely drop it */
1571 if (wanted
== 0 && used
== CEPH_CAP_FILE_CACHE
&&
1572 !(oissued
& CEPH_CAP_FILE_CACHE
)) {
1576 if ((used
| wanted
) & ~oissued
& mine
)
1577 goto out
; /* we need these caps */
1580 /* we aren't the only cap.. just remove us */
1581 __ceph_remove_cap(cap
, true);
1582 session
->s_trim_caps
--;
1584 struct dentry
*dentry
;
1585 /* try dropping referring dentries */
1586 spin_unlock(&ci
->i_ceph_lock
);
1587 dentry
= d_find_any_alias(inode
);
1588 if (dentry
&& drop_negative_children(dentry
)) {
1591 d_prune_aliases(inode
);
1592 count
= atomic_read(&inode
->i_count
);
1594 session
->s_trim_caps
--;
1595 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1604 spin_unlock(&ci
->i_ceph_lock
);
1609 * Trim session cap count down to some max number.
1611 int ceph_trim_caps(struct ceph_mds_client
*mdsc
,
1612 struct ceph_mds_session
*session
,
1615 int trim_caps
= session
->s_nr_caps
- max_caps
;
1617 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1618 session
->s_mds
, session
->s_nr_caps
, max_caps
, trim_caps
);
1619 if (trim_caps
> 0) {
1620 session
->s_trim_caps
= trim_caps
;
1621 iterate_session_caps(session
, trim_caps_cb
, session
);
1622 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1623 session
->s_mds
, session
->s_nr_caps
, max_caps
,
1624 trim_caps
- session
->s_trim_caps
);
1625 session
->s_trim_caps
= 0;
1628 ceph_send_cap_releases(mdsc
, session
);
1632 static int check_caps_flush(struct ceph_mds_client
*mdsc
,
1637 spin_lock(&mdsc
->cap_dirty_lock
);
1638 if (!list_empty(&mdsc
->cap_flush_list
)) {
1639 struct ceph_cap_flush
*cf
=
1640 list_first_entry(&mdsc
->cap_flush_list
,
1641 struct ceph_cap_flush
, g_list
);
1642 if (cf
->tid
<= want_flush_tid
) {
1643 dout("check_caps_flush still flushing tid "
1644 "%llu <= %llu\n", cf
->tid
, want_flush_tid
);
1648 spin_unlock(&mdsc
->cap_dirty_lock
);
1653 * flush all dirty inode data to disk.
1655 * returns true if we've flushed through want_flush_tid
1657 static void wait_caps_flush(struct ceph_mds_client
*mdsc
,
1660 dout("check_caps_flush want %llu\n", want_flush_tid
);
1662 wait_event(mdsc
->cap_flushing_wq
,
1663 check_caps_flush(mdsc
, want_flush_tid
));
1665 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid
);
1669 * called under s_mutex
1671 void ceph_send_cap_releases(struct ceph_mds_client
*mdsc
,
1672 struct ceph_mds_session
*session
)
1674 struct ceph_msg
*msg
= NULL
;
1675 struct ceph_mds_cap_release
*head
;
1676 struct ceph_mds_cap_item
*item
;
1677 struct ceph_osd_client
*osdc
= &mdsc
->fsc
->client
->osdc
;
1678 struct ceph_cap
*cap
;
1679 LIST_HEAD(tmp_list
);
1680 int num_cap_releases
;
1681 __le32 barrier
, *cap_barrier
;
1683 down_read(&osdc
->lock
);
1684 barrier
= cpu_to_le32(osdc
->epoch_barrier
);
1685 up_read(&osdc
->lock
);
1687 spin_lock(&session
->s_cap_lock
);
1689 list_splice_init(&session
->s_cap_releases
, &tmp_list
);
1690 num_cap_releases
= session
->s_num_cap_releases
;
1691 session
->s_num_cap_releases
= 0;
1692 spin_unlock(&session
->s_cap_lock
);
1694 while (!list_empty(&tmp_list
)) {
1696 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE
,
1697 PAGE_SIZE
, GFP_NOFS
, false);
1700 head
= msg
->front
.iov_base
;
1701 head
->num
= cpu_to_le32(0);
1702 msg
->front
.iov_len
= sizeof(*head
);
1704 msg
->hdr
.version
= cpu_to_le16(2);
1705 msg
->hdr
.compat_version
= cpu_to_le16(1);
1708 cap
= list_first_entry(&tmp_list
, struct ceph_cap
,
1710 list_del(&cap
->session_caps
);
1713 head
= msg
->front
.iov_base
;
1714 le32_add_cpu(&head
->num
, 1);
1715 item
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1716 item
->ino
= cpu_to_le64(cap
->cap_ino
);
1717 item
->cap_id
= cpu_to_le64(cap
->cap_id
);
1718 item
->migrate_seq
= cpu_to_le32(cap
->mseq
);
1719 item
->seq
= cpu_to_le32(cap
->issue_seq
);
1720 msg
->front
.iov_len
+= sizeof(*item
);
1722 ceph_put_cap(mdsc
, cap
);
1724 if (le32_to_cpu(head
->num
) == CEPH_CAPS_PER_RELEASE
) {
1725 // Append cap_barrier field
1726 cap_barrier
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1727 *cap_barrier
= barrier
;
1728 msg
->front
.iov_len
+= sizeof(*cap_barrier
);
1730 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1731 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1732 ceph_con_send(&session
->s_con
, msg
);
1737 BUG_ON(num_cap_releases
!= 0);
1739 spin_lock(&session
->s_cap_lock
);
1740 if (!list_empty(&session
->s_cap_releases
))
1742 spin_unlock(&session
->s_cap_lock
);
1745 // Append cap_barrier field
1746 cap_barrier
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1747 *cap_barrier
= barrier
;
1748 msg
->front
.iov_len
+= sizeof(*cap_barrier
);
1750 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1751 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1752 ceph_con_send(&session
->s_con
, msg
);
1756 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1758 spin_lock(&session
->s_cap_lock
);
1759 list_splice(&tmp_list
, &session
->s_cap_releases
);
1760 session
->s_num_cap_releases
+= num_cap_releases
;
1761 spin_unlock(&session
->s_cap_lock
);
1768 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request
*req
,
1771 struct ceph_inode_info
*ci
= ceph_inode(dir
);
1772 struct ceph_mds_reply_info_parsed
*rinfo
= &req
->r_reply_info
;
1773 struct ceph_mount_options
*opt
= req
->r_mdsc
->fsc
->mount_options
;
1774 size_t size
= sizeof(struct ceph_mds_reply_dir_entry
);
1775 int order
, num_entries
;
1777 spin_lock(&ci
->i_ceph_lock
);
1778 num_entries
= ci
->i_files
+ ci
->i_subdirs
;
1779 spin_unlock(&ci
->i_ceph_lock
);
1780 num_entries
= max(num_entries
, 1);
1781 num_entries
= min(num_entries
, opt
->max_readdir
);
1783 order
= get_order(size
* num_entries
);
1784 while (order
>= 0) {
1785 rinfo
->dir_entries
= (void*)__get_free_pages(GFP_KERNEL
|
1788 if (rinfo
->dir_entries
)
1792 if (!rinfo
->dir_entries
)
1795 num_entries
= (PAGE_SIZE
<< order
) / size
;
1796 num_entries
= min(num_entries
, opt
->max_readdir
);
1798 rinfo
->dir_buf_size
= PAGE_SIZE
<< order
;
1799 req
->r_num_caps
= num_entries
+ 1;
1800 req
->r_args
.readdir
.max_entries
= cpu_to_le32(num_entries
);
1801 req
->r_args
.readdir
.max_bytes
= cpu_to_le32(opt
->max_readdir_bytes
);
1806 * Create an mds request.
1808 struct ceph_mds_request
*
1809 ceph_mdsc_create_request(struct ceph_mds_client
*mdsc
, int op
, int mode
)
1811 struct ceph_mds_request
*req
= kzalloc(sizeof(*req
), GFP_NOFS
);
1812 struct timespec64 ts
;
1815 return ERR_PTR(-ENOMEM
);
1817 mutex_init(&req
->r_fill_mutex
);
1819 req
->r_started
= jiffies
;
1820 req
->r_resend_mds
= -1;
1821 INIT_LIST_HEAD(&req
->r_unsafe_dir_item
);
1822 INIT_LIST_HEAD(&req
->r_unsafe_target_item
);
1824 kref_init(&req
->r_kref
);
1825 RB_CLEAR_NODE(&req
->r_node
);
1826 INIT_LIST_HEAD(&req
->r_wait
);
1827 init_completion(&req
->r_completion
);
1828 init_completion(&req
->r_safe_completion
);
1829 INIT_LIST_HEAD(&req
->r_unsafe_item
);
1831 ktime_get_coarse_real_ts64(&ts
);
1832 req
->r_stamp
= timespec64_trunc(ts
, mdsc
->fsc
->sb
->s_time_gran
);
1835 req
->r_direct_mode
= mode
;
1840 * return oldest (lowest) request, tid in request tree, 0 if none.
1842 * called under mdsc->mutex.
1844 static struct ceph_mds_request
*__get_oldest_req(struct ceph_mds_client
*mdsc
)
1846 if (RB_EMPTY_ROOT(&mdsc
->request_tree
))
1848 return rb_entry(rb_first(&mdsc
->request_tree
),
1849 struct ceph_mds_request
, r_node
);
1852 static inline u64
__get_oldest_tid(struct ceph_mds_client
*mdsc
)
1854 return mdsc
->oldest_tid
;
1858 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1859 * on build_path_from_dentry in fs/cifs/dir.c.
1861 * If @stop_on_nosnap, generate path relative to the first non-snapped
1864 * Encode hidden .snap dirs as a double /, i.e.
1865 * foo/.snap/bar -> foo//bar
1867 char *ceph_mdsc_build_path(struct dentry
*dentry
, int *plen
, u64
*base
,
1870 struct dentry
*temp
;
1876 return ERR_PTR(-EINVAL
);
1880 seq
= read_seqbegin(&rename_lock
);
1882 for (temp
= dentry
; !IS_ROOT(temp
);) {
1883 struct inode
*inode
= d_inode(temp
);
1884 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
)
1885 len
++; /* slash only */
1886 else if (stop_on_nosnap
&& inode
&&
1887 ceph_snap(inode
) == CEPH_NOSNAP
)
1890 len
+= 1 + temp
->d_name
.len
;
1891 temp
= temp
->d_parent
;
1895 len
--; /* no leading '/' */
1897 path
= kmalloc(len
+1, GFP_NOFS
);
1899 return ERR_PTR(-ENOMEM
);
1901 path
[pos
] = 0; /* trailing null */
1903 for (temp
= dentry
; !IS_ROOT(temp
) && pos
!= 0; ) {
1904 struct inode
*inode
;
1906 spin_lock(&temp
->d_lock
);
1907 inode
= d_inode(temp
);
1908 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
) {
1909 dout("build_path path+%d: %p SNAPDIR\n",
1911 } else if (stop_on_nosnap
&& inode
&&
1912 ceph_snap(inode
) == CEPH_NOSNAP
) {
1913 spin_unlock(&temp
->d_lock
);
1916 pos
-= temp
->d_name
.len
;
1918 spin_unlock(&temp
->d_lock
);
1921 strncpy(path
+ pos
, temp
->d_name
.name
,
1924 spin_unlock(&temp
->d_lock
);
1927 temp
= temp
->d_parent
;
1930 if (pos
!= 0 || read_seqretry(&rename_lock
, seq
)) {
1931 pr_err("build_path did not end path lookup where "
1932 "expected, namelen is %d, pos is %d\n", len
, pos
);
1933 /* presumably this is only possible if racing with a
1934 rename of one of the parent directories (we can not
1935 lock the dentries above us to prevent this, but
1936 retrying should be harmless) */
1941 *base
= ceph_ino(d_inode(temp
));
1943 dout("build_path on %p %d built %llx '%.*s'\n",
1944 dentry
, d_count(dentry
), *base
, len
, path
);
1948 static int build_dentry_path(struct dentry
*dentry
, struct inode
*dir
,
1949 const char **ppath
, int *ppathlen
, u64
*pino
,
1956 dir
= d_inode_rcu(dentry
->d_parent
);
1957 if (dir
&& ceph_snap(dir
) == CEPH_NOSNAP
) {
1958 *pino
= ceph_ino(dir
);
1960 *ppath
= dentry
->d_name
.name
;
1961 *ppathlen
= dentry
->d_name
.len
;
1965 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1967 return PTR_ERR(path
);
1973 static int build_inode_path(struct inode
*inode
,
1974 const char **ppath
, int *ppathlen
, u64
*pino
,
1977 struct dentry
*dentry
;
1980 if (ceph_snap(inode
) == CEPH_NOSNAP
) {
1981 *pino
= ceph_ino(inode
);
1985 dentry
= d_find_alias(inode
);
1986 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1989 return PTR_ERR(path
);
1996 * request arguments may be specified via an inode *, a dentry *, or
1997 * an explicit ino+path.
1999 static int set_request_path_attr(struct inode
*rinode
, struct dentry
*rdentry
,
2000 struct inode
*rdiri
, const char *rpath
,
2001 u64 rino
, const char **ppath
, int *pathlen
,
2002 u64
*ino
, int *freepath
)
2007 r
= build_inode_path(rinode
, ppath
, pathlen
, ino
, freepath
);
2008 dout(" inode %p %llx.%llx\n", rinode
, ceph_ino(rinode
),
2010 } else if (rdentry
) {
2011 r
= build_dentry_path(rdentry
, rdiri
, ppath
, pathlen
, ino
,
2013 dout(" dentry %p %llx/%.*s\n", rdentry
, *ino
, *pathlen
,
2015 } else if (rpath
|| rino
) {
2018 *pathlen
= rpath
? strlen(rpath
) : 0;
2019 dout(" path %.*s\n", *pathlen
, rpath
);
2026 * called under mdsc->mutex
2028 static struct ceph_msg
*create_request_message(struct ceph_mds_client
*mdsc
,
2029 struct ceph_mds_request
*req
,
2030 int mds
, bool drop_cap_releases
)
2032 struct ceph_msg
*msg
;
2033 struct ceph_mds_request_head
*head
;
2034 const char *path1
= NULL
;
2035 const char *path2
= NULL
;
2036 u64 ino1
= 0, ino2
= 0;
2037 int pathlen1
= 0, pathlen2
= 0;
2038 int freepath1
= 0, freepath2
= 0;
2044 ret
= set_request_path_attr(req
->r_inode
, req
->r_dentry
,
2045 req
->r_parent
, req
->r_path1
, req
->r_ino1
.ino
,
2046 &path1
, &pathlen1
, &ino1
, &freepath1
);
2052 ret
= set_request_path_attr(NULL
, req
->r_old_dentry
,
2053 req
->r_old_dentry_dir
,
2054 req
->r_path2
, req
->r_ino2
.ino
,
2055 &path2
, &pathlen2
, &ino2
, &freepath2
);
2061 len
= sizeof(*head
) +
2062 pathlen1
+ pathlen2
+ 2*(1 + sizeof(u32
) + sizeof(u64
)) +
2063 sizeof(struct ceph_timespec
);
2065 /* calculate (max) length for cap releases */
2066 len
+= sizeof(struct ceph_mds_request_release
) *
2067 (!!req
->r_inode_drop
+ !!req
->r_dentry_drop
+
2068 !!req
->r_old_inode_drop
+ !!req
->r_old_dentry_drop
);
2069 if (req
->r_dentry_drop
)
2070 len
+= req
->r_dentry
->d_name
.len
;
2071 if (req
->r_old_dentry_drop
)
2072 len
+= req
->r_old_dentry
->d_name
.len
;
2074 msg
= ceph_msg_new(CEPH_MSG_CLIENT_REQUEST
, len
, GFP_NOFS
, false);
2076 msg
= ERR_PTR(-ENOMEM
);
2080 msg
->hdr
.version
= cpu_to_le16(2);
2081 msg
->hdr
.tid
= cpu_to_le64(req
->r_tid
);
2083 head
= msg
->front
.iov_base
;
2084 p
= msg
->front
.iov_base
+ sizeof(*head
);
2085 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
2087 head
->mdsmap_epoch
= cpu_to_le32(mdsc
->mdsmap
->m_epoch
);
2088 head
->op
= cpu_to_le32(req
->r_op
);
2089 head
->caller_uid
= cpu_to_le32(from_kuid(&init_user_ns
, req
->r_uid
));
2090 head
->caller_gid
= cpu_to_le32(from_kgid(&init_user_ns
, req
->r_gid
));
2091 head
->args
= req
->r_args
;
2093 ceph_encode_filepath(&p
, end
, ino1
, path1
);
2094 ceph_encode_filepath(&p
, end
, ino2
, path2
);
2096 /* make note of release offset, in case we need to replay */
2097 req
->r_request_release_offset
= p
- msg
->front
.iov_base
;
2101 if (req
->r_inode_drop
)
2102 releases
+= ceph_encode_inode_release(&p
,
2103 req
->r_inode
? req
->r_inode
: d_inode(req
->r_dentry
),
2104 mds
, req
->r_inode_drop
, req
->r_inode_unless
, 0);
2105 if (req
->r_dentry_drop
)
2106 releases
+= ceph_encode_dentry_release(&p
, req
->r_dentry
,
2107 req
->r_parent
, mds
, req
->r_dentry_drop
,
2108 req
->r_dentry_unless
);
2109 if (req
->r_old_dentry_drop
)
2110 releases
+= ceph_encode_dentry_release(&p
, req
->r_old_dentry
,
2111 req
->r_old_dentry_dir
, mds
,
2112 req
->r_old_dentry_drop
,
2113 req
->r_old_dentry_unless
);
2114 if (req
->r_old_inode_drop
)
2115 releases
+= ceph_encode_inode_release(&p
,
2116 d_inode(req
->r_old_dentry
),
2117 mds
, req
->r_old_inode_drop
, req
->r_old_inode_unless
, 0);
2119 if (drop_cap_releases
) {
2121 p
= msg
->front
.iov_base
+ req
->r_request_release_offset
;
2124 head
->num_releases
= cpu_to_le16(releases
);
2128 struct ceph_timespec ts
;
2129 ceph_encode_timespec64(&ts
, &req
->r_stamp
);
2130 ceph_encode_copy(&p
, &ts
, sizeof(ts
));
2134 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
2135 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
2137 if (req
->r_pagelist
) {
2138 struct ceph_pagelist
*pagelist
= req
->r_pagelist
;
2139 refcount_inc(&pagelist
->refcnt
);
2140 ceph_msg_data_add_pagelist(msg
, pagelist
);
2141 msg
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
2143 msg
->hdr
.data_len
= 0;
2146 msg
->hdr
.data_off
= cpu_to_le16(0);
2150 kfree((char *)path2
);
2153 kfree((char *)path1
);
2159 * called under mdsc->mutex if error, under no mutex if
2162 static void complete_request(struct ceph_mds_client
*mdsc
,
2163 struct ceph_mds_request
*req
)
2165 if (req
->r_callback
)
2166 req
->r_callback(mdsc
, req
);
2168 complete_all(&req
->r_completion
);
2172 * called under mdsc->mutex
2174 static int __prepare_send_request(struct ceph_mds_client
*mdsc
,
2175 struct ceph_mds_request
*req
,
2176 int mds
, bool drop_cap_releases
)
2178 struct ceph_mds_request_head
*rhead
;
2179 struct ceph_msg
*msg
;
2184 struct ceph_cap
*cap
=
2185 ceph_get_cap_for_mds(ceph_inode(req
->r_inode
), mds
);
2188 req
->r_sent_on_mseq
= cap
->mseq
;
2190 req
->r_sent_on_mseq
= -1;
2192 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req
,
2193 req
->r_tid
, ceph_mds_op_name(req
->r_op
), req
->r_attempts
);
2195 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
)) {
2198 * Replay. Do not regenerate message (and rebuild
2199 * paths, etc.); just use the original message.
2200 * Rebuilding paths will break for renames because
2201 * d_move mangles the src name.
2203 msg
= req
->r_request
;
2204 rhead
= msg
->front
.iov_base
;
2206 flags
= le32_to_cpu(rhead
->flags
);
2207 flags
|= CEPH_MDS_FLAG_REPLAY
;
2208 rhead
->flags
= cpu_to_le32(flags
);
2210 if (req
->r_target_inode
)
2211 rhead
->ino
= cpu_to_le64(ceph_ino(req
->r_target_inode
));
2213 rhead
->num_retry
= req
->r_attempts
- 1;
2215 /* remove cap/dentry releases from message */
2216 rhead
->num_releases
= 0;
2219 p
= msg
->front
.iov_base
+ req
->r_request_release_offset
;
2221 struct ceph_timespec ts
;
2222 ceph_encode_timespec64(&ts
, &req
->r_stamp
);
2223 ceph_encode_copy(&p
, &ts
, sizeof(ts
));
2226 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
2227 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
2231 if (req
->r_request
) {
2232 ceph_msg_put(req
->r_request
);
2233 req
->r_request
= NULL
;
2235 msg
= create_request_message(mdsc
, req
, mds
, drop_cap_releases
);
2237 req
->r_err
= PTR_ERR(msg
);
2238 return PTR_ERR(msg
);
2240 req
->r_request
= msg
;
2242 rhead
= msg
->front
.iov_base
;
2243 rhead
->oldest_client_tid
= cpu_to_le64(__get_oldest_tid(mdsc
));
2244 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
))
2245 flags
|= CEPH_MDS_FLAG_REPLAY
;
2247 flags
|= CEPH_MDS_FLAG_WANT_DENTRY
;
2248 rhead
->flags
= cpu_to_le32(flags
);
2249 rhead
->num_fwd
= req
->r_num_fwd
;
2250 rhead
->num_retry
= req
->r_attempts
- 1;
2253 dout(" r_parent = %p\n", req
->r_parent
);
2258 * send request, or put it on the appropriate wait list.
2260 static void __do_request(struct ceph_mds_client
*mdsc
,
2261 struct ceph_mds_request
*req
)
2263 struct ceph_mds_session
*session
= NULL
;
2267 if (req
->r_err
|| test_bit(CEPH_MDS_R_GOT_RESULT
, &req
->r_req_flags
)) {
2268 if (test_bit(CEPH_MDS_R_ABORTED
, &req
->r_req_flags
))
2269 __unregister_request(mdsc
, req
);
2273 if (req
->r_timeout
&&
2274 time_after_eq(jiffies
, req
->r_started
+ req
->r_timeout
)) {
2275 dout("do_request timed out\n");
2279 if (READ_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
) {
2280 dout("do_request forced umount\n");
2284 if (READ_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_MOUNTING
) {
2285 if (mdsc
->mdsmap_err
) {
2286 err
= mdsc
->mdsmap_err
;
2287 dout("do_request mdsmap err %d\n", err
);
2290 if (mdsc
->mdsmap
->m_epoch
== 0) {
2291 dout("do_request no mdsmap, waiting for map\n");
2292 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
2295 if (!(mdsc
->fsc
->mount_options
->flags
&
2296 CEPH_MOUNT_OPT_MOUNTWAIT
) &&
2297 !ceph_mdsmap_is_cluster_available(mdsc
->mdsmap
)) {
2299 pr_info("probably no mds server is up\n");
2304 put_request_session(req
);
2306 mds
= __choose_mds(mdsc
, req
);
2308 ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) < CEPH_MDS_STATE_ACTIVE
) {
2309 dout("do_request no mds or not active, waiting for map\n");
2310 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
2314 /* get, open session */
2315 session
= __ceph_lookup_mds_session(mdsc
, mds
);
2317 session
= register_session(mdsc
, mds
);
2318 if (IS_ERR(session
)) {
2319 err
= PTR_ERR(session
);
2323 req
->r_session
= get_session(session
);
2325 dout("do_request mds%d session %p state %s\n", mds
, session
,
2326 ceph_session_state_name(session
->s_state
));
2327 if (session
->s_state
!= CEPH_MDS_SESSION_OPEN
&&
2328 session
->s_state
!= CEPH_MDS_SESSION_HUNG
) {
2329 if (session
->s_state
== CEPH_MDS_SESSION_REJECTED
) {
2333 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
2334 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
2335 __open_session(mdsc
, session
);
2336 list_add(&req
->r_wait
, &session
->s_waiting
);
2341 req
->r_resend_mds
= -1; /* forget any previous mds hint */
2343 if (req
->r_request_started
== 0) /* note request start time */
2344 req
->r_request_started
= jiffies
;
2346 err
= __prepare_send_request(mdsc
, req
, mds
, false);
2348 ceph_msg_get(req
->r_request
);
2349 ceph_con_send(&session
->s_con
, req
->r_request
);
2353 ceph_put_mds_session(session
);
2356 dout("__do_request early error %d\n", err
);
2358 complete_request(mdsc
, req
);
2359 __unregister_request(mdsc
, req
);
2365 * called under mdsc->mutex
2367 static void __wake_requests(struct ceph_mds_client
*mdsc
,
2368 struct list_head
*head
)
2370 struct ceph_mds_request
*req
;
2371 LIST_HEAD(tmp_list
);
2373 list_splice_init(head
, &tmp_list
);
2375 while (!list_empty(&tmp_list
)) {
2376 req
= list_entry(tmp_list
.next
,
2377 struct ceph_mds_request
, r_wait
);
2378 list_del_init(&req
->r_wait
);
2379 dout(" wake request %p tid %llu\n", req
, req
->r_tid
);
2380 __do_request(mdsc
, req
);
2385 * Wake up threads with requests pending for @mds, so that they can
2386 * resubmit their requests to a possibly different mds.
2388 static void kick_requests(struct ceph_mds_client
*mdsc
, int mds
)
2390 struct ceph_mds_request
*req
;
2391 struct rb_node
*p
= rb_first(&mdsc
->request_tree
);
2393 dout("kick_requests mds%d\n", mds
);
2395 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2397 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
))
2399 if (req
->r_attempts
> 0)
2400 continue; /* only new requests */
2401 if (req
->r_session
&&
2402 req
->r_session
->s_mds
== mds
) {
2403 dout(" kicking tid %llu\n", req
->r_tid
);
2404 list_del_init(&req
->r_wait
);
2405 __do_request(mdsc
, req
);
2410 void ceph_mdsc_submit_request(struct ceph_mds_client
*mdsc
,
2411 struct ceph_mds_request
*req
)
2413 dout("submit_request on %p\n", req
);
2414 mutex_lock(&mdsc
->mutex
);
2415 __register_request(mdsc
, req
, NULL
);
2416 __do_request(mdsc
, req
);
2417 mutex_unlock(&mdsc
->mutex
);
2421 * Synchrously perform an mds request. Take care of all of the
2422 * session setup, forwarding, retry details.
2424 int ceph_mdsc_do_request(struct ceph_mds_client
*mdsc
,
2426 struct ceph_mds_request
*req
)
2430 dout("do_request on %p\n", req
);
2432 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2434 ceph_get_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
2436 ceph_get_cap_refs(ceph_inode(req
->r_parent
), CEPH_CAP_PIN
);
2437 if (req
->r_old_dentry_dir
)
2438 ceph_get_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
2442 mutex_lock(&mdsc
->mutex
);
2443 __register_request(mdsc
, req
, dir
);
2444 __do_request(mdsc
, req
);
2452 mutex_unlock(&mdsc
->mutex
);
2453 dout("do_request waiting\n");
2454 if (!req
->r_timeout
&& req
->r_wait_for_completion
) {
2455 err
= req
->r_wait_for_completion(mdsc
, req
);
2457 long timeleft
= wait_for_completion_killable_timeout(
2459 ceph_timeout_jiffies(req
->r_timeout
));
2463 err
= -EIO
; /* timed out */
2465 err
= timeleft
; /* killed */
2467 dout("do_request waited, got %d\n", err
);
2468 mutex_lock(&mdsc
->mutex
);
2470 /* only abort if we didn't race with a real reply */
2471 if (test_bit(CEPH_MDS_R_GOT_RESULT
, &req
->r_req_flags
)) {
2472 err
= le32_to_cpu(req
->r_reply_info
.head
->result
);
2473 } else if (err
< 0) {
2474 dout("aborted request %lld with %d\n", req
->r_tid
, err
);
2477 * ensure we aren't running concurrently with
2478 * ceph_fill_trace or ceph_readdir_prepopulate, which
2479 * rely on locks (dir mutex) held by our caller.
2481 mutex_lock(&req
->r_fill_mutex
);
2483 set_bit(CEPH_MDS_R_ABORTED
, &req
->r_req_flags
);
2484 mutex_unlock(&req
->r_fill_mutex
);
2486 if (req
->r_parent
&&
2487 (req
->r_op
& CEPH_MDS_OP_WRITE
))
2488 ceph_invalidate_dir_request(req
);
2494 mutex_unlock(&mdsc
->mutex
);
2495 dout("do_request %p done, result %d\n", req
, err
);
2500 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2501 * namespace request.
2503 void ceph_invalidate_dir_request(struct ceph_mds_request
*req
)
2505 struct inode
*dir
= req
->r_parent
;
2506 struct inode
*old_dir
= req
->r_old_dentry_dir
;
2508 dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir
, old_dir
);
2510 ceph_dir_clear_complete(dir
);
2512 ceph_dir_clear_complete(old_dir
);
2514 ceph_invalidate_dentry_lease(req
->r_dentry
);
2515 if (req
->r_old_dentry
)
2516 ceph_invalidate_dentry_lease(req
->r_old_dentry
);
2522 * We take the session mutex and parse and process the reply immediately.
2523 * This preserves the logical ordering of replies, capabilities, etc., sent
2524 * by the MDS as they are applied to our local cache.
2526 static void handle_reply(struct ceph_mds_session
*session
, struct ceph_msg
*msg
)
2528 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2529 struct ceph_mds_request
*req
;
2530 struct ceph_mds_reply_head
*head
= msg
->front
.iov_base
;
2531 struct ceph_mds_reply_info_parsed
*rinfo
; /* parsed reply info */
2532 struct ceph_snap_realm
*realm
;
2535 int mds
= session
->s_mds
;
2537 if (msg
->front
.iov_len
< sizeof(*head
)) {
2538 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2543 /* get request, session */
2544 tid
= le64_to_cpu(msg
->hdr
.tid
);
2545 mutex_lock(&mdsc
->mutex
);
2546 req
= lookup_get_request(mdsc
, tid
);
2548 dout("handle_reply on unknown tid %llu\n", tid
);
2549 mutex_unlock(&mdsc
->mutex
);
2552 dout("handle_reply %p\n", req
);
2554 /* correct session? */
2555 if (req
->r_session
!= session
) {
2556 pr_err("mdsc_handle_reply got %llu on session mds%d"
2557 " not mds%d\n", tid
, session
->s_mds
,
2558 req
->r_session
? req
->r_session
->s_mds
: -1);
2559 mutex_unlock(&mdsc
->mutex
);
2564 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
) && !head
->safe
) ||
2565 (test_bit(CEPH_MDS_R_GOT_SAFE
, &req
->r_req_flags
) && head
->safe
)) {
2566 pr_warn("got a dup %s reply on %llu from mds%d\n",
2567 head
->safe
? "safe" : "unsafe", tid
, mds
);
2568 mutex_unlock(&mdsc
->mutex
);
2571 if (test_bit(CEPH_MDS_R_GOT_SAFE
, &req
->r_req_flags
)) {
2572 pr_warn("got unsafe after safe on %llu from mds%d\n",
2574 mutex_unlock(&mdsc
->mutex
);
2578 result
= le32_to_cpu(head
->result
);
2582 * if we're not talking to the authority, send to them
2583 * if the authority has changed while we weren't looking,
2584 * send to new authority
2585 * Otherwise we just have to return an ESTALE
2587 if (result
== -ESTALE
) {
2588 dout("got ESTALE on request %llu\n", req
->r_tid
);
2589 req
->r_resend_mds
= -1;
2590 if (req
->r_direct_mode
!= USE_AUTH_MDS
) {
2591 dout("not using auth, setting for that now\n");
2592 req
->r_direct_mode
= USE_AUTH_MDS
;
2593 __do_request(mdsc
, req
);
2594 mutex_unlock(&mdsc
->mutex
);
2597 int mds
= __choose_mds(mdsc
, req
);
2598 if (mds
>= 0 && mds
!= req
->r_session
->s_mds
) {
2599 dout("but auth changed, so resending\n");
2600 __do_request(mdsc
, req
);
2601 mutex_unlock(&mdsc
->mutex
);
2605 dout("have to return ESTALE on request %llu\n", req
->r_tid
);
2610 set_bit(CEPH_MDS_R_GOT_SAFE
, &req
->r_req_flags
);
2611 __unregister_request(mdsc
, req
);
2613 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
)) {
2615 * We already handled the unsafe response, now do the
2616 * cleanup. No need to examine the response; the MDS
2617 * doesn't include any result info in the safe
2618 * response. And even if it did, there is nothing
2619 * useful we could do with a revised return value.
2621 dout("got safe reply %llu, mds%d\n", tid
, mds
);
2623 /* last unsafe request during umount? */
2624 if (mdsc
->stopping
&& !__get_oldest_req(mdsc
))
2625 complete_all(&mdsc
->safe_umount_waiters
);
2626 mutex_unlock(&mdsc
->mutex
);
2630 set_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
);
2631 list_add_tail(&req
->r_unsafe_item
, &req
->r_session
->s_unsafe
);
2632 if (req
->r_unsafe_dir
) {
2633 struct ceph_inode_info
*ci
=
2634 ceph_inode(req
->r_unsafe_dir
);
2635 spin_lock(&ci
->i_unsafe_lock
);
2636 list_add_tail(&req
->r_unsafe_dir_item
,
2637 &ci
->i_unsafe_dirops
);
2638 spin_unlock(&ci
->i_unsafe_lock
);
2642 dout("handle_reply tid %lld result %d\n", tid
, result
);
2643 rinfo
= &req
->r_reply_info
;
2644 err
= parse_reply_info(msg
, rinfo
, session
->s_con
.peer_features
);
2645 mutex_unlock(&mdsc
->mutex
);
2647 mutex_lock(&session
->s_mutex
);
2649 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds
, tid
);
2656 if (rinfo
->snapblob_len
) {
2657 down_write(&mdsc
->snap_rwsem
);
2658 ceph_update_snap_trace(mdsc
, rinfo
->snapblob
,
2659 rinfo
->snapblob
+ rinfo
->snapblob_len
,
2660 le32_to_cpu(head
->op
) == CEPH_MDS_OP_RMSNAP
,
2662 downgrade_write(&mdsc
->snap_rwsem
);
2664 down_read(&mdsc
->snap_rwsem
);
2667 /* insert trace into our cache */
2668 mutex_lock(&req
->r_fill_mutex
);
2669 current
->journal_info
= req
;
2670 err
= ceph_fill_trace(mdsc
->fsc
->sb
, req
);
2672 if (result
== 0 && (req
->r_op
== CEPH_MDS_OP_READDIR
||
2673 req
->r_op
== CEPH_MDS_OP_LSSNAP
))
2674 ceph_readdir_prepopulate(req
, req
->r_session
);
2675 ceph_unreserve_caps(mdsc
, &req
->r_caps_reservation
);
2677 current
->journal_info
= NULL
;
2678 mutex_unlock(&req
->r_fill_mutex
);
2680 up_read(&mdsc
->snap_rwsem
);
2682 ceph_put_snap_realm(mdsc
, realm
);
2684 if (err
== 0 && req
->r_target_inode
&&
2685 test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
)) {
2686 struct ceph_inode_info
*ci
= ceph_inode(req
->r_target_inode
);
2687 spin_lock(&ci
->i_unsafe_lock
);
2688 list_add_tail(&req
->r_unsafe_target_item
, &ci
->i_unsafe_iops
);
2689 spin_unlock(&ci
->i_unsafe_lock
);
2692 mutex_lock(&mdsc
->mutex
);
2693 if (!test_bit(CEPH_MDS_R_ABORTED
, &req
->r_req_flags
)) {
2697 req
->r_reply
= ceph_msg_get(msg
);
2698 set_bit(CEPH_MDS_R_GOT_RESULT
, &req
->r_req_flags
);
2701 dout("reply arrived after request %lld was aborted\n", tid
);
2703 mutex_unlock(&mdsc
->mutex
);
2705 mutex_unlock(&session
->s_mutex
);
2707 /* kick calling process */
2708 complete_request(mdsc
, req
);
2710 ceph_mdsc_put_request(req
);
2717 * handle mds notification that our request has been forwarded.
2719 static void handle_forward(struct ceph_mds_client
*mdsc
,
2720 struct ceph_mds_session
*session
,
2721 struct ceph_msg
*msg
)
2723 struct ceph_mds_request
*req
;
2724 u64 tid
= le64_to_cpu(msg
->hdr
.tid
);
2728 void *p
= msg
->front
.iov_base
;
2729 void *end
= p
+ msg
->front
.iov_len
;
2731 ceph_decode_need(&p
, end
, 2*sizeof(u32
), bad
);
2732 next_mds
= ceph_decode_32(&p
);
2733 fwd_seq
= ceph_decode_32(&p
);
2735 mutex_lock(&mdsc
->mutex
);
2736 req
= lookup_get_request(mdsc
, tid
);
2738 dout("forward tid %llu to mds%d - req dne\n", tid
, next_mds
);
2739 goto out
; /* dup reply? */
2742 if (test_bit(CEPH_MDS_R_ABORTED
, &req
->r_req_flags
)) {
2743 dout("forward tid %llu aborted, unregistering\n", tid
);
2744 __unregister_request(mdsc
, req
);
2745 } else if (fwd_seq
<= req
->r_num_fwd
) {
2746 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2747 tid
, next_mds
, req
->r_num_fwd
, fwd_seq
);
2749 /* resend. forward race not possible; mds would drop */
2750 dout("forward tid %llu to mds%d (we resend)\n", tid
, next_mds
);
2752 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT
, &req
->r_req_flags
));
2753 req
->r_attempts
= 0;
2754 req
->r_num_fwd
= fwd_seq
;
2755 req
->r_resend_mds
= next_mds
;
2756 put_request_session(req
);
2757 __do_request(mdsc
, req
);
2759 ceph_mdsc_put_request(req
);
2761 mutex_unlock(&mdsc
->mutex
);
2765 pr_err("mdsc_handle_forward decode error err=%d\n", err
);
2769 * handle a mds session control message
2771 static void handle_session(struct ceph_mds_session
*session
,
2772 struct ceph_msg
*msg
)
2774 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2777 int mds
= session
->s_mds
;
2778 struct ceph_mds_session_head
*h
= msg
->front
.iov_base
;
2782 if (msg
->front
.iov_len
< sizeof(*h
))
2784 op
= le32_to_cpu(h
->op
);
2785 seq
= le64_to_cpu(h
->seq
);
2787 mutex_lock(&mdsc
->mutex
);
2788 if (op
== CEPH_SESSION_CLOSE
) {
2789 get_session(session
);
2790 __unregister_session(mdsc
, session
);
2792 /* FIXME: this ttl calculation is generous */
2793 session
->s_ttl
= jiffies
+ HZ
*mdsc
->mdsmap
->m_session_autoclose
;
2794 mutex_unlock(&mdsc
->mutex
);
2796 mutex_lock(&session
->s_mutex
);
2798 dout("handle_session mds%d %s %p state %s seq %llu\n",
2799 mds
, ceph_session_op_name(op
), session
,
2800 ceph_session_state_name(session
->s_state
), seq
);
2802 if (session
->s_state
== CEPH_MDS_SESSION_HUNG
) {
2803 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2804 pr_info("mds%d came back\n", session
->s_mds
);
2808 case CEPH_SESSION_OPEN
:
2809 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2810 pr_info("mds%d reconnect success\n", session
->s_mds
);
2811 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2812 renewed_caps(mdsc
, session
, 0);
2815 __close_session(mdsc
, session
);
2818 case CEPH_SESSION_RENEWCAPS
:
2819 if (session
->s_renew_seq
== seq
)
2820 renewed_caps(mdsc
, session
, 1);
2823 case CEPH_SESSION_CLOSE
:
2824 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2825 pr_info("mds%d reconnect denied\n", session
->s_mds
);
2826 cleanup_session_requests(mdsc
, session
);
2827 remove_session_caps(session
);
2828 wake
= 2; /* for good measure */
2829 wake_up_all(&mdsc
->session_close_wq
);
2832 case CEPH_SESSION_STALE
:
2833 pr_info("mds%d caps went stale, renewing\n",
2835 spin_lock(&session
->s_gen_ttl_lock
);
2836 session
->s_cap_gen
++;
2837 session
->s_cap_ttl
= jiffies
- 1;
2838 spin_unlock(&session
->s_gen_ttl_lock
);
2839 send_renew_caps(mdsc
, session
);
2842 case CEPH_SESSION_RECALL_STATE
:
2843 ceph_trim_caps(mdsc
, session
, le32_to_cpu(h
->max_caps
));
2846 case CEPH_SESSION_FLUSHMSG
:
2847 send_flushmsg_ack(mdsc
, session
, seq
);
2850 case CEPH_SESSION_FORCE_RO
:
2851 dout("force_session_readonly %p\n", session
);
2852 spin_lock(&session
->s_cap_lock
);
2853 session
->s_readonly
= true;
2854 spin_unlock(&session
->s_cap_lock
);
2855 wake_up_session_caps(session
, 0);
2858 case CEPH_SESSION_REJECT
:
2859 WARN_ON(session
->s_state
!= CEPH_MDS_SESSION_OPENING
);
2860 pr_info("mds%d rejected session\n", session
->s_mds
);
2861 session
->s_state
= CEPH_MDS_SESSION_REJECTED
;
2862 cleanup_session_requests(mdsc
, session
);
2863 remove_session_caps(session
);
2864 wake
= 2; /* for good measure */
2868 pr_err("mdsc_handle_session bad op %d mds%d\n", op
, mds
);
2872 mutex_unlock(&session
->s_mutex
);
2874 mutex_lock(&mdsc
->mutex
);
2875 __wake_requests(mdsc
, &session
->s_waiting
);
2877 kick_requests(mdsc
, mds
);
2878 mutex_unlock(&mdsc
->mutex
);
2880 if (op
== CEPH_SESSION_CLOSE
)
2881 ceph_put_mds_session(session
);
2885 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds
,
2886 (int)msg
->front
.iov_len
);
2893 * called under session->mutex.
2895 static void replay_unsafe_requests(struct ceph_mds_client
*mdsc
,
2896 struct ceph_mds_session
*session
)
2898 struct ceph_mds_request
*req
, *nreq
;
2902 dout("replay_unsafe_requests mds%d\n", session
->s_mds
);
2904 mutex_lock(&mdsc
->mutex
);
2905 list_for_each_entry_safe(req
, nreq
, &session
->s_unsafe
, r_unsafe_item
) {
2906 err
= __prepare_send_request(mdsc
, req
, session
->s_mds
, true);
2908 ceph_msg_get(req
->r_request
);
2909 ceph_con_send(&session
->s_con
, req
->r_request
);
2914 * also re-send old requests when MDS enters reconnect stage. So that MDS
2915 * can process completed request in clientreplay stage.
2917 p
= rb_first(&mdsc
->request_tree
);
2919 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2921 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
))
2923 if (req
->r_attempts
== 0)
2924 continue; /* only old requests */
2925 if (req
->r_session
&&
2926 req
->r_session
->s_mds
== session
->s_mds
) {
2927 err
= __prepare_send_request(mdsc
, req
,
2928 session
->s_mds
, true);
2930 ceph_msg_get(req
->r_request
);
2931 ceph_con_send(&session
->s_con
, req
->r_request
);
2935 mutex_unlock(&mdsc
->mutex
);
2939 * Encode information about a cap for a reconnect with the MDS.
2941 static int encode_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
2945 struct ceph_mds_cap_reconnect v2
;
2946 struct ceph_mds_cap_reconnect_v1 v1
;
2948 struct ceph_inode_info
*ci
= cap
->ci
;
2949 struct ceph_reconnect_state
*recon_state
= arg
;
2950 struct ceph_pagelist
*pagelist
= recon_state
->pagelist
;
2955 struct dentry
*dentry
;
2957 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2958 inode
, ceph_vinop(inode
), cap
, cap
->cap_id
,
2959 ceph_cap_string(cap
->issued
));
2960 err
= ceph_pagelist_encode_64(pagelist
, ceph_ino(inode
));
2964 dentry
= d_find_alias(inode
);
2966 path
= ceph_mdsc_build_path(dentry
, &pathlen
, &pathbase
, 0);
2968 err
= PTR_ERR(path
);
2977 spin_lock(&ci
->i_ceph_lock
);
2978 cap
->seq
= 0; /* reset cap seq */
2979 cap
->issue_seq
= 0; /* and issue_seq */
2980 cap
->mseq
= 0; /* and migrate_seq */
2981 cap
->cap_gen
= cap
->session
->s_cap_gen
;
2983 if (recon_state
->msg_version
>= 2) {
2984 rec
.v2
.cap_id
= cpu_to_le64(cap
->cap_id
);
2985 rec
.v2
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2986 rec
.v2
.issued
= cpu_to_le32(cap
->issued
);
2987 rec
.v2
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2988 rec
.v2
.pathbase
= cpu_to_le64(pathbase
);
2989 rec
.v2
.flock_len
= (__force __le32
)
2990 ((ci
->i_ceph_flags
& CEPH_I_ERROR_FILELOCK
) ? 0 : 1);
2992 rec
.v1
.cap_id
= cpu_to_le64(cap
->cap_id
);
2993 rec
.v1
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2994 rec
.v1
.issued
= cpu_to_le32(cap
->issued
);
2995 rec
.v1
.size
= cpu_to_le64(inode
->i_size
);
2996 ceph_encode_timespec64(&rec
.v1
.mtime
, &inode
->i_mtime
);
2997 ceph_encode_timespec64(&rec
.v1
.atime
, &inode
->i_atime
);
2998 rec
.v1
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2999 rec
.v1
.pathbase
= cpu_to_le64(pathbase
);
3002 if (list_empty(&ci
->i_cap_snaps
)) {
3003 snap_follows
= ci
->i_head_snapc
? ci
->i_head_snapc
->seq
: 0;
3005 struct ceph_cap_snap
*capsnap
=
3006 list_first_entry(&ci
->i_cap_snaps
,
3007 struct ceph_cap_snap
, ci_item
);
3008 snap_follows
= capsnap
->follows
;
3010 spin_unlock(&ci
->i_ceph_lock
);
3012 if (recon_state
->msg_version
>= 2) {
3013 int num_fcntl_locks
, num_flock_locks
;
3014 struct ceph_filelock
*flocks
= NULL
;
3015 size_t struct_len
, total_len
= 0;
3019 if (rec
.v2
.flock_len
) {
3020 ceph_count_locks(inode
, &num_fcntl_locks
, &num_flock_locks
);
3022 num_fcntl_locks
= 0;
3023 num_flock_locks
= 0;
3025 if (num_fcntl_locks
+ num_flock_locks
> 0) {
3026 flocks
= kmalloc_array(num_fcntl_locks
+ num_flock_locks
,
3027 sizeof(struct ceph_filelock
),
3033 err
= ceph_encode_locks_to_buffer(inode
, flocks
,
3048 if (recon_state
->msg_version
>= 3) {
3049 /* version, compat_version and struct_len */
3050 total_len
= 2 * sizeof(u8
) + sizeof(u32
);
3054 * number of encoded locks is stable, so copy to pagelist
3056 struct_len
= 2 * sizeof(u32
) +
3057 (num_fcntl_locks
+ num_flock_locks
) *
3058 sizeof(struct ceph_filelock
);
3059 rec
.v2
.flock_len
= cpu_to_le32(struct_len
);
3061 struct_len
+= sizeof(rec
.v2
);
3062 struct_len
+= sizeof(u32
) + pathlen
;
3065 struct_len
+= sizeof(u64
); /* snap_follows */
3067 total_len
+= struct_len
;
3068 err
= ceph_pagelist_reserve(pagelist
, total_len
);
3071 if (recon_state
->msg_version
>= 3) {
3072 ceph_pagelist_encode_8(pagelist
, struct_v
);
3073 ceph_pagelist_encode_8(pagelist
, 1);
3074 ceph_pagelist_encode_32(pagelist
, struct_len
);
3076 ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
3077 ceph_pagelist_append(pagelist
, &rec
, sizeof(rec
.v2
));
3078 ceph_locks_to_pagelist(flocks
, pagelist
,
3082 ceph_pagelist_encode_64(pagelist
, snap_follows
);
3086 size_t size
= sizeof(u32
) + pathlen
+ sizeof(rec
.v1
);
3087 err
= ceph_pagelist_reserve(pagelist
, size
);
3089 ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
3090 ceph_pagelist_append(pagelist
, &rec
, sizeof(rec
.v1
));
3094 recon_state
->nr_caps
++;
3104 * If an MDS fails and recovers, clients need to reconnect in order to
3105 * reestablish shared state. This includes all caps issued through
3106 * this session _and_ the snap_realm hierarchy. Because it's not
3107 * clear which snap realms the mds cares about, we send everything we
3108 * know about.. that ensures we'll then get any new info the
3109 * recovering MDS might have.
3111 * This is a relatively heavyweight operation, but it's rare.
3113 * called with mdsc->mutex held.
3115 static void send_mds_reconnect(struct ceph_mds_client
*mdsc
,
3116 struct ceph_mds_session
*session
)
3118 struct ceph_msg
*reply
;
3120 int mds
= session
->s_mds
;
3123 struct ceph_pagelist
*pagelist
;
3124 struct ceph_reconnect_state recon_state
;
3127 pr_info("mds%d reconnect start\n", mds
);
3129 pagelist
= kmalloc(sizeof(*pagelist
), GFP_NOFS
);
3131 goto fail_nopagelist
;
3132 ceph_pagelist_init(pagelist
);
3134 reply
= ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT
, 0, GFP_NOFS
, false);
3138 mutex_lock(&session
->s_mutex
);
3139 session
->s_state
= CEPH_MDS_SESSION_RECONNECTING
;
3142 dout("session %p state %s\n", session
,
3143 ceph_session_state_name(session
->s_state
));
3145 spin_lock(&session
->s_gen_ttl_lock
);
3146 session
->s_cap_gen
++;
3147 spin_unlock(&session
->s_gen_ttl_lock
);
3149 spin_lock(&session
->s_cap_lock
);
3150 /* don't know if session is readonly */
3151 session
->s_readonly
= 0;
3153 * notify __ceph_remove_cap() that we are composing cap reconnect.
3154 * If a cap get released before being added to the cap reconnect,
3155 * __ceph_remove_cap() should skip queuing cap release.
3157 session
->s_cap_reconnect
= 1;
3158 /* drop old cap expires; we're about to reestablish that state */
3159 detach_cap_releases(session
, &dispose
);
3160 spin_unlock(&session
->s_cap_lock
);
3161 dispose_cap_releases(mdsc
, &dispose
);
3163 /* trim unused caps to reduce MDS's cache rejoin time */
3164 if (mdsc
->fsc
->sb
->s_root
)
3165 shrink_dcache_parent(mdsc
->fsc
->sb
->s_root
);
3167 ceph_con_close(&session
->s_con
);
3168 ceph_con_open(&session
->s_con
,
3169 CEPH_ENTITY_TYPE_MDS
, mds
,
3170 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
3172 /* replay unsafe requests */
3173 replay_unsafe_requests(mdsc
, session
);
3175 down_read(&mdsc
->snap_rwsem
);
3177 /* traverse this session's caps */
3178 s_nr_caps
= session
->s_nr_caps
;
3179 err
= ceph_pagelist_encode_32(pagelist
, s_nr_caps
);
3183 recon_state
.nr_caps
= 0;
3184 recon_state
.pagelist
= pagelist
;
3185 if (session
->s_con
.peer_features
& CEPH_FEATURE_MDSENC
)
3186 recon_state
.msg_version
= 3;
3187 else if (session
->s_con
.peer_features
& CEPH_FEATURE_FLOCK
)
3188 recon_state
.msg_version
= 2;
3190 recon_state
.msg_version
= 1;
3191 err
= iterate_session_caps(session
, encode_caps_cb
, &recon_state
);
3195 spin_lock(&session
->s_cap_lock
);
3196 session
->s_cap_reconnect
= 0;
3197 spin_unlock(&session
->s_cap_lock
);
3200 * snaprealms. we provide mds with the ino, seq (version), and
3201 * parent for all of our realms. If the mds has any newer info,
3204 for (p
= rb_first(&mdsc
->snap_realms
); p
; p
= rb_next(p
)) {
3205 struct ceph_snap_realm
*realm
=
3206 rb_entry(p
, struct ceph_snap_realm
, node
);
3207 struct ceph_mds_snaprealm_reconnect sr_rec
;
3209 dout(" adding snap realm %llx seq %lld parent %llx\n",
3210 realm
->ino
, realm
->seq
, realm
->parent_ino
);
3211 sr_rec
.ino
= cpu_to_le64(realm
->ino
);
3212 sr_rec
.seq
= cpu_to_le64(realm
->seq
);
3213 sr_rec
.parent
= cpu_to_le64(realm
->parent_ino
);
3214 err
= ceph_pagelist_append(pagelist
, &sr_rec
, sizeof(sr_rec
));
3219 reply
->hdr
.version
= cpu_to_le16(recon_state
.msg_version
);
3221 /* raced with cap release? */
3222 if (s_nr_caps
!= recon_state
.nr_caps
) {
3223 struct page
*page
= list_first_entry(&pagelist
->head
,
3225 __le32
*addr
= kmap_atomic(page
);
3226 *addr
= cpu_to_le32(recon_state
.nr_caps
);
3227 kunmap_atomic(addr
);
3230 reply
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
3231 ceph_msg_data_add_pagelist(reply
, pagelist
);
3233 ceph_early_kick_flushing_caps(mdsc
, session
);
3235 ceph_con_send(&session
->s_con
, reply
);
3237 mutex_unlock(&session
->s_mutex
);
3239 mutex_lock(&mdsc
->mutex
);
3240 __wake_requests(mdsc
, &session
->s_waiting
);
3241 mutex_unlock(&mdsc
->mutex
);
3243 up_read(&mdsc
->snap_rwsem
);
3247 ceph_msg_put(reply
);
3248 up_read(&mdsc
->snap_rwsem
);
3249 mutex_unlock(&session
->s_mutex
);
3251 ceph_pagelist_release(pagelist
);
3253 pr_err("error %d preparing reconnect for mds%d\n", err
, mds
);
3259 * compare old and new mdsmaps, kicking requests
3260 * and closing out old connections as necessary
3262 * called under mdsc->mutex.
3264 static void check_new_map(struct ceph_mds_client
*mdsc
,
3265 struct ceph_mdsmap
*newmap
,
3266 struct ceph_mdsmap
*oldmap
)
3269 int oldstate
, newstate
;
3270 struct ceph_mds_session
*s
;
3272 dout("check_new_map new %u old %u\n",
3273 newmap
->m_epoch
, oldmap
->m_epoch
);
3275 for (i
= 0; i
< oldmap
->m_num_mds
&& i
< mdsc
->max_sessions
; i
++) {
3276 if (!mdsc
->sessions
[i
])
3278 s
= mdsc
->sessions
[i
];
3279 oldstate
= ceph_mdsmap_get_state(oldmap
, i
);
3280 newstate
= ceph_mdsmap_get_state(newmap
, i
);
3282 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3283 i
, ceph_mds_state_name(oldstate
),
3284 ceph_mdsmap_is_laggy(oldmap
, i
) ? " (laggy)" : "",
3285 ceph_mds_state_name(newstate
),
3286 ceph_mdsmap_is_laggy(newmap
, i
) ? " (laggy)" : "",
3287 ceph_session_state_name(s
->s_state
));
3289 if (i
>= newmap
->m_num_mds
||
3290 memcmp(ceph_mdsmap_get_addr(oldmap
, i
),
3291 ceph_mdsmap_get_addr(newmap
, i
),
3292 sizeof(struct ceph_entity_addr
))) {
3293 if (s
->s_state
== CEPH_MDS_SESSION_OPENING
) {
3294 /* the session never opened, just close it
3297 __unregister_session(mdsc
, s
);
3298 __wake_requests(mdsc
, &s
->s_waiting
);
3299 ceph_put_mds_session(s
);
3300 } else if (i
>= newmap
->m_num_mds
) {
3301 /* force close session for stopped mds */
3303 __unregister_session(mdsc
, s
);
3304 __wake_requests(mdsc
, &s
->s_waiting
);
3305 kick_requests(mdsc
, i
);
3306 mutex_unlock(&mdsc
->mutex
);
3308 mutex_lock(&s
->s_mutex
);
3309 cleanup_session_requests(mdsc
, s
);
3310 remove_session_caps(s
);
3311 mutex_unlock(&s
->s_mutex
);
3313 ceph_put_mds_session(s
);
3315 mutex_lock(&mdsc
->mutex
);
3318 mutex_unlock(&mdsc
->mutex
);
3319 mutex_lock(&s
->s_mutex
);
3320 mutex_lock(&mdsc
->mutex
);
3321 ceph_con_close(&s
->s_con
);
3322 mutex_unlock(&s
->s_mutex
);
3323 s
->s_state
= CEPH_MDS_SESSION_RESTARTING
;
3325 } else if (oldstate
== newstate
) {
3326 continue; /* nothing new with this mds */
3332 if (s
->s_state
== CEPH_MDS_SESSION_RESTARTING
&&
3333 newstate
>= CEPH_MDS_STATE_RECONNECT
) {
3334 mutex_unlock(&mdsc
->mutex
);
3335 send_mds_reconnect(mdsc
, s
);
3336 mutex_lock(&mdsc
->mutex
);
3340 * kick request on any mds that has gone active.
3342 if (oldstate
< CEPH_MDS_STATE_ACTIVE
&&
3343 newstate
>= CEPH_MDS_STATE_ACTIVE
) {
3344 if (oldstate
!= CEPH_MDS_STATE_CREATING
&&
3345 oldstate
!= CEPH_MDS_STATE_STARTING
)
3346 pr_info("mds%d recovery completed\n", s
->s_mds
);
3347 kick_requests(mdsc
, i
);
3348 ceph_kick_flushing_caps(mdsc
, s
);
3349 wake_up_session_caps(s
, 1);
3353 for (i
= 0; i
< newmap
->m_num_mds
&& i
< mdsc
->max_sessions
; i
++) {
3354 s
= mdsc
->sessions
[i
];
3357 if (!ceph_mdsmap_is_laggy(newmap
, i
))
3359 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3360 s
->s_state
== CEPH_MDS_SESSION_HUNG
||
3361 s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3362 dout(" connecting to export targets of laggy mds%d\n",
3364 __open_export_target_sessions(mdsc
, s
);
3376 * caller must hold session s_mutex, dentry->d_lock
3378 void __ceph_mdsc_drop_dentry_lease(struct dentry
*dentry
)
3380 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
3382 ceph_put_mds_session(di
->lease_session
);
3383 di
->lease_session
= NULL
;
3386 static void handle_lease(struct ceph_mds_client
*mdsc
,
3387 struct ceph_mds_session
*session
,
3388 struct ceph_msg
*msg
)
3390 struct super_block
*sb
= mdsc
->fsc
->sb
;
3391 struct inode
*inode
;
3392 struct dentry
*parent
, *dentry
;
3393 struct ceph_dentry_info
*di
;
3394 int mds
= session
->s_mds
;
3395 struct ceph_mds_lease
*h
= msg
->front
.iov_base
;
3397 struct ceph_vino vino
;
3401 dout("handle_lease from mds%d\n", mds
);
3404 if (msg
->front
.iov_len
< sizeof(*h
) + sizeof(u32
))
3406 vino
.ino
= le64_to_cpu(h
->ino
);
3407 vino
.snap
= CEPH_NOSNAP
;
3408 seq
= le32_to_cpu(h
->seq
);
3409 dname
.len
= get_unaligned_le32(h
+ 1);
3410 if (msg
->front
.iov_len
< sizeof(*h
) + sizeof(u32
) + dname
.len
)
3412 dname
.name
= (void *)(h
+ 1) + sizeof(u32
);
3415 inode
= ceph_find_inode(sb
, vino
);
3416 dout("handle_lease %s, ino %llx %p %.*s\n",
3417 ceph_lease_op_name(h
->action
), vino
.ino
, inode
,
3418 dname
.len
, dname
.name
);
3420 mutex_lock(&session
->s_mutex
);
3424 dout("handle_lease no inode %llx\n", vino
.ino
);
3429 parent
= d_find_alias(inode
);
3431 dout("no parent dentry on inode %p\n", inode
);
3433 goto release
; /* hrm... */
3435 dname
.hash
= full_name_hash(parent
, dname
.name
, dname
.len
);
3436 dentry
= d_lookup(parent
, &dname
);
3441 spin_lock(&dentry
->d_lock
);
3442 di
= ceph_dentry(dentry
);
3443 switch (h
->action
) {
3444 case CEPH_MDS_LEASE_REVOKE
:
3445 if (di
->lease_session
== session
) {
3446 if (ceph_seq_cmp(di
->lease_seq
, seq
) > 0)
3447 h
->seq
= cpu_to_le32(di
->lease_seq
);
3448 __ceph_mdsc_drop_dentry_lease(dentry
);
3453 case CEPH_MDS_LEASE_RENEW
:
3454 if (di
->lease_session
== session
&&
3455 di
->lease_gen
== session
->s_cap_gen
&&
3456 di
->lease_renew_from
&&
3457 di
->lease_renew_after
== 0) {
3458 unsigned long duration
=
3459 msecs_to_jiffies(le32_to_cpu(h
->duration_ms
));
3461 di
->lease_seq
= seq
;
3462 di
->time
= di
->lease_renew_from
+ duration
;
3463 di
->lease_renew_after
= di
->lease_renew_from
+
3465 di
->lease_renew_from
= 0;
3469 spin_unlock(&dentry
->d_lock
);
3476 /* let's just reuse the same message */
3477 h
->action
= CEPH_MDS_LEASE_REVOKE_ACK
;
3479 ceph_con_send(&session
->s_con
, msg
);
3483 mutex_unlock(&session
->s_mutex
);
3487 pr_err("corrupt lease message\n");
3491 void ceph_mdsc_lease_send_msg(struct ceph_mds_session
*session
,
3492 struct inode
*inode
,
3493 struct dentry
*dentry
, char action
,
3496 struct ceph_msg
*msg
;
3497 struct ceph_mds_lease
*lease
;
3498 int len
= sizeof(*lease
) + sizeof(u32
);
3501 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3502 inode
, dentry
, ceph_lease_op_name(action
), session
->s_mds
);
3503 dnamelen
= dentry
->d_name
.len
;
3506 msg
= ceph_msg_new(CEPH_MSG_CLIENT_LEASE
, len
, GFP_NOFS
, false);
3509 lease
= msg
->front
.iov_base
;
3510 lease
->action
= action
;
3511 lease
->ino
= cpu_to_le64(ceph_vino(inode
).ino
);
3512 lease
->first
= lease
->last
= cpu_to_le64(ceph_vino(inode
).snap
);
3513 lease
->seq
= cpu_to_le32(seq
);
3514 put_unaligned_le32(dnamelen
, lease
+ 1);
3515 memcpy((void *)(lease
+ 1) + 4, dentry
->d_name
.name
, dnamelen
);
3518 * if this is a preemptive lease RELEASE, no need to
3519 * flush request stream, since the actual request will
3522 msg
->more_to_follow
= (action
== CEPH_MDS_LEASE_RELEASE
);
3524 ceph_con_send(&session
->s_con
, msg
);
3528 * lock unlock sessions, to wait ongoing session activities
3530 static void lock_unlock_sessions(struct ceph_mds_client
*mdsc
)
3534 mutex_lock(&mdsc
->mutex
);
3535 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3536 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3539 mutex_unlock(&mdsc
->mutex
);
3540 mutex_lock(&s
->s_mutex
);
3541 mutex_unlock(&s
->s_mutex
);
3542 ceph_put_mds_session(s
);
3543 mutex_lock(&mdsc
->mutex
);
3545 mutex_unlock(&mdsc
->mutex
);
3551 * delayed work -- periodically trim expired leases, renew caps with mds
3553 static void schedule_delayed(struct ceph_mds_client
*mdsc
)
3556 unsigned hz
= round_jiffies_relative(HZ
* delay
);
3557 schedule_delayed_work(&mdsc
->delayed_work
, hz
);
3560 static void delayed_work(struct work_struct
*work
)
3563 struct ceph_mds_client
*mdsc
=
3564 container_of(work
, struct ceph_mds_client
, delayed_work
.work
);
3568 dout("mdsc delayed_work\n");
3569 ceph_check_delayed_caps(mdsc
);
3571 mutex_lock(&mdsc
->mutex
);
3572 renew_interval
= mdsc
->mdsmap
->m_session_timeout
>> 2;
3573 renew_caps
= time_after_eq(jiffies
, HZ
*renew_interval
+
3574 mdsc
->last_renew_caps
);
3576 mdsc
->last_renew_caps
= jiffies
;
3578 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3579 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3582 if (s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3583 dout("resending session close request for mds%d\n",
3585 request_close_session(mdsc
, s
);
3586 ceph_put_mds_session(s
);
3589 if (s
->s_ttl
&& time_after(jiffies
, s
->s_ttl
)) {
3590 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
) {
3591 s
->s_state
= CEPH_MDS_SESSION_HUNG
;
3592 pr_info("mds%d hung\n", s
->s_mds
);
3595 if (s
->s_state
< CEPH_MDS_SESSION_OPEN
) {
3596 /* this mds is failed or recovering, just wait */
3597 ceph_put_mds_session(s
);
3600 mutex_unlock(&mdsc
->mutex
);
3602 mutex_lock(&s
->s_mutex
);
3604 send_renew_caps(mdsc
, s
);
3606 ceph_con_keepalive(&s
->s_con
);
3607 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3608 s
->s_state
== CEPH_MDS_SESSION_HUNG
)
3609 ceph_send_cap_releases(mdsc
, s
);
3610 mutex_unlock(&s
->s_mutex
);
3611 ceph_put_mds_session(s
);
3613 mutex_lock(&mdsc
->mutex
);
3615 mutex_unlock(&mdsc
->mutex
);
3617 schedule_delayed(mdsc
);
3620 int ceph_mdsc_init(struct ceph_fs_client
*fsc
)
3623 struct ceph_mds_client
*mdsc
;
3625 mdsc
= kzalloc(sizeof(struct ceph_mds_client
), GFP_NOFS
);
3629 mutex_init(&mdsc
->mutex
);
3630 mdsc
->mdsmap
= kzalloc(sizeof(*mdsc
->mdsmap
), GFP_NOFS
);
3631 if (!mdsc
->mdsmap
) {
3637 init_completion(&mdsc
->safe_umount_waiters
);
3638 init_waitqueue_head(&mdsc
->session_close_wq
);
3639 INIT_LIST_HEAD(&mdsc
->waiting_for_map
);
3640 mdsc
->sessions
= NULL
;
3641 atomic_set(&mdsc
->num_sessions
, 0);
3642 mdsc
->max_sessions
= 0;
3644 atomic64_set(&mdsc
->quotarealms_count
, 0);
3645 mdsc
->last_snap_seq
= 0;
3646 init_rwsem(&mdsc
->snap_rwsem
);
3647 mdsc
->snap_realms
= RB_ROOT
;
3648 INIT_LIST_HEAD(&mdsc
->snap_empty
);
3649 spin_lock_init(&mdsc
->snap_empty_lock
);
3651 mdsc
->oldest_tid
= 0;
3652 mdsc
->request_tree
= RB_ROOT
;
3653 INIT_DELAYED_WORK(&mdsc
->delayed_work
, delayed_work
);
3654 mdsc
->last_renew_caps
= jiffies
;
3655 INIT_LIST_HEAD(&mdsc
->cap_delay_list
);
3656 spin_lock_init(&mdsc
->cap_delay_lock
);
3657 INIT_LIST_HEAD(&mdsc
->snap_flush_list
);
3658 spin_lock_init(&mdsc
->snap_flush_lock
);
3659 mdsc
->last_cap_flush_tid
= 1;
3660 INIT_LIST_HEAD(&mdsc
->cap_flush_list
);
3661 INIT_LIST_HEAD(&mdsc
->cap_dirty
);
3662 INIT_LIST_HEAD(&mdsc
->cap_dirty_migrating
);
3663 mdsc
->num_cap_flushing
= 0;
3664 spin_lock_init(&mdsc
->cap_dirty_lock
);
3665 init_waitqueue_head(&mdsc
->cap_flushing_wq
);
3666 spin_lock_init(&mdsc
->dentry_lru_lock
);
3667 INIT_LIST_HEAD(&mdsc
->dentry_lru
);
3669 ceph_caps_init(mdsc
);
3670 ceph_adjust_min_caps(mdsc
, fsc
->min_caps
);
3672 init_rwsem(&mdsc
->pool_perm_rwsem
);
3673 mdsc
->pool_perm_tree
= RB_ROOT
;
3675 strscpy(mdsc
->nodename
, utsname()->nodename
,
3676 sizeof(mdsc
->nodename
));
3681 * Wait for safe replies on open mds requests. If we time out, drop
3682 * all requests from the tree to avoid dangling dentry refs.
3684 static void wait_requests(struct ceph_mds_client
*mdsc
)
3686 struct ceph_options
*opts
= mdsc
->fsc
->client
->options
;
3687 struct ceph_mds_request
*req
;
3689 mutex_lock(&mdsc
->mutex
);
3690 if (__get_oldest_req(mdsc
)) {
3691 mutex_unlock(&mdsc
->mutex
);
3693 dout("wait_requests waiting for requests\n");
3694 wait_for_completion_timeout(&mdsc
->safe_umount_waiters
,
3695 ceph_timeout_jiffies(opts
->mount_timeout
));
3697 /* tear down remaining requests */
3698 mutex_lock(&mdsc
->mutex
);
3699 while ((req
= __get_oldest_req(mdsc
))) {
3700 dout("wait_requests timed out on tid %llu\n",
3702 __unregister_request(mdsc
, req
);
3705 mutex_unlock(&mdsc
->mutex
);
3706 dout("wait_requests done\n");
3710 * called before mount is ro, and before dentries are torn down.
3711 * (hmm, does this still race with new lookups?)
3713 void ceph_mdsc_pre_umount(struct ceph_mds_client
*mdsc
)
3715 dout("pre_umount\n");
3718 lock_unlock_sessions(mdsc
);
3719 ceph_flush_dirty_caps(mdsc
);
3720 wait_requests(mdsc
);
3723 * wait for reply handlers to drop their request refs and
3724 * their inode/dcache refs
3730 * wait for all write mds requests to flush.
3732 static void wait_unsafe_requests(struct ceph_mds_client
*mdsc
, u64 want_tid
)
3734 struct ceph_mds_request
*req
= NULL
, *nextreq
;
3737 mutex_lock(&mdsc
->mutex
);
3738 dout("wait_unsafe_requests want %lld\n", want_tid
);
3740 req
= __get_oldest_req(mdsc
);
3741 while (req
&& req
->r_tid
<= want_tid
) {
3742 /* find next request */
3743 n
= rb_next(&req
->r_node
);
3745 nextreq
= rb_entry(n
, struct ceph_mds_request
, r_node
);
3748 if (req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
&&
3749 (req
->r_op
& CEPH_MDS_OP_WRITE
)) {
3751 ceph_mdsc_get_request(req
);
3753 ceph_mdsc_get_request(nextreq
);
3754 mutex_unlock(&mdsc
->mutex
);
3755 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3756 req
->r_tid
, want_tid
);
3757 wait_for_completion(&req
->r_safe_completion
);
3758 mutex_lock(&mdsc
->mutex
);
3759 ceph_mdsc_put_request(req
);
3761 break; /* next dne before, so we're done! */
3762 if (RB_EMPTY_NODE(&nextreq
->r_node
)) {
3763 /* next request was removed from tree */
3764 ceph_mdsc_put_request(nextreq
);
3767 ceph_mdsc_put_request(nextreq
); /* won't go away */
3771 mutex_unlock(&mdsc
->mutex
);
3772 dout("wait_unsafe_requests done\n");
3775 void ceph_mdsc_sync(struct ceph_mds_client
*mdsc
)
3777 u64 want_tid
, want_flush
;
3779 if (READ_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
3783 mutex_lock(&mdsc
->mutex
);
3784 want_tid
= mdsc
->last_tid
;
3785 mutex_unlock(&mdsc
->mutex
);
3787 ceph_flush_dirty_caps(mdsc
);
3788 spin_lock(&mdsc
->cap_dirty_lock
);
3789 want_flush
= mdsc
->last_cap_flush_tid
;
3790 if (!list_empty(&mdsc
->cap_flush_list
)) {
3791 struct ceph_cap_flush
*cf
=
3792 list_last_entry(&mdsc
->cap_flush_list
,
3793 struct ceph_cap_flush
, g_list
);
3796 spin_unlock(&mdsc
->cap_dirty_lock
);
3798 dout("sync want tid %lld flush_seq %lld\n",
3799 want_tid
, want_flush
);
3801 wait_unsafe_requests(mdsc
, want_tid
);
3802 wait_caps_flush(mdsc
, want_flush
);
3806 * true if all sessions are closed, or we force unmount
3808 static bool done_closing_sessions(struct ceph_mds_client
*mdsc
, int skipped
)
3810 if (READ_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
3812 return atomic_read(&mdsc
->num_sessions
) <= skipped
;
3816 * called after sb is ro.
3818 void ceph_mdsc_close_sessions(struct ceph_mds_client
*mdsc
)
3820 struct ceph_options
*opts
= mdsc
->fsc
->client
->options
;
3821 struct ceph_mds_session
*session
;
3825 dout("close_sessions\n");
3827 /* close sessions */
3828 mutex_lock(&mdsc
->mutex
);
3829 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3830 session
= __ceph_lookup_mds_session(mdsc
, i
);
3833 mutex_unlock(&mdsc
->mutex
);
3834 mutex_lock(&session
->s_mutex
);
3835 if (__close_session(mdsc
, session
) <= 0)
3837 mutex_unlock(&session
->s_mutex
);
3838 ceph_put_mds_session(session
);
3839 mutex_lock(&mdsc
->mutex
);
3841 mutex_unlock(&mdsc
->mutex
);
3843 dout("waiting for sessions to close\n");
3844 wait_event_timeout(mdsc
->session_close_wq
,
3845 done_closing_sessions(mdsc
, skipped
),
3846 ceph_timeout_jiffies(opts
->mount_timeout
));
3848 /* tear down remaining sessions */
3849 mutex_lock(&mdsc
->mutex
);
3850 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3851 if (mdsc
->sessions
[i
]) {
3852 session
= get_session(mdsc
->sessions
[i
]);
3853 __unregister_session(mdsc
, session
);
3854 mutex_unlock(&mdsc
->mutex
);
3855 mutex_lock(&session
->s_mutex
);
3856 remove_session_caps(session
);
3857 mutex_unlock(&session
->s_mutex
);
3858 ceph_put_mds_session(session
);
3859 mutex_lock(&mdsc
->mutex
);
3862 WARN_ON(!list_empty(&mdsc
->cap_delay_list
));
3863 mutex_unlock(&mdsc
->mutex
);
3865 ceph_cleanup_empty_realms(mdsc
);
3867 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3872 void ceph_mdsc_force_umount(struct ceph_mds_client
*mdsc
)
3874 struct ceph_mds_session
*session
;
3877 dout("force umount\n");
3879 mutex_lock(&mdsc
->mutex
);
3880 for (mds
= 0; mds
< mdsc
->max_sessions
; mds
++) {
3881 session
= __ceph_lookup_mds_session(mdsc
, mds
);
3884 mutex_unlock(&mdsc
->mutex
);
3885 mutex_lock(&session
->s_mutex
);
3886 __close_session(mdsc
, session
);
3887 if (session
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3888 cleanup_session_requests(mdsc
, session
);
3889 remove_session_caps(session
);
3891 mutex_unlock(&session
->s_mutex
);
3892 ceph_put_mds_session(session
);
3893 mutex_lock(&mdsc
->mutex
);
3894 kick_requests(mdsc
, mds
);
3896 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3897 mutex_unlock(&mdsc
->mutex
);
3900 static void ceph_mdsc_stop(struct ceph_mds_client
*mdsc
)
3903 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3905 ceph_mdsmap_destroy(mdsc
->mdsmap
);
3906 kfree(mdsc
->sessions
);
3907 ceph_caps_finalize(mdsc
);
3908 ceph_pool_perm_destroy(mdsc
);
3911 void ceph_mdsc_destroy(struct ceph_fs_client
*fsc
)
3913 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
3914 dout("mdsc_destroy %p\n", mdsc
);
3919 /* flush out any connection work with references to us */
3922 ceph_mdsc_stop(mdsc
);
3926 dout("mdsc_destroy %p done\n", mdsc
);
3929 void ceph_mdsc_handle_fsmap(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3931 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3932 const char *mds_namespace
= fsc
->mount_options
->mds_namespace
;
3933 void *p
= msg
->front
.iov_base
;
3934 void *end
= p
+ msg
->front
.iov_len
;
3938 u32 mount_fscid
= (u32
)-1;
3939 u8 struct_v
, struct_cv
;
3942 ceph_decode_need(&p
, end
, sizeof(u32
), bad
);
3943 epoch
= ceph_decode_32(&p
);
3945 dout("handle_fsmap epoch %u\n", epoch
);
3947 ceph_decode_need(&p
, end
, 2 + sizeof(u32
), bad
);
3948 struct_v
= ceph_decode_8(&p
);
3949 struct_cv
= ceph_decode_8(&p
);
3950 map_len
= ceph_decode_32(&p
);
3952 ceph_decode_need(&p
, end
, sizeof(u32
) * 3, bad
);
3953 p
+= sizeof(u32
) * 2; /* skip epoch and legacy_client_fscid */
3955 num_fs
= ceph_decode_32(&p
);
3956 while (num_fs
-- > 0) {
3957 void *info_p
, *info_end
;
3962 ceph_decode_need(&p
, end
, 2 + sizeof(u32
), bad
);
3963 info_v
= ceph_decode_8(&p
);
3964 info_cv
= ceph_decode_8(&p
);
3965 info_len
= ceph_decode_32(&p
);
3966 ceph_decode_need(&p
, end
, info_len
, bad
);
3968 info_end
= p
+ info_len
;
3971 ceph_decode_need(&info_p
, info_end
, sizeof(u32
) * 2, bad
);
3972 fscid
= ceph_decode_32(&info_p
);
3973 namelen
= ceph_decode_32(&info_p
);
3974 ceph_decode_need(&info_p
, info_end
, namelen
, bad
);
3976 if (mds_namespace
&&
3977 strlen(mds_namespace
) == namelen
&&
3978 !strncmp(mds_namespace
, (char *)info_p
, namelen
)) {
3979 mount_fscid
= fscid
;
3984 ceph_monc_got_map(&fsc
->client
->monc
, CEPH_SUB_FSMAP
, epoch
);
3985 if (mount_fscid
!= (u32
)-1) {
3986 fsc
->client
->monc
.fs_cluster_id
= mount_fscid
;
3987 ceph_monc_want_map(&fsc
->client
->monc
, CEPH_SUB_MDSMAP
,
3989 ceph_monc_renew_subs(&fsc
->client
->monc
);
3997 pr_err("error decoding fsmap\n");
3999 mutex_lock(&mdsc
->mutex
);
4000 mdsc
->mdsmap_err
= err
;
4001 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
4002 mutex_unlock(&mdsc
->mutex
);
4006 * handle mds map update.
4008 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
4012 void *p
= msg
->front
.iov_base
;
4013 void *end
= p
+ msg
->front
.iov_len
;
4014 struct ceph_mdsmap
*newmap
, *oldmap
;
4015 struct ceph_fsid fsid
;
4018 ceph_decode_need(&p
, end
, sizeof(fsid
)+2*sizeof(u32
), bad
);
4019 ceph_decode_copy(&p
, &fsid
, sizeof(fsid
));
4020 if (ceph_check_fsid(mdsc
->fsc
->client
, &fsid
) < 0)
4022 epoch
= ceph_decode_32(&p
);
4023 maplen
= ceph_decode_32(&p
);
4024 dout("handle_map epoch %u len %d\n", epoch
, (int)maplen
);
4026 /* do we need it? */
4027 mutex_lock(&mdsc
->mutex
);
4028 if (mdsc
->mdsmap
&& epoch
<= mdsc
->mdsmap
->m_epoch
) {
4029 dout("handle_map epoch %u <= our %u\n",
4030 epoch
, mdsc
->mdsmap
->m_epoch
);
4031 mutex_unlock(&mdsc
->mutex
);
4035 newmap
= ceph_mdsmap_decode(&p
, end
);
4036 if (IS_ERR(newmap
)) {
4037 err
= PTR_ERR(newmap
);
4041 /* swap into place */
4043 oldmap
= mdsc
->mdsmap
;
4044 mdsc
->mdsmap
= newmap
;
4045 check_new_map(mdsc
, newmap
, oldmap
);
4046 ceph_mdsmap_destroy(oldmap
);
4048 mdsc
->mdsmap
= newmap
; /* first mds map */
4050 mdsc
->fsc
->max_file_size
= min((loff_t
)mdsc
->mdsmap
->m_max_file_size
,
4053 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
4054 ceph_monc_got_map(&mdsc
->fsc
->client
->monc
, CEPH_SUB_MDSMAP
,
4055 mdsc
->mdsmap
->m_epoch
);
4057 mutex_unlock(&mdsc
->mutex
);
4058 schedule_delayed(mdsc
);
4062 mutex_unlock(&mdsc
->mutex
);
4064 pr_err("error decoding mdsmap %d\n", err
);
4068 static struct ceph_connection
*con_get(struct ceph_connection
*con
)
4070 struct ceph_mds_session
*s
= con
->private;
4072 if (get_session(s
)) {
4073 dout("mdsc con_get %p ok (%d)\n", s
, refcount_read(&s
->s_ref
));
4076 dout("mdsc con_get %p FAIL\n", s
);
4080 static void con_put(struct ceph_connection
*con
)
4082 struct ceph_mds_session
*s
= con
->private;
4084 dout("mdsc con_put %p (%d)\n", s
, refcount_read(&s
->s_ref
) - 1);
4085 ceph_put_mds_session(s
);
4089 * if the client is unresponsive for long enough, the mds will kill
4090 * the session entirely.
4092 static void peer_reset(struct ceph_connection
*con
)
4094 struct ceph_mds_session
*s
= con
->private;
4095 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4097 pr_warn("mds%d closed our session\n", s
->s_mds
);
4098 send_mds_reconnect(mdsc
, s
);
4101 static void dispatch(struct ceph_connection
*con
, struct ceph_msg
*msg
)
4103 struct ceph_mds_session
*s
= con
->private;
4104 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4105 int type
= le16_to_cpu(msg
->hdr
.type
);
4107 mutex_lock(&mdsc
->mutex
);
4108 if (__verify_registered_session(mdsc
, s
) < 0) {
4109 mutex_unlock(&mdsc
->mutex
);
4112 mutex_unlock(&mdsc
->mutex
);
4115 case CEPH_MSG_MDS_MAP
:
4116 ceph_mdsc_handle_mdsmap(mdsc
, msg
);
4118 case CEPH_MSG_FS_MAP_USER
:
4119 ceph_mdsc_handle_fsmap(mdsc
, msg
);
4121 case CEPH_MSG_CLIENT_SESSION
:
4122 handle_session(s
, msg
);
4124 case CEPH_MSG_CLIENT_REPLY
:
4125 handle_reply(s
, msg
);
4127 case CEPH_MSG_CLIENT_REQUEST_FORWARD
:
4128 handle_forward(mdsc
, s
, msg
);
4130 case CEPH_MSG_CLIENT_CAPS
:
4131 ceph_handle_caps(s
, msg
);
4133 case CEPH_MSG_CLIENT_SNAP
:
4134 ceph_handle_snap(mdsc
, s
, msg
);
4136 case CEPH_MSG_CLIENT_LEASE
:
4137 handle_lease(mdsc
, s
, msg
);
4139 case CEPH_MSG_CLIENT_QUOTA
:
4140 ceph_handle_quota(mdsc
, s
, msg
);
4144 pr_err("received unknown message type %d %s\n", type
,
4145 ceph_msg_type_name(type
));
4156 * Note: returned pointer is the address of a structure that's
4157 * managed separately. Caller must *not* attempt to free it.
4159 static struct ceph_auth_handshake
*get_authorizer(struct ceph_connection
*con
,
4160 int *proto
, int force_new
)
4162 struct ceph_mds_session
*s
= con
->private;
4163 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4164 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
4165 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
4167 if (force_new
&& auth
->authorizer
) {
4168 ceph_auth_destroy_authorizer(auth
->authorizer
);
4169 auth
->authorizer
= NULL
;
4171 if (!auth
->authorizer
) {
4172 int ret
= ceph_auth_create_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
4175 return ERR_PTR(ret
);
4177 int ret
= ceph_auth_update_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
4180 return ERR_PTR(ret
);
4182 *proto
= ac
->protocol
;
4187 static int add_authorizer_challenge(struct ceph_connection
*con
,
4188 void *challenge_buf
, int challenge_buf_len
)
4190 struct ceph_mds_session
*s
= con
->private;
4191 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4192 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
4194 return ceph_auth_add_authorizer_challenge(ac
, s
->s_auth
.authorizer
,
4195 challenge_buf
, challenge_buf_len
);
4198 static int verify_authorizer_reply(struct ceph_connection
*con
)
4200 struct ceph_mds_session
*s
= con
->private;
4201 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4202 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
4204 return ceph_auth_verify_authorizer_reply(ac
, s
->s_auth
.authorizer
);
4207 static int invalidate_authorizer(struct ceph_connection
*con
)
4209 struct ceph_mds_session
*s
= con
->private;
4210 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4211 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
4213 ceph_auth_invalidate_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
);
4215 return ceph_monc_validate_auth(&mdsc
->fsc
->client
->monc
);
4218 static struct ceph_msg
*mds_alloc_msg(struct ceph_connection
*con
,
4219 struct ceph_msg_header
*hdr
, int *skip
)
4221 struct ceph_msg
*msg
;
4222 int type
= (int) le16_to_cpu(hdr
->type
);
4223 int front_len
= (int) le32_to_cpu(hdr
->front_len
);
4229 msg
= ceph_msg_new(type
, front_len
, GFP_NOFS
, false);
4231 pr_err("unable to allocate msg type %d len %d\n",
4239 static int mds_sign_message(struct ceph_msg
*msg
)
4241 struct ceph_mds_session
*s
= msg
->con
->private;
4242 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
4244 return ceph_auth_sign_message(auth
, msg
);
4247 static int mds_check_message_signature(struct ceph_msg
*msg
)
4249 struct ceph_mds_session
*s
= msg
->con
->private;
4250 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
4252 return ceph_auth_check_message_signature(auth
, msg
);
4255 static const struct ceph_connection_operations mds_con_ops
= {
4258 .dispatch
= dispatch
,
4259 .get_authorizer
= get_authorizer
,
4260 .add_authorizer_challenge
= add_authorizer_challenge
,
4261 .verify_authorizer_reply
= verify_authorizer_reply
,
4262 .invalidate_authorizer
= invalidate_authorizer
,
4263 .peer_reset
= peer_reset
,
4264 .alloc_msg
= mds_alloc_msg
,
4265 .sign_message
= mds_sign_message
,
4266 .check_message_signature
= mds_check_message_signature
,