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45 * are no longer valid.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 info
->pool_ns_len
= 0;
104 info
->pool_ns_data
= NULL
;
105 if (features
& CEPH_FEATURE_FS_FILE_LAYOUT_V2
) {
106 ceph_decode_32_safe(p
, end
, info
->pool_ns_len
, bad
);
107 if (info
->pool_ns_len
> 0) {
108 ceph_decode_need(p
, end
, info
->pool_ns_len
, bad
);
109 info
->pool_ns_data
= *p
;
110 *p
+= info
->pool_ns_len
;
120 * parse a normal reply, which may contain a (dir+)dentry and/or a
123 static int parse_reply_info_trace(void **p
, void *end
,
124 struct ceph_mds_reply_info_parsed
*info
,
129 if (info
->head
->is_dentry
) {
130 err
= parse_reply_info_in(p
, end
, &info
->diri
, features
);
134 if (unlikely(*p
+ sizeof(*info
->dirfrag
) > end
))
137 *p
+= sizeof(*info
->dirfrag
) +
138 sizeof(u32
)*le32_to_cpu(info
->dirfrag
->ndist
);
139 if (unlikely(*p
> end
))
142 ceph_decode_32_safe(p
, end
, info
->dname_len
, bad
);
143 ceph_decode_need(p
, end
, info
->dname_len
, bad
);
145 *p
+= info
->dname_len
;
147 *p
+= sizeof(*info
->dlease
);
150 if (info
->head
->is_target
) {
151 err
= parse_reply_info_in(p
, end
, &info
->targeti
, features
);
156 if (unlikely(*p
!= end
))
163 pr_err("problem parsing mds trace %d\n", err
);
168 * parse readdir results
170 static int parse_reply_info_dir(void **p
, void *end
,
171 struct ceph_mds_reply_info_parsed
*info
,
178 if (*p
+ sizeof(*info
->dir_dir
) > end
)
180 *p
+= sizeof(*info
->dir_dir
) +
181 sizeof(u32
)*le32_to_cpu(info
->dir_dir
->ndist
);
185 ceph_decode_need(p
, end
, sizeof(num
) + 2, bad
);
186 num
= ceph_decode_32(p
);
188 u16 flags
= ceph_decode_16(p
);
189 info
->dir_end
= !!(flags
& CEPH_READDIR_FRAG_END
);
190 info
->dir_complete
= !!(flags
& CEPH_READDIR_FRAG_COMPLETE
);
191 info
->hash_order
= !!(flags
& CEPH_READDIR_HASH_ORDER
);
192 info
->offset_hash
= !!(flags
& CEPH_READDIR_OFFSET_HASH
);
197 BUG_ON(!info
->dir_entries
);
198 if ((unsigned long)(info
->dir_entries
+ num
) >
199 (unsigned long)info
->dir_entries
+ info
->dir_buf_size
) {
200 pr_err("dir contents are larger than expected\n");
207 struct ceph_mds_reply_dir_entry
*rde
= info
->dir_entries
+ i
;
209 ceph_decode_need(p
, end
, sizeof(u32
)*2, bad
);
210 rde
->name_len
= ceph_decode_32(p
);
211 ceph_decode_need(p
, end
, rde
->name_len
, bad
);
214 dout("parsed dir dname '%.*s'\n", rde
->name_len
, rde
->name
);
216 *p
+= sizeof(struct ceph_mds_reply_lease
);
219 err
= parse_reply_info_in(p
, end
, &rde
->inode
, features
);
222 /* ceph_readdir_prepopulate() will update it */
236 pr_err("problem parsing dir contents %d\n", err
);
241 * parse fcntl F_GETLK results
243 static int parse_reply_info_filelock(void **p
, void *end
,
244 struct ceph_mds_reply_info_parsed
*info
,
247 if (*p
+ sizeof(*info
->filelock_reply
) > end
)
250 info
->filelock_reply
= *p
;
251 *p
+= sizeof(*info
->filelock_reply
);
253 if (unlikely(*p
!= end
))
262 * parse create results
264 static int parse_reply_info_create(void **p
, void *end
,
265 struct ceph_mds_reply_info_parsed
*info
,
268 if (features
& CEPH_FEATURE_REPLY_CREATE_INODE
) {
270 info
->has_create_ino
= false;
272 info
->has_create_ino
= true;
273 info
->ino
= ceph_decode_64(p
);
277 if (unlikely(*p
!= end
))
286 * parse extra results
288 static int parse_reply_info_extra(void **p
, void *end
,
289 struct ceph_mds_reply_info_parsed
*info
,
292 u32 op
= le32_to_cpu(info
->head
->op
);
294 if (op
== CEPH_MDS_OP_GETFILELOCK
)
295 return parse_reply_info_filelock(p
, end
, info
, features
);
296 else if (op
== CEPH_MDS_OP_READDIR
|| op
== CEPH_MDS_OP_LSSNAP
)
297 return parse_reply_info_dir(p
, end
, info
, features
);
298 else if (op
== CEPH_MDS_OP_CREATE
)
299 return parse_reply_info_create(p
, end
, info
, features
);
305 * parse entire mds reply
307 static int parse_reply_info(struct ceph_msg
*msg
,
308 struct ceph_mds_reply_info_parsed
*info
,
315 info
->head
= msg
->front
.iov_base
;
316 p
= msg
->front
.iov_base
+ sizeof(struct ceph_mds_reply_head
);
317 end
= p
+ msg
->front
.iov_len
- sizeof(struct ceph_mds_reply_head
);
320 ceph_decode_32_safe(&p
, end
, len
, bad
);
322 ceph_decode_need(&p
, end
, len
, bad
);
323 err
= parse_reply_info_trace(&p
, p
+len
, info
, features
);
329 ceph_decode_32_safe(&p
, end
, len
, bad
);
331 ceph_decode_need(&p
, end
, len
, bad
);
332 err
= parse_reply_info_extra(&p
, p
+len
, info
, features
);
338 ceph_decode_32_safe(&p
, end
, len
, bad
);
339 info
->snapblob_len
= len
;
350 pr_err("mds parse_reply err %d\n", err
);
354 static void destroy_reply_info(struct ceph_mds_reply_info_parsed
*info
)
356 if (!info
->dir_entries
)
358 free_pages((unsigned long)info
->dir_entries
, get_order(info
->dir_buf_size
));
365 const char *ceph_session_state_name(int s
)
368 case CEPH_MDS_SESSION_NEW
: return "new";
369 case CEPH_MDS_SESSION_OPENING
: return "opening";
370 case CEPH_MDS_SESSION_OPEN
: return "open";
371 case CEPH_MDS_SESSION_HUNG
: return "hung";
372 case CEPH_MDS_SESSION_CLOSING
: return "closing";
373 case CEPH_MDS_SESSION_RESTARTING
: return "restarting";
374 case CEPH_MDS_SESSION_RECONNECTING
: return "reconnecting";
375 case CEPH_MDS_SESSION_REJECTED
: return "rejected";
376 default: return "???";
380 static struct ceph_mds_session
*get_session(struct ceph_mds_session
*s
)
382 if (refcount_inc_not_zero(&s
->s_ref
)) {
383 dout("mdsc get_session %p %d -> %d\n", s
,
384 refcount_read(&s
->s_ref
)-1, refcount_read(&s
->s_ref
));
387 dout("mdsc get_session %p 0 -- FAIL", s
);
392 void ceph_put_mds_session(struct ceph_mds_session
*s
)
394 dout("mdsc put_session %p %d -> %d\n", s
,
395 refcount_read(&s
->s_ref
), refcount_read(&s
->s_ref
)-1);
396 if (refcount_dec_and_test(&s
->s_ref
)) {
397 if (s
->s_auth
.authorizer
)
398 ceph_auth_destroy_authorizer(s
->s_auth
.authorizer
);
404 * called under mdsc->mutex
406 struct ceph_mds_session
*__ceph_lookup_mds_session(struct ceph_mds_client
*mdsc
,
409 struct ceph_mds_session
*session
;
411 if (mds
>= mdsc
->max_sessions
|| !mdsc
->sessions
[mds
])
413 session
= mdsc
->sessions
[mds
];
414 dout("lookup_mds_session %p %d\n", session
,
415 refcount_read(&session
->s_ref
));
416 get_session(session
);
420 static bool __have_session(struct ceph_mds_client
*mdsc
, int mds
)
422 if (mds
>= mdsc
->max_sessions
)
424 return mdsc
->sessions
[mds
];
427 static int __verify_registered_session(struct ceph_mds_client
*mdsc
,
428 struct ceph_mds_session
*s
)
430 if (s
->s_mds
>= mdsc
->max_sessions
||
431 mdsc
->sessions
[s
->s_mds
] != s
)
437 * create+register a new session for given mds.
438 * called under mdsc->mutex.
440 static struct ceph_mds_session
*register_session(struct ceph_mds_client
*mdsc
,
443 struct ceph_mds_session
*s
;
445 if (mds
>= mdsc
->mdsmap
->m_num_mds
)
446 return ERR_PTR(-EINVAL
);
448 s
= kzalloc(sizeof(*s
), GFP_NOFS
);
450 return ERR_PTR(-ENOMEM
);
453 s
->s_state
= CEPH_MDS_SESSION_NEW
;
456 mutex_init(&s
->s_mutex
);
458 ceph_con_init(&s
->s_con
, s
, &mds_con_ops
, &mdsc
->fsc
->client
->msgr
);
460 spin_lock_init(&s
->s_gen_ttl_lock
);
462 s
->s_cap_ttl
= jiffies
- 1;
464 spin_lock_init(&s
->s_cap_lock
);
465 s
->s_renew_requested
= 0;
467 INIT_LIST_HEAD(&s
->s_caps
);
470 refcount_set(&s
->s_ref
, 1);
471 INIT_LIST_HEAD(&s
->s_waiting
);
472 INIT_LIST_HEAD(&s
->s_unsafe
);
473 s
->s_num_cap_releases
= 0;
474 s
->s_cap_reconnect
= 0;
475 s
->s_cap_iterator
= NULL
;
476 INIT_LIST_HEAD(&s
->s_cap_releases
);
477 INIT_LIST_HEAD(&s
->s_cap_flushing
);
479 dout("register_session mds%d\n", mds
);
480 if (mds
>= mdsc
->max_sessions
) {
481 int newmax
= 1 << get_count_order(mds
+1);
482 struct ceph_mds_session
**sa
;
484 dout("register_session realloc to %d\n", newmax
);
485 sa
= kcalloc(newmax
, sizeof(void *), GFP_NOFS
);
488 if (mdsc
->sessions
) {
489 memcpy(sa
, mdsc
->sessions
,
490 mdsc
->max_sessions
* sizeof(void *));
491 kfree(mdsc
->sessions
);
494 mdsc
->max_sessions
= newmax
;
496 mdsc
->sessions
[mds
] = s
;
497 atomic_inc(&mdsc
->num_sessions
);
498 refcount_inc(&s
->s_ref
); /* one ref to sessions[], one to caller */
500 ceph_con_open(&s
->s_con
, CEPH_ENTITY_TYPE_MDS
, mds
,
501 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
507 return ERR_PTR(-ENOMEM
);
511 * called under mdsc->mutex
513 static void __unregister_session(struct ceph_mds_client
*mdsc
,
514 struct ceph_mds_session
*s
)
516 dout("__unregister_session mds%d %p\n", s
->s_mds
, s
);
517 BUG_ON(mdsc
->sessions
[s
->s_mds
] != s
);
518 mdsc
->sessions
[s
->s_mds
] = NULL
;
519 ceph_con_close(&s
->s_con
);
520 ceph_put_mds_session(s
);
521 atomic_dec(&mdsc
->num_sessions
);
525 * drop session refs in request.
527 * should be last request ref, or hold mdsc->mutex
529 static void put_request_session(struct ceph_mds_request
*req
)
531 if (req
->r_session
) {
532 ceph_put_mds_session(req
->r_session
);
533 req
->r_session
= NULL
;
537 void ceph_mdsc_release_request(struct kref
*kref
)
539 struct ceph_mds_request
*req
= container_of(kref
,
540 struct ceph_mds_request
,
542 destroy_reply_info(&req
->r_reply_info
);
544 ceph_msg_put(req
->r_request
);
546 ceph_msg_put(req
->r_reply
);
548 ceph_put_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
552 ceph_put_cap_refs(ceph_inode(req
->r_parent
), CEPH_CAP_PIN
);
553 iput(req
->r_target_inode
);
556 if (req
->r_old_dentry
)
557 dput(req
->r_old_dentry
);
558 if (req
->r_old_dentry_dir
) {
560 * track (and drop pins for) r_old_dentry_dir
561 * separately, since r_old_dentry's d_parent may have
562 * changed between the dir mutex being dropped and
563 * this request being freed.
565 ceph_put_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
567 iput(req
->r_old_dentry_dir
);
572 ceph_pagelist_release(req
->r_pagelist
);
573 put_request_session(req
);
574 ceph_unreserve_caps(req
->r_mdsc
, &req
->r_caps_reservation
);
578 DEFINE_RB_FUNCS(request
, struct ceph_mds_request
, r_tid
, r_node
)
581 * lookup session, bump ref if found.
583 * called under mdsc->mutex.
585 static struct ceph_mds_request
*
586 lookup_get_request(struct ceph_mds_client
*mdsc
, u64 tid
)
588 struct ceph_mds_request
*req
;
590 req
= lookup_request(&mdsc
->request_tree
, tid
);
592 ceph_mdsc_get_request(req
);
598 * Register an in-flight request, and assign a tid. Link to directory
599 * are modifying (if any).
601 * Called under mdsc->mutex.
603 static void __register_request(struct ceph_mds_client
*mdsc
,
604 struct ceph_mds_request
*req
,
607 req
->r_tid
= ++mdsc
->last_tid
;
609 ceph_reserve_caps(mdsc
, &req
->r_caps_reservation
,
611 dout("__register_request %p tid %lld\n", req
, req
->r_tid
);
612 ceph_mdsc_get_request(req
);
613 insert_request(&mdsc
->request_tree
, req
);
615 req
->r_uid
= current_fsuid();
616 req
->r_gid
= current_fsgid();
618 if (mdsc
->oldest_tid
== 0 && req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
)
619 mdsc
->oldest_tid
= req
->r_tid
;
623 req
->r_unsafe_dir
= dir
;
627 static void __unregister_request(struct ceph_mds_client
*mdsc
,
628 struct ceph_mds_request
*req
)
630 dout("__unregister_request %p tid %lld\n", req
, req
->r_tid
);
632 /* Never leave an unregistered request on an unsafe list! */
633 list_del_init(&req
->r_unsafe_item
);
635 if (req
->r_tid
== mdsc
->oldest_tid
) {
636 struct rb_node
*p
= rb_next(&req
->r_node
);
637 mdsc
->oldest_tid
= 0;
639 struct ceph_mds_request
*next_req
=
640 rb_entry(p
, struct ceph_mds_request
, r_node
);
641 if (next_req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
) {
642 mdsc
->oldest_tid
= next_req
->r_tid
;
649 erase_request(&mdsc
->request_tree
, req
);
651 if (req
->r_unsafe_dir
&&
652 test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
)) {
653 struct ceph_inode_info
*ci
= ceph_inode(req
->r_unsafe_dir
);
654 spin_lock(&ci
->i_unsafe_lock
);
655 list_del_init(&req
->r_unsafe_dir_item
);
656 spin_unlock(&ci
->i_unsafe_lock
);
658 if (req
->r_target_inode
&&
659 test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
)) {
660 struct ceph_inode_info
*ci
= ceph_inode(req
->r_target_inode
);
661 spin_lock(&ci
->i_unsafe_lock
);
662 list_del_init(&req
->r_unsafe_target_item
);
663 spin_unlock(&ci
->i_unsafe_lock
);
666 if (req
->r_unsafe_dir
) {
667 iput(req
->r_unsafe_dir
);
668 req
->r_unsafe_dir
= NULL
;
671 complete_all(&req
->r_safe_completion
);
673 ceph_mdsc_put_request(req
);
677 * Walk back up the dentry tree until we hit a dentry representing a
678 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
679 * when calling this) to ensure that the objects won't disappear while we're
680 * working with them. Once we hit a candidate dentry, we attempt to take a
681 * reference to it, and return that as the result.
683 static struct inode
*get_nonsnap_parent(struct dentry
*dentry
)
685 struct inode
*inode
= NULL
;
687 while (dentry
&& !IS_ROOT(dentry
)) {
688 inode
= d_inode_rcu(dentry
);
689 if (!inode
|| ceph_snap(inode
) == CEPH_NOSNAP
)
691 dentry
= dentry
->d_parent
;
694 inode
= igrab(inode
);
699 * Choose mds to send request to next. If there is a hint set in the
700 * request (e.g., due to a prior forward hint from the mds), use that.
701 * Otherwise, consult frag tree and/or caps to identify the
702 * appropriate mds. If all else fails, choose randomly.
704 * Called under mdsc->mutex.
706 static int __choose_mds(struct ceph_mds_client
*mdsc
,
707 struct ceph_mds_request
*req
)
710 struct ceph_inode_info
*ci
;
711 struct ceph_cap
*cap
;
712 int mode
= req
->r_direct_mode
;
714 u32 hash
= req
->r_direct_hash
;
715 bool is_hash
= test_bit(CEPH_MDS_R_DIRECT_IS_HASH
, &req
->r_req_flags
);
718 * is there a specific mds we should try? ignore hint if we have
719 * no session and the mds is not up (active or recovering).
721 if (req
->r_resend_mds
>= 0 &&
722 (__have_session(mdsc
, req
->r_resend_mds
) ||
723 ceph_mdsmap_get_state(mdsc
->mdsmap
, req
->r_resend_mds
) > 0)) {
724 dout("choose_mds using resend_mds mds%d\n",
726 return req
->r_resend_mds
;
729 if (mode
== USE_RANDOM_MDS
)
734 if (ceph_snap(req
->r_inode
) != CEPH_SNAPDIR
) {
735 inode
= req
->r_inode
;
738 /* req->r_dentry is non-null for LSSNAP request */
740 inode
= get_nonsnap_parent(req
->r_dentry
);
742 dout("__choose_mds using snapdir's parent %p\n", inode
);
744 } else if (req
->r_dentry
) {
745 /* ignore race with rename; old or new d_parent is okay */
746 struct dentry
*parent
;
750 parent
= req
->r_dentry
->d_parent
;
751 dir
= req
->r_parent
? : d_inode_rcu(parent
);
753 if (!dir
|| dir
->i_sb
!= mdsc
->fsc
->sb
) {
754 /* not this fs or parent went negative */
755 inode
= d_inode(req
->r_dentry
);
758 } else if (ceph_snap(dir
) != CEPH_NOSNAP
) {
759 /* direct snapped/virtual snapdir requests
760 * based on parent dir inode */
761 inode
= get_nonsnap_parent(parent
);
762 dout("__choose_mds using nonsnap parent %p\n", inode
);
765 inode
= d_inode(req
->r_dentry
);
766 if (!inode
|| mode
== USE_AUTH_MDS
) {
769 hash
= ceph_dentry_hash(dir
, req
->r_dentry
);
778 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode
, (int)is_hash
,
782 ci
= ceph_inode(inode
);
784 if (is_hash
&& S_ISDIR(inode
->i_mode
)) {
785 struct ceph_inode_frag frag
;
788 ceph_choose_frag(ci
, hash
, &frag
, &found
);
790 if (mode
== USE_ANY_MDS
&& frag
.ndist
> 0) {
793 /* choose a random replica */
794 get_random_bytes(&r
, 1);
797 dout("choose_mds %p %llx.%llx "
798 "frag %u mds%d (%d/%d)\n",
799 inode
, ceph_vinop(inode
),
802 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
803 CEPH_MDS_STATE_ACTIVE
)
807 /* since this file/dir wasn't known to be
808 * replicated, then we want to look for the
809 * authoritative mds. */
812 /* choose auth mds */
814 dout("choose_mds %p %llx.%llx "
815 "frag %u mds%d (auth)\n",
816 inode
, ceph_vinop(inode
), frag
.frag
, mds
);
817 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
818 CEPH_MDS_STATE_ACTIVE
)
824 spin_lock(&ci
->i_ceph_lock
);
826 if (mode
== USE_AUTH_MDS
)
827 cap
= ci
->i_auth_cap
;
828 if (!cap
&& !RB_EMPTY_ROOT(&ci
->i_caps
))
829 cap
= rb_entry(rb_first(&ci
->i_caps
), struct ceph_cap
, ci_node
);
831 spin_unlock(&ci
->i_ceph_lock
);
835 mds
= cap
->session
->s_mds
;
836 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
837 inode
, ceph_vinop(inode
), mds
,
838 cap
== ci
->i_auth_cap
? "auth " : "", cap
);
839 spin_unlock(&ci
->i_ceph_lock
);
845 mds
= ceph_mdsmap_get_random_mds(mdsc
->mdsmap
);
846 dout("choose_mds chose random mds%d\n", mds
);
854 static struct ceph_msg
*create_session_msg(u32 op
, u64 seq
)
856 struct ceph_msg
*msg
;
857 struct ceph_mds_session_head
*h
;
859 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
), GFP_NOFS
,
862 pr_err("create_session_msg ENOMEM creating msg\n");
865 h
= msg
->front
.iov_base
;
866 h
->op
= cpu_to_le32(op
);
867 h
->seq
= cpu_to_le64(seq
);
873 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
874 * to include additional client metadata fields.
876 static struct ceph_msg
*create_session_open_msg(struct ceph_mds_client
*mdsc
, u64 seq
)
878 struct ceph_msg
*msg
;
879 struct ceph_mds_session_head
*h
;
881 int metadata_bytes
= 0;
882 int metadata_key_count
= 0;
883 struct ceph_options
*opt
= mdsc
->fsc
->client
->options
;
884 struct ceph_mount_options
*fsopt
= mdsc
->fsc
->mount_options
;
887 const char* metadata
[][2] = {
888 {"hostname", mdsc
->nodename
},
889 {"kernel_version", init_utsname()->release
},
890 {"entity_id", opt
->name
? : ""},
891 {"root", fsopt
->server_path
? : "/"},
895 /* Calculate serialized length of metadata */
896 metadata_bytes
= 4; /* map length */
897 for (i
= 0; metadata
[i
][0]; ++i
) {
898 metadata_bytes
+= 8 + strlen(metadata
[i
][0]) +
899 strlen(metadata
[i
][1]);
900 metadata_key_count
++;
903 /* Allocate the message */
904 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
) + metadata_bytes
,
907 pr_err("create_session_msg ENOMEM creating msg\n");
910 h
= msg
->front
.iov_base
;
911 h
->op
= cpu_to_le32(CEPH_SESSION_REQUEST_OPEN
);
912 h
->seq
= cpu_to_le64(seq
);
915 * Serialize client metadata into waiting buffer space, using
916 * the format that userspace expects for map<string, string>
918 * ClientSession messages with metadata are v2
920 msg
->hdr
.version
= cpu_to_le16(2);
921 msg
->hdr
.compat_version
= cpu_to_le16(1);
923 /* The write pointer, following the session_head structure */
924 p
= msg
->front
.iov_base
+ sizeof(*h
);
926 /* Number of entries in the map */
927 ceph_encode_32(&p
, metadata_key_count
);
929 /* Two length-prefixed strings for each entry in the map */
930 for (i
= 0; metadata
[i
][0]; ++i
) {
931 size_t const key_len
= strlen(metadata
[i
][0]);
932 size_t const val_len
= strlen(metadata
[i
][1]);
934 ceph_encode_32(&p
, key_len
);
935 memcpy(p
, metadata
[i
][0], key_len
);
937 ceph_encode_32(&p
, val_len
);
938 memcpy(p
, metadata
[i
][1], val_len
);
946 * send session open request.
948 * called under mdsc->mutex
950 static int __open_session(struct ceph_mds_client
*mdsc
,
951 struct ceph_mds_session
*session
)
953 struct ceph_msg
*msg
;
955 int mds
= session
->s_mds
;
957 /* wait for mds to go active? */
958 mstate
= ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
);
959 dout("open_session to mds%d (%s)\n", mds
,
960 ceph_mds_state_name(mstate
));
961 session
->s_state
= CEPH_MDS_SESSION_OPENING
;
962 session
->s_renew_requested
= jiffies
;
964 /* send connect message */
965 msg
= create_session_open_msg(mdsc
, session
->s_seq
);
968 ceph_con_send(&session
->s_con
, msg
);
973 * open sessions for any export targets for the given mds
975 * called under mdsc->mutex
977 static struct ceph_mds_session
*
978 __open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
980 struct ceph_mds_session
*session
;
982 session
= __ceph_lookup_mds_session(mdsc
, target
);
984 session
= register_session(mdsc
, target
);
988 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
989 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
990 __open_session(mdsc
, session
);
995 struct ceph_mds_session
*
996 ceph_mdsc_open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
998 struct ceph_mds_session
*session
;
1000 dout("open_export_target_session to mds%d\n", target
);
1002 mutex_lock(&mdsc
->mutex
);
1003 session
= __open_export_target_session(mdsc
, target
);
1004 mutex_unlock(&mdsc
->mutex
);
1009 static void __open_export_target_sessions(struct ceph_mds_client
*mdsc
,
1010 struct ceph_mds_session
*session
)
1012 struct ceph_mds_info
*mi
;
1013 struct ceph_mds_session
*ts
;
1014 int i
, mds
= session
->s_mds
;
1016 if (mds
>= mdsc
->mdsmap
->m_num_mds
)
1019 mi
= &mdsc
->mdsmap
->m_info
[mds
];
1020 dout("open_export_target_sessions for mds%d (%d targets)\n",
1021 session
->s_mds
, mi
->num_export_targets
);
1023 for (i
= 0; i
< mi
->num_export_targets
; i
++) {
1024 ts
= __open_export_target_session(mdsc
, mi
->export_targets
[i
]);
1026 ceph_put_mds_session(ts
);
1030 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client
*mdsc
,
1031 struct ceph_mds_session
*session
)
1033 mutex_lock(&mdsc
->mutex
);
1034 __open_export_target_sessions(mdsc
, session
);
1035 mutex_unlock(&mdsc
->mutex
);
1042 static void detach_cap_releases(struct ceph_mds_session
*session
,
1043 struct list_head
*target
)
1045 lockdep_assert_held(&session
->s_cap_lock
);
1047 list_splice_init(&session
->s_cap_releases
, target
);
1048 session
->s_num_cap_releases
= 0;
1049 dout("dispose_cap_releases mds%d\n", session
->s_mds
);
1052 static void dispose_cap_releases(struct ceph_mds_client
*mdsc
,
1053 struct list_head
*dispose
)
1055 while (!list_empty(dispose
)) {
1056 struct ceph_cap
*cap
;
1057 /* zero out the in-progress message */
1058 cap
= list_first_entry(dispose
, struct ceph_cap
, session_caps
);
1059 list_del(&cap
->session_caps
);
1060 ceph_put_cap(mdsc
, cap
);
1064 static void cleanup_session_requests(struct ceph_mds_client
*mdsc
,
1065 struct ceph_mds_session
*session
)
1067 struct ceph_mds_request
*req
;
1070 dout("cleanup_session_requests mds%d\n", session
->s_mds
);
1071 mutex_lock(&mdsc
->mutex
);
1072 while (!list_empty(&session
->s_unsafe
)) {
1073 req
= list_first_entry(&session
->s_unsafe
,
1074 struct ceph_mds_request
, r_unsafe_item
);
1075 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1077 __unregister_request(mdsc
, req
);
1079 /* zero r_attempts, so kick_requests() will re-send requests */
1080 p
= rb_first(&mdsc
->request_tree
);
1082 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
1084 if (req
->r_session
&&
1085 req
->r_session
->s_mds
== session
->s_mds
)
1086 req
->r_attempts
= 0;
1088 mutex_unlock(&mdsc
->mutex
);
1092 * Helper to safely iterate over all caps associated with a session, with
1093 * special care taken to handle a racing __ceph_remove_cap().
1095 * Caller must hold session s_mutex.
1097 static int iterate_session_caps(struct ceph_mds_session
*session
,
1098 int (*cb
)(struct inode
*, struct ceph_cap
*,
1101 struct list_head
*p
;
1102 struct ceph_cap
*cap
;
1103 struct inode
*inode
, *last_inode
= NULL
;
1104 struct ceph_cap
*old_cap
= NULL
;
1107 dout("iterate_session_caps %p mds%d\n", session
, session
->s_mds
);
1108 spin_lock(&session
->s_cap_lock
);
1109 p
= session
->s_caps
.next
;
1110 while (p
!= &session
->s_caps
) {
1111 cap
= list_entry(p
, struct ceph_cap
, session_caps
);
1112 inode
= igrab(&cap
->ci
->vfs_inode
);
1117 session
->s_cap_iterator
= cap
;
1118 spin_unlock(&session
->s_cap_lock
);
1125 ceph_put_cap(session
->s_mdsc
, old_cap
);
1129 ret
= cb(inode
, cap
, arg
);
1132 spin_lock(&session
->s_cap_lock
);
1135 dout("iterate_session_caps finishing cap %p removal\n",
1137 BUG_ON(cap
->session
!= session
);
1138 cap
->session
= NULL
;
1139 list_del_init(&cap
->session_caps
);
1140 session
->s_nr_caps
--;
1141 if (cap
->queue_release
) {
1142 list_add_tail(&cap
->session_caps
,
1143 &session
->s_cap_releases
);
1144 session
->s_num_cap_releases
++;
1146 old_cap
= cap
; /* put_cap it w/o locks held */
1154 session
->s_cap_iterator
= NULL
;
1155 spin_unlock(&session
->s_cap_lock
);
1159 ceph_put_cap(session
->s_mdsc
, old_cap
);
1164 static int remove_session_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1167 struct ceph_fs_client
*fsc
= (struct ceph_fs_client
*)arg
;
1168 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1169 LIST_HEAD(to_remove
);
1171 bool invalidate
= false;
1173 dout("removing cap %p, ci is %p, inode is %p\n",
1174 cap
, ci
, &ci
->vfs_inode
);
1175 spin_lock(&ci
->i_ceph_lock
);
1176 __ceph_remove_cap(cap
, false);
1177 if (!ci
->i_auth_cap
) {
1178 struct ceph_cap_flush
*cf
;
1179 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
1181 ci
->i_ceph_flags
|= CEPH_I_CAP_DROPPED
;
1183 if (ci
->i_wrbuffer_ref
> 0 &&
1184 READ_ONCE(fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
1187 while (!list_empty(&ci
->i_cap_flush_list
)) {
1188 cf
= list_first_entry(&ci
->i_cap_flush_list
,
1189 struct ceph_cap_flush
, i_list
);
1190 list_move(&cf
->i_list
, &to_remove
);
1193 spin_lock(&mdsc
->cap_dirty_lock
);
1195 list_for_each_entry(cf
, &to_remove
, i_list
)
1196 list_del(&cf
->g_list
);
1198 if (!list_empty(&ci
->i_dirty_item
)) {
1199 pr_warn_ratelimited(
1200 " dropping dirty %s state for %p %lld\n",
1201 ceph_cap_string(ci
->i_dirty_caps
),
1202 inode
, ceph_ino(inode
));
1203 ci
->i_dirty_caps
= 0;
1204 list_del_init(&ci
->i_dirty_item
);
1207 if (!list_empty(&ci
->i_flushing_item
)) {
1208 pr_warn_ratelimited(
1209 " dropping dirty+flushing %s state for %p %lld\n",
1210 ceph_cap_string(ci
->i_flushing_caps
),
1211 inode
, ceph_ino(inode
));
1212 ci
->i_flushing_caps
= 0;
1213 list_del_init(&ci
->i_flushing_item
);
1214 mdsc
->num_cap_flushing
--;
1217 spin_unlock(&mdsc
->cap_dirty_lock
);
1219 if (atomic_read(&ci
->i_filelock_ref
) > 0) {
1220 /* make further file lock syscall return -EIO */
1221 ci
->i_ceph_flags
|= CEPH_I_ERROR_FILELOCK
;
1222 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1223 inode
, ceph_ino(inode
));
1226 if (!ci
->i_dirty_caps
&& ci
->i_prealloc_cap_flush
) {
1227 list_add(&ci
->i_prealloc_cap_flush
->i_list
, &to_remove
);
1228 ci
->i_prealloc_cap_flush
= NULL
;
1231 spin_unlock(&ci
->i_ceph_lock
);
1232 while (!list_empty(&to_remove
)) {
1233 struct ceph_cap_flush
*cf
;
1234 cf
= list_first_entry(&to_remove
,
1235 struct ceph_cap_flush
, i_list
);
1236 list_del(&cf
->i_list
);
1237 ceph_free_cap_flush(cf
);
1240 wake_up_all(&ci
->i_cap_wq
);
1242 ceph_queue_invalidate(inode
);
1249 * caller must hold session s_mutex
1251 static void remove_session_caps(struct ceph_mds_session
*session
)
1253 struct ceph_fs_client
*fsc
= session
->s_mdsc
->fsc
;
1254 struct super_block
*sb
= fsc
->sb
;
1257 dout("remove_session_caps on %p\n", session
);
1258 iterate_session_caps(session
, remove_session_caps_cb
, fsc
);
1260 wake_up_all(&fsc
->mdsc
->cap_flushing_wq
);
1262 spin_lock(&session
->s_cap_lock
);
1263 if (session
->s_nr_caps
> 0) {
1264 struct inode
*inode
;
1265 struct ceph_cap
*cap
, *prev
= NULL
;
1266 struct ceph_vino vino
;
1268 * iterate_session_caps() skips inodes that are being
1269 * deleted, we need to wait until deletions are complete.
1270 * __wait_on_freeing_inode() is designed for the job,
1271 * but it is not exported, so use lookup inode function
1274 while (!list_empty(&session
->s_caps
)) {
1275 cap
= list_entry(session
->s_caps
.next
,
1276 struct ceph_cap
, session_caps
);
1280 vino
= cap
->ci
->i_vino
;
1281 spin_unlock(&session
->s_cap_lock
);
1283 inode
= ceph_find_inode(sb
, vino
);
1286 spin_lock(&session
->s_cap_lock
);
1290 // drop cap expires and unlock s_cap_lock
1291 detach_cap_releases(session
, &dispose
);
1293 BUG_ON(session
->s_nr_caps
> 0);
1294 BUG_ON(!list_empty(&session
->s_cap_flushing
));
1295 spin_unlock(&session
->s_cap_lock
);
1296 dispose_cap_releases(session
->s_mdsc
, &dispose
);
1300 * wake up any threads waiting on this session's caps. if the cap is
1301 * old (didn't get renewed on the client reconnect), remove it now.
1303 * caller must hold s_mutex.
1305 static int wake_up_session_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1308 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1311 spin_lock(&ci
->i_ceph_lock
);
1312 ci
->i_wanted_max_size
= 0;
1313 ci
->i_requested_max_size
= 0;
1314 spin_unlock(&ci
->i_ceph_lock
);
1316 wake_up_all(&ci
->i_cap_wq
);
1320 static void wake_up_session_caps(struct ceph_mds_session
*session
,
1323 dout("wake_up_session_caps %p mds%d\n", session
, session
->s_mds
);
1324 iterate_session_caps(session
, wake_up_session_cb
,
1325 (void *)(unsigned long)reconnect
);
1329 * Send periodic message to MDS renewing all currently held caps. The
1330 * ack will reset the expiration for all caps from this session.
1332 * caller holds s_mutex
1334 static int send_renew_caps(struct ceph_mds_client
*mdsc
,
1335 struct ceph_mds_session
*session
)
1337 struct ceph_msg
*msg
;
1340 if (time_after_eq(jiffies
, session
->s_cap_ttl
) &&
1341 time_after_eq(session
->s_cap_ttl
, session
->s_renew_requested
))
1342 pr_info("mds%d caps stale\n", session
->s_mds
);
1343 session
->s_renew_requested
= jiffies
;
1345 /* do not try to renew caps until a recovering mds has reconnected
1346 * with its clients. */
1347 state
= ceph_mdsmap_get_state(mdsc
->mdsmap
, session
->s_mds
);
1348 if (state
< CEPH_MDS_STATE_RECONNECT
) {
1349 dout("send_renew_caps ignoring mds%d (%s)\n",
1350 session
->s_mds
, ceph_mds_state_name(state
));
1354 dout("send_renew_caps to mds%d (%s)\n", session
->s_mds
,
1355 ceph_mds_state_name(state
));
1356 msg
= create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS
,
1357 ++session
->s_renew_seq
);
1360 ceph_con_send(&session
->s_con
, msg
);
1364 static int send_flushmsg_ack(struct ceph_mds_client
*mdsc
,
1365 struct ceph_mds_session
*session
, u64 seq
)
1367 struct ceph_msg
*msg
;
1369 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1370 session
->s_mds
, ceph_session_state_name(session
->s_state
), seq
);
1371 msg
= create_session_msg(CEPH_SESSION_FLUSHMSG_ACK
, seq
);
1374 ceph_con_send(&session
->s_con
, msg
);
1380 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1382 * Called under session->s_mutex
1384 static void renewed_caps(struct ceph_mds_client
*mdsc
,
1385 struct ceph_mds_session
*session
, int is_renew
)
1390 spin_lock(&session
->s_cap_lock
);
1391 was_stale
= is_renew
&& time_after_eq(jiffies
, session
->s_cap_ttl
);
1393 session
->s_cap_ttl
= session
->s_renew_requested
+
1394 mdsc
->mdsmap
->m_session_timeout
*HZ
;
1397 if (time_before(jiffies
, session
->s_cap_ttl
)) {
1398 pr_info("mds%d caps renewed\n", session
->s_mds
);
1401 pr_info("mds%d caps still stale\n", session
->s_mds
);
1404 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1405 session
->s_mds
, session
->s_cap_ttl
, was_stale
? "stale" : "fresh",
1406 time_before(jiffies
, session
->s_cap_ttl
) ? "stale" : "fresh");
1407 spin_unlock(&session
->s_cap_lock
);
1410 wake_up_session_caps(session
, 0);
1414 * send a session close request
1416 static int request_close_session(struct ceph_mds_client
*mdsc
,
1417 struct ceph_mds_session
*session
)
1419 struct ceph_msg
*msg
;
1421 dout("request_close_session mds%d state %s seq %lld\n",
1422 session
->s_mds
, ceph_session_state_name(session
->s_state
),
1424 msg
= create_session_msg(CEPH_SESSION_REQUEST_CLOSE
, session
->s_seq
);
1427 ceph_con_send(&session
->s_con
, msg
);
1432 * Called with s_mutex held.
1434 static int __close_session(struct ceph_mds_client
*mdsc
,
1435 struct ceph_mds_session
*session
)
1437 if (session
->s_state
>= CEPH_MDS_SESSION_CLOSING
)
1439 session
->s_state
= CEPH_MDS_SESSION_CLOSING
;
1440 return request_close_session(mdsc
, session
);
1443 static bool drop_negative_children(struct dentry
*dentry
)
1445 struct dentry
*child
;
1446 bool all_negative
= true;
1448 if (!d_is_dir(dentry
))
1451 spin_lock(&dentry
->d_lock
);
1452 list_for_each_entry(child
, &dentry
->d_subdirs
, d_child
) {
1453 if (d_really_is_positive(child
)) {
1454 all_negative
= false;
1458 spin_unlock(&dentry
->d_lock
);
1461 shrink_dcache_parent(dentry
);
1463 return all_negative
;
1467 * Trim old(er) caps.
1469 * Because we can't cache an inode without one or more caps, we do
1470 * this indirectly: if a cap is unused, we prune its aliases, at which
1471 * point the inode will hopefully get dropped to.
1473 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1474 * memory pressure from the MDS, though, so it needn't be perfect.
1476 static int trim_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
, void *arg
)
1478 struct ceph_mds_session
*session
= arg
;
1479 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1480 int used
, wanted
, oissued
, mine
;
1482 if (session
->s_trim_caps
<= 0)
1485 spin_lock(&ci
->i_ceph_lock
);
1486 mine
= cap
->issued
| cap
->implemented
;
1487 used
= __ceph_caps_used(ci
);
1488 wanted
= __ceph_caps_file_wanted(ci
);
1489 oissued
= __ceph_caps_issued_other(ci
, cap
);
1491 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1492 inode
, cap
, ceph_cap_string(mine
), ceph_cap_string(oissued
),
1493 ceph_cap_string(used
), ceph_cap_string(wanted
));
1494 if (cap
== ci
->i_auth_cap
) {
1495 if (ci
->i_dirty_caps
|| ci
->i_flushing_caps
||
1496 !list_empty(&ci
->i_cap_snaps
))
1498 if ((used
| wanted
) & CEPH_CAP_ANY_WR
)
1500 /* Note: it's possible that i_filelock_ref becomes non-zero
1501 * after dropping auth caps. It doesn't hurt because reply
1502 * of lock mds request will re-add auth caps. */
1503 if (atomic_read(&ci
->i_filelock_ref
) > 0)
1506 /* The inode has cached pages, but it's no longer used.
1507 * we can safely drop it */
1508 if (wanted
== 0 && used
== CEPH_CAP_FILE_CACHE
&&
1509 !(oissued
& CEPH_CAP_FILE_CACHE
)) {
1513 if ((used
| wanted
) & ~oissued
& mine
)
1514 goto out
; /* we need these caps */
1517 /* we aren't the only cap.. just remove us */
1518 __ceph_remove_cap(cap
, true);
1519 session
->s_trim_caps
--;
1521 struct dentry
*dentry
;
1522 /* try dropping referring dentries */
1523 spin_unlock(&ci
->i_ceph_lock
);
1524 dentry
= d_find_any_alias(inode
);
1525 if (dentry
&& drop_negative_children(dentry
)) {
1528 d_prune_aliases(inode
);
1529 count
= atomic_read(&inode
->i_count
);
1531 session
->s_trim_caps
--;
1532 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1541 spin_unlock(&ci
->i_ceph_lock
);
1546 * Trim session cap count down to some max number.
1548 static int trim_caps(struct ceph_mds_client
*mdsc
,
1549 struct ceph_mds_session
*session
,
1552 int trim_caps
= session
->s_nr_caps
- max_caps
;
1554 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1555 session
->s_mds
, session
->s_nr_caps
, max_caps
, trim_caps
);
1556 if (trim_caps
> 0) {
1557 session
->s_trim_caps
= trim_caps
;
1558 iterate_session_caps(session
, trim_caps_cb
, session
);
1559 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1560 session
->s_mds
, session
->s_nr_caps
, max_caps
,
1561 trim_caps
- session
->s_trim_caps
);
1562 session
->s_trim_caps
= 0;
1565 ceph_send_cap_releases(mdsc
, session
);
1569 static int check_caps_flush(struct ceph_mds_client
*mdsc
,
1574 spin_lock(&mdsc
->cap_dirty_lock
);
1575 if (!list_empty(&mdsc
->cap_flush_list
)) {
1576 struct ceph_cap_flush
*cf
=
1577 list_first_entry(&mdsc
->cap_flush_list
,
1578 struct ceph_cap_flush
, g_list
);
1579 if (cf
->tid
<= want_flush_tid
) {
1580 dout("check_caps_flush still flushing tid "
1581 "%llu <= %llu\n", cf
->tid
, want_flush_tid
);
1585 spin_unlock(&mdsc
->cap_dirty_lock
);
1590 * flush all dirty inode data to disk.
1592 * returns true if we've flushed through want_flush_tid
1594 static void wait_caps_flush(struct ceph_mds_client
*mdsc
,
1597 dout("check_caps_flush want %llu\n", want_flush_tid
);
1599 wait_event(mdsc
->cap_flushing_wq
,
1600 check_caps_flush(mdsc
, want_flush_tid
));
1602 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid
);
1606 * called under s_mutex
1608 void ceph_send_cap_releases(struct ceph_mds_client
*mdsc
,
1609 struct ceph_mds_session
*session
)
1611 struct ceph_msg
*msg
= NULL
;
1612 struct ceph_mds_cap_release
*head
;
1613 struct ceph_mds_cap_item
*item
;
1614 struct ceph_osd_client
*osdc
= &mdsc
->fsc
->client
->osdc
;
1615 struct ceph_cap
*cap
;
1616 LIST_HEAD(tmp_list
);
1617 int num_cap_releases
;
1618 __le32 barrier
, *cap_barrier
;
1620 down_read(&osdc
->lock
);
1621 barrier
= cpu_to_le32(osdc
->epoch_barrier
);
1622 up_read(&osdc
->lock
);
1624 spin_lock(&session
->s_cap_lock
);
1626 list_splice_init(&session
->s_cap_releases
, &tmp_list
);
1627 num_cap_releases
= session
->s_num_cap_releases
;
1628 session
->s_num_cap_releases
= 0;
1629 spin_unlock(&session
->s_cap_lock
);
1631 while (!list_empty(&tmp_list
)) {
1633 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE
,
1634 PAGE_SIZE
, GFP_NOFS
, false);
1637 head
= msg
->front
.iov_base
;
1638 head
->num
= cpu_to_le32(0);
1639 msg
->front
.iov_len
= sizeof(*head
);
1641 msg
->hdr
.version
= cpu_to_le16(2);
1642 msg
->hdr
.compat_version
= cpu_to_le16(1);
1645 cap
= list_first_entry(&tmp_list
, struct ceph_cap
,
1647 list_del(&cap
->session_caps
);
1650 head
= msg
->front
.iov_base
;
1651 le32_add_cpu(&head
->num
, 1);
1652 item
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1653 item
->ino
= cpu_to_le64(cap
->cap_ino
);
1654 item
->cap_id
= cpu_to_le64(cap
->cap_id
);
1655 item
->migrate_seq
= cpu_to_le32(cap
->mseq
);
1656 item
->seq
= cpu_to_le32(cap
->issue_seq
);
1657 msg
->front
.iov_len
+= sizeof(*item
);
1659 ceph_put_cap(mdsc
, cap
);
1661 if (le32_to_cpu(head
->num
) == CEPH_CAPS_PER_RELEASE
) {
1662 // Append cap_barrier field
1663 cap_barrier
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1664 *cap_barrier
= barrier
;
1665 msg
->front
.iov_len
+= sizeof(*cap_barrier
);
1667 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1668 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1669 ceph_con_send(&session
->s_con
, msg
);
1674 BUG_ON(num_cap_releases
!= 0);
1676 spin_lock(&session
->s_cap_lock
);
1677 if (!list_empty(&session
->s_cap_releases
))
1679 spin_unlock(&session
->s_cap_lock
);
1682 // Append cap_barrier field
1683 cap_barrier
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1684 *cap_barrier
= barrier
;
1685 msg
->front
.iov_len
+= sizeof(*cap_barrier
);
1687 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1688 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1689 ceph_con_send(&session
->s_con
, msg
);
1693 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1695 spin_lock(&session
->s_cap_lock
);
1696 list_splice(&tmp_list
, &session
->s_cap_releases
);
1697 session
->s_num_cap_releases
+= num_cap_releases
;
1698 spin_unlock(&session
->s_cap_lock
);
1705 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request
*req
,
1708 struct ceph_inode_info
*ci
= ceph_inode(dir
);
1709 struct ceph_mds_reply_info_parsed
*rinfo
= &req
->r_reply_info
;
1710 struct ceph_mount_options
*opt
= req
->r_mdsc
->fsc
->mount_options
;
1711 size_t size
= sizeof(struct ceph_mds_reply_dir_entry
);
1712 int order
, num_entries
;
1714 spin_lock(&ci
->i_ceph_lock
);
1715 num_entries
= ci
->i_files
+ ci
->i_subdirs
;
1716 spin_unlock(&ci
->i_ceph_lock
);
1717 num_entries
= max(num_entries
, 1);
1718 num_entries
= min(num_entries
, opt
->max_readdir
);
1720 order
= get_order(size
* num_entries
);
1721 while (order
>= 0) {
1722 rinfo
->dir_entries
= (void*)__get_free_pages(GFP_KERNEL
|
1725 if (rinfo
->dir_entries
)
1729 if (!rinfo
->dir_entries
)
1732 num_entries
= (PAGE_SIZE
<< order
) / size
;
1733 num_entries
= min(num_entries
, opt
->max_readdir
);
1735 rinfo
->dir_buf_size
= PAGE_SIZE
<< order
;
1736 req
->r_num_caps
= num_entries
+ 1;
1737 req
->r_args
.readdir
.max_entries
= cpu_to_le32(num_entries
);
1738 req
->r_args
.readdir
.max_bytes
= cpu_to_le32(opt
->max_readdir_bytes
);
1743 * Create an mds request.
1745 struct ceph_mds_request
*
1746 ceph_mdsc_create_request(struct ceph_mds_client
*mdsc
, int op
, int mode
)
1748 struct ceph_mds_request
*req
= kzalloc(sizeof(*req
), GFP_NOFS
);
1751 return ERR_PTR(-ENOMEM
);
1753 mutex_init(&req
->r_fill_mutex
);
1755 req
->r_started
= jiffies
;
1756 req
->r_resend_mds
= -1;
1757 INIT_LIST_HEAD(&req
->r_unsafe_dir_item
);
1758 INIT_LIST_HEAD(&req
->r_unsafe_target_item
);
1760 kref_init(&req
->r_kref
);
1761 RB_CLEAR_NODE(&req
->r_node
);
1762 INIT_LIST_HEAD(&req
->r_wait
);
1763 init_completion(&req
->r_completion
);
1764 init_completion(&req
->r_safe_completion
);
1765 INIT_LIST_HEAD(&req
->r_unsafe_item
);
1767 req
->r_stamp
= timespec_trunc(current_kernel_time(), mdsc
->fsc
->sb
->s_time_gran
);
1770 req
->r_direct_mode
= mode
;
1775 * return oldest (lowest) request, tid in request tree, 0 if none.
1777 * called under mdsc->mutex.
1779 static struct ceph_mds_request
*__get_oldest_req(struct ceph_mds_client
*mdsc
)
1781 if (RB_EMPTY_ROOT(&mdsc
->request_tree
))
1783 return rb_entry(rb_first(&mdsc
->request_tree
),
1784 struct ceph_mds_request
, r_node
);
1787 static inline u64
__get_oldest_tid(struct ceph_mds_client
*mdsc
)
1789 return mdsc
->oldest_tid
;
1793 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1794 * on build_path_from_dentry in fs/cifs/dir.c.
1796 * If @stop_on_nosnap, generate path relative to the first non-snapped
1799 * Encode hidden .snap dirs as a double /, i.e.
1800 * foo/.snap/bar -> foo//bar
1802 char *ceph_mdsc_build_path(struct dentry
*dentry
, int *plen
, u64
*base
,
1805 struct dentry
*temp
;
1811 return ERR_PTR(-EINVAL
);
1815 seq
= read_seqbegin(&rename_lock
);
1817 for (temp
= dentry
; !IS_ROOT(temp
);) {
1818 struct inode
*inode
= d_inode(temp
);
1819 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
)
1820 len
++; /* slash only */
1821 else if (stop_on_nosnap
&& inode
&&
1822 ceph_snap(inode
) == CEPH_NOSNAP
)
1825 len
+= 1 + temp
->d_name
.len
;
1826 temp
= temp
->d_parent
;
1830 len
--; /* no leading '/' */
1832 path
= kmalloc(len
+1, GFP_NOFS
);
1834 return ERR_PTR(-ENOMEM
);
1836 path
[pos
] = 0; /* trailing null */
1838 for (temp
= dentry
; !IS_ROOT(temp
) && pos
!= 0; ) {
1839 struct inode
*inode
;
1841 spin_lock(&temp
->d_lock
);
1842 inode
= d_inode(temp
);
1843 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
) {
1844 dout("build_path path+%d: %p SNAPDIR\n",
1846 } else if (stop_on_nosnap
&& inode
&&
1847 ceph_snap(inode
) == CEPH_NOSNAP
) {
1848 spin_unlock(&temp
->d_lock
);
1851 pos
-= temp
->d_name
.len
;
1853 spin_unlock(&temp
->d_lock
);
1856 strncpy(path
+ pos
, temp
->d_name
.name
,
1859 spin_unlock(&temp
->d_lock
);
1862 temp
= temp
->d_parent
;
1865 if (pos
!= 0 || read_seqretry(&rename_lock
, seq
)) {
1866 pr_err("build_path did not end path lookup where "
1867 "expected, namelen is %d, pos is %d\n", len
, pos
);
1868 /* presumably this is only possible if racing with a
1869 rename of one of the parent directories (we can not
1870 lock the dentries above us to prevent this, but
1871 retrying should be harmless) */
1876 *base
= ceph_ino(d_inode(temp
));
1878 dout("build_path on %p %d built %llx '%.*s'\n",
1879 dentry
, d_count(dentry
), *base
, len
, path
);
1883 static int build_dentry_path(struct dentry
*dentry
, struct inode
*dir
,
1884 const char **ppath
, int *ppathlen
, u64
*pino
,
1891 dir
= d_inode_rcu(dentry
->d_parent
);
1892 if (dir
&& ceph_snap(dir
) == CEPH_NOSNAP
) {
1893 *pino
= ceph_ino(dir
);
1895 *ppath
= dentry
->d_name
.name
;
1896 *ppathlen
= dentry
->d_name
.len
;
1900 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1902 return PTR_ERR(path
);
1908 static int build_inode_path(struct inode
*inode
,
1909 const char **ppath
, int *ppathlen
, u64
*pino
,
1912 struct dentry
*dentry
;
1915 if (ceph_snap(inode
) == CEPH_NOSNAP
) {
1916 *pino
= ceph_ino(inode
);
1920 dentry
= d_find_alias(inode
);
1921 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1924 return PTR_ERR(path
);
1931 * request arguments may be specified via an inode *, a dentry *, or
1932 * an explicit ino+path.
1934 static int set_request_path_attr(struct inode
*rinode
, struct dentry
*rdentry
,
1935 struct inode
*rdiri
, const char *rpath
,
1936 u64 rino
, const char **ppath
, int *pathlen
,
1937 u64
*ino
, int *freepath
)
1942 r
= build_inode_path(rinode
, ppath
, pathlen
, ino
, freepath
);
1943 dout(" inode %p %llx.%llx\n", rinode
, ceph_ino(rinode
),
1945 } else if (rdentry
) {
1946 r
= build_dentry_path(rdentry
, rdiri
, ppath
, pathlen
, ino
,
1948 dout(" dentry %p %llx/%.*s\n", rdentry
, *ino
, *pathlen
,
1950 } else if (rpath
|| rino
) {
1953 *pathlen
= rpath
? strlen(rpath
) : 0;
1954 dout(" path %.*s\n", *pathlen
, rpath
);
1961 * called under mdsc->mutex
1963 static struct ceph_msg
*create_request_message(struct ceph_mds_client
*mdsc
,
1964 struct ceph_mds_request
*req
,
1965 int mds
, bool drop_cap_releases
)
1967 struct ceph_msg
*msg
;
1968 struct ceph_mds_request_head
*head
;
1969 const char *path1
= NULL
;
1970 const char *path2
= NULL
;
1971 u64 ino1
= 0, ino2
= 0;
1972 int pathlen1
= 0, pathlen2
= 0;
1973 int freepath1
= 0, freepath2
= 0;
1979 ret
= set_request_path_attr(req
->r_inode
, req
->r_dentry
,
1980 req
->r_parent
, req
->r_path1
, req
->r_ino1
.ino
,
1981 &path1
, &pathlen1
, &ino1
, &freepath1
);
1987 ret
= set_request_path_attr(NULL
, req
->r_old_dentry
,
1988 req
->r_old_dentry_dir
,
1989 req
->r_path2
, req
->r_ino2
.ino
,
1990 &path2
, &pathlen2
, &ino2
, &freepath2
);
1996 len
= sizeof(*head
) +
1997 pathlen1
+ pathlen2
+ 2*(1 + sizeof(u32
) + sizeof(u64
)) +
1998 sizeof(struct ceph_timespec
);
2000 /* calculate (max) length for cap releases */
2001 len
+= sizeof(struct ceph_mds_request_release
) *
2002 (!!req
->r_inode_drop
+ !!req
->r_dentry_drop
+
2003 !!req
->r_old_inode_drop
+ !!req
->r_old_dentry_drop
);
2004 if (req
->r_dentry_drop
)
2005 len
+= req
->r_dentry
->d_name
.len
;
2006 if (req
->r_old_dentry_drop
)
2007 len
+= req
->r_old_dentry
->d_name
.len
;
2009 msg
= ceph_msg_new(CEPH_MSG_CLIENT_REQUEST
, len
, GFP_NOFS
, false);
2011 msg
= ERR_PTR(-ENOMEM
);
2015 msg
->hdr
.version
= cpu_to_le16(2);
2016 msg
->hdr
.tid
= cpu_to_le64(req
->r_tid
);
2018 head
= msg
->front
.iov_base
;
2019 p
= msg
->front
.iov_base
+ sizeof(*head
);
2020 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
2022 head
->mdsmap_epoch
= cpu_to_le32(mdsc
->mdsmap
->m_epoch
);
2023 head
->op
= cpu_to_le32(req
->r_op
);
2024 head
->caller_uid
= cpu_to_le32(from_kuid(&init_user_ns
, req
->r_uid
));
2025 head
->caller_gid
= cpu_to_le32(from_kgid(&init_user_ns
, req
->r_gid
));
2026 head
->args
= req
->r_args
;
2028 ceph_encode_filepath(&p
, end
, ino1
, path1
);
2029 ceph_encode_filepath(&p
, end
, ino2
, path2
);
2031 /* make note of release offset, in case we need to replay */
2032 req
->r_request_release_offset
= p
- msg
->front
.iov_base
;
2036 if (req
->r_inode_drop
)
2037 releases
+= ceph_encode_inode_release(&p
,
2038 req
->r_inode
? req
->r_inode
: d_inode(req
->r_dentry
),
2039 mds
, req
->r_inode_drop
, req
->r_inode_unless
, 0);
2040 if (req
->r_dentry_drop
)
2041 releases
+= ceph_encode_dentry_release(&p
, req
->r_dentry
,
2042 req
->r_parent
, mds
, req
->r_dentry_drop
,
2043 req
->r_dentry_unless
);
2044 if (req
->r_old_dentry_drop
)
2045 releases
+= ceph_encode_dentry_release(&p
, req
->r_old_dentry
,
2046 req
->r_old_dentry_dir
, mds
,
2047 req
->r_old_dentry_drop
,
2048 req
->r_old_dentry_unless
);
2049 if (req
->r_old_inode_drop
)
2050 releases
+= ceph_encode_inode_release(&p
,
2051 d_inode(req
->r_old_dentry
),
2052 mds
, req
->r_old_inode_drop
, req
->r_old_inode_unless
, 0);
2054 if (drop_cap_releases
) {
2056 p
= msg
->front
.iov_base
+ req
->r_request_release_offset
;
2059 head
->num_releases
= cpu_to_le16(releases
);
2063 struct ceph_timespec ts
;
2064 ceph_encode_timespec(&ts
, &req
->r_stamp
);
2065 ceph_encode_copy(&p
, &ts
, sizeof(ts
));
2069 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
2070 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
2072 if (req
->r_pagelist
) {
2073 struct ceph_pagelist
*pagelist
= req
->r_pagelist
;
2074 refcount_inc(&pagelist
->refcnt
);
2075 ceph_msg_data_add_pagelist(msg
, pagelist
);
2076 msg
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
2078 msg
->hdr
.data_len
= 0;
2081 msg
->hdr
.data_off
= cpu_to_le16(0);
2085 kfree((char *)path2
);
2088 kfree((char *)path1
);
2094 * called under mdsc->mutex if error, under no mutex if
2097 static void complete_request(struct ceph_mds_client
*mdsc
,
2098 struct ceph_mds_request
*req
)
2100 if (req
->r_callback
)
2101 req
->r_callback(mdsc
, req
);
2103 complete_all(&req
->r_completion
);
2107 * called under mdsc->mutex
2109 static int __prepare_send_request(struct ceph_mds_client
*mdsc
,
2110 struct ceph_mds_request
*req
,
2111 int mds
, bool drop_cap_releases
)
2113 struct ceph_mds_request_head
*rhead
;
2114 struct ceph_msg
*msg
;
2119 struct ceph_cap
*cap
=
2120 ceph_get_cap_for_mds(ceph_inode(req
->r_inode
), mds
);
2123 req
->r_sent_on_mseq
= cap
->mseq
;
2125 req
->r_sent_on_mseq
= -1;
2127 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req
,
2128 req
->r_tid
, ceph_mds_op_name(req
->r_op
), req
->r_attempts
);
2130 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
)) {
2133 * Replay. Do not regenerate message (and rebuild
2134 * paths, etc.); just use the original message.
2135 * Rebuilding paths will break for renames because
2136 * d_move mangles the src name.
2138 msg
= req
->r_request
;
2139 rhead
= msg
->front
.iov_base
;
2141 flags
= le32_to_cpu(rhead
->flags
);
2142 flags
|= CEPH_MDS_FLAG_REPLAY
;
2143 rhead
->flags
= cpu_to_le32(flags
);
2145 if (req
->r_target_inode
)
2146 rhead
->ino
= cpu_to_le64(ceph_ino(req
->r_target_inode
));
2148 rhead
->num_retry
= req
->r_attempts
- 1;
2150 /* remove cap/dentry releases from message */
2151 rhead
->num_releases
= 0;
2154 p
= msg
->front
.iov_base
+ req
->r_request_release_offset
;
2156 struct ceph_timespec ts
;
2157 ceph_encode_timespec(&ts
, &req
->r_stamp
);
2158 ceph_encode_copy(&p
, &ts
, sizeof(ts
));
2161 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
2162 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
2166 if (req
->r_request
) {
2167 ceph_msg_put(req
->r_request
);
2168 req
->r_request
= NULL
;
2170 msg
= create_request_message(mdsc
, req
, mds
, drop_cap_releases
);
2172 req
->r_err
= PTR_ERR(msg
);
2173 return PTR_ERR(msg
);
2175 req
->r_request
= msg
;
2177 rhead
= msg
->front
.iov_base
;
2178 rhead
->oldest_client_tid
= cpu_to_le64(__get_oldest_tid(mdsc
));
2179 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
))
2180 flags
|= CEPH_MDS_FLAG_REPLAY
;
2182 flags
|= CEPH_MDS_FLAG_WANT_DENTRY
;
2183 rhead
->flags
= cpu_to_le32(flags
);
2184 rhead
->num_fwd
= req
->r_num_fwd
;
2185 rhead
->num_retry
= req
->r_attempts
- 1;
2188 dout(" r_parent = %p\n", req
->r_parent
);
2193 * send request, or put it on the appropriate wait list.
2195 static int __do_request(struct ceph_mds_client
*mdsc
,
2196 struct ceph_mds_request
*req
)
2198 struct ceph_mds_session
*session
= NULL
;
2202 if (req
->r_err
|| test_bit(CEPH_MDS_R_GOT_RESULT
, &req
->r_req_flags
)) {
2203 if (test_bit(CEPH_MDS_R_ABORTED
, &req
->r_req_flags
))
2204 __unregister_request(mdsc
, req
);
2208 if (req
->r_timeout
&&
2209 time_after_eq(jiffies
, req
->r_started
+ req
->r_timeout
)) {
2210 dout("do_request timed out\n");
2214 if (READ_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
) {
2215 dout("do_request forced umount\n");
2219 if (READ_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_MOUNTING
) {
2220 if (mdsc
->mdsmap_err
) {
2221 err
= mdsc
->mdsmap_err
;
2222 dout("do_request mdsmap err %d\n", err
);
2225 if (mdsc
->mdsmap
->m_epoch
== 0) {
2226 dout("do_request no mdsmap, waiting for map\n");
2227 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
2230 if (!(mdsc
->fsc
->mount_options
->flags
&
2231 CEPH_MOUNT_OPT_MOUNTWAIT
) &&
2232 !ceph_mdsmap_is_cluster_available(mdsc
->mdsmap
)) {
2234 pr_info("probably no mds server is up\n");
2239 put_request_session(req
);
2241 mds
= __choose_mds(mdsc
, req
);
2243 ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) < CEPH_MDS_STATE_ACTIVE
) {
2244 dout("do_request no mds or not active, waiting for map\n");
2245 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
2249 /* get, open session */
2250 session
= __ceph_lookup_mds_session(mdsc
, mds
);
2252 session
= register_session(mdsc
, mds
);
2253 if (IS_ERR(session
)) {
2254 err
= PTR_ERR(session
);
2258 req
->r_session
= get_session(session
);
2260 dout("do_request mds%d session %p state %s\n", mds
, session
,
2261 ceph_session_state_name(session
->s_state
));
2262 if (session
->s_state
!= CEPH_MDS_SESSION_OPEN
&&
2263 session
->s_state
!= CEPH_MDS_SESSION_HUNG
) {
2264 if (session
->s_state
== CEPH_MDS_SESSION_REJECTED
) {
2268 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
2269 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
2270 __open_session(mdsc
, session
);
2271 list_add(&req
->r_wait
, &session
->s_waiting
);
2276 req
->r_resend_mds
= -1; /* forget any previous mds hint */
2278 if (req
->r_request_started
== 0) /* note request start time */
2279 req
->r_request_started
= jiffies
;
2281 err
= __prepare_send_request(mdsc
, req
, mds
, false);
2283 ceph_msg_get(req
->r_request
);
2284 ceph_con_send(&session
->s_con
, req
->r_request
);
2288 ceph_put_mds_session(session
);
2291 dout("__do_request early error %d\n", err
);
2293 complete_request(mdsc
, req
);
2294 __unregister_request(mdsc
, req
);
2301 * called under mdsc->mutex
2303 static void __wake_requests(struct ceph_mds_client
*mdsc
,
2304 struct list_head
*head
)
2306 struct ceph_mds_request
*req
;
2307 LIST_HEAD(tmp_list
);
2309 list_splice_init(head
, &tmp_list
);
2311 while (!list_empty(&tmp_list
)) {
2312 req
= list_entry(tmp_list
.next
,
2313 struct ceph_mds_request
, r_wait
);
2314 list_del_init(&req
->r_wait
);
2315 dout(" wake request %p tid %llu\n", req
, req
->r_tid
);
2316 __do_request(mdsc
, req
);
2321 * Wake up threads with requests pending for @mds, so that they can
2322 * resubmit their requests to a possibly different mds.
2324 static void kick_requests(struct ceph_mds_client
*mdsc
, int mds
)
2326 struct ceph_mds_request
*req
;
2327 struct rb_node
*p
= rb_first(&mdsc
->request_tree
);
2329 dout("kick_requests mds%d\n", mds
);
2331 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2333 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
))
2335 if (req
->r_attempts
> 0)
2336 continue; /* only new requests */
2337 if (req
->r_session
&&
2338 req
->r_session
->s_mds
== mds
) {
2339 dout(" kicking tid %llu\n", req
->r_tid
);
2340 list_del_init(&req
->r_wait
);
2341 __do_request(mdsc
, req
);
2346 void ceph_mdsc_submit_request(struct ceph_mds_client
*mdsc
,
2347 struct ceph_mds_request
*req
)
2349 dout("submit_request on %p\n", req
);
2350 mutex_lock(&mdsc
->mutex
);
2351 __register_request(mdsc
, req
, NULL
);
2352 __do_request(mdsc
, req
);
2353 mutex_unlock(&mdsc
->mutex
);
2357 * Synchrously perform an mds request. Take care of all of the
2358 * session setup, forwarding, retry details.
2360 int ceph_mdsc_do_request(struct ceph_mds_client
*mdsc
,
2362 struct ceph_mds_request
*req
)
2366 dout("do_request on %p\n", req
);
2368 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2370 ceph_get_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
2372 ceph_get_cap_refs(ceph_inode(req
->r_parent
), CEPH_CAP_PIN
);
2373 if (req
->r_old_dentry_dir
)
2374 ceph_get_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
2378 mutex_lock(&mdsc
->mutex
);
2379 __register_request(mdsc
, req
, dir
);
2380 __do_request(mdsc
, req
);
2388 mutex_unlock(&mdsc
->mutex
);
2389 dout("do_request waiting\n");
2390 if (!req
->r_timeout
&& req
->r_wait_for_completion
) {
2391 err
= req
->r_wait_for_completion(mdsc
, req
);
2393 long timeleft
= wait_for_completion_killable_timeout(
2395 ceph_timeout_jiffies(req
->r_timeout
));
2399 err
= -EIO
; /* timed out */
2401 err
= timeleft
; /* killed */
2403 dout("do_request waited, got %d\n", err
);
2404 mutex_lock(&mdsc
->mutex
);
2406 /* only abort if we didn't race with a real reply */
2407 if (test_bit(CEPH_MDS_R_GOT_RESULT
, &req
->r_req_flags
)) {
2408 err
= le32_to_cpu(req
->r_reply_info
.head
->result
);
2409 } else if (err
< 0) {
2410 dout("aborted request %lld with %d\n", req
->r_tid
, err
);
2413 * ensure we aren't running concurrently with
2414 * ceph_fill_trace or ceph_readdir_prepopulate, which
2415 * rely on locks (dir mutex) held by our caller.
2417 mutex_lock(&req
->r_fill_mutex
);
2419 set_bit(CEPH_MDS_R_ABORTED
, &req
->r_req_flags
);
2420 mutex_unlock(&req
->r_fill_mutex
);
2422 if (req
->r_parent
&&
2423 (req
->r_op
& CEPH_MDS_OP_WRITE
))
2424 ceph_invalidate_dir_request(req
);
2430 mutex_unlock(&mdsc
->mutex
);
2431 dout("do_request %p done, result %d\n", req
, err
);
2436 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2437 * namespace request.
2439 void ceph_invalidate_dir_request(struct ceph_mds_request
*req
)
2441 struct inode
*inode
= req
->r_parent
;
2443 dout("invalidate_dir_request %p (complete, lease(s))\n", inode
);
2445 ceph_dir_clear_complete(inode
);
2447 ceph_invalidate_dentry_lease(req
->r_dentry
);
2448 if (req
->r_old_dentry
)
2449 ceph_invalidate_dentry_lease(req
->r_old_dentry
);
2455 * We take the session mutex and parse and process the reply immediately.
2456 * This preserves the logical ordering of replies, capabilities, etc., sent
2457 * by the MDS as they are applied to our local cache.
2459 static void handle_reply(struct ceph_mds_session
*session
, struct ceph_msg
*msg
)
2461 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2462 struct ceph_mds_request
*req
;
2463 struct ceph_mds_reply_head
*head
= msg
->front
.iov_base
;
2464 struct ceph_mds_reply_info_parsed
*rinfo
; /* parsed reply info */
2465 struct ceph_snap_realm
*realm
;
2468 int mds
= session
->s_mds
;
2470 if (msg
->front
.iov_len
< sizeof(*head
)) {
2471 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2476 /* get request, session */
2477 tid
= le64_to_cpu(msg
->hdr
.tid
);
2478 mutex_lock(&mdsc
->mutex
);
2479 req
= lookup_get_request(mdsc
, tid
);
2481 dout("handle_reply on unknown tid %llu\n", tid
);
2482 mutex_unlock(&mdsc
->mutex
);
2485 dout("handle_reply %p\n", req
);
2487 /* correct session? */
2488 if (req
->r_session
!= session
) {
2489 pr_err("mdsc_handle_reply got %llu on session mds%d"
2490 " not mds%d\n", tid
, session
->s_mds
,
2491 req
->r_session
? req
->r_session
->s_mds
: -1);
2492 mutex_unlock(&mdsc
->mutex
);
2497 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
) && !head
->safe
) ||
2498 (test_bit(CEPH_MDS_R_GOT_SAFE
, &req
->r_req_flags
) && head
->safe
)) {
2499 pr_warn("got a dup %s reply on %llu from mds%d\n",
2500 head
->safe
? "safe" : "unsafe", tid
, mds
);
2501 mutex_unlock(&mdsc
->mutex
);
2504 if (test_bit(CEPH_MDS_R_GOT_SAFE
, &req
->r_req_flags
)) {
2505 pr_warn("got unsafe after safe on %llu from mds%d\n",
2507 mutex_unlock(&mdsc
->mutex
);
2511 result
= le32_to_cpu(head
->result
);
2515 * if we're not talking to the authority, send to them
2516 * if the authority has changed while we weren't looking,
2517 * send to new authority
2518 * Otherwise we just have to return an ESTALE
2520 if (result
== -ESTALE
) {
2521 dout("got ESTALE on request %llu", req
->r_tid
);
2522 req
->r_resend_mds
= -1;
2523 if (req
->r_direct_mode
!= USE_AUTH_MDS
) {
2524 dout("not using auth, setting for that now");
2525 req
->r_direct_mode
= USE_AUTH_MDS
;
2526 __do_request(mdsc
, req
);
2527 mutex_unlock(&mdsc
->mutex
);
2530 int mds
= __choose_mds(mdsc
, req
);
2531 if (mds
>= 0 && mds
!= req
->r_session
->s_mds
) {
2532 dout("but auth changed, so resending");
2533 __do_request(mdsc
, req
);
2534 mutex_unlock(&mdsc
->mutex
);
2538 dout("have to return ESTALE on request %llu", req
->r_tid
);
2543 set_bit(CEPH_MDS_R_GOT_SAFE
, &req
->r_req_flags
);
2544 __unregister_request(mdsc
, req
);
2546 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
)) {
2548 * We already handled the unsafe response, now do the
2549 * cleanup. No need to examine the response; the MDS
2550 * doesn't include any result info in the safe
2551 * response. And even if it did, there is nothing
2552 * useful we could do with a revised return value.
2554 dout("got safe reply %llu, mds%d\n", tid
, mds
);
2556 /* last unsafe request during umount? */
2557 if (mdsc
->stopping
&& !__get_oldest_req(mdsc
))
2558 complete_all(&mdsc
->safe_umount_waiters
);
2559 mutex_unlock(&mdsc
->mutex
);
2563 set_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
);
2564 list_add_tail(&req
->r_unsafe_item
, &req
->r_session
->s_unsafe
);
2565 if (req
->r_unsafe_dir
) {
2566 struct ceph_inode_info
*ci
=
2567 ceph_inode(req
->r_unsafe_dir
);
2568 spin_lock(&ci
->i_unsafe_lock
);
2569 list_add_tail(&req
->r_unsafe_dir_item
,
2570 &ci
->i_unsafe_dirops
);
2571 spin_unlock(&ci
->i_unsafe_lock
);
2575 dout("handle_reply tid %lld result %d\n", tid
, result
);
2576 rinfo
= &req
->r_reply_info
;
2577 err
= parse_reply_info(msg
, rinfo
, session
->s_con
.peer_features
);
2578 mutex_unlock(&mdsc
->mutex
);
2580 mutex_lock(&session
->s_mutex
);
2582 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds
, tid
);
2589 if (rinfo
->snapblob_len
) {
2590 down_write(&mdsc
->snap_rwsem
);
2591 ceph_update_snap_trace(mdsc
, rinfo
->snapblob
,
2592 rinfo
->snapblob
+ rinfo
->snapblob_len
,
2593 le32_to_cpu(head
->op
) == CEPH_MDS_OP_RMSNAP
,
2595 downgrade_write(&mdsc
->snap_rwsem
);
2597 down_read(&mdsc
->snap_rwsem
);
2600 /* insert trace into our cache */
2601 mutex_lock(&req
->r_fill_mutex
);
2602 current
->journal_info
= req
;
2603 err
= ceph_fill_trace(mdsc
->fsc
->sb
, req
);
2605 if (result
== 0 && (req
->r_op
== CEPH_MDS_OP_READDIR
||
2606 req
->r_op
== CEPH_MDS_OP_LSSNAP
))
2607 ceph_readdir_prepopulate(req
, req
->r_session
);
2608 ceph_unreserve_caps(mdsc
, &req
->r_caps_reservation
);
2610 current
->journal_info
= NULL
;
2611 mutex_unlock(&req
->r_fill_mutex
);
2613 up_read(&mdsc
->snap_rwsem
);
2615 ceph_put_snap_realm(mdsc
, realm
);
2617 if (err
== 0 && req
->r_target_inode
&&
2618 test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
)) {
2619 struct ceph_inode_info
*ci
= ceph_inode(req
->r_target_inode
);
2620 spin_lock(&ci
->i_unsafe_lock
);
2621 list_add_tail(&req
->r_unsafe_target_item
, &ci
->i_unsafe_iops
);
2622 spin_unlock(&ci
->i_unsafe_lock
);
2625 mutex_lock(&mdsc
->mutex
);
2626 if (!test_bit(CEPH_MDS_R_ABORTED
, &req
->r_req_flags
)) {
2630 req
->r_reply
= ceph_msg_get(msg
);
2631 set_bit(CEPH_MDS_R_GOT_RESULT
, &req
->r_req_flags
);
2634 dout("reply arrived after request %lld was aborted\n", tid
);
2636 mutex_unlock(&mdsc
->mutex
);
2638 mutex_unlock(&session
->s_mutex
);
2640 /* kick calling process */
2641 complete_request(mdsc
, req
);
2643 ceph_mdsc_put_request(req
);
2650 * handle mds notification that our request has been forwarded.
2652 static void handle_forward(struct ceph_mds_client
*mdsc
,
2653 struct ceph_mds_session
*session
,
2654 struct ceph_msg
*msg
)
2656 struct ceph_mds_request
*req
;
2657 u64 tid
= le64_to_cpu(msg
->hdr
.tid
);
2661 void *p
= msg
->front
.iov_base
;
2662 void *end
= p
+ msg
->front
.iov_len
;
2664 ceph_decode_need(&p
, end
, 2*sizeof(u32
), bad
);
2665 next_mds
= ceph_decode_32(&p
);
2666 fwd_seq
= ceph_decode_32(&p
);
2668 mutex_lock(&mdsc
->mutex
);
2669 req
= lookup_get_request(mdsc
, tid
);
2671 dout("forward tid %llu to mds%d - req dne\n", tid
, next_mds
);
2672 goto out
; /* dup reply? */
2675 if (test_bit(CEPH_MDS_R_ABORTED
, &req
->r_req_flags
)) {
2676 dout("forward tid %llu aborted, unregistering\n", tid
);
2677 __unregister_request(mdsc
, req
);
2678 } else if (fwd_seq
<= req
->r_num_fwd
) {
2679 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2680 tid
, next_mds
, req
->r_num_fwd
, fwd_seq
);
2682 /* resend. forward race not possible; mds would drop */
2683 dout("forward tid %llu to mds%d (we resend)\n", tid
, next_mds
);
2685 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT
, &req
->r_req_flags
));
2686 req
->r_attempts
= 0;
2687 req
->r_num_fwd
= fwd_seq
;
2688 req
->r_resend_mds
= next_mds
;
2689 put_request_session(req
);
2690 __do_request(mdsc
, req
);
2692 ceph_mdsc_put_request(req
);
2694 mutex_unlock(&mdsc
->mutex
);
2698 pr_err("mdsc_handle_forward decode error err=%d\n", err
);
2702 * handle a mds session control message
2704 static void handle_session(struct ceph_mds_session
*session
,
2705 struct ceph_msg
*msg
)
2707 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2710 int mds
= session
->s_mds
;
2711 struct ceph_mds_session_head
*h
= msg
->front
.iov_base
;
2715 if (msg
->front
.iov_len
!= sizeof(*h
))
2717 op
= le32_to_cpu(h
->op
);
2718 seq
= le64_to_cpu(h
->seq
);
2720 mutex_lock(&mdsc
->mutex
);
2721 if (op
== CEPH_SESSION_CLOSE
) {
2722 get_session(session
);
2723 __unregister_session(mdsc
, session
);
2725 /* FIXME: this ttl calculation is generous */
2726 session
->s_ttl
= jiffies
+ HZ
*mdsc
->mdsmap
->m_session_autoclose
;
2727 mutex_unlock(&mdsc
->mutex
);
2729 mutex_lock(&session
->s_mutex
);
2731 dout("handle_session mds%d %s %p state %s seq %llu\n",
2732 mds
, ceph_session_op_name(op
), session
,
2733 ceph_session_state_name(session
->s_state
), seq
);
2735 if (session
->s_state
== CEPH_MDS_SESSION_HUNG
) {
2736 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2737 pr_info("mds%d came back\n", session
->s_mds
);
2741 case CEPH_SESSION_OPEN
:
2742 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2743 pr_info("mds%d reconnect success\n", session
->s_mds
);
2744 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2745 renewed_caps(mdsc
, session
, 0);
2748 __close_session(mdsc
, session
);
2751 case CEPH_SESSION_RENEWCAPS
:
2752 if (session
->s_renew_seq
== seq
)
2753 renewed_caps(mdsc
, session
, 1);
2756 case CEPH_SESSION_CLOSE
:
2757 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2758 pr_info("mds%d reconnect denied\n", session
->s_mds
);
2759 cleanup_session_requests(mdsc
, session
);
2760 remove_session_caps(session
);
2761 wake
= 2; /* for good measure */
2762 wake_up_all(&mdsc
->session_close_wq
);
2765 case CEPH_SESSION_STALE
:
2766 pr_info("mds%d caps went stale, renewing\n",
2768 spin_lock(&session
->s_gen_ttl_lock
);
2769 session
->s_cap_gen
++;
2770 session
->s_cap_ttl
= jiffies
- 1;
2771 spin_unlock(&session
->s_gen_ttl_lock
);
2772 send_renew_caps(mdsc
, session
);
2775 case CEPH_SESSION_RECALL_STATE
:
2776 trim_caps(mdsc
, session
, le32_to_cpu(h
->max_caps
));
2779 case CEPH_SESSION_FLUSHMSG
:
2780 send_flushmsg_ack(mdsc
, session
, seq
);
2783 case CEPH_SESSION_FORCE_RO
:
2784 dout("force_session_readonly %p\n", session
);
2785 spin_lock(&session
->s_cap_lock
);
2786 session
->s_readonly
= true;
2787 spin_unlock(&session
->s_cap_lock
);
2788 wake_up_session_caps(session
, 0);
2791 case CEPH_SESSION_REJECT
:
2792 WARN_ON(session
->s_state
!= CEPH_MDS_SESSION_OPENING
);
2793 pr_info("mds%d rejected session\n", session
->s_mds
);
2794 session
->s_state
= CEPH_MDS_SESSION_REJECTED
;
2795 cleanup_session_requests(mdsc
, session
);
2796 remove_session_caps(session
);
2797 wake
= 2; /* for good measure */
2801 pr_err("mdsc_handle_session bad op %d mds%d\n", op
, mds
);
2805 mutex_unlock(&session
->s_mutex
);
2807 mutex_lock(&mdsc
->mutex
);
2808 __wake_requests(mdsc
, &session
->s_waiting
);
2810 kick_requests(mdsc
, mds
);
2811 mutex_unlock(&mdsc
->mutex
);
2813 if (op
== CEPH_SESSION_CLOSE
)
2814 ceph_put_mds_session(session
);
2818 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds
,
2819 (int)msg
->front
.iov_len
);
2826 * called under session->mutex.
2828 static void replay_unsafe_requests(struct ceph_mds_client
*mdsc
,
2829 struct ceph_mds_session
*session
)
2831 struct ceph_mds_request
*req
, *nreq
;
2835 dout("replay_unsafe_requests mds%d\n", session
->s_mds
);
2837 mutex_lock(&mdsc
->mutex
);
2838 list_for_each_entry_safe(req
, nreq
, &session
->s_unsafe
, r_unsafe_item
) {
2839 err
= __prepare_send_request(mdsc
, req
, session
->s_mds
, true);
2841 ceph_msg_get(req
->r_request
);
2842 ceph_con_send(&session
->s_con
, req
->r_request
);
2847 * also re-send old requests when MDS enters reconnect stage. So that MDS
2848 * can process completed request in clientreplay stage.
2850 p
= rb_first(&mdsc
->request_tree
);
2852 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2854 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
))
2856 if (req
->r_attempts
== 0)
2857 continue; /* only old requests */
2858 if (req
->r_session
&&
2859 req
->r_session
->s_mds
== session
->s_mds
) {
2860 err
= __prepare_send_request(mdsc
, req
,
2861 session
->s_mds
, true);
2863 ceph_msg_get(req
->r_request
);
2864 ceph_con_send(&session
->s_con
, req
->r_request
);
2868 mutex_unlock(&mdsc
->mutex
);
2872 * Encode information about a cap for a reconnect with the MDS.
2874 static int encode_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
2878 struct ceph_mds_cap_reconnect v2
;
2879 struct ceph_mds_cap_reconnect_v1 v1
;
2881 struct ceph_inode_info
*ci
= cap
->ci
;
2882 struct ceph_reconnect_state
*recon_state
= arg
;
2883 struct ceph_pagelist
*pagelist
= recon_state
->pagelist
;
2888 struct dentry
*dentry
;
2890 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2891 inode
, ceph_vinop(inode
), cap
, cap
->cap_id
,
2892 ceph_cap_string(cap
->issued
));
2893 err
= ceph_pagelist_encode_64(pagelist
, ceph_ino(inode
));
2897 dentry
= d_find_alias(inode
);
2899 path
= ceph_mdsc_build_path(dentry
, &pathlen
, &pathbase
, 0);
2901 err
= PTR_ERR(path
);
2910 spin_lock(&ci
->i_ceph_lock
);
2911 cap
->seq
= 0; /* reset cap seq */
2912 cap
->issue_seq
= 0; /* and issue_seq */
2913 cap
->mseq
= 0; /* and migrate_seq */
2914 cap
->cap_gen
= cap
->session
->s_cap_gen
;
2916 if (recon_state
->msg_version
>= 2) {
2917 rec
.v2
.cap_id
= cpu_to_le64(cap
->cap_id
);
2918 rec
.v2
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2919 rec
.v2
.issued
= cpu_to_le32(cap
->issued
);
2920 rec
.v2
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2921 rec
.v2
.pathbase
= cpu_to_le64(pathbase
);
2922 rec
.v2
.flock_len
= (__force __le32
)
2923 ((ci
->i_ceph_flags
& CEPH_I_ERROR_FILELOCK
) ? 0 : 1);
2925 rec
.v1
.cap_id
= cpu_to_le64(cap
->cap_id
);
2926 rec
.v1
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2927 rec
.v1
.issued
= cpu_to_le32(cap
->issued
);
2928 rec
.v1
.size
= cpu_to_le64(inode
->i_size
);
2929 ceph_encode_timespec(&rec
.v1
.mtime
, &inode
->i_mtime
);
2930 ceph_encode_timespec(&rec
.v1
.atime
, &inode
->i_atime
);
2931 rec
.v1
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2932 rec
.v1
.pathbase
= cpu_to_le64(pathbase
);
2935 if (list_empty(&ci
->i_cap_snaps
)) {
2936 snap_follows
= ci
->i_head_snapc
? ci
->i_head_snapc
->seq
: 0;
2938 struct ceph_cap_snap
*capsnap
=
2939 list_first_entry(&ci
->i_cap_snaps
,
2940 struct ceph_cap_snap
, ci_item
);
2941 snap_follows
= capsnap
->follows
;
2943 spin_unlock(&ci
->i_ceph_lock
);
2945 if (recon_state
->msg_version
>= 2) {
2946 int num_fcntl_locks
, num_flock_locks
;
2947 struct ceph_filelock
*flocks
= NULL
;
2948 size_t struct_len
, total_len
= 0;
2952 if (rec
.v2
.flock_len
) {
2953 ceph_count_locks(inode
, &num_fcntl_locks
, &num_flock_locks
);
2955 num_fcntl_locks
= 0;
2956 num_flock_locks
= 0;
2958 if (num_fcntl_locks
+ num_flock_locks
> 0) {
2959 flocks
= kmalloc((num_fcntl_locks
+ num_flock_locks
) *
2960 sizeof(struct ceph_filelock
), GFP_NOFS
);
2965 err
= ceph_encode_locks_to_buffer(inode
, flocks
,
2980 if (recon_state
->msg_version
>= 3) {
2981 /* version, compat_version and struct_len */
2982 total_len
= 2 * sizeof(u8
) + sizeof(u32
);
2986 * number of encoded locks is stable, so copy to pagelist
2988 struct_len
= 2 * sizeof(u32
) +
2989 (num_fcntl_locks
+ num_flock_locks
) *
2990 sizeof(struct ceph_filelock
);
2991 rec
.v2
.flock_len
= cpu_to_le32(struct_len
);
2993 struct_len
+= sizeof(rec
.v2
);
2994 struct_len
+= sizeof(u32
) + pathlen
;
2997 struct_len
+= sizeof(u64
); /* snap_follows */
2999 total_len
+= struct_len
;
3000 err
= ceph_pagelist_reserve(pagelist
, total_len
);
3003 if (recon_state
->msg_version
>= 3) {
3004 ceph_pagelist_encode_8(pagelist
, struct_v
);
3005 ceph_pagelist_encode_8(pagelist
, 1);
3006 ceph_pagelist_encode_32(pagelist
, struct_len
);
3008 ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
3009 ceph_pagelist_append(pagelist
, &rec
, sizeof(rec
.v2
));
3010 ceph_locks_to_pagelist(flocks
, pagelist
,
3014 ceph_pagelist_encode_64(pagelist
, snap_follows
);
3018 size_t size
= sizeof(u32
) + pathlen
+ sizeof(rec
.v1
);
3019 err
= ceph_pagelist_reserve(pagelist
, size
);
3021 ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
3022 ceph_pagelist_append(pagelist
, &rec
, sizeof(rec
.v1
));
3026 recon_state
->nr_caps
++;
3036 * If an MDS fails and recovers, clients need to reconnect in order to
3037 * reestablish shared state. This includes all caps issued through
3038 * this session _and_ the snap_realm hierarchy. Because it's not
3039 * clear which snap realms the mds cares about, we send everything we
3040 * know about.. that ensures we'll then get any new info the
3041 * recovering MDS might have.
3043 * This is a relatively heavyweight operation, but it's rare.
3045 * called with mdsc->mutex held.
3047 static void send_mds_reconnect(struct ceph_mds_client
*mdsc
,
3048 struct ceph_mds_session
*session
)
3050 struct ceph_msg
*reply
;
3052 int mds
= session
->s_mds
;
3055 struct ceph_pagelist
*pagelist
;
3056 struct ceph_reconnect_state recon_state
;
3059 pr_info("mds%d reconnect start\n", mds
);
3061 pagelist
= kmalloc(sizeof(*pagelist
), GFP_NOFS
);
3063 goto fail_nopagelist
;
3064 ceph_pagelist_init(pagelist
);
3066 reply
= ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT
, 0, GFP_NOFS
, false);
3070 mutex_lock(&session
->s_mutex
);
3071 session
->s_state
= CEPH_MDS_SESSION_RECONNECTING
;
3074 dout("session %p state %s\n", session
,
3075 ceph_session_state_name(session
->s_state
));
3077 spin_lock(&session
->s_gen_ttl_lock
);
3078 session
->s_cap_gen
++;
3079 spin_unlock(&session
->s_gen_ttl_lock
);
3081 spin_lock(&session
->s_cap_lock
);
3082 /* don't know if session is readonly */
3083 session
->s_readonly
= 0;
3085 * notify __ceph_remove_cap() that we are composing cap reconnect.
3086 * If a cap get released before being added to the cap reconnect,
3087 * __ceph_remove_cap() should skip queuing cap release.
3089 session
->s_cap_reconnect
= 1;
3090 /* drop old cap expires; we're about to reestablish that state */
3091 detach_cap_releases(session
, &dispose
);
3092 spin_unlock(&session
->s_cap_lock
);
3093 dispose_cap_releases(mdsc
, &dispose
);
3095 /* trim unused caps to reduce MDS's cache rejoin time */
3096 if (mdsc
->fsc
->sb
->s_root
)
3097 shrink_dcache_parent(mdsc
->fsc
->sb
->s_root
);
3099 ceph_con_close(&session
->s_con
);
3100 ceph_con_open(&session
->s_con
,
3101 CEPH_ENTITY_TYPE_MDS
, mds
,
3102 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
3104 /* replay unsafe requests */
3105 replay_unsafe_requests(mdsc
, session
);
3107 down_read(&mdsc
->snap_rwsem
);
3109 /* traverse this session's caps */
3110 s_nr_caps
= session
->s_nr_caps
;
3111 err
= ceph_pagelist_encode_32(pagelist
, s_nr_caps
);
3115 recon_state
.nr_caps
= 0;
3116 recon_state
.pagelist
= pagelist
;
3117 if (session
->s_con
.peer_features
& CEPH_FEATURE_MDSENC
)
3118 recon_state
.msg_version
= 3;
3119 else if (session
->s_con
.peer_features
& CEPH_FEATURE_FLOCK
)
3120 recon_state
.msg_version
= 2;
3122 recon_state
.msg_version
= 1;
3123 err
= iterate_session_caps(session
, encode_caps_cb
, &recon_state
);
3127 spin_lock(&session
->s_cap_lock
);
3128 session
->s_cap_reconnect
= 0;
3129 spin_unlock(&session
->s_cap_lock
);
3132 * snaprealms. we provide mds with the ino, seq (version), and
3133 * parent for all of our realms. If the mds has any newer info,
3136 for (p
= rb_first(&mdsc
->snap_realms
); p
; p
= rb_next(p
)) {
3137 struct ceph_snap_realm
*realm
=
3138 rb_entry(p
, struct ceph_snap_realm
, node
);
3139 struct ceph_mds_snaprealm_reconnect sr_rec
;
3141 dout(" adding snap realm %llx seq %lld parent %llx\n",
3142 realm
->ino
, realm
->seq
, realm
->parent_ino
);
3143 sr_rec
.ino
= cpu_to_le64(realm
->ino
);
3144 sr_rec
.seq
= cpu_to_le64(realm
->seq
);
3145 sr_rec
.parent
= cpu_to_le64(realm
->parent_ino
);
3146 err
= ceph_pagelist_append(pagelist
, &sr_rec
, sizeof(sr_rec
));
3151 reply
->hdr
.version
= cpu_to_le16(recon_state
.msg_version
);
3153 /* raced with cap release? */
3154 if (s_nr_caps
!= recon_state
.nr_caps
) {
3155 struct page
*page
= list_first_entry(&pagelist
->head
,
3157 __le32
*addr
= kmap_atomic(page
);
3158 *addr
= cpu_to_le32(recon_state
.nr_caps
);
3159 kunmap_atomic(addr
);
3162 reply
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
3163 ceph_msg_data_add_pagelist(reply
, pagelist
);
3165 ceph_early_kick_flushing_caps(mdsc
, session
);
3167 ceph_con_send(&session
->s_con
, reply
);
3169 mutex_unlock(&session
->s_mutex
);
3171 mutex_lock(&mdsc
->mutex
);
3172 __wake_requests(mdsc
, &session
->s_waiting
);
3173 mutex_unlock(&mdsc
->mutex
);
3175 up_read(&mdsc
->snap_rwsem
);
3179 ceph_msg_put(reply
);
3180 up_read(&mdsc
->snap_rwsem
);
3181 mutex_unlock(&session
->s_mutex
);
3183 ceph_pagelist_release(pagelist
);
3185 pr_err("error %d preparing reconnect for mds%d\n", err
, mds
);
3191 * compare old and new mdsmaps, kicking requests
3192 * and closing out old connections as necessary
3194 * called under mdsc->mutex.
3196 static void check_new_map(struct ceph_mds_client
*mdsc
,
3197 struct ceph_mdsmap
*newmap
,
3198 struct ceph_mdsmap
*oldmap
)
3201 int oldstate
, newstate
;
3202 struct ceph_mds_session
*s
;
3204 dout("check_new_map new %u old %u\n",
3205 newmap
->m_epoch
, oldmap
->m_epoch
);
3207 for (i
= 0; i
< oldmap
->m_num_mds
&& i
< mdsc
->max_sessions
; i
++) {
3208 if (!mdsc
->sessions
[i
])
3210 s
= mdsc
->sessions
[i
];
3211 oldstate
= ceph_mdsmap_get_state(oldmap
, i
);
3212 newstate
= ceph_mdsmap_get_state(newmap
, i
);
3214 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3215 i
, ceph_mds_state_name(oldstate
),
3216 ceph_mdsmap_is_laggy(oldmap
, i
) ? " (laggy)" : "",
3217 ceph_mds_state_name(newstate
),
3218 ceph_mdsmap_is_laggy(newmap
, i
) ? " (laggy)" : "",
3219 ceph_session_state_name(s
->s_state
));
3221 if (i
>= newmap
->m_num_mds
||
3222 memcmp(ceph_mdsmap_get_addr(oldmap
, i
),
3223 ceph_mdsmap_get_addr(newmap
, i
),
3224 sizeof(struct ceph_entity_addr
))) {
3225 if (s
->s_state
== CEPH_MDS_SESSION_OPENING
) {
3226 /* the session never opened, just close it
3229 __unregister_session(mdsc
, s
);
3230 __wake_requests(mdsc
, &s
->s_waiting
);
3231 ceph_put_mds_session(s
);
3232 } else if (i
>= newmap
->m_num_mds
) {
3233 /* force close session for stopped mds */
3235 __unregister_session(mdsc
, s
);
3236 __wake_requests(mdsc
, &s
->s_waiting
);
3237 kick_requests(mdsc
, i
);
3238 mutex_unlock(&mdsc
->mutex
);
3240 mutex_lock(&s
->s_mutex
);
3241 cleanup_session_requests(mdsc
, s
);
3242 remove_session_caps(s
);
3243 mutex_unlock(&s
->s_mutex
);
3245 ceph_put_mds_session(s
);
3247 mutex_lock(&mdsc
->mutex
);
3250 mutex_unlock(&mdsc
->mutex
);
3251 mutex_lock(&s
->s_mutex
);
3252 mutex_lock(&mdsc
->mutex
);
3253 ceph_con_close(&s
->s_con
);
3254 mutex_unlock(&s
->s_mutex
);
3255 s
->s_state
= CEPH_MDS_SESSION_RESTARTING
;
3257 } else if (oldstate
== newstate
) {
3258 continue; /* nothing new with this mds */
3264 if (s
->s_state
== CEPH_MDS_SESSION_RESTARTING
&&
3265 newstate
>= CEPH_MDS_STATE_RECONNECT
) {
3266 mutex_unlock(&mdsc
->mutex
);
3267 send_mds_reconnect(mdsc
, s
);
3268 mutex_lock(&mdsc
->mutex
);
3272 * kick request on any mds that has gone active.
3274 if (oldstate
< CEPH_MDS_STATE_ACTIVE
&&
3275 newstate
>= CEPH_MDS_STATE_ACTIVE
) {
3276 if (oldstate
!= CEPH_MDS_STATE_CREATING
&&
3277 oldstate
!= CEPH_MDS_STATE_STARTING
)
3278 pr_info("mds%d recovery completed\n", s
->s_mds
);
3279 kick_requests(mdsc
, i
);
3280 ceph_kick_flushing_caps(mdsc
, s
);
3281 wake_up_session_caps(s
, 1);
3285 for (i
= 0; i
< newmap
->m_num_mds
&& i
< mdsc
->max_sessions
; i
++) {
3286 s
= mdsc
->sessions
[i
];
3289 if (!ceph_mdsmap_is_laggy(newmap
, i
))
3291 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3292 s
->s_state
== CEPH_MDS_SESSION_HUNG
||
3293 s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3294 dout(" connecting to export targets of laggy mds%d\n",
3296 __open_export_target_sessions(mdsc
, s
);
3308 * caller must hold session s_mutex, dentry->d_lock
3310 void __ceph_mdsc_drop_dentry_lease(struct dentry
*dentry
)
3312 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
3314 ceph_put_mds_session(di
->lease_session
);
3315 di
->lease_session
= NULL
;
3318 static void handle_lease(struct ceph_mds_client
*mdsc
,
3319 struct ceph_mds_session
*session
,
3320 struct ceph_msg
*msg
)
3322 struct super_block
*sb
= mdsc
->fsc
->sb
;
3323 struct inode
*inode
;
3324 struct dentry
*parent
, *dentry
;
3325 struct ceph_dentry_info
*di
;
3326 int mds
= session
->s_mds
;
3327 struct ceph_mds_lease
*h
= msg
->front
.iov_base
;
3329 struct ceph_vino vino
;
3333 dout("handle_lease from mds%d\n", mds
);
3336 if (msg
->front
.iov_len
< sizeof(*h
) + sizeof(u32
))
3338 vino
.ino
= le64_to_cpu(h
->ino
);
3339 vino
.snap
= CEPH_NOSNAP
;
3340 seq
= le32_to_cpu(h
->seq
);
3341 dname
.name
= (void *)h
+ sizeof(*h
) + sizeof(u32
);
3342 dname
.len
= msg
->front
.iov_len
- sizeof(*h
) - sizeof(u32
);
3343 if (dname
.len
!= get_unaligned_le32(h
+1))
3347 inode
= ceph_find_inode(sb
, vino
);
3348 dout("handle_lease %s, ino %llx %p %.*s\n",
3349 ceph_lease_op_name(h
->action
), vino
.ino
, inode
,
3350 dname
.len
, dname
.name
);
3352 mutex_lock(&session
->s_mutex
);
3356 dout("handle_lease no inode %llx\n", vino
.ino
);
3361 parent
= d_find_alias(inode
);
3363 dout("no parent dentry on inode %p\n", inode
);
3365 goto release
; /* hrm... */
3367 dname
.hash
= full_name_hash(parent
, dname
.name
, dname
.len
);
3368 dentry
= d_lookup(parent
, &dname
);
3373 spin_lock(&dentry
->d_lock
);
3374 di
= ceph_dentry(dentry
);
3375 switch (h
->action
) {
3376 case CEPH_MDS_LEASE_REVOKE
:
3377 if (di
->lease_session
== session
) {
3378 if (ceph_seq_cmp(di
->lease_seq
, seq
) > 0)
3379 h
->seq
= cpu_to_le32(di
->lease_seq
);
3380 __ceph_mdsc_drop_dentry_lease(dentry
);
3385 case CEPH_MDS_LEASE_RENEW
:
3386 if (di
->lease_session
== session
&&
3387 di
->lease_gen
== session
->s_cap_gen
&&
3388 di
->lease_renew_from
&&
3389 di
->lease_renew_after
== 0) {
3390 unsigned long duration
=
3391 msecs_to_jiffies(le32_to_cpu(h
->duration_ms
));
3393 di
->lease_seq
= seq
;
3394 di
->time
= di
->lease_renew_from
+ duration
;
3395 di
->lease_renew_after
= di
->lease_renew_from
+
3397 di
->lease_renew_from
= 0;
3401 spin_unlock(&dentry
->d_lock
);
3408 /* let's just reuse the same message */
3409 h
->action
= CEPH_MDS_LEASE_REVOKE_ACK
;
3411 ceph_con_send(&session
->s_con
, msg
);
3415 mutex_unlock(&session
->s_mutex
);
3419 pr_err("corrupt lease message\n");
3423 void ceph_mdsc_lease_send_msg(struct ceph_mds_session
*session
,
3424 struct inode
*inode
,
3425 struct dentry
*dentry
, char action
,
3428 struct ceph_msg
*msg
;
3429 struct ceph_mds_lease
*lease
;
3430 int len
= sizeof(*lease
) + sizeof(u32
);
3433 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3434 inode
, dentry
, ceph_lease_op_name(action
), session
->s_mds
);
3435 dnamelen
= dentry
->d_name
.len
;
3438 msg
= ceph_msg_new(CEPH_MSG_CLIENT_LEASE
, len
, GFP_NOFS
, false);
3441 lease
= msg
->front
.iov_base
;
3442 lease
->action
= action
;
3443 lease
->ino
= cpu_to_le64(ceph_vino(inode
).ino
);
3444 lease
->first
= lease
->last
= cpu_to_le64(ceph_vino(inode
).snap
);
3445 lease
->seq
= cpu_to_le32(seq
);
3446 put_unaligned_le32(dnamelen
, lease
+ 1);
3447 memcpy((void *)(lease
+ 1) + 4, dentry
->d_name
.name
, dnamelen
);
3450 * if this is a preemptive lease RELEASE, no need to
3451 * flush request stream, since the actual request will
3454 msg
->more_to_follow
= (action
== CEPH_MDS_LEASE_RELEASE
);
3456 ceph_con_send(&session
->s_con
, msg
);
3460 * drop all leases (and dentry refs) in preparation for umount
3462 static void drop_leases(struct ceph_mds_client
*mdsc
)
3466 dout("drop_leases\n");
3467 mutex_lock(&mdsc
->mutex
);
3468 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3469 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3472 mutex_unlock(&mdsc
->mutex
);
3473 mutex_lock(&s
->s_mutex
);
3474 mutex_unlock(&s
->s_mutex
);
3475 ceph_put_mds_session(s
);
3476 mutex_lock(&mdsc
->mutex
);
3478 mutex_unlock(&mdsc
->mutex
);
3484 * delayed work -- periodically trim expired leases, renew caps with mds
3486 static void schedule_delayed(struct ceph_mds_client
*mdsc
)
3489 unsigned hz
= round_jiffies_relative(HZ
* delay
);
3490 schedule_delayed_work(&mdsc
->delayed_work
, hz
);
3493 static void delayed_work(struct work_struct
*work
)
3496 struct ceph_mds_client
*mdsc
=
3497 container_of(work
, struct ceph_mds_client
, delayed_work
.work
);
3501 dout("mdsc delayed_work\n");
3502 ceph_check_delayed_caps(mdsc
);
3504 mutex_lock(&mdsc
->mutex
);
3505 renew_interval
= mdsc
->mdsmap
->m_session_timeout
>> 2;
3506 renew_caps
= time_after_eq(jiffies
, HZ
*renew_interval
+
3507 mdsc
->last_renew_caps
);
3509 mdsc
->last_renew_caps
= jiffies
;
3511 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3512 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3515 if (s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3516 dout("resending session close request for mds%d\n",
3518 request_close_session(mdsc
, s
);
3519 ceph_put_mds_session(s
);
3522 if (s
->s_ttl
&& time_after(jiffies
, s
->s_ttl
)) {
3523 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
) {
3524 s
->s_state
= CEPH_MDS_SESSION_HUNG
;
3525 pr_info("mds%d hung\n", s
->s_mds
);
3528 if (s
->s_state
< CEPH_MDS_SESSION_OPEN
) {
3529 /* this mds is failed or recovering, just wait */
3530 ceph_put_mds_session(s
);
3533 mutex_unlock(&mdsc
->mutex
);
3535 mutex_lock(&s
->s_mutex
);
3537 send_renew_caps(mdsc
, s
);
3539 ceph_con_keepalive(&s
->s_con
);
3540 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3541 s
->s_state
== CEPH_MDS_SESSION_HUNG
)
3542 ceph_send_cap_releases(mdsc
, s
);
3543 mutex_unlock(&s
->s_mutex
);
3544 ceph_put_mds_session(s
);
3546 mutex_lock(&mdsc
->mutex
);
3548 mutex_unlock(&mdsc
->mutex
);
3550 schedule_delayed(mdsc
);
3553 int ceph_mdsc_init(struct ceph_fs_client
*fsc
)
3556 struct ceph_mds_client
*mdsc
;
3558 mdsc
= kzalloc(sizeof(struct ceph_mds_client
), GFP_NOFS
);
3563 mutex_init(&mdsc
->mutex
);
3564 mdsc
->mdsmap
= kzalloc(sizeof(*mdsc
->mdsmap
), GFP_NOFS
);
3565 if (!mdsc
->mdsmap
) {
3570 init_completion(&mdsc
->safe_umount_waiters
);
3571 init_waitqueue_head(&mdsc
->session_close_wq
);
3572 INIT_LIST_HEAD(&mdsc
->waiting_for_map
);
3573 mdsc
->sessions
= NULL
;
3574 atomic_set(&mdsc
->num_sessions
, 0);
3575 mdsc
->max_sessions
= 0;
3577 mdsc
->last_snap_seq
= 0;
3578 init_rwsem(&mdsc
->snap_rwsem
);
3579 mdsc
->snap_realms
= RB_ROOT
;
3580 INIT_LIST_HEAD(&mdsc
->snap_empty
);
3581 spin_lock_init(&mdsc
->snap_empty_lock
);
3583 mdsc
->oldest_tid
= 0;
3584 mdsc
->request_tree
= RB_ROOT
;
3585 INIT_DELAYED_WORK(&mdsc
->delayed_work
, delayed_work
);
3586 mdsc
->last_renew_caps
= jiffies
;
3587 INIT_LIST_HEAD(&mdsc
->cap_delay_list
);
3588 spin_lock_init(&mdsc
->cap_delay_lock
);
3589 INIT_LIST_HEAD(&mdsc
->snap_flush_list
);
3590 spin_lock_init(&mdsc
->snap_flush_lock
);
3591 mdsc
->last_cap_flush_tid
= 1;
3592 INIT_LIST_HEAD(&mdsc
->cap_flush_list
);
3593 INIT_LIST_HEAD(&mdsc
->cap_dirty
);
3594 INIT_LIST_HEAD(&mdsc
->cap_dirty_migrating
);
3595 mdsc
->num_cap_flushing
= 0;
3596 spin_lock_init(&mdsc
->cap_dirty_lock
);
3597 init_waitqueue_head(&mdsc
->cap_flushing_wq
);
3598 spin_lock_init(&mdsc
->dentry_lru_lock
);
3599 INIT_LIST_HEAD(&mdsc
->dentry_lru
);
3601 ceph_caps_init(mdsc
);
3602 ceph_adjust_min_caps(mdsc
, fsc
->min_caps
);
3604 init_rwsem(&mdsc
->pool_perm_rwsem
);
3605 mdsc
->pool_perm_tree
= RB_ROOT
;
3607 strncpy(mdsc
->nodename
, utsname()->nodename
,
3608 sizeof(mdsc
->nodename
) - 1);
3613 * Wait for safe replies on open mds requests. If we time out, drop
3614 * all requests from the tree to avoid dangling dentry refs.
3616 static void wait_requests(struct ceph_mds_client
*mdsc
)
3618 struct ceph_options
*opts
= mdsc
->fsc
->client
->options
;
3619 struct ceph_mds_request
*req
;
3621 mutex_lock(&mdsc
->mutex
);
3622 if (__get_oldest_req(mdsc
)) {
3623 mutex_unlock(&mdsc
->mutex
);
3625 dout("wait_requests waiting for requests\n");
3626 wait_for_completion_timeout(&mdsc
->safe_umount_waiters
,
3627 ceph_timeout_jiffies(opts
->mount_timeout
));
3629 /* tear down remaining requests */
3630 mutex_lock(&mdsc
->mutex
);
3631 while ((req
= __get_oldest_req(mdsc
))) {
3632 dout("wait_requests timed out on tid %llu\n",
3634 __unregister_request(mdsc
, req
);
3637 mutex_unlock(&mdsc
->mutex
);
3638 dout("wait_requests done\n");
3642 * called before mount is ro, and before dentries are torn down.
3643 * (hmm, does this still race with new lookups?)
3645 void ceph_mdsc_pre_umount(struct ceph_mds_client
*mdsc
)
3647 dout("pre_umount\n");
3651 ceph_flush_dirty_caps(mdsc
);
3652 wait_requests(mdsc
);
3655 * wait for reply handlers to drop their request refs and
3656 * their inode/dcache refs
3662 * wait for all write mds requests to flush.
3664 static void wait_unsafe_requests(struct ceph_mds_client
*mdsc
, u64 want_tid
)
3666 struct ceph_mds_request
*req
= NULL
, *nextreq
;
3669 mutex_lock(&mdsc
->mutex
);
3670 dout("wait_unsafe_requests want %lld\n", want_tid
);
3672 req
= __get_oldest_req(mdsc
);
3673 while (req
&& req
->r_tid
<= want_tid
) {
3674 /* find next request */
3675 n
= rb_next(&req
->r_node
);
3677 nextreq
= rb_entry(n
, struct ceph_mds_request
, r_node
);
3680 if (req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
&&
3681 (req
->r_op
& CEPH_MDS_OP_WRITE
)) {
3683 ceph_mdsc_get_request(req
);
3685 ceph_mdsc_get_request(nextreq
);
3686 mutex_unlock(&mdsc
->mutex
);
3687 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3688 req
->r_tid
, want_tid
);
3689 wait_for_completion(&req
->r_safe_completion
);
3690 mutex_lock(&mdsc
->mutex
);
3691 ceph_mdsc_put_request(req
);
3693 break; /* next dne before, so we're done! */
3694 if (RB_EMPTY_NODE(&nextreq
->r_node
)) {
3695 /* next request was removed from tree */
3696 ceph_mdsc_put_request(nextreq
);
3699 ceph_mdsc_put_request(nextreq
); /* won't go away */
3703 mutex_unlock(&mdsc
->mutex
);
3704 dout("wait_unsafe_requests done\n");
3707 void ceph_mdsc_sync(struct ceph_mds_client
*mdsc
)
3709 u64 want_tid
, want_flush
;
3711 if (READ_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
3715 mutex_lock(&mdsc
->mutex
);
3716 want_tid
= mdsc
->last_tid
;
3717 mutex_unlock(&mdsc
->mutex
);
3719 ceph_flush_dirty_caps(mdsc
);
3720 spin_lock(&mdsc
->cap_dirty_lock
);
3721 want_flush
= mdsc
->last_cap_flush_tid
;
3722 if (!list_empty(&mdsc
->cap_flush_list
)) {
3723 struct ceph_cap_flush
*cf
=
3724 list_last_entry(&mdsc
->cap_flush_list
,
3725 struct ceph_cap_flush
, g_list
);
3728 spin_unlock(&mdsc
->cap_dirty_lock
);
3730 dout("sync want tid %lld flush_seq %lld\n",
3731 want_tid
, want_flush
);
3733 wait_unsafe_requests(mdsc
, want_tid
);
3734 wait_caps_flush(mdsc
, want_flush
);
3738 * true if all sessions are closed, or we force unmount
3740 static bool done_closing_sessions(struct ceph_mds_client
*mdsc
, int skipped
)
3742 if (READ_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
3744 return atomic_read(&mdsc
->num_sessions
) <= skipped
;
3748 * called after sb is ro.
3750 void ceph_mdsc_close_sessions(struct ceph_mds_client
*mdsc
)
3752 struct ceph_options
*opts
= mdsc
->fsc
->client
->options
;
3753 struct ceph_mds_session
*session
;
3757 dout("close_sessions\n");
3759 /* close sessions */
3760 mutex_lock(&mdsc
->mutex
);
3761 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3762 session
= __ceph_lookup_mds_session(mdsc
, i
);
3765 mutex_unlock(&mdsc
->mutex
);
3766 mutex_lock(&session
->s_mutex
);
3767 if (__close_session(mdsc
, session
) <= 0)
3769 mutex_unlock(&session
->s_mutex
);
3770 ceph_put_mds_session(session
);
3771 mutex_lock(&mdsc
->mutex
);
3773 mutex_unlock(&mdsc
->mutex
);
3775 dout("waiting for sessions to close\n");
3776 wait_event_timeout(mdsc
->session_close_wq
,
3777 done_closing_sessions(mdsc
, skipped
),
3778 ceph_timeout_jiffies(opts
->mount_timeout
));
3780 /* tear down remaining sessions */
3781 mutex_lock(&mdsc
->mutex
);
3782 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3783 if (mdsc
->sessions
[i
]) {
3784 session
= get_session(mdsc
->sessions
[i
]);
3785 __unregister_session(mdsc
, session
);
3786 mutex_unlock(&mdsc
->mutex
);
3787 mutex_lock(&session
->s_mutex
);
3788 remove_session_caps(session
);
3789 mutex_unlock(&session
->s_mutex
);
3790 ceph_put_mds_session(session
);
3791 mutex_lock(&mdsc
->mutex
);
3794 WARN_ON(!list_empty(&mdsc
->cap_delay_list
));
3795 mutex_unlock(&mdsc
->mutex
);
3797 ceph_cleanup_empty_realms(mdsc
);
3799 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3804 void ceph_mdsc_force_umount(struct ceph_mds_client
*mdsc
)
3806 struct ceph_mds_session
*session
;
3809 dout("force umount\n");
3811 mutex_lock(&mdsc
->mutex
);
3812 for (mds
= 0; mds
< mdsc
->max_sessions
; mds
++) {
3813 session
= __ceph_lookup_mds_session(mdsc
, mds
);
3816 mutex_unlock(&mdsc
->mutex
);
3817 mutex_lock(&session
->s_mutex
);
3818 __close_session(mdsc
, session
);
3819 if (session
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3820 cleanup_session_requests(mdsc
, session
);
3821 remove_session_caps(session
);
3823 mutex_unlock(&session
->s_mutex
);
3824 ceph_put_mds_session(session
);
3825 mutex_lock(&mdsc
->mutex
);
3826 kick_requests(mdsc
, mds
);
3828 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3829 mutex_unlock(&mdsc
->mutex
);
3832 static void ceph_mdsc_stop(struct ceph_mds_client
*mdsc
)
3835 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3837 ceph_mdsmap_destroy(mdsc
->mdsmap
);
3838 kfree(mdsc
->sessions
);
3839 ceph_caps_finalize(mdsc
);
3840 ceph_pool_perm_destroy(mdsc
);
3843 void ceph_mdsc_destroy(struct ceph_fs_client
*fsc
)
3845 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
3846 dout("mdsc_destroy %p\n", mdsc
);
3848 /* flush out any connection work with references to us */
3851 ceph_mdsc_stop(mdsc
);
3855 dout("mdsc_destroy %p done\n", mdsc
);
3858 void ceph_mdsc_handle_fsmap(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3860 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3861 const char *mds_namespace
= fsc
->mount_options
->mds_namespace
;
3862 void *p
= msg
->front
.iov_base
;
3863 void *end
= p
+ msg
->front
.iov_len
;
3867 u32 mount_fscid
= (u32
)-1;
3868 u8 struct_v
, struct_cv
;
3871 ceph_decode_need(&p
, end
, sizeof(u32
), bad
);
3872 epoch
= ceph_decode_32(&p
);
3874 dout("handle_fsmap epoch %u\n", epoch
);
3876 ceph_decode_need(&p
, end
, 2 + sizeof(u32
), bad
);
3877 struct_v
= ceph_decode_8(&p
);
3878 struct_cv
= ceph_decode_8(&p
);
3879 map_len
= ceph_decode_32(&p
);
3881 ceph_decode_need(&p
, end
, sizeof(u32
) * 3, bad
);
3882 p
+= sizeof(u32
) * 2; /* skip epoch and legacy_client_fscid */
3884 num_fs
= ceph_decode_32(&p
);
3885 while (num_fs
-- > 0) {
3886 void *info_p
, *info_end
;
3891 ceph_decode_need(&p
, end
, 2 + sizeof(u32
), bad
);
3892 info_v
= ceph_decode_8(&p
);
3893 info_cv
= ceph_decode_8(&p
);
3894 info_len
= ceph_decode_32(&p
);
3895 ceph_decode_need(&p
, end
, info_len
, bad
);
3897 info_end
= p
+ info_len
;
3900 ceph_decode_need(&info_p
, info_end
, sizeof(u32
) * 2, bad
);
3901 fscid
= ceph_decode_32(&info_p
);
3902 namelen
= ceph_decode_32(&info_p
);
3903 ceph_decode_need(&info_p
, info_end
, namelen
, bad
);
3905 if (mds_namespace
&&
3906 strlen(mds_namespace
) == namelen
&&
3907 !strncmp(mds_namespace
, (char *)info_p
, namelen
)) {
3908 mount_fscid
= fscid
;
3913 ceph_monc_got_map(&fsc
->client
->monc
, CEPH_SUB_FSMAP
, epoch
);
3914 if (mount_fscid
!= (u32
)-1) {
3915 fsc
->client
->monc
.fs_cluster_id
= mount_fscid
;
3916 ceph_monc_want_map(&fsc
->client
->monc
, CEPH_SUB_MDSMAP
,
3918 ceph_monc_renew_subs(&fsc
->client
->monc
);
3926 pr_err("error decoding fsmap\n");
3928 mutex_lock(&mdsc
->mutex
);
3929 mdsc
->mdsmap_err
= err
;
3930 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3931 mutex_unlock(&mdsc
->mutex
);
3935 * handle mds map update.
3937 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3941 void *p
= msg
->front
.iov_base
;
3942 void *end
= p
+ msg
->front
.iov_len
;
3943 struct ceph_mdsmap
*newmap
, *oldmap
;
3944 struct ceph_fsid fsid
;
3947 ceph_decode_need(&p
, end
, sizeof(fsid
)+2*sizeof(u32
), bad
);
3948 ceph_decode_copy(&p
, &fsid
, sizeof(fsid
));
3949 if (ceph_check_fsid(mdsc
->fsc
->client
, &fsid
) < 0)
3951 epoch
= ceph_decode_32(&p
);
3952 maplen
= ceph_decode_32(&p
);
3953 dout("handle_map epoch %u len %d\n", epoch
, (int)maplen
);
3955 /* do we need it? */
3956 mutex_lock(&mdsc
->mutex
);
3957 if (mdsc
->mdsmap
&& epoch
<= mdsc
->mdsmap
->m_epoch
) {
3958 dout("handle_map epoch %u <= our %u\n",
3959 epoch
, mdsc
->mdsmap
->m_epoch
);
3960 mutex_unlock(&mdsc
->mutex
);
3964 newmap
= ceph_mdsmap_decode(&p
, end
);
3965 if (IS_ERR(newmap
)) {
3966 err
= PTR_ERR(newmap
);
3970 /* swap into place */
3972 oldmap
= mdsc
->mdsmap
;
3973 mdsc
->mdsmap
= newmap
;
3974 check_new_map(mdsc
, newmap
, oldmap
);
3975 ceph_mdsmap_destroy(oldmap
);
3977 mdsc
->mdsmap
= newmap
; /* first mds map */
3979 mdsc
->fsc
->sb
->s_maxbytes
= mdsc
->mdsmap
->m_max_file_size
;
3981 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3982 ceph_monc_got_map(&mdsc
->fsc
->client
->monc
, CEPH_SUB_MDSMAP
,
3983 mdsc
->mdsmap
->m_epoch
);
3985 mutex_unlock(&mdsc
->mutex
);
3986 schedule_delayed(mdsc
);
3990 mutex_unlock(&mdsc
->mutex
);
3992 pr_err("error decoding mdsmap %d\n", err
);
3996 static struct ceph_connection
*con_get(struct ceph_connection
*con
)
3998 struct ceph_mds_session
*s
= con
->private;
4000 if (get_session(s
)) {
4001 dout("mdsc con_get %p ok (%d)\n", s
, refcount_read(&s
->s_ref
));
4004 dout("mdsc con_get %p FAIL\n", s
);
4008 static void con_put(struct ceph_connection
*con
)
4010 struct ceph_mds_session
*s
= con
->private;
4012 dout("mdsc con_put %p (%d)\n", s
, refcount_read(&s
->s_ref
) - 1);
4013 ceph_put_mds_session(s
);
4017 * if the client is unresponsive for long enough, the mds will kill
4018 * the session entirely.
4020 static void peer_reset(struct ceph_connection
*con
)
4022 struct ceph_mds_session
*s
= con
->private;
4023 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4025 pr_warn("mds%d closed our session\n", s
->s_mds
);
4026 send_mds_reconnect(mdsc
, s
);
4029 static void dispatch(struct ceph_connection
*con
, struct ceph_msg
*msg
)
4031 struct ceph_mds_session
*s
= con
->private;
4032 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4033 int type
= le16_to_cpu(msg
->hdr
.type
);
4035 mutex_lock(&mdsc
->mutex
);
4036 if (__verify_registered_session(mdsc
, s
) < 0) {
4037 mutex_unlock(&mdsc
->mutex
);
4040 mutex_unlock(&mdsc
->mutex
);
4043 case CEPH_MSG_MDS_MAP
:
4044 ceph_mdsc_handle_mdsmap(mdsc
, msg
);
4046 case CEPH_MSG_FS_MAP_USER
:
4047 ceph_mdsc_handle_fsmap(mdsc
, msg
);
4049 case CEPH_MSG_CLIENT_SESSION
:
4050 handle_session(s
, msg
);
4052 case CEPH_MSG_CLIENT_REPLY
:
4053 handle_reply(s
, msg
);
4055 case CEPH_MSG_CLIENT_REQUEST_FORWARD
:
4056 handle_forward(mdsc
, s
, msg
);
4058 case CEPH_MSG_CLIENT_CAPS
:
4059 ceph_handle_caps(s
, msg
);
4061 case CEPH_MSG_CLIENT_SNAP
:
4062 ceph_handle_snap(mdsc
, s
, msg
);
4064 case CEPH_MSG_CLIENT_LEASE
:
4065 handle_lease(mdsc
, s
, msg
);
4069 pr_err("received unknown message type %d %s\n", type
,
4070 ceph_msg_type_name(type
));
4081 * Note: returned pointer is the address of a structure that's
4082 * managed separately. Caller must *not* attempt to free it.
4084 static struct ceph_auth_handshake
*get_authorizer(struct ceph_connection
*con
,
4085 int *proto
, int force_new
)
4087 struct ceph_mds_session
*s
= con
->private;
4088 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4089 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
4090 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
4092 if (force_new
&& auth
->authorizer
) {
4093 ceph_auth_destroy_authorizer(auth
->authorizer
);
4094 auth
->authorizer
= NULL
;
4096 if (!auth
->authorizer
) {
4097 int ret
= ceph_auth_create_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
4100 return ERR_PTR(ret
);
4102 int ret
= ceph_auth_update_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
4105 return ERR_PTR(ret
);
4107 *proto
= ac
->protocol
;
4113 static int verify_authorizer_reply(struct ceph_connection
*con
)
4115 struct ceph_mds_session
*s
= con
->private;
4116 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4117 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
4119 return ceph_auth_verify_authorizer_reply(ac
, s
->s_auth
.authorizer
);
4122 static int invalidate_authorizer(struct ceph_connection
*con
)
4124 struct ceph_mds_session
*s
= con
->private;
4125 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4126 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
4128 ceph_auth_invalidate_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
);
4130 return ceph_monc_validate_auth(&mdsc
->fsc
->client
->monc
);
4133 static struct ceph_msg
*mds_alloc_msg(struct ceph_connection
*con
,
4134 struct ceph_msg_header
*hdr
, int *skip
)
4136 struct ceph_msg
*msg
;
4137 int type
= (int) le16_to_cpu(hdr
->type
);
4138 int front_len
= (int) le32_to_cpu(hdr
->front_len
);
4144 msg
= ceph_msg_new(type
, front_len
, GFP_NOFS
, false);
4146 pr_err("unable to allocate msg type %d len %d\n",
4154 static int mds_sign_message(struct ceph_msg
*msg
)
4156 struct ceph_mds_session
*s
= msg
->con
->private;
4157 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
4159 return ceph_auth_sign_message(auth
, msg
);
4162 static int mds_check_message_signature(struct ceph_msg
*msg
)
4164 struct ceph_mds_session
*s
= msg
->con
->private;
4165 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
4167 return ceph_auth_check_message_signature(auth
, msg
);
4170 static const struct ceph_connection_operations mds_con_ops
= {
4173 .dispatch
= dispatch
,
4174 .get_authorizer
= get_authorizer
,
4175 .verify_authorizer_reply
= verify_authorizer_reply
,
4176 .invalidate_authorizer
= invalidate_authorizer
,
4177 .peer_reset
= peer_reset
,
4178 .alloc_msg
= mds_alloc_msg
,
4179 .sign_message
= mds_sign_message
,
4180 .check_message_signature
= mds_check_message_signature
,