1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
7 #include <linux/sched.h>
8 #include <linux/debugfs.h>
9 #include <linux/seq_file.h>
10 #include <linux/utsname.h>
11 #include <linux/ratelimit.h>
14 #include "mds_client.h"
16 #include <linux/ceph/ceph_features.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/pagelist.h>
20 #include <linux/ceph/auth.h>
21 #include <linux/ceph/debugfs.h>
24 * A cluster of MDS (metadata server) daemons is responsible for
25 * managing the file system namespace (the directory hierarchy and
26 * inodes) and for coordinating shared access to storage. Metadata is
27 * partitioning hierarchically across a number of servers, and that
28 * partition varies over time as the cluster adjusts the distribution
29 * in order to balance load.
31 * The MDS client is primarily responsible to managing synchronous
32 * metadata requests for operations like open, unlink, and so forth.
33 * If there is a MDS failure, we find out about it when we (possibly
34 * request and) receive a new MDS map, and can resubmit affected
37 * For the most part, though, we take advantage of a lossless
38 * communications channel to the MDS, and do not need to worry about
39 * timing out or resubmitting requests.
41 * We maintain a stateful "session" with each MDS we interact with.
42 * Within each session, we sent periodic heartbeat messages to ensure
43 * any capabilities or leases we have been issues remain valid. If
44 * the session times out and goes stale, our leases and capabilities
45 * are no longer valid.
48 struct ceph_reconnect_state
{
50 struct ceph_pagelist
*pagelist
;
54 static void __wake_requests(struct ceph_mds_client
*mdsc
,
55 struct list_head
*head
);
57 static const struct ceph_connection_operations mds_con_ops
;
65 * parse individual inode info
67 static int parse_reply_info_in(void **p
, void *end
,
68 struct ceph_mds_reply_info_in
*info
,
74 *p
+= sizeof(struct ceph_mds_reply_inode
) +
75 sizeof(*info
->in
->fragtree
.splits
) *
76 le32_to_cpu(info
->in
->fragtree
.nsplits
);
78 ceph_decode_32_safe(p
, end
, info
->symlink_len
, bad
);
79 ceph_decode_need(p
, end
, info
->symlink_len
, bad
);
81 *p
+= info
->symlink_len
;
83 if (features
& CEPH_FEATURE_DIRLAYOUTHASH
)
84 ceph_decode_copy_safe(p
, end
, &info
->dir_layout
,
85 sizeof(info
->dir_layout
), bad
);
87 memset(&info
->dir_layout
, 0, sizeof(info
->dir_layout
));
89 ceph_decode_32_safe(p
, end
, info
->xattr_len
, bad
);
90 ceph_decode_need(p
, end
, info
->xattr_len
, bad
);
91 info
->xattr_data
= *p
;
92 *p
+= info
->xattr_len
;
94 if (features
& CEPH_FEATURE_MDS_INLINE_DATA
) {
95 ceph_decode_64_safe(p
, end
, info
->inline_version
, bad
);
96 ceph_decode_32_safe(p
, end
, info
->inline_len
, bad
);
97 ceph_decode_need(p
, end
, info
->inline_len
, bad
);
98 info
->inline_data
= *p
;
99 *p
+= info
->inline_len
;
101 info
->inline_version
= CEPH_INLINE_NONE
;
103 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
] == NULL
)
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 inode
= req
->r_inode
;
736 } else if (req
->r_dentry
) {
737 /* ignore race with rename; old or new d_parent is okay */
738 struct dentry
*parent
;
742 parent
= req
->r_dentry
->d_parent
;
743 dir
= req
->r_parent
? : d_inode_rcu(parent
);
745 if (!dir
|| dir
->i_sb
!= mdsc
->fsc
->sb
) {
746 /* not this fs or parent went negative */
747 inode
= d_inode(req
->r_dentry
);
750 } else if (ceph_snap(dir
) != CEPH_NOSNAP
) {
751 /* direct snapped/virtual snapdir requests
752 * based on parent dir inode */
753 inode
= get_nonsnap_parent(parent
);
754 dout("__choose_mds using nonsnap parent %p\n", inode
);
757 inode
= d_inode(req
->r_dentry
);
758 if (!inode
|| mode
== USE_AUTH_MDS
) {
761 hash
= ceph_dentry_hash(dir
, req
->r_dentry
);
770 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode
, (int)is_hash
,
774 ci
= ceph_inode(inode
);
776 if (is_hash
&& S_ISDIR(inode
->i_mode
)) {
777 struct ceph_inode_frag frag
;
780 ceph_choose_frag(ci
, hash
, &frag
, &found
);
782 if (mode
== USE_ANY_MDS
&& frag
.ndist
> 0) {
785 /* choose a random replica */
786 get_random_bytes(&r
, 1);
789 dout("choose_mds %p %llx.%llx "
790 "frag %u mds%d (%d/%d)\n",
791 inode
, ceph_vinop(inode
),
794 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
795 CEPH_MDS_STATE_ACTIVE
)
799 /* since this file/dir wasn't known to be
800 * replicated, then we want to look for the
801 * authoritative mds. */
804 /* choose auth mds */
806 dout("choose_mds %p %llx.%llx "
807 "frag %u mds%d (auth)\n",
808 inode
, ceph_vinop(inode
), frag
.frag
, mds
);
809 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
810 CEPH_MDS_STATE_ACTIVE
)
816 spin_lock(&ci
->i_ceph_lock
);
818 if (mode
== USE_AUTH_MDS
)
819 cap
= ci
->i_auth_cap
;
820 if (!cap
&& !RB_EMPTY_ROOT(&ci
->i_caps
))
821 cap
= rb_entry(rb_first(&ci
->i_caps
), struct ceph_cap
, ci_node
);
823 spin_unlock(&ci
->i_ceph_lock
);
827 mds
= cap
->session
->s_mds
;
828 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
829 inode
, ceph_vinop(inode
), mds
,
830 cap
== ci
->i_auth_cap
? "auth " : "", cap
);
831 spin_unlock(&ci
->i_ceph_lock
);
837 mds
= ceph_mdsmap_get_random_mds(mdsc
->mdsmap
);
838 dout("choose_mds chose random mds%d\n", mds
);
846 static struct ceph_msg
*create_session_msg(u32 op
, u64 seq
)
848 struct ceph_msg
*msg
;
849 struct ceph_mds_session_head
*h
;
851 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
), GFP_NOFS
,
854 pr_err("create_session_msg ENOMEM creating msg\n");
857 h
= msg
->front
.iov_base
;
858 h
->op
= cpu_to_le32(op
);
859 h
->seq
= cpu_to_le64(seq
);
865 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
866 * to include additional client metadata fields.
868 static struct ceph_msg
*create_session_open_msg(struct ceph_mds_client
*mdsc
, u64 seq
)
870 struct ceph_msg
*msg
;
871 struct ceph_mds_session_head
*h
;
873 int metadata_bytes
= 0;
874 int metadata_key_count
= 0;
875 struct ceph_options
*opt
= mdsc
->fsc
->client
->options
;
876 struct ceph_mount_options
*fsopt
= mdsc
->fsc
->mount_options
;
879 const char* metadata
[][2] = {
880 {"hostname", utsname()->nodename
},
881 {"kernel_version", utsname()->release
},
882 {"entity_id", opt
->name
? : ""},
883 {"root", fsopt
->server_path
? : "/"},
887 /* Calculate serialized length of metadata */
888 metadata_bytes
= 4; /* map length */
889 for (i
= 0; metadata
[i
][0] != NULL
; ++i
) {
890 metadata_bytes
+= 8 + strlen(metadata
[i
][0]) +
891 strlen(metadata
[i
][1]);
892 metadata_key_count
++;
895 /* Allocate the message */
896 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
) + metadata_bytes
,
899 pr_err("create_session_msg ENOMEM creating msg\n");
902 h
= msg
->front
.iov_base
;
903 h
->op
= cpu_to_le32(CEPH_SESSION_REQUEST_OPEN
);
904 h
->seq
= cpu_to_le64(seq
);
907 * Serialize client metadata into waiting buffer space, using
908 * the format that userspace expects for map<string, string>
910 * ClientSession messages with metadata are v2
912 msg
->hdr
.version
= cpu_to_le16(2);
913 msg
->hdr
.compat_version
= cpu_to_le16(1);
915 /* The write pointer, following the session_head structure */
916 p
= msg
->front
.iov_base
+ sizeof(*h
);
918 /* Number of entries in the map */
919 ceph_encode_32(&p
, metadata_key_count
);
921 /* Two length-prefixed strings for each entry in the map */
922 for (i
= 0; metadata
[i
][0] != NULL
; ++i
) {
923 size_t const key_len
= strlen(metadata
[i
][0]);
924 size_t const val_len
= strlen(metadata
[i
][1]);
926 ceph_encode_32(&p
, key_len
);
927 memcpy(p
, metadata
[i
][0], key_len
);
929 ceph_encode_32(&p
, val_len
);
930 memcpy(p
, metadata
[i
][1], val_len
);
938 * send session open request.
940 * called under mdsc->mutex
942 static int __open_session(struct ceph_mds_client
*mdsc
,
943 struct ceph_mds_session
*session
)
945 struct ceph_msg
*msg
;
947 int mds
= session
->s_mds
;
949 /* wait for mds to go active? */
950 mstate
= ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
);
951 dout("open_session to mds%d (%s)\n", mds
,
952 ceph_mds_state_name(mstate
));
953 session
->s_state
= CEPH_MDS_SESSION_OPENING
;
954 session
->s_renew_requested
= jiffies
;
956 /* send connect message */
957 msg
= create_session_open_msg(mdsc
, session
->s_seq
);
960 ceph_con_send(&session
->s_con
, msg
);
965 * open sessions for any export targets for the given mds
967 * called under mdsc->mutex
969 static struct ceph_mds_session
*
970 __open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
972 struct ceph_mds_session
*session
;
974 session
= __ceph_lookup_mds_session(mdsc
, target
);
976 session
= register_session(mdsc
, target
);
980 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
981 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
982 __open_session(mdsc
, session
);
987 struct ceph_mds_session
*
988 ceph_mdsc_open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
990 struct ceph_mds_session
*session
;
992 dout("open_export_target_session to mds%d\n", target
);
994 mutex_lock(&mdsc
->mutex
);
995 session
= __open_export_target_session(mdsc
, target
);
996 mutex_unlock(&mdsc
->mutex
);
1001 static void __open_export_target_sessions(struct ceph_mds_client
*mdsc
,
1002 struct ceph_mds_session
*session
)
1004 struct ceph_mds_info
*mi
;
1005 struct ceph_mds_session
*ts
;
1006 int i
, mds
= session
->s_mds
;
1008 if (mds
>= mdsc
->mdsmap
->m_num_mds
)
1011 mi
= &mdsc
->mdsmap
->m_info
[mds
];
1012 dout("open_export_target_sessions for mds%d (%d targets)\n",
1013 session
->s_mds
, mi
->num_export_targets
);
1015 for (i
= 0; i
< mi
->num_export_targets
; i
++) {
1016 ts
= __open_export_target_session(mdsc
, mi
->export_targets
[i
]);
1018 ceph_put_mds_session(ts
);
1022 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client
*mdsc
,
1023 struct ceph_mds_session
*session
)
1025 mutex_lock(&mdsc
->mutex
);
1026 __open_export_target_sessions(mdsc
, session
);
1027 mutex_unlock(&mdsc
->mutex
);
1034 /* caller holds s_cap_lock, we drop it */
1035 static void cleanup_cap_releases(struct ceph_mds_client
*mdsc
,
1036 struct ceph_mds_session
*session
)
1037 __releases(session
->s_cap_lock
)
1039 LIST_HEAD(tmp_list
);
1040 list_splice_init(&session
->s_cap_releases
, &tmp_list
);
1041 session
->s_num_cap_releases
= 0;
1042 spin_unlock(&session
->s_cap_lock
);
1044 dout("cleanup_cap_releases mds%d\n", session
->s_mds
);
1045 while (!list_empty(&tmp_list
)) {
1046 struct ceph_cap
*cap
;
1047 /* zero out the in-progress message */
1048 cap
= list_first_entry(&tmp_list
,
1049 struct ceph_cap
, session_caps
);
1050 list_del(&cap
->session_caps
);
1051 ceph_put_cap(mdsc
, cap
);
1055 static void cleanup_session_requests(struct ceph_mds_client
*mdsc
,
1056 struct ceph_mds_session
*session
)
1058 struct ceph_mds_request
*req
;
1061 dout("cleanup_session_requests mds%d\n", session
->s_mds
);
1062 mutex_lock(&mdsc
->mutex
);
1063 while (!list_empty(&session
->s_unsafe
)) {
1064 req
= list_first_entry(&session
->s_unsafe
,
1065 struct ceph_mds_request
, r_unsafe_item
);
1066 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1068 __unregister_request(mdsc
, req
);
1070 /* zero r_attempts, so kick_requests() will re-send requests */
1071 p
= rb_first(&mdsc
->request_tree
);
1073 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
1075 if (req
->r_session
&&
1076 req
->r_session
->s_mds
== session
->s_mds
)
1077 req
->r_attempts
= 0;
1079 mutex_unlock(&mdsc
->mutex
);
1083 * Helper to safely iterate over all caps associated with a session, with
1084 * special care taken to handle a racing __ceph_remove_cap().
1086 * Caller must hold session s_mutex.
1088 static int iterate_session_caps(struct ceph_mds_session
*session
,
1089 int (*cb
)(struct inode
*, struct ceph_cap
*,
1092 struct list_head
*p
;
1093 struct ceph_cap
*cap
;
1094 struct inode
*inode
, *last_inode
= NULL
;
1095 struct ceph_cap
*old_cap
= NULL
;
1098 dout("iterate_session_caps %p mds%d\n", session
, session
->s_mds
);
1099 spin_lock(&session
->s_cap_lock
);
1100 p
= session
->s_caps
.next
;
1101 while (p
!= &session
->s_caps
) {
1102 cap
= list_entry(p
, struct ceph_cap
, session_caps
);
1103 inode
= igrab(&cap
->ci
->vfs_inode
);
1108 session
->s_cap_iterator
= cap
;
1109 spin_unlock(&session
->s_cap_lock
);
1116 ceph_put_cap(session
->s_mdsc
, old_cap
);
1120 ret
= cb(inode
, cap
, arg
);
1123 spin_lock(&session
->s_cap_lock
);
1125 if (cap
->ci
== NULL
) {
1126 dout("iterate_session_caps finishing cap %p removal\n",
1128 BUG_ON(cap
->session
!= session
);
1129 cap
->session
= NULL
;
1130 list_del_init(&cap
->session_caps
);
1131 session
->s_nr_caps
--;
1132 if (cap
->queue_release
) {
1133 list_add_tail(&cap
->session_caps
,
1134 &session
->s_cap_releases
);
1135 session
->s_num_cap_releases
++;
1137 old_cap
= cap
; /* put_cap it w/o locks held */
1145 session
->s_cap_iterator
= NULL
;
1146 spin_unlock(&session
->s_cap_lock
);
1150 ceph_put_cap(session
->s_mdsc
, old_cap
);
1155 static int remove_session_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1158 struct ceph_fs_client
*fsc
= (struct ceph_fs_client
*)arg
;
1159 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1160 LIST_HEAD(to_remove
);
1162 bool invalidate
= false;
1164 dout("removing cap %p, ci is %p, inode is %p\n",
1165 cap
, ci
, &ci
->vfs_inode
);
1166 spin_lock(&ci
->i_ceph_lock
);
1167 __ceph_remove_cap(cap
, false);
1168 if (!ci
->i_auth_cap
) {
1169 struct ceph_cap_flush
*cf
;
1170 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
1172 ci
->i_ceph_flags
|= CEPH_I_CAP_DROPPED
;
1174 if (ci
->i_wrbuffer_ref
> 0 &&
1175 READ_ONCE(fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
1178 while (!list_empty(&ci
->i_cap_flush_list
)) {
1179 cf
= list_first_entry(&ci
->i_cap_flush_list
,
1180 struct ceph_cap_flush
, i_list
);
1181 list_move(&cf
->i_list
, &to_remove
);
1184 spin_lock(&mdsc
->cap_dirty_lock
);
1186 list_for_each_entry(cf
, &to_remove
, i_list
)
1187 list_del(&cf
->g_list
);
1189 if (!list_empty(&ci
->i_dirty_item
)) {
1190 pr_warn_ratelimited(
1191 " dropping dirty %s state for %p %lld\n",
1192 ceph_cap_string(ci
->i_dirty_caps
),
1193 inode
, ceph_ino(inode
));
1194 ci
->i_dirty_caps
= 0;
1195 list_del_init(&ci
->i_dirty_item
);
1198 if (!list_empty(&ci
->i_flushing_item
)) {
1199 pr_warn_ratelimited(
1200 " dropping dirty+flushing %s state for %p %lld\n",
1201 ceph_cap_string(ci
->i_flushing_caps
),
1202 inode
, ceph_ino(inode
));
1203 ci
->i_flushing_caps
= 0;
1204 list_del_init(&ci
->i_flushing_item
);
1205 mdsc
->num_cap_flushing
--;
1208 spin_unlock(&mdsc
->cap_dirty_lock
);
1210 if (!ci
->i_dirty_caps
&& ci
->i_prealloc_cap_flush
) {
1211 list_add(&ci
->i_prealloc_cap_flush
->i_list
, &to_remove
);
1212 ci
->i_prealloc_cap_flush
= NULL
;
1215 spin_unlock(&ci
->i_ceph_lock
);
1216 while (!list_empty(&to_remove
)) {
1217 struct ceph_cap_flush
*cf
;
1218 cf
= list_first_entry(&to_remove
,
1219 struct ceph_cap_flush
, i_list
);
1220 list_del(&cf
->i_list
);
1221 ceph_free_cap_flush(cf
);
1224 wake_up_all(&ci
->i_cap_wq
);
1226 ceph_queue_invalidate(inode
);
1233 * caller must hold session s_mutex
1235 static void remove_session_caps(struct ceph_mds_session
*session
)
1237 struct ceph_fs_client
*fsc
= session
->s_mdsc
->fsc
;
1238 struct super_block
*sb
= fsc
->sb
;
1239 dout("remove_session_caps on %p\n", session
);
1240 iterate_session_caps(session
, remove_session_caps_cb
, fsc
);
1242 wake_up_all(&fsc
->mdsc
->cap_flushing_wq
);
1244 spin_lock(&session
->s_cap_lock
);
1245 if (session
->s_nr_caps
> 0) {
1246 struct inode
*inode
;
1247 struct ceph_cap
*cap
, *prev
= NULL
;
1248 struct ceph_vino vino
;
1250 * iterate_session_caps() skips inodes that are being
1251 * deleted, we need to wait until deletions are complete.
1252 * __wait_on_freeing_inode() is designed for the job,
1253 * but it is not exported, so use lookup inode function
1256 while (!list_empty(&session
->s_caps
)) {
1257 cap
= list_entry(session
->s_caps
.next
,
1258 struct ceph_cap
, session_caps
);
1262 vino
= cap
->ci
->i_vino
;
1263 spin_unlock(&session
->s_cap_lock
);
1265 inode
= ceph_find_inode(sb
, vino
);
1268 spin_lock(&session
->s_cap_lock
);
1272 // drop cap expires and unlock s_cap_lock
1273 cleanup_cap_releases(session
->s_mdsc
, session
);
1275 BUG_ON(session
->s_nr_caps
> 0);
1276 BUG_ON(!list_empty(&session
->s_cap_flushing
));
1280 * wake up any threads waiting on this session's caps. if the cap is
1281 * old (didn't get renewed on the client reconnect), remove it now.
1283 * caller must hold s_mutex.
1285 static int wake_up_session_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1288 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1291 spin_lock(&ci
->i_ceph_lock
);
1292 ci
->i_wanted_max_size
= 0;
1293 ci
->i_requested_max_size
= 0;
1294 spin_unlock(&ci
->i_ceph_lock
);
1296 wake_up_all(&ci
->i_cap_wq
);
1300 static void wake_up_session_caps(struct ceph_mds_session
*session
,
1303 dout("wake_up_session_caps %p mds%d\n", session
, session
->s_mds
);
1304 iterate_session_caps(session
, wake_up_session_cb
,
1305 (void *)(unsigned long)reconnect
);
1309 * Send periodic message to MDS renewing all currently held caps. The
1310 * ack will reset the expiration for all caps from this session.
1312 * caller holds s_mutex
1314 static int send_renew_caps(struct ceph_mds_client
*mdsc
,
1315 struct ceph_mds_session
*session
)
1317 struct ceph_msg
*msg
;
1320 if (time_after_eq(jiffies
, session
->s_cap_ttl
) &&
1321 time_after_eq(session
->s_cap_ttl
, session
->s_renew_requested
))
1322 pr_info("mds%d caps stale\n", session
->s_mds
);
1323 session
->s_renew_requested
= jiffies
;
1325 /* do not try to renew caps until a recovering mds has reconnected
1326 * with its clients. */
1327 state
= ceph_mdsmap_get_state(mdsc
->mdsmap
, session
->s_mds
);
1328 if (state
< CEPH_MDS_STATE_RECONNECT
) {
1329 dout("send_renew_caps ignoring mds%d (%s)\n",
1330 session
->s_mds
, ceph_mds_state_name(state
));
1334 dout("send_renew_caps to mds%d (%s)\n", session
->s_mds
,
1335 ceph_mds_state_name(state
));
1336 msg
= create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS
,
1337 ++session
->s_renew_seq
);
1340 ceph_con_send(&session
->s_con
, msg
);
1344 static int send_flushmsg_ack(struct ceph_mds_client
*mdsc
,
1345 struct ceph_mds_session
*session
, u64 seq
)
1347 struct ceph_msg
*msg
;
1349 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1350 session
->s_mds
, ceph_session_state_name(session
->s_state
), seq
);
1351 msg
= create_session_msg(CEPH_SESSION_FLUSHMSG_ACK
, seq
);
1354 ceph_con_send(&session
->s_con
, msg
);
1360 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1362 * Called under session->s_mutex
1364 static void renewed_caps(struct ceph_mds_client
*mdsc
,
1365 struct ceph_mds_session
*session
, int is_renew
)
1370 spin_lock(&session
->s_cap_lock
);
1371 was_stale
= is_renew
&& time_after_eq(jiffies
, session
->s_cap_ttl
);
1373 session
->s_cap_ttl
= session
->s_renew_requested
+
1374 mdsc
->mdsmap
->m_session_timeout
*HZ
;
1377 if (time_before(jiffies
, session
->s_cap_ttl
)) {
1378 pr_info("mds%d caps renewed\n", session
->s_mds
);
1381 pr_info("mds%d caps still stale\n", session
->s_mds
);
1384 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1385 session
->s_mds
, session
->s_cap_ttl
, was_stale
? "stale" : "fresh",
1386 time_before(jiffies
, session
->s_cap_ttl
) ? "stale" : "fresh");
1387 spin_unlock(&session
->s_cap_lock
);
1390 wake_up_session_caps(session
, 0);
1394 * send a session close request
1396 static int request_close_session(struct ceph_mds_client
*mdsc
,
1397 struct ceph_mds_session
*session
)
1399 struct ceph_msg
*msg
;
1401 dout("request_close_session mds%d state %s seq %lld\n",
1402 session
->s_mds
, ceph_session_state_name(session
->s_state
),
1404 msg
= create_session_msg(CEPH_SESSION_REQUEST_CLOSE
, session
->s_seq
);
1407 ceph_con_send(&session
->s_con
, msg
);
1412 * Called with s_mutex held.
1414 static int __close_session(struct ceph_mds_client
*mdsc
,
1415 struct ceph_mds_session
*session
)
1417 if (session
->s_state
>= CEPH_MDS_SESSION_CLOSING
)
1419 session
->s_state
= CEPH_MDS_SESSION_CLOSING
;
1420 return request_close_session(mdsc
, session
);
1424 * Trim old(er) caps.
1426 * Because we can't cache an inode without one or more caps, we do
1427 * this indirectly: if a cap is unused, we prune its aliases, at which
1428 * point the inode will hopefully get dropped to.
1430 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1431 * memory pressure from the MDS, though, so it needn't be perfect.
1433 static int trim_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
, void *arg
)
1435 struct ceph_mds_session
*session
= arg
;
1436 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1437 int used
, wanted
, oissued
, mine
;
1439 if (session
->s_trim_caps
<= 0)
1442 spin_lock(&ci
->i_ceph_lock
);
1443 mine
= cap
->issued
| cap
->implemented
;
1444 used
= __ceph_caps_used(ci
);
1445 wanted
= __ceph_caps_file_wanted(ci
);
1446 oissued
= __ceph_caps_issued_other(ci
, cap
);
1448 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1449 inode
, cap
, ceph_cap_string(mine
), ceph_cap_string(oissued
),
1450 ceph_cap_string(used
), ceph_cap_string(wanted
));
1451 if (cap
== ci
->i_auth_cap
) {
1452 if (ci
->i_dirty_caps
|| ci
->i_flushing_caps
||
1453 !list_empty(&ci
->i_cap_snaps
))
1455 if ((used
| wanted
) & CEPH_CAP_ANY_WR
)
1458 /* The inode has cached pages, but it's no longer used.
1459 * we can safely drop it */
1460 if (wanted
== 0 && used
== CEPH_CAP_FILE_CACHE
&&
1461 !(oissued
& CEPH_CAP_FILE_CACHE
)) {
1465 if ((used
| wanted
) & ~oissued
& mine
)
1466 goto out
; /* we need these caps */
1468 session
->s_trim_caps
--;
1470 /* we aren't the only cap.. just remove us */
1471 __ceph_remove_cap(cap
, true);
1473 /* try dropping referring dentries */
1474 spin_unlock(&ci
->i_ceph_lock
);
1475 d_prune_aliases(inode
);
1476 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1477 inode
, cap
, atomic_read(&inode
->i_count
));
1482 spin_unlock(&ci
->i_ceph_lock
);
1487 * Trim session cap count down to some max number.
1489 static int trim_caps(struct ceph_mds_client
*mdsc
,
1490 struct ceph_mds_session
*session
,
1493 int trim_caps
= session
->s_nr_caps
- max_caps
;
1495 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1496 session
->s_mds
, session
->s_nr_caps
, max_caps
, trim_caps
);
1497 if (trim_caps
> 0) {
1498 session
->s_trim_caps
= trim_caps
;
1499 iterate_session_caps(session
, trim_caps_cb
, session
);
1500 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1501 session
->s_mds
, session
->s_nr_caps
, max_caps
,
1502 trim_caps
- session
->s_trim_caps
);
1503 session
->s_trim_caps
= 0;
1506 ceph_send_cap_releases(mdsc
, session
);
1510 static int check_caps_flush(struct ceph_mds_client
*mdsc
,
1515 spin_lock(&mdsc
->cap_dirty_lock
);
1516 if (!list_empty(&mdsc
->cap_flush_list
)) {
1517 struct ceph_cap_flush
*cf
=
1518 list_first_entry(&mdsc
->cap_flush_list
,
1519 struct ceph_cap_flush
, g_list
);
1520 if (cf
->tid
<= want_flush_tid
) {
1521 dout("check_caps_flush still flushing tid "
1522 "%llu <= %llu\n", cf
->tid
, want_flush_tid
);
1526 spin_unlock(&mdsc
->cap_dirty_lock
);
1531 * flush all dirty inode data to disk.
1533 * returns true if we've flushed through want_flush_tid
1535 static void wait_caps_flush(struct ceph_mds_client
*mdsc
,
1538 dout("check_caps_flush want %llu\n", want_flush_tid
);
1540 wait_event(mdsc
->cap_flushing_wq
,
1541 check_caps_flush(mdsc
, want_flush_tid
));
1543 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid
);
1547 * called under s_mutex
1549 void ceph_send_cap_releases(struct ceph_mds_client
*mdsc
,
1550 struct ceph_mds_session
*session
)
1552 struct ceph_msg
*msg
= NULL
;
1553 struct ceph_mds_cap_release
*head
;
1554 struct ceph_mds_cap_item
*item
;
1555 struct ceph_osd_client
*osdc
= &mdsc
->fsc
->client
->osdc
;
1556 struct ceph_cap
*cap
;
1557 LIST_HEAD(tmp_list
);
1558 int num_cap_releases
;
1559 __le32 barrier
, *cap_barrier
;
1561 down_read(&osdc
->lock
);
1562 barrier
= cpu_to_le32(osdc
->epoch_barrier
);
1563 up_read(&osdc
->lock
);
1565 spin_lock(&session
->s_cap_lock
);
1567 list_splice_init(&session
->s_cap_releases
, &tmp_list
);
1568 num_cap_releases
= session
->s_num_cap_releases
;
1569 session
->s_num_cap_releases
= 0;
1570 spin_unlock(&session
->s_cap_lock
);
1572 while (!list_empty(&tmp_list
)) {
1574 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE
,
1575 PAGE_SIZE
, GFP_NOFS
, false);
1578 head
= msg
->front
.iov_base
;
1579 head
->num
= cpu_to_le32(0);
1580 msg
->front
.iov_len
= sizeof(*head
);
1582 msg
->hdr
.version
= cpu_to_le16(2);
1583 msg
->hdr
.compat_version
= cpu_to_le16(1);
1586 cap
= list_first_entry(&tmp_list
, struct ceph_cap
,
1588 list_del(&cap
->session_caps
);
1591 head
= msg
->front
.iov_base
;
1592 le32_add_cpu(&head
->num
, 1);
1593 item
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1594 item
->ino
= cpu_to_le64(cap
->cap_ino
);
1595 item
->cap_id
= cpu_to_le64(cap
->cap_id
);
1596 item
->migrate_seq
= cpu_to_le32(cap
->mseq
);
1597 item
->seq
= cpu_to_le32(cap
->issue_seq
);
1598 msg
->front
.iov_len
+= sizeof(*item
);
1600 ceph_put_cap(mdsc
, cap
);
1602 if (le32_to_cpu(head
->num
) == CEPH_CAPS_PER_RELEASE
) {
1603 // Append cap_barrier field
1604 cap_barrier
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1605 *cap_barrier
= barrier
;
1606 msg
->front
.iov_len
+= sizeof(*cap_barrier
);
1608 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1609 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1610 ceph_con_send(&session
->s_con
, msg
);
1615 BUG_ON(num_cap_releases
!= 0);
1617 spin_lock(&session
->s_cap_lock
);
1618 if (!list_empty(&session
->s_cap_releases
))
1620 spin_unlock(&session
->s_cap_lock
);
1623 // Append cap_barrier field
1624 cap_barrier
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1625 *cap_barrier
= barrier
;
1626 msg
->front
.iov_len
+= sizeof(*cap_barrier
);
1628 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1629 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1630 ceph_con_send(&session
->s_con
, msg
);
1634 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1636 spin_lock(&session
->s_cap_lock
);
1637 list_splice(&tmp_list
, &session
->s_cap_releases
);
1638 session
->s_num_cap_releases
+= num_cap_releases
;
1639 spin_unlock(&session
->s_cap_lock
);
1646 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request
*req
,
1649 struct ceph_inode_info
*ci
= ceph_inode(dir
);
1650 struct ceph_mds_reply_info_parsed
*rinfo
= &req
->r_reply_info
;
1651 struct ceph_mount_options
*opt
= req
->r_mdsc
->fsc
->mount_options
;
1652 size_t size
= sizeof(struct ceph_mds_reply_dir_entry
);
1653 int order
, num_entries
;
1655 spin_lock(&ci
->i_ceph_lock
);
1656 num_entries
= ci
->i_files
+ ci
->i_subdirs
;
1657 spin_unlock(&ci
->i_ceph_lock
);
1658 num_entries
= max(num_entries
, 1);
1659 num_entries
= min(num_entries
, opt
->max_readdir
);
1661 order
= get_order(size
* num_entries
);
1662 while (order
>= 0) {
1663 rinfo
->dir_entries
= (void*)__get_free_pages(GFP_KERNEL
|
1666 if (rinfo
->dir_entries
)
1670 if (!rinfo
->dir_entries
)
1673 num_entries
= (PAGE_SIZE
<< order
) / size
;
1674 num_entries
= min(num_entries
, opt
->max_readdir
);
1676 rinfo
->dir_buf_size
= PAGE_SIZE
<< order
;
1677 req
->r_num_caps
= num_entries
+ 1;
1678 req
->r_args
.readdir
.max_entries
= cpu_to_le32(num_entries
);
1679 req
->r_args
.readdir
.max_bytes
= cpu_to_le32(opt
->max_readdir_bytes
);
1684 * Create an mds request.
1686 struct ceph_mds_request
*
1687 ceph_mdsc_create_request(struct ceph_mds_client
*mdsc
, int op
, int mode
)
1689 struct ceph_mds_request
*req
= kzalloc(sizeof(*req
), GFP_NOFS
);
1693 return ERR_PTR(-ENOMEM
);
1695 mutex_init(&req
->r_fill_mutex
);
1697 req
->r_started
= jiffies
;
1698 req
->r_resend_mds
= -1;
1699 INIT_LIST_HEAD(&req
->r_unsafe_dir_item
);
1700 INIT_LIST_HEAD(&req
->r_unsafe_target_item
);
1702 kref_init(&req
->r_kref
);
1703 RB_CLEAR_NODE(&req
->r_node
);
1704 INIT_LIST_HEAD(&req
->r_wait
);
1705 init_completion(&req
->r_completion
);
1706 init_completion(&req
->r_safe_completion
);
1707 INIT_LIST_HEAD(&req
->r_unsafe_item
);
1709 ktime_get_real_ts(&ts
);
1710 req
->r_stamp
= timespec_trunc(ts
, mdsc
->fsc
->sb
->s_time_gran
);
1713 req
->r_direct_mode
= mode
;
1718 * return oldest (lowest) request, tid in request tree, 0 if none.
1720 * called under mdsc->mutex.
1722 static struct ceph_mds_request
*__get_oldest_req(struct ceph_mds_client
*mdsc
)
1724 if (RB_EMPTY_ROOT(&mdsc
->request_tree
))
1726 return rb_entry(rb_first(&mdsc
->request_tree
),
1727 struct ceph_mds_request
, r_node
);
1730 static inline u64
__get_oldest_tid(struct ceph_mds_client
*mdsc
)
1732 return mdsc
->oldest_tid
;
1736 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1737 * on build_path_from_dentry in fs/cifs/dir.c.
1739 * If @stop_on_nosnap, generate path relative to the first non-snapped
1742 * Encode hidden .snap dirs as a double /, i.e.
1743 * foo/.snap/bar -> foo//bar
1745 char *ceph_mdsc_build_path(struct dentry
*dentry
, int *plen
, u64
*base
,
1748 struct dentry
*temp
;
1754 return ERR_PTR(-EINVAL
);
1758 seq
= read_seqbegin(&rename_lock
);
1760 for (temp
= dentry
; !IS_ROOT(temp
);) {
1761 struct inode
*inode
= d_inode(temp
);
1762 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
)
1763 len
++; /* slash only */
1764 else if (stop_on_nosnap
&& inode
&&
1765 ceph_snap(inode
) == CEPH_NOSNAP
)
1768 len
+= 1 + temp
->d_name
.len
;
1769 temp
= temp
->d_parent
;
1773 len
--; /* no leading '/' */
1775 path
= kmalloc(len
+1, GFP_NOFS
);
1777 return ERR_PTR(-ENOMEM
);
1779 path
[pos
] = 0; /* trailing null */
1781 for (temp
= dentry
; !IS_ROOT(temp
) && pos
!= 0; ) {
1782 struct inode
*inode
;
1784 spin_lock(&temp
->d_lock
);
1785 inode
= d_inode(temp
);
1786 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
) {
1787 dout("build_path path+%d: %p SNAPDIR\n",
1789 } else if (stop_on_nosnap
&& inode
&&
1790 ceph_snap(inode
) == CEPH_NOSNAP
) {
1791 spin_unlock(&temp
->d_lock
);
1794 pos
-= temp
->d_name
.len
;
1796 spin_unlock(&temp
->d_lock
);
1799 strncpy(path
+ pos
, temp
->d_name
.name
,
1802 spin_unlock(&temp
->d_lock
);
1805 temp
= temp
->d_parent
;
1808 if (pos
!= 0 || read_seqretry(&rename_lock
, seq
)) {
1809 pr_err("build_path did not end path lookup where "
1810 "expected, namelen is %d, pos is %d\n", len
, pos
);
1811 /* presumably this is only possible if racing with a
1812 rename of one of the parent directories (we can not
1813 lock the dentries above us to prevent this, but
1814 retrying should be harmless) */
1819 *base
= ceph_ino(d_inode(temp
));
1821 dout("build_path on %p %d built %llx '%.*s'\n",
1822 dentry
, d_count(dentry
), *base
, len
, path
);
1826 static int build_dentry_path(struct dentry
*dentry
, struct inode
*dir
,
1827 const char **ppath
, int *ppathlen
, u64
*pino
,
1834 dir
= d_inode_rcu(dentry
->d_parent
);
1835 if (dir
&& ceph_snap(dir
) == CEPH_NOSNAP
) {
1836 *pino
= ceph_ino(dir
);
1838 *ppath
= dentry
->d_name
.name
;
1839 *ppathlen
= dentry
->d_name
.len
;
1843 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1845 return PTR_ERR(path
);
1851 static int build_inode_path(struct inode
*inode
,
1852 const char **ppath
, int *ppathlen
, u64
*pino
,
1855 struct dentry
*dentry
;
1858 if (ceph_snap(inode
) == CEPH_NOSNAP
) {
1859 *pino
= ceph_ino(inode
);
1863 dentry
= d_find_alias(inode
);
1864 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1867 return PTR_ERR(path
);
1874 * request arguments may be specified via an inode *, a dentry *, or
1875 * an explicit ino+path.
1877 static int set_request_path_attr(struct inode
*rinode
, struct dentry
*rdentry
,
1878 struct inode
*rdiri
, const char *rpath
,
1879 u64 rino
, const char **ppath
, int *pathlen
,
1880 u64
*ino
, int *freepath
)
1885 r
= build_inode_path(rinode
, ppath
, pathlen
, ino
, freepath
);
1886 dout(" inode %p %llx.%llx\n", rinode
, ceph_ino(rinode
),
1888 } else if (rdentry
) {
1889 r
= build_dentry_path(rdentry
, rdiri
, ppath
, pathlen
, ino
,
1891 dout(" dentry %p %llx/%.*s\n", rdentry
, *ino
, *pathlen
,
1893 } else if (rpath
|| rino
) {
1896 *pathlen
= rpath
? strlen(rpath
) : 0;
1897 dout(" path %.*s\n", *pathlen
, rpath
);
1904 * called under mdsc->mutex
1906 static struct ceph_msg
*create_request_message(struct ceph_mds_client
*mdsc
,
1907 struct ceph_mds_request
*req
,
1908 int mds
, bool drop_cap_releases
)
1910 struct ceph_msg
*msg
;
1911 struct ceph_mds_request_head
*head
;
1912 const char *path1
= NULL
;
1913 const char *path2
= NULL
;
1914 u64 ino1
= 0, ino2
= 0;
1915 int pathlen1
= 0, pathlen2
= 0;
1916 int freepath1
= 0, freepath2
= 0;
1922 ret
= set_request_path_attr(req
->r_inode
, req
->r_dentry
,
1923 req
->r_parent
, req
->r_path1
, req
->r_ino1
.ino
,
1924 &path1
, &pathlen1
, &ino1
, &freepath1
);
1930 ret
= set_request_path_attr(NULL
, req
->r_old_dentry
,
1931 req
->r_old_dentry_dir
,
1932 req
->r_path2
, req
->r_ino2
.ino
,
1933 &path2
, &pathlen2
, &ino2
, &freepath2
);
1939 len
= sizeof(*head
) +
1940 pathlen1
+ pathlen2
+ 2*(1 + sizeof(u32
) + sizeof(u64
)) +
1941 sizeof(struct ceph_timespec
);
1943 /* calculate (max) length for cap releases */
1944 len
+= sizeof(struct ceph_mds_request_release
) *
1945 (!!req
->r_inode_drop
+ !!req
->r_dentry_drop
+
1946 !!req
->r_old_inode_drop
+ !!req
->r_old_dentry_drop
);
1947 if (req
->r_dentry_drop
)
1948 len
+= req
->r_dentry
->d_name
.len
;
1949 if (req
->r_old_dentry_drop
)
1950 len
+= req
->r_old_dentry
->d_name
.len
;
1952 msg
= ceph_msg_new(CEPH_MSG_CLIENT_REQUEST
, len
, GFP_NOFS
, false);
1954 msg
= ERR_PTR(-ENOMEM
);
1958 msg
->hdr
.version
= cpu_to_le16(2);
1959 msg
->hdr
.tid
= cpu_to_le64(req
->r_tid
);
1961 head
= msg
->front
.iov_base
;
1962 p
= msg
->front
.iov_base
+ sizeof(*head
);
1963 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1965 head
->mdsmap_epoch
= cpu_to_le32(mdsc
->mdsmap
->m_epoch
);
1966 head
->op
= cpu_to_le32(req
->r_op
);
1967 head
->caller_uid
= cpu_to_le32(from_kuid(&init_user_ns
, req
->r_uid
));
1968 head
->caller_gid
= cpu_to_le32(from_kgid(&init_user_ns
, req
->r_gid
));
1969 head
->args
= req
->r_args
;
1971 ceph_encode_filepath(&p
, end
, ino1
, path1
);
1972 ceph_encode_filepath(&p
, end
, ino2
, path2
);
1974 /* make note of release offset, in case we need to replay */
1975 req
->r_request_release_offset
= p
- msg
->front
.iov_base
;
1979 if (req
->r_inode_drop
)
1980 releases
+= ceph_encode_inode_release(&p
,
1981 req
->r_inode
? req
->r_inode
: d_inode(req
->r_dentry
),
1982 mds
, req
->r_inode_drop
, req
->r_inode_unless
, 0);
1983 if (req
->r_dentry_drop
)
1984 releases
+= ceph_encode_dentry_release(&p
, req
->r_dentry
,
1985 req
->r_parent
, mds
, req
->r_dentry_drop
,
1986 req
->r_dentry_unless
);
1987 if (req
->r_old_dentry_drop
)
1988 releases
+= ceph_encode_dentry_release(&p
, req
->r_old_dentry
,
1989 req
->r_old_dentry_dir
, mds
,
1990 req
->r_old_dentry_drop
,
1991 req
->r_old_dentry_unless
);
1992 if (req
->r_old_inode_drop
)
1993 releases
+= ceph_encode_inode_release(&p
,
1994 d_inode(req
->r_old_dentry
),
1995 mds
, req
->r_old_inode_drop
, req
->r_old_inode_unless
, 0);
1997 if (drop_cap_releases
) {
1999 p
= msg
->front
.iov_base
+ req
->r_request_release_offset
;
2002 head
->num_releases
= cpu_to_le16(releases
);
2006 struct ceph_timespec ts
;
2007 ceph_encode_timespec(&ts
, &req
->r_stamp
);
2008 ceph_encode_copy(&p
, &ts
, sizeof(ts
));
2012 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
2013 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
2015 if (req
->r_pagelist
) {
2016 struct ceph_pagelist
*pagelist
= req
->r_pagelist
;
2017 refcount_inc(&pagelist
->refcnt
);
2018 ceph_msg_data_add_pagelist(msg
, pagelist
);
2019 msg
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
2021 msg
->hdr
.data_len
= 0;
2024 msg
->hdr
.data_off
= cpu_to_le16(0);
2028 kfree((char *)path2
);
2031 kfree((char *)path1
);
2037 * called under mdsc->mutex if error, under no mutex if
2040 static void complete_request(struct ceph_mds_client
*mdsc
,
2041 struct ceph_mds_request
*req
)
2043 if (req
->r_callback
)
2044 req
->r_callback(mdsc
, req
);
2046 complete_all(&req
->r_completion
);
2050 * called under mdsc->mutex
2052 static int __prepare_send_request(struct ceph_mds_client
*mdsc
,
2053 struct ceph_mds_request
*req
,
2054 int mds
, bool drop_cap_releases
)
2056 struct ceph_mds_request_head
*rhead
;
2057 struct ceph_msg
*msg
;
2062 struct ceph_cap
*cap
=
2063 ceph_get_cap_for_mds(ceph_inode(req
->r_inode
), mds
);
2066 req
->r_sent_on_mseq
= cap
->mseq
;
2068 req
->r_sent_on_mseq
= -1;
2070 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req
,
2071 req
->r_tid
, ceph_mds_op_name(req
->r_op
), req
->r_attempts
);
2073 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
)) {
2076 * Replay. Do not regenerate message (and rebuild
2077 * paths, etc.); just use the original message.
2078 * Rebuilding paths will break for renames because
2079 * d_move mangles the src name.
2081 msg
= req
->r_request
;
2082 rhead
= msg
->front
.iov_base
;
2084 flags
= le32_to_cpu(rhead
->flags
);
2085 flags
|= CEPH_MDS_FLAG_REPLAY
;
2086 rhead
->flags
= cpu_to_le32(flags
);
2088 if (req
->r_target_inode
)
2089 rhead
->ino
= cpu_to_le64(ceph_ino(req
->r_target_inode
));
2091 rhead
->num_retry
= req
->r_attempts
- 1;
2093 /* remove cap/dentry releases from message */
2094 rhead
->num_releases
= 0;
2097 p
= msg
->front
.iov_base
+ req
->r_request_release_offset
;
2099 struct ceph_timespec ts
;
2100 ceph_encode_timespec(&ts
, &req
->r_stamp
);
2101 ceph_encode_copy(&p
, &ts
, sizeof(ts
));
2104 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
2105 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
2109 if (req
->r_request
) {
2110 ceph_msg_put(req
->r_request
);
2111 req
->r_request
= NULL
;
2113 msg
= create_request_message(mdsc
, req
, mds
, drop_cap_releases
);
2115 req
->r_err
= PTR_ERR(msg
);
2116 return PTR_ERR(msg
);
2118 req
->r_request
= msg
;
2120 rhead
= msg
->front
.iov_base
;
2121 rhead
->oldest_client_tid
= cpu_to_le64(__get_oldest_tid(mdsc
));
2122 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
))
2123 flags
|= CEPH_MDS_FLAG_REPLAY
;
2125 flags
|= CEPH_MDS_FLAG_WANT_DENTRY
;
2126 rhead
->flags
= cpu_to_le32(flags
);
2127 rhead
->num_fwd
= req
->r_num_fwd
;
2128 rhead
->num_retry
= req
->r_attempts
- 1;
2131 dout(" r_parent = %p\n", req
->r_parent
);
2136 * send request, or put it on the appropriate wait list.
2138 static int __do_request(struct ceph_mds_client
*mdsc
,
2139 struct ceph_mds_request
*req
)
2141 struct ceph_mds_session
*session
= NULL
;
2145 if (req
->r_err
|| test_bit(CEPH_MDS_R_GOT_RESULT
, &req
->r_req_flags
)) {
2146 if (test_bit(CEPH_MDS_R_ABORTED
, &req
->r_req_flags
))
2147 __unregister_request(mdsc
, req
);
2151 if (req
->r_timeout
&&
2152 time_after_eq(jiffies
, req
->r_started
+ req
->r_timeout
)) {
2153 dout("do_request timed out\n");
2157 if (READ_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
) {
2158 dout("do_request forced umount\n");
2162 if (READ_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_MOUNTING
) {
2163 if (mdsc
->mdsmap_err
) {
2164 err
= mdsc
->mdsmap_err
;
2165 dout("do_request mdsmap err %d\n", err
);
2168 if (mdsc
->mdsmap
->m_epoch
== 0) {
2169 dout("do_request no mdsmap, waiting for map\n");
2170 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
2173 if (!(mdsc
->fsc
->mount_options
->flags
&
2174 CEPH_MOUNT_OPT_MOUNTWAIT
) &&
2175 !ceph_mdsmap_is_cluster_available(mdsc
->mdsmap
)) {
2177 pr_info("probably no mds server is up\n");
2182 put_request_session(req
);
2184 mds
= __choose_mds(mdsc
, req
);
2186 ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) < CEPH_MDS_STATE_ACTIVE
) {
2187 dout("do_request no mds or not active, waiting for map\n");
2188 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
2192 /* get, open session */
2193 session
= __ceph_lookup_mds_session(mdsc
, mds
);
2195 session
= register_session(mdsc
, mds
);
2196 if (IS_ERR(session
)) {
2197 err
= PTR_ERR(session
);
2201 req
->r_session
= get_session(session
);
2203 dout("do_request mds%d session %p state %s\n", mds
, session
,
2204 ceph_session_state_name(session
->s_state
));
2205 if (session
->s_state
!= CEPH_MDS_SESSION_OPEN
&&
2206 session
->s_state
!= CEPH_MDS_SESSION_HUNG
) {
2207 if (session
->s_state
== CEPH_MDS_SESSION_REJECTED
) {
2211 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
2212 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
2213 __open_session(mdsc
, session
);
2214 list_add(&req
->r_wait
, &session
->s_waiting
);
2219 req
->r_resend_mds
= -1; /* forget any previous mds hint */
2221 if (req
->r_request_started
== 0) /* note request start time */
2222 req
->r_request_started
= jiffies
;
2224 err
= __prepare_send_request(mdsc
, req
, mds
, false);
2226 ceph_msg_get(req
->r_request
);
2227 ceph_con_send(&session
->s_con
, req
->r_request
);
2231 ceph_put_mds_session(session
);
2234 dout("__do_request early error %d\n", err
);
2236 complete_request(mdsc
, req
);
2237 __unregister_request(mdsc
, req
);
2244 * called under mdsc->mutex
2246 static void __wake_requests(struct ceph_mds_client
*mdsc
,
2247 struct list_head
*head
)
2249 struct ceph_mds_request
*req
;
2250 LIST_HEAD(tmp_list
);
2252 list_splice_init(head
, &tmp_list
);
2254 while (!list_empty(&tmp_list
)) {
2255 req
= list_entry(tmp_list
.next
,
2256 struct ceph_mds_request
, r_wait
);
2257 list_del_init(&req
->r_wait
);
2258 dout(" wake request %p tid %llu\n", req
, req
->r_tid
);
2259 __do_request(mdsc
, req
);
2264 * Wake up threads with requests pending for @mds, so that they can
2265 * resubmit their requests to a possibly different mds.
2267 static void kick_requests(struct ceph_mds_client
*mdsc
, int mds
)
2269 struct ceph_mds_request
*req
;
2270 struct rb_node
*p
= rb_first(&mdsc
->request_tree
);
2272 dout("kick_requests mds%d\n", mds
);
2274 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2276 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
))
2278 if (req
->r_attempts
> 0)
2279 continue; /* only new requests */
2280 if (req
->r_session
&&
2281 req
->r_session
->s_mds
== mds
) {
2282 dout(" kicking tid %llu\n", req
->r_tid
);
2283 list_del_init(&req
->r_wait
);
2284 __do_request(mdsc
, req
);
2289 void ceph_mdsc_submit_request(struct ceph_mds_client
*mdsc
,
2290 struct ceph_mds_request
*req
)
2292 dout("submit_request on %p\n", req
);
2293 mutex_lock(&mdsc
->mutex
);
2294 __register_request(mdsc
, req
, NULL
);
2295 __do_request(mdsc
, req
);
2296 mutex_unlock(&mdsc
->mutex
);
2300 * Synchrously perform an mds request. Take care of all of the
2301 * session setup, forwarding, retry details.
2303 int ceph_mdsc_do_request(struct ceph_mds_client
*mdsc
,
2305 struct ceph_mds_request
*req
)
2309 dout("do_request on %p\n", req
);
2311 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2313 ceph_get_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
2315 ceph_get_cap_refs(ceph_inode(req
->r_parent
), CEPH_CAP_PIN
);
2316 if (req
->r_old_dentry_dir
)
2317 ceph_get_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
2321 mutex_lock(&mdsc
->mutex
);
2322 __register_request(mdsc
, req
, dir
);
2323 __do_request(mdsc
, req
);
2331 mutex_unlock(&mdsc
->mutex
);
2332 dout("do_request waiting\n");
2333 if (!req
->r_timeout
&& req
->r_wait_for_completion
) {
2334 err
= req
->r_wait_for_completion(mdsc
, req
);
2336 long timeleft
= wait_for_completion_killable_timeout(
2338 ceph_timeout_jiffies(req
->r_timeout
));
2342 err
= -EIO
; /* timed out */
2344 err
= timeleft
; /* killed */
2346 dout("do_request waited, got %d\n", err
);
2347 mutex_lock(&mdsc
->mutex
);
2349 /* only abort if we didn't race with a real reply */
2350 if (test_bit(CEPH_MDS_R_GOT_RESULT
, &req
->r_req_flags
)) {
2351 err
= le32_to_cpu(req
->r_reply_info
.head
->result
);
2352 } else if (err
< 0) {
2353 dout("aborted request %lld with %d\n", req
->r_tid
, err
);
2356 * ensure we aren't running concurrently with
2357 * ceph_fill_trace or ceph_readdir_prepopulate, which
2358 * rely on locks (dir mutex) held by our caller.
2360 mutex_lock(&req
->r_fill_mutex
);
2362 set_bit(CEPH_MDS_R_ABORTED
, &req
->r_req_flags
);
2363 mutex_unlock(&req
->r_fill_mutex
);
2365 if (req
->r_parent
&&
2366 (req
->r_op
& CEPH_MDS_OP_WRITE
))
2367 ceph_invalidate_dir_request(req
);
2373 mutex_unlock(&mdsc
->mutex
);
2374 dout("do_request %p done, result %d\n", req
, err
);
2379 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2380 * namespace request.
2382 void ceph_invalidate_dir_request(struct ceph_mds_request
*req
)
2384 struct inode
*inode
= req
->r_parent
;
2386 dout("invalidate_dir_request %p (complete, lease(s))\n", inode
);
2388 ceph_dir_clear_complete(inode
);
2390 ceph_invalidate_dentry_lease(req
->r_dentry
);
2391 if (req
->r_old_dentry
)
2392 ceph_invalidate_dentry_lease(req
->r_old_dentry
);
2398 * We take the session mutex and parse and process the reply immediately.
2399 * This preserves the logical ordering of replies, capabilities, etc., sent
2400 * by the MDS as they are applied to our local cache.
2402 static void handle_reply(struct ceph_mds_session
*session
, struct ceph_msg
*msg
)
2404 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2405 struct ceph_mds_request
*req
;
2406 struct ceph_mds_reply_head
*head
= msg
->front
.iov_base
;
2407 struct ceph_mds_reply_info_parsed
*rinfo
; /* parsed reply info */
2408 struct ceph_snap_realm
*realm
;
2411 int mds
= session
->s_mds
;
2413 if (msg
->front
.iov_len
< sizeof(*head
)) {
2414 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2419 /* get request, session */
2420 tid
= le64_to_cpu(msg
->hdr
.tid
);
2421 mutex_lock(&mdsc
->mutex
);
2422 req
= lookup_get_request(mdsc
, tid
);
2424 dout("handle_reply on unknown tid %llu\n", tid
);
2425 mutex_unlock(&mdsc
->mutex
);
2428 dout("handle_reply %p\n", req
);
2430 /* correct session? */
2431 if (req
->r_session
!= session
) {
2432 pr_err("mdsc_handle_reply got %llu on session mds%d"
2433 " not mds%d\n", tid
, session
->s_mds
,
2434 req
->r_session
? req
->r_session
->s_mds
: -1);
2435 mutex_unlock(&mdsc
->mutex
);
2440 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
) && !head
->safe
) ||
2441 (test_bit(CEPH_MDS_R_GOT_SAFE
, &req
->r_req_flags
) && head
->safe
)) {
2442 pr_warn("got a dup %s reply on %llu from mds%d\n",
2443 head
->safe
? "safe" : "unsafe", tid
, mds
);
2444 mutex_unlock(&mdsc
->mutex
);
2447 if (test_bit(CEPH_MDS_R_GOT_SAFE
, &req
->r_req_flags
)) {
2448 pr_warn("got unsafe after safe on %llu from mds%d\n",
2450 mutex_unlock(&mdsc
->mutex
);
2454 result
= le32_to_cpu(head
->result
);
2458 * if we're not talking to the authority, send to them
2459 * if the authority has changed while we weren't looking,
2460 * send to new authority
2461 * Otherwise we just have to return an ESTALE
2463 if (result
== -ESTALE
) {
2464 dout("got ESTALE on request %llu", req
->r_tid
);
2465 req
->r_resend_mds
= -1;
2466 if (req
->r_direct_mode
!= USE_AUTH_MDS
) {
2467 dout("not using auth, setting for that now");
2468 req
->r_direct_mode
= USE_AUTH_MDS
;
2469 __do_request(mdsc
, req
);
2470 mutex_unlock(&mdsc
->mutex
);
2473 int mds
= __choose_mds(mdsc
, req
);
2474 if (mds
>= 0 && mds
!= req
->r_session
->s_mds
) {
2475 dout("but auth changed, so resending");
2476 __do_request(mdsc
, req
);
2477 mutex_unlock(&mdsc
->mutex
);
2481 dout("have to return ESTALE on request %llu", req
->r_tid
);
2486 set_bit(CEPH_MDS_R_GOT_SAFE
, &req
->r_req_flags
);
2487 __unregister_request(mdsc
, req
);
2489 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
)) {
2491 * We already handled the unsafe response, now do the
2492 * cleanup. No need to examine the response; the MDS
2493 * doesn't include any result info in the safe
2494 * response. And even if it did, there is nothing
2495 * useful we could do with a revised return value.
2497 dout("got safe reply %llu, mds%d\n", tid
, mds
);
2499 /* last unsafe request during umount? */
2500 if (mdsc
->stopping
&& !__get_oldest_req(mdsc
))
2501 complete_all(&mdsc
->safe_umount_waiters
);
2502 mutex_unlock(&mdsc
->mutex
);
2506 set_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
);
2507 list_add_tail(&req
->r_unsafe_item
, &req
->r_session
->s_unsafe
);
2508 if (req
->r_unsafe_dir
) {
2509 struct ceph_inode_info
*ci
=
2510 ceph_inode(req
->r_unsafe_dir
);
2511 spin_lock(&ci
->i_unsafe_lock
);
2512 list_add_tail(&req
->r_unsafe_dir_item
,
2513 &ci
->i_unsafe_dirops
);
2514 spin_unlock(&ci
->i_unsafe_lock
);
2518 dout("handle_reply tid %lld result %d\n", tid
, result
);
2519 rinfo
= &req
->r_reply_info
;
2520 err
= parse_reply_info(msg
, rinfo
, session
->s_con
.peer_features
);
2521 mutex_unlock(&mdsc
->mutex
);
2523 mutex_lock(&session
->s_mutex
);
2525 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds
, tid
);
2532 if (rinfo
->snapblob_len
) {
2533 down_write(&mdsc
->snap_rwsem
);
2534 ceph_update_snap_trace(mdsc
, rinfo
->snapblob
,
2535 rinfo
->snapblob
+ rinfo
->snapblob_len
,
2536 le32_to_cpu(head
->op
) == CEPH_MDS_OP_RMSNAP
,
2538 downgrade_write(&mdsc
->snap_rwsem
);
2540 down_read(&mdsc
->snap_rwsem
);
2543 /* insert trace into our cache */
2544 mutex_lock(&req
->r_fill_mutex
);
2545 current
->journal_info
= req
;
2546 err
= ceph_fill_trace(mdsc
->fsc
->sb
, req
);
2548 if (result
== 0 && (req
->r_op
== CEPH_MDS_OP_READDIR
||
2549 req
->r_op
== CEPH_MDS_OP_LSSNAP
))
2550 ceph_readdir_prepopulate(req
, req
->r_session
);
2551 ceph_unreserve_caps(mdsc
, &req
->r_caps_reservation
);
2553 current
->journal_info
= NULL
;
2554 mutex_unlock(&req
->r_fill_mutex
);
2556 up_read(&mdsc
->snap_rwsem
);
2558 ceph_put_snap_realm(mdsc
, realm
);
2560 if (err
== 0 && req
->r_target_inode
&&
2561 test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
)) {
2562 struct ceph_inode_info
*ci
= ceph_inode(req
->r_target_inode
);
2563 spin_lock(&ci
->i_unsafe_lock
);
2564 list_add_tail(&req
->r_unsafe_target_item
, &ci
->i_unsafe_iops
);
2565 spin_unlock(&ci
->i_unsafe_lock
);
2568 mutex_lock(&mdsc
->mutex
);
2569 if (!test_bit(CEPH_MDS_R_ABORTED
, &req
->r_req_flags
)) {
2573 req
->r_reply
= ceph_msg_get(msg
);
2574 set_bit(CEPH_MDS_R_GOT_RESULT
, &req
->r_req_flags
);
2577 dout("reply arrived after request %lld was aborted\n", tid
);
2579 mutex_unlock(&mdsc
->mutex
);
2581 mutex_unlock(&session
->s_mutex
);
2583 /* kick calling process */
2584 complete_request(mdsc
, req
);
2586 ceph_mdsc_put_request(req
);
2593 * handle mds notification that our request has been forwarded.
2595 static void handle_forward(struct ceph_mds_client
*mdsc
,
2596 struct ceph_mds_session
*session
,
2597 struct ceph_msg
*msg
)
2599 struct ceph_mds_request
*req
;
2600 u64 tid
= le64_to_cpu(msg
->hdr
.tid
);
2604 void *p
= msg
->front
.iov_base
;
2605 void *end
= p
+ msg
->front
.iov_len
;
2607 ceph_decode_need(&p
, end
, 2*sizeof(u32
), bad
);
2608 next_mds
= ceph_decode_32(&p
);
2609 fwd_seq
= ceph_decode_32(&p
);
2611 mutex_lock(&mdsc
->mutex
);
2612 req
= lookup_get_request(mdsc
, tid
);
2614 dout("forward tid %llu to mds%d - req dne\n", tid
, next_mds
);
2615 goto out
; /* dup reply? */
2618 if (test_bit(CEPH_MDS_R_ABORTED
, &req
->r_req_flags
)) {
2619 dout("forward tid %llu aborted, unregistering\n", tid
);
2620 __unregister_request(mdsc
, req
);
2621 } else if (fwd_seq
<= req
->r_num_fwd
) {
2622 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2623 tid
, next_mds
, req
->r_num_fwd
, fwd_seq
);
2625 /* resend. forward race not possible; mds would drop */
2626 dout("forward tid %llu to mds%d (we resend)\n", tid
, next_mds
);
2628 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT
, &req
->r_req_flags
));
2629 req
->r_attempts
= 0;
2630 req
->r_num_fwd
= fwd_seq
;
2631 req
->r_resend_mds
= next_mds
;
2632 put_request_session(req
);
2633 __do_request(mdsc
, req
);
2635 ceph_mdsc_put_request(req
);
2637 mutex_unlock(&mdsc
->mutex
);
2641 pr_err("mdsc_handle_forward decode error err=%d\n", err
);
2645 * handle a mds session control message
2647 static void handle_session(struct ceph_mds_session
*session
,
2648 struct ceph_msg
*msg
)
2650 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2653 int mds
= session
->s_mds
;
2654 struct ceph_mds_session_head
*h
= msg
->front
.iov_base
;
2658 if (msg
->front
.iov_len
!= sizeof(*h
))
2660 op
= le32_to_cpu(h
->op
);
2661 seq
= le64_to_cpu(h
->seq
);
2663 mutex_lock(&mdsc
->mutex
);
2664 if (op
== CEPH_SESSION_CLOSE
) {
2665 get_session(session
);
2666 __unregister_session(mdsc
, session
);
2668 /* FIXME: this ttl calculation is generous */
2669 session
->s_ttl
= jiffies
+ HZ
*mdsc
->mdsmap
->m_session_autoclose
;
2670 mutex_unlock(&mdsc
->mutex
);
2672 mutex_lock(&session
->s_mutex
);
2674 dout("handle_session mds%d %s %p state %s seq %llu\n",
2675 mds
, ceph_session_op_name(op
), session
,
2676 ceph_session_state_name(session
->s_state
), seq
);
2678 if (session
->s_state
== CEPH_MDS_SESSION_HUNG
) {
2679 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2680 pr_info("mds%d came back\n", session
->s_mds
);
2684 case CEPH_SESSION_OPEN
:
2685 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2686 pr_info("mds%d reconnect success\n", session
->s_mds
);
2687 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2688 renewed_caps(mdsc
, session
, 0);
2691 __close_session(mdsc
, session
);
2694 case CEPH_SESSION_RENEWCAPS
:
2695 if (session
->s_renew_seq
== seq
)
2696 renewed_caps(mdsc
, session
, 1);
2699 case CEPH_SESSION_CLOSE
:
2700 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2701 pr_info("mds%d reconnect denied\n", session
->s_mds
);
2702 cleanup_session_requests(mdsc
, session
);
2703 remove_session_caps(session
);
2704 wake
= 2; /* for good measure */
2705 wake_up_all(&mdsc
->session_close_wq
);
2708 case CEPH_SESSION_STALE
:
2709 pr_info("mds%d caps went stale, renewing\n",
2711 spin_lock(&session
->s_gen_ttl_lock
);
2712 session
->s_cap_gen
++;
2713 session
->s_cap_ttl
= jiffies
- 1;
2714 spin_unlock(&session
->s_gen_ttl_lock
);
2715 send_renew_caps(mdsc
, session
);
2718 case CEPH_SESSION_RECALL_STATE
:
2719 trim_caps(mdsc
, session
, le32_to_cpu(h
->max_caps
));
2722 case CEPH_SESSION_FLUSHMSG
:
2723 send_flushmsg_ack(mdsc
, session
, seq
);
2726 case CEPH_SESSION_FORCE_RO
:
2727 dout("force_session_readonly %p\n", session
);
2728 spin_lock(&session
->s_cap_lock
);
2729 session
->s_readonly
= true;
2730 spin_unlock(&session
->s_cap_lock
);
2731 wake_up_session_caps(session
, 0);
2734 case CEPH_SESSION_REJECT
:
2735 WARN_ON(session
->s_state
!= CEPH_MDS_SESSION_OPENING
);
2736 pr_info("mds%d rejected session\n", session
->s_mds
);
2737 session
->s_state
= CEPH_MDS_SESSION_REJECTED
;
2738 cleanup_session_requests(mdsc
, session
);
2739 remove_session_caps(session
);
2740 wake
= 2; /* for good measure */
2744 pr_err("mdsc_handle_session bad op %d mds%d\n", op
, mds
);
2748 mutex_unlock(&session
->s_mutex
);
2750 mutex_lock(&mdsc
->mutex
);
2751 __wake_requests(mdsc
, &session
->s_waiting
);
2753 kick_requests(mdsc
, mds
);
2754 mutex_unlock(&mdsc
->mutex
);
2756 if (op
== CEPH_SESSION_CLOSE
)
2757 ceph_put_mds_session(session
);
2761 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds
,
2762 (int)msg
->front
.iov_len
);
2769 * called under session->mutex.
2771 static void replay_unsafe_requests(struct ceph_mds_client
*mdsc
,
2772 struct ceph_mds_session
*session
)
2774 struct ceph_mds_request
*req
, *nreq
;
2778 dout("replay_unsafe_requests mds%d\n", session
->s_mds
);
2780 mutex_lock(&mdsc
->mutex
);
2781 list_for_each_entry_safe(req
, nreq
, &session
->s_unsafe
, r_unsafe_item
) {
2782 err
= __prepare_send_request(mdsc
, req
, session
->s_mds
, true);
2784 ceph_msg_get(req
->r_request
);
2785 ceph_con_send(&session
->s_con
, req
->r_request
);
2790 * also re-send old requests when MDS enters reconnect stage. So that MDS
2791 * can process completed request in clientreplay stage.
2793 p
= rb_first(&mdsc
->request_tree
);
2795 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2797 if (test_bit(CEPH_MDS_R_GOT_UNSAFE
, &req
->r_req_flags
))
2799 if (req
->r_attempts
== 0)
2800 continue; /* only old requests */
2801 if (req
->r_session
&&
2802 req
->r_session
->s_mds
== session
->s_mds
) {
2803 err
= __prepare_send_request(mdsc
, req
,
2804 session
->s_mds
, true);
2806 ceph_msg_get(req
->r_request
);
2807 ceph_con_send(&session
->s_con
, req
->r_request
);
2811 mutex_unlock(&mdsc
->mutex
);
2815 * Encode information about a cap for a reconnect with the MDS.
2817 static int encode_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
2821 struct ceph_mds_cap_reconnect v2
;
2822 struct ceph_mds_cap_reconnect_v1 v1
;
2824 struct ceph_inode_info
*ci
;
2825 struct ceph_reconnect_state
*recon_state
= arg
;
2826 struct ceph_pagelist
*pagelist
= recon_state
->pagelist
;
2831 struct dentry
*dentry
;
2835 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2836 inode
, ceph_vinop(inode
), cap
, cap
->cap_id
,
2837 ceph_cap_string(cap
->issued
));
2838 err
= ceph_pagelist_encode_64(pagelist
, ceph_ino(inode
));
2842 dentry
= d_find_alias(inode
);
2844 path
= ceph_mdsc_build_path(dentry
, &pathlen
, &pathbase
, 0);
2846 err
= PTR_ERR(path
);
2855 spin_lock(&ci
->i_ceph_lock
);
2856 cap
->seq
= 0; /* reset cap seq */
2857 cap
->issue_seq
= 0; /* and issue_seq */
2858 cap
->mseq
= 0; /* and migrate_seq */
2859 cap
->cap_gen
= cap
->session
->s_cap_gen
;
2861 if (recon_state
->msg_version
>= 2) {
2862 rec
.v2
.cap_id
= cpu_to_le64(cap
->cap_id
);
2863 rec
.v2
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2864 rec
.v2
.issued
= cpu_to_le32(cap
->issued
);
2865 rec
.v2
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2866 rec
.v2
.pathbase
= cpu_to_le64(pathbase
);
2867 rec
.v2
.flock_len
= 0;
2869 rec
.v1
.cap_id
= cpu_to_le64(cap
->cap_id
);
2870 rec
.v1
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2871 rec
.v1
.issued
= cpu_to_le32(cap
->issued
);
2872 rec
.v1
.size
= cpu_to_le64(inode
->i_size
);
2873 ceph_encode_timespec(&rec
.v1
.mtime
, &inode
->i_mtime
);
2874 ceph_encode_timespec(&rec
.v1
.atime
, &inode
->i_atime
);
2875 rec
.v1
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2876 rec
.v1
.pathbase
= cpu_to_le64(pathbase
);
2879 if (list_empty(&ci
->i_cap_snaps
)) {
2882 struct ceph_cap_snap
*capsnap
=
2883 list_first_entry(&ci
->i_cap_snaps
,
2884 struct ceph_cap_snap
, ci_item
);
2885 snap_follows
= capsnap
->follows
;
2887 spin_unlock(&ci
->i_ceph_lock
);
2889 if (recon_state
->msg_version
>= 2) {
2890 int num_fcntl_locks
, num_flock_locks
;
2891 struct ceph_filelock
*flocks
;
2892 size_t struct_len
, total_len
= 0;
2896 ceph_count_locks(inode
, &num_fcntl_locks
, &num_flock_locks
);
2897 flocks
= kmalloc((num_fcntl_locks
+num_flock_locks
) *
2898 sizeof(struct ceph_filelock
), GFP_NOFS
);
2903 err
= ceph_encode_locks_to_buffer(inode
, flocks
,
2913 if (recon_state
->msg_version
>= 3) {
2914 /* version, compat_version and struct_len */
2915 total_len
= 2 * sizeof(u8
) + sizeof(u32
);
2919 * number of encoded locks is stable, so copy to pagelist
2921 struct_len
= 2 * sizeof(u32
) +
2922 (num_fcntl_locks
+ num_flock_locks
) *
2923 sizeof(struct ceph_filelock
);
2924 rec
.v2
.flock_len
= cpu_to_le32(struct_len
);
2926 struct_len
+= sizeof(rec
.v2
);
2927 struct_len
+= sizeof(u32
) + pathlen
;
2930 struct_len
+= sizeof(u64
); /* snap_follows */
2932 total_len
+= struct_len
;
2933 err
= ceph_pagelist_reserve(pagelist
, total_len
);
2936 if (recon_state
->msg_version
>= 3) {
2937 ceph_pagelist_encode_8(pagelist
, struct_v
);
2938 ceph_pagelist_encode_8(pagelist
, 1);
2939 ceph_pagelist_encode_32(pagelist
, struct_len
);
2941 ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
2942 ceph_pagelist_append(pagelist
, &rec
, sizeof(rec
.v2
));
2943 ceph_locks_to_pagelist(flocks
, pagelist
,
2947 ceph_pagelist_encode_64(pagelist
, snap_follows
);
2951 size_t size
= sizeof(u32
) + pathlen
+ sizeof(rec
.v1
);
2952 err
= ceph_pagelist_reserve(pagelist
, size
);
2954 ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
2955 ceph_pagelist_append(pagelist
, &rec
, sizeof(rec
.v1
));
2959 recon_state
->nr_caps
++;
2969 * If an MDS fails and recovers, clients need to reconnect in order to
2970 * reestablish shared state. This includes all caps issued through
2971 * this session _and_ the snap_realm hierarchy. Because it's not
2972 * clear which snap realms the mds cares about, we send everything we
2973 * know about.. that ensures we'll then get any new info the
2974 * recovering MDS might have.
2976 * This is a relatively heavyweight operation, but it's rare.
2978 * called with mdsc->mutex held.
2980 static void send_mds_reconnect(struct ceph_mds_client
*mdsc
,
2981 struct ceph_mds_session
*session
)
2983 struct ceph_msg
*reply
;
2985 int mds
= session
->s_mds
;
2988 struct ceph_pagelist
*pagelist
;
2989 struct ceph_reconnect_state recon_state
;
2991 pr_info("mds%d reconnect start\n", mds
);
2993 pagelist
= kmalloc(sizeof(*pagelist
), GFP_NOFS
);
2995 goto fail_nopagelist
;
2996 ceph_pagelist_init(pagelist
);
2998 reply
= ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT
, 0, GFP_NOFS
, false);
3002 mutex_lock(&session
->s_mutex
);
3003 session
->s_state
= CEPH_MDS_SESSION_RECONNECTING
;
3006 dout("session %p state %s\n", session
,
3007 ceph_session_state_name(session
->s_state
));
3009 spin_lock(&session
->s_gen_ttl_lock
);
3010 session
->s_cap_gen
++;
3011 spin_unlock(&session
->s_gen_ttl_lock
);
3013 spin_lock(&session
->s_cap_lock
);
3014 /* don't know if session is readonly */
3015 session
->s_readonly
= 0;
3017 * notify __ceph_remove_cap() that we are composing cap reconnect.
3018 * If a cap get released before being added to the cap reconnect,
3019 * __ceph_remove_cap() should skip queuing cap release.
3021 session
->s_cap_reconnect
= 1;
3022 /* drop old cap expires; we're about to reestablish that state */
3023 cleanup_cap_releases(mdsc
, session
);
3025 /* trim unused caps to reduce MDS's cache rejoin time */
3026 if (mdsc
->fsc
->sb
->s_root
)
3027 shrink_dcache_parent(mdsc
->fsc
->sb
->s_root
);
3029 ceph_con_close(&session
->s_con
);
3030 ceph_con_open(&session
->s_con
,
3031 CEPH_ENTITY_TYPE_MDS
, mds
,
3032 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
3034 /* replay unsafe requests */
3035 replay_unsafe_requests(mdsc
, session
);
3037 down_read(&mdsc
->snap_rwsem
);
3039 /* traverse this session's caps */
3040 s_nr_caps
= session
->s_nr_caps
;
3041 err
= ceph_pagelist_encode_32(pagelist
, s_nr_caps
);
3045 recon_state
.nr_caps
= 0;
3046 recon_state
.pagelist
= pagelist
;
3047 if (session
->s_con
.peer_features
& CEPH_FEATURE_MDSENC
)
3048 recon_state
.msg_version
= 3;
3049 else if (session
->s_con
.peer_features
& CEPH_FEATURE_FLOCK
)
3050 recon_state
.msg_version
= 2;
3052 recon_state
.msg_version
= 1;
3053 err
= iterate_session_caps(session
, encode_caps_cb
, &recon_state
);
3057 spin_lock(&session
->s_cap_lock
);
3058 session
->s_cap_reconnect
= 0;
3059 spin_unlock(&session
->s_cap_lock
);
3062 * snaprealms. we provide mds with the ino, seq (version), and
3063 * parent for all of our realms. If the mds has any newer info,
3066 for (p
= rb_first(&mdsc
->snap_realms
); p
; p
= rb_next(p
)) {
3067 struct ceph_snap_realm
*realm
=
3068 rb_entry(p
, struct ceph_snap_realm
, node
);
3069 struct ceph_mds_snaprealm_reconnect sr_rec
;
3071 dout(" adding snap realm %llx seq %lld parent %llx\n",
3072 realm
->ino
, realm
->seq
, realm
->parent_ino
);
3073 sr_rec
.ino
= cpu_to_le64(realm
->ino
);
3074 sr_rec
.seq
= cpu_to_le64(realm
->seq
);
3075 sr_rec
.parent
= cpu_to_le64(realm
->parent_ino
);
3076 err
= ceph_pagelist_append(pagelist
, &sr_rec
, sizeof(sr_rec
));
3081 reply
->hdr
.version
= cpu_to_le16(recon_state
.msg_version
);
3083 /* raced with cap release? */
3084 if (s_nr_caps
!= recon_state
.nr_caps
) {
3085 struct page
*page
= list_first_entry(&pagelist
->head
,
3087 __le32
*addr
= kmap_atomic(page
);
3088 *addr
= cpu_to_le32(recon_state
.nr_caps
);
3089 kunmap_atomic(addr
);
3092 reply
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
3093 ceph_msg_data_add_pagelist(reply
, pagelist
);
3095 ceph_early_kick_flushing_caps(mdsc
, session
);
3097 ceph_con_send(&session
->s_con
, reply
);
3099 mutex_unlock(&session
->s_mutex
);
3101 mutex_lock(&mdsc
->mutex
);
3102 __wake_requests(mdsc
, &session
->s_waiting
);
3103 mutex_unlock(&mdsc
->mutex
);
3105 up_read(&mdsc
->snap_rwsem
);
3109 ceph_msg_put(reply
);
3110 up_read(&mdsc
->snap_rwsem
);
3111 mutex_unlock(&session
->s_mutex
);
3113 ceph_pagelist_release(pagelist
);
3115 pr_err("error %d preparing reconnect for mds%d\n", err
, mds
);
3121 * compare old and new mdsmaps, kicking requests
3122 * and closing out old connections as necessary
3124 * called under mdsc->mutex.
3126 static void check_new_map(struct ceph_mds_client
*mdsc
,
3127 struct ceph_mdsmap
*newmap
,
3128 struct ceph_mdsmap
*oldmap
)
3131 int oldstate
, newstate
;
3132 struct ceph_mds_session
*s
;
3134 dout("check_new_map new %u old %u\n",
3135 newmap
->m_epoch
, oldmap
->m_epoch
);
3137 for (i
= 0; i
< oldmap
->m_num_mds
&& i
< mdsc
->max_sessions
; i
++) {
3138 if (mdsc
->sessions
[i
] == NULL
)
3140 s
= mdsc
->sessions
[i
];
3141 oldstate
= ceph_mdsmap_get_state(oldmap
, i
);
3142 newstate
= ceph_mdsmap_get_state(newmap
, i
);
3144 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3145 i
, ceph_mds_state_name(oldstate
),
3146 ceph_mdsmap_is_laggy(oldmap
, i
) ? " (laggy)" : "",
3147 ceph_mds_state_name(newstate
),
3148 ceph_mdsmap_is_laggy(newmap
, i
) ? " (laggy)" : "",
3149 ceph_session_state_name(s
->s_state
));
3151 if (i
>= newmap
->m_num_mds
||
3152 memcmp(ceph_mdsmap_get_addr(oldmap
, i
),
3153 ceph_mdsmap_get_addr(newmap
, i
),
3154 sizeof(struct ceph_entity_addr
))) {
3155 if (s
->s_state
== CEPH_MDS_SESSION_OPENING
) {
3156 /* the session never opened, just close it
3159 __unregister_session(mdsc
, s
);
3160 __wake_requests(mdsc
, &s
->s_waiting
);
3161 ceph_put_mds_session(s
);
3162 } else if (i
>= newmap
->m_num_mds
) {
3163 /* force close session for stopped mds */
3165 __unregister_session(mdsc
, s
);
3166 __wake_requests(mdsc
, &s
->s_waiting
);
3167 kick_requests(mdsc
, i
);
3168 mutex_unlock(&mdsc
->mutex
);
3170 mutex_lock(&s
->s_mutex
);
3171 cleanup_session_requests(mdsc
, s
);
3172 remove_session_caps(s
);
3173 mutex_unlock(&s
->s_mutex
);
3175 ceph_put_mds_session(s
);
3177 mutex_lock(&mdsc
->mutex
);
3180 mutex_unlock(&mdsc
->mutex
);
3181 mutex_lock(&s
->s_mutex
);
3182 mutex_lock(&mdsc
->mutex
);
3183 ceph_con_close(&s
->s_con
);
3184 mutex_unlock(&s
->s_mutex
);
3185 s
->s_state
= CEPH_MDS_SESSION_RESTARTING
;
3187 } else if (oldstate
== newstate
) {
3188 continue; /* nothing new with this mds */
3194 if (s
->s_state
== CEPH_MDS_SESSION_RESTARTING
&&
3195 newstate
>= CEPH_MDS_STATE_RECONNECT
) {
3196 mutex_unlock(&mdsc
->mutex
);
3197 send_mds_reconnect(mdsc
, s
);
3198 mutex_lock(&mdsc
->mutex
);
3202 * kick request on any mds that has gone active.
3204 if (oldstate
< CEPH_MDS_STATE_ACTIVE
&&
3205 newstate
>= CEPH_MDS_STATE_ACTIVE
) {
3206 if (oldstate
!= CEPH_MDS_STATE_CREATING
&&
3207 oldstate
!= CEPH_MDS_STATE_STARTING
)
3208 pr_info("mds%d recovery completed\n", s
->s_mds
);
3209 kick_requests(mdsc
, i
);
3210 ceph_kick_flushing_caps(mdsc
, s
);
3211 wake_up_session_caps(s
, 1);
3215 for (i
= 0; i
< newmap
->m_num_mds
&& i
< mdsc
->max_sessions
; i
++) {
3216 s
= mdsc
->sessions
[i
];
3219 if (!ceph_mdsmap_is_laggy(newmap
, i
))
3221 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3222 s
->s_state
== CEPH_MDS_SESSION_HUNG
||
3223 s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3224 dout(" connecting to export targets of laggy mds%d\n",
3226 __open_export_target_sessions(mdsc
, s
);
3238 * caller must hold session s_mutex, dentry->d_lock
3240 void __ceph_mdsc_drop_dentry_lease(struct dentry
*dentry
)
3242 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
3244 ceph_put_mds_session(di
->lease_session
);
3245 di
->lease_session
= NULL
;
3248 static void handle_lease(struct ceph_mds_client
*mdsc
,
3249 struct ceph_mds_session
*session
,
3250 struct ceph_msg
*msg
)
3252 struct super_block
*sb
= mdsc
->fsc
->sb
;
3253 struct inode
*inode
;
3254 struct dentry
*parent
, *dentry
;
3255 struct ceph_dentry_info
*di
;
3256 int mds
= session
->s_mds
;
3257 struct ceph_mds_lease
*h
= msg
->front
.iov_base
;
3259 struct ceph_vino vino
;
3263 dout("handle_lease from mds%d\n", mds
);
3266 if (msg
->front
.iov_len
< sizeof(*h
) + sizeof(u32
))
3268 vino
.ino
= le64_to_cpu(h
->ino
);
3269 vino
.snap
= CEPH_NOSNAP
;
3270 seq
= le32_to_cpu(h
->seq
);
3271 dname
.name
= (void *)h
+ sizeof(*h
) + sizeof(u32
);
3272 dname
.len
= msg
->front
.iov_len
- sizeof(*h
) - sizeof(u32
);
3273 if (dname
.len
!= get_unaligned_le32(h
+1))
3277 inode
= ceph_find_inode(sb
, vino
);
3278 dout("handle_lease %s, ino %llx %p %.*s\n",
3279 ceph_lease_op_name(h
->action
), vino
.ino
, inode
,
3280 dname
.len
, dname
.name
);
3282 mutex_lock(&session
->s_mutex
);
3285 if (inode
== NULL
) {
3286 dout("handle_lease no inode %llx\n", vino
.ino
);
3291 parent
= d_find_alias(inode
);
3293 dout("no parent dentry on inode %p\n", inode
);
3295 goto release
; /* hrm... */
3297 dname
.hash
= full_name_hash(parent
, dname
.name
, dname
.len
);
3298 dentry
= d_lookup(parent
, &dname
);
3303 spin_lock(&dentry
->d_lock
);
3304 di
= ceph_dentry(dentry
);
3305 switch (h
->action
) {
3306 case CEPH_MDS_LEASE_REVOKE
:
3307 if (di
->lease_session
== session
) {
3308 if (ceph_seq_cmp(di
->lease_seq
, seq
) > 0)
3309 h
->seq
= cpu_to_le32(di
->lease_seq
);
3310 __ceph_mdsc_drop_dentry_lease(dentry
);
3315 case CEPH_MDS_LEASE_RENEW
:
3316 if (di
->lease_session
== session
&&
3317 di
->lease_gen
== session
->s_cap_gen
&&
3318 di
->lease_renew_from
&&
3319 di
->lease_renew_after
== 0) {
3320 unsigned long duration
=
3321 msecs_to_jiffies(le32_to_cpu(h
->duration_ms
));
3323 di
->lease_seq
= seq
;
3324 di
->time
= di
->lease_renew_from
+ duration
;
3325 di
->lease_renew_after
= di
->lease_renew_from
+
3327 di
->lease_renew_from
= 0;
3331 spin_unlock(&dentry
->d_lock
);
3338 /* let's just reuse the same message */
3339 h
->action
= CEPH_MDS_LEASE_REVOKE_ACK
;
3341 ceph_con_send(&session
->s_con
, msg
);
3345 mutex_unlock(&session
->s_mutex
);
3349 pr_err("corrupt lease message\n");
3353 void ceph_mdsc_lease_send_msg(struct ceph_mds_session
*session
,
3354 struct inode
*inode
,
3355 struct dentry
*dentry
, char action
,
3358 struct ceph_msg
*msg
;
3359 struct ceph_mds_lease
*lease
;
3360 int len
= sizeof(*lease
) + sizeof(u32
);
3363 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3364 inode
, dentry
, ceph_lease_op_name(action
), session
->s_mds
);
3365 dnamelen
= dentry
->d_name
.len
;
3368 msg
= ceph_msg_new(CEPH_MSG_CLIENT_LEASE
, len
, GFP_NOFS
, false);
3371 lease
= msg
->front
.iov_base
;
3372 lease
->action
= action
;
3373 lease
->ino
= cpu_to_le64(ceph_vino(inode
).ino
);
3374 lease
->first
= lease
->last
= cpu_to_le64(ceph_vino(inode
).snap
);
3375 lease
->seq
= cpu_to_le32(seq
);
3376 put_unaligned_le32(dnamelen
, lease
+ 1);
3377 memcpy((void *)(lease
+ 1) + 4, dentry
->d_name
.name
, dnamelen
);
3380 * if this is a preemptive lease RELEASE, no need to
3381 * flush request stream, since the actual request will
3384 msg
->more_to_follow
= (action
== CEPH_MDS_LEASE_RELEASE
);
3386 ceph_con_send(&session
->s_con
, msg
);
3390 * drop all leases (and dentry refs) in preparation for umount
3392 static void drop_leases(struct ceph_mds_client
*mdsc
)
3396 dout("drop_leases\n");
3397 mutex_lock(&mdsc
->mutex
);
3398 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3399 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3402 mutex_unlock(&mdsc
->mutex
);
3403 mutex_lock(&s
->s_mutex
);
3404 mutex_unlock(&s
->s_mutex
);
3405 ceph_put_mds_session(s
);
3406 mutex_lock(&mdsc
->mutex
);
3408 mutex_unlock(&mdsc
->mutex
);
3414 * delayed work -- periodically trim expired leases, renew caps with mds
3416 static void schedule_delayed(struct ceph_mds_client
*mdsc
)
3419 unsigned hz
= round_jiffies_relative(HZ
* delay
);
3420 schedule_delayed_work(&mdsc
->delayed_work
, hz
);
3423 static void delayed_work(struct work_struct
*work
)
3426 struct ceph_mds_client
*mdsc
=
3427 container_of(work
, struct ceph_mds_client
, delayed_work
.work
);
3431 dout("mdsc delayed_work\n");
3432 ceph_check_delayed_caps(mdsc
);
3434 mutex_lock(&mdsc
->mutex
);
3435 renew_interval
= mdsc
->mdsmap
->m_session_timeout
>> 2;
3436 renew_caps
= time_after_eq(jiffies
, HZ
*renew_interval
+
3437 mdsc
->last_renew_caps
);
3439 mdsc
->last_renew_caps
= jiffies
;
3441 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3442 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3445 if (s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3446 dout("resending session close request for mds%d\n",
3448 request_close_session(mdsc
, s
);
3449 ceph_put_mds_session(s
);
3452 if (s
->s_ttl
&& time_after(jiffies
, s
->s_ttl
)) {
3453 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
) {
3454 s
->s_state
= CEPH_MDS_SESSION_HUNG
;
3455 pr_info("mds%d hung\n", s
->s_mds
);
3458 if (s
->s_state
< CEPH_MDS_SESSION_OPEN
) {
3459 /* this mds is failed or recovering, just wait */
3460 ceph_put_mds_session(s
);
3463 mutex_unlock(&mdsc
->mutex
);
3465 mutex_lock(&s
->s_mutex
);
3467 send_renew_caps(mdsc
, s
);
3469 ceph_con_keepalive(&s
->s_con
);
3470 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3471 s
->s_state
== CEPH_MDS_SESSION_HUNG
)
3472 ceph_send_cap_releases(mdsc
, s
);
3473 mutex_unlock(&s
->s_mutex
);
3474 ceph_put_mds_session(s
);
3476 mutex_lock(&mdsc
->mutex
);
3478 mutex_unlock(&mdsc
->mutex
);
3480 schedule_delayed(mdsc
);
3483 int ceph_mdsc_init(struct ceph_fs_client
*fsc
)
3486 struct ceph_mds_client
*mdsc
;
3488 mdsc
= kzalloc(sizeof(struct ceph_mds_client
), GFP_NOFS
);
3493 mutex_init(&mdsc
->mutex
);
3494 mdsc
->mdsmap
= kzalloc(sizeof(*mdsc
->mdsmap
), GFP_NOFS
);
3495 if (mdsc
->mdsmap
== NULL
) {
3500 init_completion(&mdsc
->safe_umount_waiters
);
3501 init_waitqueue_head(&mdsc
->session_close_wq
);
3502 INIT_LIST_HEAD(&mdsc
->waiting_for_map
);
3503 mdsc
->sessions
= NULL
;
3504 atomic_set(&mdsc
->num_sessions
, 0);
3505 mdsc
->max_sessions
= 0;
3507 mdsc
->last_snap_seq
= 0;
3508 init_rwsem(&mdsc
->snap_rwsem
);
3509 mdsc
->snap_realms
= RB_ROOT
;
3510 INIT_LIST_HEAD(&mdsc
->snap_empty
);
3511 spin_lock_init(&mdsc
->snap_empty_lock
);
3513 mdsc
->oldest_tid
= 0;
3514 mdsc
->request_tree
= RB_ROOT
;
3515 INIT_DELAYED_WORK(&mdsc
->delayed_work
, delayed_work
);
3516 mdsc
->last_renew_caps
= jiffies
;
3517 INIT_LIST_HEAD(&mdsc
->cap_delay_list
);
3518 spin_lock_init(&mdsc
->cap_delay_lock
);
3519 INIT_LIST_HEAD(&mdsc
->snap_flush_list
);
3520 spin_lock_init(&mdsc
->snap_flush_lock
);
3521 mdsc
->last_cap_flush_tid
= 1;
3522 INIT_LIST_HEAD(&mdsc
->cap_flush_list
);
3523 INIT_LIST_HEAD(&mdsc
->cap_dirty
);
3524 INIT_LIST_HEAD(&mdsc
->cap_dirty_migrating
);
3525 mdsc
->num_cap_flushing
= 0;
3526 spin_lock_init(&mdsc
->cap_dirty_lock
);
3527 init_waitqueue_head(&mdsc
->cap_flushing_wq
);
3528 spin_lock_init(&mdsc
->dentry_lru_lock
);
3529 INIT_LIST_HEAD(&mdsc
->dentry_lru
);
3531 ceph_caps_init(mdsc
);
3532 ceph_adjust_min_caps(mdsc
, fsc
->min_caps
);
3534 init_rwsem(&mdsc
->pool_perm_rwsem
);
3535 mdsc
->pool_perm_tree
= RB_ROOT
;
3541 * Wait for safe replies on open mds requests. If we time out, drop
3542 * all requests from the tree to avoid dangling dentry refs.
3544 static void wait_requests(struct ceph_mds_client
*mdsc
)
3546 struct ceph_options
*opts
= mdsc
->fsc
->client
->options
;
3547 struct ceph_mds_request
*req
;
3549 mutex_lock(&mdsc
->mutex
);
3550 if (__get_oldest_req(mdsc
)) {
3551 mutex_unlock(&mdsc
->mutex
);
3553 dout("wait_requests waiting for requests\n");
3554 wait_for_completion_timeout(&mdsc
->safe_umount_waiters
,
3555 ceph_timeout_jiffies(opts
->mount_timeout
));
3557 /* tear down remaining requests */
3558 mutex_lock(&mdsc
->mutex
);
3559 while ((req
= __get_oldest_req(mdsc
))) {
3560 dout("wait_requests timed out on tid %llu\n",
3562 __unregister_request(mdsc
, req
);
3565 mutex_unlock(&mdsc
->mutex
);
3566 dout("wait_requests done\n");
3570 * called before mount is ro, and before dentries are torn down.
3571 * (hmm, does this still race with new lookups?)
3573 void ceph_mdsc_pre_umount(struct ceph_mds_client
*mdsc
)
3575 dout("pre_umount\n");
3579 ceph_flush_dirty_caps(mdsc
);
3580 wait_requests(mdsc
);
3583 * wait for reply handlers to drop their request refs and
3584 * their inode/dcache refs
3590 * wait for all write mds requests to flush.
3592 static void wait_unsafe_requests(struct ceph_mds_client
*mdsc
, u64 want_tid
)
3594 struct ceph_mds_request
*req
= NULL
, *nextreq
;
3597 mutex_lock(&mdsc
->mutex
);
3598 dout("wait_unsafe_requests want %lld\n", want_tid
);
3600 req
= __get_oldest_req(mdsc
);
3601 while (req
&& req
->r_tid
<= want_tid
) {
3602 /* find next request */
3603 n
= rb_next(&req
->r_node
);
3605 nextreq
= rb_entry(n
, struct ceph_mds_request
, r_node
);
3608 if (req
->r_op
!= CEPH_MDS_OP_SETFILELOCK
&&
3609 (req
->r_op
& CEPH_MDS_OP_WRITE
)) {
3611 ceph_mdsc_get_request(req
);
3613 ceph_mdsc_get_request(nextreq
);
3614 mutex_unlock(&mdsc
->mutex
);
3615 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3616 req
->r_tid
, want_tid
);
3617 wait_for_completion(&req
->r_safe_completion
);
3618 mutex_lock(&mdsc
->mutex
);
3619 ceph_mdsc_put_request(req
);
3621 break; /* next dne before, so we're done! */
3622 if (RB_EMPTY_NODE(&nextreq
->r_node
)) {
3623 /* next request was removed from tree */
3624 ceph_mdsc_put_request(nextreq
);
3627 ceph_mdsc_put_request(nextreq
); /* won't go away */
3631 mutex_unlock(&mdsc
->mutex
);
3632 dout("wait_unsafe_requests done\n");
3635 void ceph_mdsc_sync(struct ceph_mds_client
*mdsc
)
3637 u64 want_tid
, want_flush
;
3639 if (READ_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
3643 mutex_lock(&mdsc
->mutex
);
3644 want_tid
= mdsc
->last_tid
;
3645 mutex_unlock(&mdsc
->mutex
);
3647 ceph_flush_dirty_caps(mdsc
);
3648 spin_lock(&mdsc
->cap_dirty_lock
);
3649 want_flush
= mdsc
->last_cap_flush_tid
;
3650 if (!list_empty(&mdsc
->cap_flush_list
)) {
3651 struct ceph_cap_flush
*cf
=
3652 list_last_entry(&mdsc
->cap_flush_list
,
3653 struct ceph_cap_flush
, g_list
);
3656 spin_unlock(&mdsc
->cap_dirty_lock
);
3658 dout("sync want tid %lld flush_seq %lld\n",
3659 want_tid
, want_flush
);
3661 wait_unsafe_requests(mdsc
, want_tid
);
3662 wait_caps_flush(mdsc
, want_flush
);
3666 * true if all sessions are closed, or we force unmount
3668 static bool done_closing_sessions(struct ceph_mds_client
*mdsc
, int skipped
)
3670 if (READ_ONCE(mdsc
->fsc
->mount_state
) == CEPH_MOUNT_SHUTDOWN
)
3672 return atomic_read(&mdsc
->num_sessions
) <= skipped
;
3676 * called after sb is ro.
3678 void ceph_mdsc_close_sessions(struct ceph_mds_client
*mdsc
)
3680 struct ceph_options
*opts
= mdsc
->fsc
->client
->options
;
3681 struct ceph_mds_session
*session
;
3685 dout("close_sessions\n");
3687 /* close sessions */
3688 mutex_lock(&mdsc
->mutex
);
3689 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3690 session
= __ceph_lookup_mds_session(mdsc
, i
);
3693 mutex_unlock(&mdsc
->mutex
);
3694 mutex_lock(&session
->s_mutex
);
3695 if (__close_session(mdsc
, session
) <= 0)
3697 mutex_unlock(&session
->s_mutex
);
3698 ceph_put_mds_session(session
);
3699 mutex_lock(&mdsc
->mutex
);
3701 mutex_unlock(&mdsc
->mutex
);
3703 dout("waiting for sessions to close\n");
3704 wait_event_timeout(mdsc
->session_close_wq
,
3705 done_closing_sessions(mdsc
, skipped
),
3706 ceph_timeout_jiffies(opts
->mount_timeout
));
3708 /* tear down remaining sessions */
3709 mutex_lock(&mdsc
->mutex
);
3710 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3711 if (mdsc
->sessions
[i
]) {
3712 session
= get_session(mdsc
->sessions
[i
]);
3713 __unregister_session(mdsc
, session
);
3714 mutex_unlock(&mdsc
->mutex
);
3715 mutex_lock(&session
->s_mutex
);
3716 remove_session_caps(session
);
3717 mutex_unlock(&session
->s_mutex
);
3718 ceph_put_mds_session(session
);
3719 mutex_lock(&mdsc
->mutex
);
3722 WARN_ON(!list_empty(&mdsc
->cap_delay_list
));
3723 mutex_unlock(&mdsc
->mutex
);
3725 ceph_cleanup_empty_realms(mdsc
);
3727 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3732 void ceph_mdsc_force_umount(struct ceph_mds_client
*mdsc
)
3734 struct ceph_mds_session
*session
;
3737 dout("force umount\n");
3739 mutex_lock(&mdsc
->mutex
);
3740 for (mds
= 0; mds
< mdsc
->max_sessions
; mds
++) {
3741 session
= __ceph_lookup_mds_session(mdsc
, mds
);
3744 mutex_unlock(&mdsc
->mutex
);
3745 mutex_lock(&session
->s_mutex
);
3746 __close_session(mdsc
, session
);
3747 if (session
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3748 cleanup_session_requests(mdsc
, session
);
3749 remove_session_caps(session
);
3751 mutex_unlock(&session
->s_mutex
);
3752 ceph_put_mds_session(session
);
3753 mutex_lock(&mdsc
->mutex
);
3754 kick_requests(mdsc
, mds
);
3756 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3757 mutex_unlock(&mdsc
->mutex
);
3760 static void ceph_mdsc_stop(struct ceph_mds_client
*mdsc
)
3763 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3765 ceph_mdsmap_destroy(mdsc
->mdsmap
);
3766 kfree(mdsc
->sessions
);
3767 ceph_caps_finalize(mdsc
);
3768 ceph_pool_perm_destroy(mdsc
);
3771 void ceph_mdsc_destroy(struct ceph_fs_client
*fsc
)
3773 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
3775 dout("mdsc_destroy %p\n", mdsc
);
3776 ceph_mdsc_stop(mdsc
);
3778 /* flush out any connection work with references to us */
3783 dout("mdsc_destroy %p done\n", mdsc
);
3786 void ceph_mdsc_handle_fsmap(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3788 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3789 const char *mds_namespace
= fsc
->mount_options
->mds_namespace
;
3790 void *p
= msg
->front
.iov_base
;
3791 void *end
= p
+ msg
->front
.iov_len
;
3795 u32 mount_fscid
= (u32
)-1;
3796 u8 struct_v
, struct_cv
;
3799 ceph_decode_need(&p
, end
, sizeof(u32
), bad
);
3800 epoch
= ceph_decode_32(&p
);
3802 dout("handle_fsmap epoch %u\n", epoch
);
3804 ceph_decode_need(&p
, end
, 2 + sizeof(u32
), bad
);
3805 struct_v
= ceph_decode_8(&p
);
3806 struct_cv
= ceph_decode_8(&p
);
3807 map_len
= ceph_decode_32(&p
);
3809 ceph_decode_need(&p
, end
, sizeof(u32
) * 3, bad
);
3810 p
+= sizeof(u32
) * 2; /* skip epoch and legacy_client_fscid */
3812 num_fs
= ceph_decode_32(&p
);
3813 while (num_fs
-- > 0) {
3814 void *info_p
, *info_end
;
3819 ceph_decode_need(&p
, end
, 2 + sizeof(u32
), bad
);
3820 info_v
= ceph_decode_8(&p
);
3821 info_cv
= ceph_decode_8(&p
);
3822 info_len
= ceph_decode_32(&p
);
3823 ceph_decode_need(&p
, end
, info_len
, bad
);
3825 info_end
= p
+ info_len
;
3828 ceph_decode_need(&info_p
, info_end
, sizeof(u32
) * 2, bad
);
3829 fscid
= ceph_decode_32(&info_p
);
3830 namelen
= ceph_decode_32(&info_p
);
3831 ceph_decode_need(&info_p
, info_end
, namelen
, bad
);
3833 if (mds_namespace
&&
3834 strlen(mds_namespace
) == namelen
&&
3835 !strncmp(mds_namespace
, (char *)info_p
, namelen
)) {
3836 mount_fscid
= fscid
;
3841 ceph_monc_got_map(&fsc
->client
->monc
, CEPH_SUB_FSMAP
, epoch
);
3842 if (mount_fscid
!= (u32
)-1) {
3843 fsc
->client
->monc
.fs_cluster_id
= mount_fscid
;
3844 ceph_monc_want_map(&fsc
->client
->monc
, CEPH_SUB_MDSMAP
,
3846 ceph_monc_renew_subs(&fsc
->client
->monc
);
3853 pr_err("error decoding fsmap\n");
3855 mutex_lock(&mdsc
->mutex
);
3856 mdsc
->mdsmap_err
= -ENOENT
;
3857 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3858 mutex_unlock(&mdsc
->mutex
);
3863 * handle mds map update.
3865 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3869 void *p
= msg
->front
.iov_base
;
3870 void *end
= p
+ msg
->front
.iov_len
;
3871 struct ceph_mdsmap
*newmap
, *oldmap
;
3872 struct ceph_fsid fsid
;
3875 ceph_decode_need(&p
, end
, sizeof(fsid
)+2*sizeof(u32
), bad
);
3876 ceph_decode_copy(&p
, &fsid
, sizeof(fsid
));
3877 if (ceph_check_fsid(mdsc
->fsc
->client
, &fsid
) < 0)
3879 epoch
= ceph_decode_32(&p
);
3880 maplen
= ceph_decode_32(&p
);
3881 dout("handle_map epoch %u len %d\n", epoch
, (int)maplen
);
3883 /* do we need it? */
3884 mutex_lock(&mdsc
->mutex
);
3885 if (mdsc
->mdsmap
&& epoch
<= mdsc
->mdsmap
->m_epoch
) {
3886 dout("handle_map epoch %u <= our %u\n",
3887 epoch
, mdsc
->mdsmap
->m_epoch
);
3888 mutex_unlock(&mdsc
->mutex
);
3892 newmap
= ceph_mdsmap_decode(&p
, end
);
3893 if (IS_ERR(newmap
)) {
3894 err
= PTR_ERR(newmap
);
3898 /* swap into place */
3900 oldmap
= mdsc
->mdsmap
;
3901 mdsc
->mdsmap
= newmap
;
3902 check_new_map(mdsc
, newmap
, oldmap
);
3903 ceph_mdsmap_destroy(oldmap
);
3905 mdsc
->mdsmap
= newmap
; /* first mds map */
3907 mdsc
->fsc
->sb
->s_maxbytes
= mdsc
->mdsmap
->m_max_file_size
;
3909 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3910 ceph_monc_got_map(&mdsc
->fsc
->client
->monc
, CEPH_SUB_MDSMAP
,
3911 mdsc
->mdsmap
->m_epoch
);
3913 mutex_unlock(&mdsc
->mutex
);
3914 schedule_delayed(mdsc
);
3918 mutex_unlock(&mdsc
->mutex
);
3920 pr_err("error decoding mdsmap %d\n", err
);
3924 static struct ceph_connection
*con_get(struct ceph_connection
*con
)
3926 struct ceph_mds_session
*s
= con
->private;
3928 if (get_session(s
)) {
3929 dout("mdsc con_get %p ok (%d)\n", s
, refcount_read(&s
->s_ref
));
3932 dout("mdsc con_get %p FAIL\n", s
);
3936 static void con_put(struct ceph_connection
*con
)
3938 struct ceph_mds_session
*s
= con
->private;
3940 dout("mdsc con_put %p (%d)\n", s
, refcount_read(&s
->s_ref
) - 1);
3941 ceph_put_mds_session(s
);
3945 * if the client is unresponsive for long enough, the mds will kill
3946 * the session entirely.
3948 static void peer_reset(struct ceph_connection
*con
)
3950 struct ceph_mds_session
*s
= con
->private;
3951 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3953 pr_warn("mds%d closed our session\n", s
->s_mds
);
3954 send_mds_reconnect(mdsc
, s
);
3957 static void dispatch(struct ceph_connection
*con
, struct ceph_msg
*msg
)
3959 struct ceph_mds_session
*s
= con
->private;
3960 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3961 int type
= le16_to_cpu(msg
->hdr
.type
);
3963 mutex_lock(&mdsc
->mutex
);
3964 if (__verify_registered_session(mdsc
, s
) < 0) {
3965 mutex_unlock(&mdsc
->mutex
);
3968 mutex_unlock(&mdsc
->mutex
);
3971 case CEPH_MSG_MDS_MAP
:
3972 ceph_mdsc_handle_mdsmap(mdsc
, msg
);
3974 case CEPH_MSG_FS_MAP_USER
:
3975 ceph_mdsc_handle_fsmap(mdsc
, msg
);
3977 case CEPH_MSG_CLIENT_SESSION
:
3978 handle_session(s
, msg
);
3980 case CEPH_MSG_CLIENT_REPLY
:
3981 handle_reply(s
, msg
);
3983 case CEPH_MSG_CLIENT_REQUEST_FORWARD
:
3984 handle_forward(mdsc
, s
, msg
);
3986 case CEPH_MSG_CLIENT_CAPS
:
3987 ceph_handle_caps(s
, msg
);
3989 case CEPH_MSG_CLIENT_SNAP
:
3990 ceph_handle_snap(mdsc
, s
, msg
);
3992 case CEPH_MSG_CLIENT_LEASE
:
3993 handle_lease(mdsc
, s
, msg
);
3997 pr_err("received unknown message type %d %s\n", type
,
3998 ceph_msg_type_name(type
));
4009 * Note: returned pointer is the address of a structure that's
4010 * managed separately. Caller must *not* attempt to free it.
4012 static struct ceph_auth_handshake
*get_authorizer(struct ceph_connection
*con
,
4013 int *proto
, int force_new
)
4015 struct ceph_mds_session
*s
= con
->private;
4016 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4017 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
4018 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
4020 if (force_new
&& auth
->authorizer
) {
4021 ceph_auth_destroy_authorizer(auth
->authorizer
);
4022 auth
->authorizer
= NULL
;
4024 if (!auth
->authorizer
) {
4025 int ret
= ceph_auth_create_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
4028 return ERR_PTR(ret
);
4030 int ret
= ceph_auth_update_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
4033 return ERR_PTR(ret
);
4035 *proto
= ac
->protocol
;
4041 static int verify_authorizer_reply(struct ceph_connection
*con
)
4043 struct ceph_mds_session
*s
= con
->private;
4044 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4045 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
4047 return ceph_auth_verify_authorizer_reply(ac
, s
->s_auth
.authorizer
);
4050 static int invalidate_authorizer(struct ceph_connection
*con
)
4052 struct ceph_mds_session
*s
= con
->private;
4053 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
4054 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
4056 ceph_auth_invalidate_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
);
4058 return ceph_monc_validate_auth(&mdsc
->fsc
->client
->monc
);
4061 static struct ceph_msg
*mds_alloc_msg(struct ceph_connection
*con
,
4062 struct ceph_msg_header
*hdr
, int *skip
)
4064 struct ceph_msg
*msg
;
4065 int type
= (int) le16_to_cpu(hdr
->type
);
4066 int front_len
= (int) le32_to_cpu(hdr
->front_len
);
4072 msg
= ceph_msg_new(type
, front_len
, GFP_NOFS
, false);
4074 pr_err("unable to allocate msg type %d len %d\n",
4082 static int mds_sign_message(struct ceph_msg
*msg
)
4084 struct ceph_mds_session
*s
= msg
->con
->private;
4085 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
4087 return ceph_auth_sign_message(auth
, msg
);
4090 static int mds_check_message_signature(struct ceph_msg
*msg
)
4092 struct ceph_mds_session
*s
= msg
->con
->private;
4093 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
4095 return ceph_auth_check_message_signature(auth
, msg
);
4098 static const struct ceph_connection_operations mds_con_ops
= {
4101 .dispatch
= dispatch
,
4102 .get_authorizer
= get_authorizer
,
4103 .verify_authorizer_reply
= verify_authorizer_reply
,
4104 .invalidate_authorizer
= invalidate_authorizer
,
4105 .peer_reset
= peer_reset
,
4106 .alloc_msg
= mds_alloc_msg
,
4107 .sign_message
= mds_sign_message
,
4108 .check_message_signature
= mds_check_message_signature
,