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>
12 #include "mds_client.h"
14 #include <linux/ceph/ceph_features.h>
15 #include <linux/ceph/messenger.h>
16 #include <linux/ceph/decode.h>
17 #include <linux/ceph/pagelist.h>
18 #include <linux/ceph/auth.h>
19 #include <linux/ceph/debugfs.h>
22 * A cluster of MDS (metadata server) daemons is responsible for
23 * managing the file system namespace (the directory hierarchy and
24 * inodes) and for coordinating shared access to storage. Metadata is
25 * partitioning hierarchically across a number of servers, and that
26 * partition varies over time as the cluster adjusts the distribution
27 * in order to balance load.
29 * The MDS client is primarily responsible to managing synchronous
30 * metadata requests for operations like open, unlink, and so forth.
31 * If there is a MDS failure, we find out about it when we (possibly
32 * request and) receive a new MDS map, and can resubmit affected
35 * For the most part, though, we take advantage of a lossless
36 * communications channel to the MDS, and do not need to worry about
37 * timing out or resubmitting requests.
39 * We maintain a stateful "session" with each MDS we interact with.
40 * Within each session, we sent periodic heartbeat messages to ensure
41 * any capabilities or leases we have been issues remain valid. If
42 * the session times out and goes stale, our leases and capabilities
43 * are no longer valid.
46 struct ceph_reconnect_state
{
48 struct ceph_pagelist
*pagelist
;
52 static void __wake_requests(struct ceph_mds_client
*mdsc
,
53 struct list_head
*head
);
55 static const struct ceph_connection_operations mds_con_ops
;
63 * parse individual inode info
65 static int parse_reply_info_in(void **p
, void *end
,
66 struct ceph_mds_reply_info_in
*info
,
72 *p
+= sizeof(struct ceph_mds_reply_inode
) +
73 sizeof(*info
->in
->fragtree
.splits
) *
74 le32_to_cpu(info
->in
->fragtree
.nsplits
);
76 ceph_decode_32_safe(p
, end
, info
->symlink_len
, bad
);
77 ceph_decode_need(p
, end
, info
->symlink_len
, bad
);
79 *p
+= info
->symlink_len
;
81 if (features
& CEPH_FEATURE_DIRLAYOUTHASH
)
82 ceph_decode_copy_safe(p
, end
, &info
->dir_layout
,
83 sizeof(info
->dir_layout
), bad
);
85 memset(&info
->dir_layout
, 0, sizeof(info
->dir_layout
));
87 ceph_decode_32_safe(p
, end
, info
->xattr_len
, bad
);
88 ceph_decode_need(p
, end
, info
->xattr_len
, bad
);
89 info
->xattr_data
= *p
;
90 *p
+= info
->xattr_len
;
97 * parse a normal reply, which may contain a (dir+)dentry and/or a
100 static int parse_reply_info_trace(void **p
, void *end
,
101 struct ceph_mds_reply_info_parsed
*info
,
106 if (info
->head
->is_dentry
) {
107 err
= parse_reply_info_in(p
, end
, &info
->diri
, features
);
111 if (unlikely(*p
+ sizeof(*info
->dirfrag
) > end
))
114 *p
+= sizeof(*info
->dirfrag
) +
115 sizeof(u32
)*le32_to_cpu(info
->dirfrag
->ndist
);
116 if (unlikely(*p
> end
))
119 ceph_decode_32_safe(p
, end
, info
->dname_len
, bad
);
120 ceph_decode_need(p
, end
, info
->dname_len
, bad
);
122 *p
+= info
->dname_len
;
124 *p
+= sizeof(*info
->dlease
);
127 if (info
->head
->is_target
) {
128 err
= parse_reply_info_in(p
, end
, &info
->targeti
, features
);
133 if (unlikely(*p
!= end
))
140 pr_err("problem parsing mds trace %d\n", err
);
145 * parse readdir results
147 static int parse_reply_info_dir(void **p
, void *end
,
148 struct ceph_mds_reply_info_parsed
*info
,
155 if (*p
+ sizeof(*info
->dir_dir
) > end
)
157 *p
+= sizeof(*info
->dir_dir
) +
158 sizeof(u32
)*le32_to_cpu(info
->dir_dir
->ndist
);
162 ceph_decode_need(p
, end
, sizeof(num
) + 2, bad
);
163 num
= ceph_decode_32(p
);
164 info
->dir_end
= ceph_decode_8(p
);
165 info
->dir_complete
= ceph_decode_8(p
);
169 BUG_ON(!info
->dir_in
);
170 info
->dir_dname
= (void *)(info
->dir_in
+ num
);
171 info
->dir_dname_len
= (void *)(info
->dir_dname
+ num
);
172 info
->dir_dlease
= (void *)(info
->dir_dname_len
+ num
);
173 if ((unsigned long)(info
->dir_dlease
+ num
) >
174 (unsigned long)info
->dir_in
+ info
->dir_buf_size
) {
175 pr_err("dir contents are larger than expected\n");
183 ceph_decode_need(p
, end
, sizeof(u32
)*2, bad
);
184 info
->dir_dname_len
[i
] = ceph_decode_32(p
);
185 ceph_decode_need(p
, end
, info
->dir_dname_len
[i
], bad
);
186 info
->dir_dname
[i
] = *p
;
187 *p
+= info
->dir_dname_len
[i
];
188 dout("parsed dir dname '%.*s'\n", info
->dir_dname_len
[i
],
190 info
->dir_dlease
[i
] = *p
;
191 *p
+= sizeof(struct ceph_mds_reply_lease
);
194 err
= parse_reply_info_in(p
, end
, &info
->dir_in
[i
], features
);
209 pr_err("problem parsing dir contents %d\n", err
);
214 * parse fcntl F_GETLK results
216 static int parse_reply_info_filelock(void **p
, void *end
,
217 struct ceph_mds_reply_info_parsed
*info
,
220 if (*p
+ sizeof(*info
->filelock_reply
) > end
)
223 info
->filelock_reply
= *p
;
224 *p
+= sizeof(*info
->filelock_reply
);
226 if (unlikely(*p
!= end
))
235 * parse create results
237 static int parse_reply_info_create(void **p
, void *end
,
238 struct ceph_mds_reply_info_parsed
*info
,
241 if (features
& CEPH_FEATURE_REPLY_CREATE_INODE
) {
243 info
->has_create_ino
= false;
245 info
->has_create_ino
= true;
246 info
->ino
= ceph_decode_64(p
);
250 if (unlikely(*p
!= end
))
259 * parse extra results
261 static int parse_reply_info_extra(void **p
, void *end
,
262 struct ceph_mds_reply_info_parsed
*info
,
265 if (info
->head
->op
== CEPH_MDS_OP_GETFILELOCK
)
266 return parse_reply_info_filelock(p
, end
, info
, features
);
267 else if (info
->head
->op
== CEPH_MDS_OP_READDIR
||
268 info
->head
->op
== CEPH_MDS_OP_LSSNAP
)
269 return parse_reply_info_dir(p
, end
, info
, features
);
270 else if (info
->head
->op
== CEPH_MDS_OP_CREATE
)
271 return parse_reply_info_create(p
, end
, info
, features
);
277 * parse entire mds reply
279 static int parse_reply_info(struct ceph_msg
*msg
,
280 struct ceph_mds_reply_info_parsed
*info
,
287 info
->head
= msg
->front
.iov_base
;
288 p
= msg
->front
.iov_base
+ sizeof(struct ceph_mds_reply_head
);
289 end
= p
+ msg
->front
.iov_len
- sizeof(struct ceph_mds_reply_head
);
292 ceph_decode_32_safe(&p
, end
, len
, bad
);
294 ceph_decode_need(&p
, end
, len
, bad
);
295 err
= parse_reply_info_trace(&p
, p
+len
, info
, features
);
301 ceph_decode_32_safe(&p
, end
, len
, bad
);
303 ceph_decode_need(&p
, end
, len
, bad
);
304 err
= parse_reply_info_extra(&p
, p
+len
, info
, features
);
310 ceph_decode_32_safe(&p
, end
, len
, bad
);
311 info
->snapblob_len
= len
;
322 pr_err("mds parse_reply err %d\n", err
);
326 static void destroy_reply_info(struct ceph_mds_reply_info_parsed
*info
)
330 free_pages((unsigned long)info
->dir_in
, get_order(info
->dir_buf_size
));
337 static const char *session_state_name(int s
)
340 case CEPH_MDS_SESSION_NEW
: return "new";
341 case CEPH_MDS_SESSION_OPENING
: return "opening";
342 case CEPH_MDS_SESSION_OPEN
: return "open";
343 case CEPH_MDS_SESSION_HUNG
: return "hung";
344 case CEPH_MDS_SESSION_CLOSING
: return "closing";
345 case CEPH_MDS_SESSION_RESTARTING
: return "restarting";
346 case CEPH_MDS_SESSION_RECONNECTING
: return "reconnecting";
347 default: return "???";
351 static struct ceph_mds_session
*get_session(struct ceph_mds_session
*s
)
353 if (atomic_inc_not_zero(&s
->s_ref
)) {
354 dout("mdsc get_session %p %d -> %d\n", s
,
355 atomic_read(&s
->s_ref
)-1, atomic_read(&s
->s_ref
));
358 dout("mdsc get_session %p 0 -- FAIL", s
);
363 void ceph_put_mds_session(struct ceph_mds_session
*s
)
365 dout("mdsc put_session %p %d -> %d\n", s
,
366 atomic_read(&s
->s_ref
), atomic_read(&s
->s_ref
)-1);
367 if (atomic_dec_and_test(&s
->s_ref
)) {
368 if (s
->s_auth
.authorizer
)
369 ceph_auth_destroy_authorizer(
370 s
->s_mdsc
->fsc
->client
->monc
.auth
,
371 s
->s_auth
.authorizer
);
377 * called under mdsc->mutex
379 struct ceph_mds_session
*__ceph_lookup_mds_session(struct ceph_mds_client
*mdsc
,
382 struct ceph_mds_session
*session
;
384 if (mds
>= mdsc
->max_sessions
|| mdsc
->sessions
[mds
] == NULL
)
386 session
= mdsc
->sessions
[mds
];
387 dout("lookup_mds_session %p %d\n", session
,
388 atomic_read(&session
->s_ref
));
389 get_session(session
);
393 static bool __have_session(struct ceph_mds_client
*mdsc
, int mds
)
395 if (mds
>= mdsc
->max_sessions
)
397 return mdsc
->sessions
[mds
];
400 static int __verify_registered_session(struct ceph_mds_client
*mdsc
,
401 struct ceph_mds_session
*s
)
403 if (s
->s_mds
>= mdsc
->max_sessions
||
404 mdsc
->sessions
[s
->s_mds
] != s
)
410 * create+register a new session for given mds.
411 * called under mdsc->mutex.
413 static struct ceph_mds_session
*register_session(struct ceph_mds_client
*mdsc
,
416 struct ceph_mds_session
*s
;
418 if (mds
>= mdsc
->mdsmap
->m_max_mds
)
419 return ERR_PTR(-EINVAL
);
421 s
= kzalloc(sizeof(*s
), GFP_NOFS
);
423 return ERR_PTR(-ENOMEM
);
426 s
->s_state
= CEPH_MDS_SESSION_NEW
;
429 mutex_init(&s
->s_mutex
);
431 ceph_con_init(&s
->s_con
, s
, &mds_con_ops
, &mdsc
->fsc
->client
->msgr
);
433 spin_lock_init(&s
->s_gen_ttl_lock
);
435 s
->s_cap_ttl
= jiffies
- 1;
437 spin_lock_init(&s
->s_cap_lock
);
438 s
->s_renew_requested
= 0;
440 INIT_LIST_HEAD(&s
->s_caps
);
443 atomic_set(&s
->s_ref
, 1);
444 INIT_LIST_HEAD(&s
->s_waiting
);
445 INIT_LIST_HEAD(&s
->s_unsafe
);
446 s
->s_num_cap_releases
= 0;
447 s
->s_cap_reconnect
= 0;
448 s
->s_cap_iterator
= NULL
;
449 INIT_LIST_HEAD(&s
->s_cap_releases
);
450 INIT_LIST_HEAD(&s
->s_cap_releases_done
);
451 INIT_LIST_HEAD(&s
->s_cap_flushing
);
452 INIT_LIST_HEAD(&s
->s_cap_snaps_flushing
);
454 dout("register_session mds%d\n", mds
);
455 if (mds
>= mdsc
->max_sessions
) {
456 int newmax
= 1 << get_count_order(mds
+1);
457 struct ceph_mds_session
**sa
;
459 dout("register_session realloc to %d\n", newmax
);
460 sa
= kcalloc(newmax
, sizeof(void *), GFP_NOFS
);
463 if (mdsc
->sessions
) {
464 memcpy(sa
, mdsc
->sessions
,
465 mdsc
->max_sessions
* sizeof(void *));
466 kfree(mdsc
->sessions
);
469 mdsc
->max_sessions
= newmax
;
471 mdsc
->sessions
[mds
] = s
;
472 atomic_inc(&s
->s_ref
); /* one ref to sessions[], one to caller */
474 ceph_con_open(&s
->s_con
, CEPH_ENTITY_TYPE_MDS
, mds
,
475 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
481 return ERR_PTR(-ENOMEM
);
485 * called under mdsc->mutex
487 static void __unregister_session(struct ceph_mds_client
*mdsc
,
488 struct ceph_mds_session
*s
)
490 dout("__unregister_session mds%d %p\n", s
->s_mds
, s
);
491 BUG_ON(mdsc
->sessions
[s
->s_mds
] != s
);
492 mdsc
->sessions
[s
->s_mds
] = NULL
;
493 ceph_con_close(&s
->s_con
);
494 ceph_put_mds_session(s
);
498 * drop session refs in request.
500 * should be last request ref, or hold mdsc->mutex
502 static void put_request_session(struct ceph_mds_request
*req
)
504 if (req
->r_session
) {
505 ceph_put_mds_session(req
->r_session
);
506 req
->r_session
= NULL
;
510 void ceph_mdsc_release_request(struct kref
*kref
)
512 struct ceph_mds_request
*req
= container_of(kref
,
513 struct ceph_mds_request
,
515 destroy_reply_info(&req
->r_reply_info
);
517 ceph_msg_put(req
->r_request
);
519 ceph_msg_put(req
->r_reply
);
521 ceph_put_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
524 if (req
->r_locked_dir
)
525 ceph_put_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
526 if (req
->r_target_inode
)
527 iput(req
->r_target_inode
);
530 if (req
->r_old_dentry
)
531 dput(req
->r_old_dentry
);
532 if (req
->r_old_dentry_dir
) {
534 * track (and drop pins for) r_old_dentry_dir
535 * separately, since r_old_dentry's d_parent may have
536 * changed between the dir mutex being dropped and
537 * this request being freed.
539 ceph_put_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
541 iput(req
->r_old_dentry_dir
);
545 put_request_session(req
);
546 ceph_unreserve_caps(req
->r_mdsc
, &req
->r_caps_reservation
);
551 * lookup session, bump ref if found.
553 * called under mdsc->mutex.
555 static struct ceph_mds_request
*__lookup_request(struct ceph_mds_client
*mdsc
,
558 struct ceph_mds_request
*req
;
559 struct rb_node
*n
= mdsc
->request_tree
.rb_node
;
562 req
= rb_entry(n
, struct ceph_mds_request
, r_node
);
563 if (tid
< req
->r_tid
)
565 else if (tid
> req
->r_tid
)
568 ceph_mdsc_get_request(req
);
575 static void __insert_request(struct ceph_mds_client
*mdsc
,
576 struct ceph_mds_request
*new)
578 struct rb_node
**p
= &mdsc
->request_tree
.rb_node
;
579 struct rb_node
*parent
= NULL
;
580 struct ceph_mds_request
*req
= NULL
;
584 req
= rb_entry(parent
, struct ceph_mds_request
, r_node
);
585 if (new->r_tid
< req
->r_tid
)
587 else if (new->r_tid
> req
->r_tid
)
593 rb_link_node(&new->r_node
, parent
, p
);
594 rb_insert_color(&new->r_node
, &mdsc
->request_tree
);
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
, req
);
615 req
->r_uid
= current_fsuid();
616 req
->r_gid
= current_fsgid();
619 struct ceph_inode_info
*ci
= ceph_inode(dir
);
622 spin_lock(&ci
->i_unsafe_lock
);
623 req
->r_unsafe_dir
= dir
;
624 list_add_tail(&req
->r_unsafe_dir_item
, &ci
->i_unsafe_dirops
);
625 spin_unlock(&ci
->i_unsafe_lock
);
629 static void __unregister_request(struct ceph_mds_client
*mdsc
,
630 struct ceph_mds_request
*req
)
632 dout("__unregister_request %p tid %lld\n", req
, req
->r_tid
);
633 rb_erase(&req
->r_node
, &mdsc
->request_tree
);
634 RB_CLEAR_NODE(&req
->r_node
);
636 if (req
->r_unsafe_dir
) {
637 struct ceph_inode_info
*ci
= ceph_inode(req
->r_unsafe_dir
);
639 spin_lock(&ci
->i_unsafe_lock
);
640 list_del_init(&req
->r_unsafe_dir_item
);
641 spin_unlock(&ci
->i_unsafe_lock
);
643 iput(req
->r_unsafe_dir
);
644 req
->r_unsafe_dir
= NULL
;
647 complete_all(&req
->r_safe_completion
);
649 ceph_mdsc_put_request(req
);
653 * Choose mds to send request to next. If there is a hint set in the
654 * request (e.g., due to a prior forward hint from the mds), use that.
655 * Otherwise, consult frag tree and/or caps to identify the
656 * appropriate mds. If all else fails, choose randomly.
658 * Called under mdsc->mutex.
660 static struct dentry
*get_nonsnap_parent(struct dentry
*dentry
)
663 * we don't need to worry about protecting the d_parent access
664 * here because we never renaming inside the snapped namespace
665 * except to resplice to another snapdir, and either the old or new
666 * result is a valid result.
668 while (!IS_ROOT(dentry
) && ceph_snap(dentry
->d_inode
) != CEPH_NOSNAP
)
669 dentry
= dentry
->d_parent
;
673 static int __choose_mds(struct ceph_mds_client
*mdsc
,
674 struct ceph_mds_request
*req
)
677 struct ceph_inode_info
*ci
;
678 struct ceph_cap
*cap
;
679 int mode
= req
->r_direct_mode
;
681 u32 hash
= req
->r_direct_hash
;
682 bool is_hash
= req
->r_direct_is_hash
;
685 * is there a specific mds we should try? ignore hint if we have
686 * no session and the mds is not up (active or recovering).
688 if (req
->r_resend_mds
>= 0 &&
689 (__have_session(mdsc
, req
->r_resend_mds
) ||
690 ceph_mdsmap_get_state(mdsc
->mdsmap
, req
->r_resend_mds
) > 0)) {
691 dout("choose_mds using resend_mds mds%d\n",
693 return req
->r_resend_mds
;
696 if (mode
== USE_RANDOM_MDS
)
701 inode
= req
->r_inode
;
702 } else if (req
->r_dentry
) {
703 /* ignore race with rename; old or new d_parent is okay */
704 struct dentry
*parent
= req
->r_dentry
->d_parent
;
705 struct inode
*dir
= parent
->d_inode
;
707 if (dir
->i_sb
!= mdsc
->fsc
->sb
) {
709 inode
= req
->r_dentry
->d_inode
;
710 } else if (ceph_snap(dir
) != CEPH_NOSNAP
) {
711 /* direct snapped/virtual snapdir requests
712 * based on parent dir inode */
713 struct dentry
*dn
= get_nonsnap_parent(parent
);
715 dout("__choose_mds using nonsnap parent %p\n", inode
);
718 inode
= req
->r_dentry
->d_inode
;
719 if (!inode
|| mode
== USE_AUTH_MDS
) {
722 hash
= ceph_dentry_hash(dir
, req
->r_dentry
);
728 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode
, (int)is_hash
,
732 ci
= ceph_inode(inode
);
734 if (is_hash
&& S_ISDIR(inode
->i_mode
)) {
735 struct ceph_inode_frag frag
;
738 ceph_choose_frag(ci
, hash
, &frag
, &found
);
740 if (mode
== USE_ANY_MDS
&& frag
.ndist
> 0) {
743 /* choose a random replica */
744 get_random_bytes(&r
, 1);
747 dout("choose_mds %p %llx.%llx "
748 "frag %u mds%d (%d/%d)\n",
749 inode
, ceph_vinop(inode
),
752 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
753 CEPH_MDS_STATE_ACTIVE
)
757 /* since this file/dir wasn't known to be
758 * replicated, then we want to look for the
759 * authoritative mds. */
762 /* choose auth mds */
764 dout("choose_mds %p %llx.%llx "
765 "frag %u mds%d (auth)\n",
766 inode
, ceph_vinop(inode
), frag
.frag
, mds
);
767 if (ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) >=
768 CEPH_MDS_STATE_ACTIVE
)
774 spin_lock(&ci
->i_ceph_lock
);
776 if (mode
== USE_AUTH_MDS
)
777 cap
= ci
->i_auth_cap
;
778 if (!cap
&& !RB_EMPTY_ROOT(&ci
->i_caps
))
779 cap
= rb_entry(rb_first(&ci
->i_caps
), struct ceph_cap
, ci_node
);
781 spin_unlock(&ci
->i_ceph_lock
);
784 mds
= cap
->session
->s_mds
;
785 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
786 inode
, ceph_vinop(inode
), mds
,
787 cap
== ci
->i_auth_cap
? "auth " : "", cap
);
788 spin_unlock(&ci
->i_ceph_lock
);
792 mds
= ceph_mdsmap_get_random_mds(mdsc
->mdsmap
);
793 dout("choose_mds chose random mds%d\n", mds
);
801 static struct ceph_msg
*create_session_msg(u32 op
, u64 seq
)
803 struct ceph_msg
*msg
;
804 struct ceph_mds_session_head
*h
;
806 msg
= ceph_msg_new(CEPH_MSG_CLIENT_SESSION
, sizeof(*h
), GFP_NOFS
,
809 pr_err("create_session_msg ENOMEM creating msg\n");
812 h
= msg
->front
.iov_base
;
813 h
->op
= cpu_to_le32(op
);
814 h
->seq
= cpu_to_le64(seq
);
819 * send session open request.
821 * called under mdsc->mutex
823 static int __open_session(struct ceph_mds_client
*mdsc
,
824 struct ceph_mds_session
*session
)
826 struct ceph_msg
*msg
;
828 int mds
= session
->s_mds
;
830 /* wait for mds to go active? */
831 mstate
= ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
);
832 dout("open_session to mds%d (%s)\n", mds
,
833 ceph_mds_state_name(mstate
));
834 session
->s_state
= CEPH_MDS_SESSION_OPENING
;
835 session
->s_renew_requested
= jiffies
;
837 /* send connect message */
838 msg
= create_session_msg(CEPH_SESSION_REQUEST_OPEN
, session
->s_seq
);
841 ceph_con_send(&session
->s_con
, msg
);
846 * open sessions for any export targets for the given mds
848 * called under mdsc->mutex
850 static struct ceph_mds_session
*
851 __open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
853 struct ceph_mds_session
*session
;
855 session
= __ceph_lookup_mds_session(mdsc
, target
);
857 session
= register_session(mdsc
, target
);
861 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
862 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
863 __open_session(mdsc
, session
);
868 struct ceph_mds_session
*
869 ceph_mdsc_open_export_target_session(struct ceph_mds_client
*mdsc
, int target
)
871 struct ceph_mds_session
*session
;
873 dout("open_export_target_session to mds%d\n", target
);
875 mutex_lock(&mdsc
->mutex
);
876 session
= __open_export_target_session(mdsc
, target
);
877 mutex_unlock(&mdsc
->mutex
);
882 static void __open_export_target_sessions(struct ceph_mds_client
*mdsc
,
883 struct ceph_mds_session
*session
)
885 struct ceph_mds_info
*mi
;
886 struct ceph_mds_session
*ts
;
887 int i
, mds
= session
->s_mds
;
889 if (mds
>= mdsc
->mdsmap
->m_max_mds
)
892 mi
= &mdsc
->mdsmap
->m_info
[mds
];
893 dout("open_export_target_sessions for mds%d (%d targets)\n",
894 session
->s_mds
, mi
->num_export_targets
);
896 for (i
= 0; i
< mi
->num_export_targets
; i
++) {
897 ts
= __open_export_target_session(mdsc
, mi
->export_targets
[i
]);
899 ceph_put_mds_session(ts
);
903 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client
*mdsc
,
904 struct ceph_mds_session
*session
)
906 mutex_lock(&mdsc
->mutex
);
907 __open_export_target_sessions(mdsc
, session
);
908 mutex_unlock(&mdsc
->mutex
);
916 * Free preallocated cap messages assigned to this session
918 static void cleanup_cap_releases(struct ceph_mds_session
*session
)
920 struct ceph_msg
*msg
;
922 spin_lock(&session
->s_cap_lock
);
923 while (!list_empty(&session
->s_cap_releases
)) {
924 msg
= list_first_entry(&session
->s_cap_releases
,
925 struct ceph_msg
, list_head
);
926 list_del_init(&msg
->list_head
);
929 while (!list_empty(&session
->s_cap_releases_done
)) {
930 msg
= list_first_entry(&session
->s_cap_releases_done
,
931 struct ceph_msg
, list_head
);
932 list_del_init(&msg
->list_head
);
935 spin_unlock(&session
->s_cap_lock
);
939 * Helper to safely iterate over all caps associated with a session, with
940 * special care taken to handle a racing __ceph_remove_cap().
942 * Caller must hold session s_mutex.
944 static int iterate_session_caps(struct ceph_mds_session
*session
,
945 int (*cb
)(struct inode
*, struct ceph_cap
*,
949 struct ceph_cap
*cap
;
950 struct inode
*inode
, *last_inode
= NULL
;
951 struct ceph_cap
*old_cap
= NULL
;
954 dout("iterate_session_caps %p mds%d\n", session
, session
->s_mds
);
955 spin_lock(&session
->s_cap_lock
);
956 p
= session
->s_caps
.next
;
957 while (p
!= &session
->s_caps
) {
958 cap
= list_entry(p
, struct ceph_cap
, session_caps
);
959 inode
= igrab(&cap
->ci
->vfs_inode
);
964 session
->s_cap_iterator
= cap
;
965 spin_unlock(&session
->s_cap_lock
);
972 ceph_put_cap(session
->s_mdsc
, old_cap
);
976 ret
= cb(inode
, cap
, arg
);
979 spin_lock(&session
->s_cap_lock
);
981 if (cap
->ci
== NULL
) {
982 dout("iterate_session_caps finishing cap %p removal\n",
984 BUG_ON(cap
->session
!= session
);
985 list_del_init(&cap
->session_caps
);
986 session
->s_nr_caps
--;
988 old_cap
= cap
; /* put_cap it w/o locks held */
995 session
->s_cap_iterator
= NULL
;
996 spin_unlock(&session
->s_cap_lock
);
1001 ceph_put_cap(session
->s_mdsc
, old_cap
);
1006 static int remove_session_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1009 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1012 dout("removing cap %p, ci is %p, inode is %p\n",
1013 cap
, ci
, &ci
->vfs_inode
);
1014 spin_lock(&ci
->i_ceph_lock
);
1015 __ceph_remove_cap(cap
, false);
1016 if (!__ceph_is_any_real_caps(ci
)) {
1017 struct ceph_mds_client
*mdsc
=
1018 ceph_sb_to_client(inode
->i_sb
)->mdsc
;
1020 spin_lock(&mdsc
->cap_dirty_lock
);
1021 if (!list_empty(&ci
->i_dirty_item
)) {
1022 pr_info(" dropping dirty %s state for %p %lld\n",
1023 ceph_cap_string(ci
->i_dirty_caps
),
1024 inode
, ceph_ino(inode
));
1025 ci
->i_dirty_caps
= 0;
1026 list_del_init(&ci
->i_dirty_item
);
1029 if (!list_empty(&ci
->i_flushing_item
)) {
1030 pr_info(" dropping dirty+flushing %s state for %p %lld\n",
1031 ceph_cap_string(ci
->i_flushing_caps
),
1032 inode
, ceph_ino(inode
));
1033 ci
->i_flushing_caps
= 0;
1034 list_del_init(&ci
->i_flushing_item
);
1035 mdsc
->num_cap_flushing
--;
1038 if (drop
&& ci
->i_wrbuffer_ref
) {
1039 pr_info(" dropping dirty data for %p %lld\n",
1040 inode
, ceph_ino(inode
));
1041 ci
->i_wrbuffer_ref
= 0;
1042 ci
->i_wrbuffer_ref_head
= 0;
1045 spin_unlock(&mdsc
->cap_dirty_lock
);
1047 spin_unlock(&ci
->i_ceph_lock
);
1054 * caller must hold session s_mutex
1056 static void remove_session_caps(struct ceph_mds_session
*session
)
1058 dout("remove_session_caps on %p\n", session
);
1059 iterate_session_caps(session
, remove_session_caps_cb
, NULL
);
1061 spin_lock(&session
->s_cap_lock
);
1062 if (session
->s_nr_caps
> 0) {
1063 struct super_block
*sb
= session
->s_mdsc
->fsc
->sb
;
1064 struct inode
*inode
;
1065 struct ceph_cap
*cap
, *prev
= NULL
;
1066 struct ceph_vino vino
;
1068 * iterate_session_caps() skips inodes that are being
1069 * deleted, we need to wait until deletions are complete.
1070 * __wait_on_freeing_inode() is designed for the job,
1071 * but it is not exported, so use lookup inode function
1074 while (!list_empty(&session
->s_caps
)) {
1075 cap
= list_entry(session
->s_caps
.next
,
1076 struct ceph_cap
, session_caps
);
1080 vino
= cap
->ci
->i_vino
;
1081 spin_unlock(&session
->s_cap_lock
);
1083 inode
= ceph_find_inode(sb
, vino
);
1086 spin_lock(&session
->s_cap_lock
);
1089 spin_unlock(&session
->s_cap_lock
);
1091 BUG_ON(session
->s_nr_caps
> 0);
1092 BUG_ON(!list_empty(&session
->s_cap_flushing
));
1093 cleanup_cap_releases(session
);
1097 * wake up any threads waiting on this session's caps. if the cap is
1098 * old (didn't get renewed on the client reconnect), remove it now.
1100 * caller must hold s_mutex.
1102 static int wake_up_session_cb(struct inode
*inode
, struct ceph_cap
*cap
,
1105 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1107 wake_up_all(&ci
->i_cap_wq
);
1109 spin_lock(&ci
->i_ceph_lock
);
1110 ci
->i_wanted_max_size
= 0;
1111 ci
->i_requested_max_size
= 0;
1112 spin_unlock(&ci
->i_ceph_lock
);
1117 static void wake_up_session_caps(struct ceph_mds_session
*session
,
1120 dout("wake_up_session_caps %p mds%d\n", session
, session
->s_mds
);
1121 iterate_session_caps(session
, wake_up_session_cb
,
1122 (void *)(unsigned long)reconnect
);
1126 * Send periodic message to MDS renewing all currently held caps. The
1127 * ack will reset the expiration for all caps from this session.
1129 * caller holds s_mutex
1131 static int send_renew_caps(struct ceph_mds_client
*mdsc
,
1132 struct ceph_mds_session
*session
)
1134 struct ceph_msg
*msg
;
1137 if (time_after_eq(jiffies
, session
->s_cap_ttl
) &&
1138 time_after_eq(session
->s_cap_ttl
, session
->s_renew_requested
))
1139 pr_info("mds%d caps stale\n", session
->s_mds
);
1140 session
->s_renew_requested
= jiffies
;
1142 /* do not try to renew caps until a recovering mds has reconnected
1143 * with its clients. */
1144 state
= ceph_mdsmap_get_state(mdsc
->mdsmap
, session
->s_mds
);
1145 if (state
< CEPH_MDS_STATE_RECONNECT
) {
1146 dout("send_renew_caps ignoring mds%d (%s)\n",
1147 session
->s_mds
, ceph_mds_state_name(state
));
1151 dout("send_renew_caps to mds%d (%s)\n", session
->s_mds
,
1152 ceph_mds_state_name(state
));
1153 msg
= create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS
,
1154 ++session
->s_renew_seq
);
1157 ceph_con_send(&session
->s_con
, msg
);
1161 static int send_flushmsg_ack(struct ceph_mds_client
*mdsc
,
1162 struct ceph_mds_session
*session
, u64 seq
)
1164 struct ceph_msg
*msg
;
1166 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1167 session
->s_mds
, session_state_name(session
->s_state
), seq
);
1168 msg
= create_session_msg(CEPH_SESSION_FLUSHMSG_ACK
, seq
);
1171 ceph_con_send(&session
->s_con
, msg
);
1177 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1179 * Called under session->s_mutex
1181 static void renewed_caps(struct ceph_mds_client
*mdsc
,
1182 struct ceph_mds_session
*session
, int is_renew
)
1187 spin_lock(&session
->s_cap_lock
);
1188 was_stale
= is_renew
&& time_after_eq(jiffies
, session
->s_cap_ttl
);
1190 session
->s_cap_ttl
= session
->s_renew_requested
+
1191 mdsc
->mdsmap
->m_session_timeout
*HZ
;
1194 if (time_before(jiffies
, session
->s_cap_ttl
)) {
1195 pr_info("mds%d caps renewed\n", session
->s_mds
);
1198 pr_info("mds%d caps still stale\n", session
->s_mds
);
1201 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1202 session
->s_mds
, session
->s_cap_ttl
, was_stale
? "stale" : "fresh",
1203 time_before(jiffies
, session
->s_cap_ttl
) ? "stale" : "fresh");
1204 spin_unlock(&session
->s_cap_lock
);
1207 wake_up_session_caps(session
, 0);
1211 * send a session close request
1213 static int request_close_session(struct ceph_mds_client
*mdsc
,
1214 struct ceph_mds_session
*session
)
1216 struct ceph_msg
*msg
;
1218 dout("request_close_session mds%d state %s seq %lld\n",
1219 session
->s_mds
, session_state_name(session
->s_state
),
1221 msg
= create_session_msg(CEPH_SESSION_REQUEST_CLOSE
, session
->s_seq
);
1224 ceph_con_send(&session
->s_con
, msg
);
1229 * Called with s_mutex held.
1231 static int __close_session(struct ceph_mds_client
*mdsc
,
1232 struct ceph_mds_session
*session
)
1234 if (session
->s_state
>= CEPH_MDS_SESSION_CLOSING
)
1236 session
->s_state
= CEPH_MDS_SESSION_CLOSING
;
1237 return request_close_session(mdsc
, session
);
1241 * Trim old(er) caps.
1243 * Because we can't cache an inode without one or more caps, we do
1244 * this indirectly: if a cap is unused, we prune its aliases, at which
1245 * point the inode will hopefully get dropped to.
1247 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1248 * memory pressure from the MDS, though, so it needn't be perfect.
1250 static int trim_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
, void *arg
)
1252 struct ceph_mds_session
*session
= arg
;
1253 struct ceph_inode_info
*ci
= ceph_inode(inode
);
1254 int used
, wanted
, oissued
, mine
;
1256 if (session
->s_trim_caps
<= 0)
1259 spin_lock(&ci
->i_ceph_lock
);
1260 mine
= cap
->issued
| cap
->implemented
;
1261 used
= __ceph_caps_used(ci
);
1262 wanted
= __ceph_caps_file_wanted(ci
);
1263 oissued
= __ceph_caps_issued_other(ci
, cap
);
1265 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1266 inode
, cap
, ceph_cap_string(mine
), ceph_cap_string(oissued
),
1267 ceph_cap_string(used
), ceph_cap_string(wanted
));
1268 if (cap
== ci
->i_auth_cap
) {
1269 if (ci
->i_dirty_caps
| ci
->i_flushing_caps
)
1271 if ((used
| wanted
) & CEPH_CAP_ANY_WR
)
1274 if ((used
| wanted
) & ~oissued
& mine
)
1275 goto out
; /* we need these caps */
1277 session
->s_trim_caps
--;
1279 /* we aren't the only cap.. just remove us */
1280 __ceph_remove_cap(cap
, true);
1282 /* try to drop referring dentries */
1283 spin_unlock(&ci
->i_ceph_lock
);
1284 d_prune_aliases(inode
);
1285 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1286 inode
, cap
, atomic_read(&inode
->i_count
));
1291 spin_unlock(&ci
->i_ceph_lock
);
1296 * Trim session cap count down to some max number.
1298 static int trim_caps(struct ceph_mds_client
*mdsc
,
1299 struct ceph_mds_session
*session
,
1302 int trim_caps
= session
->s_nr_caps
- max_caps
;
1304 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1305 session
->s_mds
, session
->s_nr_caps
, max_caps
, trim_caps
);
1306 if (trim_caps
> 0) {
1307 session
->s_trim_caps
= trim_caps
;
1308 iterate_session_caps(session
, trim_caps_cb
, session
);
1309 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1310 session
->s_mds
, session
->s_nr_caps
, max_caps
,
1311 trim_caps
- session
->s_trim_caps
);
1312 session
->s_trim_caps
= 0;
1315 ceph_add_cap_releases(mdsc
, session
);
1316 ceph_send_cap_releases(mdsc
, session
);
1321 * Allocate cap_release messages. If there is a partially full message
1322 * in the queue, try to allocate enough to cover it's remainder, so that
1323 * we can send it immediately.
1325 * Called under s_mutex.
1327 int ceph_add_cap_releases(struct ceph_mds_client
*mdsc
,
1328 struct ceph_mds_session
*session
)
1330 struct ceph_msg
*msg
, *partial
= NULL
;
1331 struct ceph_mds_cap_release
*head
;
1333 int extra
= mdsc
->fsc
->mount_options
->cap_release_safety
;
1336 dout("add_cap_releases %p mds%d extra %d\n", session
, session
->s_mds
,
1339 spin_lock(&session
->s_cap_lock
);
1341 if (!list_empty(&session
->s_cap_releases
)) {
1342 msg
= list_first_entry(&session
->s_cap_releases
,
1345 head
= msg
->front
.iov_base
;
1346 num
= le32_to_cpu(head
->num
);
1348 dout(" partial %p with (%d/%d)\n", msg
, num
,
1349 (int)CEPH_CAPS_PER_RELEASE
);
1350 extra
+= CEPH_CAPS_PER_RELEASE
- num
;
1354 while (session
->s_num_cap_releases
< session
->s_nr_caps
+ extra
) {
1355 spin_unlock(&session
->s_cap_lock
);
1356 msg
= ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE
, PAGE_CACHE_SIZE
,
1360 dout("add_cap_releases %p msg %p now %d\n", session
, msg
,
1361 (int)msg
->front
.iov_len
);
1362 head
= msg
->front
.iov_base
;
1363 head
->num
= cpu_to_le32(0);
1364 msg
->front
.iov_len
= sizeof(*head
);
1365 spin_lock(&session
->s_cap_lock
);
1366 list_add(&msg
->list_head
, &session
->s_cap_releases
);
1367 session
->s_num_cap_releases
+= CEPH_CAPS_PER_RELEASE
;
1371 head
= partial
->front
.iov_base
;
1372 num
= le32_to_cpu(head
->num
);
1373 dout(" queueing partial %p with %d/%d\n", partial
, num
,
1374 (int)CEPH_CAPS_PER_RELEASE
);
1375 list_move_tail(&partial
->list_head
,
1376 &session
->s_cap_releases_done
);
1377 session
->s_num_cap_releases
-= CEPH_CAPS_PER_RELEASE
- num
;
1380 spin_unlock(&session
->s_cap_lock
);
1386 * flush all dirty inode data to disk.
1388 * returns true if we've flushed through want_flush_seq
1390 static int check_cap_flush(struct ceph_mds_client
*mdsc
, u64 want_flush_seq
)
1394 dout("check_cap_flush want %lld\n", want_flush_seq
);
1395 mutex_lock(&mdsc
->mutex
);
1396 for (mds
= 0; ret
&& mds
< mdsc
->max_sessions
; mds
++) {
1397 struct ceph_mds_session
*session
= mdsc
->sessions
[mds
];
1401 get_session(session
);
1402 mutex_unlock(&mdsc
->mutex
);
1404 mutex_lock(&session
->s_mutex
);
1405 if (!list_empty(&session
->s_cap_flushing
)) {
1406 struct ceph_inode_info
*ci
=
1407 list_entry(session
->s_cap_flushing
.next
,
1408 struct ceph_inode_info
,
1410 struct inode
*inode
= &ci
->vfs_inode
;
1412 spin_lock(&ci
->i_ceph_lock
);
1413 if (ci
->i_cap_flush_seq
<= want_flush_seq
) {
1414 dout("check_cap_flush still flushing %p "
1415 "seq %lld <= %lld to mds%d\n", inode
,
1416 ci
->i_cap_flush_seq
, want_flush_seq
,
1420 spin_unlock(&ci
->i_ceph_lock
);
1422 mutex_unlock(&session
->s_mutex
);
1423 ceph_put_mds_session(session
);
1427 mutex_lock(&mdsc
->mutex
);
1430 mutex_unlock(&mdsc
->mutex
);
1431 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq
);
1436 * called under s_mutex
1438 void ceph_send_cap_releases(struct ceph_mds_client
*mdsc
,
1439 struct ceph_mds_session
*session
)
1441 struct ceph_msg
*msg
;
1443 dout("send_cap_releases mds%d\n", session
->s_mds
);
1444 spin_lock(&session
->s_cap_lock
);
1445 while (!list_empty(&session
->s_cap_releases_done
)) {
1446 msg
= list_first_entry(&session
->s_cap_releases_done
,
1447 struct ceph_msg
, list_head
);
1448 list_del_init(&msg
->list_head
);
1449 spin_unlock(&session
->s_cap_lock
);
1450 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1451 dout("send_cap_releases mds%d %p\n", session
->s_mds
, msg
);
1452 ceph_con_send(&session
->s_con
, msg
);
1453 spin_lock(&session
->s_cap_lock
);
1455 spin_unlock(&session
->s_cap_lock
);
1458 static void discard_cap_releases(struct ceph_mds_client
*mdsc
,
1459 struct ceph_mds_session
*session
)
1461 struct ceph_msg
*msg
;
1462 struct ceph_mds_cap_release
*head
;
1465 dout("discard_cap_releases mds%d\n", session
->s_mds
);
1467 if (!list_empty(&session
->s_cap_releases
)) {
1468 /* zero out the in-progress message */
1469 msg
= list_first_entry(&session
->s_cap_releases
,
1470 struct ceph_msg
, list_head
);
1471 head
= msg
->front
.iov_base
;
1472 num
= le32_to_cpu(head
->num
);
1473 dout("discard_cap_releases mds%d %p %u\n",
1474 session
->s_mds
, msg
, num
);
1475 head
->num
= cpu_to_le32(0);
1476 msg
->front
.iov_len
= sizeof(*head
);
1477 session
->s_num_cap_releases
+= num
;
1480 /* requeue completed messages */
1481 while (!list_empty(&session
->s_cap_releases_done
)) {
1482 msg
= list_first_entry(&session
->s_cap_releases_done
,
1483 struct ceph_msg
, list_head
);
1484 list_del_init(&msg
->list_head
);
1486 head
= msg
->front
.iov_base
;
1487 num
= le32_to_cpu(head
->num
);
1488 dout("discard_cap_releases mds%d %p %u\n", session
->s_mds
, msg
,
1490 session
->s_num_cap_releases
+= num
;
1491 head
->num
= cpu_to_le32(0);
1492 msg
->front
.iov_len
= sizeof(*head
);
1493 list_add(&msg
->list_head
, &session
->s_cap_releases
);
1501 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request
*req
,
1504 struct ceph_inode_info
*ci
= ceph_inode(dir
);
1505 struct ceph_mds_reply_info_parsed
*rinfo
= &req
->r_reply_info
;
1506 struct ceph_mount_options
*opt
= req
->r_mdsc
->fsc
->mount_options
;
1507 size_t size
= sizeof(*rinfo
->dir_in
) + sizeof(*rinfo
->dir_dname_len
) +
1508 sizeof(*rinfo
->dir_dname
) + sizeof(*rinfo
->dir_dlease
);
1509 int order
, num_entries
;
1511 spin_lock(&ci
->i_ceph_lock
);
1512 num_entries
= ci
->i_files
+ ci
->i_subdirs
;
1513 spin_unlock(&ci
->i_ceph_lock
);
1514 num_entries
= max(num_entries
, 1);
1515 num_entries
= min(num_entries
, opt
->max_readdir
);
1517 order
= get_order(size
* num_entries
);
1518 while (order
>= 0) {
1519 rinfo
->dir_in
= (void*)__get_free_pages(GFP_NOFS
| __GFP_NOWARN
,
1528 num_entries
= (PAGE_SIZE
<< order
) / size
;
1529 num_entries
= min(num_entries
, opt
->max_readdir
);
1531 rinfo
->dir_buf_size
= PAGE_SIZE
<< order
;
1532 req
->r_num_caps
= num_entries
+ 1;
1533 req
->r_args
.readdir
.max_entries
= cpu_to_le32(num_entries
);
1534 req
->r_args
.readdir
.max_bytes
= cpu_to_le32(opt
->max_readdir_bytes
);
1539 * Create an mds request.
1541 struct ceph_mds_request
*
1542 ceph_mdsc_create_request(struct ceph_mds_client
*mdsc
, int op
, int mode
)
1544 struct ceph_mds_request
*req
= kzalloc(sizeof(*req
), GFP_NOFS
);
1547 return ERR_PTR(-ENOMEM
);
1549 mutex_init(&req
->r_fill_mutex
);
1551 req
->r_started
= jiffies
;
1552 req
->r_resend_mds
= -1;
1553 INIT_LIST_HEAD(&req
->r_unsafe_dir_item
);
1555 kref_init(&req
->r_kref
);
1556 INIT_LIST_HEAD(&req
->r_wait
);
1557 init_completion(&req
->r_completion
);
1558 init_completion(&req
->r_safe_completion
);
1559 INIT_LIST_HEAD(&req
->r_unsafe_item
);
1561 req
->r_stamp
= CURRENT_TIME
;
1564 req
->r_direct_mode
= mode
;
1569 * return oldest (lowest) request, tid in request tree, 0 if none.
1571 * called under mdsc->mutex.
1573 static struct ceph_mds_request
*__get_oldest_req(struct ceph_mds_client
*mdsc
)
1575 if (RB_EMPTY_ROOT(&mdsc
->request_tree
))
1577 return rb_entry(rb_first(&mdsc
->request_tree
),
1578 struct ceph_mds_request
, r_node
);
1581 static u64
__get_oldest_tid(struct ceph_mds_client
*mdsc
)
1583 struct ceph_mds_request
*req
= __get_oldest_req(mdsc
);
1591 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1592 * on build_path_from_dentry in fs/cifs/dir.c.
1594 * If @stop_on_nosnap, generate path relative to the first non-snapped
1597 * Encode hidden .snap dirs as a double /, i.e.
1598 * foo/.snap/bar -> foo//bar
1600 char *ceph_mdsc_build_path(struct dentry
*dentry
, int *plen
, u64
*base
,
1603 struct dentry
*temp
;
1609 return ERR_PTR(-EINVAL
);
1613 seq
= read_seqbegin(&rename_lock
);
1615 for (temp
= dentry
; !IS_ROOT(temp
);) {
1616 struct inode
*inode
= temp
->d_inode
;
1617 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
)
1618 len
++; /* slash only */
1619 else if (stop_on_nosnap
&& inode
&&
1620 ceph_snap(inode
) == CEPH_NOSNAP
)
1623 len
+= 1 + temp
->d_name
.len
;
1624 temp
= temp
->d_parent
;
1628 len
--; /* no leading '/' */
1630 path
= kmalloc(len
+1, GFP_NOFS
);
1632 return ERR_PTR(-ENOMEM
);
1634 path
[pos
] = 0; /* trailing null */
1636 for (temp
= dentry
; !IS_ROOT(temp
) && pos
!= 0; ) {
1637 struct inode
*inode
;
1639 spin_lock(&temp
->d_lock
);
1640 inode
= temp
->d_inode
;
1641 if (inode
&& ceph_snap(inode
) == CEPH_SNAPDIR
) {
1642 dout("build_path path+%d: %p SNAPDIR\n",
1644 } else if (stop_on_nosnap
&& inode
&&
1645 ceph_snap(inode
) == CEPH_NOSNAP
) {
1646 spin_unlock(&temp
->d_lock
);
1649 pos
-= temp
->d_name
.len
;
1651 spin_unlock(&temp
->d_lock
);
1654 strncpy(path
+ pos
, temp
->d_name
.name
,
1657 spin_unlock(&temp
->d_lock
);
1660 temp
= temp
->d_parent
;
1663 if (pos
!= 0 || read_seqretry(&rename_lock
, seq
)) {
1664 pr_err("build_path did not end path lookup where "
1665 "expected, namelen is %d, pos is %d\n", len
, pos
);
1666 /* presumably this is only possible if racing with a
1667 rename of one of the parent directories (we can not
1668 lock the dentries above us to prevent this, but
1669 retrying should be harmless) */
1674 *base
= ceph_ino(temp
->d_inode
);
1676 dout("build_path on %p %d built %llx '%.*s'\n",
1677 dentry
, d_count(dentry
), *base
, len
, path
);
1681 static int build_dentry_path(struct dentry
*dentry
,
1682 const char **ppath
, int *ppathlen
, u64
*pino
,
1687 if (ceph_snap(dentry
->d_parent
->d_inode
) == CEPH_NOSNAP
) {
1688 *pino
= ceph_ino(dentry
->d_parent
->d_inode
);
1689 *ppath
= dentry
->d_name
.name
;
1690 *ppathlen
= dentry
->d_name
.len
;
1693 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1695 return PTR_ERR(path
);
1701 static int build_inode_path(struct inode
*inode
,
1702 const char **ppath
, int *ppathlen
, u64
*pino
,
1705 struct dentry
*dentry
;
1708 if (ceph_snap(inode
) == CEPH_NOSNAP
) {
1709 *pino
= ceph_ino(inode
);
1713 dentry
= d_find_alias(inode
);
1714 path
= ceph_mdsc_build_path(dentry
, ppathlen
, pino
, 1);
1717 return PTR_ERR(path
);
1724 * request arguments may be specified via an inode *, a dentry *, or
1725 * an explicit ino+path.
1727 static int set_request_path_attr(struct inode
*rinode
, struct dentry
*rdentry
,
1728 const char *rpath
, u64 rino
,
1729 const char **ppath
, int *pathlen
,
1730 u64
*ino
, int *freepath
)
1735 r
= build_inode_path(rinode
, ppath
, pathlen
, ino
, freepath
);
1736 dout(" inode %p %llx.%llx\n", rinode
, ceph_ino(rinode
),
1738 } else if (rdentry
) {
1739 r
= build_dentry_path(rdentry
, ppath
, pathlen
, ino
, freepath
);
1740 dout(" dentry %p %llx/%.*s\n", rdentry
, *ino
, *pathlen
,
1742 } else if (rpath
|| rino
) {
1745 *pathlen
= rpath
? strlen(rpath
) : 0;
1746 dout(" path %.*s\n", *pathlen
, rpath
);
1753 * called under mdsc->mutex
1755 static struct ceph_msg
*create_request_message(struct ceph_mds_client
*mdsc
,
1756 struct ceph_mds_request
*req
,
1759 struct ceph_msg
*msg
;
1760 struct ceph_mds_request_head
*head
;
1761 const char *path1
= NULL
;
1762 const char *path2
= NULL
;
1763 u64 ino1
= 0, ino2
= 0;
1764 int pathlen1
= 0, pathlen2
= 0;
1765 int freepath1
= 0, freepath2
= 0;
1771 ret
= set_request_path_attr(req
->r_inode
, req
->r_dentry
,
1772 req
->r_path1
, req
->r_ino1
.ino
,
1773 &path1
, &pathlen1
, &ino1
, &freepath1
);
1779 ret
= set_request_path_attr(NULL
, req
->r_old_dentry
,
1780 req
->r_path2
, req
->r_ino2
.ino
,
1781 &path2
, &pathlen2
, &ino2
, &freepath2
);
1787 len
= sizeof(*head
) +
1788 pathlen1
+ pathlen2
+ 2*(1 + sizeof(u32
) + sizeof(u64
)) +
1789 sizeof(struct timespec
);
1791 /* calculate (max) length for cap releases */
1792 len
+= sizeof(struct ceph_mds_request_release
) *
1793 (!!req
->r_inode_drop
+ !!req
->r_dentry_drop
+
1794 !!req
->r_old_inode_drop
+ !!req
->r_old_dentry_drop
);
1795 if (req
->r_dentry_drop
)
1796 len
+= req
->r_dentry
->d_name
.len
;
1797 if (req
->r_old_dentry_drop
)
1798 len
+= req
->r_old_dentry
->d_name
.len
;
1800 msg
= ceph_msg_new(CEPH_MSG_CLIENT_REQUEST
, len
, GFP_NOFS
, false);
1802 msg
= ERR_PTR(-ENOMEM
);
1806 msg
->hdr
.version
= 2;
1807 msg
->hdr
.tid
= cpu_to_le64(req
->r_tid
);
1809 head
= msg
->front
.iov_base
;
1810 p
= msg
->front
.iov_base
+ sizeof(*head
);
1811 end
= msg
->front
.iov_base
+ msg
->front
.iov_len
;
1813 head
->mdsmap_epoch
= cpu_to_le32(mdsc
->mdsmap
->m_epoch
);
1814 head
->op
= cpu_to_le32(req
->r_op
);
1815 head
->caller_uid
= cpu_to_le32(from_kuid(&init_user_ns
, req
->r_uid
));
1816 head
->caller_gid
= cpu_to_le32(from_kgid(&init_user_ns
, req
->r_gid
));
1817 head
->args
= req
->r_args
;
1819 ceph_encode_filepath(&p
, end
, ino1
, path1
);
1820 ceph_encode_filepath(&p
, end
, ino2
, path2
);
1822 /* make note of release offset, in case we need to replay */
1823 req
->r_request_release_offset
= p
- msg
->front
.iov_base
;
1827 if (req
->r_inode_drop
)
1828 releases
+= ceph_encode_inode_release(&p
,
1829 req
->r_inode
? req
->r_inode
: req
->r_dentry
->d_inode
,
1830 mds
, req
->r_inode_drop
, req
->r_inode_unless
, 0);
1831 if (req
->r_dentry_drop
)
1832 releases
+= ceph_encode_dentry_release(&p
, req
->r_dentry
,
1833 mds
, req
->r_dentry_drop
, req
->r_dentry_unless
);
1834 if (req
->r_old_dentry_drop
)
1835 releases
+= ceph_encode_dentry_release(&p
, req
->r_old_dentry
,
1836 mds
, req
->r_old_dentry_drop
, req
->r_old_dentry_unless
);
1837 if (req
->r_old_inode_drop
)
1838 releases
+= ceph_encode_inode_release(&p
,
1839 req
->r_old_dentry
->d_inode
,
1840 mds
, req
->r_old_inode_drop
, req
->r_old_inode_unless
, 0);
1841 head
->num_releases
= cpu_to_le16(releases
);
1844 ceph_encode_copy(&p
, &req
->r_stamp
, sizeof(req
->r_stamp
));
1847 msg
->front
.iov_len
= p
- msg
->front
.iov_base
;
1848 msg
->hdr
.front_len
= cpu_to_le32(msg
->front
.iov_len
);
1850 if (req
->r_data_len
) {
1851 /* outbound data set only by ceph_sync_setxattr() */
1852 BUG_ON(!req
->r_pages
);
1853 ceph_msg_data_add_pages(msg
, req
->r_pages
, req
->r_data_len
, 0);
1856 msg
->hdr
.data_len
= cpu_to_le32(req
->r_data_len
);
1857 msg
->hdr
.data_off
= cpu_to_le16(0);
1861 kfree((char *)path2
);
1864 kfree((char *)path1
);
1870 * called under mdsc->mutex if error, under no mutex if
1873 static void complete_request(struct ceph_mds_client
*mdsc
,
1874 struct ceph_mds_request
*req
)
1876 if (req
->r_callback
)
1877 req
->r_callback(mdsc
, req
);
1879 complete_all(&req
->r_completion
);
1883 * called under mdsc->mutex
1885 static int __prepare_send_request(struct ceph_mds_client
*mdsc
,
1886 struct ceph_mds_request
*req
,
1889 struct ceph_mds_request_head
*rhead
;
1890 struct ceph_msg
*msg
;
1895 struct ceph_cap
*cap
=
1896 ceph_get_cap_for_mds(ceph_inode(req
->r_inode
), mds
);
1899 req
->r_sent_on_mseq
= cap
->mseq
;
1901 req
->r_sent_on_mseq
= -1;
1903 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req
,
1904 req
->r_tid
, ceph_mds_op_name(req
->r_op
), req
->r_attempts
);
1906 if (req
->r_got_unsafe
) {
1908 * Replay. Do not regenerate message (and rebuild
1909 * paths, etc.); just use the original message.
1910 * Rebuilding paths will break for renames because
1911 * d_move mangles the src name.
1913 msg
= req
->r_request
;
1914 rhead
= msg
->front
.iov_base
;
1916 flags
= le32_to_cpu(rhead
->flags
);
1917 flags
|= CEPH_MDS_FLAG_REPLAY
;
1918 rhead
->flags
= cpu_to_le32(flags
);
1920 if (req
->r_target_inode
)
1921 rhead
->ino
= cpu_to_le64(ceph_ino(req
->r_target_inode
));
1923 rhead
->num_retry
= req
->r_attempts
- 1;
1925 /* remove cap/dentry releases from message */
1926 rhead
->num_releases
= 0;
1927 msg
->hdr
.front_len
= cpu_to_le32(req
->r_request_release_offset
);
1928 msg
->front
.iov_len
= req
->r_request_release_offset
;
1932 if (req
->r_request
) {
1933 ceph_msg_put(req
->r_request
);
1934 req
->r_request
= NULL
;
1936 msg
= create_request_message(mdsc
, req
, mds
);
1938 req
->r_err
= PTR_ERR(msg
);
1939 complete_request(mdsc
, req
);
1940 return PTR_ERR(msg
);
1942 req
->r_request
= msg
;
1944 rhead
= msg
->front
.iov_base
;
1945 rhead
->oldest_client_tid
= cpu_to_le64(__get_oldest_tid(mdsc
));
1946 if (req
->r_got_unsafe
)
1947 flags
|= CEPH_MDS_FLAG_REPLAY
;
1948 if (req
->r_locked_dir
)
1949 flags
|= CEPH_MDS_FLAG_WANT_DENTRY
;
1950 rhead
->flags
= cpu_to_le32(flags
);
1951 rhead
->num_fwd
= req
->r_num_fwd
;
1952 rhead
->num_retry
= req
->r_attempts
- 1;
1955 dout(" r_locked_dir = %p\n", req
->r_locked_dir
);
1960 * send request, or put it on the appropriate wait list.
1962 static int __do_request(struct ceph_mds_client
*mdsc
,
1963 struct ceph_mds_request
*req
)
1965 struct ceph_mds_session
*session
= NULL
;
1969 if (req
->r_err
|| req
->r_got_result
) {
1971 __unregister_request(mdsc
, req
);
1975 if (req
->r_timeout
&&
1976 time_after_eq(jiffies
, req
->r_started
+ req
->r_timeout
)) {
1977 dout("do_request timed out\n");
1982 put_request_session(req
);
1984 mds
= __choose_mds(mdsc
, req
);
1986 ceph_mdsmap_get_state(mdsc
->mdsmap
, mds
) < CEPH_MDS_STATE_ACTIVE
) {
1987 dout("do_request no mds or not active, waiting for map\n");
1988 list_add(&req
->r_wait
, &mdsc
->waiting_for_map
);
1992 /* get, open session */
1993 session
= __ceph_lookup_mds_session(mdsc
, mds
);
1995 session
= register_session(mdsc
, mds
);
1996 if (IS_ERR(session
)) {
1997 err
= PTR_ERR(session
);
2001 req
->r_session
= get_session(session
);
2003 dout("do_request mds%d session %p state %s\n", mds
, session
,
2004 session_state_name(session
->s_state
));
2005 if (session
->s_state
!= CEPH_MDS_SESSION_OPEN
&&
2006 session
->s_state
!= CEPH_MDS_SESSION_HUNG
) {
2007 if (session
->s_state
== CEPH_MDS_SESSION_NEW
||
2008 session
->s_state
== CEPH_MDS_SESSION_CLOSING
)
2009 __open_session(mdsc
, session
);
2010 list_add(&req
->r_wait
, &session
->s_waiting
);
2015 req
->r_resend_mds
= -1; /* forget any previous mds hint */
2017 if (req
->r_request_started
== 0) /* note request start time */
2018 req
->r_request_started
= jiffies
;
2020 err
= __prepare_send_request(mdsc
, req
, mds
);
2022 ceph_msg_get(req
->r_request
);
2023 ceph_con_send(&session
->s_con
, req
->r_request
);
2027 ceph_put_mds_session(session
);
2033 complete_request(mdsc
, req
);
2038 * called under mdsc->mutex
2040 static void __wake_requests(struct ceph_mds_client
*mdsc
,
2041 struct list_head
*head
)
2043 struct ceph_mds_request
*req
;
2044 LIST_HEAD(tmp_list
);
2046 list_splice_init(head
, &tmp_list
);
2048 while (!list_empty(&tmp_list
)) {
2049 req
= list_entry(tmp_list
.next
,
2050 struct ceph_mds_request
, r_wait
);
2051 list_del_init(&req
->r_wait
);
2052 dout(" wake request %p tid %llu\n", req
, req
->r_tid
);
2053 __do_request(mdsc
, req
);
2058 * Wake up threads with requests pending for @mds, so that they can
2059 * resubmit their requests to a possibly different mds.
2061 static void kick_requests(struct ceph_mds_client
*mdsc
, int mds
)
2063 struct ceph_mds_request
*req
;
2066 dout("kick_requests mds%d\n", mds
);
2067 for (p
= rb_first(&mdsc
->request_tree
); p
; p
= rb_next(p
)) {
2068 req
= rb_entry(p
, struct ceph_mds_request
, r_node
);
2069 if (req
->r_got_unsafe
)
2071 if (req
->r_session
&&
2072 req
->r_session
->s_mds
== mds
) {
2073 dout(" kicking tid %llu\n", req
->r_tid
);
2074 __do_request(mdsc
, req
);
2079 void ceph_mdsc_submit_request(struct ceph_mds_client
*mdsc
,
2080 struct ceph_mds_request
*req
)
2082 dout("submit_request on %p\n", req
);
2083 mutex_lock(&mdsc
->mutex
);
2084 __register_request(mdsc
, req
, NULL
);
2085 __do_request(mdsc
, req
);
2086 mutex_unlock(&mdsc
->mutex
);
2090 * Synchrously perform an mds request. Take care of all of the
2091 * session setup, forwarding, retry details.
2093 int ceph_mdsc_do_request(struct ceph_mds_client
*mdsc
,
2095 struct ceph_mds_request
*req
)
2099 dout("do_request on %p\n", req
);
2101 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2103 ceph_get_cap_refs(ceph_inode(req
->r_inode
), CEPH_CAP_PIN
);
2104 if (req
->r_locked_dir
)
2105 ceph_get_cap_refs(ceph_inode(req
->r_locked_dir
), CEPH_CAP_PIN
);
2106 if (req
->r_old_dentry_dir
)
2107 ceph_get_cap_refs(ceph_inode(req
->r_old_dentry_dir
),
2111 mutex_lock(&mdsc
->mutex
);
2112 __register_request(mdsc
, req
, dir
);
2113 __do_request(mdsc
, req
);
2117 __unregister_request(mdsc
, req
);
2118 dout("do_request early error %d\n", err
);
2123 mutex_unlock(&mdsc
->mutex
);
2124 dout("do_request waiting\n");
2125 if (req
->r_timeout
) {
2126 err
= (long)wait_for_completion_killable_timeout(
2127 &req
->r_completion
, req
->r_timeout
);
2131 err
= wait_for_completion_killable(&req
->r_completion
);
2133 dout("do_request waited, got %d\n", err
);
2134 mutex_lock(&mdsc
->mutex
);
2136 /* only abort if we didn't race with a real reply */
2137 if (req
->r_got_result
) {
2138 err
= le32_to_cpu(req
->r_reply_info
.head
->result
);
2139 } else if (err
< 0) {
2140 dout("aborted request %lld with %d\n", req
->r_tid
, err
);
2143 * ensure we aren't running concurrently with
2144 * ceph_fill_trace or ceph_readdir_prepopulate, which
2145 * rely on locks (dir mutex) held by our caller.
2147 mutex_lock(&req
->r_fill_mutex
);
2149 req
->r_aborted
= true;
2150 mutex_unlock(&req
->r_fill_mutex
);
2152 if (req
->r_locked_dir
&&
2153 (req
->r_op
& CEPH_MDS_OP_WRITE
))
2154 ceph_invalidate_dir_request(req
);
2160 mutex_unlock(&mdsc
->mutex
);
2161 dout("do_request %p done, result %d\n", req
, err
);
2166 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2167 * namespace request.
2169 void ceph_invalidate_dir_request(struct ceph_mds_request
*req
)
2171 struct inode
*inode
= req
->r_locked_dir
;
2173 dout("invalidate_dir_request %p (complete, lease(s))\n", inode
);
2175 ceph_dir_clear_complete(inode
);
2177 ceph_invalidate_dentry_lease(req
->r_dentry
);
2178 if (req
->r_old_dentry
)
2179 ceph_invalidate_dentry_lease(req
->r_old_dentry
);
2185 * We take the session mutex and parse and process the reply immediately.
2186 * This preserves the logical ordering of replies, capabilities, etc., sent
2187 * by the MDS as they are applied to our local cache.
2189 static void handle_reply(struct ceph_mds_session
*session
, struct ceph_msg
*msg
)
2191 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2192 struct ceph_mds_request
*req
;
2193 struct ceph_mds_reply_head
*head
= msg
->front
.iov_base
;
2194 struct ceph_mds_reply_info_parsed
*rinfo
; /* parsed reply info */
2197 int mds
= session
->s_mds
;
2199 if (msg
->front
.iov_len
< sizeof(*head
)) {
2200 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2205 /* get request, session */
2206 tid
= le64_to_cpu(msg
->hdr
.tid
);
2207 mutex_lock(&mdsc
->mutex
);
2208 req
= __lookup_request(mdsc
, tid
);
2210 dout("handle_reply on unknown tid %llu\n", tid
);
2211 mutex_unlock(&mdsc
->mutex
);
2214 dout("handle_reply %p\n", req
);
2216 /* correct session? */
2217 if (req
->r_session
!= session
) {
2218 pr_err("mdsc_handle_reply got %llu on session mds%d"
2219 " not mds%d\n", tid
, session
->s_mds
,
2220 req
->r_session
? req
->r_session
->s_mds
: -1);
2221 mutex_unlock(&mdsc
->mutex
);
2226 if ((req
->r_got_unsafe
&& !head
->safe
) ||
2227 (req
->r_got_safe
&& head
->safe
)) {
2228 pr_warn("got a dup %s reply on %llu from mds%d\n",
2229 head
->safe
? "safe" : "unsafe", tid
, mds
);
2230 mutex_unlock(&mdsc
->mutex
);
2233 if (req
->r_got_safe
&& !head
->safe
) {
2234 pr_warn("got unsafe after safe on %llu from mds%d\n",
2236 mutex_unlock(&mdsc
->mutex
);
2240 result
= le32_to_cpu(head
->result
);
2244 * if we're not talking to the authority, send to them
2245 * if the authority has changed while we weren't looking,
2246 * send to new authority
2247 * Otherwise we just have to return an ESTALE
2249 if (result
== -ESTALE
) {
2250 dout("got ESTALE on request %llu", req
->r_tid
);
2251 if (req
->r_direct_mode
!= USE_AUTH_MDS
) {
2252 dout("not using auth, setting for that now");
2253 req
->r_direct_mode
= USE_AUTH_MDS
;
2254 __do_request(mdsc
, req
);
2255 mutex_unlock(&mdsc
->mutex
);
2258 int mds
= __choose_mds(mdsc
, req
);
2259 if (mds
>= 0 && mds
!= req
->r_session
->s_mds
) {
2260 dout("but auth changed, so resending");
2261 __do_request(mdsc
, req
);
2262 mutex_unlock(&mdsc
->mutex
);
2266 dout("have to return ESTALE on request %llu", req
->r_tid
);
2271 req
->r_got_safe
= true;
2272 __unregister_request(mdsc
, req
);
2274 if (req
->r_got_unsafe
) {
2276 * We already handled the unsafe response, now do the
2277 * cleanup. No need to examine the response; the MDS
2278 * doesn't include any result info in the safe
2279 * response. And even if it did, there is nothing
2280 * useful we could do with a revised return value.
2282 dout("got safe reply %llu, mds%d\n", tid
, mds
);
2283 list_del_init(&req
->r_unsafe_item
);
2285 /* last unsafe request during umount? */
2286 if (mdsc
->stopping
&& !__get_oldest_req(mdsc
))
2287 complete_all(&mdsc
->safe_umount_waiters
);
2288 mutex_unlock(&mdsc
->mutex
);
2292 req
->r_got_unsafe
= true;
2293 list_add_tail(&req
->r_unsafe_item
, &req
->r_session
->s_unsafe
);
2296 dout("handle_reply tid %lld result %d\n", tid
, result
);
2297 rinfo
= &req
->r_reply_info
;
2298 err
= parse_reply_info(msg
, rinfo
, session
->s_con
.peer_features
);
2299 mutex_unlock(&mdsc
->mutex
);
2301 mutex_lock(&session
->s_mutex
);
2303 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds
, tid
);
2309 if (rinfo
->snapblob_len
) {
2310 down_write(&mdsc
->snap_rwsem
);
2311 ceph_update_snap_trace(mdsc
, rinfo
->snapblob
,
2312 rinfo
->snapblob
+ rinfo
->snapblob_len
,
2313 le32_to_cpu(head
->op
) == CEPH_MDS_OP_RMSNAP
);
2314 downgrade_write(&mdsc
->snap_rwsem
);
2316 down_read(&mdsc
->snap_rwsem
);
2319 /* insert trace into our cache */
2320 mutex_lock(&req
->r_fill_mutex
);
2321 err
= ceph_fill_trace(mdsc
->fsc
->sb
, req
, req
->r_session
);
2323 if (result
== 0 && (req
->r_op
== CEPH_MDS_OP_READDIR
||
2324 req
->r_op
== CEPH_MDS_OP_LSSNAP
))
2325 ceph_readdir_prepopulate(req
, req
->r_session
);
2326 ceph_unreserve_caps(mdsc
, &req
->r_caps_reservation
);
2328 mutex_unlock(&req
->r_fill_mutex
);
2330 up_read(&mdsc
->snap_rwsem
);
2332 mutex_lock(&mdsc
->mutex
);
2333 if (!req
->r_aborted
) {
2339 req
->r_got_result
= true;
2342 dout("reply arrived after request %lld was aborted\n", tid
);
2344 mutex_unlock(&mdsc
->mutex
);
2346 ceph_add_cap_releases(mdsc
, req
->r_session
);
2347 mutex_unlock(&session
->s_mutex
);
2349 /* kick calling process */
2350 complete_request(mdsc
, req
);
2352 ceph_mdsc_put_request(req
);
2359 * handle mds notification that our request has been forwarded.
2361 static void handle_forward(struct ceph_mds_client
*mdsc
,
2362 struct ceph_mds_session
*session
,
2363 struct ceph_msg
*msg
)
2365 struct ceph_mds_request
*req
;
2366 u64 tid
= le64_to_cpu(msg
->hdr
.tid
);
2370 void *p
= msg
->front
.iov_base
;
2371 void *end
= p
+ msg
->front
.iov_len
;
2373 ceph_decode_need(&p
, end
, 2*sizeof(u32
), bad
);
2374 next_mds
= ceph_decode_32(&p
);
2375 fwd_seq
= ceph_decode_32(&p
);
2377 mutex_lock(&mdsc
->mutex
);
2378 req
= __lookup_request(mdsc
, tid
);
2380 dout("forward tid %llu to mds%d - req dne\n", tid
, next_mds
);
2381 goto out
; /* dup reply? */
2384 if (req
->r_aborted
) {
2385 dout("forward tid %llu aborted, unregistering\n", tid
);
2386 __unregister_request(mdsc
, req
);
2387 } else if (fwd_seq
<= req
->r_num_fwd
) {
2388 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2389 tid
, next_mds
, req
->r_num_fwd
, fwd_seq
);
2391 /* resend. forward race not possible; mds would drop */
2392 dout("forward tid %llu to mds%d (we resend)\n", tid
, next_mds
);
2394 BUG_ON(req
->r_got_result
);
2395 req
->r_num_fwd
= fwd_seq
;
2396 req
->r_resend_mds
= next_mds
;
2397 put_request_session(req
);
2398 __do_request(mdsc
, req
);
2400 ceph_mdsc_put_request(req
);
2402 mutex_unlock(&mdsc
->mutex
);
2406 pr_err("mdsc_handle_forward decode error err=%d\n", err
);
2410 * handle a mds session control message
2412 static void handle_session(struct ceph_mds_session
*session
,
2413 struct ceph_msg
*msg
)
2415 struct ceph_mds_client
*mdsc
= session
->s_mdsc
;
2418 int mds
= session
->s_mds
;
2419 struct ceph_mds_session_head
*h
= msg
->front
.iov_base
;
2423 if (msg
->front
.iov_len
!= sizeof(*h
))
2425 op
= le32_to_cpu(h
->op
);
2426 seq
= le64_to_cpu(h
->seq
);
2428 mutex_lock(&mdsc
->mutex
);
2429 if (op
== CEPH_SESSION_CLOSE
)
2430 __unregister_session(mdsc
, session
);
2431 /* FIXME: this ttl calculation is generous */
2432 session
->s_ttl
= jiffies
+ HZ
*mdsc
->mdsmap
->m_session_autoclose
;
2433 mutex_unlock(&mdsc
->mutex
);
2435 mutex_lock(&session
->s_mutex
);
2437 dout("handle_session mds%d %s %p state %s seq %llu\n",
2438 mds
, ceph_session_op_name(op
), session
,
2439 session_state_name(session
->s_state
), seq
);
2441 if (session
->s_state
== CEPH_MDS_SESSION_HUNG
) {
2442 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2443 pr_info("mds%d came back\n", session
->s_mds
);
2447 case CEPH_SESSION_OPEN
:
2448 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2449 pr_info("mds%d reconnect success\n", session
->s_mds
);
2450 session
->s_state
= CEPH_MDS_SESSION_OPEN
;
2451 renewed_caps(mdsc
, session
, 0);
2454 __close_session(mdsc
, session
);
2457 case CEPH_SESSION_RENEWCAPS
:
2458 if (session
->s_renew_seq
== seq
)
2459 renewed_caps(mdsc
, session
, 1);
2462 case CEPH_SESSION_CLOSE
:
2463 if (session
->s_state
== CEPH_MDS_SESSION_RECONNECTING
)
2464 pr_info("mds%d reconnect denied\n", session
->s_mds
);
2465 remove_session_caps(session
);
2466 wake
= 1; /* for good measure */
2467 wake_up_all(&mdsc
->session_close_wq
);
2468 kick_requests(mdsc
, mds
);
2471 case CEPH_SESSION_STALE
:
2472 pr_info("mds%d caps went stale, renewing\n",
2474 spin_lock(&session
->s_gen_ttl_lock
);
2475 session
->s_cap_gen
++;
2476 session
->s_cap_ttl
= jiffies
- 1;
2477 spin_unlock(&session
->s_gen_ttl_lock
);
2478 send_renew_caps(mdsc
, session
);
2481 case CEPH_SESSION_RECALL_STATE
:
2482 trim_caps(mdsc
, session
, le32_to_cpu(h
->max_caps
));
2485 case CEPH_SESSION_FLUSHMSG
:
2486 send_flushmsg_ack(mdsc
, session
, seq
);
2490 pr_err("mdsc_handle_session bad op %d mds%d\n", op
, mds
);
2494 mutex_unlock(&session
->s_mutex
);
2496 mutex_lock(&mdsc
->mutex
);
2497 __wake_requests(mdsc
, &session
->s_waiting
);
2498 mutex_unlock(&mdsc
->mutex
);
2503 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds
,
2504 (int)msg
->front
.iov_len
);
2511 * called under session->mutex.
2513 static void replay_unsafe_requests(struct ceph_mds_client
*mdsc
,
2514 struct ceph_mds_session
*session
)
2516 struct ceph_mds_request
*req
, *nreq
;
2519 dout("replay_unsafe_requests mds%d\n", session
->s_mds
);
2521 mutex_lock(&mdsc
->mutex
);
2522 list_for_each_entry_safe(req
, nreq
, &session
->s_unsafe
, r_unsafe_item
) {
2523 err
= __prepare_send_request(mdsc
, req
, session
->s_mds
);
2525 ceph_msg_get(req
->r_request
);
2526 ceph_con_send(&session
->s_con
, req
->r_request
);
2529 mutex_unlock(&mdsc
->mutex
);
2533 * Encode information about a cap for a reconnect with the MDS.
2535 static int encode_caps_cb(struct inode
*inode
, struct ceph_cap
*cap
,
2539 struct ceph_mds_cap_reconnect v2
;
2540 struct ceph_mds_cap_reconnect_v1 v1
;
2543 struct ceph_inode_info
*ci
;
2544 struct ceph_reconnect_state
*recon_state
= arg
;
2545 struct ceph_pagelist
*pagelist
= recon_state
->pagelist
;
2549 struct dentry
*dentry
;
2553 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2554 inode
, ceph_vinop(inode
), cap
, cap
->cap_id
,
2555 ceph_cap_string(cap
->issued
));
2556 err
= ceph_pagelist_encode_64(pagelist
, ceph_ino(inode
));
2560 dentry
= d_find_alias(inode
);
2562 path
= ceph_mdsc_build_path(dentry
, &pathlen
, &pathbase
, 0);
2564 err
= PTR_ERR(path
);
2571 err
= ceph_pagelist_encode_string(pagelist
, path
, pathlen
);
2575 spin_lock(&ci
->i_ceph_lock
);
2576 cap
->seq
= 0; /* reset cap seq */
2577 cap
->issue_seq
= 0; /* and issue_seq */
2578 cap
->mseq
= 0; /* and migrate_seq */
2579 cap
->cap_gen
= cap
->session
->s_cap_gen
;
2581 if (recon_state
->flock
) {
2582 rec
.v2
.cap_id
= cpu_to_le64(cap
->cap_id
);
2583 rec
.v2
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2584 rec
.v2
.issued
= cpu_to_le32(cap
->issued
);
2585 rec
.v2
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2586 rec
.v2
.pathbase
= cpu_to_le64(pathbase
);
2587 rec
.v2
.flock_len
= 0;
2588 reclen
= sizeof(rec
.v2
);
2590 rec
.v1
.cap_id
= cpu_to_le64(cap
->cap_id
);
2591 rec
.v1
.wanted
= cpu_to_le32(__ceph_caps_wanted(ci
));
2592 rec
.v1
.issued
= cpu_to_le32(cap
->issued
);
2593 rec
.v1
.size
= cpu_to_le64(inode
->i_size
);
2594 ceph_encode_timespec(&rec
.v1
.mtime
, &inode
->i_mtime
);
2595 ceph_encode_timespec(&rec
.v1
.atime
, &inode
->i_atime
);
2596 rec
.v1
.snaprealm
= cpu_to_le64(ci
->i_snap_realm
->ino
);
2597 rec
.v1
.pathbase
= cpu_to_le64(pathbase
);
2598 reclen
= sizeof(rec
.v1
);
2600 spin_unlock(&ci
->i_ceph_lock
);
2602 if (recon_state
->flock
) {
2603 int num_fcntl_locks
, num_flock_locks
;
2604 struct ceph_filelock
*flocks
;
2607 spin_lock(&inode
->i_lock
);
2608 ceph_count_locks(inode
, &num_fcntl_locks
, &num_flock_locks
);
2609 spin_unlock(&inode
->i_lock
);
2610 flocks
= kmalloc((num_fcntl_locks
+num_flock_locks
) *
2611 sizeof(struct ceph_filelock
), GFP_NOFS
);
2616 spin_lock(&inode
->i_lock
);
2617 err
= ceph_encode_locks_to_buffer(inode
, flocks
,
2620 spin_unlock(&inode
->i_lock
);
2628 * number of encoded locks is stable, so copy to pagelist
2630 rec
.v2
.flock_len
= cpu_to_le32(2*sizeof(u32
) +
2631 (num_fcntl_locks
+num_flock_locks
) *
2632 sizeof(struct ceph_filelock
));
2633 err
= ceph_pagelist_append(pagelist
, &rec
, reclen
);
2635 err
= ceph_locks_to_pagelist(flocks
, pagelist
,
2640 err
= ceph_pagelist_append(pagelist
, &rec
, reclen
);
2643 recon_state
->nr_caps
++;
2653 * If an MDS fails and recovers, clients need to reconnect in order to
2654 * reestablish shared state. This includes all caps issued through
2655 * this session _and_ the snap_realm hierarchy. Because it's not
2656 * clear which snap realms the mds cares about, we send everything we
2657 * know about.. that ensures we'll then get any new info the
2658 * recovering MDS might have.
2660 * This is a relatively heavyweight operation, but it's rare.
2662 * called with mdsc->mutex held.
2664 static void send_mds_reconnect(struct ceph_mds_client
*mdsc
,
2665 struct ceph_mds_session
*session
)
2667 struct ceph_msg
*reply
;
2669 int mds
= session
->s_mds
;
2672 struct ceph_pagelist
*pagelist
;
2673 struct ceph_reconnect_state recon_state
;
2675 pr_info("mds%d reconnect start\n", mds
);
2677 pagelist
= kmalloc(sizeof(*pagelist
), GFP_NOFS
);
2679 goto fail_nopagelist
;
2680 ceph_pagelist_init(pagelist
);
2682 reply
= ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT
, 0, GFP_NOFS
, false);
2686 mutex_lock(&session
->s_mutex
);
2687 session
->s_state
= CEPH_MDS_SESSION_RECONNECTING
;
2690 ceph_con_close(&session
->s_con
);
2691 ceph_con_open(&session
->s_con
,
2692 CEPH_ENTITY_TYPE_MDS
, mds
,
2693 ceph_mdsmap_get_addr(mdsc
->mdsmap
, mds
));
2695 /* replay unsafe requests */
2696 replay_unsafe_requests(mdsc
, session
);
2698 down_read(&mdsc
->snap_rwsem
);
2700 dout("session %p state %s\n", session
,
2701 session_state_name(session
->s_state
));
2703 spin_lock(&session
->s_gen_ttl_lock
);
2704 session
->s_cap_gen
++;
2705 spin_unlock(&session
->s_gen_ttl_lock
);
2707 spin_lock(&session
->s_cap_lock
);
2709 * notify __ceph_remove_cap() that we are composing cap reconnect.
2710 * If a cap get released before being added to the cap reconnect,
2711 * __ceph_remove_cap() should skip queuing cap release.
2713 session
->s_cap_reconnect
= 1;
2714 /* drop old cap expires; we're about to reestablish that state */
2715 discard_cap_releases(mdsc
, session
);
2716 spin_unlock(&session
->s_cap_lock
);
2718 /* traverse this session's caps */
2719 s_nr_caps
= session
->s_nr_caps
;
2720 err
= ceph_pagelist_encode_32(pagelist
, s_nr_caps
);
2724 recon_state
.nr_caps
= 0;
2725 recon_state
.pagelist
= pagelist
;
2726 recon_state
.flock
= session
->s_con
.peer_features
& CEPH_FEATURE_FLOCK
;
2727 err
= iterate_session_caps(session
, encode_caps_cb
, &recon_state
);
2731 spin_lock(&session
->s_cap_lock
);
2732 session
->s_cap_reconnect
= 0;
2733 spin_unlock(&session
->s_cap_lock
);
2736 * snaprealms. we provide mds with the ino, seq (version), and
2737 * parent for all of our realms. If the mds has any newer info,
2740 for (p
= rb_first(&mdsc
->snap_realms
); p
; p
= rb_next(p
)) {
2741 struct ceph_snap_realm
*realm
=
2742 rb_entry(p
, struct ceph_snap_realm
, node
);
2743 struct ceph_mds_snaprealm_reconnect sr_rec
;
2745 dout(" adding snap realm %llx seq %lld parent %llx\n",
2746 realm
->ino
, realm
->seq
, realm
->parent_ino
);
2747 sr_rec
.ino
= cpu_to_le64(realm
->ino
);
2748 sr_rec
.seq
= cpu_to_le64(realm
->seq
);
2749 sr_rec
.parent
= cpu_to_le64(realm
->parent_ino
);
2750 err
= ceph_pagelist_append(pagelist
, &sr_rec
, sizeof(sr_rec
));
2755 if (recon_state
.flock
)
2756 reply
->hdr
.version
= cpu_to_le16(2);
2758 /* raced with cap release? */
2759 if (s_nr_caps
!= recon_state
.nr_caps
) {
2760 struct page
*page
= list_first_entry(&pagelist
->head
,
2762 __le32
*addr
= kmap_atomic(page
);
2763 *addr
= cpu_to_le32(recon_state
.nr_caps
);
2764 kunmap_atomic(addr
);
2767 reply
->hdr
.data_len
= cpu_to_le32(pagelist
->length
);
2768 ceph_msg_data_add_pagelist(reply
, pagelist
);
2769 ceph_con_send(&session
->s_con
, reply
);
2771 mutex_unlock(&session
->s_mutex
);
2773 mutex_lock(&mdsc
->mutex
);
2774 __wake_requests(mdsc
, &session
->s_waiting
);
2775 mutex_unlock(&mdsc
->mutex
);
2777 up_read(&mdsc
->snap_rwsem
);
2781 ceph_msg_put(reply
);
2782 up_read(&mdsc
->snap_rwsem
);
2783 mutex_unlock(&session
->s_mutex
);
2785 ceph_pagelist_release(pagelist
);
2788 pr_err("error %d preparing reconnect for mds%d\n", err
, mds
);
2794 * compare old and new mdsmaps, kicking requests
2795 * and closing out old connections as necessary
2797 * called under mdsc->mutex.
2799 static void check_new_map(struct ceph_mds_client
*mdsc
,
2800 struct ceph_mdsmap
*newmap
,
2801 struct ceph_mdsmap
*oldmap
)
2804 int oldstate
, newstate
;
2805 struct ceph_mds_session
*s
;
2807 dout("check_new_map new %u old %u\n",
2808 newmap
->m_epoch
, oldmap
->m_epoch
);
2810 for (i
= 0; i
< oldmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
2811 if (mdsc
->sessions
[i
] == NULL
)
2813 s
= mdsc
->sessions
[i
];
2814 oldstate
= ceph_mdsmap_get_state(oldmap
, i
);
2815 newstate
= ceph_mdsmap_get_state(newmap
, i
);
2817 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2818 i
, ceph_mds_state_name(oldstate
),
2819 ceph_mdsmap_is_laggy(oldmap
, i
) ? " (laggy)" : "",
2820 ceph_mds_state_name(newstate
),
2821 ceph_mdsmap_is_laggy(newmap
, i
) ? " (laggy)" : "",
2822 session_state_name(s
->s_state
));
2824 if (i
>= newmap
->m_max_mds
||
2825 memcmp(ceph_mdsmap_get_addr(oldmap
, i
),
2826 ceph_mdsmap_get_addr(newmap
, i
),
2827 sizeof(struct ceph_entity_addr
))) {
2828 if (s
->s_state
== CEPH_MDS_SESSION_OPENING
) {
2829 /* the session never opened, just close it
2831 __wake_requests(mdsc
, &s
->s_waiting
);
2832 __unregister_session(mdsc
, s
);
2835 mutex_unlock(&mdsc
->mutex
);
2836 mutex_lock(&s
->s_mutex
);
2837 mutex_lock(&mdsc
->mutex
);
2838 ceph_con_close(&s
->s_con
);
2839 mutex_unlock(&s
->s_mutex
);
2840 s
->s_state
= CEPH_MDS_SESSION_RESTARTING
;
2843 /* kick any requests waiting on the recovering mds */
2844 kick_requests(mdsc
, i
);
2845 } else if (oldstate
== newstate
) {
2846 continue; /* nothing new with this mds */
2852 if (s
->s_state
== CEPH_MDS_SESSION_RESTARTING
&&
2853 newstate
>= CEPH_MDS_STATE_RECONNECT
) {
2854 mutex_unlock(&mdsc
->mutex
);
2855 send_mds_reconnect(mdsc
, s
);
2856 mutex_lock(&mdsc
->mutex
);
2860 * kick request on any mds that has gone active.
2862 if (oldstate
< CEPH_MDS_STATE_ACTIVE
&&
2863 newstate
>= CEPH_MDS_STATE_ACTIVE
) {
2864 if (oldstate
!= CEPH_MDS_STATE_CREATING
&&
2865 oldstate
!= CEPH_MDS_STATE_STARTING
)
2866 pr_info("mds%d recovery completed\n", s
->s_mds
);
2867 kick_requests(mdsc
, i
);
2868 ceph_kick_flushing_caps(mdsc
, s
);
2869 wake_up_session_caps(s
, 1);
2873 for (i
= 0; i
< newmap
->m_max_mds
&& i
< mdsc
->max_sessions
; i
++) {
2874 s
= mdsc
->sessions
[i
];
2877 if (!ceph_mdsmap_is_laggy(newmap
, i
))
2879 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
2880 s
->s_state
== CEPH_MDS_SESSION_HUNG
||
2881 s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
2882 dout(" connecting to export targets of laggy mds%d\n",
2884 __open_export_target_sessions(mdsc
, s
);
2896 * caller must hold session s_mutex, dentry->d_lock
2898 void __ceph_mdsc_drop_dentry_lease(struct dentry
*dentry
)
2900 struct ceph_dentry_info
*di
= ceph_dentry(dentry
);
2902 ceph_put_mds_session(di
->lease_session
);
2903 di
->lease_session
= NULL
;
2906 static void handle_lease(struct ceph_mds_client
*mdsc
,
2907 struct ceph_mds_session
*session
,
2908 struct ceph_msg
*msg
)
2910 struct super_block
*sb
= mdsc
->fsc
->sb
;
2911 struct inode
*inode
;
2912 struct dentry
*parent
, *dentry
;
2913 struct ceph_dentry_info
*di
;
2914 int mds
= session
->s_mds
;
2915 struct ceph_mds_lease
*h
= msg
->front
.iov_base
;
2917 struct ceph_vino vino
;
2921 dout("handle_lease from mds%d\n", mds
);
2924 if (msg
->front
.iov_len
< sizeof(*h
) + sizeof(u32
))
2926 vino
.ino
= le64_to_cpu(h
->ino
);
2927 vino
.snap
= CEPH_NOSNAP
;
2928 seq
= le32_to_cpu(h
->seq
);
2929 dname
.name
= (void *)h
+ sizeof(*h
) + sizeof(u32
);
2930 dname
.len
= msg
->front
.iov_len
- sizeof(*h
) - sizeof(u32
);
2931 if (dname
.len
!= get_unaligned_le32(h
+1))
2934 mutex_lock(&session
->s_mutex
);
2938 inode
= ceph_find_inode(sb
, vino
);
2939 dout("handle_lease %s, ino %llx %p %.*s\n",
2940 ceph_lease_op_name(h
->action
), vino
.ino
, inode
,
2941 dname
.len
, dname
.name
);
2942 if (inode
== NULL
) {
2943 dout("handle_lease no inode %llx\n", vino
.ino
);
2948 parent
= d_find_alias(inode
);
2950 dout("no parent dentry on inode %p\n", inode
);
2952 goto release
; /* hrm... */
2954 dname
.hash
= full_name_hash(dname
.name
, dname
.len
);
2955 dentry
= d_lookup(parent
, &dname
);
2960 spin_lock(&dentry
->d_lock
);
2961 di
= ceph_dentry(dentry
);
2962 switch (h
->action
) {
2963 case CEPH_MDS_LEASE_REVOKE
:
2964 if (di
->lease_session
== session
) {
2965 if (ceph_seq_cmp(di
->lease_seq
, seq
) > 0)
2966 h
->seq
= cpu_to_le32(di
->lease_seq
);
2967 __ceph_mdsc_drop_dentry_lease(dentry
);
2972 case CEPH_MDS_LEASE_RENEW
:
2973 if (di
->lease_session
== session
&&
2974 di
->lease_gen
== session
->s_cap_gen
&&
2975 di
->lease_renew_from
&&
2976 di
->lease_renew_after
== 0) {
2977 unsigned long duration
=
2978 le32_to_cpu(h
->duration_ms
) * HZ
/ 1000;
2980 di
->lease_seq
= seq
;
2981 dentry
->d_time
= di
->lease_renew_from
+ duration
;
2982 di
->lease_renew_after
= di
->lease_renew_from
+
2984 di
->lease_renew_from
= 0;
2988 spin_unlock(&dentry
->d_lock
);
2995 /* let's just reuse the same message */
2996 h
->action
= CEPH_MDS_LEASE_REVOKE_ACK
;
2998 ceph_con_send(&session
->s_con
, msg
);
3002 mutex_unlock(&session
->s_mutex
);
3006 pr_err("corrupt lease message\n");
3010 void ceph_mdsc_lease_send_msg(struct ceph_mds_session
*session
,
3011 struct inode
*inode
,
3012 struct dentry
*dentry
, char action
,
3015 struct ceph_msg
*msg
;
3016 struct ceph_mds_lease
*lease
;
3017 int len
= sizeof(*lease
) + sizeof(u32
);
3020 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3021 inode
, dentry
, ceph_lease_op_name(action
), session
->s_mds
);
3022 dnamelen
= dentry
->d_name
.len
;
3025 msg
= ceph_msg_new(CEPH_MSG_CLIENT_LEASE
, len
, GFP_NOFS
, false);
3028 lease
= msg
->front
.iov_base
;
3029 lease
->action
= action
;
3030 lease
->ino
= cpu_to_le64(ceph_vino(inode
).ino
);
3031 lease
->first
= lease
->last
= cpu_to_le64(ceph_vino(inode
).snap
);
3032 lease
->seq
= cpu_to_le32(seq
);
3033 put_unaligned_le32(dnamelen
, lease
+ 1);
3034 memcpy((void *)(lease
+ 1) + 4, dentry
->d_name
.name
, dnamelen
);
3037 * if this is a preemptive lease RELEASE, no need to
3038 * flush request stream, since the actual request will
3041 msg
->more_to_follow
= (action
== CEPH_MDS_LEASE_RELEASE
);
3043 ceph_con_send(&session
->s_con
, msg
);
3047 * Preemptively release a lease we expect to invalidate anyway.
3048 * Pass @inode always, @dentry is optional.
3050 void ceph_mdsc_lease_release(struct ceph_mds_client
*mdsc
, struct inode
*inode
,
3051 struct dentry
*dentry
)
3053 struct ceph_dentry_info
*di
;
3054 struct ceph_mds_session
*session
;
3057 BUG_ON(inode
== NULL
);
3058 BUG_ON(dentry
== NULL
);
3060 /* is dentry lease valid? */
3061 spin_lock(&dentry
->d_lock
);
3062 di
= ceph_dentry(dentry
);
3063 if (!di
|| !di
->lease_session
||
3064 di
->lease_session
->s_mds
< 0 ||
3065 di
->lease_gen
!= di
->lease_session
->s_cap_gen
||
3066 !time_before(jiffies
, dentry
->d_time
)) {
3067 dout("lease_release inode %p dentry %p -- "
3070 spin_unlock(&dentry
->d_lock
);
3074 /* we do have a lease on this dentry; note mds and seq */
3075 session
= ceph_get_mds_session(di
->lease_session
);
3076 seq
= di
->lease_seq
;
3077 __ceph_mdsc_drop_dentry_lease(dentry
);
3078 spin_unlock(&dentry
->d_lock
);
3080 dout("lease_release inode %p dentry %p to mds%d\n",
3081 inode
, dentry
, session
->s_mds
);
3082 ceph_mdsc_lease_send_msg(session
, inode
, dentry
,
3083 CEPH_MDS_LEASE_RELEASE
, seq
);
3084 ceph_put_mds_session(session
);
3088 * drop all leases (and dentry refs) in preparation for umount
3090 static void drop_leases(struct ceph_mds_client
*mdsc
)
3094 dout("drop_leases\n");
3095 mutex_lock(&mdsc
->mutex
);
3096 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3097 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3100 mutex_unlock(&mdsc
->mutex
);
3101 mutex_lock(&s
->s_mutex
);
3102 mutex_unlock(&s
->s_mutex
);
3103 ceph_put_mds_session(s
);
3104 mutex_lock(&mdsc
->mutex
);
3106 mutex_unlock(&mdsc
->mutex
);
3112 * delayed work -- periodically trim expired leases, renew caps with mds
3114 static void schedule_delayed(struct ceph_mds_client
*mdsc
)
3117 unsigned hz
= round_jiffies_relative(HZ
* delay
);
3118 schedule_delayed_work(&mdsc
->delayed_work
, hz
);
3121 static void delayed_work(struct work_struct
*work
)
3124 struct ceph_mds_client
*mdsc
=
3125 container_of(work
, struct ceph_mds_client
, delayed_work
.work
);
3129 dout("mdsc delayed_work\n");
3130 ceph_check_delayed_caps(mdsc
);
3132 mutex_lock(&mdsc
->mutex
);
3133 renew_interval
= mdsc
->mdsmap
->m_session_timeout
>> 2;
3134 renew_caps
= time_after_eq(jiffies
, HZ
*renew_interval
+
3135 mdsc
->last_renew_caps
);
3137 mdsc
->last_renew_caps
= jiffies
;
3139 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3140 struct ceph_mds_session
*s
= __ceph_lookup_mds_session(mdsc
, i
);
3143 if (s
->s_state
== CEPH_MDS_SESSION_CLOSING
) {
3144 dout("resending session close request for mds%d\n",
3146 request_close_session(mdsc
, s
);
3147 ceph_put_mds_session(s
);
3150 if (s
->s_ttl
&& time_after(jiffies
, s
->s_ttl
)) {
3151 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
) {
3152 s
->s_state
= CEPH_MDS_SESSION_HUNG
;
3153 pr_info("mds%d hung\n", s
->s_mds
);
3156 if (s
->s_state
< CEPH_MDS_SESSION_OPEN
) {
3157 /* this mds is failed or recovering, just wait */
3158 ceph_put_mds_session(s
);
3161 mutex_unlock(&mdsc
->mutex
);
3163 mutex_lock(&s
->s_mutex
);
3165 send_renew_caps(mdsc
, s
);
3167 ceph_con_keepalive(&s
->s_con
);
3168 ceph_add_cap_releases(mdsc
, s
);
3169 if (s
->s_state
== CEPH_MDS_SESSION_OPEN
||
3170 s
->s_state
== CEPH_MDS_SESSION_HUNG
)
3171 ceph_send_cap_releases(mdsc
, s
);
3172 mutex_unlock(&s
->s_mutex
);
3173 ceph_put_mds_session(s
);
3175 mutex_lock(&mdsc
->mutex
);
3177 mutex_unlock(&mdsc
->mutex
);
3179 schedule_delayed(mdsc
);
3182 int ceph_mdsc_init(struct ceph_fs_client
*fsc
)
3185 struct ceph_mds_client
*mdsc
;
3187 mdsc
= kzalloc(sizeof(struct ceph_mds_client
), GFP_NOFS
);
3192 mutex_init(&mdsc
->mutex
);
3193 mdsc
->mdsmap
= kzalloc(sizeof(*mdsc
->mdsmap
), GFP_NOFS
);
3194 if (mdsc
->mdsmap
== NULL
) {
3199 init_completion(&mdsc
->safe_umount_waiters
);
3200 init_waitqueue_head(&mdsc
->session_close_wq
);
3201 INIT_LIST_HEAD(&mdsc
->waiting_for_map
);
3202 mdsc
->sessions
= NULL
;
3203 mdsc
->max_sessions
= 0;
3205 init_rwsem(&mdsc
->snap_rwsem
);
3206 mdsc
->snap_realms
= RB_ROOT
;
3207 INIT_LIST_HEAD(&mdsc
->snap_empty
);
3208 spin_lock_init(&mdsc
->snap_empty_lock
);
3210 mdsc
->request_tree
= RB_ROOT
;
3211 INIT_DELAYED_WORK(&mdsc
->delayed_work
, delayed_work
);
3212 mdsc
->last_renew_caps
= jiffies
;
3213 INIT_LIST_HEAD(&mdsc
->cap_delay_list
);
3214 spin_lock_init(&mdsc
->cap_delay_lock
);
3215 INIT_LIST_HEAD(&mdsc
->snap_flush_list
);
3216 spin_lock_init(&mdsc
->snap_flush_lock
);
3217 mdsc
->cap_flush_seq
= 0;
3218 INIT_LIST_HEAD(&mdsc
->cap_dirty
);
3219 INIT_LIST_HEAD(&mdsc
->cap_dirty_migrating
);
3220 mdsc
->num_cap_flushing
= 0;
3221 spin_lock_init(&mdsc
->cap_dirty_lock
);
3222 init_waitqueue_head(&mdsc
->cap_flushing_wq
);
3223 spin_lock_init(&mdsc
->dentry_lru_lock
);
3224 INIT_LIST_HEAD(&mdsc
->dentry_lru
);
3226 ceph_caps_init(mdsc
);
3227 ceph_adjust_min_caps(mdsc
, fsc
->min_caps
);
3233 * Wait for safe replies on open mds requests. If we time out, drop
3234 * all requests from the tree to avoid dangling dentry refs.
3236 static void wait_requests(struct ceph_mds_client
*mdsc
)
3238 struct ceph_mds_request
*req
;
3239 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3241 mutex_lock(&mdsc
->mutex
);
3242 if (__get_oldest_req(mdsc
)) {
3243 mutex_unlock(&mdsc
->mutex
);
3245 dout("wait_requests waiting for requests\n");
3246 wait_for_completion_timeout(&mdsc
->safe_umount_waiters
,
3247 fsc
->client
->options
->mount_timeout
* HZ
);
3249 /* tear down remaining requests */
3250 mutex_lock(&mdsc
->mutex
);
3251 while ((req
= __get_oldest_req(mdsc
))) {
3252 dout("wait_requests timed out on tid %llu\n",
3254 __unregister_request(mdsc
, req
);
3257 mutex_unlock(&mdsc
->mutex
);
3258 dout("wait_requests done\n");
3262 * called before mount is ro, and before dentries are torn down.
3263 * (hmm, does this still race with new lookups?)
3265 void ceph_mdsc_pre_umount(struct ceph_mds_client
*mdsc
)
3267 dout("pre_umount\n");
3271 ceph_flush_dirty_caps(mdsc
);
3272 wait_requests(mdsc
);
3275 * wait for reply handlers to drop their request refs and
3276 * their inode/dcache refs
3282 * wait for all write mds requests to flush.
3284 static void wait_unsafe_requests(struct ceph_mds_client
*mdsc
, u64 want_tid
)
3286 struct ceph_mds_request
*req
= NULL
, *nextreq
;
3289 mutex_lock(&mdsc
->mutex
);
3290 dout("wait_unsafe_requests want %lld\n", want_tid
);
3292 req
= __get_oldest_req(mdsc
);
3293 while (req
&& req
->r_tid
<= want_tid
) {
3294 /* find next request */
3295 n
= rb_next(&req
->r_node
);
3297 nextreq
= rb_entry(n
, struct ceph_mds_request
, r_node
);
3300 if ((req
->r_op
& CEPH_MDS_OP_WRITE
)) {
3302 ceph_mdsc_get_request(req
);
3304 ceph_mdsc_get_request(nextreq
);
3305 mutex_unlock(&mdsc
->mutex
);
3306 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3307 req
->r_tid
, want_tid
);
3308 wait_for_completion(&req
->r_safe_completion
);
3309 mutex_lock(&mdsc
->mutex
);
3310 ceph_mdsc_put_request(req
);
3312 break; /* next dne before, so we're done! */
3313 if (RB_EMPTY_NODE(&nextreq
->r_node
)) {
3314 /* next request was removed from tree */
3315 ceph_mdsc_put_request(nextreq
);
3318 ceph_mdsc_put_request(nextreq
); /* won't go away */
3322 mutex_unlock(&mdsc
->mutex
);
3323 dout("wait_unsafe_requests done\n");
3326 void ceph_mdsc_sync(struct ceph_mds_client
*mdsc
)
3328 u64 want_tid
, want_flush
;
3330 if (mdsc
->fsc
->mount_state
== CEPH_MOUNT_SHUTDOWN
)
3334 mutex_lock(&mdsc
->mutex
);
3335 want_tid
= mdsc
->last_tid
;
3336 want_flush
= mdsc
->cap_flush_seq
;
3337 mutex_unlock(&mdsc
->mutex
);
3338 dout("sync want tid %lld flush_seq %lld\n", want_tid
, want_flush
);
3340 ceph_flush_dirty_caps(mdsc
);
3342 wait_unsafe_requests(mdsc
, want_tid
);
3343 wait_event(mdsc
->cap_flushing_wq
, check_cap_flush(mdsc
, want_flush
));
3347 * true if all sessions are closed, or we force unmount
3349 static bool done_closing_sessions(struct ceph_mds_client
*mdsc
)
3353 if (mdsc
->fsc
->mount_state
== CEPH_MOUNT_SHUTDOWN
)
3356 mutex_lock(&mdsc
->mutex
);
3357 for (i
= 0; i
< mdsc
->max_sessions
; i
++)
3358 if (mdsc
->sessions
[i
])
3360 mutex_unlock(&mdsc
->mutex
);
3365 * called after sb is ro.
3367 void ceph_mdsc_close_sessions(struct ceph_mds_client
*mdsc
)
3369 struct ceph_mds_session
*session
;
3371 struct ceph_fs_client
*fsc
= mdsc
->fsc
;
3372 unsigned long timeout
= fsc
->client
->options
->mount_timeout
* HZ
;
3374 dout("close_sessions\n");
3376 /* close sessions */
3377 mutex_lock(&mdsc
->mutex
);
3378 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3379 session
= __ceph_lookup_mds_session(mdsc
, i
);
3382 mutex_unlock(&mdsc
->mutex
);
3383 mutex_lock(&session
->s_mutex
);
3384 __close_session(mdsc
, session
);
3385 mutex_unlock(&session
->s_mutex
);
3386 ceph_put_mds_session(session
);
3387 mutex_lock(&mdsc
->mutex
);
3389 mutex_unlock(&mdsc
->mutex
);
3391 dout("waiting for sessions to close\n");
3392 wait_event_timeout(mdsc
->session_close_wq
, done_closing_sessions(mdsc
),
3395 /* tear down remaining sessions */
3396 mutex_lock(&mdsc
->mutex
);
3397 for (i
= 0; i
< mdsc
->max_sessions
; i
++) {
3398 if (mdsc
->sessions
[i
]) {
3399 session
= get_session(mdsc
->sessions
[i
]);
3400 __unregister_session(mdsc
, session
);
3401 mutex_unlock(&mdsc
->mutex
);
3402 mutex_lock(&session
->s_mutex
);
3403 remove_session_caps(session
);
3404 mutex_unlock(&session
->s_mutex
);
3405 ceph_put_mds_session(session
);
3406 mutex_lock(&mdsc
->mutex
);
3409 WARN_ON(!list_empty(&mdsc
->cap_delay_list
));
3410 mutex_unlock(&mdsc
->mutex
);
3412 ceph_cleanup_empty_realms(mdsc
);
3414 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3419 static void ceph_mdsc_stop(struct ceph_mds_client
*mdsc
)
3422 cancel_delayed_work_sync(&mdsc
->delayed_work
); /* cancel timer */
3424 ceph_mdsmap_destroy(mdsc
->mdsmap
);
3425 kfree(mdsc
->sessions
);
3426 ceph_caps_finalize(mdsc
);
3429 void ceph_mdsc_destroy(struct ceph_fs_client
*fsc
)
3431 struct ceph_mds_client
*mdsc
= fsc
->mdsc
;
3433 dout("mdsc_destroy %p\n", mdsc
);
3434 ceph_mdsc_stop(mdsc
);
3436 /* flush out any connection work with references to us */
3441 dout("mdsc_destroy %p done\n", mdsc
);
3446 * handle mds map update.
3448 void ceph_mdsc_handle_map(struct ceph_mds_client
*mdsc
, struct ceph_msg
*msg
)
3452 void *p
= msg
->front
.iov_base
;
3453 void *end
= p
+ msg
->front
.iov_len
;
3454 struct ceph_mdsmap
*newmap
, *oldmap
;
3455 struct ceph_fsid fsid
;
3458 ceph_decode_need(&p
, end
, sizeof(fsid
)+2*sizeof(u32
), bad
);
3459 ceph_decode_copy(&p
, &fsid
, sizeof(fsid
));
3460 if (ceph_check_fsid(mdsc
->fsc
->client
, &fsid
) < 0)
3462 epoch
= ceph_decode_32(&p
);
3463 maplen
= ceph_decode_32(&p
);
3464 dout("handle_map epoch %u len %d\n", epoch
, (int)maplen
);
3466 /* do we need it? */
3467 ceph_monc_got_mdsmap(&mdsc
->fsc
->client
->monc
, epoch
);
3468 mutex_lock(&mdsc
->mutex
);
3469 if (mdsc
->mdsmap
&& epoch
<= mdsc
->mdsmap
->m_epoch
) {
3470 dout("handle_map epoch %u <= our %u\n",
3471 epoch
, mdsc
->mdsmap
->m_epoch
);
3472 mutex_unlock(&mdsc
->mutex
);
3476 newmap
= ceph_mdsmap_decode(&p
, end
);
3477 if (IS_ERR(newmap
)) {
3478 err
= PTR_ERR(newmap
);
3482 /* swap into place */
3484 oldmap
= mdsc
->mdsmap
;
3485 mdsc
->mdsmap
= newmap
;
3486 check_new_map(mdsc
, newmap
, oldmap
);
3487 ceph_mdsmap_destroy(oldmap
);
3489 mdsc
->mdsmap
= newmap
; /* first mds map */
3491 mdsc
->fsc
->sb
->s_maxbytes
= mdsc
->mdsmap
->m_max_file_size
;
3493 __wake_requests(mdsc
, &mdsc
->waiting_for_map
);
3495 mutex_unlock(&mdsc
->mutex
);
3496 schedule_delayed(mdsc
);
3500 mutex_unlock(&mdsc
->mutex
);
3502 pr_err("error decoding mdsmap %d\n", err
);
3506 static struct ceph_connection
*con_get(struct ceph_connection
*con
)
3508 struct ceph_mds_session
*s
= con
->private;
3510 if (get_session(s
)) {
3511 dout("mdsc con_get %p ok (%d)\n", s
, atomic_read(&s
->s_ref
));
3514 dout("mdsc con_get %p FAIL\n", s
);
3518 static void con_put(struct ceph_connection
*con
)
3520 struct ceph_mds_session
*s
= con
->private;
3522 dout("mdsc con_put %p (%d)\n", s
, atomic_read(&s
->s_ref
) - 1);
3523 ceph_put_mds_session(s
);
3527 * if the client is unresponsive for long enough, the mds will kill
3528 * the session entirely.
3530 static void peer_reset(struct ceph_connection
*con
)
3532 struct ceph_mds_session
*s
= con
->private;
3533 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3535 pr_warn("mds%d closed our session\n", s
->s_mds
);
3536 send_mds_reconnect(mdsc
, s
);
3539 static void dispatch(struct ceph_connection
*con
, struct ceph_msg
*msg
)
3541 struct ceph_mds_session
*s
= con
->private;
3542 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3543 int type
= le16_to_cpu(msg
->hdr
.type
);
3545 mutex_lock(&mdsc
->mutex
);
3546 if (__verify_registered_session(mdsc
, s
) < 0) {
3547 mutex_unlock(&mdsc
->mutex
);
3550 mutex_unlock(&mdsc
->mutex
);
3553 case CEPH_MSG_MDS_MAP
:
3554 ceph_mdsc_handle_map(mdsc
, msg
);
3556 case CEPH_MSG_CLIENT_SESSION
:
3557 handle_session(s
, msg
);
3559 case CEPH_MSG_CLIENT_REPLY
:
3560 handle_reply(s
, msg
);
3562 case CEPH_MSG_CLIENT_REQUEST_FORWARD
:
3563 handle_forward(mdsc
, s
, msg
);
3565 case CEPH_MSG_CLIENT_CAPS
:
3566 ceph_handle_caps(s
, msg
);
3568 case CEPH_MSG_CLIENT_SNAP
:
3569 ceph_handle_snap(mdsc
, s
, msg
);
3571 case CEPH_MSG_CLIENT_LEASE
:
3572 handle_lease(mdsc
, s
, msg
);
3576 pr_err("received unknown message type %d %s\n", type
,
3577 ceph_msg_type_name(type
));
3588 * Note: returned pointer is the address of a structure that's
3589 * managed separately. Caller must *not* attempt to free it.
3591 static struct ceph_auth_handshake
*get_authorizer(struct ceph_connection
*con
,
3592 int *proto
, int force_new
)
3594 struct ceph_mds_session
*s
= con
->private;
3595 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3596 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3597 struct ceph_auth_handshake
*auth
= &s
->s_auth
;
3599 if (force_new
&& auth
->authorizer
) {
3600 ceph_auth_destroy_authorizer(ac
, auth
->authorizer
);
3601 auth
->authorizer
= NULL
;
3603 if (!auth
->authorizer
) {
3604 int ret
= ceph_auth_create_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3607 return ERR_PTR(ret
);
3609 int ret
= ceph_auth_update_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
,
3612 return ERR_PTR(ret
);
3614 *proto
= ac
->protocol
;
3620 static int verify_authorizer_reply(struct ceph_connection
*con
, int len
)
3622 struct ceph_mds_session
*s
= con
->private;
3623 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3624 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3626 return ceph_auth_verify_authorizer_reply(ac
, s
->s_auth
.authorizer
, len
);
3629 static int invalidate_authorizer(struct ceph_connection
*con
)
3631 struct ceph_mds_session
*s
= con
->private;
3632 struct ceph_mds_client
*mdsc
= s
->s_mdsc
;
3633 struct ceph_auth_client
*ac
= mdsc
->fsc
->client
->monc
.auth
;
3635 ceph_auth_invalidate_authorizer(ac
, CEPH_ENTITY_TYPE_MDS
);
3637 return ceph_monc_validate_auth(&mdsc
->fsc
->client
->monc
);
3640 static struct ceph_msg
*mds_alloc_msg(struct ceph_connection
*con
,
3641 struct ceph_msg_header
*hdr
, int *skip
)
3643 struct ceph_msg
*msg
;
3644 int type
= (int) le16_to_cpu(hdr
->type
);
3645 int front_len
= (int) le32_to_cpu(hdr
->front_len
);
3651 msg
= ceph_msg_new(type
, front_len
, GFP_NOFS
, false);
3653 pr_err("unable to allocate msg type %d len %d\n",
3661 static const struct ceph_connection_operations mds_con_ops
= {
3664 .dispatch
= dispatch
,
3665 .get_authorizer
= get_authorizer
,
3666 .verify_authorizer_reply
= verify_authorizer_reply
,
3667 .invalidate_authorizer
= invalidate_authorizer
,
3668 .peer_reset
= peer_reset
,
3669 .alloc_msg
= mds_alloc_msg
,