1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
6 #include <linux/sched.h>
7 #include <linux/debugfs.h>
8 #include <linux/seq_file.h>
11 #include "mds_client.h"
13 #include <linux/ceph/ceph_features.h>
14 #include <linux/ceph/messenger.h>
15 #include <linux/ceph/decode.h>
16 #include <linux/ceph/pagelist.h>
17 #include <linux/ceph/auth.h>
18 #include <linux/ceph/debugfs.h>
21 * A cluster of MDS (metadata server) daemons is responsible for
22 * managing the file system namespace (the directory hierarchy and
23 * inodes) and for coordinating shared access to storage. Metadata is
24 * partitioning hierarchically across a number of servers, and that
25 * partition varies over time as the cluster adjusts the distribution
26 * in order to balance load.
28 * The MDS client is primarily responsible to managing synchronous
29 * metadata requests for operations like open, unlink, and so forth.
30 * If there is a MDS failure, we find out about it when we (possibly
31 * request and) receive a new MDS map, and can resubmit affected
34 * For the most part, though, we take advantage of a lossless
35 * communications channel to the MDS, and do not need to worry about
36 * timing out or resubmitting requests.
38 * We maintain a stateful "session" with each MDS we interact with.
39 * Within each session, we sent periodic heartbeat messages to ensure
40 * any capabilities or leases we have been issues remain valid. If
41 * the session times out and goes stale, our leases and capabilities
42 * are no longer valid.
45 struct ceph_reconnect_state {
47 struct ceph_pagelist *pagelist;
51 static void __wake_requests(struct ceph_mds_client *mdsc,
52 struct list_head *head);
54 static const struct ceph_connection_operations mds_con_ops;
62 * parse individual inode info
64 static int parse_reply_info_in(void **p, void *end,
65 struct ceph_mds_reply_info_in *info,
71 *p += sizeof(struct ceph_mds_reply_inode) +
72 sizeof(*info->in->fragtree.splits) *
73 le32_to_cpu(info->in->fragtree.nsplits);
75 ceph_decode_32_safe(p, end, info->symlink_len, bad);
76 ceph_decode_need(p, end, info->symlink_len, bad);
78 *p += info->symlink_len;
80 if (features & CEPH_FEATURE_DIRLAYOUTHASH)
81 ceph_decode_copy_safe(p, end, &info->dir_layout,
82 sizeof(info->dir_layout), bad);
84 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
86 ceph_decode_32_safe(p, end, info->xattr_len, bad);
87 ceph_decode_need(p, end, info->xattr_len, bad);
88 info->xattr_data = *p;
89 *p += info->xattr_len;
96 * parse a normal reply, which may contain a (dir+)dentry and/or a
99 static int parse_reply_info_trace(void **p, void *end,
100 struct ceph_mds_reply_info_parsed *info,
105 if (info->head->is_dentry) {
106 err = parse_reply_info_in(p, end, &info->diri, features);
110 if (unlikely(*p + sizeof(*info->dirfrag) > end))
113 *p += sizeof(*info->dirfrag) +
114 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
115 if (unlikely(*p > end))
118 ceph_decode_32_safe(p, end, info->dname_len, bad);
119 ceph_decode_need(p, end, info->dname_len, bad);
121 *p += info->dname_len;
123 *p += sizeof(*info->dlease);
126 if (info->head->is_target) {
127 err = parse_reply_info_in(p, end, &info->targeti, features);
132 if (unlikely(*p != end))
139 pr_err("problem parsing mds trace %d\n", err);
144 * parse readdir results
146 static int parse_reply_info_dir(void **p, void *end,
147 struct ceph_mds_reply_info_parsed *info,
154 if (*p + sizeof(*info->dir_dir) > end)
156 *p += sizeof(*info->dir_dir) +
157 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
161 ceph_decode_need(p, end, sizeof(num) + 2, bad);
162 num = ceph_decode_32(p);
163 info->dir_end = ceph_decode_8(p);
164 info->dir_complete = ceph_decode_8(p);
168 /* alloc large array */
170 info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
171 sizeof(*info->dir_dname) +
172 sizeof(*info->dir_dname_len) +
173 sizeof(*info->dir_dlease),
175 if (info->dir_in == NULL) {
179 info->dir_dname = (void *)(info->dir_in + num);
180 info->dir_dname_len = (void *)(info->dir_dname + num);
181 info->dir_dlease = (void *)(info->dir_dname_len + num);
185 ceph_decode_need(p, end, sizeof(u32)*2, bad);
186 info->dir_dname_len[i] = ceph_decode_32(p);
187 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
188 info->dir_dname[i] = *p;
189 *p += info->dir_dname_len[i];
190 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
192 info->dir_dlease[i] = *p;
193 *p += sizeof(struct ceph_mds_reply_lease);
196 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
211 pr_err("problem parsing dir contents %d\n", err);
216 * parse fcntl F_GETLK results
218 static int parse_reply_info_filelock(void **p, void *end,
219 struct ceph_mds_reply_info_parsed *info,
222 if (*p + sizeof(*info->filelock_reply) > end)
225 info->filelock_reply = *p;
226 *p += sizeof(*info->filelock_reply);
228 if (unlikely(*p != end))
237 * parse create results
239 static int parse_reply_info_create(void **p, void *end,
240 struct ceph_mds_reply_info_parsed *info,
243 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
245 info->has_create_ino = false;
247 info->has_create_ino = true;
248 info->ino = ceph_decode_64(p);
252 if (unlikely(*p != end))
261 * parse extra results
263 static int parse_reply_info_extra(void **p, void *end,
264 struct ceph_mds_reply_info_parsed *info,
267 if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
268 return parse_reply_info_filelock(p, end, info, features);
269 else if (info->head->op == CEPH_MDS_OP_READDIR ||
270 info->head->op == CEPH_MDS_OP_LSSNAP)
271 return parse_reply_info_dir(p, end, info, features);
272 else if (info->head->op == CEPH_MDS_OP_CREATE)
273 return parse_reply_info_create(p, end, info, features);
279 * parse entire mds reply
281 static int parse_reply_info(struct ceph_msg *msg,
282 struct ceph_mds_reply_info_parsed *info,
289 info->head = msg->front.iov_base;
290 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
291 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
294 ceph_decode_32_safe(&p, end, len, bad);
296 ceph_decode_need(&p, end, len, bad);
297 err = parse_reply_info_trace(&p, p+len, info, features);
303 ceph_decode_32_safe(&p, end, len, bad);
305 ceph_decode_need(&p, end, len, bad);
306 err = parse_reply_info_extra(&p, p+len, info, features);
312 ceph_decode_32_safe(&p, end, len, bad);
313 info->snapblob_len = len;
324 pr_err("mds parse_reply err %d\n", err);
328 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
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,
516 ceph_msg_put(req->r_request);
518 ceph_msg_put(req->r_reply);
519 destroy_reply_info(&req->r_reply_info);
522 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
525 if (req->r_locked_dir)
526 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
527 if (req->r_target_inode)
528 iput(req->r_target_inode);
531 if (req->r_old_dentry) {
533 * track (and drop pins for) r_old_dentry_dir
534 * separately, since r_old_dentry's d_parent may have
535 * changed between the dir mutex being dropped and
536 * this request being freed.
538 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
540 dput(req->r_old_dentry);
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 ceph_mdsc_put_request(req);
651 * Choose mds to send request to next. If there is a hint set in the
652 * request (e.g., due to a prior forward hint from the mds), use that.
653 * Otherwise, consult frag tree and/or caps to identify the
654 * appropriate mds. If all else fails, choose randomly.
656 * Called under mdsc->mutex.
658 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
661 * we don't need to worry about protecting the d_parent access
662 * here because we never renaming inside the snapped namespace
663 * except to resplice to another snapdir, and either the old or new
664 * result is a valid result.
666 while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
667 dentry = dentry->d_parent;
671 static int __choose_mds(struct ceph_mds_client *mdsc,
672 struct ceph_mds_request *req)
675 struct ceph_inode_info *ci;
676 struct ceph_cap *cap;
677 int mode = req->r_direct_mode;
679 u32 hash = req->r_direct_hash;
680 bool is_hash = req->r_direct_is_hash;
683 * is there a specific mds we should try? ignore hint if we have
684 * no session and the mds is not up (active or recovering).
686 if (req->r_resend_mds >= 0 &&
687 (__have_session(mdsc, req->r_resend_mds) ||
688 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
689 dout("choose_mds using resend_mds mds%d\n",
691 return req->r_resend_mds;
694 if (mode == USE_RANDOM_MDS)
699 inode = req->r_inode;
700 } else if (req->r_dentry) {
701 /* ignore race with rename; old or new d_parent is okay */
702 struct dentry *parent = req->r_dentry->d_parent;
703 struct inode *dir = parent->d_inode;
705 if (dir->i_sb != mdsc->fsc->sb) {
707 inode = req->r_dentry->d_inode;
708 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
709 /* direct snapped/virtual snapdir requests
710 * based on parent dir inode */
711 struct dentry *dn = get_nonsnap_parent(parent);
713 dout("__choose_mds using nonsnap parent %p\n", inode);
714 } else if (req->r_dentry->d_inode) {
716 inode = req->r_dentry->d_inode;
720 hash = ceph_dentry_hash(dir, req->r_dentry);
725 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
729 ci = ceph_inode(inode);
731 if (is_hash && S_ISDIR(inode->i_mode)) {
732 struct ceph_inode_frag frag;
735 ceph_choose_frag(ci, hash, &frag, &found);
737 if (mode == USE_ANY_MDS && frag.ndist > 0) {
740 /* choose a random replica */
741 get_random_bytes(&r, 1);
744 dout("choose_mds %p %llx.%llx "
745 "frag %u mds%d (%d/%d)\n",
746 inode, ceph_vinop(inode),
749 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
750 CEPH_MDS_STATE_ACTIVE)
754 /* since this file/dir wasn't known to be
755 * replicated, then we want to look for the
756 * authoritative mds. */
759 /* choose auth mds */
761 dout("choose_mds %p %llx.%llx "
762 "frag %u mds%d (auth)\n",
763 inode, ceph_vinop(inode), frag.frag, mds);
764 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
765 CEPH_MDS_STATE_ACTIVE)
771 spin_lock(&ci->i_ceph_lock);
773 if (mode == USE_AUTH_MDS)
774 cap = ci->i_auth_cap;
775 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
776 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
778 spin_unlock(&ci->i_ceph_lock);
781 mds = cap->session->s_mds;
782 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
783 inode, ceph_vinop(inode), mds,
784 cap == ci->i_auth_cap ? "auth " : "", cap);
785 spin_unlock(&ci->i_ceph_lock);
789 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
790 dout("choose_mds chose random mds%d\n", mds);
798 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
800 struct ceph_msg *msg;
801 struct ceph_mds_session_head *h;
803 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
806 pr_err("create_session_msg ENOMEM creating msg\n");
809 h = msg->front.iov_base;
810 h->op = cpu_to_le32(op);
811 h->seq = cpu_to_le64(seq);
816 * send session open request.
818 * called under mdsc->mutex
820 static int __open_session(struct ceph_mds_client *mdsc,
821 struct ceph_mds_session *session)
823 struct ceph_msg *msg;
825 int mds = session->s_mds;
827 /* wait for mds to go active? */
828 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
829 dout("open_session to mds%d (%s)\n", mds,
830 ceph_mds_state_name(mstate));
831 session->s_state = CEPH_MDS_SESSION_OPENING;
832 session->s_renew_requested = jiffies;
834 /* send connect message */
835 msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
838 ceph_con_send(&session->s_con, msg);
843 * open sessions for any export targets for the given mds
845 * called under mdsc->mutex
847 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
848 struct ceph_mds_session *session)
850 struct ceph_mds_info *mi;
851 struct ceph_mds_session *ts;
852 int i, mds = session->s_mds;
855 if (mds >= mdsc->mdsmap->m_max_mds)
857 mi = &mdsc->mdsmap->m_info[mds];
858 dout("open_export_target_sessions for mds%d (%d targets)\n",
859 session->s_mds, mi->num_export_targets);
861 for (i = 0; i < mi->num_export_targets; i++) {
862 target = mi->export_targets[i];
863 ts = __ceph_lookup_mds_session(mdsc, target);
865 ts = register_session(mdsc, target);
869 if (session->s_state == CEPH_MDS_SESSION_NEW ||
870 session->s_state == CEPH_MDS_SESSION_CLOSING)
871 __open_session(mdsc, session);
873 dout(" mds%d target mds%d %p is %s\n", session->s_mds,
874 i, ts, session_state_name(ts->s_state));
875 ceph_put_mds_session(ts);
879 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
880 struct ceph_mds_session *session)
882 mutex_lock(&mdsc->mutex);
883 __open_export_target_sessions(mdsc, session);
884 mutex_unlock(&mdsc->mutex);
892 * Free preallocated cap messages assigned to this session
894 static void cleanup_cap_releases(struct ceph_mds_session *session)
896 struct ceph_msg *msg;
898 spin_lock(&session->s_cap_lock);
899 while (!list_empty(&session->s_cap_releases)) {
900 msg = list_first_entry(&session->s_cap_releases,
901 struct ceph_msg, list_head);
902 list_del_init(&msg->list_head);
905 while (!list_empty(&session->s_cap_releases_done)) {
906 msg = list_first_entry(&session->s_cap_releases_done,
907 struct ceph_msg, list_head);
908 list_del_init(&msg->list_head);
911 spin_unlock(&session->s_cap_lock);
915 * Helper to safely iterate over all caps associated with a session, with
916 * special care taken to handle a racing __ceph_remove_cap().
918 * Caller must hold session s_mutex.
920 static int iterate_session_caps(struct ceph_mds_session *session,
921 int (*cb)(struct inode *, struct ceph_cap *,
925 struct ceph_cap *cap;
926 struct inode *inode, *last_inode = NULL;
927 struct ceph_cap *old_cap = NULL;
930 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
931 spin_lock(&session->s_cap_lock);
932 p = session->s_caps.next;
933 while (p != &session->s_caps) {
934 cap = list_entry(p, struct ceph_cap, session_caps);
935 inode = igrab(&cap->ci->vfs_inode);
940 session->s_cap_iterator = cap;
941 spin_unlock(&session->s_cap_lock);
948 ceph_put_cap(session->s_mdsc, old_cap);
952 ret = cb(inode, cap, arg);
955 spin_lock(&session->s_cap_lock);
957 if (cap->ci == NULL) {
958 dout("iterate_session_caps finishing cap %p removal\n",
960 BUG_ON(cap->session != session);
961 list_del_init(&cap->session_caps);
962 session->s_nr_caps--;
964 old_cap = cap; /* put_cap it w/o locks held */
971 session->s_cap_iterator = NULL;
972 spin_unlock(&session->s_cap_lock);
977 ceph_put_cap(session->s_mdsc, old_cap);
982 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
985 struct ceph_inode_info *ci = ceph_inode(inode);
988 dout("removing cap %p, ci is %p, inode is %p\n",
989 cap, ci, &ci->vfs_inode);
990 spin_lock(&ci->i_ceph_lock);
991 __ceph_remove_cap(cap, false);
992 if (!__ceph_is_any_real_caps(ci)) {
993 struct ceph_mds_client *mdsc =
994 ceph_sb_to_client(inode->i_sb)->mdsc;
996 spin_lock(&mdsc->cap_dirty_lock);
997 if (!list_empty(&ci->i_dirty_item)) {
998 pr_info(" dropping dirty %s state for %p %lld\n",
999 ceph_cap_string(ci->i_dirty_caps),
1000 inode, ceph_ino(inode));
1001 ci->i_dirty_caps = 0;
1002 list_del_init(&ci->i_dirty_item);
1005 if (!list_empty(&ci->i_flushing_item)) {
1006 pr_info(" dropping dirty+flushing %s state for %p %lld\n",
1007 ceph_cap_string(ci->i_flushing_caps),
1008 inode, ceph_ino(inode));
1009 ci->i_flushing_caps = 0;
1010 list_del_init(&ci->i_flushing_item);
1011 mdsc->num_cap_flushing--;
1014 if (drop && ci->i_wrbuffer_ref) {
1015 pr_info(" dropping dirty data for %p %lld\n",
1016 inode, ceph_ino(inode));
1017 ci->i_wrbuffer_ref = 0;
1018 ci->i_wrbuffer_ref_head = 0;
1021 spin_unlock(&mdsc->cap_dirty_lock);
1023 spin_unlock(&ci->i_ceph_lock);
1030 * caller must hold session s_mutex
1032 static void remove_session_caps(struct ceph_mds_session *session)
1034 dout("remove_session_caps on %p\n", session);
1035 iterate_session_caps(session, remove_session_caps_cb, NULL);
1037 spin_lock(&session->s_cap_lock);
1038 if (session->s_nr_caps > 0) {
1039 struct super_block *sb = session->s_mdsc->fsc->sb;
1040 struct inode *inode;
1041 struct ceph_cap *cap, *prev = NULL;
1042 struct ceph_vino vino;
1044 * iterate_session_caps() skips inodes that are being
1045 * deleted, we need to wait until deletions are complete.
1046 * __wait_on_freeing_inode() is designed for the job,
1047 * but it is not exported, so use lookup inode function
1050 while (!list_empty(&session->s_caps)) {
1051 cap = list_entry(session->s_caps.next,
1052 struct ceph_cap, session_caps);
1056 vino = cap->ci->i_vino;
1057 spin_unlock(&session->s_cap_lock);
1059 inode = ceph_find_inode(sb, vino);
1062 spin_lock(&session->s_cap_lock);
1065 spin_unlock(&session->s_cap_lock);
1067 BUG_ON(session->s_nr_caps > 0);
1068 BUG_ON(!list_empty(&session->s_cap_flushing));
1069 cleanup_cap_releases(session);
1073 * wake up any threads waiting on this session's caps. if the cap is
1074 * old (didn't get renewed on the client reconnect), remove it now.
1076 * caller must hold s_mutex.
1078 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1081 struct ceph_inode_info *ci = ceph_inode(inode);
1083 wake_up_all(&ci->i_cap_wq);
1085 spin_lock(&ci->i_ceph_lock);
1086 ci->i_wanted_max_size = 0;
1087 ci->i_requested_max_size = 0;
1088 spin_unlock(&ci->i_ceph_lock);
1093 static void wake_up_session_caps(struct ceph_mds_session *session,
1096 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1097 iterate_session_caps(session, wake_up_session_cb,
1098 (void *)(unsigned long)reconnect);
1102 * Send periodic message to MDS renewing all currently held caps. The
1103 * ack will reset the expiration for all caps from this session.
1105 * caller holds s_mutex
1107 static int send_renew_caps(struct ceph_mds_client *mdsc,
1108 struct ceph_mds_session *session)
1110 struct ceph_msg *msg;
1113 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1114 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1115 pr_info("mds%d caps stale\n", session->s_mds);
1116 session->s_renew_requested = jiffies;
1118 /* do not try to renew caps until a recovering mds has reconnected
1119 * with its clients. */
1120 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1121 if (state < CEPH_MDS_STATE_RECONNECT) {
1122 dout("send_renew_caps ignoring mds%d (%s)\n",
1123 session->s_mds, ceph_mds_state_name(state));
1127 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1128 ceph_mds_state_name(state));
1129 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1130 ++session->s_renew_seq);
1133 ceph_con_send(&session->s_con, msg);
1138 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1140 * Called under session->s_mutex
1142 static void renewed_caps(struct ceph_mds_client *mdsc,
1143 struct ceph_mds_session *session, int is_renew)
1148 spin_lock(&session->s_cap_lock);
1149 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1151 session->s_cap_ttl = session->s_renew_requested +
1152 mdsc->mdsmap->m_session_timeout*HZ;
1155 if (time_before(jiffies, session->s_cap_ttl)) {
1156 pr_info("mds%d caps renewed\n", session->s_mds);
1159 pr_info("mds%d caps still stale\n", session->s_mds);
1162 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1163 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1164 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1165 spin_unlock(&session->s_cap_lock);
1168 wake_up_session_caps(session, 0);
1172 * send a session close request
1174 static int request_close_session(struct ceph_mds_client *mdsc,
1175 struct ceph_mds_session *session)
1177 struct ceph_msg *msg;
1179 dout("request_close_session mds%d state %s seq %lld\n",
1180 session->s_mds, session_state_name(session->s_state),
1182 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1185 ceph_con_send(&session->s_con, msg);
1190 * Called with s_mutex held.
1192 static int __close_session(struct ceph_mds_client *mdsc,
1193 struct ceph_mds_session *session)
1195 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1197 session->s_state = CEPH_MDS_SESSION_CLOSING;
1198 return request_close_session(mdsc, session);
1202 * Trim old(er) caps.
1204 * Because we can't cache an inode without one or more caps, we do
1205 * this indirectly: if a cap is unused, we prune its aliases, at which
1206 * point the inode will hopefully get dropped to.
1208 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1209 * memory pressure from the MDS, though, so it needn't be perfect.
1211 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1213 struct ceph_mds_session *session = arg;
1214 struct ceph_inode_info *ci = ceph_inode(inode);
1215 int used, oissued, mine;
1217 if (session->s_trim_caps <= 0)
1220 spin_lock(&ci->i_ceph_lock);
1221 mine = cap->issued | cap->implemented;
1222 used = __ceph_caps_used(ci);
1223 oissued = __ceph_caps_issued_other(ci, cap);
1225 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1226 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1227 ceph_cap_string(used));
1228 if (ci->i_dirty_caps)
1229 goto out; /* dirty caps */
1230 if ((used & ~oissued) & mine)
1231 goto out; /* we need these caps */
1233 session->s_trim_caps--;
1235 /* we aren't the only cap.. just remove us */
1236 __ceph_remove_cap(cap, true);
1238 /* try to drop referring dentries */
1239 spin_unlock(&ci->i_ceph_lock);
1240 d_prune_aliases(inode);
1241 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1242 inode, cap, atomic_read(&inode->i_count));
1247 spin_unlock(&ci->i_ceph_lock);
1252 * Trim session cap count down to some max number.
1254 static int trim_caps(struct ceph_mds_client *mdsc,
1255 struct ceph_mds_session *session,
1258 int trim_caps = session->s_nr_caps - max_caps;
1260 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1261 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1262 if (trim_caps > 0) {
1263 session->s_trim_caps = trim_caps;
1264 iterate_session_caps(session, trim_caps_cb, session);
1265 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1266 session->s_mds, session->s_nr_caps, max_caps,
1267 trim_caps - session->s_trim_caps);
1268 session->s_trim_caps = 0;
1274 * Allocate cap_release messages. If there is a partially full message
1275 * in the queue, try to allocate enough to cover it's remainder, so that
1276 * we can send it immediately.
1278 * Called under s_mutex.
1280 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1281 struct ceph_mds_session *session)
1283 struct ceph_msg *msg, *partial = NULL;
1284 struct ceph_mds_cap_release *head;
1286 int extra = mdsc->fsc->mount_options->cap_release_safety;
1289 dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1292 spin_lock(&session->s_cap_lock);
1294 if (!list_empty(&session->s_cap_releases)) {
1295 msg = list_first_entry(&session->s_cap_releases,
1298 head = msg->front.iov_base;
1299 num = le32_to_cpu(head->num);
1301 dout(" partial %p with (%d/%d)\n", msg, num,
1302 (int)CEPH_CAPS_PER_RELEASE);
1303 extra += CEPH_CAPS_PER_RELEASE - num;
1307 while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1308 spin_unlock(&session->s_cap_lock);
1309 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1313 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1314 (int)msg->front.iov_len);
1315 head = msg->front.iov_base;
1316 head->num = cpu_to_le32(0);
1317 msg->front.iov_len = sizeof(*head);
1318 spin_lock(&session->s_cap_lock);
1319 list_add(&msg->list_head, &session->s_cap_releases);
1320 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1324 head = partial->front.iov_base;
1325 num = le32_to_cpu(head->num);
1326 dout(" queueing partial %p with %d/%d\n", partial, num,
1327 (int)CEPH_CAPS_PER_RELEASE);
1328 list_move_tail(&partial->list_head,
1329 &session->s_cap_releases_done);
1330 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1333 spin_unlock(&session->s_cap_lock);
1339 * flush all dirty inode data to disk.
1341 * returns true if we've flushed through want_flush_seq
1343 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1347 dout("check_cap_flush want %lld\n", want_flush_seq);
1348 mutex_lock(&mdsc->mutex);
1349 for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1350 struct ceph_mds_session *session = mdsc->sessions[mds];
1354 get_session(session);
1355 mutex_unlock(&mdsc->mutex);
1357 mutex_lock(&session->s_mutex);
1358 if (!list_empty(&session->s_cap_flushing)) {
1359 struct ceph_inode_info *ci =
1360 list_entry(session->s_cap_flushing.next,
1361 struct ceph_inode_info,
1363 struct inode *inode = &ci->vfs_inode;
1365 spin_lock(&ci->i_ceph_lock);
1366 if (ci->i_cap_flush_seq <= want_flush_seq) {
1367 dout("check_cap_flush still flushing %p "
1368 "seq %lld <= %lld to mds%d\n", inode,
1369 ci->i_cap_flush_seq, want_flush_seq,
1373 spin_unlock(&ci->i_ceph_lock);
1375 mutex_unlock(&session->s_mutex);
1376 ceph_put_mds_session(session);
1380 mutex_lock(&mdsc->mutex);
1383 mutex_unlock(&mdsc->mutex);
1384 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1389 * called under s_mutex
1391 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1392 struct ceph_mds_session *session)
1394 struct ceph_msg *msg;
1396 dout("send_cap_releases mds%d\n", session->s_mds);
1397 spin_lock(&session->s_cap_lock);
1398 while (!list_empty(&session->s_cap_releases_done)) {
1399 msg = list_first_entry(&session->s_cap_releases_done,
1400 struct ceph_msg, list_head);
1401 list_del_init(&msg->list_head);
1402 spin_unlock(&session->s_cap_lock);
1403 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1404 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1405 ceph_con_send(&session->s_con, msg);
1406 spin_lock(&session->s_cap_lock);
1408 spin_unlock(&session->s_cap_lock);
1411 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1412 struct ceph_mds_session *session)
1414 struct ceph_msg *msg;
1415 struct ceph_mds_cap_release *head;
1418 dout("discard_cap_releases mds%d\n", session->s_mds);
1420 /* zero out the in-progress message */
1421 msg = list_first_entry(&session->s_cap_releases,
1422 struct ceph_msg, list_head);
1423 head = msg->front.iov_base;
1424 num = le32_to_cpu(head->num);
1425 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1426 head->num = cpu_to_le32(0);
1427 msg->front.iov_len = sizeof(*head);
1428 session->s_num_cap_releases += num;
1430 /* requeue completed messages */
1431 while (!list_empty(&session->s_cap_releases_done)) {
1432 msg = list_first_entry(&session->s_cap_releases_done,
1433 struct ceph_msg, list_head);
1434 list_del_init(&msg->list_head);
1436 head = msg->front.iov_base;
1437 num = le32_to_cpu(head->num);
1438 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1440 session->s_num_cap_releases += num;
1441 head->num = cpu_to_le32(0);
1442 msg->front.iov_len = sizeof(*head);
1443 list_add(&msg->list_head, &session->s_cap_releases);
1452 * Create an mds request.
1454 struct ceph_mds_request *
1455 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1457 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1460 return ERR_PTR(-ENOMEM);
1462 mutex_init(&req->r_fill_mutex);
1464 req->r_started = jiffies;
1465 req->r_resend_mds = -1;
1466 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1468 kref_init(&req->r_kref);
1469 INIT_LIST_HEAD(&req->r_wait);
1470 init_completion(&req->r_completion);
1471 init_completion(&req->r_safe_completion);
1472 INIT_LIST_HEAD(&req->r_unsafe_item);
1475 req->r_direct_mode = mode;
1480 * return oldest (lowest) request, tid in request tree, 0 if none.
1482 * called under mdsc->mutex.
1484 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1486 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1488 return rb_entry(rb_first(&mdsc->request_tree),
1489 struct ceph_mds_request, r_node);
1492 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1494 struct ceph_mds_request *req = __get_oldest_req(mdsc);
1502 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1503 * on build_path_from_dentry in fs/cifs/dir.c.
1505 * If @stop_on_nosnap, generate path relative to the first non-snapped
1508 * Encode hidden .snap dirs as a double /, i.e.
1509 * foo/.snap/bar -> foo//bar
1511 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1514 struct dentry *temp;
1520 return ERR_PTR(-EINVAL);
1524 seq = read_seqbegin(&rename_lock);
1526 for (temp = dentry; !IS_ROOT(temp);) {
1527 struct inode *inode = temp->d_inode;
1528 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1529 len++; /* slash only */
1530 else if (stop_on_nosnap && inode &&
1531 ceph_snap(inode) == CEPH_NOSNAP)
1534 len += 1 + temp->d_name.len;
1535 temp = temp->d_parent;
1539 len--; /* no leading '/' */
1541 path = kmalloc(len+1, GFP_NOFS);
1543 return ERR_PTR(-ENOMEM);
1545 path[pos] = 0; /* trailing null */
1547 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1548 struct inode *inode;
1550 spin_lock(&temp->d_lock);
1551 inode = temp->d_inode;
1552 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1553 dout("build_path path+%d: %p SNAPDIR\n",
1555 } else if (stop_on_nosnap && inode &&
1556 ceph_snap(inode) == CEPH_NOSNAP) {
1557 spin_unlock(&temp->d_lock);
1560 pos -= temp->d_name.len;
1562 spin_unlock(&temp->d_lock);
1565 strncpy(path + pos, temp->d_name.name,
1568 spin_unlock(&temp->d_lock);
1571 temp = temp->d_parent;
1574 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1575 pr_err("build_path did not end path lookup where "
1576 "expected, namelen is %d, pos is %d\n", len, pos);
1577 /* presumably this is only possible if racing with a
1578 rename of one of the parent directories (we can not
1579 lock the dentries above us to prevent this, but
1580 retrying should be harmless) */
1585 *base = ceph_ino(temp->d_inode);
1587 dout("build_path on %p %d built %llx '%.*s'\n",
1588 dentry, d_count(dentry), *base, len, path);
1592 static int build_dentry_path(struct dentry *dentry,
1593 const char **ppath, int *ppathlen, u64 *pino,
1598 if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1599 *pino = ceph_ino(dentry->d_parent->d_inode);
1600 *ppath = dentry->d_name.name;
1601 *ppathlen = dentry->d_name.len;
1604 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1606 return PTR_ERR(path);
1612 static int build_inode_path(struct inode *inode,
1613 const char **ppath, int *ppathlen, u64 *pino,
1616 struct dentry *dentry;
1619 if (ceph_snap(inode) == CEPH_NOSNAP) {
1620 *pino = ceph_ino(inode);
1624 dentry = d_find_alias(inode);
1625 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1628 return PTR_ERR(path);
1635 * request arguments may be specified via an inode *, a dentry *, or
1636 * an explicit ino+path.
1638 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1639 const char *rpath, u64 rino,
1640 const char **ppath, int *pathlen,
1641 u64 *ino, int *freepath)
1646 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1647 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1649 } else if (rdentry) {
1650 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1651 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1653 } else if (rpath || rino) {
1656 *pathlen = rpath ? strlen(rpath) : 0;
1657 dout(" path %.*s\n", *pathlen, rpath);
1664 * called under mdsc->mutex
1666 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1667 struct ceph_mds_request *req,
1670 struct ceph_msg *msg;
1671 struct ceph_mds_request_head *head;
1672 const char *path1 = NULL;
1673 const char *path2 = NULL;
1674 u64 ino1 = 0, ino2 = 0;
1675 int pathlen1 = 0, pathlen2 = 0;
1676 int freepath1 = 0, freepath2 = 0;
1682 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1683 req->r_path1, req->r_ino1.ino,
1684 &path1, &pathlen1, &ino1, &freepath1);
1690 ret = set_request_path_attr(NULL, req->r_old_dentry,
1691 req->r_path2, req->r_ino2.ino,
1692 &path2, &pathlen2, &ino2, &freepath2);
1698 len = sizeof(*head) +
1699 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1701 /* calculate (max) length for cap releases */
1702 len += sizeof(struct ceph_mds_request_release) *
1703 (!!req->r_inode_drop + !!req->r_dentry_drop +
1704 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1705 if (req->r_dentry_drop)
1706 len += req->r_dentry->d_name.len;
1707 if (req->r_old_dentry_drop)
1708 len += req->r_old_dentry->d_name.len;
1710 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1712 msg = ERR_PTR(-ENOMEM);
1716 msg->hdr.tid = cpu_to_le64(req->r_tid);
1718 head = msg->front.iov_base;
1719 p = msg->front.iov_base + sizeof(*head);
1720 end = msg->front.iov_base + msg->front.iov_len;
1722 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1723 head->op = cpu_to_le32(req->r_op);
1724 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1725 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1726 head->args = req->r_args;
1728 ceph_encode_filepath(&p, end, ino1, path1);
1729 ceph_encode_filepath(&p, end, ino2, path2);
1731 /* make note of release offset, in case we need to replay */
1732 req->r_request_release_offset = p - msg->front.iov_base;
1736 if (req->r_inode_drop)
1737 releases += ceph_encode_inode_release(&p,
1738 req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1739 mds, req->r_inode_drop, req->r_inode_unless, 0);
1740 if (req->r_dentry_drop)
1741 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1742 mds, req->r_dentry_drop, req->r_dentry_unless);
1743 if (req->r_old_dentry_drop)
1744 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1745 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1746 if (req->r_old_inode_drop)
1747 releases += ceph_encode_inode_release(&p,
1748 req->r_old_dentry->d_inode,
1749 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1750 head->num_releases = cpu_to_le16(releases);
1753 msg->front.iov_len = p - msg->front.iov_base;
1754 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1756 if (req->r_data_len) {
1757 /* outbound data set only by ceph_sync_setxattr() */
1758 BUG_ON(!req->r_pages);
1759 ceph_msg_data_add_pages(msg, req->r_pages, req->r_data_len, 0);
1762 msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1763 msg->hdr.data_off = cpu_to_le16(0);
1767 kfree((char *)path2);
1770 kfree((char *)path1);
1776 * called under mdsc->mutex if error, under no mutex if
1779 static void complete_request(struct ceph_mds_client *mdsc,
1780 struct ceph_mds_request *req)
1782 if (req->r_callback)
1783 req->r_callback(mdsc, req);
1785 complete_all(&req->r_completion);
1789 * called under mdsc->mutex
1791 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1792 struct ceph_mds_request *req,
1795 struct ceph_mds_request_head *rhead;
1796 struct ceph_msg *msg;
1801 struct ceph_cap *cap =
1802 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1805 req->r_sent_on_mseq = cap->mseq;
1807 req->r_sent_on_mseq = -1;
1809 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1810 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1812 if (req->r_got_unsafe) {
1814 * Replay. Do not regenerate message (and rebuild
1815 * paths, etc.); just use the original message.
1816 * Rebuilding paths will break for renames because
1817 * d_move mangles the src name.
1819 msg = req->r_request;
1820 rhead = msg->front.iov_base;
1822 flags = le32_to_cpu(rhead->flags);
1823 flags |= CEPH_MDS_FLAG_REPLAY;
1824 rhead->flags = cpu_to_le32(flags);
1826 if (req->r_target_inode)
1827 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1829 rhead->num_retry = req->r_attempts - 1;
1831 /* remove cap/dentry releases from message */
1832 rhead->num_releases = 0;
1833 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1834 msg->front.iov_len = req->r_request_release_offset;
1838 if (req->r_request) {
1839 ceph_msg_put(req->r_request);
1840 req->r_request = NULL;
1842 msg = create_request_message(mdsc, req, mds);
1844 req->r_err = PTR_ERR(msg);
1845 complete_request(mdsc, req);
1846 return PTR_ERR(msg);
1848 req->r_request = msg;
1850 rhead = msg->front.iov_base;
1851 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1852 if (req->r_got_unsafe)
1853 flags |= CEPH_MDS_FLAG_REPLAY;
1854 if (req->r_locked_dir)
1855 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1856 rhead->flags = cpu_to_le32(flags);
1857 rhead->num_fwd = req->r_num_fwd;
1858 rhead->num_retry = req->r_attempts - 1;
1861 dout(" r_locked_dir = %p\n", req->r_locked_dir);
1866 * send request, or put it on the appropriate wait list.
1868 static int __do_request(struct ceph_mds_client *mdsc,
1869 struct ceph_mds_request *req)
1871 struct ceph_mds_session *session = NULL;
1875 if (req->r_err || req->r_got_result) {
1877 __unregister_request(mdsc, req);
1881 if (req->r_timeout &&
1882 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1883 dout("do_request timed out\n");
1888 put_request_session(req);
1890 mds = __choose_mds(mdsc, req);
1892 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1893 dout("do_request no mds or not active, waiting for map\n");
1894 list_add(&req->r_wait, &mdsc->waiting_for_map);
1898 /* get, open session */
1899 session = __ceph_lookup_mds_session(mdsc, mds);
1901 session = register_session(mdsc, mds);
1902 if (IS_ERR(session)) {
1903 err = PTR_ERR(session);
1907 req->r_session = get_session(session);
1909 dout("do_request mds%d session %p state %s\n", mds, session,
1910 session_state_name(session->s_state));
1911 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1912 session->s_state != CEPH_MDS_SESSION_HUNG) {
1913 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1914 session->s_state == CEPH_MDS_SESSION_CLOSING)
1915 __open_session(mdsc, session);
1916 list_add(&req->r_wait, &session->s_waiting);
1921 req->r_resend_mds = -1; /* forget any previous mds hint */
1923 if (req->r_request_started == 0) /* note request start time */
1924 req->r_request_started = jiffies;
1926 err = __prepare_send_request(mdsc, req, mds);
1928 ceph_msg_get(req->r_request);
1929 ceph_con_send(&session->s_con, req->r_request);
1933 ceph_put_mds_session(session);
1939 complete_request(mdsc, req);
1944 * called under mdsc->mutex
1946 static void __wake_requests(struct ceph_mds_client *mdsc,
1947 struct list_head *head)
1949 struct ceph_mds_request *req;
1950 LIST_HEAD(tmp_list);
1952 list_splice_init(head, &tmp_list);
1954 while (!list_empty(&tmp_list)) {
1955 req = list_entry(tmp_list.next,
1956 struct ceph_mds_request, r_wait);
1957 list_del_init(&req->r_wait);
1958 dout(" wake request %p tid %llu\n", req, req->r_tid);
1959 __do_request(mdsc, req);
1964 * Wake up threads with requests pending for @mds, so that they can
1965 * resubmit their requests to a possibly different mds.
1967 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1969 struct ceph_mds_request *req;
1972 dout("kick_requests mds%d\n", mds);
1973 for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1974 req = rb_entry(p, struct ceph_mds_request, r_node);
1975 if (req->r_got_unsafe)
1977 if (req->r_session &&
1978 req->r_session->s_mds == mds) {
1979 dout(" kicking tid %llu\n", req->r_tid);
1980 __do_request(mdsc, req);
1985 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1986 struct ceph_mds_request *req)
1988 dout("submit_request on %p\n", req);
1989 mutex_lock(&mdsc->mutex);
1990 __register_request(mdsc, req, NULL);
1991 __do_request(mdsc, req);
1992 mutex_unlock(&mdsc->mutex);
1996 * Synchrously perform an mds request. Take care of all of the
1997 * session setup, forwarding, retry details.
1999 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2001 struct ceph_mds_request *req)
2005 dout("do_request on %p\n", req);
2007 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2009 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2010 if (req->r_locked_dir)
2011 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2012 if (req->r_old_dentry)
2013 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2017 mutex_lock(&mdsc->mutex);
2018 __register_request(mdsc, req, dir);
2019 __do_request(mdsc, req);
2023 __unregister_request(mdsc, req);
2024 dout("do_request early error %d\n", err);
2029 mutex_unlock(&mdsc->mutex);
2030 dout("do_request waiting\n");
2031 if (req->r_timeout) {
2032 err = (long)wait_for_completion_killable_timeout(
2033 &req->r_completion, req->r_timeout);
2037 err = wait_for_completion_killable(&req->r_completion);
2039 dout("do_request waited, got %d\n", err);
2040 mutex_lock(&mdsc->mutex);
2042 /* only abort if we didn't race with a real reply */
2043 if (req->r_got_result) {
2044 err = le32_to_cpu(req->r_reply_info.head->result);
2045 } else if (err < 0) {
2046 dout("aborted request %lld with %d\n", req->r_tid, err);
2049 * ensure we aren't running concurrently with
2050 * ceph_fill_trace or ceph_readdir_prepopulate, which
2051 * rely on locks (dir mutex) held by our caller.
2053 mutex_lock(&req->r_fill_mutex);
2055 req->r_aborted = true;
2056 mutex_unlock(&req->r_fill_mutex);
2058 if (req->r_locked_dir &&
2059 (req->r_op & CEPH_MDS_OP_WRITE))
2060 ceph_invalidate_dir_request(req);
2066 mutex_unlock(&mdsc->mutex);
2067 dout("do_request %p done, result %d\n", req, err);
2072 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2073 * namespace request.
2075 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2077 struct inode *inode = req->r_locked_dir;
2079 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2081 ceph_dir_clear_complete(inode);
2083 ceph_invalidate_dentry_lease(req->r_dentry);
2084 if (req->r_old_dentry)
2085 ceph_invalidate_dentry_lease(req->r_old_dentry);
2091 * We take the session mutex and parse and process the reply immediately.
2092 * This preserves the logical ordering of replies, capabilities, etc., sent
2093 * by the MDS as they are applied to our local cache.
2095 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2097 struct ceph_mds_client *mdsc = session->s_mdsc;
2098 struct ceph_mds_request *req;
2099 struct ceph_mds_reply_head *head = msg->front.iov_base;
2100 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2103 int mds = session->s_mds;
2105 if (msg->front.iov_len < sizeof(*head)) {
2106 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2111 /* get request, session */
2112 tid = le64_to_cpu(msg->hdr.tid);
2113 mutex_lock(&mdsc->mutex);
2114 req = __lookup_request(mdsc, tid);
2116 dout("handle_reply on unknown tid %llu\n", tid);
2117 mutex_unlock(&mdsc->mutex);
2120 dout("handle_reply %p\n", req);
2122 /* correct session? */
2123 if (req->r_session != session) {
2124 pr_err("mdsc_handle_reply got %llu on session mds%d"
2125 " not mds%d\n", tid, session->s_mds,
2126 req->r_session ? req->r_session->s_mds : -1);
2127 mutex_unlock(&mdsc->mutex);
2132 if ((req->r_got_unsafe && !head->safe) ||
2133 (req->r_got_safe && head->safe)) {
2134 pr_warning("got a dup %s reply on %llu from mds%d\n",
2135 head->safe ? "safe" : "unsafe", tid, mds);
2136 mutex_unlock(&mdsc->mutex);
2139 if (req->r_got_safe && !head->safe) {
2140 pr_warning("got unsafe after safe on %llu from mds%d\n",
2142 mutex_unlock(&mdsc->mutex);
2146 result = le32_to_cpu(head->result);
2150 * if we're not talking to the authority, send to them
2151 * if the authority has changed while we weren't looking,
2152 * send to new authority
2153 * Otherwise we just have to return an ESTALE
2155 if (result == -ESTALE) {
2156 dout("got ESTALE on request %llu", req->r_tid);
2157 if (!req->r_inode) {
2158 /* do nothing; not an authority problem */
2159 } else if (req->r_direct_mode != USE_AUTH_MDS) {
2160 dout("not using auth, setting for that now");
2161 req->r_direct_mode = USE_AUTH_MDS;
2162 __do_request(mdsc, req);
2163 mutex_unlock(&mdsc->mutex);
2166 struct ceph_inode_info *ci = ceph_inode(req->r_inode);
2167 struct ceph_cap *cap = NULL;
2170 cap = ceph_get_cap_for_mds(ci,
2171 req->r_session->s_mds);
2173 dout("already using auth");
2174 if ((!cap || cap != ci->i_auth_cap) ||
2175 (cap->mseq != req->r_sent_on_mseq)) {
2176 dout("but cap changed, so resending");
2177 __do_request(mdsc, req);
2178 mutex_unlock(&mdsc->mutex);
2182 dout("have to return ESTALE on request %llu", req->r_tid);
2187 req->r_got_safe = true;
2188 __unregister_request(mdsc, req);
2189 complete_all(&req->r_safe_completion);
2191 if (req->r_got_unsafe) {
2193 * We already handled the unsafe response, now do the
2194 * cleanup. No need to examine the response; the MDS
2195 * doesn't include any result info in the safe
2196 * response. And even if it did, there is nothing
2197 * useful we could do with a revised return value.
2199 dout("got safe reply %llu, mds%d\n", tid, mds);
2200 list_del_init(&req->r_unsafe_item);
2202 /* last unsafe request during umount? */
2203 if (mdsc->stopping && !__get_oldest_req(mdsc))
2204 complete_all(&mdsc->safe_umount_waiters);
2205 mutex_unlock(&mdsc->mutex);
2209 req->r_got_unsafe = true;
2210 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2213 dout("handle_reply tid %lld result %d\n", tid, result);
2214 rinfo = &req->r_reply_info;
2215 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2216 mutex_unlock(&mdsc->mutex);
2218 mutex_lock(&session->s_mutex);
2220 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2226 if (rinfo->snapblob_len) {
2227 down_write(&mdsc->snap_rwsem);
2228 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2229 rinfo->snapblob + rinfo->snapblob_len,
2230 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2231 downgrade_write(&mdsc->snap_rwsem);
2233 down_read(&mdsc->snap_rwsem);
2236 /* insert trace into our cache */
2237 mutex_lock(&req->r_fill_mutex);
2238 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2240 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2241 req->r_op == CEPH_MDS_OP_LSSNAP))
2242 ceph_readdir_prepopulate(req, req->r_session);
2243 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2245 mutex_unlock(&req->r_fill_mutex);
2247 up_read(&mdsc->snap_rwsem);
2249 mutex_lock(&mdsc->mutex);
2250 if (!req->r_aborted) {
2256 req->r_got_result = true;
2259 dout("reply arrived after request %lld was aborted\n", tid);
2261 mutex_unlock(&mdsc->mutex);
2263 ceph_add_cap_releases(mdsc, req->r_session);
2264 mutex_unlock(&session->s_mutex);
2266 /* kick calling process */
2267 complete_request(mdsc, req);
2269 ceph_mdsc_put_request(req);
2276 * handle mds notification that our request has been forwarded.
2278 static void handle_forward(struct ceph_mds_client *mdsc,
2279 struct ceph_mds_session *session,
2280 struct ceph_msg *msg)
2282 struct ceph_mds_request *req;
2283 u64 tid = le64_to_cpu(msg->hdr.tid);
2287 void *p = msg->front.iov_base;
2288 void *end = p + msg->front.iov_len;
2290 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2291 next_mds = ceph_decode_32(&p);
2292 fwd_seq = ceph_decode_32(&p);
2294 mutex_lock(&mdsc->mutex);
2295 req = __lookup_request(mdsc, tid);
2297 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2298 goto out; /* dup reply? */
2301 if (req->r_aborted) {
2302 dout("forward tid %llu aborted, unregistering\n", tid);
2303 __unregister_request(mdsc, req);
2304 } else if (fwd_seq <= req->r_num_fwd) {
2305 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2306 tid, next_mds, req->r_num_fwd, fwd_seq);
2308 /* resend. forward race not possible; mds would drop */
2309 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2311 BUG_ON(req->r_got_result);
2312 req->r_num_fwd = fwd_seq;
2313 req->r_resend_mds = next_mds;
2314 put_request_session(req);
2315 __do_request(mdsc, req);
2317 ceph_mdsc_put_request(req);
2319 mutex_unlock(&mdsc->mutex);
2323 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2327 * handle a mds session control message
2329 static void handle_session(struct ceph_mds_session *session,
2330 struct ceph_msg *msg)
2332 struct ceph_mds_client *mdsc = session->s_mdsc;
2335 int mds = session->s_mds;
2336 struct ceph_mds_session_head *h = msg->front.iov_base;
2340 if (msg->front.iov_len != sizeof(*h))
2342 op = le32_to_cpu(h->op);
2343 seq = le64_to_cpu(h->seq);
2345 mutex_lock(&mdsc->mutex);
2346 if (op == CEPH_SESSION_CLOSE)
2347 __unregister_session(mdsc, session);
2348 /* FIXME: this ttl calculation is generous */
2349 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2350 mutex_unlock(&mdsc->mutex);
2352 mutex_lock(&session->s_mutex);
2354 dout("handle_session mds%d %s %p state %s seq %llu\n",
2355 mds, ceph_session_op_name(op), session,
2356 session_state_name(session->s_state), seq);
2358 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2359 session->s_state = CEPH_MDS_SESSION_OPEN;
2360 pr_info("mds%d came back\n", session->s_mds);
2364 case CEPH_SESSION_OPEN:
2365 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2366 pr_info("mds%d reconnect success\n", session->s_mds);
2367 session->s_state = CEPH_MDS_SESSION_OPEN;
2368 renewed_caps(mdsc, session, 0);
2371 __close_session(mdsc, session);
2374 case CEPH_SESSION_RENEWCAPS:
2375 if (session->s_renew_seq == seq)
2376 renewed_caps(mdsc, session, 1);
2379 case CEPH_SESSION_CLOSE:
2380 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2381 pr_info("mds%d reconnect denied\n", session->s_mds);
2382 remove_session_caps(session);
2383 wake = 1; /* for good measure */
2384 wake_up_all(&mdsc->session_close_wq);
2385 kick_requests(mdsc, mds);
2388 case CEPH_SESSION_STALE:
2389 pr_info("mds%d caps went stale, renewing\n",
2391 spin_lock(&session->s_gen_ttl_lock);
2392 session->s_cap_gen++;
2393 session->s_cap_ttl = jiffies - 1;
2394 spin_unlock(&session->s_gen_ttl_lock);
2395 send_renew_caps(mdsc, session);
2398 case CEPH_SESSION_RECALL_STATE:
2399 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2403 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2407 mutex_unlock(&session->s_mutex);
2409 mutex_lock(&mdsc->mutex);
2410 __wake_requests(mdsc, &session->s_waiting);
2411 mutex_unlock(&mdsc->mutex);
2416 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2417 (int)msg->front.iov_len);
2424 * called under session->mutex.
2426 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2427 struct ceph_mds_session *session)
2429 struct ceph_mds_request *req, *nreq;
2432 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2434 mutex_lock(&mdsc->mutex);
2435 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2436 err = __prepare_send_request(mdsc, req, session->s_mds);
2438 ceph_msg_get(req->r_request);
2439 ceph_con_send(&session->s_con, req->r_request);
2442 mutex_unlock(&mdsc->mutex);
2446 * Encode information about a cap for a reconnect with the MDS.
2448 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2452 struct ceph_mds_cap_reconnect v2;
2453 struct ceph_mds_cap_reconnect_v1 v1;
2456 struct ceph_inode_info *ci;
2457 struct ceph_reconnect_state *recon_state = arg;
2458 struct ceph_pagelist *pagelist = recon_state->pagelist;
2462 struct dentry *dentry;
2466 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2467 inode, ceph_vinop(inode), cap, cap->cap_id,
2468 ceph_cap_string(cap->issued));
2469 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2473 dentry = d_find_alias(inode);
2475 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2477 err = PTR_ERR(path);
2484 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2488 spin_lock(&ci->i_ceph_lock);
2489 cap->seq = 0; /* reset cap seq */
2490 cap->issue_seq = 0; /* and issue_seq */
2491 cap->mseq = 0; /* and migrate_seq */
2492 cap->cap_gen = cap->session->s_cap_gen;
2494 if (recon_state->flock) {
2495 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2496 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2497 rec.v2.issued = cpu_to_le32(cap->issued);
2498 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2499 rec.v2.pathbase = cpu_to_le64(pathbase);
2500 rec.v2.flock_len = 0;
2501 reclen = sizeof(rec.v2);
2503 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2504 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2505 rec.v1.issued = cpu_to_le32(cap->issued);
2506 rec.v1.size = cpu_to_le64(inode->i_size);
2507 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2508 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2509 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2510 rec.v1.pathbase = cpu_to_le64(pathbase);
2511 reclen = sizeof(rec.v1);
2513 spin_unlock(&ci->i_ceph_lock);
2515 if (recon_state->flock) {
2516 int num_fcntl_locks, num_flock_locks;
2517 struct ceph_filelock *flocks;
2520 spin_lock(&inode->i_lock);
2521 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2522 spin_unlock(&inode->i_lock);
2523 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2524 sizeof(struct ceph_filelock), GFP_NOFS);
2529 spin_lock(&inode->i_lock);
2530 err = ceph_encode_locks_to_buffer(inode, flocks,
2533 spin_unlock(&inode->i_lock);
2541 * number of encoded locks is stable, so copy to pagelist
2543 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2544 (num_fcntl_locks+num_flock_locks) *
2545 sizeof(struct ceph_filelock));
2546 err = ceph_pagelist_append(pagelist, &rec, reclen);
2548 err = ceph_locks_to_pagelist(flocks, pagelist,
2553 err = ceph_pagelist_append(pagelist, &rec, reclen);
2556 recon_state->nr_caps++;
2566 * If an MDS fails and recovers, clients need to reconnect in order to
2567 * reestablish shared state. This includes all caps issued through
2568 * this session _and_ the snap_realm hierarchy. Because it's not
2569 * clear which snap realms the mds cares about, we send everything we
2570 * know about.. that ensures we'll then get any new info the
2571 * recovering MDS might have.
2573 * This is a relatively heavyweight operation, but it's rare.
2575 * called with mdsc->mutex held.
2577 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2578 struct ceph_mds_session *session)
2580 struct ceph_msg *reply;
2582 int mds = session->s_mds;
2585 struct ceph_pagelist *pagelist;
2586 struct ceph_reconnect_state recon_state;
2588 pr_info("mds%d reconnect start\n", mds);
2590 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2592 goto fail_nopagelist;
2593 ceph_pagelist_init(pagelist);
2595 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2599 mutex_lock(&session->s_mutex);
2600 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2603 ceph_con_close(&session->s_con);
2604 ceph_con_open(&session->s_con,
2605 CEPH_ENTITY_TYPE_MDS, mds,
2606 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2608 /* replay unsafe requests */
2609 replay_unsafe_requests(mdsc, session);
2611 down_read(&mdsc->snap_rwsem);
2613 dout("session %p state %s\n", session,
2614 session_state_name(session->s_state));
2616 spin_lock(&session->s_gen_ttl_lock);
2617 session->s_cap_gen++;
2618 spin_unlock(&session->s_gen_ttl_lock);
2620 spin_lock(&session->s_cap_lock);
2622 * notify __ceph_remove_cap() that we are composing cap reconnect.
2623 * If a cap get released before being added to the cap reconnect,
2624 * __ceph_remove_cap() should skip queuing cap release.
2626 session->s_cap_reconnect = 1;
2627 /* drop old cap expires; we're about to reestablish that state */
2628 discard_cap_releases(mdsc, session);
2629 spin_unlock(&session->s_cap_lock);
2631 /* traverse this session's caps */
2632 s_nr_caps = session->s_nr_caps;
2633 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2637 recon_state.nr_caps = 0;
2638 recon_state.pagelist = pagelist;
2639 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2640 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2644 spin_lock(&session->s_cap_lock);
2645 session->s_cap_reconnect = 0;
2646 spin_unlock(&session->s_cap_lock);
2649 * snaprealms. we provide mds with the ino, seq (version), and
2650 * parent for all of our realms. If the mds has any newer info,
2653 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2654 struct ceph_snap_realm *realm =
2655 rb_entry(p, struct ceph_snap_realm, node);
2656 struct ceph_mds_snaprealm_reconnect sr_rec;
2658 dout(" adding snap realm %llx seq %lld parent %llx\n",
2659 realm->ino, realm->seq, realm->parent_ino);
2660 sr_rec.ino = cpu_to_le64(realm->ino);
2661 sr_rec.seq = cpu_to_le64(realm->seq);
2662 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2663 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2668 if (recon_state.flock)
2669 reply->hdr.version = cpu_to_le16(2);
2671 /* raced with cap release? */
2672 if (s_nr_caps != recon_state.nr_caps) {
2673 struct page *page = list_first_entry(&pagelist->head,
2675 __le32 *addr = kmap_atomic(page);
2676 *addr = cpu_to_le32(recon_state.nr_caps);
2677 kunmap_atomic(addr);
2680 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2681 ceph_msg_data_add_pagelist(reply, pagelist);
2682 ceph_con_send(&session->s_con, reply);
2684 mutex_unlock(&session->s_mutex);
2686 mutex_lock(&mdsc->mutex);
2687 __wake_requests(mdsc, &session->s_waiting);
2688 mutex_unlock(&mdsc->mutex);
2690 up_read(&mdsc->snap_rwsem);
2694 ceph_msg_put(reply);
2695 up_read(&mdsc->snap_rwsem);
2696 mutex_unlock(&session->s_mutex);
2698 ceph_pagelist_release(pagelist);
2701 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2707 * compare old and new mdsmaps, kicking requests
2708 * and closing out old connections as necessary
2710 * called under mdsc->mutex.
2712 static void check_new_map(struct ceph_mds_client *mdsc,
2713 struct ceph_mdsmap *newmap,
2714 struct ceph_mdsmap *oldmap)
2717 int oldstate, newstate;
2718 struct ceph_mds_session *s;
2720 dout("check_new_map new %u old %u\n",
2721 newmap->m_epoch, oldmap->m_epoch);
2723 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2724 if (mdsc->sessions[i] == NULL)
2726 s = mdsc->sessions[i];
2727 oldstate = ceph_mdsmap_get_state(oldmap, i);
2728 newstate = ceph_mdsmap_get_state(newmap, i);
2730 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2731 i, ceph_mds_state_name(oldstate),
2732 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2733 ceph_mds_state_name(newstate),
2734 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2735 session_state_name(s->s_state));
2737 if (i >= newmap->m_max_mds ||
2738 memcmp(ceph_mdsmap_get_addr(oldmap, i),
2739 ceph_mdsmap_get_addr(newmap, i),
2740 sizeof(struct ceph_entity_addr))) {
2741 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2742 /* the session never opened, just close it
2744 __wake_requests(mdsc, &s->s_waiting);
2745 __unregister_session(mdsc, s);
2748 mutex_unlock(&mdsc->mutex);
2749 mutex_lock(&s->s_mutex);
2750 mutex_lock(&mdsc->mutex);
2751 ceph_con_close(&s->s_con);
2752 mutex_unlock(&s->s_mutex);
2753 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2756 /* kick any requests waiting on the recovering mds */
2757 kick_requests(mdsc, i);
2758 } else if (oldstate == newstate) {
2759 continue; /* nothing new with this mds */
2765 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2766 newstate >= CEPH_MDS_STATE_RECONNECT) {
2767 mutex_unlock(&mdsc->mutex);
2768 send_mds_reconnect(mdsc, s);
2769 mutex_lock(&mdsc->mutex);
2773 * kick request on any mds that has gone active.
2775 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2776 newstate >= CEPH_MDS_STATE_ACTIVE) {
2777 if (oldstate != CEPH_MDS_STATE_CREATING &&
2778 oldstate != CEPH_MDS_STATE_STARTING)
2779 pr_info("mds%d recovery completed\n", s->s_mds);
2780 kick_requests(mdsc, i);
2781 ceph_kick_flushing_caps(mdsc, s);
2782 wake_up_session_caps(s, 1);
2786 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2787 s = mdsc->sessions[i];
2790 if (!ceph_mdsmap_is_laggy(newmap, i))
2792 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2793 s->s_state == CEPH_MDS_SESSION_HUNG ||
2794 s->s_state == CEPH_MDS_SESSION_CLOSING) {
2795 dout(" connecting to export targets of laggy mds%d\n",
2797 __open_export_target_sessions(mdsc, s);
2809 * caller must hold session s_mutex, dentry->d_lock
2811 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2813 struct ceph_dentry_info *di = ceph_dentry(dentry);
2815 ceph_put_mds_session(di->lease_session);
2816 di->lease_session = NULL;
2819 static void handle_lease(struct ceph_mds_client *mdsc,
2820 struct ceph_mds_session *session,
2821 struct ceph_msg *msg)
2823 struct super_block *sb = mdsc->fsc->sb;
2824 struct inode *inode;
2825 struct dentry *parent, *dentry;
2826 struct ceph_dentry_info *di;
2827 int mds = session->s_mds;
2828 struct ceph_mds_lease *h = msg->front.iov_base;
2830 struct ceph_vino vino;
2834 dout("handle_lease from mds%d\n", mds);
2837 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2839 vino.ino = le64_to_cpu(h->ino);
2840 vino.snap = CEPH_NOSNAP;
2841 seq = le32_to_cpu(h->seq);
2842 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2843 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2844 if (dname.len != get_unaligned_le32(h+1))
2847 mutex_lock(&session->s_mutex);
2851 inode = ceph_find_inode(sb, vino);
2852 dout("handle_lease %s, ino %llx %p %.*s\n",
2853 ceph_lease_op_name(h->action), vino.ino, inode,
2854 dname.len, dname.name);
2855 if (inode == NULL) {
2856 dout("handle_lease no inode %llx\n", vino.ino);
2861 parent = d_find_alias(inode);
2863 dout("no parent dentry on inode %p\n", inode);
2865 goto release; /* hrm... */
2867 dname.hash = full_name_hash(dname.name, dname.len);
2868 dentry = d_lookup(parent, &dname);
2873 spin_lock(&dentry->d_lock);
2874 di = ceph_dentry(dentry);
2875 switch (h->action) {
2876 case CEPH_MDS_LEASE_REVOKE:
2877 if (di->lease_session == session) {
2878 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2879 h->seq = cpu_to_le32(di->lease_seq);
2880 __ceph_mdsc_drop_dentry_lease(dentry);
2885 case CEPH_MDS_LEASE_RENEW:
2886 if (di->lease_session == session &&
2887 di->lease_gen == session->s_cap_gen &&
2888 di->lease_renew_from &&
2889 di->lease_renew_after == 0) {
2890 unsigned long duration =
2891 le32_to_cpu(h->duration_ms) * HZ / 1000;
2893 di->lease_seq = seq;
2894 dentry->d_time = di->lease_renew_from + duration;
2895 di->lease_renew_after = di->lease_renew_from +
2897 di->lease_renew_from = 0;
2901 spin_unlock(&dentry->d_lock);
2908 /* let's just reuse the same message */
2909 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2911 ceph_con_send(&session->s_con, msg);
2915 mutex_unlock(&session->s_mutex);
2919 pr_err("corrupt lease message\n");
2923 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2924 struct inode *inode,
2925 struct dentry *dentry, char action,
2928 struct ceph_msg *msg;
2929 struct ceph_mds_lease *lease;
2930 int len = sizeof(*lease) + sizeof(u32);
2933 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2934 inode, dentry, ceph_lease_op_name(action), session->s_mds);
2935 dnamelen = dentry->d_name.len;
2938 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
2941 lease = msg->front.iov_base;
2942 lease->action = action;
2943 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2944 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2945 lease->seq = cpu_to_le32(seq);
2946 put_unaligned_le32(dnamelen, lease + 1);
2947 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2950 * if this is a preemptive lease RELEASE, no need to
2951 * flush request stream, since the actual request will
2954 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2956 ceph_con_send(&session->s_con, msg);
2960 * Preemptively release a lease we expect to invalidate anyway.
2961 * Pass @inode always, @dentry is optional.
2963 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2964 struct dentry *dentry)
2966 struct ceph_dentry_info *di;
2967 struct ceph_mds_session *session;
2970 BUG_ON(inode == NULL);
2971 BUG_ON(dentry == NULL);
2973 /* is dentry lease valid? */
2974 spin_lock(&dentry->d_lock);
2975 di = ceph_dentry(dentry);
2976 if (!di || !di->lease_session ||
2977 di->lease_session->s_mds < 0 ||
2978 di->lease_gen != di->lease_session->s_cap_gen ||
2979 !time_before(jiffies, dentry->d_time)) {
2980 dout("lease_release inode %p dentry %p -- "
2983 spin_unlock(&dentry->d_lock);
2987 /* we do have a lease on this dentry; note mds and seq */
2988 session = ceph_get_mds_session(di->lease_session);
2989 seq = di->lease_seq;
2990 __ceph_mdsc_drop_dentry_lease(dentry);
2991 spin_unlock(&dentry->d_lock);
2993 dout("lease_release inode %p dentry %p to mds%d\n",
2994 inode, dentry, session->s_mds);
2995 ceph_mdsc_lease_send_msg(session, inode, dentry,
2996 CEPH_MDS_LEASE_RELEASE, seq);
2997 ceph_put_mds_session(session);
3001 * drop all leases (and dentry refs) in preparation for umount
3003 static void drop_leases(struct ceph_mds_client *mdsc)
3007 dout("drop_leases\n");
3008 mutex_lock(&mdsc->mutex);
3009 for (i = 0; i < mdsc->max_sessions; i++) {
3010 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3013 mutex_unlock(&mdsc->mutex);
3014 mutex_lock(&s->s_mutex);
3015 mutex_unlock(&s->s_mutex);
3016 ceph_put_mds_session(s);
3017 mutex_lock(&mdsc->mutex);
3019 mutex_unlock(&mdsc->mutex);
3025 * delayed work -- periodically trim expired leases, renew caps with mds
3027 static void schedule_delayed(struct ceph_mds_client *mdsc)
3030 unsigned hz = round_jiffies_relative(HZ * delay);
3031 schedule_delayed_work(&mdsc->delayed_work, hz);
3034 static void delayed_work(struct work_struct *work)
3037 struct ceph_mds_client *mdsc =
3038 container_of(work, struct ceph_mds_client, delayed_work.work);
3042 dout("mdsc delayed_work\n");
3043 ceph_check_delayed_caps(mdsc);
3045 mutex_lock(&mdsc->mutex);
3046 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3047 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3048 mdsc->last_renew_caps);
3050 mdsc->last_renew_caps = jiffies;
3052 for (i = 0; i < mdsc->max_sessions; i++) {
3053 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3056 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3057 dout("resending session close request for mds%d\n",
3059 request_close_session(mdsc, s);
3060 ceph_put_mds_session(s);
3063 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3064 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3065 s->s_state = CEPH_MDS_SESSION_HUNG;
3066 pr_info("mds%d hung\n", s->s_mds);
3069 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3070 /* this mds is failed or recovering, just wait */
3071 ceph_put_mds_session(s);
3074 mutex_unlock(&mdsc->mutex);
3076 mutex_lock(&s->s_mutex);
3078 send_renew_caps(mdsc, s);
3080 ceph_con_keepalive(&s->s_con);
3081 ceph_add_cap_releases(mdsc, s);
3082 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3083 s->s_state == CEPH_MDS_SESSION_HUNG)
3084 ceph_send_cap_releases(mdsc, s);
3085 mutex_unlock(&s->s_mutex);
3086 ceph_put_mds_session(s);
3088 mutex_lock(&mdsc->mutex);
3090 mutex_unlock(&mdsc->mutex);
3092 schedule_delayed(mdsc);
3095 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3098 struct ceph_mds_client *mdsc;
3100 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3105 mutex_init(&mdsc->mutex);
3106 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3107 if (mdsc->mdsmap == NULL) {
3112 init_completion(&mdsc->safe_umount_waiters);
3113 init_waitqueue_head(&mdsc->session_close_wq);
3114 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3115 mdsc->sessions = NULL;
3116 mdsc->max_sessions = 0;
3118 init_rwsem(&mdsc->snap_rwsem);
3119 mdsc->snap_realms = RB_ROOT;
3120 INIT_LIST_HEAD(&mdsc->snap_empty);
3121 spin_lock_init(&mdsc->snap_empty_lock);
3123 mdsc->request_tree = RB_ROOT;
3124 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3125 mdsc->last_renew_caps = jiffies;
3126 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3127 spin_lock_init(&mdsc->cap_delay_lock);
3128 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3129 spin_lock_init(&mdsc->snap_flush_lock);
3130 mdsc->cap_flush_seq = 0;
3131 INIT_LIST_HEAD(&mdsc->cap_dirty);
3132 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3133 mdsc->num_cap_flushing = 0;
3134 spin_lock_init(&mdsc->cap_dirty_lock);
3135 init_waitqueue_head(&mdsc->cap_flushing_wq);
3136 spin_lock_init(&mdsc->dentry_lru_lock);
3137 INIT_LIST_HEAD(&mdsc->dentry_lru);
3139 ceph_caps_init(mdsc);
3140 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3146 * Wait for safe replies on open mds requests. If we time out, drop
3147 * all requests from the tree to avoid dangling dentry refs.
3149 static void wait_requests(struct ceph_mds_client *mdsc)
3151 struct ceph_mds_request *req;
3152 struct ceph_fs_client *fsc = mdsc->fsc;
3154 mutex_lock(&mdsc->mutex);
3155 if (__get_oldest_req(mdsc)) {
3156 mutex_unlock(&mdsc->mutex);
3158 dout("wait_requests waiting for requests\n");
3159 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3160 fsc->client->options->mount_timeout * HZ);
3162 /* tear down remaining requests */
3163 mutex_lock(&mdsc->mutex);
3164 while ((req = __get_oldest_req(mdsc))) {
3165 dout("wait_requests timed out on tid %llu\n",
3167 __unregister_request(mdsc, req);
3170 mutex_unlock(&mdsc->mutex);
3171 dout("wait_requests done\n");
3175 * called before mount is ro, and before dentries are torn down.
3176 * (hmm, does this still race with new lookups?)
3178 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3180 dout("pre_umount\n");
3184 ceph_flush_dirty_caps(mdsc);
3185 wait_requests(mdsc);
3188 * wait for reply handlers to drop their request refs and
3189 * their inode/dcache refs
3195 * wait for all write mds requests to flush.
3197 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3199 struct ceph_mds_request *req = NULL, *nextreq;
3202 mutex_lock(&mdsc->mutex);
3203 dout("wait_unsafe_requests want %lld\n", want_tid);
3205 req = __get_oldest_req(mdsc);
3206 while (req && req->r_tid <= want_tid) {
3207 /* find next request */
3208 n = rb_next(&req->r_node);
3210 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3213 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3215 ceph_mdsc_get_request(req);
3217 ceph_mdsc_get_request(nextreq);
3218 mutex_unlock(&mdsc->mutex);
3219 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3220 req->r_tid, want_tid);
3221 wait_for_completion(&req->r_safe_completion);
3222 mutex_lock(&mdsc->mutex);
3223 ceph_mdsc_put_request(req);
3225 break; /* next dne before, so we're done! */
3226 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3227 /* next request was removed from tree */
3228 ceph_mdsc_put_request(nextreq);
3231 ceph_mdsc_put_request(nextreq); /* won't go away */
3235 mutex_unlock(&mdsc->mutex);
3236 dout("wait_unsafe_requests done\n");
3239 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3241 u64 want_tid, want_flush;
3243 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3247 mutex_lock(&mdsc->mutex);
3248 want_tid = mdsc->last_tid;
3249 want_flush = mdsc->cap_flush_seq;
3250 mutex_unlock(&mdsc->mutex);
3251 dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3253 ceph_flush_dirty_caps(mdsc);
3255 wait_unsafe_requests(mdsc, want_tid);
3256 wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3260 * true if all sessions are closed, or we force unmount
3262 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3266 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3269 mutex_lock(&mdsc->mutex);
3270 for (i = 0; i < mdsc->max_sessions; i++)
3271 if (mdsc->sessions[i])
3273 mutex_unlock(&mdsc->mutex);
3278 * called after sb is ro.
3280 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3282 struct ceph_mds_session *session;
3284 struct ceph_fs_client *fsc = mdsc->fsc;
3285 unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3287 dout("close_sessions\n");
3289 /* close sessions */
3290 mutex_lock(&mdsc->mutex);
3291 for (i = 0; i < mdsc->max_sessions; i++) {
3292 session = __ceph_lookup_mds_session(mdsc, i);
3295 mutex_unlock(&mdsc->mutex);
3296 mutex_lock(&session->s_mutex);
3297 __close_session(mdsc, session);
3298 mutex_unlock(&session->s_mutex);
3299 ceph_put_mds_session(session);
3300 mutex_lock(&mdsc->mutex);
3302 mutex_unlock(&mdsc->mutex);
3304 dout("waiting for sessions to close\n");
3305 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3308 /* tear down remaining sessions */
3309 mutex_lock(&mdsc->mutex);
3310 for (i = 0; i < mdsc->max_sessions; i++) {
3311 if (mdsc->sessions[i]) {
3312 session = get_session(mdsc->sessions[i]);
3313 __unregister_session(mdsc, session);
3314 mutex_unlock(&mdsc->mutex);
3315 mutex_lock(&session->s_mutex);
3316 remove_session_caps(session);
3317 mutex_unlock(&session->s_mutex);
3318 ceph_put_mds_session(session);
3319 mutex_lock(&mdsc->mutex);
3322 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3323 mutex_unlock(&mdsc->mutex);
3325 ceph_cleanup_empty_realms(mdsc);
3327 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3332 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3335 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3337 ceph_mdsmap_destroy(mdsc->mdsmap);
3338 kfree(mdsc->sessions);
3339 ceph_caps_finalize(mdsc);
3342 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3344 struct ceph_mds_client *mdsc = fsc->mdsc;
3346 dout("mdsc_destroy %p\n", mdsc);
3347 ceph_mdsc_stop(mdsc);
3349 /* flush out any connection work with references to us */
3354 dout("mdsc_destroy %p done\n", mdsc);
3359 * handle mds map update.
3361 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3365 void *p = msg->front.iov_base;
3366 void *end = p + msg->front.iov_len;
3367 struct ceph_mdsmap *newmap, *oldmap;
3368 struct ceph_fsid fsid;
3371 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3372 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3373 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3375 epoch = ceph_decode_32(&p);
3376 maplen = ceph_decode_32(&p);
3377 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3379 /* do we need it? */
3380 ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3381 mutex_lock(&mdsc->mutex);
3382 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3383 dout("handle_map epoch %u <= our %u\n",
3384 epoch, mdsc->mdsmap->m_epoch);
3385 mutex_unlock(&mdsc->mutex);
3389 newmap = ceph_mdsmap_decode(&p, end);
3390 if (IS_ERR(newmap)) {
3391 err = PTR_ERR(newmap);
3395 /* swap into place */
3397 oldmap = mdsc->mdsmap;
3398 mdsc->mdsmap = newmap;
3399 check_new_map(mdsc, newmap, oldmap);
3400 ceph_mdsmap_destroy(oldmap);
3402 mdsc->mdsmap = newmap; /* first mds map */
3404 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3406 __wake_requests(mdsc, &mdsc->waiting_for_map);
3408 mutex_unlock(&mdsc->mutex);
3409 schedule_delayed(mdsc);
3413 mutex_unlock(&mdsc->mutex);
3415 pr_err("error decoding mdsmap %d\n", err);
3419 static struct ceph_connection *con_get(struct ceph_connection *con)
3421 struct ceph_mds_session *s = con->private;
3423 if (get_session(s)) {
3424 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3427 dout("mdsc con_get %p FAIL\n", s);
3431 static void con_put(struct ceph_connection *con)
3433 struct ceph_mds_session *s = con->private;
3435 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3436 ceph_put_mds_session(s);
3440 * if the client is unresponsive for long enough, the mds will kill
3441 * the session entirely.
3443 static void peer_reset(struct ceph_connection *con)
3445 struct ceph_mds_session *s = con->private;
3446 struct ceph_mds_client *mdsc = s->s_mdsc;
3448 pr_warning("mds%d closed our session\n", s->s_mds);
3449 send_mds_reconnect(mdsc, s);
3452 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3454 struct ceph_mds_session *s = con->private;
3455 struct ceph_mds_client *mdsc = s->s_mdsc;
3456 int type = le16_to_cpu(msg->hdr.type);
3458 mutex_lock(&mdsc->mutex);
3459 if (__verify_registered_session(mdsc, s) < 0) {
3460 mutex_unlock(&mdsc->mutex);
3463 mutex_unlock(&mdsc->mutex);
3466 case CEPH_MSG_MDS_MAP:
3467 ceph_mdsc_handle_map(mdsc, msg);
3469 case CEPH_MSG_CLIENT_SESSION:
3470 handle_session(s, msg);
3472 case CEPH_MSG_CLIENT_REPLY:
3473 handle_reply(s, msg);
3475 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3476 handle_forward(mdsc, s, msg);
3478 case CEPH_MSG_CLIENT_CAPS:
3479 ceph_handle_caps(s, msg);
3481 case CEPH_MSG_CLIENT_SNAP:
3482 ceph_handle_snap(mdsc, s, msg);
3484 case CEPH_MSG_CLIENT_LEASE:
3485 handle_lease(mdsc, s, msg);
3489 pr_err("received unknown message type %d %s\n", type,
3490 ceph_msg_type_name(type));
3501 * Note: returned pointer is the address of a structure that's
3502 * managed separately. Caller must *not* attempt to free it.
3504 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3505 int *proto, int force_new)
3507 struct ceph_mds_session *s = con->private;
3508 struct ceph_mds_client *mdsc = s->s_mdsc;
3509 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3510 struct ceph_auth_handshake *auth = &s->s_auth;
3512 if (force_new && auth->authorizer) {
3513 ceph_auth_destroy_authorizer(ac, auth->authorizer);
3514 auth->authorizer = NULL;
3516 if (!auth->authorizer) {
3517 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3520 return ERR_PTR(ret);
3522 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3525 return ERR_PTR(ret);
3527 *proto = ac->protocol;
3533 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3535 struct ceph_mds_session *s = con->private;
3536 struct ceph_mds_client *mdsc = s->s_mdsc;
3537 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3539 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3542 static int invalidate_authorizer(struct ceph_connection *con)
3544 struct ceph_mds_session *s = con->private;
3545 struct ceph_mds_client *mdsc = s->s_mdsc;
3546 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3548 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3550 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3553 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3554 struct ceph_msg_header *hdr, int *skip)
3556 struct ceph_msg *msg;
3557 int type = (int) le16_to_cpu(hdr->type);
3558 int front_len = (int) le32_to_cpu(hdr->front_len);
3564 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3566 pr_err("unable to allocate msg type %d len %d\n",
3574 static const struct ceph_connection_operations mds_con_ops = {
3577 .dispatch = dispatch,
3578 .get_authorizer = get_authorizer,
3579 .verify_authorizer_reply = verify_authorizer_reply,
3580 .invalidate_authorizer = invalidate_authorizer,
3581 .peer_reset = peer_reset,
3582 .alloc_msg = mds_alloc_msg,