1 #include "ceph_debug.h"
3 #include <linux/wait.h>
4 #include <linux/slab.h>
5 #include <linux/sched.h>
7 #include "mds_client.h"
8 #include "mon_client.h"
10 #include "messenger.h"
16 * A cluster of MDS (metadata server) daemons is responsible for
17 * managing the file system namespace (the directory hierarchy and
18 * inodes) and for coordinating shared access to storage. Metadata is
19 * partitioning hierarchically across a number of servers, and that
20 * partition varies over time as the cluster adjusts the distribution
21 * in order to balance load.
23 * The MDS client is primarily responsible to managing synchronous
24 * metadata requests for operations like open, unlink, and so forth.
25 * If there is a MDS failure, we find out about it when we (possibly
26 * request and) receive a new MDS map, and can resubmit affected
29 * For the most part, though, we take advantage of a lossless
30 * communications channel to the MDS, and do not need to worry about
31 * timing out or resubmitting requests.
33 * We maintain a stateful "session" with each MDS we interact with.
34 * Within each session, we sent periodic heartbeat messages to ensure
35 * any capabilities or leases we have been issues remain valid. If
36 * the session times out and goes stale, our leases and capabilities
37 * are no longer valid.
40 static void __wake_requests(struct ceph_mds_client *mdsc,
41 struct list_head *head);
43 const static struct ceph_connection_operations mds_con_ops;
51 * parse individual inode info
53 static int parse_reply_info_in(void **p, void *end,
54 struct ceph_mds_reply_info_in *info)
59 *p += sizeof(struct ceph_mds_reply_inode) +
60 sizeof(*info->in->fragtree.splits) *
61 le32_to_cpu(info->in->fragtree.nsplits);
63 ceph_decode_32_safe(p, end, info->symlink_len, bad);
64 ceph_decode_need(p, end, info->symlink_len, bad);
66 *p += info->symlink_len;
68 ceph_decode_32_safe(p, end, info->xattr_len, bad);
69 ceph_decode_need(p, end, info->xattr_len, bad);
70 info->xattr_data = *p;
71 *p += info->xattr_len;
78 * parse a normal reply, which may contain a (dir+)dentry and/or a
81 static int parse_reply_info_trace(void **p, void *end,
82 struct ceph_mds_reply_info_parsed *info)
86 if (info->head->is_dentry) {
87 err = parse_reply_info_in(p, end, &info->diri);
91 if (unlikely(*p + sizeof(*info->dirfrag) > end))
94 *p += sizeof(*info->dirfrag) +
95 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
96 if (unlikely(*p > end))
99 ceph_decode_32_safe(p, end, info->dname_len, bad);
100 ceph_decode_need(p, end, info->dname_len, bad);
102 *p += info->dname_len;
104 *p += sizeof(*info->dlease);
107 if (info->head->is_target) {
108 err = parse_reply_info_in(p, end, &info->targeti);
113 if (unlikely(*p != end))
120 pr_err("problem parsing mds trace %d\n", err);
125 * parse readdir results
127 static int parse_reply_info_dir(void **p, void *end,
128 struct ceph_mds_reply_info_parsed *info)
134 if (*p + sizeof(*info->dir_dir) > end)
136 *p += sizeof(*info->dir_dir) +
137 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
141 ceph_decode_need(p, end, sizeof(num) + 2, bad);
142 num = ceph_decode_32(p);
143 info->dir_end = ceph_decode_8(p);
144 info->dir_complete = ceph_decode_8(p);
148 /* alloc large array */
150 info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
151 sizeof(*info->dir_dname) +
152 sizeof(*info->dir_dname_len) +
153 sizeof(*info->dir_dlease),
155 if (info->dir_in == NULL) {
159 info->dir_dname = (void *)(info->dir_in + num);
160 info->dir_dname_len = (void *)(info->dir_dname + num);
161 info->dir_dlease = (void *)(info->dir_dname_len + num);
165 ceph_decode_need(p, end, sizeof(u32)*2, bad);
166 info->dir_dname_len[i] = ceph_decode_32(p);
167 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
168 info->dir_dname[i] = *p;
169 *p += info->dir_dname_len[i];
170 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
172 info->dir_dlease[i] = *p;
173 *p += sizeof(struct ceph_mds_reply_lease);
176 err = parse_reply_info_in(p, end, &info->dir_in[i]);
191 pr_err("problem parsing dir contents %d\n", err);
196 * parse entire mds reply
198 static int parse_reply_info(struct ceph_msg *msg,
199 struct ceph_mds_reply_info_parsed *info)
205 info->head = msg->front.iov_base;
206 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
207 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
210 ceph_decode_32_safe(&p, end, len, bad);
212 err = parse_reply_info_trace(&p, p+len, info);
218 ceph_decode_32_safe(&p, end, len, bad);
220 err = parse_reply_info_dir(&p, p+len, info);
226 ceph_decode_32_safe(&p, end, len, bad);
227 info->snapblob_len = len;
238 pr_err("mds parse_reply err %d\n", err);
242 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
251 static const char *session_state_name(int s)
254 case CEPH_MDS_SESSION_NEW: return "new";
255 case CEPH_MDS_SESSION_OPENING: return "opening";
256 case CEPH_MDS_SESSION_OPEN: return "open";
257 case CEPH_MDS_SESSION_HUNG: return "hung";
258 case CEPH_MDS_SESSION_CLOSING: return "closing";
259 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
260 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
261 default: return "???";
265 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
267 if (atomic_inc_not_zero(&s->s_ref)) {
268 dout("mdsc get_session %p %d -> %d\n", s,
269 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
272 dout("mdsc get_session %p 0 -- FAIL", s);
277 void ceph_put_mds_session(struct ceph_mds_session *s)
279 dout("mdsc put_session %p %d -> %d\n", s,
280 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
281 if (atomic_dec_and_test(&s->s_ref)) {
283 s->s_mdsc->client->monc.auth->ops->destroy_authorizer(
284 s->s_mdsc->client->monc.auth, s->s_authorizer);
290 * called under mdsc->mutex
292 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
295 struct ceph_mds_session *session;
297 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
299 session = mdsc->sessions[mds];
300 dout("lookup_mds_session %p %d\n", session,
301 atomic_read(&session->s_ref));
302 get_session(session);
306 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
308 if (mds >= mdsc->max_sessions)
310 return mdsc->sessions[mds];
313 static int __verify_registered_session(struct ceph_mds_client *mdsc,
314 struct ceph_mds_session *s)
316 if (s->s_mds >= mdsc->max_sessions ||
317 mdsc->sessions[s->s_mds] != s)
323 * create+register a new session for given mds.
324 * called under mdsc->mutex.
326 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
329 struct ceph_mds_session *s;
331 s = kzalloc(sizeof(*s), GFP_NOFS);
333 return ERR_PTR(-ENOMEM);
336 s->s_state = CEPH_MDS_SESSION_NEW;
339 mutex_init(&s->s_mutex);
341 ceph_con_init(mdsc->client->msgr, &s->s_con);
342 s->s_con.private = s;
343 s->s_con.ops = &mds_con_ops;
344 s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
345 s->s_con.peer_name.num = cpu_to_le64(mds);
347 spin_lock_init(&s->s_cap_lock);
350 s->s_renew_requested = 0;
352 INIT_LIST_HEAD(&s->s_caps);
355 atomic_set(&s->s_ref, 1);
356 INIT_LIST_HEAD(&s->s_waiting);
357 INIT_LIST_HEAD(&s->s_unsafe);
358 s->s_num_cap_releases = 0;
359 s->s_cap_iterator = NULL;
360 INIT_LIST_HEAD(&s->s_cap_releases);
361 INIT_LIST_HEAD(&s->s_cap_releases_done);
362 INIT_LIST_HEAD(&s->s_cap_flushing);
363 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
365 dout("register_session mds%d\n", mds);
366 if (mds >= mdsc->max_sessions) {
367 int newmax = 1 << get_count_order(mds+1);
368 struct ceph_mds_session **sa;
370 dout("register_session realloc to %d\n", newmax);
371 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
374 if (mdsc->sessions) {
375 memcpy(sa, mdsc->sessions,
376 mdsc->max_sessions * sizeof(void *));
377 kfree(mdsc->sessions);
380 mdsc->max_sessions = newmax;
382 mdsc->sessions[mds] = s;
383 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
385 ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
391 return ERR_PTR(-ENOMEM);
395 * called under mdsc->mutex
397 static void __unregister_session(struct ceph_mds_client *mdsc,
398 struct ceph_mds_session *s)
400 dout("__unregister_session mds%d %p\n", s->s_mds, s);
401 BUG_ON(mdsc->sessions[s->s_mds] != s);
402 mdsc->sessions[s->s_mds] = NULL;
403 ceph_con_close(&s->s_con);
404 ceph_put_mds_session(s);
408 * drop session refs in request.
410 * should be last request ref, or hold mdsc->mutex
412 static void put_request_session(struct ceph_mds_request *req)
414 if (req->r_session) {
415 ceph_put_mds_session(req->r_session);
416 req->r_session = NULL;
420 void ceph_mdsc_release_request(struct kref *kref)
422 struct ceph_mds_request *req = container_of(kref,
423 struct ceph_mds_request,
426 ceph_msg_put(req->r_request);
428 ceph_msg_put(req->r_reply);
429 destroy_reply_info(&req->r_reply_info);
432 ceph_put_cap_refs(ceph_inode(req->r_inode),
436 if (req->r_locked_dir)
437 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
439 if (req->r_target_inode)
440 iput(req->r_target_inode);
443 if (req->r_old_dentry) {
445 ceph_inode(req->r_old_dentry->d_parent->d_inode),
447 dput(req->r_old_dentry);
451 put_request_session(req);
452 ceph_unreserve_caps(&req->r_caps_reservation);
457 * lookup session, bump ref if found.
459 * called under mdsc->mutex.
461 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
464 struct ceph_mds_request *req;
465 struct rb_node *n = mdsc->request_tree.rb_node;
468 req = rb_entry(n, struct ceph_mds_request, r_node);
469 if (tid < req->r_tid)
471 else if (tid > req->r_tid)
474 ceph_mdsc_get_request(req);
481 static void __insert_request(struct ceph_mds_client *mdsc,
482 struct ceph_mds_request *new)
484 struct rb_node **p = &mdsc->request_tree.rb_node;
485 struct rb_node *parent = NULL;
486 struct ceph_mds_request *req = NULL;
490 req = rb_entry(parent, struct ceph_mds_request, r_node);
491 if (new->r_tid < req->r_tid)
493 else if (new->r_tid > req->r_tid)
499 rb_link_node(&new->r_node, parent, p);
500 rb_insert_color(&new->r_node, &mdsc->request_tree);
504 * Register an in-flight request, and assign a tid. Link to directory
505 * are modifying (if any).
507 * Called under mdsc->mutex.
509 static void __register_request(struct ceph_mds_client *mdsc,
510 struct ceph_mds_request *req,
513 req->r_tid = ++mdsc->last_tid;
515 ceph_reserve_caps(&req->r_caps_reservation, req->r_num_caps);
516 dout("__register_request %p tid %lld\n", req, req->r_tid);
517 ceph_mdsc_get_request(req);
518 __insert_request(mdsc, req);
521 struct ceph_inode_info *ci = ceph_inode(dir);
523 spin_lock(&ci->i_unsafe_lock);
524 req->r_unsafe_dir = dir;
525 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
526 spin_unlock(&ci->i_unsafe_lock);
530 static void __unregister_request(struct ceph_mds_client *mdsc,
531 struct ceph_mds_request *req)
533 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
534 rb_erase(&req->r_node, &mdsc->request_tree);
535 RB_CLEAR_NODE(&req->r_node);
537 if (req->r_unsafe_dir) {
538 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
540 spin_lock(&ci->i_unsafe_lock);
541 list_del_init(&req->r_unsafe_dir_item);
542 spin_unlock(&ci->i_unsafe_lock);
545 ceph_mdsc_put_request(req);
549 * Choose mds to send request to next. If there is a hint set in the
550 * request (e.g., due to a prior forward hint from the mds), use that.
551 * Otherwise, consult frag tree and/or caps to identify the
552 * appropriate mds. If all else fails, choose randomly.
554 * Called under mdsc->mutex.
556 static int __choose_mds(struct ceph_mds_client *mdsc,
557 struct ceph_mds_request *req)
560 struct ceph_inode_info *ci;
561 struct ceph_cap *cap;
562 int mode = req->r_direct_mode;
564 u32 hash = req->r_direct_hash;
565 bool is_hash = req->r_direct_is_hash;
568 * is there a specific mds we should try? ignore hint if we have
569 * no session and the mds is not up (active or recovering).
571 if (req->r_resend_mds >= 0 &&
572 (__have_session(mdsc, req->r_resend_mds) ||
573 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
574 dout("choose_mds using resend_mds mds%d\n",
576 return req->r_resend_mds;
579 if (mode == USE_RANDOM_MDS)
584 inode = req->r_inode;
585 } else if (req->r_dentry) {
586 if (req->r_dentry->d_inode) {
587 inode = req->r_dentry->d_inode;
589 inode = req->r_dentry->d_parent->d_inode;
590 hash = req->r_dentry->d_name.hash;
594 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
598 ci = ceph_inode(inode);
600 if (is_hash && S_ISDIR(inode->i_mode)) {
601 struct ceph_inode_frag frag;
604 ceph_choose_frag(ci, hash, &frag, &found);
606 if (mode == USE_ANY_MDS && frag.ndist > 0) {
609 /* choose a random replica */
610 get_random_bytes(&r, 1);
613 dout("choose_mds %p %llx.%llx "
614 "frag %u mds%d (%d/%d)\n",
615 inode, ceph_vinop(inode),
621 /* since this file/dir wasn't known to be
622 * replicated, then we want to look for the
623 * authoritative mds. */
626 /* choose auth mds */
628 dout("choose_mds %p %llx.%llx "
629 "frag %u mds%d (auth)\n",
630 inode, ceph_vinop(inode), frag.frag, mds);
636 spin_lock(&inode->i_lock);
638 if (mode == USE_AUTH_MDS)
639 cap = ci->i_auth_cap;
640 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
641 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
643 spin_unlock(&inode->i_lock);
646 mds = cap->session->s_mds;
647 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
648 inode, ceph_vinop(inode), mds,
649 cap == ci->i_auth_cap ? "auth " : "", cap);
650 spin_unlock(&inode->i_lock);
654 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
655 dout("choose_mds chose random mds%d\n", mds);
663 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
665 struct ceph_msg *msg;
666 struct ceph_mds_session_head *h;
668 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), 0, 0, NULL);
670 pr_err("create_session_msg ENOMEM creating msg\n");
671 return ERR_PTR(PTR_ERR(msg));
673 h = msg->front.iov_base;
674 h->op = cpu_to_le32(op);
675 h->seq = cpu_to_le64(seq);
680 * send session open request.
682 * called under mdsc->mutex
684 static int __open_session(struct ceph_mds_client *mdsc,
685 struct ceph_mds_session *session)
687 struct ceph_msg *msg;
689 int mds = session->s_mds;
692 /* wait for mds to go active? */
693 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
694 dout("open_session to mds%d (%s)\n", mds,
695 ceph_mds_state_name(mstate));
696 session->s_state = CEPH_MDS_SESSION_OPENING;
697 session->s_renew_requested = jiffies;
699 /* send connect message */
700 msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
705 ceph_con_send(&session->s_con, msg);
716 * Free preallocated cap messages assigned to this session
718 static void cleanup_cap_releases(struct ceph_mds_session *session)
720 struct ceph_msg *msg;
722 spin_lock(&session->s_cap_lock);
723 while (!list_empty(&session->s_cap_releases)) {
724 msg = list_first_entry(&session->s_cap_releases,
725 struct ceph_msg, list_head);
726 list_del_init(&msg->list_head);
729 while (!list_empty(&session->s_cap_releases_done)) {
730 msg = list_first_entry(&session->s_cap_releases_done,
731 struct ceph_msg, list_head);
732 list_del_init(&msg->list_head);
735 spin_unlock(&session->s_cap_lock);
739 * Helper to safely iterate over all caps associated with a session, with
740 * special care taken to handle a racing __ceph_remove_cap().
742 * Caller must hold session s_mutex.
744 static int iterate_session_caps(struct ceph_mds_session *session,
745 int (*cb)(struct inode *, struct ceph_cap *,
749 struct ceph_cap *cap;
750 struct inode *inode, *last_inode = NULL;
751 struct ceph_cap *old_cap = NULL;
754 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
755 spin_lock(&session->s_cap_lock);
756 p = session->s_caps.next;
757 while (p != &session->s_caps) {
758 cap = list_entry(p, struct ceph_cap, session_caps);
759 inode = igrab(&cap->ci->vfs_inode);
764 session->s_cap_iterator = cap;
765 spin_unlock(&session->s_cap_lock);
772 ceph_put_cap(old_cap);
776 ret = cb(inode, cap, arg);
779 spin_lock(&session->s_cap_lock);
781 if (cap->ci == NULL) {
782 dout("iterate_session_caps finishing cap %p removal\n",
784 BUG_ON(cap->session != session);
785 list_del_init(&cap->session_caps);
786 session->s_nr_caps--;
788 old_cap = cap; /* put_cap it w/o locks held */
795 session->s_cap_iterator = NULL;
796 spin_unlock(&session->s_cap_lock);
801 ceph_put_cap(old_cap);
806 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
809 struct ceph_inode_info *ci = ceph_inode(inode);
810 dout("removing cap %p, ci is %p, inode is %p\n",
811 cap, ci, &ci->vfs_inode);
812 ceph_remove_cap(cap);
817 * caller must hold session s_mutex
819 static void remove_session_caps(struct ceph_mds_session *session)
821 dout("remove_session_caps on %p\n", session);
822 iterate_session_caps(session, remove_session_caps_cb, NULL);
823 BUG_ON(session->s_nr_caps > 0);
824 cleanup_cap_releases(session);
828 * wake up any threads waiting on this session's caps. if the cap is
829 * old (didn't get renewed on the client reconnect), remove it now.
831 * caller must hold s_mutex.
833 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
836 struct ceph_inode_info *ci = ceph_inode(inode);
838 wake_up(&ci->i_cap_wq);
840 spin_lock(&inode->i_lock);
841 ci->i_wanted_max_size = 0;
842 ci->i_requested_max_size = 0;
843 spin_unlock(&inode->i_lock);
848 static void wake_up_session_caps(struct ceph_mds_session *session,
851 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
852 iterate_session_caps(session, wake_up_session_cb,
853 (void *)(unsigned long)reconnect);
857 * Send periodic message to MDS renewing all currently held caps. The
858 * ack will reset the expiration for all caps from this session.
860 * caller holds s_mutex
862 static int send_renew_caps(struct ceph_mds_client *mdsc,
863 struct ceph_mds_session *session)
865 struct ceph_msg *msg;
868 if (time_after_eq(jiffies, session->s_cap_ttl) &&
869 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
870 pr_info("mds%d caps stale\n", session->s_mds);
871 session->s_renew_requested = jiffies;
873 /* do not try to renew caps until a recovering mds has reconnected
874 * with its clients. */
875 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
876 if (state < CEPH_MDS_STATE_RECONNECT) {
877 dout("send_renew_caps ignoring mds%d (%s)\n",
878 session->s_mds, ceph_mds_state_name(state));
882 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
883 ceph_mds_state_name(state));
884 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
885 ++session->s_renew_seq);
888 ceph_con_send(&session->s_con, msg);
893 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
895 * Called under session->s_mutex
897 static void renewed_caps(struct ceph_mds_client *mdsc,
898 struct ceph_mds_session *session, int is_renew)
903 spin_lock(&session->s_cap_lock);
904 was_stale = is_renew && (session->s_cap_ttl == 0 ||
905 time_after_eq(jiffies, session->s_cap_ttl));
907 session->s_cap_ttl = session->s_renew_requested +
908 mdsc->mdsmap->m_session_timeout*HZ;
911 if (time_before(jiffies, session->s_cap_ttl)) {
912 pr_info("mds%d caps renewed\n", session->s_mds);
915 pr_info("mds%d caps still stale\n", session->s_mds);
918 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
919 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
920 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
921 spin_unlock(&session->s_cap_lock);
924 wake_up_session_caps(session, 0);
928 * send a session close request
930 static int request_close_session(struct ceph_mds_client *mdsc,
931 struct ceph_mds_session *session)
933 struct ceph_msg *msg;
936 dout("request_close_session mds%d state %s seq %lld\n",
937 session->s_mds, session_state_name(session->s_state),
939 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
943 ceph_con_send(&session->s_con, msg);
948 * Called with s_mutex held.
950 static int __close_session(struct ceph_mds_client *mdsc,
951 struct ceph_mds_session *session)
953 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
955 session->s_state = CEPH_MDS_SESSION_CLOSING;
956 return request_close_session(mdsc, session);
962 * Because we can't cache an inode without one or more caps, we do
963 * this indirectly: if a cap is unused, we prune its aliases, at which
964 * point the inode will hopefully get dropped to.
966 * Yes, this is a bit sloppy. Our only real goal here is to respond to
967 * memory pressure from the MDS, though, so it needn't be perfect.
969 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
971 struct ceph_mds_session *session = arg;
972 struct ceph_inode_info *ci = ceph_inode(inode);
973 int used, oissued, mine;
975 if (session->s_trim_caps <= 0)
978 spin_lock(&inode->i_lock);
979 mine = cap->issued | cap->implemented;
980 used = __ceph_caps_used(ci);
981 oissued = __ceph_caps_issued_other(ci, cap);
983 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
984 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
985 ceph_cap_string(used));
986 if (ci->i_dirty_caps)
987 goto out; /* dirty caps */
988 if ((used & ~oissued) & mine)
989 goto out; /* we need these caps */
991 session->s_trim_caps--;
993 /* we aren't the only cap.. just remove us */
994 __ceph_remove_cap(cap);
996 /* try to drop referring dentries */
997 spin_unlock(&inode->i_lock);
998 d_prune_aliases(inode);
999 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1000 inode, cap, atomic_read(&inode->i_count));
1005 spin_unlock(&inode->i_lock);
1010 * Trim session cap count down to some max number.
1012 static int trim_caps(struct ceph_mds_client *mdsc,
1013 struct ceph_mds_session *session,
1016 int trim_caps = session->s_nr_caps - max_caps;
1018 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1019 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1020 if (trim_caps > 0) {
1021 session->s_trim_caps = trim_caps;
1022 iterate_session_caps(session, trim_caps_cb, session);
1023 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1024 session->s_mds, session->s_nr_caps, max_caps,
1025 trim_caps - session->s_trim_caps);
1026 session->s_trim_caps = 0;
1032 * Allocate cap_release messages. If there is a partially full message
1033 * in the queue, try to allocate enough to cover it's remainder, so that
1034 * we can send it immediately.
1036 * Called under s_mutex.
1038 static int add_cap_releases(struct ceph_mds_client *mdsc,
1039 struct ceph_mds_session *session,
1042 struct ceph_msg *msg;
1043 struct ceph_mds_cap_release *head;
1047 extra = mdsc->client->mount_args->cap_release_safety;
1049 spin_lock(&session->s_cap_lock);
1051 if (!list_empty(&session->s_cap_releases)) {
1052 msg = list_first_entry(&session->s_cap_releases,
1055 head = msg->front.iov_base;
1056 extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1059 while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1060 spin_unlock(&session->s_cap_lock);
1061 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1065 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1066 (int)msg->front.iov_len);
1067 head = msg->front.iov_base;
1068 head->num = cpu_to_le32(0);
1069 msg->front.iov_len = sizeof(*head);
1070 spin_lock(&session->s_cap_lock);
1071 list_add(&msg->list_head, &session->s_cap_releases);
1072 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1075 if (!list_empty(&session->s_cap_releases)) {
1076 msg = list_first_entry(&session->s_cap_releases,
1079 head = msg->front.iov_base;
1081 dout(" queueing non-full %p (%d)\n", msg,
1082 le32_to_cpu(head->num));
1083 list_move_tail(&msg->list_head,
1084 &session->s_cap_releases_done);
1085 session->s_num_cap_releases -=
1086 CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1090 spin_unlock(&session->s_cap_lock);
1096 * flush all dirty inode data to disk.
1098 * returns true if we've flushed through want_flush_seq
1100 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1104 dout("check_cap_flush want %lld\n", want_flush_seq);
1105 mutex_lock(&mdsc->mutex);
1106 for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1107 struct ceph_mds_session *session = mdsc->sessions[mds];
1111 get_session(session);
1112 mutex_unlock(&mdsc->mutex);
1114 mutex_lock(&session->s_mutex);
1115 if (!list_empty(&session->s_cap_flushing)) {
1116 struct ceph_inode_info *ci =
1117 list_entry(session->s_cap_flushing.next,
1118 struct ceph_inode_info,
1120 struct inode *inode = &ci->vfs_inode;
1122 spin_lock(&inode->i_lock);
1123 if (ci->i_cap_flush_seq <= want_flush_seq) {
1124 dout("check_cap_flush still flushing %p "
1125 "seq %lld <= %lld to mds%d\n", inode,
1126 ci->i_cap_flush_seq, want_flush_seq,
1130 spin_unlock(&inode->i_lock);
1132 mutex_unlock(&session->s_mutex);
1133 ceph_put_mds_session(session);
1137 mutex_lock(&mdsc->mutex);
1140 mutex_unlock(&mdsc->mutex);
1141 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1146 * called under s_mutex
1148 static void send_cap_releases(struct ceph_mds_client *mdsc,
1149 struct ceph_mds_session *session)
1151 struct ceph_msg *msg;
1153 dout("send_cap_releases mds%d\n", session->s_mds);
1155 spin_lock(&session->s_cap_lock);
1156 if (list_empty(&session->s_cap_releases_done))
1158 msg = list_first_entry(&session->s_cap_releases_done,
1159 struct ceph_msg, list_head);
1160 list_del_init(&msg->list_head);
1161 spin_unlock(&session->s_cap_lock);
1162 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1163 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1164 ceph_con_send(&session->s_con, msg);
1166 spin_unlock(&session->s_cap_lock);
1174 * Create an mds request.
1176 struct ceph_mds_request *
1177 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1179 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1182 return ERR_PTR(-ENOMEM);
1184 mutex_init(&req->r_fill_mutex);
1185 req->r_started = jiffies;
1186 req->r_resend_mds = -1;
1187 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1189 kref_init(&req->r_kref);
1190 INIT_LIST_HEAD(&req->r_wait);
1191 init_completion(&req->r_completion);
1192 init_completion(&req->r_safe_completion);
1193 INIT_LIST_HEAD(&req->r_unsafe_item);
1196 req->r_direct_mode = mode;
1201 * return oldest (lowest) request, tid in request tree, 0 if none.
1203 * called under mdsc->mutex.
1205 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1207 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1209 return rb_entry(rb_first(&mdsc->request_tree),
1210 struct ceph_mds_request, r_node);
1213 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1215 struct ceph_mds_request *req = __get_oldest_req(mdsc);
1223 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1224 * on build_path_from_dentry in fs/cifs/dir.c.
1226 * If @stop_on_nosnap, generate path relative to the first non-snapped
1229 * Encode hidden .snap dirs as a double /, i.e.
1230 * foo/.snap/bar -> foo//bar
1232 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1235 struct dentry *temp;
1240 return ERR_PTR(-EINVAL);
1244 for (temp = dentry; !IS_ROOT(temp);) {
1245 struct inode *inode = temp->d_inode;
1246 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1247 len++; /* slash only */
1248 else if (stop_on_nosnap && inode &&
1249 ceph_snap(inode) == CEPH_NOSNAP)
1252 len += 1 + temp->d_name.len;
1253 temp = temp->d_parent;
1255 pr_err("build_path_dentry corrupt dentry %p\n", dentry);
1256 return ERR_PTR(-EINVAL);
1260 len--; /* no leading '/' */
1262 path = kmalloc(len+1, GFP_NOFS);
1264 return ERR_PTR(-ENOMEM);
1266 path[pos] = 0; /* trailing null */
1267 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1268 struct inode *inode = temp->d_inode;
1270 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1271 dout("build_path path+%d: %p SNAPDIR\n",
1273 } else if (stop_on_nosnap && inode &&
1274 ceph_snap(inode) == CEPH_NOSNAP) {
1277 pos -= temp->d_name.len;
1280 strncpy(path + pos, temp->d_name.name,
1285 temp = temp->d_parent;
1287 pr_err("build_path corrupt dentry\n");
1289 return ERR_PTR(-EINVAL);
1293 pr_err("build_path did not end path lookup where "
1294 "expected, namelen is %d, pos is %d\n", len, pos);
1295 /* presumably this is only possible if racing with a
1296 rename of one of the parent directories (we can not
1297 lock the dentries above us to prevent this, but
1298 retrying should be harmless) */
1303 *base = ceph_ino(temp->d_inode);
1305 dout("build_path on %p %d built %llx '%.*s'\n",
1306 dentry, atomic_read(&dentry->d_count), *base, len, path);
1310 static int build_dentry_path(struct dentry *dentry,
1311 const char **ppath, int *ppathlen, u64 *pino,
1316 if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1317 *pino = ceph_ino(dentry->d_parent->d_inode);
1318 *ppath = dentry->d_name.name;
1319 *ppathlen = dentry->d_name.len;
1322 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1324 return PTR_ERR(path);
1330 static int build_inode_path(struct inode *inode,
1331 const char **ppath, int *ppathlen, u64 *pino,
1334 struct dentry *dentry;
1337 if (ceph_snap(inode) == CEPH_NOSNAP) {
1338 *pino = ceph_ino(inode);
1342 dentry = d_find_alias(inode);
1343 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1346 return PTR_ERR(path);
1353 * request arguments may be specified via an inode *, a dentry *, or
1354 * an explicit ino+path.
1356 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1357 const char *rpath, u64 rino,
1358 const char **ppath, int *pathlen,
1359 u64 *ino, int *freepath)
1364 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1365 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1367 } else if (rdentry) {
1368 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1369 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1374 *pathlen = strlen(rpath);
1375 dout(" path %.*s\n", *pathlen, rpath);
1382 * called under mdsc->mutex
1384 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1385 struct ceph_mds_request *req,
1388 struct ceph_msg *msg;
1389 struct ceph_mds_request_head *head;
1390 const char *path1 = NULL;
1391 const char *path2 = NULL;
1392 u64 ino1 = 0, ino2 = 0;
1393 int pathlen1 = 0, pathlen2 = 0;
1394 int freepath1 = 0, freepath2 = 0;
1400 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1401 req->r_path1, req->r_ino1.ino,
1402 &path1, &pathlen1, &ino1, &freepath1);
1408 ret = set_request_path_attr(NULL, req->r_old_dentry,
1409 req->r_path2, req->r_ino2.ino,
1410 &path2, &pathlen2, &ino2, &freepath2);
1416 len = sizeof(*head) +
1417 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1419 /* calculate (max) length for cap releases */
1420 len += sizeof(struct ceph_mds_request_release) *
1421 (!!req->r_inode_drop + !!req->r_dentry_drop +
1422 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1423 if (req->r_dentry_drop)
1424 len += req->r_dentry->d_name.len;
1425 if (req->r_old_dentry_drop)
1426 len += req->r_old_dentry->d_name.len;
1428 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, 0, 0, NULL);
1432 msg->hdr.tid = cpu_to_le64(req->r_tid);
1434 head = msg->front.iov_base;
1435 p = msg->front.iov_base + sizeof(*head);
1436 end = msg->front.iov_base + msg->front.iov_len;
1438 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1439 head->op = cpu_to_le32(req->r_op);
1440 head->caller_uid = cpu_to_le32(current_fsuid());
1441 head->caller_gid = cpu_to_le32(current_fsgid());
1442 head->args = req->r_args;
1444 ceph_encode_filepath(&p, end, ino1, path1);
1445 ceph_encode_filepath(&p, end, ino2, path2);
1449 if (req->r_inode_drop)
1450 releases += ceph_encode_inode_release(&p,
1451 req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1452 mds, req->r_inode_drop, req->r_inode_unless, 0);
1453 if (req->r_dentry_drop)
1454 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1455 mds, req->r_dentry_drop, req->r_dentry_unless);
1456 if (req->r_old_dentry_drop)
1457 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1458 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1459 if (req->r_old_inode_drop)
1460 releases += ceph_encode_inode_release(&p,
1461 req->r_old_dentry->d_inode,
1462 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1463 head->num_releases = cpu_to_le16(releases);
1466 msg->front.iov_len = p - msg->front.iov_base;
1467 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1469 msg->pages = req->r_pages;
1470 msg->nr_pages = req->r_num_pages;
1471 msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1472 msg->hdr.data_off = cpu_to_le16(0);
1476 kfree((char *)path2);
1479 kfree((char *)path1);
1485 * called under mdsc->mutex if error, under no mutex if
1488 static void complete_request(struct ceph_mds_client *mdsc,
1489 struct ceph_mds_request *req)
1491 if (req->r_callback)
1492 req->r_callback(mdsc, req);
1494 complete(&req->r_completion);
1498 * called under mdsc->mutex
1500 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1501 struct ceph_mds_request *req,
1504 struct ceph_mds_request_head *rhead;
1505 struct ceph_msg *msg;
1510 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1511 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1513 if (req->r_request) {
1514 ceph_msg_put(req->r_request);
1515 req->r_request = NULL;
1517 msg = create_request_message(mdsc, req, mds);
1519 req->r_err = PTR_ERR(msg);
1520 complete_request(mdsc, req);
1521 return -PTR_ERR(msg);
1523 req->r_request = msg;
1525 rhead = msg->front.iov_base;
1526 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1527 if (req->r_got_unsafe)
1528 flags |= CEPH_MDS_FLAG_REPLAY;
1529 if (req->r_locked_dir)
1530 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1531 rhead->flags = cpu_to_le32(flags);
1532 rhead->num_fwd = req->r_num_fwd;
1533 rhead->num_retry = req->r_attempts - 1;
1535 dout(" r_locked_dir = %p\n", req->r_locked_dir);
1537 if (req->r_target_inode && req->r_got_unsafe)
1538 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1545 * send request, or put it on the appropriate wait list.
1547 static int __do_request(struct ceph_mds_client *mdsc,
1548 struct ceph_mds_request *req)
1550 struct ceph_mds_session *session = NULL;
1554 if (req->r_err || req->r_got_result)
1557 if (req->r_timeout &&
1558 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1559 dout("do_request timed out\n");
1564 mds = __choose_mds(mdsc, req);
1566 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1567 dout("do_request no mds or not active, waiting for map\n");
1568 list_add(&req->r_wait, &mdsc->waiting_for_map);
1572 /* get, open session */
1573 session = __ceph_lookup_mds_session(mdsc, mds);
1575 session = register_session(mdsc, mds);
1576 if (IS_ERR(session)) {
1577 err = PTR_ERR(session);
1581 dout("do_request mds%d session %p state %s\n", mds, session,
1582 session_state_name(session->s_state));
1583 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1584 session->s_state != CEPH_MDS_SESSION_HUNG) {
1585 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1586 session->s_state == CEPH_MDS_SESSION_CLOSING)
1587 __open_session(mdsc, session);
1588 list_add(&req->r_wait, &session->s_waiting);
1593 req->r_session = get_session(session);
1594 req->r_resend_mds = -1; /* forget any previous mds hint */
1596 if (req->r_request_started == 0) /* note request start time */
1597 req->r_request_started = jiffies;
1599 err = __prepare_send_request(mdsc, req, mds);
1601 ceph_msg_get(req->r_request);
1602 ceph_con_send(&session->s_con, req->r_request);
1606 ceph_put_mds_session(session);
1612 complete_request(mdsc, req);
1617 * called under mdsc->mutex
1619 static void __wake_requests(struct ceph_mds_client *mdsc,
1620 struct list_head *head)
1622 struct ceph_mds_request *req, *nreq;
1624 list_for_each_entry_safe(req, nreq, head, r_wait) {
1625 list_del_init(&req->r_wait);
1626 __do_request(mdsc, req);
1631 * Wake up threads with requests pending for @mds, so that they can
1632 * resubmit their requests to a possibly different mds. If @all is set,
1633 * wake up if their requests has been forwarded to @mds, too.
1635 static void kick_requests(struct ceph_mds_client *mdsc, int mds, int all)
1637 struct ceph_mds_request *req;
1640 dout("kick_requests mds%d\n", mds);
1641 for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1642 req = rb_entry(p, struct ceph_mds_request, r_node);
1643 if (req->r_got_unsafe)
1645 if (req->r_session &&
1646 req->r_session->s_mds == mds) {
1647 dout(" kicking tid %llu\n", req->r_tid);
1648 put_request_session(req);
1649 __do_request(mdsc, req);
1654 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1655 struct ceph_mds_request *req)
1657 dout("submit_request on %p\n", req);
1658 mutex_lock(&mdsc->mutex);
1659 __register_request(mdsc, req, NULL);
1660 __do_request(mdsc, req);
1661 mutex_unlock(&mdsc->mutex);
1665 * Synchrously perform an mds request. Take care of all of the
1666 * session setup, forwarding, retry details.
1668 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1670 struct ceph_mds_request *req)
1674 dout("do_request on %p\n", req);
1676 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1678 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1679 if (req->r_locked_dir)
1680 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1681 if (req->r_old_dentry)
1683 ceph_inode(req->r_old_dentry->d_parent->d_inode),
1687 mutex_lock(&mdsc->mutex);
1688 __register_request(mdsc, req, dir);
1689 __do_request(mdsc, req);
1693 __unregister_request(mdsc, req);
1694 dout("do_request early error %d\n", err);
1699 mutex_unlock(&mdsc->mutex);
1700 dout("do_request waiting\n");
1701 if (req->r_timeout) {
1702 err = (long)wait_for_completion_interruptible_timeout(
1703 &req->r_completion, req->r_timeout);
1707 err = wait_for_completion_interruptible(&req->r_completion);
1709 dout("do_request waited, got %d\n", err);
1710 mutex_lock(&mdsc->mutex);
1712 /* only abort if we didn't race with a real reply */
1713 if (req->r_got_result) {
1714 err = le32_to_cpu(req->r_reply_info.head->result);
1715 } else if (err < 0) {
1716 dout("aborted request %lld with %d\n", req->r_tid, err);
1719 * ensure we aren't running concurrently with
1720 * ceph_fill_trace or ceph_readdir_prepopulate, which
1721 * rely on locks (dir mutex) held by our caller.
1723 mutex_lock(&req->r_fill_mutex);
1725 req->r_aborted = true;
1726 mutex_unlock(&req->r_fill_mutex);
1728 if (req->r_locked_dir &&
1729 (req->r_op & CEPH_MDS_OP_WRITE)) {
1730 struct ceph_inode_info *ci =
1731 ceph_inode(req->r_locked_dir);
1733 dout("aborted, clearing I_COMPLETE on %p, leases\n",
1735 spin_lock(&req->r_locked_dir->i_lock);
1736 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1737 ci->i_release_count++;
1738 spin_unlock(&req->r_locked_dir->i_lock);
1741 ceph_invalidate_dentry_lease(req->r_dentry);
1742 if (req->r_old_dentry)
1743 ceph_invalidate_dentry_lease(req->r_old_dentry);
1750 mutex_unlock(&mdsc->mutex);
1751 dout("do_request %p done, result %d\n", req, err);
1758 * We take the session mutex and parse and process the reply immediately.
1759 * This preserves the logical ordering of replies, capabilities, etc., sent
1760 * by the MDS as they are applied to our local cache.
1762 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1764 struct ceph_mds_client *mdsc = session->s_mdsc;
1765 struct ceph_mds_request *req;
1766 struct ceph_mds_reply_head *head = msg->front.iov_base;
1767 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
1770 int mds = session->s_mds;
1772 if (msg->front.iov_len < sizeof(*head)) {
1773 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1778 /* get request, session */
1779 tid = le64_to_cpu(msg->hdr.tid);
1780 mutex_lock(&mdsc->mutex);
1781 req = __lookup_request(mdsc, tid);
1783 dout("handle_reply on unknown tid %llu\n", tid);
1784 mutex_unlock(&mdsc->mutex);
1787 dout("handle_reply %p\n", req);
1789 /* correct session? */
1790 if (req->r_session != session) {
1791 pr_err("mdsc_handle_reply got %llu on session mds%d"
1792 " not mds%d\n", tid, session->s_mds,
1793 req->r_session ? req->r_session->s_mds : -1);
1794 mutex_unlock(&mdsc->mutex);
1799 if ((req->r_got_unsafe && !head->safe) ||
1800 (req->r_got_safe && head->safe)) {
1801 pr_warning("got a dup %s reply on %llu from mds%d\n",
1802 head->safe ? "safe" : "unsafe", tid, mds);
1803 mutex_unlock(&mdsc->mutex);
1807 result = le32_to_cpu(head->result);
1810 * Tolerate 2 consecutive ESTALEs from the same mds.
1811 * FIXME: we should be looking at the cap migrate_seq.
1813 if (result == -ESTALE) {
1814 req->r_direct_mode = USE_AUTH_MDS;
1816 if (req->r_num_stale <= 2) {
1817 __do_request(mdsc, req);
1818 mutex_unlock(&mdsc->mutex);
1822 req->r_num_stale = 0;
1826 req->r_got_safe = true;
1827 __unregister_request(mdsc, req);
1828 complete(&req->r_safe_completion);
1830 if (req->r_got_unsafe) {
1832 * We already handled the unsafe response, now do the
1833 * cleanup. No need to examine the response; the MDS
1834 * doesn't include any result info in the safe
1835 * response. And even if it did, there is nothing
1836 * useful we could do with a revised return value.
1838 dout("got safe reply %llu, mds%d\n", tid, mds);
1839 list_del_init(&req->r_unsafe_item);
1841 /* last unsafe request during umount? */
1842 if (mdsc->stopping && !__get_oldest_req(mdsc))
1843 complete(&mdsc->safe_umount_waiters);
1844 mutex_unlock(&mdsc->mutex);
1848 req->r_got_unsafe = true;
1849 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1852 dout("handle_reply tid %lld result %d\n", tid, result);
1853 rinfo = &req->r_reply_info;
1854 err = parse_reply_info(msg, rinfo);
1855 mutex_unlock(&mdsc->mutex);
1857 mutex_lock(&session->s_mutex);
1859 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1865 if (rinfo->snapblob_len) {
1866 down_write(&mdsc->snap_rwsem);
1867 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1868 rinfo->snapblob + rinfo->snapblob_len,
1869 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1870 downgrade_write(&mdsc->snap_rwsem);
1872 down_read(&mdsc->snap_rwsem);
1875 /* insert trace into our cache */
1876 mutex_lock(&req->r_fill_mutex);
1877 err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1879 if (result == 0 && rinfo->dir_nr)
1880 ceph_readdir_prepopulate(req, req->r_session);
1881 ceph_unreserve_caps(&req->r_caps_reservation);
1883 mutex_unlock(&req->r_fill_mutex);
1885 up_read(&mdsc->snap_rwsem);
1887 mutex_lock(&mdsc->mutex);
1888 if (!req->r_aborted) {
1894 req->r_got_result = true;
1897 dout("reply arrived after request %lld was aborted\n", tid);
1899 mutex_unlock(&mdsc->mutex);
1901 add_cap_releases(mdsc, req->r_session, -1);
1902 mutex_unlock(&session->s_mutex);
1904 /* kick calling process */
1905 complete_request(mdsc, req);
1907 ceph_mdsc_put_request(req);
1914 * handle mds notification that our request has been forwarded.
1916 static void handle_forward(struct ceph_mds_client *mdsc,
1917 struct ceph_mds_session *session,
1918 struct ceph_msg *msg)
1920 struct ceph_mds_request *req;
1921 u64 tid = le64_to_cpu(msg->hdr.tid);
1925 void *p = msg->front.iov_base;
1926 void *end = p + msg->front.iov_len;
1928 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1929 next_mds = ceph_decode_32(&p);
1930 fwd_seq = ceph_decode_32(&p);
1932 mutex_lock(&mdsc->mutex);
1933 req = __lookup_request(mdsc, tid);
1935 dout("forward %llu to mds%d - req dne\n", tid, next_mds);
1936 goto out; /* dup reply? */
1939 if (fwd_seq <= req->r_num_fwd) {
1940 dout("forward %llu to mds%d - old seq %d <= %d\n",
1941 tid, next_mds, req->r_num_fwd, fwd_seq);
1943 /* resend. forward race not possible; mds would drop */
1944 dout("forward %llu to mds%d (we resend)\n", tid, next_mds);
1945 req->r_num_fwd = fwd_seq;
1946 req->r_resend_mds = next_mds;
1947 put_request_session(req);
1948 __do_request(mdsc, req);
1950 ceph_mdsc_put_request(req);
1952 mutex_unlock(&mdsc->mutex);
1956 pr_err("mdsc_handle_forward decode error err=%d\n", err);
1960 * handle a mds session control message
1962 static void handle_session(struct ceph_mds_session *session,
1963 struct ceph_msg *msg)
1965 struct ceph_mds_client *mdsc = session->s_mdsc;
1968 int mds = session->s_mds;
1969 struct ceph_mds_session_head *h = msg->front.iov_base;
1973 if (msg->front.iov_len != sizeof(*h))
1975 op = le32_to_cpu(h->op);
1976 seq = le64_to_cpu(h->seq);
1978 mutex_lock(&mdsc->mutex);
1979 if (op == CEPH_SESSION_CLOSE)
1980 __unregister_session(mdsc, session);
1981 /* FIXME: this ttl calculation is generous */
1982 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
1983 mutex_unlock(&mdsc->mutex);
1985 mutex_lock(&session->s_mutex);
1987 dout("handle_session mds%d %s %p state %s seq %llu\n",
1988 mds, ceph_session_op_name(op), session,
1989 session_state_name(session->s_state), seq);
1991 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
1992 session->s_state = CEPH_MDS_SESSION_OPEN;
1993 pr_info("mds%d came back\n", session->s_mds);
1997 case CEPH_SESSION_OPEN:
1998 session->s_state = CEPH_MDS_SESSION_OPEN;
1999 renewed_caps(mdsc, session, 0);
2002 __close_session(mdsc, session);
2005 case CEPH_SESSION_RENEWCAPS:
2006 if (session->s_renew_seq == seq)
2007 renewed_caps(mdsc, session, 1);
2010 case CEPH_SESSION_CLOSE:
2011 remove_session_caps(session);
2012 wake = 1; /* for good measure */
2013 complete(&mdsc->session_close_waiters);
2014 kick_requests(mdsc, mds, 0); /* cur only */
2017 case CEPH_SESSION_STALE:
2018 pr_info("mds%d caps went stale, renewing\n",
2020 spin_lock(&session->s_cap_lock);
2021 session->s_cap_gen++;
2022 session->s_cap_ttl = 0;
2023 spin_unlock(&session->s_cap_lock);
2024 send_renew_caps(mdsc, session);
2027 case CEPH_SESSION_RECALL_STATE:
2028 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2032 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2036 mutex_unlock(&session->s_mutex);
2038 mutex_lock(&mdsc->mutex);
2039 __wake_requests(mdsc, &session->s_waiting);
2040 mutex_unlock(&mdsc->mutex);
2045 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2046 (int)msg->front.iov_len);
2053 * called under session->mutex.
2055 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2056 struct ceph_mds_session *session)
2058 struct ceph_mds_request *req, *nreq;
2061 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2063 mutex_lock(&mdsc->mutex);
2064 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2065 err = __prepare_send_request(mdsc, req, session->s_mds);
2067 ceph_msg_get(req->r_request);
2068 ceph_con_send(&session->s_con, req->r_request);
2071 mutex_unlock(&mdsc->mutex);
2075 * Encode information about a cap for a reconnect with the MDS.
2077 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2080 struct ceph_mds_cap_reconnect rec;
2081 struct ceph_inode_info *ci;
2082 struct ceph_pagelist *pagelist = arg;
2086 struct dentry *dentry;
2090 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2091 inode, ceph_vinop(inode), cap, cap->cap_id,
2092 ceph_cap_string(cap->issued));
2093 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2097 dentry = d_find_alias(inode);
2099 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2101 err = PTR_ERR(path);
2108 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2112 spin_lock(&inode->i_lock);
2113 cap->seq = 0; /* reset cap seq */
2114 cap->issue_seq = 0; /* and issue_seq */
2115 rec.cap_id = cpu_to_le64(cap->cap_id);
2116 rec.pathbase = cpu_to_le64(pathbase);
2117 rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2118 rec.issued = cpu_to_le32(cap->issued);
2119 rec.size = cpu_to_le64(inode->i_size);
2120 ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2121 ceph_encode_timespec(&rec.atime, &inode->i_atime);
2122 rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2123 spin_unlock(&inode->i_lock);
2125 err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2135 * If an MDS fails and recovers, clients need to reconnect in order to
2136 * reestablish shared state. This includes all caps issued through
2137 * this session _and_ the snap_realm hierarchy. Because it's not
2138 * clear which snap realms the mds cares about, we send everything we
2139 * know about.. that ensures we'll then get any new info the
2140 * recovering MDS might have.
2142 * This is a relatively heavyweight operation, but it's rare.
2144 * called with mdsc->mutex held.
2146 static void send_mds_reconnect(struct ceph_mds_client *mdsc, int mds)
2148 struct ceph_mds_session *session = NULL;
2149 struct ceph_msg *reply;
2152 struct ceph_pagelist *pagelist;
2154 pr_info("reconnect to recovering mds%d\n", mds);
2156 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2158 goto fail_nopagelist;
2159 ceph_pagelist_init(pagelist);
2161 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, 0, 0, NULL);
2162 if (IS_ERR(reply)) {
2163 err = PTR_ERR(reply);
2168 session = __ceph_lookup_mds_session(mdsc, mds);
2169 mutex_unlock(&mdsc->mutex); /* drop lock for duration */
2172 mutex_lock(&session->s_mutex);
2174 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2177 ceph_con_open(&session->s_con,
2178 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2180 /* replay unsafe requests */
2181 replay_unsafe_requests(mdsc, session);
2183 dout("no session for mds%d, will send short reconnect\n",
2187 down_read(&mdsc->snap_rwsem);
2191 dout("session %p state %s\n", session,
2192 session_state_name(session->s_state));
2194 /* traverse this session's caps */
2195 err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2198 err = iterate_session_caps(session, encode_caps_cb, pagelist);
2203 * snaprealms. we provide mds with the ino, seq (version), and
2204 * parent for all of our realms. If the mds has any newer info,
2207 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2208 struct ceph_snap_realm *realm =
2209 rb_entry(p, struct ceph_snap_realm, node);
2210 struct ceph_mds_snaprealm_reconnect sr_rec;
2212 dout(" adding snap realm %llx seq %lld parent %llx\n",
2213 realm->ino, realm->seq, realm->parent_ino);
2214 sr_rec.ino = cpu_to_le64(realm->ino);
2215 sr_rec.seq = cpu_to_le64(realm->seq);
2216 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2217 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2223 reply->pagelist = pagelist;
2224 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2225 reply->nr_pages = calc_pages_for(0, pagelist->length);
2226 ceph_con_send(&session->s_con, reply);
2228 session->s_state = CEPH_MDS_SESSION_OPEN;
2229 mutex_unlock(&session->s_mutex);
2231 mutex_lock(&mdsc->mutex);
2232 __wake_requests(mdsc, &session->s_waiting);
2233 mutex_unlock(&mdsc->mutex);
2235 ceph_put_mds_session(session);
2237 up_read(&mdsc->snap_rwsem);
2238 mutex_lock(&mdsc->mutex);
2242 ceph_msg_put(reply);
2243 up_read(&mdsc->snap_rwsem);
2244 mutex_unlock(&session->s_mutex);
2245 ceph_put_mds_session(session);
2247 ceph_pagelist_release(pagelist);
2250 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2251 mutex_lock(&mdsc->mutex);
2257 * compare old and new mdsmaps, kicking requests
2258 * and closing out old connections as necessary
2260 * called under mdsc->mutex.
2262 static void check_new_map(struct ceph_mds_client *mdsc,
2263 struct ceph_mdsmap *newmap,
2264 struct ceph_mdsmap *oldmap)
2267 int oldstate, newstate;
2268 struct ceph_mds_session *s;
2270 dout("check_new_map new %u old %u\n",
2271 newmap->m_epoch, oldmap->m_epoch);
2273 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2274 if (mdsc->sessions[i] == NULL)
2276 s = mdsc->sessions[i];
2277 oldstate = ceph_mdsmap_get_state(oldmap, i);
2278 newstate = ceph_mdsmap_get_state(newmap, i);
2280 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2281 i, ceph_mds_state_name(oldstate),
2282 ceph_mds_state_name(newstate),
2283 session_state_name(s->s_state));
2285 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2286 ceph_mdsmap_get_addr(newmap, i),
2287 sizeof(struct ceph_entity_addr))) {
2288 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2289 /* the session never opened, just close it
2291 __wake_requests(mdsc, &s->s_waiting);
2292 __unregister_session(mdsc, s);
2295 mutex_unlock(&mdsc->mutex);
2296 mutex_lock(&s->s_mutex);
2297 mutex_lock(&mdsc->mutex);
2298 ceph_con_close(&s->s_con);
2299 mutex_unlock(&s->s_mutex);
2300 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2303 /* kick any requests waiting on the recovering mds */
2304 kick_requests(mdsc, i, 1);
2305 } else if (oldstate == newstate) {
2306 continue; /* nothing new with this mds */
2312 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2313 newstate >= CEPH_MDS_STATE_RECONNECT)
2314 send_mds_reconnect(mdsc, i);
2317 * kick requests on any mds that has gone active.
2319 * kick requests on cur or forwarder: we may have sent
2320 * the request to mds1, mds1 told us it forwarded it
2321 * to mds2, but then we learn mds1 failed and can't be
2322 * sure it successfully forwarded our request before
2325 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2326 newstate >= CEPH_MDS_STATE_ACTIVE) {
2327 pr_info("mds%d reconnect completed\n", s->s_mds);
2328 kick_requests(mdsc, i, 1);
2329 ceph_kick_flushing_caps(mdsc, s);
2330 wake_up_session_caps(s, 1);
2342 * caller must hold session s_mutex, dentry->d_lock
2344 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2346 struct ceph_dentry_info *di = ceph_dentry(dentry);
2348 ceph_put_mds_session(di->lease_session);
2349 di->lease_session = NULL;
2352 static void handle_lease(struct ceph_mds_client *mdsc,
2353 struct ceph_mds_session *session,
2354 struct ceph_msg *msg)
2356 struct super_block *sb = mdsc->client->sb;
2357 struct inode *inode;
2358 struct ceph_inode_info *ci;
2359 struct dentry *parent, *dentry;
2360 struct ceph_dentry_info *di;
2361 int mds = session->s_mds;
2362 struct ceph_mds_lease *h = msg->front.iov_base;
2363 struct ceph_vino vino;
2368 dout("handle_lease from mds%d\n", mds);
2371 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2373 vino.ino = le64_to_cpu(h->ino);
2374 vino.snap = CEPH_NOSNAP;
2375 mask = le16_to_cpu(h->mask);
2376 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2377 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2378 if (dname.len != get_unaligned_le32(h+1))
2381 mutex_lock(&session->s_mutex);
2385 inode = ceph_find_inode(sb, vino);
2386 dout("handle_lease '%s', mask %d, ino %llx %p\n",
2387 ceph_lease_op_name(h->action), mask, vino.ino, inode);
2388 if (inode == NULL) {
2389 dout("handle_lease no inode %llx\n", vino.ino);
2392 ci = ceph_inode(inode);
2395 parent = d_find_alias(inode);
2397 dout("no parent dentry on inode %p\n", inode);
2399 goto release; /* hrm... */
2401 dname.hash = full_name_hash(dname.name, dname.len);
2402 dentry = d_lookup(parent, &dname);
2407 spin_lock(&dentry->d_lock);
2408 di = ceph_dentry(dentry);
2409 switch (h->action) {
2410 case CEPH_MDS_LEASE_REVOKE:
2411 if (di && di->lease_session == session) {
2412 h->seq = cpu_to_le32(di->lease_seq);
2413 __ceph_mdsc_drop_dentry_lease(dentry);
2418 case CEPH_MDS_LEASE_RENEW:
2419 if (di && di->lease_session == session &&
2420 di->lease_gen == session->s_cap_gen &&
2421 di->lease_renew_from &&
2422 di->lease_renew_after == 0) {
2423 unsigned long duration =
2424 le32_to_cpu(h->duration_ms) * HZ / 1000;
2426 di->lease_seq = le32_to_cpu(h->seq);
2427 dentry->d_time = di->lease_renew_from + duration;
2428 di->lease_renew_after = di->lease_renew_from +
2430 di->lease_renew_from = 0;
2434 spin_unlock(&dentry->d_lock);
2441 /* let's just reuse the same message */
2442 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2444 ceph_con_send(&session->s_con, msg);
2448 mutex_unlock(&session->s_mutex);
2452 pr_err("corrupt lease message\n");
2456 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2457 struct inode *inode,
2458 struct dentry *dentry, char action,
2461 struct ceph_msg *msg;
2462 struct ceph_mds_lease *lease;
2463 int len = sizeof(*lease) + sizeof(u32);
2466 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2467 inode, dentry, ceph_lease_op_name(action), session->s_mds);
2468 dnamelen = dentry->d_name.len;
2471 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, 0, 0, NULL);
2474 lease = msg->front.iov_base;
2475 lease->action = action;
2476 lease->mask = cpu_to_le16(CEPH_LOCK_DN);
2477 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2478 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2479 lease->seq = cpu_to_le32(seq);
2480 put_unaligned_le32(dnamelen, lease + 1);
2481 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2484 * if this is a preemptive lease RELEASE, no need to
2485 * flush request stream, since the actual request will
2488 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2490 ceph_con_send(&session->s_con, msg);
2494 * Preemptively release a lease we expect to invalidate anyway.
2495 * Pass @inode always, @dentry is optional.
2497 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2498 struct dentry *dentry, int mask)
2500 struct ceph_dentry_info *di;
2501 struct ceph_mds_session *session;
2504 BUG_ON(inode == NULL);
2505 BUG_ON(dentry == NULL);
2506 BUG_ON(mask != CEPH_LOCK_DN);
2508 /* is dentry lease valid? */
2509 spin_lock(&dentry->d_lock);
2510 di = ceph_dentry(dentry);
2511 if (!di || !di->lease_session ||
2512 di->lease_session->s_mds < 0 ||
2513 di->lease_gen != di->lease_session->s_cap_gen ||
2514 !time_before(jiffies, dentry->d_time)) {
2515 dout("lease_release inode %p dentry %p -- "
2517 inode, dentry, mask);
2518 spin_unlock(&dentry->d_lock);
2522 /* we do have a lease on this dentry; note mds and seq */
2523 session = ceph_get_mds_session(di->lease_session);
2524 seq = di->lease_seq;
2525 __ceph_mdsc_drop_dentry_lease(dentry);
2526 spin_unlock(&dentry->d_lock);
2528 dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2529 inode, dentry, mask, session->s_mds);
2530 ceph_mdsc_lease_send_msg(session, inode, dentry,
2531 CEPH_MDS_LEASE_RELEASE, seq);
2532 ceph_put_mds_session(session);
2536 * drop all leases (and dentry refs) in preparation for umount
2538 static void drop_leases(struct ceph_mds_client *mdsc)
2542 dout("drop_leases\n");
2543 mutex_lock(&mdsc->mutex);
2544 for (i = 0; i < mdsc->max_sessions; i++) {
2545 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2548 mutex_unlock(&mdsc->mutex);
2549 mutex_lock(&s->s_mutex);
2550 mutex_unlock(&s->s_mutex);
2551 ceph_put_mds_session(s);
2552 mutex_lock(&mdsc->mutex);
2554 mutex_unlock(&mdsc->mutex);
2560 * delayed work -- periodically trim expired leases, renew caps with mds
2562 static void schedule_delayed(struct ceph_mds_client *mdsc)
2565 unsigned hz = round_jiffies_relative(HZ * delay);
2566 schedule_delayed_work(&mdsc->delayed_work, hz);
2569 static void delayed_work(struct work_struct *work)
2572 struct ceph_mds_client *mdsc =
2573 container_of(work, struct ceph_mds_client, delayed_work.work);
2577 dout("mdsc delayed_work\n");
2578 ceph_check_delayed_caps(mdsc);
2580 mutex_lock(&mdsc->mutex);
2581 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2582 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2583 mdsc->last_renew_caps);
2585 mdsc->last_renew_caps = jiffies;
2587 for (i = 0; i < mdsc->max_sessions; i++) {
2588 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2591 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2592 dout("resending session close request for mds%d\n",
2594 request_close_session(mdsc, s);
2595 ceph_put_mds_session(s);
2598 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2599 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2600 s->s_state = CEPH_MDS_SESSION_HUNG;
2601 pr_info("mds%d hung\n", s->s_mds);
2604 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2605 /* this mds is failed or recovering, just wait */
2606 ceph_put_mds_session(s);
2609 mutex_unlock(&mdsc->mutex);
2611 mutex_lock(&s->s_mutex);
2613 send_renew_caps(mdsc, s);
2615 ceph_con_keepalive(&s->s_con);
2616 add_cap_releases(mdsc, s, -1);
2617 send_cap_releases(mdsc, s);
2618 mutex_unlock(&s->s_mutex);
2619 ceph_put_mds_session(s);
2621 mutex_lock(&mdsc->mutex);
2623 mutex_unlock(&mdsc->mutex);
2625 schedule_delayed(mdsc);
2629 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2631 mdsc->client = client;
2632 mutex_init(&mdsc->mutex);
2633 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2634 if (mdsc->mdsmap == NULL)
2637 init_completion(&mdsc->safe_umount_waiters);
2638 init_completion(&mdsc->session_close_waiters);
2639 INIT_LIST_HEAD(&mdsc->waiting_for_map);
2640 mdsc->sessions = NULL;
2641 mdsc->max_sessions = 0;
2643 init_rwsem(&mdsc->snap_rwsem);
2644 mdsc->snap_realms = RB_ROOT;
2645 INIT_LIST_HEAD(&mdsc->snap_empty);
2646 spin_lock_init(&mdsc->snap_empty_lock);
2648 mdsc->request_tree = RB_ROOT;
2649 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2650 mdsc->last_renew_caps = jiffies;
2651 INIT_LIST_HEAD(&mdsc->cap_delay_list);
2652 spin_lock_init(&mdsc->cap_delay_lock);
2653 INIT_LIST_HEAD(&mdsc->snap_flush_list);
2654 spin_lock_init(&mdsc->snap_flush_lock);
2655 mdsc->cap_flush_seq = 0;
2656 INIT_LIST_HEAD(&mdsc->cap_dirty);
2657 mdsc->num_cap_flushing = 0;
2658 spin_lock_init(&mdsc->cap_dirty_lock);
2659 init_waitqueue_head(&mdsc->cap_flushing_wq);
2660 spin_lock_init(&mdsc->dentry_lru_lock);
2661 INIT_LIST_HEAD(&mdsc->dentry_lru);
2667 * Wait for safe replies on open mds requests. If we time out, drop
2668 * all requests from the tree to avoid dangling dentry refs.
2670 static void wait_requests(struct ceph_mds_client *mdsc)
2672 struct ceph_mds_request *req;
2673 struct ceph_client *client = mdsc->client;
2675 mutex_lock(&mdsc->mutex);
2676 if (__get_oldest_req(mdsc)) {
2677 mutex_unlock(&mdsc->mutex);
2679 dout("wait_requests waiting for requests\n");
2680 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2681 client->mount_args->mount_timeout * HZ);
2683 /* tear down remaining requests */
2684 mutex_lock(&mdsc->mutex);
2685 while ((req = __get_oldest_req(mdsc))) {
2686 dout("wait_requests timed out on tid %llu\n",
2688 __unregister_request(mdsc, req);
2691 mutex_unlock(&mdsc->mutex);
2692 dout("wait_requests done\n");
2696 * called before mount is ro, and before dentries are torn down.
2697 * (hmm, does this still race with new lookups?)
2699 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2701 dout("pre_umount\n");
2705 ceph_flush_dirty_caps(mdsc);
2706 wait_requests(mdsc);
2710 * wait for all write mds requests to flush.
2712 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2714 struct ceph_mds_request *req = NULL, *nextreq;
2717 mutex_lock(&mdsc->mutex);
2718 dout("wait_unsafe_requests want %lld\n", want_tid);
2720 req = __get_oldest_req(mdsc);
2721 while (req && req->r_tid <= want_tid) {
2722 /* find next request */
2723 n = rb_next(&req->r_node);
2725 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2728 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2730 ceph_mdsc_get_request(req);
2732 ceph_mdsc_get_request(nextreq);
2733 mutex_unlock(&mdsc->mutex);
2734 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
2735 req->r_tid, want_tid);
2736 wait_for_completion(&req->r_safe_completion);
2737 mutex_lock(&mdsc->mutex);
2738 ceph_mdsc_put_request(req);
2740 break; /* next dne before, so we're done! */
2741 if (RB_EMPTY_NODE(&nextreq->r_node)) {
2742 /* next request was removed from tree */
2743 ceph_mdsc_put_request(nextreq);
2746 ceph_mdsc_put_request(nextreq); /* won't go away */
2750 mutex_unlock(&mdsc->mutex);
2751 dout("wait_unsafe_requests done\n");
2754 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2756 u64 want_tid, want_flush;
2759 mutex_lock(&mdsc->mutex);
2760 want_tid = mdsc->last_tid;
2761 want_flush = mdsc->cap_flush_seq;
2762 mutex_unlock(&mdsc->mutex);
2763 dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2765 ceph_flush_dirty_caps(mdsc);
2767 wait_unsafe_requests(mdsc, want_tid);
2768 wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2773 * called after sb is ro.
2775 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2777 struct ceph_mds_session *session;
2780 struct ceph_client *client = mdsc->client;
2781 unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2783 dout("close_sessions\n");
2785 mutex_lock(&mdsc->mutex);
2787 /* close sessions */
2789 while (time_before(jiffies, started + timeout)) {
2790 dout("closing sessions\n");
2792 for (i = 0; i < mdsc->max_sessions; i++) {
2793 session = __ceph_lookup_mds_session(mdsc, i);
2796 mutex_unlock(&mdsc->mutex);
2797 mutex_lock(&session->s_mutex);
2798 __close_session(mdsc, session);
2799 mutex_unlock(&session->s_mutex);
2800 ceph_put_mds_session(session);
2801 mutex_lock(&mdsc->mutex);
2807 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2810 dout("waiting for sessions to close\n");
2811 mutex_unlock(&mdsc->mutex);
2812 wait_for_completion_timeout(&mdsc->session_close_waiters,
2814 mutex_lock(&mdsc->mutex);
2817 /* tear down remaining sessions */
2818 for (i = 0; i < mdsc->max_sessions; i++) {
2819 if (mdsc->sessions[i]) {
2820 session = get_session(mdsc->sessions[i]);
2821 __unregister_session(mdsc, session);
2822 mutex_unlock(&mdsc->mutex);
2823 mutex_lock(&session->s_mutex);
2824 remove_session_caps(session);
2825 mutex_unlock(&session->s_mutex);
2826 ceph_put_mds_session(session);
2827 mutex_lock(&mdsc->mutex);
2831 WARN_ON(!list_empty(&mdsc->cap_delay_list));
2833 mutex_unlock(&mdsc->mutex);
2835 ceph_cleanup_empty_realms(mdsc);
2837 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2842 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2845 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2847 ceph_mdsmap_destroy(mdsc->mdsmap);
2848 kfree(mdsc->sessions);
2853 * handle mds map update.
2855 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2859 void *p = msg->front.iov_base;
2860 void *end = p + msg->front.iov_len;
2861 struct ceph_mdsmap *newmap, *oldmap;
2862 struct ceph_fsid fsid;
2865 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2866 ceph_decode_copy(&p, &fsid, sizeof(fsid));
2867 if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2869 epoch = ceph_decode_32(&p);
2870 maplen = ceph_decode_32(&p);
2871 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2873 /* do we need it? */
2874 ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2875 mutex_lock(&mdsc->mutex);
2876 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2877 dout("handle_map epoch %u <= our %u\n",
2878 epoch, mdsc->mdsmap->m_epoch);
2879 mutex_unlock(&mdsc->mutex);
2883 newmap = ceph_mdsmap_decode(&p, end);
2884 if (IS_ERR(newmap)) {
2885 err = PTR_ERR(newmap);
2889 /* swap into place */
2891 oldmap = mdsc->mdsmap;
2892 mdsc->mdsmap = newmap;
2893 check_new_map(mdsc, newmap, oldmap);
2894 ceph_mdsmap_destroy(oldmap);
2896 mdsc->mdsmap = newmap; /* first mds map */
2898 mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2900 __wake_requests(mdsc, &mdsc->waiting_for_map);
2902 mutex_unlock(&mdsc->mutex);
2903 schedule_delayed(mdsc);
2907 mutex_unlock(&mdsc->mutex);
2909 pr_err("error decoding mdsmap %d\n", err);
2913 static struct ceph_connection *con_get(struct ceph_connection *con)
2915 struct ceph_mds_session *s = con->private;
2917 if (get_session(s)) {
2918 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
2921 dout("mdsc con_get %p FAIL\n", s);
2925 static void con_put(struct ceph_connection *con)
2927 struct ceph_mds_session *s = con->private;
2929 ceph_put_mds_session(s);
2930 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
2934 * if the client is unresponsive for long enough, the mds will kill
2935 * the session entirely.
2937 static void peer_reset(struct ceph_connection *con)
2939 struct ceph_mds_session *s = con->private;
2941 pr_err("mds%d gave us the boot. IMPLEMENT RECONNECT.\n",
2945 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2947 struct ceph_mds_session *s = con->private;
2948 struct ceph_mds_client *mdsc = s->s_mdsc;
2949 int type = le16_to_cpu(msg->hdr.type);
2951 mutex_lock(&mdsc->mutex);
2952 if (__verify_registered_session(mdsc, s) < 0) {
2953 mutex_unlock(&mdsc->mutex);
2956 mutex_unlock(&mdsc->mutex);
2959 case CEPH_MSG_MDS_MAP:
2960 ceph_mdsc_handle_map(mdsc, msg);
2962 case CEPH_MSG_CLIENT_SESSION:
2963 handle_session(s, msg);
2965 case CEPH_MSG_CLIENT_REPLY:
2966 handle_reply(s, msg);
2968 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
2969 handle_forward(mdsc, s, msg);
2971 case CEPH_MSG_CLIENT_CAPS:
2972 ceph_handle_caps(s, msg);
2974 case CEPH_MSG_CLIENT_SNAP:
2975 ceph_handle_snap(mdsc, s, msg);
2977 case CEPH_MSG_CLIENT_LEASE:
2978 handle_lease(mdsc, s, msg);
2982 pr_err("received unknown message type %d %s\n", type,
2983 ceph_msg_type_name(type));
2992 static int get_authorizer(struct ceph_connection *con,
2993 void **buf, int *len, int *proto,
2994 void **reply_buf, int *reply_len, int force_new)
2996 struct ceph_mds_session *s = con->private;
2997 struct ceph_mds_client *mdsc = s->s_mdsc;
2998 struct ceph_auth_client *ac = mdsc->client->monc.auth;
3001 if (force_new && s->s_authorizer) {
3002 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3003 s->s_authorizer = NULL;
3005 if (s->s_authorizer == NULL) {
3006 if (ac->ops->create_authorizer) {
3007 ret = ac->ops->create_authorizer(
3008 ac, CEPH_ENTITY_TYPE_MDS,
3010 &s->s_authorizer_buf,
3011 &s->s_authorizer_buf_len,
3012 &s->s_authorizer_reply_buf,
3013 &s->s_authorizer_reply_buf_len);
3019 *proto = ac->protocol;
3020 *buf = s->s_authorizer_buf;
3021 *len = s->s_authorizer_buf_len;
3022 *reply_buf = s->s_authorizer_reply_buf;
3023 *reply_len = s->s_authorizer_reply_buf_len;
3028 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3030 struct ceph_mds_session *s = con->private;
3031 struct ceph_mds_client *mdsc = s->s_mdsc;
3032 struct ceph_auth_client *ac = mdsc->client->monc.auth;
3034 return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3037 static int invalidate_authorizer(struct ceph_connection *con)
3039 struct ceph_mds_session *s = con->private;
3040 struct ceph_mds_client *mdsc = s->s_mdsc;
3041 struct ceph_auth_client *ac = mdsc->client->monc.auth;
3043 if (ac->ops->invalidate_authorizer)
3044 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3046 return ceph_monc_validate_auth(&mdsc->client->monc);
3049 const static struct ceph_connection_operations mds_con_ops = {
3052 .dispatch = dispatch,
3053 .get_authorizer = get_authorizer,
3054 .verify_authorizer_reply = verify_authorizer_reply,
3055 .invalidate_authorizer = invalidate_authorizer,
3056 .peer_reset = peer_reset,