1 #include "ceph_debug.h"
3 #include <linux/wait.h>
4 #include <linux/sched.h>
6 #include "mds_client.h"
7 #include "mon_client.h"
15 * A cluster of MDS (metadata server) daemons is responsible for
16 * managing the file system namespace (the directory hierarchy and
17 * inodes) and for coordinating shared access to storage. Metadata is
18 * partitioning hierarchically across a number of servers, and that
19 * partition varies over time as the cluster adjusts the distribution
20 * in order to balance load.
22 * The MDS client is primarily responsible to managing synchronous
23 * metadata requests for operations like open, unlink, and so forth.
24 * If there is a MDS failure, we find out about it when we (possibly
25 * request and) receive a new MDS map, and can resubmit affected
28 * For the most part, though, we take advantage of a lossless
29 * communications channel to the MDS, and do not need to worry about
30 * timing out or resubmitting requests.
32 * We maintain a stateful "session" with each MDS we interact with.
33 * Within each session, we sent periodic heartbeat messages to ensure
34 * any capabilities or leases we have been issues remain valid. If
35 * the session times out and goes stale, our leases and capabilities
36 * are no longer valid.
39 static void __wake_requests(struct ceph_mds_client *mdsc,
40 struct list_head *head);
42 const static struct ceph_connection_operations mds_con_ops;
50 * parse individual inode info
52 static int parse_reply_info_in(void **p, void *end,
53 struct ceph_mds_reply_info_in *info)
58 *p += sizeof(struct ceph_mds_reply_inode) +
59 sizeof(*info->in->fragtree.splits) *
60 le32_to_cpu(info->in->fragtree.nsplits);
62 ceph_decode_32_safe(p, end, info->symlink_len, bad);
63 ceph_decode_need(p, end, info->symlink_len, bad);
65 *p += info->symlink_len;
67 ceph_decode_32_safe(p, end, info->xattr_len, bad);
68 ceph_decode_need(p, end, info->xattr_len, bad);
69 info->xattr_data = *p;
70 *p += info->xattr_len;
77 * parse a normal reply, which may contain a (dir+)dentry and/or a
80 static int parse_reply_info_trace(void **p, void *end,
81 struct ceph_mds_reply_info_parsed *info)
85 if (info->head->is_dentry) {
86 err = parse_reply_info_in(p, end, &info->diri);
90 if (unlikely(*p + sizeof(*info->dirfrag) > end))
93 *p += sizeof(*info->dirfrag) +
94 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
95 if (unlikely(*p > end))
98 ceph_decode_32_safe(p, end, info->dname_len, bad);
99 ceph_decode_need(p, end, info->dname_len, bad);
101 *p += info->dname_len;
103 *p += sizeof(*info->dlease);
106 if (info->head->is_target) {
107 err = parse_reply_info_in(p, end, &info->targeti);
112 if (unlikely(*p != end))
119 pr_err("problem parsing mds trace %d\n", err);
124 * parse readdir results
126 static int parse_reply_info_dir(void **p, void *end,
127 struct ceph_mds_reply_info_parsed *info)
133 if (*p + sizeof(*info->dir_dir) > end)
135 *p += sizeof(*info->dir_dir) +
136 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
140 ceph_decode_need(p, end, sizeof(num) + 2, bad);
141 num = ceph_decode_32(p);
142 info->dir_end = ceph_decode_8(p);
143 info->dir_complete = ceph_decode_8(p);
147 /* alloc large array */
149 info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
150 sizeof(*info->dir_dname) +
151 sizeof(*info->dir_dname_len) +
152 sizeof(*info->dir_dlease),
154 if (info->dir_in == NULL) {
158 info->dir_dname = (void *)(info->dir_in + num);
159 info->dir_dname_len = (void *)(info->dir_dname + num);
160 info->dir_dlease = (void *)(info->dir_dname_len + num);
164 ceph_decode_need(p, end, sizeof(u32)*2, bad);
165 info->dir_dname_len[i] = ceph_decode_32(p);
166 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
167 info->dir_dname[i] = *p;
168 *p += info->dir_dname_len[i];
169 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
171 info->dir_dlease[i] = *p;
172 *p += sizeof(struct ceph_mds_reply_lease);
175 err = parse_reply_info_in(p, end, &info->dir_in[i]);
190 pr_err("problem parsing dir contents %d\n", err);
195 * parse entire mds reply
197 static int parse_reply_info(struct ceph_msg *msg,
198 struct ceph_mds_reply_info_parsed *info)
204 info->head = msg->front.iov_base;
205 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
206 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
209 ceph_decode_32_safe(&p, end, len, bad);
211 err = parse_reply_info_trace(&p, p+len, info);
217 ceph_decode_32_safe(&p, end, len, bad);
219 err = parse_reply_info_dir(&p, p+len, info);
225 ceph_decode_32_safe(&p, end, len, bad);
226 info->snapblob_len = len;
237 pr_err("mds parse_reply err %d\n", err);
241 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
250 static const char *session_state_name(int s)
253 case CEPH_MDS_SESSION_NEW: return "new";
254 case CEPH_MDS_SESSION_OPENING: return "opening";
255 case CEPH_MDS_SESSION_OPEN: return "open";
256 case CEPH_MDS_SESSION_HUNG: return "hung";
257 case CEPH_MDS_SESSION_CLOSING: return "closing";
258 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
259 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
260 default: return "???";
264 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
266 if (atomic_inc_not_zero(&s->s_ref)) {
267 dout("mdsc get_session %p %d -> %d\n", s,
268 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
271 dout("mdsc get_session %p 0 -- FAIL", s);
276 void ceph_put_mds_session(struct ceph_mds_session *s)
278 dout("mdsc put_session %p %d -> %d\n", s,
279 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
280 if (atomic_dec_and_test(&s->s_ref)) {
282 s->s_mdsc->client->monc.auth->ops->destroy_authorizer(
283 s->s_mdsc->client->monc.auth, s->s_authorizer);
289 * called under mdsc->mutex
291 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
294 struct ceph_mds_session *session;
296 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
298 session = mdsc->sessions[mds];
299 dout("lookup_mds_session %p %d\n", session,
300 atomic_read(&session->s_ref));
301 get_session(session);
305 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
307 if (mds >= mdsc->max_sessions)
309 return mdsc->sessions[mds];
312 static int __verify_registered_session(struct ceph_mds_client *mdsc,
313 struct ceph_mds_session *s)
315 if (s->s_mds >= mdsc->max_sessions ||
316 mdsc->sessions[s->s_mds] != s)
322 * create+register a new session for given mds.
323 * called under mdsc->mutex.
325 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
328 struct ceph_mds_session *s;
330 s = kzalloc(sizeof(*s), GFP_NOFS);
332 return ERR_PTR(-ENOMEM);
335 s->s_state = CEPH_MDS_SESSION_NEW;
338 mutex_init(&s->s_mutex);
340 ceph_con_init(mdsc->client->msgr, &s->s_con);
341 s->s_con.private = s;
342 s->s_con.ops = &mds_con_ops;
343 s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
344 s->s_con.peer_name.num = cpu_to_le64(mds);
346 spin_lock_init(&s->s_cap_lock);
349 s->s_renew_requested = 0;
351 INIT_LIST_HEAD(&s->s_caps);
354 atomic_set(&s->s_ref, 1);
355 INIT_LIST_HEAD(&s->s_waiting);
356 INIT_LIST_HEAD(&s->s_unsafe);
357 s->s_num_cap_releases = 0;
358 s->s_cap_iterator = NULL;
359 INIT_LIST_HEAD(&s->s_cap_releases);
360 INIT_LIST_HEAD(&s->s_cap_releases_done);
361 INIT_LIST_HEAD(&s->s_cap_flushing);
362 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
364 dout("register_session mds%d\n", mds);
365 if (mds >= mdsc->max_sessions) {
366 int newmax = 1 << get_count_order(mds+1);
367 struct ceph_mds_session **sa;
369 dout("register_session realloc to %d\n", newmax);
370 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
373 if (mdsc->sessions) {
374 memcpy(sa, mdsc->sessions,
375 mdsc->max_sessions * sizeof(void *));
376 kfree(mdsc->sessions);
379 mdsc->max_sessions = newmax;
381 mdsc->sessions[mds] = s;
382 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
384 ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
390 return ERR_PTR(-ENOMEM);
394 * called under mdsc->mutex
396 static void __unregister_session(struct ceph_mds_client *mdsc,
397 struct ceph_mds_session *s)
399 dout("__unregister_session mds%d %p\n", s->s_mds, s);
400 BUG_ON(mdsc->sessions[s->s_mds] != s);
401 mdsc->sessions[s->s_mds] = NULL;
402 ceph_con_close(&s->s_con);
403 ceph_put_mds_session(s);
407 * drop session refs in request.
409 * should be last request ref, or hold mdsc->mutex
411 static void put_request_session(struct ceph_mds_request *req)
413 if (req->r_session) {
414 ceph_put_mds_session(req->r_session);
415 req->r_session = NULL;
419 void ceph_mdsc_release_request(struct kref *kref)
421 struct ceph_mds_request *req = container_of(kref,
422 struct ceph_mds_request,
425 ceph_msg_put(req->r_request);
427 ceph_msg_put(req->r_reply);
428 destroy_reply_info(&req->r_reply_info);
431 ceph_put_cap_refs(ceph_inode(req->r_inode),
435 if (req->r_locked_dir)
436 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
438 if (req->r_target_inode)
439 iput(req->r_target_inode);
442 if (req->r_old_dentry) {
444 ceph_inode(req->r_old_dentry->d_parent->d_inode),
446 dput(req->r_old_dentry);
450 put_request_session(req);
451 ceph_unreserve_caps(&req->r_caps_reservation);
456 * lookup session, bump ref if found.
458 * called under mdsc->mutex.
460 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
463 struct ceph_mds_request *req;
464 struct rb_node *n = mdsc->request_tree.rb_node;
467 req = rb_entry(n, struct ceph_mds_request, r_node);
468 if (tid < req->r_tid)
470 else if (tid > req->r_tid)
473 ceph_mdsc_get_request(req);
480 static void __insert_request(struct ceph_mds_client *mdsc,
481 struct ceph_mds_request *new)
483 struct rb_node **p = &mdsc->request_tree.rb_node;
484 struct rb_node *parent = NULL;
485 struct ceph_mds_request *req = NULL;
489 req = rb_entry(parent, struct ceph_mds_request, r_node);
490 if (new->r_tid < req->r_tid)
492 else if (new->r_tid > req->r_tid)
498 rb_link_node(&new->r_node, parent, p);
499 rb_insert_color(&new->r_node, &mdsc->request_tree);
503 * Register an in-flight request, and assign a tid. Link to directory
504 * are modifying (if any).
506 * Called under mdsc->mutex.
508 static void __register_request(struct ceph_mds_client *mdsc,
509 struct ceph_mds_request *req,
512 req->r_tid = ++mdsc->last_tid;
514 ceph_reserve_caps(&req->r_caps_reservation, req->r_num_caps);
515 dout("__register_request %p tid %lld\n", req, req->r_tid);
516 ceph_mdsc_get_request(req);
517 __insert_request(mdsc, req);
520 struct ceph_inode_info *ci = ceph_inode(dir);
522 spin_lock(&ci->i_unsafe_lock);
523 req->r_unsafe_dir = dir;
524 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
525 spin_unlock(&ci->i_unsafe_lock);
529 static void __unregister_request(struct ceph_mds_client *mdsc,
530 struct ceph_mds_request *req)
532 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
533 rb_erase(&req->r_node, &mdsc->request_tree);
534 RB_CLEAR_NODE(&req->r_node);
536 if (req->r_unsafe_dir) {
537 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
539 spin_lock(&ci->i_unsafe_lock);
540 list_del_init(&req->r_unsafe_dir_item);
541 spin_unlock(&ci->i_unsafe_lock);
544 ceph_mdsc_put_request(req);
548 * Choose mds to send request to next. If there is a hint set in the
549 * request (e.g., due to a prior forward hint from the mds), use that.
550 * Otherwise, consult frag tree and/or caps to identify the
551 * appropriate mds. If all else fails, choose randomly.
553 * Called under mdsc->mutex.
555 static int __choose_mds(struct ceph_mds_client *mdsc,
556 struct ceph_mds_request *req)
559 struct ceph_inode_info *ci;
560 struct ceph_cap *cap;
561 int mode = req->r_direct_mode;
563 u32 hash = req->r_direct_hash;
564 bool is_hash = req->r_direct_is_hash;
567 * is there a specific mds we should try? ignore hint if we have
568 * no session and the mds is not up (active or recovering).
570 if (req->r_resend_mds >= 0 &&
571 (__have_session(mdsc, req->r_resend_mds) ||
572 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
573 dout("choose_mds using resend_mds mds%d\n",
575 return req->r_resend_mds;
578 if (mode == USE_RANDOM_MDS)
583 inode = req->r_inode;
584 } else if (req->r_dentry) {
585 if (req->r_dentry->d_inode) {
586 inode = req->r_dentry->d_inode;
588 inode = req->r_dentry->d_parent->d_inode;
589 hash = req->r_dentry->d_name.hash;
593 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
597 ci = ceph_inode(inode);
599 if (is_hash && S_ISDIR(inode->i_mode)) {
600 struct ceph_inode_frag frag;
603 ceph_choose_frag(ci, hash, &frag, &found);
605 if (mode == USE_ANY_MDS && frag.ndist > 0) {
608 /* choose a random replica */
609 get_random_bytes(&r, 1);
612 dout("choose_mds %p %llx.%llx "
613 "frag %u mds%d (%d/%d)\n",
614 inode, ceph_vinop(inode),
620 /* since this file/dir wasn't known to be
621 * replicated, then we want to look for the
622 * authoritative mds. */
625 /* choose auth mds */
627 dout("choose_mds %p %llx.%llx "
628 "frag %u mds%d (auth)\n",
629 inode, ceph_vinop(inode), frag.frag, mds);
635 spin_lock(&inode->i_lock);
637 if (mode == USE_AUTH_MDS)
638 cap = ci->i_auth_cap;
639 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
640 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
642 spin_unlock(&inode->i_lock);
645 mds = cap->session->s_mds;
646 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
647 inode, ceph_vinop(inode), mds,
648 cap == ci->i_auth_cap ? "auth " : "", cap);
649 spin_unlock(&inode->i_lock);
653 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
654 dout("choose_mds chose random mds%d\n", mds);
662 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
664 struct ceph_msg *msg;
665 struct ceph_mds_session_head *h;
667 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), 0, 0, NULL);
669 pr_err("create_session_msg ENOMEM creating msg\n");
670 return ERR_PTR(PTR_ERR(msg));
672 h = msg->front.iov_base;
673 h->op = cpu_to_le32(op);
674 h->seq = cpu_to_le64(seq);
679 * send session open request.
681 * called under mdsc->mutex
683 static int __open_session(struct ceph_mds_client *mdsc,
684 struct ceph_mds_session *session)
686 struct ceph_msg *msg;
688 int mds = session->s_mds;
691 /* wait for mds to go active? */
692 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
693 dout("open_session to mds%d (%s)\n", mds,
694 ceph_mds_state_name(mstate));
695 session->s_state = CEPH_MDS_SESSION_OPENING;
696 session->s_renew_requested = jiffies;
698 /* send connect message */
699 msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
704 ceph_con_send(&session->s_con, msg);
715 * Free preallocated cap messages assigned to this session
717 static void cleanup_cap_releases(struct ceph_mds_session *session)
719 struct ceph_msg *msg;
721 spin_lock(&session->s_cap_lock);
722 while (!list_empty(&session->s_cap_releases)) {
723 msg = list_first_entry(&session->s_cap_releases,
724 struct ceph_msg, list_head);
725 list_del_init(&msg->list_head);
728 while (!list_empty(&session->s_cap_releases_done)) {
729 msg = list_first_entry(&session->s_cap_releases_done,
730 struct ceph_msg, list_head);
731 list_del_init(&msg->list_head);
734 spin_unlock(&session->s_cap_lock);
738 * Helper to safely iterate over all caps associated with a session.
740 * caller must hold session s_mutex
742 static int iterate_session_caps(struct ceph_mds_session *session,
743 int (*cb)(struct inode *, struct ceph_cap *,
747 struct ceph_cap *cap;
748 struct inode *inode, *last_inode = NULL;
749 struct ceph_cap *old_cap = NULL;
752 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
753 spin_lock(&session->s_cap_lock);
754 p = session->s_caps.next;
755 while (p != &session->s_caps) {
756 cap = list_entry(p, struct ceph_cap, session_caps);
757 inode = igrab(&cap->ci->vfs_inode);
762 session->s_cap_iterator = cap;
763 spin_unlock(&session->s_cap_lock);
770 ceph_put_cap(old_cap);
774 ret = cb(inode, cap, arg);
777 spin_lock(&session->s_cap_lock);
779 if (cap->ci == NULL) {
780 dout("iterate_session_caps finishing cap %p removal\n",
782 BUG_ON(cap->session != session);
783 list_del_init(&cap->session_caps);
784 session->s_nr_caps--;
786 old_cap = cap; /* put_cap it w/o locks held */
793 session->s_cap_iterator = NULL;
794 spin_unlock(&session->s_cap_lock);
799 ceph_put_cap(old_cap);
804 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
807 struct ceph_inode_info *ci = ceph_inode(inode);
808 dout("removing cap %p, ci is %p, inode is %p\n",
809 cap, ci, &ci->vfs_inode);
810 ceph_remove_cap(cap);
815 * caller must hold session s_mutex
817 static void remove_session_caps(struct ceph_mds_session *session)
819 dout("remove_session_caps on %p\n", session);
820 iterate_session_caps(session, remove_session_caps_cb, NULL);
821 BUG_ON(session->s_nr_caps > 0);
822 cleanup_cap_releases(session);
826 * wake up any threads waiting on this session's caps. if the cap is
827 * old (didn't get renewed on the client reconnect), remove it now.
829 * caller must hold s_mutex.
831 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
834 struct ceph_inode_info *ci = ceph_inode(inode);
836 wake_up(&ci->i_cap_wq);
838 spin_lock(&inode->i_lock);
839 ci->i_wanted_max_size = 0;
840 ci->i_requested_max_size = 0;
841 spin_unlock(&inode->i_lock);
846 static void wake_up_session_caps(struct ceph_mds_session *session,
849 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
850 iterate_session_caps(session, wake_up_session_cb,
851 (void *)(unsigned long)reconnect);
855 * Send periodic message to MDS renewing all currently held caps. The
856 * ack will reset the expiration for all caps from this session.
858 * caller holds s_mutex
860 static int send_renew_caps(struct ceph_mds_client *mdsc,
861 struct ceph_mds_session *session)
863 struct ceph_msg *msg;
866 if (time_after_eq(jiffies, session->s_cap_ttl) &&
867 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
868 pr_info("mds%d caps stale\n", session->s_mds);
869 session->s_renew_requested = jiffies;
871 /* do not try to renew caps until a recovering mds has reconnected
872 * with its clients. */
873 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
874 if (state < CEPH_MDS_STATE_RECONNECT) {
875 dout("send_renew_caps ignoring mds%d (%s)\n",
876 session->s_mds, ceph_mds_state_name(state));
880 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
881 ceph_mds_state_name(state));
882 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
883 ++session->s_renew_seq);
886 ceph_con_send(&session->s_con, msg);
891 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
893 * Called under session->s_mutex
895 static void renewed_caps(struct ceph_mds_client *mdsc,
896 struct ceph_mds_session *session, int is_renew)
901 spin_lock(&session->s_cap_lock);
902 was_stale = is_renew && (session->s_cap_ttl == 0 ||
903 time_after_eq(jiffies, session->s_cap_ttl));
905 session->s_cap_ttl = session->s_renew_requested +
906 mdsc->mdsmap->m_session_timeout*HZ;
909 if (time_before(jiffies, session->s_cap_ttl)) {
910 pr_info("mds%d caps renewed\n", session->s_mds);
913 pr_info("mds%d caps still stale\n", session->s_mds);
916 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
917 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
918 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
919 spin_unlock(&session->s_cap_lock);
922 wake_up_session_caps(session, 0);
926 * send a session close request
928 static int request_close_session(struct ceph_mds_client *mdsc,
929 struct ceph_mds_session *session)
931 struct ceph_msg *msg;
934 dout("request_close_session mds%d state %s seq %lld\n",
935 session->s_mds, session_state_name(session->s_state),
937 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
941 ceph_con_send(&session->s_con, msg);
946 * Called with s_mutex held.
948 static int __close_session(struct ceph_mds_client *mdsc,
949 struct ceph_mds_session *session)
951 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
953 session->s_state = CEPH_MDS_SESSION_CLOSING;
954 return request_close_session(mdsc, session);
960 * Because we can't cache an inode without one or more caps, we do
961 * this indirectly: if a cap is unused, we prune its aliases, at which
962 * point the inode will hopefully get dropped to.
964 * Yes, this is a bit sloppy. Our only real goal here is to respond to
965 * memory pressure from the MDS, though, so it needn't be perfect.
967 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
969 struct ceph_mds_session *session = arg;
970 struct ceph_inode_info *ci = ceph_inode(inode);
971 int used, oissued, mine;
973 if (session->s_trim_caps <= 0)
976 spin_lock(&inode->i_lock);
977 mine = cap->issued | cap->implemented;
978 used = __ceph_caps_used(ci);
979 oissued = __ceph_caps_issued_other(ci, cap);
981 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
982 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
983 ceph_cap_string(used));
984 if (ci->i_dirty_caps)
985 goto out; /* dirty caps */
986 if ((used & ~oissued) & mine)
987 goto out; /* we need these caps */
989 session->s_trim_caps--;
991 /* we aren't the only cap.. just remove us */
992 __ceph_remove_cap(cap);
994 /* try to drop referring dentries */
995 spin_unlock(&inode->i_lock);
996 d_prune_aliases(inode);
997 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
998 inode, cap, atomic_read(&inode->i_count));
1003 spin_unlock(&inode->i_lock);
1008 * Trim session cap count down to some max number.
1010 static int trim_caps(struct ceph_mds_client *mdsc,
1011 struct ceph_mds_session *session,
1014 int trim_caps = session->s_nr_caps - max_caps;
1016 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1017 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1018 if (trim_caps > 0) {
1019 session->s_trim_caps = trim_caps;
1020 iterate_session_caps(session, trim_caps_cb, session);
1021 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1022 session->s_mds, session->s_nr_caps, max_caps,
1023 trim_caps - session->s_trim_caps);
1024 session->s_trim_caps = 0;
1030 * Allocate cap_release messages. If there is a partially full message
1031 * in the queue, try to allocate enough to cover it's remainder, so that
1032 * we can send it immediately.
1034 * Called under s_mutex.
1036 static int add_cap_releases(struct ceph_mds_client *mdsc,
1037 struct ceph_mds_session *session,
1040 struct ceph_msg *msg;
1041 struct ceph_mds_cap_release *head;
1045 extra = mdsc->client->mount_args->cap_release_safety;
1047 spin_lock(&session->s_cap_lock);
1049 if (!list_empty(&session->s_cap_releases)) {
1050 msg = list_first_entry(&session->s_cap_releases,
1053 head = msg->front.iov_base;
1054 extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1057 while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1058 spin_unlock(&session->s_cap_lock);
1059 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1063 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1064 (int)msg->front.iov_len);
1065 head = msg->front.iov_base;
1066 head->num = cpu_to_le32(0);
1067 msg->front.iov_len = sizeof(*head);
1068 spin_lock(&session->s_cap_lock);
1069 list_add(&msg->list_head, &session->s_cap_releases);
1070 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1073 if (!list_empty(&session->s_cap_releases)) {
1074 msg = list_first_entry(&session->s_cap_releases,
1077 head = msg->front.iov_base;
1079 dout(" queueing non-full %p (%d)\n", msg,
1080 le32_to_cpu(head->num));
1081 list_move_tail(&msg->list_head,
1082 &session->s_cap_releases_done);
1083 session->s_num_cap_releases -=
1084 CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1088 spin_unlock(&session->s_cap_lock);
1094 * flush all dirty inode data to disk.
1096 * returns true if we've flushed through want_flush_seq
1098 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1102 dout("check_cap_flush want %lld\n", want_flush_seq);
1103 mutex_lock(&mdsc->mutex);
1104 for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1105 struct ceph_mds_session *session = mdsc->sessions[mds];
1109 get_session(session);
1110 mutex_unlock(&mdsc->mutex);
1112 mutex_lock(&session->s_mutex);
1113 if (!list_empty(&session->s_cap_flushing)) {
1114 struct ceph_inode_info *ci =
1115 list_entry(session->s_cap_flushing.next,
1116 struct ceph_inode_info,
1118 struct inode *inode = &ci->vfs_inode;
1120 spin_lock(&inode->i_lock);
1121 if (ci->i_cap_flush_seq <= want_flush_seq) {
1122 dout("check_cap_flush still flushing %p "
1123 "seq %lld <= %lld to mds%d\n", inode,
1124 ci->i_cap_flush_seq, want_flush_seq,
1128 spin_unlock(&inode->i_lock);
1130 mutex_unlock(&session->s_mutex);
1131 ceph_put_mds_session(session);
1135 mutex_lock(&mdsc->mutex);
1138 mutex_unlock(&mdsc->mutex);
1139 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1144 * called under s_mutex
1146 static void send_cap_releases(struct ceph_mds_client *mdsc,
1147 struct ceph_mds_session *session)
1149 struct ceph_msg *msg;
1151 dout("send_cap_releases mds%d\n", session->s_mds);
1153 spin_lock(&session->s_cap_lock);
1154 if (list_empty(&session->s_cap_releases_done))
1156 msg = list_first_entry(&session->s_cap_releases_done,
1157 struct ceph_msg, list_head);
1158 list_del_init(&msg->list_head);
1159 spin_unlock(&session->s_cap_lock);
1160 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1161 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1162 ceph_con_send(&session->s_con, msg);
1164 spin_unlock(&session->s_cap_lock);
1172 * Create an mds request.
1174 struct ceph_mds_request *
1175 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1177 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1180 return ERR_PTR(-ENOMEM);
1182 req->r_started = jiffies;
1183 req->r_resend_mds = -1;
1184 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1186 kref_init(&req->r_kref);
1187 INIT_LIST_HEAD(&req->r_wait);
1188 init_completion(&req->r_completion);
1189 init_completion(&req->r_safe_completion);
1190 INIT_LIST_HEAD(&req->r_unsafe_item);
1193 req->r_direct_mode = mode;
1198 * return oldest (lowest) request, tid in request tree, 0 if none.
1200 * called under mdsc->mutex.
1202 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1204 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1206 return rb_entry(rb_first(&mdsc->request_tree),
1207 struct ceph_mds_request, r_node);
1210 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1212 struct ceph_mds_request *req = __get_oldest_req(mdsc);
1220 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1221 * on build_path_from_dentry in fs/cifs/dir.c.
1223 * If @stop_on_nosnap, generate path relative to the first non-snapped
1226 * Encode hidden .snap dirs as a double /, i.e.
1227 * foo/.snap/bar -> foo//bar
1229 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1232 struct dentry *temp;
1237 return ERR_PTR(-EINVAL);
1241 for (temp = dentry; !IS_ROOT(temp);) {
1242 struct inode *inode = temp->d_inode;
1243 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1244 len++; /* slash only */
1245 else if (stop_on_nosnap && inode &&
1246 ceph_snap(inode) == CEPH_NOSNAP)
1249 len += 1 + temp->d_name.len;
1250 temp = temp->d_parent;
1252 pr_err("build_path_dentry corrupt dentry %p\n", dentry);
1253 return ERR_PTR(-EINVAL);
1257 len--; /* no leading '/' */
1259 path = kmalloc(len+1, GFP_NOFS);
1261 return ERR_PTR(-ENOMEM);
1263 path[pos] = 0; /* trailing null */
1264 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1265 struct inode *inode = temp->d_inode;
1267 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1268 dout("build_path_dentry path+%d: %p SNAPDIR\n",
1270 } else if (stop_on_nosnap && inode &&
1271 ceph_snap(inode) == CEPH_NOSNAP) {
1274 pos -= temp->d_name.len;
1277 strncpy(path + pos, temp->d_name.name,
1279 dout("build_path_dentry path+%d: %p '%.*s'\n",
1280 pos, temp, temp->d_name.len, path + pos);
1284 temp = temp->d_parent;
1286 pr_err("build_path_dentry corrupt dentry\n");
1288 return ERR_PTR(-EINVAL);
1292 pr_err("build_path_dentry did not end path lookup where "
1293 "expected, namelen is %d, pos is %d\n", len, pos);
1294 /* presumably this is only possible if racing with a
1295 rename of one of the parent directories (we can not
1296 lock the dentries above us to prevent this, but
1297 retrying should be harmless) */
1302 *base = ceph_ino(temp->d_inode);
1304 dout("build_path_dentry on %p %d built %llx '%.*s'\n",
1305 dentry, atomic_read(&dentry->d_count), *base, len, path);
1309 static int build_dentry_path(struct dentry *dentry,
1310 const char **ppath, int *ppathlen, u64 *pino,
1315 if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1316 *pino = ceph_ino(dentry->d_parent->d_inode);
1317 *ppath = dentry->d_name.name;
1318 *ppathlen = dentry->d_name.len;
1321 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1323 return PTR_ERR(path);
1329 static int build_inode_path(struct inode *inode,
1330 const char **ppath, int *ppathlen, u64 *pino,
1333 struct dentry *dentry;
1336 if (ceph_snap(inode) == CEPH_NOSNAP) {
1337 *pino = ceph_ino(inode);
1341 dentry = d_find_alias(inode);
1342 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1345 return PTR_ERR(path);
1352 * request arguments may be specified via an inode *, a dentry *, or
1353 * an explicit ino+path.
1355 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1356 const char *rpath, u64 rino,
1357 const char **ppath, int *pathlen,
1358 u64 *ino, int *freepath)
1363 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1364 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1366 } else if (rdentry) {
1367 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1368 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1373 *pathlen = strlen(rpath);
1374 dout(" path %.*s\n", *pathlen, rpath);
1381 * called under mdsc->mutex
1383 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1384 struct ceph_mds_request *req,
1387 struct ceph_msg *msg;
1388 struct ceph_mds_request_head *head;
1389 const char *path1 = NULL;
1390 const char *path2 = NULL;
1391 u64 ino1 = 0, ino2 = 0;
1392 int pathlen1 = 0, pathlen2 = 0;
1393 int freepath1 = 0, freepath2 = 0;
1399 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1400 req->r_path1, req->r_ino1.ino,
1401 &path1, &pathlen1, &ino1, &freepath1);
1407 ret = set_request_path_attr(NULL, req->r_old_dentry,
1408 req->r_path2, req->r_ino2.ino,
1409 &path2, &pathlen2, &ino2, &freepath2);
1415 len = sizeof(*head) +
1416 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1418 /* calculate (max) length for cap releases */
1419 len += sizeof(struct ceph_mds_request_release) *
1420 (!!req->r_inode_drop + !!req->r_dentry_drop +
1421 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1422 if (req->r_dentry_drop)
1423 len += req->r_dentry->d_name.len;
1424 if (req->r_old_dentry_drop)
1425 len += req->r_old_dentry->d_name.len;
1427 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, 0, 0, NULL);
1431 msg->hdr.tid = cpu_to_le64(req->r_tid);
1433 head = msg->front.iov_base;
1434 p = msg->front.iov_base + sizeof(*head);
1435 end = msg->front.iov_base + msg->front.iov_len;
1437 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1438 head->op = cpu_to_le32(req->r_op);
1439 head->caller_uid = cpu_to_le32(current_fsuid());
1440 head->caller_gid = cpu_to_le32(current_fsgid());
1441 head->args = req->r_args;
1443 ceph_encode_filepath(&p, end, ino1, path1);
1444 ceph_encode_filepath(&p, end, ino2, path2);
1448 if (req->r_inode_drop)
1449 releases += ceph_encode_inode_release(&p,
1450 req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1451 mds, req->r_inode_drop, req->r_inode_unless, 0);
1452 if (req->r_dentry_drop)
1453 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1454 mds, req->r_dentry_drop, req->r_dentry_unless);
1455 if (req->r_old_dentry_drop)
1456 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1457 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1458 if (req->r_old_inode_drop)
1459 releases += ceph_encode_inode_release(&p,
1460 req->r_old_dentry->d_inode,
1461 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1462 head->num_releases = cpu_to_le16(releases);
1465 msg->front.iov_len = p - msg->front.iov_base;
1466 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1468 msg->pages = req->r_pages;
1469 msg->nr_pages = req->r_num_pages;
1470 msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1471 msg->hdr.data_off = cpu_to_le16(0);
1475 kfree((char *)path2);
1478 kfree((char *)path1);
1484 * called under mdsc->mutex if error, under no mutex if
1487 static void complete_request(struct ceph_mds_client *mdsc,
1488 struct ceph_mds_request *req)
1490 if (req->r_callback)
1491 req->r_callback(mdsc, req);
1493 complete(&req->r_completion);
1497 * called under mdsc->mutex
1499 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1500 struct ceph_mds_request *req,
1503 struct ceph_mds_request_head *rhead;
1504 struct ceph_msg *msg;
1509 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1510 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1512 if (req->r_request) {
1513 ceph_msg_put(req->r_request);
1514 req->r_request = NULL;
1516 msg = create_request_message(mdsc, req, mds);
1518 req->r_reply = ERR_PTR(PTR_ERR(msg));
1519 complete_request(mdsc, req);
1520 return -PTR_ERR(msg);
1522 req->r_request = msg;
1524 rhead = msg->front.iov_base;
1525 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1526 if (req->r_got_unsafe)
1527 flags |= CEPH_MDS_FLAG_REPLAY;
1528 if (req->r_locked_dir)
1529 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1530 rhead->flags = cpu_to_le32(flags);
1531 rhead->num_fwd = req->r_num_fwd;
1532 rhead->num_retry = req->r_attempts - 1;
1534 dout(" r_locked_dir = %p\n", req->r_locked_dir);
1536 if (req->r_target_inode && req->r_got_unsafe)
1537 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1544 * send request, or put it on the appropriate wait list.
1546 static int __do_request(struct ceph_mds_client *mdsc,
1547 struct ceph_mds_request *req)
1549 struct ceph_mds_session *session = NULL;
1556 if (req->r_timeout &&
1557 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1558 dout("do_request timed out\n");
1563 mds = __choose_mds(mdsc, req);
1565 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1566 dout("do_request no mds or not active, waiting for map\n");
1567 list_add(&req->r_wait, &mdsc->waiting_for_map);
1571 /* get, open session */
1572 session = __ceph_lookup_mds_session(mdsc, mds);
1574 session = register_session(mdsc, mds);
1575 if (IS_ERR(session)) {
1576 err = PTR_ERR(session);
1580 dout("do_request mds%d session %p state %s\n", mds, session,
1581 session_state_name(session->s_state));
1582 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1583 session->s_state != CEPH_MDS_SESSION_HUNG) {
1584 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1585 session->s_state == CEPH_MDS_SESSION_CLOSING)
1586 __open_session(mdsc, session);
1587 list_add(&req->r_wait, &session->s_waiting);
1592 req->r_session = get_session(session);
1593 req->r_resend_mds = -1; /* forget any previous mds hint */
1595 if (req->r_request_started == 0) /* note request start time */
1596 req->r_request_started = jiffies;
1598 err = __prepare_send_request(mdsc, req, mds);
1600 ceph_msg_get(req->r_request);
1601 ceph_con_send(&session->s_con, req->r_request);
1605 ceph_put_mds_session(session);
1610 req->r_reply = ERR_PTR(err);
1611 complete_request(mdsc, req);
1616 * called under mdsc->mutex
1618 static void __wake_requests(struct ceph_mds_client *mdsc,
1619 struct list_head *head)
1621 struct ceph_mds_request *req, *nreq;
1623 list_for_each_entry_safe(req, nreq, head, r_wait) {
1624 list_del_init(&req->r_wait);
1625 __do_request(mdsc, req);
1630 * Wake up threads with requests pending for @mds, so that they can
1631 * resubmit their requests to a possibly different mds. If @all is set,
1632 * wake up if their requests has been forwarded to @mds, too.
1634 static void kick_requests(struct ceph_mds_client *mdsc, int mds, int all)
1636 struct ceph_mds_request *req;
1639 dout("kick_requests mds%d\n", mds);
1640 for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1641 req = rb_entry(p, struct ceph_mds_request, r_node);
1642 if (req->r_got_unsafe)
1644 if (req->r_session &&
1645 req->r_session->s_mds == mds) {
1646 dout(" kicking tid %llu\n", req->r_tid);
1647 put_request_session(req);
1648 __do_request(mdsc, req);
1653 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1654 struct ceph_mds_request *req)
1656 dout("submit_request on %p\n", req);
1657 mutex_lock(&mdsc->mutex);
1658 __register_request(mdsc, req, NULL);
1659 __do_request(mdsc, req);
1660 mutex_unlock(&mdsc->mutex);
1664 * Synchrously perform an mds request. Take care of all of the
1665 * session setup, forwarding, retry details.
1667 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1669 struct ceph_mds_request *req)
1673 dout("do_request on %p\n", req);
1675 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1677 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1678 if (req->r_locked_dir)
1679 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1680 if (req->r_old_dentry)
1682 ceph_inode(req->r_old_dentry->d_parent->d_inode),
1686 mutex_lock(&mdsc->mutex);
1687 __register_request(mdsc, req, dir);
1688 __do_request(mdsc, req);
1691 if (!req->r_reply) {
1692 mutex_unlock(&mdsc->mutex);
1693 if (req->r_timeout) {
1694 err = (long)wait_for_completion_interruptible_timeout(
1695 &req->r_completion, req->r_timeout);
1697 req->r_reply = ERR_PTR(-EIO);
1699 req->r_reply = ERR_PTR(err);
1701 err = wait_for_completion_interruptible(
1702 &req->r_completion);
1704 req->r_reply = ERR_PTR(err);
1706 mutex_lock(&mdsc->mutex);
1709 if (IS_ERR(req->r_reply)) {
1710 err = PTR_ERR(req->r_reply);
1711 req->r_reply = NULL;
1713 if (err == -ERESTARTSYS) {
1715 req->r_aborted = true;
1717 if (req->r_locked_dir &&
1718 (req->r_op & CEPH_MDS_OP_WRITE)) {
1719 struct ceph_inode_info *ci =
1720 ceph_inode(req->r_locked_dir);
1722 dout("aborted, clearing I_COMPLETE on %p\n",
1724 spin_lock(&req->r_locked_dir->i_lock);
1725 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1726 ci->i_release_count++;
1727 spin_unlock(&req->r_locked_dir->i_lock);
1730 /* clean up this request */
1731 __unregister_request(mdsc, req);
1732 if (!list_empty(&req->r_unsafe_item))
1733 list_del_init(&req->r_unsafe_item);
1734 complete(&req->r_safe_completion);
1736 } else if (req->r_err) {
1739 err = le32_to_cpu(req->r_reply_info.head->result);
1741 mutex_unlock(&mdsc->mutex);
1743 dout("do_request %p done, result %d\n", req, err);
1750 * We take the session mutex and parse and process the reply immediately.
1751 * This preserves the logical ordering of replies, capabilities, etc., sent
1752 * by the MDS as they are applied to our local cache.
1754 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1756 struct ceph_mds_client *mdsc = session->s_mdsc;
1757 struct ceph_mds_request *req;
1758 struct ceph_mds_reply_head *head = msg->front.iov_base;
1759 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
1762 int mds = session->s_mds;
1764 if (msg->front.iov_len < sizeof(*head)) {
1765 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1770 /* get request, session */
1771 tid = le64_to_cpu(msg->hdr.tid);
1772 mutex_lock(&mdsc->mutex);
1773 req = __lookup_request(mdsc, tid);
1775 dout("handle_reply on unknown tid %llu\n", tid);
1776 mutex_unlock(&mdsc->mutex);
1779 dout("handle_reply %p\n", req);
1781 /* correct session? */
1782 if (req->r_session != session) {
1783 pr_err("mdsc_handle_reply got %llu on session mds%d"
1784 " not mds%d\n", tid, session->s_mds,
1785 req->r_session ? req->r_session->s_mds : -1);
1786 mutex_unlock(&mdsc->mutex);
1791 if ((req->r_got_unsafe && !head->safe) ||
1792 (req->r_got_safe && head->safe)) {
1793 pr_warning("got a dup %s reply on %llu from mds%d\n",
1794 head->safe ? "safe" : "unsafe", tid, mds);
1795 mutex_unlock(&mdsc->mutex);
1799 result = le32_to_cpu(head->result);
1802 * Tolerate 2 consecutive ESTALEs from the same mds.
1803 * FIXME: we should be looking at the cap migrate_seq.
1805 if (result == -ESTALE) {
1806 req->r_direct_mode = USE_AUTH_MDS;
1808 if (req->r_num_stale <= 2) {
1809 __do_request(mdsc, req);
1810 mutex_unlock(&mdsc->mutex);
1814 req->r_num_stale = 0;
1818 req->r_got_safe = true;
1819 __unregister_request(mdsc, req);
1820 complete(&req->r_safe_completion);
1822 if (req->r_got_unsafe) {
1824 * We already handled the unsafe response, now do the
1825 * cleanup. No need to examine the response; the MDS
1826 * doesn't include any result info in the safe
1827 * response. And even if it did, there is nothing
1828 * useful we could do with a revised return value.
1830 dout("got safe reply %llu, mds%d\n", tid, mds);
1831 list_del_init(&req->r_unsafe_item);
1833 /* last unsafe request during umount? */
1834 if (mdsc->stopping && !__get_oldest_req(mdsc))
1835 complete(&mdsc->safe_umount_waiters);
1836 mutex_unlock(&mdsc->mutex);
1841 BUG_ON(req->r_reply);
1844 req->r_got_unsafe = true;
1845 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1848 dout("handle_reply tid %lld result %d\n", tid, result);
1849 rinfo = &req->r_reply_info;
1850 err = parse_reply_info(msg, rinfo);
1851 mutex_unlock(&mdsc->mutex);
1853 mutex_lock(&session->s_mutex);
1855 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1861 if (rinfo->snapblob_len) {
1862 down_write(&mdsc->snap_rwsem);
1863 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1864 rinfo->snapblob + rinfo->snapblob_len,
1865 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1866 downgrade_write(&mdsc->snap_rwsem);
1868 down_read(&mdsc->snap_rwsem);
1871 /* insert trace into our cache */
1872 err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1874 if (result == 0 && rinfo->dir_nr)
1875 ceph_readdir_prepopulate(req, req->r_session);
1876 ceph_unreserve_caps(&req->r_caps_reservation);
1879 up_read(&mdsc->snap_rwsem);
1888 add_cap_releases(mdsc, req->r_session, -1);
1889 mutex_unlock(&session->s_mutex);
1891 /* kick calling process */
1892 complete_request(mdsc, req);
1894 ceph_mdsc_put_request(req);
1901 * handle mds notification that our request has been forwarded.
1903 static void handle_forward(struct ceph_mds_client *mdsc,
1904 struct ceph_mds_session *session,
1905 struct ceph_msg *msg)
1907 struct ceph_mds_request *req;
1908 u64 tid = le64_to_cpu(msg->hdr.tid);
1912 void *p = msg->front.iov_base;
1913 void *end = p + msg->front.iov_len;
1915 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1916 next_mds = ceph_decode_32(&p);
1917 fwd_seq = ceph_decode_32(&p);
1919 mutex_lock(&mdsc->mutex);
1920 req = __lookup_request(mdsc, tid);
1922 dout("forward %llu to mds%d - req dne\n", tid, next_mds);
1923 goto out; /* dup reply? */
1926 if (fwd_seq <= req->r_num_fwd) {
1927 dout("forward %llu to mds%d - old seq %d <= %d\n",
1928 tid, next_mds, req->r_num_fwd, fwd_seq);
1930 /* resend. forward race not possible; mds would drop */
1931 dout("forward %llu to mds%d (we resend)\n", tid, next_mds);
1932 req->r_num_fwd = fwd_seq;
1933 req->r_resend_mds = next_mds;
1934 put_request_session(req);
1935 __do_request(mdsc, req);
1937 ceph_mdsc_put_request(req);
1939 mutex_unlock(&mdsc->mutex);
1943 pr_err("mdsc_handle_forward decode error err=%d\n", err);
1947 * handle a mds session control message
1949 static void handle_session(struct ceph_mds_session *session,
1950 struct ceph_msg *msg)
1952 struct ceph_mds_client *mdsc = session->s_mdsc;
1955 int mds = session->s_mds;
1956 struct ceph_mds_session_head *h = msg->front.iov_base;
1960 if (msg->front.iov_len != sizeof(*h))
1962 op = le32_to_cpu(h->op);
1963 seq = le64_to_cpu(h->seq);
1965 mutex_lock(&mdsc->mutex);
1966 if (op == CEPH_SESSION_CLOSE)
1967 __unregister_session(mdsc, session);
1968 /* FIXME: this ttl calculation is generous */
1969 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
1970 mutex_unlock(&mdsc->mutex);
1972 mutex_lock(&session->s_mutex);
1974 dout("handle_session mds%d %s %p state %s seq %llu\n",
1975 mds, ceph_session_op_name(op), session,
1976 session_state_name(session->s_state), seq);
1978 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
1979 session->s_state = CEPH_MDS_SESSION_OPEN;
1980 pr_info("mds%d came back\n", session->s_mds);
1984 case CEPH_SESSION_OPEN:
1985 session->s_state = CEPH_MDS_SESSION_OPEN;
1986 renewed_caps(mdsc, session, 0);
1989 __close_session(mdsc, session);
1992 case CEPH_SESSION_RENEWCAPS:
1993 if (session->s_renew_seq == seq)
1994 renewed_caps(mdsc, session, 1);
1997 case CEPH_SESSION_CLOSE:
1998 remove_session_caps(session);
1999 wake = 1; /* for good measure */
2000 complete(&mdsc->session_close_waiters);
2001 kick_requests(mdsc, mds, 0); /* cur only */
2004 case CEPH_SESSION_STALE:
2005 pr_info("mds%d caps went stale, renewing\n",
2007 spin_lock(&session->s_cap_lock);
2008 session->s_cap_gen++;
2009 session->s_cap_ttl = 0;
2010 spin_unlock(&session->s_cap_lock);
2011 send_renew_caps(mdsc, session);
2014 case CEPH_SESSION_RECALL_STATE:
2015 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2019 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2023 mutex_unlock(&session->s_mutex);
2025 mutex_lock(&mdsc->mutex);
2026 __wake_requests(mdsc, &session->s_waiting);
2027 mutex_unlock(&mdsc->mutex);
2032 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2033 (int)msg->front.iov_len);
2040 * called under session->mutex.
2042 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2043 struct ceph_mds_session *session)
2045 struct ceph_mds_request *req, *nreq;
2048 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2050 mutex_lock(&mdsc->mutex);
2051 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2052 err = __prepare_send_request(mdsc, req, session->s_mds);
2054 ceph_msg_get(req->r_request);
2055 ceph_con_send(&session->s_con, req->r_request);
2058 mutex_unlock(&mdsc->mutex);
2062 * Encode information about a cap for a reconnect with the MDS.
2064 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2067 struct ceph_mds_cap_reconnect rec;
2068 struct ceph_inode_info *ci;
2069 struct ceph_pagelist *pagelist = arg;
2073 struct dentry *dentry;
2077 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2078 inode, ceph_vinop(inode), cap, cap->cap_id,
2079 ceph_cap_string(cap->issued));
2080 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2084 dentry = d_find_alias(inode);
2086 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2088 err = PTR_ERR(path);
2095 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2099 spin_lock(&inode->i_lock);
2100 cap->seq = 0; /* reset cap seq */
2101 cap->issue_seq = 0; /* and issue_seq */
2102 rec.cap_id = cpu_to_le64(cap->cap_id);
2103 rec.pathbase = cpu_to_le64(pathbase);
2104 rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2105 rec.issued = cpu_to_le32(cap->issued);
2106 rec.size = cpu_to_le64(inode->i_size);
2107 ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2108 ceph_encode_timespec(&rec.atime, &inode->i_atime);
2109 rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2110 spin_unlock(&inode->i_lock);
2112 err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2122 * If an MDS fails and recovers, clients need to reconnect in order to
2123 * reestablish shared state. This includes all caps issued through
2124 * this session _and_ the snap_realm hierarchy. Because it's not
2125 * clear which snap realms the mds cares about, we send everything we
2126 * know about.. that ensures we'll then get any new info the
2127 * recovering MDS might have.
2129 * This is a relatively heavyweight operation, but it's rare.
2131 * called with mdsc->mutex held.
2133 static void send_mds_reconnect(struct ceph_mds_client *mdsc, int mds)
2135 struct ceph_mds_session *session = NULL;
2136 struct ceph_msg *reply;
2139 struct ceph_pagelist *pagelist;
2141 pr_info("reconnect to recovering mds%d\n", mds);
2143 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2145 goto fail_nopagelist;
2146 ceph_pagelist_init(pagelist);
2148 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, 0, 0, NULL);
2149 if (IS_ERR(reply)) {
2150 err = PTR_ERR(reply);
2155 session = __ceph_lookup_mds_session(mdsc, mds);
2156 mutex_unlock(&mdsc->mutex); /* drop lock for duration */
2159 mutex_lock(&session->s_mutex);
2161 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2164 ceph_con_open(&session->s_con,
2165 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2167 /* replay unsafe requests */
2168 replay_unsafe_requests(mdsc, session);
2170 dout("no session for mds%d, will send short reconnect\n",
2174 down_read(&mdsc->snap_rwsem);
2178 dout("session %p state %s\n", session,
2179 session_state_name(session->s_state));
2181 /* traverse this session's caps */
2182 err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2185 err = iterate_session_caps(session, encode_caps_cb, pagelist);
2190 * snaprealms. we provide mds with the ino, seq (version), and
2191 * parent for all of our realms. If the mds has any newer info,
2194 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2195 struct ceph_snap_realm *realm =
2196 rb_entry(p, struct ceph_snap_realm, node);
2197 struct ceph_mds_snaprealm_reconnect sr_rec;
2199 dout(" adding snap realm %llx seq %lld parent %llx\n",
2200 realm->ino, realm->seq, realm->parent_ino);
2201 sr_rec.ino = cpu_to_le64(realm->ino);
2202 sr_rec.seq = cpu_to_le64(realm->seq);
2203 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2204 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2210 reply->pagelist = pagelist;
2211 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2212 reply->nr_pages = calc_pages_for(0, pagelist->length);
2213 ceph_con_send(&session->s_con, reply);
2216 session->s_state = CEPH_MDS_SESSION_OPEN;
2217 __wake_requests(mdsc, &session->s_waiting);
2221 up_read(&mdsc->snap_rwsem);
2223 mutex_unlock(&session->s_mutex);
2224 ceph_put_mds_session(session);
2226 mutex_lock(&mdsc->mutex);
2230 ceph_msg_put(reply);
2232 ceph_pagelist_release(pagelist);
2235 pr_err("ENOMEM preparing reconnect for mds%d\n", mds);
2241 * compare old and new mdsmaps, kicking requests
2242 * and closing out old connections as necessary
2244 * called under mdsc->mutex.
2246 static void check_new_map(struct ceph_mds_client *mdsc,
2247 struct ceph_mdsmap *newmap,
2248 struct ceph_mdsmap *oldmap)
2251 int oldstate, newstate;
2252 struct ceph_mds_session *s;
2254 dout("check_new_map new %u old %u\n",
2255 newmap->m_epoch, oldmap->m_epoch);
2257 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2258 if (mdsc->sessions[i] == NULL)
2260 s = mdsc->sessions[i];
2261 oldstate = ceph_mdsmap_get_state(oldmap, i);
2262 newstate = ceph_mdsmap_get_state(newmap, i);
2264 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2265 i, ceph_mds_state_name(oldstate),
2266 ceph_mds_state_name(newstate),
2267 session_state_name(s->s_state));
2269 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2270 ceph_mdsmap_get_addr(newmap, i),
2271 sizeof(struct ceph_entity_addr))) {
2272 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2273 /* the session never opened, just close it
2275 __wake_requests(mdsc, &s->s_waiting);
2276 __unregister_session(mdsc, s);
2279 mutex_unlock(&mdsc->mutex);
2280 mutex_lock(&s->s_mutex);
2281 mutex_lock(&mdsc->mutex);
2282 ceph_con_close(&s->s_con);
2283 mutex_unlock(&s->s_mutex);
2284 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2287 /* kick any requests waiting on the recovering mds */
2288 kick_requests(mdsc, i, 1);
2289 } else if (oldstate == newstate) {
2290 continue; /* nothing new with this mds */
2296 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2297 newstate >= CEPH_MDS_STATE_RECONNECT)
2298 send_mds_reconnect(mdsc, i);
2301 * kick requests on any mds that has gone active.
2303 * kick requests on cur or forwarder: we may have sent
2304 * the request to mds1, mds1 told us it forwarded it
2305 * to mds2, but then we learn mds1 failed and can't be
2306 * sure it successfully forwarded our request before
2309 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2310 newstate >= CEPH_MDS_STATE_ACTIVE) {
2311 pr_info("mds%d reconnect completed\n", s->s_mds);
2312 kick_requests(mdsc, i, 1);
2313 ceph_kick_flushing_caps(mdsc, s);
2314 wake_up_session_caps(s, 1);
2326 * caller must hold session s_mutex, dentry->d_lock
2328 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2330 struct ceph_dentry_info *di = ceph_dentry(dentry);
2332 ceph_put_mds_session(di->lease_session);
2333 di->lease_session = NULL;
2336 static void handle_lease(struct ceph_mds_client *mdsc,
2337 struct ceph_mds_session *session,
2338 struct ceph_msg *msg)
2340 struct super_block *sb = mdsc->client->sb;
2341 struct inode *inode;
2342 struct ceph_inode_info *ci;
2343 struct dentry *parent, *dentry;
2344 struct ceph_dentry_info *di;
2345 int mds = session->s_mds;
2346 struct ceph_mds_lease *h = msg->front.iov_base;
2347 struct ceph_vino vino;
2352 dout("handle_lease from mds%d\n", mds);
2355 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2357 vino.ino = le64_to_cpu(h->ino);
2358 vino.snap = CEPH_NOSNAP;
2359 mask = le16_to_cpu(h->mask);
2360 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2361 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2362 if (dname.len != get_unaligned_le32(h+1))
2365 mutex_lock(&session->s_mutex);
2369 inode = ceph_find_inode(sb, vino);
2370 dout("handle_lease '%s', mask %d, ino %llx %p\n",
2371 ceph_lease_op_name(h->action), mask, vino.ino, inode);
2372 if (inode == NULL) {
2373 dout("handle_lease no inode %llx\n", vino.ino);
2376 ci = ceph_inode(inode);
2379 parent = d_find_alias(inode);
2381 dout("no parent dentry on inode %p\n", inode);
2383 goto release; /* hrm... */
2385 dname.hash = full_name_hash(dname.name, dname.len);
2386 dentry = d_lookup(parent, &dname);
2391 spin_lock(&dentry->d_lock);
2392 di = ceph_dentry(dentry);
2393 switch (h->action) {
2394 case CEPH_MDS_LEASE_REVOKE:
2395 if (di && di->lease_session == session) {
2396 h->seq = cpu_to_le32(di->lease_seq);
2397 __ceph_mdsc_drop_dentry_lease(dentry);
2402 case CEPH_MDS_LEASE_RENEW:
2403 if (di && di->lease_session == session &&
2404 di->lease_gen == session->s_cap_gen &&
2405 di->lease_renew_from &&
2406 di->lease_renew_after == 0) {
2407 unsigned long duration =
2408 le32_to_cpu(h->duration_ms) * HZ / 1000;
2410 di->lease_seq = le32_to_cpu(h->seq);
2411 dentry->d_time = di->lease_renew_from + duration;
2412 di->lease_renew_after = di->lease_renew_from +
2414 di->lease_renew_from = 0;
2418 spin_unlock(&dentry->d_lock);
2425 /* let's just reuse the same message */
2426 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2428 ceph_con_send(&session->s_con, msg);
2432 mutex_unlock(&session->s_mutex);
2436 pr_err("corrupt lease message\n");
2440 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2441 struct inode *inode,
2442 struct dentry *dentry, char action,
2445 struct ceph_msg *msg;
2446 struct ceph_mds_lease *lease;
2447 int len = sizeof(*lease) + sizeof(u32);
2450 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2451 inode, dentry, ceph_lease_op_name(action), session->s_mds);
2452 dnamelen = dentry->d_name.len;
2455 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, 0, 0, NULL);
2458 lease = msg->front.iov_base;
2459 lease->action = action;
2460 lease->mask = cpu_to_le16(CEPH_LOCK_DN);
2461 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2462 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2463 lease->seq = cpu_to_le32(seq);
2464 put_unaligned_le32(dnamelen, lease + 1);
2465 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2468 * if this is a preemptive lease RELEASE, no need to
2469 * flush request stream, since the actual request will
2472 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2474 ceph_con_send(&session->s_con, msg);
2478 * Preemptively release a lease we expect to invalidate anyway.
2479 * Pass @inode always, @dentry is optional.
2481 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2482 struct dentry *dentry, int mask)
2484 struct ceph_dentry_info *di;
2485 struct ceph_mds_session *session;
2488 BUG_ON(inode == NULL);
2489 BUG_ON(dentry == NULL);
2490 BUG_ON(mask != CEPH_LOCK_DN);
2492 /* is dentry lease valid? */
2493 spin_lock(&dentry->d_lock);
2494 di = ceph_dentry(dentry);
2495 if (!di || !di->lease_session ||
2496 di->lease_session->s_mds < 0 ||
2497 di->lease_gen != di->lease_session->s_cap_gen ||
2498 !time_before(jiffies, dentry->d_time)) {
2499 dout("lease_release inode %p dentry %p -- "
2501 inode, dentry, mask);
2502 spin_unlock(&dentry->d_lock);
2506 /* we do have a lease on this dentry; note mds and seq */
2507 session = ceph_get_mds_session(di->lease_session);
2508 seq = di->lease_seq;
2509 __ceph_mdsc_drop_dentry_lease(dentry);
2510 spin_unlock(&dentry->d_lock);
2512 dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2513 inode, dentry, mask, session->s_mds);
2514 ceph_mdsc_lease_send_msg(session, inode, dentry,
2515 CEPH_MDS_LEASE_RELEASE, seq);
2516 ceph_put_mds_session(session);
2520 * drop all leases (and dentry refs) in preparation for umount
2522 static void drop_leases(struct ceph_mds_client *mdsc)
2526 dout("drop_leases\n");
2527 mutex_lock(&mdsc->mutex);
2528 for (i = 0; i < mdsc->max_sessions; i++) {
2529 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2532 mutex_unlock(&mdsc->mutex);
2533 mutex_lock(&s->s_mutex);
2534 mutex_unlock(&s->s_mutex);
2535 ceph_put_mds_session(s);
2536 mutex_lock(&mdsc->mutex);
2538 mutex_unlock(&mdsc->mutex);
2544 * delayed work -- periodically trim expired leases, renew caps with mds
2546 static void schedule_delayed(struct ceph_mds_client *mdsc)
2549 unsigned hz = round_jiffies_relative(HZ * delay);
2550 schedule_delayed_work(&mdsc->delayed_work, hz);
2553 static void delayed_work(struct work_struct *work)
2556 struct ceph_mds_client *mdsc =
2557 container_of(work, struct ceph_mds_client, delayed_work.work);
2561 dout("mdsc delayed_work\n");
2562 ceph_check_delayed_caps(mdsc);
2564 mutex_lock(&mdsc->mutex);
2565 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2566 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2567 mdsc->last_renew_caps);
2569 mdsc->last_renew_caps = jiffies;
2571 for (i = 0; i < mdsc->max_sessions; i++) {
2572 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2575 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2576 dout("resending session close request for mds%d\n",
2578 request_close_session(mdsc, s);
2579 ceph_put_mds_session(s);
2582 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2583 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2584 s->s_state = CEPH_MDS_SESSION_HUNG;
2585 pr_info("mds%d hung\n", s->s_mds);
2588 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2589 /* this mds is failed or recovering, just wait */
2590 ceph_put_mds_session(s);
2593 mutex_unlock(&mdsc->mutex);
2595 mutex_lock(&s->s_mutex);
2597 send_renew_caps(mdsc, s);
2599 ceph_con_keepalive(&s->s_con);
2600 add_cap_releases(mdsc, s, -1);
2601 send_cap_releases(mdsc, s);
2602 mutex_unlock(&s->s_mutex);
2603 ceph_put_mds_session(s);
2605 mutex_lock(&mdsc->mutex);
2607 mutex_unlock(&mdsc->mutex);
2609 schedule_delayed(mdsc);
2613 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2615 mdsc->client = client;
2616 mutex_init(&mdsc->mutex);
2617 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2618 init_completion(&mdsc->safe_umount_waiters);
2619 init_completion(&mdsc->session_close_waiters);
2620 INIT_LIST_HEAD(&mdsc->waiting_for_map);
2621 mdsc->sessions = NULL;
2622 mdsc->max_sessions = 0;
2624 init_rwsem(&mdsc->snap_rwsem);
2625 mdsc->snap_realms = RB_ROOT;
2626 INIT_LIST_HEAD(&mdsc->snap_empty);
2627 spin_lock_init(&mdsc->snap_empty_lock);
2629 mdsc->request_tree = RB_ROOT;
2630 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2631 mdsc->last_renew_caps = jiffies;
2632 INIT_LIST_HEAD(&mdsc->cap_delay_list);
2633 spin_lock_init(&mdsc->cap_delay_lock);
2634 INIT_LIST_HEAD(&mdsc->snap_flush_list);
2635 spin_lock_init(&mdsc->snap_flush_lock);
2636 mdsc->cap_flush_seq = 0;
2637 INIT_LIST_HEAD(&mdsc->cap_dirty);
2638 mdsc->num_cap_flushing = 0;
2639 spin_lock_init(&mdsc->cap_dirty_lock);
2640 init_waitqueue_head(&mdsc->cap_flushing_wq);
2641 spin_lock_init(&mdsc->dentry_lru_lock);
2642 INIT_LIST_HEAD(&mdsc->dentry_lru);
2647 * Wait for safe replies on open mds requests. If we time out, drop
2648 * all requests from the tree to avoid dangling dentry refs.
2650 static void wait_requests(struct ceph_mds_client *mdsc)
2652 struct ceph_mds_request *req;
2653 struct ceph_client *client = mdsc->client;
2655 mutex_lock(&mdsc->mutex);
2656 if (__get_oldest_req(mdsc)) {
2657 mutex_unlock(&mdsc->mutex);
2659 dout("wait_requests waiting for requests\n");
2660 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2661 client->mount_args->mount_timeout * HZ);
2663 /* tear down remaining requests */
2664 mutex_lock(&mdsc->mutex);
2665 while ((req = __get_oldest_req(mdsc))) {
2666 dout("wait_requests timed out on tid %llu\n",
2668 __unregister_request(mdsc, req);
2671 mutex_unlock(&mdsc->mutex);
2672 dout("wait_requests done\n");
2676 * called before mount is ro, and before dentries are torn down.
2677 * (hmm, does this still race with new lookups?)
2679 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2681 dout("pre_umount\n");
2685 ceph_flush_dirty_caps(mdsc);
2686 wait_requests(mdsc);
2690 * wait for all write mds requests to flush.
2692 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2694 struct ceph_mds_request *req = NULL, *nextreq;
2697 mutex_lock(&mdsc->mutex);
2698 dout("wait_unsafe_requests want %lld\n", want_tid);
2700 req = __get_oldest_req(mdsc);
2701 while (req && req->r_tid <= want_tid) {
2702 /* find next request */
2703 n = rb_next(&req->r_node);
2705 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2708 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2710 ceph_mdsc_get_request(req);
2712 ceph_mdsc_get_request(nextreq);
2713 mutex_unlock(&mdsc->mutex);
2714 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
2715 req->r_tid, want_tid);
2716 wait_for_completion(&req->r_safe_completion);
2717 mutex_lock(&mdsc->mutex);
2718 ceph_mdsc_put_request(req);
2720 break; /* next dne before, so we're done! */
2721 if (RB_EMPTY_NODE(&nextreq->r_node)) {
2722 /* next request was removed from tree */
2723 ceph_mdsc_put_request(nextreq);
2726 ceph_mdsc_put_request(nextreq); /* won't go away */
2730 mutex_unlock(&mdsc->mutex);
2731 dout("wait_unsafe_requests done\n");
2734 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2736 u64 want_tid, want_flush;
2739 mutex_lock(&mdsc->mutex);
2740 want_tid = mdsc->last_tid;
2741 want_flush = mdsc->cap_flush_seq;
2742 mutex_unlock(&mdsc->mutex);
2743 dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2745 ceph_flush_dirty_caps(mdsc);
2747 wait_unsafe_requests(mdsc, want_tid);
2748 wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2753 * called after sb is ro.
2755 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2757 struct ceph_mds_session *session;
2760 struct ceph_client *client = mdsc->client;
2761 unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2763 dout("close_sessions\n");
2765 mutex_lock(&mdsc->mutex);
2767 /* close sessions */
2769 while (time_before(jiffies, started + timeout)) {
2770 dout("closing sessions\n");
2772 for (i = 0; i < mdsc->max_sessions; i++) {
2773 session = __ceph_lookup_mds_session(mdsc, i);
2776 mutex_unlock(&mdsc->mutex);
2777 mutex_lock(&session->s_mutex);
2778 __close_session(mdsc, session);
2779 mutex_unlock(&session->s_mutex);
2780 ceph_put_mds_session(session);
2781 mutex_lock(&mdsc->mutex);
2787 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2790 dout("waiting for sessions to close\n");
2791 mutex_unlock(&mdsc->mutex);
2792 wait_for_completion_timeout(&mdsc->session_close_waiters,
2794 mutex_lock(&mdsc->mutex);
2797 /* tear down remaining sessions */
2798 for (i = 0; i < mdsc->max_sessions; i++) {
2799 if (mdsc->sessions[i]) {
2800 session = get_session(mdsc->sessions[i]);
2801 __unregister_session(mdsc, session);
2802 mutex_unlock(&mdsc->mutex);
2803 mutex_lock(&session->s_mutex);
2804 remove_session_caps(session);
2805 mutex_unlock(&session->s_mutex);
2806 ceph_put_mds_session(session);
2807 mutex_lock(&mdsc->mutex);
2811 WARN_ON(!list_empty(&mdsc->cap_delay_list));
2813 mutex_unlock(&mdsc->mutex);
2815 ceph_cleanup_empty_realms(mdsc);
2817 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2822 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2825 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2827 ceph_mdsmap_destroy(mdsc->mdsmap);
2828 kfree(mdsc->sessions);
2833 * handle mds map update.
2835 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2839 void *p = msg->front.iov_base;
2840 void *end = p + msg->front.iov_len;
2841 struct ceph_mdsmap *newmap, *oldmap;
2842 struct ceph_fsid fsid;
2845 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2846 ceph_decode_copy(&p, &fsid, sizeof(fsid));
2847 if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2849 epoch = ceph_decode_32(&p);
2850 maplen = ceph_decode_32(&p);
2851 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2853 /* do we need it? */
2854 ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2855 mutex_lock(&mdsc->mutex);
2856 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2857 dout("handle_map epoch %u <= our %u\n",
2858 epoch, mdsc->mdsmap->m_epoch);
2859 mutex_unlock(&mdsc->mutex);
2863 newmap = ceph_mdsmap_decode(&p, end);
2864 if (IS_ERR(newmap)) {
2865 err = PTR_ERR(newmap);
2869 /* swap into place */
2871 oldmap = mdsc->mdsmap;
2872 mdsc->mdsmap = newmap;
2873 check_new_map(mdsc, newmap, oldmap);
2874 ceph_mdsmap_destroy(oldmap);
2876 mdsc->mdsmap = newmap; /* first mds map */
2878 mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2880 __wake_requests(mdsc, &mdsc->waiting_for_map);
2882 mutex_unlock(&mdsc->mutex);
2883 schedule_delayed(mdsc);
2887 mutex_unlock(&mdsc->mutex);
2889 pr_err("error decoding mdsmap %d\n", err);
2893 static struct ceph_connection *con_get(struct ceph_connection *con)
2895 struct ceph_mds_session *s = con->private;
2897 if (get_session(s)) {
2898 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
2901 dout("mdsc con_get %p FAIL\n", s);
2905 static void con_put(struct ceph_connection *con)
2907 struct ceph_mds_session *s = con->private;
2909 ceph_put_mds_session(s);
2910 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
2914 * if the client is unresponsive for long enough, the mds will kill
2915 * the session entirely.
2917 static void peer_reset(struct ceph_connection *con)
2919 struct ceph_mds_session *s = con->private;
2921 pr_err("mds%d gave us the boot. IMPLEMENT RECONNECT.\n",
2925 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2927 struct ceph_mds_session *s = con->private;
2928 struct ceph_mds_client *mdsc = s->s_mdsc;
2929 int type = le16_to_cpu(msg->hdr.type);
2931 mutex_lock(&mdsc->mutex);
2932 if (__verify_registered_session(mdsc, s) < 0) {
2933 mutex_unlock(&mdsc->mutex);
2936 mutex_unlock(&mdsc->mutex);
2939 case CEPH_MSG_MDS_MAP:
2940 ceph_mdsc_handle_map(mdsc, msg);
2942 case CEPH_MSG_CLIENT_SESSION:
2943 handle_session(s, msg);
2945 case CEPH_MSG_CLIENT_REPLY:
2946 handle_reply(s, msg);
2948 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
2949 handle_forward(mdsc, s, msg);
2951 case CEPH_MSG_CLIENT_CAPS:
2952 ceph_handle_caps(s, msg);
2954 case CEPH_MSG_CLIENT_SNAP:
2955 ceph_handle_snap(mdsc, s, msg);
2957 case CEPH_MSG_CLIENT_LEASE:
2958 handle_lease(mdsc, s, msg);
2962 pr_err("received unknown message type %d %s\n", type,
2963 ceph_msg_type_name(type));
2972 static int get_authorizer(struct ceph_connection *con,
2973 void **buf, int *len, int *proto,
2974 void **reply_buf, int *reply_len, int force_new)
2976 struct ceph_mds_session *s = con->private;
2977 struct ceph_mds_client *mdsc = s->s_mdsc;
2978 struct ceph_auth_client *ac = mdsc->client->monc.auth;
2981 if (force_new && s->s_authorizer) {
2982 ac->ops->destroy_authorizer(ac, s->s_authorizer);
2983 s->s_authorizer = NULL;
2985 if (s->s_authorizer == NULL) {
2986 if (ac->ops->create_authorizer) {
2987 ret = ac->ops->create_authorizer(
2988 ac, CEPH_ENTITY_TYPE_MDS,
2990 &s->s_authorizer_buf,
2991 &s->s_authorizer_buf_len,
2992 &s->s_authorizer_reply_buf,
2993 &s->s_authorizer_reply_buf_len);
2999 *proto = ac->protocol;
3000 *buf = s->s_authorizer_buf;
3001 *len = s->s_authorizer_buf_len;
3002 *reply_buf = s->s_authorizer_reply_buf;
3003 *reply_len = s->s_authorizer_reply_buf_len;
3008 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3010 struct ceph_mds_session *s = con->private;
3011 struct ceph_mds_client *mdsc = s->s_mdsc;
3012 struct ceph_auth_client *ac = mdsc->client->monc.auth;
3014 return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3017 static int invalidate_authorizer(struct ceph_connection *con)
3019 struct ceph_mds_session *s = con->private;
3020 struct ceph_mds_client *mdsc = s->s_mdsc;
3021 struct ceph_auth_client *ac = mdsc->client->monc.auth;
3023 if (ac->ops->invalidate_authorizer)
3024 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3026 return ceph_monc_validate_auth(&mdsc->client->monc);
3029 const static struct ceph_connection_operations mds_con_ops = {
3032 .dispatch = dispatch,
3033 .get_authorizer = get_authorizer,
3034 .verify_authorizer_reply = verify_authorizer_reply,
3035 .invalidate_authorizer = invalidate_authorizer,
3036 .peer_reset = peer_reset,