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36 * are no longer valid.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_RECONNECTING: return "reconnecting";
259 default: return "???";
263 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
265 if (atomic_inc_not_zero(&s->s_ref)) {
266 dout("mdsc get_session %p %d -> %d\n", s,
267 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
270 dout("mdsc get_session %p 0 -- FAIL", s);
275 void ceph_put_mds_session(struct ceph_mds_session *s)
277 dout("mdsc put_session %p %d -> %d\n", s,
278 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
279 if (atomic_dec_and_test(&s->s_ref)) {
281 s->s_mdsc->client->monc.auth->ops->destroy_authorizer(
282 s->s_mdsc->client->monc.auth, s->s_authorizer);
288 * called under mdsc->mutex
290 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
293 struct ceph_mds_session *session;
295 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
297 session = mdsc->sessions[mds];
298 dout("lookup_mds_session %p %d\n", session,
299 atomic_read(&session->s_ref));
300 get_session(session);
304 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
306 if (mds >= mdsc->max_sessions)
308 return mdsc->sessions[mds];
312 * create+register a new session for given mds.
313 * called under mdsc->mutex.
315 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
318 struct ceph_mds_session *s;
320 s = kzalloc(sizeof(*s), GFP_NOFS);
323 s->s_state = CEPH_MDS_SESSION_NEW;
326 mutex_init(&s->s_mutex);
328 ceph_con_init(mdsc->client->msgr, &s->s_con);
329 s->s_con.private = s;
330 s->s_con.ops = &mds_con_ops;
331 s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
332 s->s_con.peer_name.num = cpu_to_le64(mds);
334 spin_lock_init(&s->s_cap_lock);
337 s->s_renew_requested = 0;
339 INIT_LIST_HEAD(&s->s_caps);
342 atomic_set(&s->s_ref, 1);
343 INIT_LIST_HEAD(&s->s_waiting);
344 INIT_LIST_HEAD(&s->s_unsafe);
345 s->s_num_cap_releases = 0;
346 s->s_iterating_caps = false;
347 INIT_LIST_HEAD(&s->s_cap_releases);
348 INIT_LIST_HEAD(&s->s_cap_releases_done);
349 INIT_LIST_HEAD(&s->s_cap_flushing);
350 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
352 dout("register_session mds%d\n", mds);
353 if (mds >= mdsc->max_sessions) {
354 int newmax = 1 << get_count_order(mds+1);
355 struct ceph_mds_session **sa;
357 dout("register_session realloc to %d\n", newmax);
358 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
361 if (mdsc->sessions) {
362 memcpy(sa, mdsc->sessions,
363 mdsc->max_sessions * sizeof(void *));
364 kfree(mdsc->sessions);
367 mdsc->max_sessions = newmax;
369 mdsc->sessions[mds] = s;
370 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
372 ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
378 return ERR_PTR(-ENOMEM);
382 * called under mdsc->mutex
384 static void unregister_session(struct ceph_mds_client *mdsc,
385 struct ceph_mds_session *s)
387 dout("unregister_session mds%d %p\n", s->s_mds, s);
388 mdsc->sessions[s->s_mds] = NULL;
389 ceph_con_close(&s->s_con);
390 ceph_put_mds_session(s);
394 * drop session refs in request.
396 * should be last request ref, or hold mdsc->mutex
398 static void put_request_session(struct ceph_mds_request *req)
400 if (req->r_session) {
401 ceph_put_mds_session(req->r_session);
402 req->r_session = NULL;
406 void ceph_mdsc_release_request(struct kref *kref)
408 struct ceph_mds_request *req = container_of(kref,
409 struct ceph_mds_request,
412 ceph_msg_put(req->r_request);
414 ceph_msg_put(req->r_reply);
415 destroy_reply_info(&req->r_reply_info);
418 ceph_put_cap_refs(ceph_inode(req->r_inode),
422 if (req->r_locked_dir)
423 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
425 if (req->r_target_inode)
426 iput(req->r_target_inode);
429 if (req->r_old_dentry) {
431 ceph_inode(req->r_old_dentry->d_parent->d_inode),
433 dput(req->r_old_dentry);
437 put_request_session(req);
438 ceph_unreserve_caps(&req->r_caps_reservation);
443 * lookup session, bump ref if found.
445 * called under mdsc->mutex.
447 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
450 struct ceph_mds_request *req;
451 req = radix_tree_lookup(&mdsc->request_tree, tid);
453 ceph_mdsc_get_request(req);
458 * Register an in-flight request, and assign a tid. Link to directory
459 * are modifying (if any).
461 * Called under mdsc->mutex.
463 static void __register_request(struct ceph_mds_client *mdsc,
464 struct ceph_mds_request *req,
467 req->r_tid = ++mdsc->last_tid;
469 ceph_reserve_caps(&req->r_caps_reservation, req->r_num_caps);
470 dout("__register_request %p tid %lld\n", req, req->r_tid);
471 ceph_mdsc_get_request(req);
472 radix_tree_insert(&mdsc->request_tree, req->r_tid, (void *)req);
475 struct ceph_inode_info *ci = ceph_inode(dir);
477 spin_lock(&ci->i_unsafe_lock);
478 req->r_unsafe_dir = dir;
479 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
480 spin_unlock(&ci->i_unsafe_lock);
484 static void __unregister_request(struct ceph_mds_client *mdsc,
485 struct ceph_mds_request *req)
487 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
488 radix_tree_delete(&mdsc->request_tree, req->r_tid);
489 ceph_mdsc_put_request(req);
491 if (req->r_unsafe_dir) {
492 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
494 spin_lock(&ci->i_unsafe_lock);
495 list_del_init(&req->r_unsafe_dir_item);
496 spin_unlock(&ci->i_unsafe_lock);
501 * Choose mds to send request to next. If there is a hint set in the
502 * request (e.g., due to a prior forward hint from the mds), use that.
503 * Otherwise, consult frag tree and/or caps to identify the
504 * appropriate mds. If all else fails, choose randomly.
506 * Called under mdsc->mutex.
508 static int __choose_mds(struct ceph_mds_client *mdsc,
509 struct ceph_mds_request *req)
512 struct ceph_inode_info *ci;
513 struct ceph_cap *cap;
514 int mode = req->r_direct_mode;
516 u32 hash = req->r_direct_hash;
517 bool is_hash = req->r_direct_is_hash;
520 * is there a specific mds we should try? ignore hint if we have
521 * no session and the mds is not up (active or recovering).
523 if (req->r_resend_mds >= 0 &&
524 (__have_session(mdsc, req->r_resend_mds) ||
525 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
526 dout("choose_mds using resend_mds mds%d\n",
528 return req->r_resend_mds;
531 if (mode == USE_RANDOM_MDS)
536 inode = req->r_inode;
537 } else if (req->r_dentry) {
538 if (req->r_dentry->d_inode) {
539 inode = req->r_dentry->d_inode;
541 inode = req->r_dentry->d_parent->d_inode;
542 hash = req->r_dentry->d_name.hash;
546 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
550 ci = ceph_inode(inode);
552 if (is_hash && S_ISDIR(inode->i_mode)) {
553 struct ceph_inode_frag frag;
556 ceph_choose_frag(ci, hash, &frag, &found);
558 if (mode == USE_ANY_MDS && frag.ndist > 0) {
561 /* choose a random replica */
562 get_random_bytes(&r, 1);
565 dout("choose_mds %p %llx.%llx "
566 "frag %u mds%d (%d/%d)\n",
567 inode, ceph_vinop(inode),
573 /* since this file/dir wasn't known to be
574 * replicated, then we want to look for the
575 * authoritative mds. */
578 /* choose auth mds */
580 dout("choose_mds %p %llx.%llx "
581 "frag %u mds%d (auth)\n",
582 inode, ceph_vinop(inode), frag.frag, mds);
588 spin_lock(&inode->i_lock);
590 if (mode == USE_AUTH_MDS)
591 cap = ci->i_auth_cap;
592 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
593 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
595 spin_unlock(&inode->i_lock);
598 mds = cap->session->s_mds;
599 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
600 inode, ceph_vinop(inode), mds,
601 cap == ci->i_auth_cap ? "auth " : "", cap);
602 spin_unlock(&inode->i_lock);
606 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
607 dout("choose_mds chose random mds%d\n", mds);
615 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
617 struct ceph_msg *msg;
618 struct ceph_mds_session_head *h;
620 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), 0, 0, NULL);
622 pr_err("create_session_msg ENOMEM creating msg\n");
623 return ERR_PTR(PTR_ERR(msg));
625 h = msg->front.iov_base;
626 h->op = cpu_to_le32(op);
627 h->seq = cpu_to_le64(seq);
632 * send session open request.
634 * called under mdsc->mutex
636 static int __open_session(struct ceph_mds_client *mdsc,
637 struct ceph_mds_session *session)
639 struct ceph_msg *msg;
641 int mds = session->s_mds;
644 /* wait for mds to go active? */
645 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
646 dout("open_session to mds%d (%s)\n", mds,
647 ceph_mds_state_name(mstate));
648 session->s_state = CEPH_MDS_SESSION_OPENING;
649 session->s_renew_requested = jiffies;
651 /* send connect message */
652 msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
657 ceph_con_send(&session->s_con, msg);
668 * Free preallocated cap messages assigned to this session
670 static void cleanup_cap_releases(struct ceph_mds_session *session)
672 struct ceph_msg *msg;
674 spin_lock(&session->s_cap_lock);
675 while (!list_empty(&session->s_cap_releases)) {
676 msg = list_first_entry(&session->s_cap_releases,
677 struct ceph_msg, list_head);
678 list_del_init(&msg->list_head);
681 while (!list_empty(&session->s_cap_releases_done)) {
682 msg = list_first_entry(&session->s_cap_releases_done,
683 struct ceph_msg, list_head);
684 list_del_init(&msg->list_head);
687 spin_unlock(&session->s_cap_lock);
691 * Helper to safely iterate over all caps associated with a session.
693 * caller must hold session s_mutex
695 static int iterate_session_caps(struct ceph_mds_session *session,
696 int (*cb)(struct inode *, struct ceph_cap *,
699 struct ceph_cap *cap, *ncap;
703 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
704 spin_lock(&session->s_cap_lock);
705 session->s_iterating_caps = true;
706 list_for_each_entry_safe(cap, ncap, &session->s_caps, session_caps) {
707 inode = igrab(&cap->ci->vfs_inode);
710 spin_unlock(&session->s_cap_lock);
711 ret = cb(inode, cap, arg);
713 spin_lock(&session->s_cap_lock);
719 session->s_iterating_caps = false;
720 spin_unlock(&session->s_cap_lock);
724 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
727 struct ceph_inode_info *ci = ceph_inode(inode);
728 dout("removing cap %p, ci is %p, inode is %p\n",
729 cap, ci, &ci->vfs_inode);
730 ceph_remove_cap(cap);
735 * caller must hold session s_mutex
737 static void remove_session_caps(struct ceph_mds_session *session)
739 dout("remove_session_caps on %p\n", session);
740 iterate_session_caps(session, remove_session_caps_cb, NULL);
741 BUG_ON(session->s_nr_caps > 0);
742 cleanup_cap_releases(session);
746 * wake up any threads waiting on this session's caps. if the cap is
747 * old (didn't get renewed on the client reconnect), remove it now.
749 * caller must hold s_mutex.
751 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
754 struct ceph_inode_info *ci = ceph_inode(inode);
756 wake_up(&ci->i_cap_wq);
758 spin_lock(&inode->i_lock);
759 ci->i_wanted_max_size = 0;
760 ci->i_requested_max_size = 0;
761 spin_unlock(&inode->i_lock);
766 static void wake_up_session_caps(struct ceph_mds_session *session,
769 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
770 iterate_session_caps(session, wake_up_session_cb,
771 (void *)(unsigned long)reconnect);
775 * Send periodic message to MDS renewing all currently held caps. The
776 * ack will reset the expiration for all caps from this session.
778 * caller holds s_mutex
780 static int send_renew_caps(struct ceph_mds_client *mdsc,
781 struct ceph_mds_session *session)
783 struct ceph_msg *msg;
786 if (time_after_eq(jiffies, session->s_cap_ttl) &&
787 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
788 pr_info("mds%d caps stale\n", session->s_mds);
790 /* do not try to renew caps until a recovering mds has reconnected
791 * with its clients. */
792 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
793 if (state < CEPH_MDS_STATE_RECONNECT) {
794 dout("send_renew_caps ignoring mds%d (%s)\n",
795 session->s_mds, ceph_mds_state_name(state));
799 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
800 ceph_mds_state_name(state));
801 session->s_renew_requested = jiffies;
802 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
803 ++session->s_renew_seq);
806 ceph_con_send(&session->s_con, msg);
811 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
813 * Called under session->s_mutex
815 static void renewed_caps(struct ceph_mds_client *mdsc,
816 struct ceph_mds_session *session, int is_renew)
821 spin_lock(&session->s_cap_lock);
822 was_stale = is_renew && (session->s_cap_ttl == 0 ||
823 time_after_eq(jiffies, session->s_cap_ttl));
825 session->s_cap_ttl = session->s_renew_requested +
826 mdsc->mdsmap->m_session_timeout*HZ;
829 if (time_before(jiffies, session->s_cap_ttl)) {
830 pr_info("mds%d caps renewed\n", session->s_mds);
833 pr_info("mds%d caps still stale\n", session->s_mds);
836 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
837 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
838 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
839 spin_unlock(&session->s_cap_lock);
842 wake_up_session_caps(session, 0);
846 * send a session close request
848 static int request_close_session(struct ceph_mds_client *mdsc,
849 struct ceph_mds_session *session)
851 struct ceph_msg *msg;
854 dout("request_close_session mds%d state %s seq %lld\n",
855 session->s_mds, session_state_name(session->s_state),
857 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
861 ceph_con_send(&session->s_con, msg);
866 * Called with s_mutex held.
868 static int __close_session(struct ceph_mds_client *mdsc,
869 struct ceph_mds_session *session)
871 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
873 session->s_state = CEPH_MDS_SESSION_CLOSING;
874 return request_close_session(mdsc, session);
880 * Because we can't cache an inode without one or more caps, we do
881 * this indirectly: if a cap is unused, we prune its aliases, at which
882 * point the inode will hopefully get dropped to.
884 * Yes, this is a bit sloppy. Our only real goal here is to respond to
885 * memory pressure from the MDS, though, so it needn't be perfect.
887 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
889 struct ceph_mds_session *session = arg;
890 struct ceph_inode_info *ci = ceph_inode(inode);
891 int used, oissued, mine;
893 if (session->s_trim_caps <= 0)
896 spin_lock(&inode->i_lock);
897 mine = cap->issued | cap->implemented;
898 used = __ceph_caps_used(ci);
899 oissued = __ceph_caps_issued_other(ci, cap);
901 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
902 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
903 ceph_cap_string(used));
904 if (ci->i_dirty_caps)
905 goto out; /* dirty caps */
906 if ((used & ~oissued) & mine)
907 goto out; /* we need these caps */
909 session->s_trim_caps--;
911 /* we aren't the only cap.. just remove us */
912 __ceph_remove_cap(cap, NULL);
914 /* try to drop referring dentries */
915 spin_unlock(&inode->i_lock);
916 d_prune_aliases(inode);
917 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
918 inode, cap, atomic_read(&inode->i_count));
923 spin_unlock(&inode->i_lock);
928 * Trim session cap count down to some max number.
930 static int trim_caps(struct ceph_mds_client *mdsc,
931 struct ceph_mds_session *session,
934 int trim_caps = session->s_nr_caps - max_caps;
936 dout("trim_caps mds%d start: %d / %d, trim %d\n",
937 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
939 session->s_trim_caps = trim_caps;
940 iterate_session_caps(session, trim_caps_cb, session);
941 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
942 session->s_mds, session->s_nr_caps, max_caps,
943 trim_caps - session->s_trim_caps);
944 session->s_trim_caps = 0;
950 * Allocate cap_release messages. If there is a partially full message
951 * in the queue, try to allocate enough to cover it's remainder, so that
952 * we can send it immediately.
954 * Called under s_mutex.
956 static int add_cap_releases(struct ceph_mds_client *mdsc,
957 struct ceph_mds_session *session,
960 struct ceph_msg *msg;
961 struct ceph_mds_cap_release *head;
965 extra = mdsc->client->mount_args->cap_release_safety;
967 spin_lock(&session->s_cap_lock);
969 if (!list_empty(&session->s_cap_releases)) {
970 msg = list_first_entry(&session->s_cap_releases,
973 head = msg->front.iov_base;
974 extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
977 while (session->s_num_cap_releases < session->s_nr_caps + extra) {
978 spin_unlock(&session->s_cap_lock);
979 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
983 dout("add_cap_releases %p msg %p now %d\n", session, msg,
984 (int)msg->front.iov_len);
985 head = msg->front.iov_base;
986 head->num = cpu_to_le32(0);
987 msg->front.iov_len = sizeof(*head);
988 spin_lock(&session->s_cap_lock);
989 list_add(&msg->list_head, &session->s_cap_releases);
990 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
993 if (!list_empty(&session->s_cap_releases)) {
994 msg = list_first_entry(&session->s_cap_releases,
997 head = msg->front.iov_base;
999 dout(" queueing non-full %p (%d)\n", msg,
1000 le32_to_cpu(head->num));
1001 list_move_tail(&msg->list_head,
1002 &session->s_cap_releases_done);
1003 session->s_num_cap_releases -=
1004 CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1008 spin_unlock(&session->s_cap_lock);
1014 * flush all dirty inode data to disk.
1016 * returns true if we've flushed through want_flush_seq
1018 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1022 dout("check_cap_flush want %lld\n", want_flush_seq);
1023 mutex_lock(&mdsc->mutex);
1024 for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1025 struct ceph_mds_session *session = mdsc->sessions[mds];
1029 get_session(session);
1030 mutex_unlock(&mdsc->mutex);
1032 mutex_lock(&session->s_mutex);
1033 if (!list_empty(&session->s_cap_flushing)) {
1034 struct ceph_inode_info *ci =
1035 list_entry(session->s_cap_flushing.next,
1036 struct ceph_inode_info,
1038 struct inode *inode = &ci->vfs_inode;
1040 spin_lock(&inode->i_lock);
1041 if (ci->i_cap_flush_seq <= want_flush_seq) {
1042 dout("check_cap_flush still flushing %p "
1043 "seq %lld <= %lld to mds%d\n", inode,
1044 ci->i_cap_flush_seq, want_flush_seq,
1048 spin_unlock(&inode->i_lock);
1050 mutex_unlock(&session->s_mutex);
1051 ceph_put_mds_session(session);
1055 mutex_lock(&mdsc->mutex);
1058 mutex_unlock(&mdsc->mutex);
1059 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1064 * called under s_mutex
1066 static void send_cap_releases(struct ceph_mds_client *mdsc,
1067 struct ceph_mds_session *session)
1069 struct ceph_msg *msg;
1071 dout("send_cap_releases mds%d\n", session->s_mds);
1073 spin_lock(&session->s_cap_lock);
1074 if (list_empty(&session->s_cap_releases_done))
1076 msg = list_first_entry(&session->s_cap_releases_done,
1077 struct ceph_msg, list_head);
1078 list_del_init(&msg->list_head);
1079 spin_unlock(&session->s_cap_lock);
1080 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1081 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1082 ceph_con_send(&session->s_con, msg);
1084 spin_unlock(&session->s_cap_lock);
1092 * Create an mds request.
1094 struct ceph_mds_request *
1095 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1097 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1100 return ERR_PTR(-ENOMEM);
1102 req->r_started = jiffies;
1103 req->r_resend_mds = -1;
1104 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1106 kref_init(&req->r_kref);
1107 INIT_LIST_HEAD(&req->r_wait);
1108 init_completion(&req->r_completion);
1109 init_completion(&req->r_safe_completion);
1110 INIT_LIST_HEAD(&req->r_unsafe_item);
1113 req->r_direct_mode = mode;
1118 * return oldest (lowest) tid in request tree, 0 if none.
1120 * called under mdsc->mutex.
1122 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1124 struct ceph_mds_request *first;
1125 if (radix_tree_gang_lookup(&mdsc->request_tree,
1126 (void **)&first, 0, 1) <= 0)
1128 return first->r_tid;
1132 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1133 * on build_path_from_dentry in fs/cifs/dir.c.
1135 * If @stop_on_nosnap, generate path relative to the first non-snapped
1138 * Encode hidden .snap dirs as a double /, i.e.
1139 * foo/.snap/bar -> foo//bar
1141 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1144 struct dentry *temp;
1149 return ERR_PTR(-EINVAL);
1153 for (temp = dentry; !IS_ROOT(temp);) {
1154 struct inode *inode = temp->d_inode;
1155 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1156 len++; /* slash only */
1157 else if (stop_on_nosnap && inode &&
1158 ceph_snap(inode) == CEPH_NOSNAP)
1161 len += 1 + temp->d_name.len;
1162 temp = temp->d_parent;
1164 pr_err("build_path_dentry corrupt dentry %p\n", dentry);
1165 return ERR_PTR(-EINVAL);
1169 len--; /* no leading '/' */
1171 path = kmalloc(len+1, GFP_NOFS);
1173 return ERR_PTR(-ENOMEM);
1175 path[pos] = 0; /* trailing null */
1176 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1177 struct inode *inode = temp->d_inode;
1179 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1180 dout("build_path_dentry path+%d: %p SNAPDIR\n",
1182 } else if (stop_on_nosnap && inode &&
1183 ceph_snap(inode) == CEPH_NOSNAP) {
1186 pos -= temp->d_name.len;
1189 strncpy(path + pos, temp->d_name.name,
1191 dout("build_path_dentry path+%d: %p '%.*s'\n",
1192 pos, temp, temp->d_name.len, path + pos);
1196 temp = temp->d_parent;
1198 pr_err("build_path_dentry corrupt dentry\n");
1200 return ERR_PTR(-EINVAL);
1204 pr_err("build_path_dentry did not end path lookup where "
1205 "expected, namelen is %d, pos is %d\n", len, pos);
1206 /* presumably this is only possible if racing with a
1207 rename of one of the parent directories (we can not
1208 lock the dentries above us to prevent this, but
1209 retrying should be harmless) */
1214 *base = ceph_ino(temp->d_inode);
1216 dout("build_path_dentry on %p %d built %llx '%.*s'\n",
1217 dentry, atomic_read(&dentry->d_count), *base, len, path);
1221 static int build_dentry_path(struct dentry *dentry,
1222 const char **ppath, int *ppathlen, u64 *pino,
1227 if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1228 *pino = ceph_ino(dentry->d_parent->d_inode);
1229 *ppath = dentry->d_name.name;
1230 *ppathlen = dentry->d_name.len;
1233 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1235 return PTR_ERR(path);
1241 static int build_inode_path(struct inode *inode,
1242 const char **ppath, int *ppathlen, u64 *pino,
1245 struct dentry *dentry;
1248 if (ceph_snap(inode) == CEPH_NOSNAP) {
1249 *pino = ceph_ino(inode);
1253 dentry = d_find_alias(inode);
1254 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1257 return PTR_ERR(path);
1264 * request arguments may be specified via an inode *, a dentry *, or
1265 * an explicit ino+path.
1267 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1268 const char *rpath, u64 rino,
1269 const char **ppath, int *pathlen,
1270 u64 *ino, int *freepath)
1275 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1276 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1278 } else if (rdentry) {
1279 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1280 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1285 *pathlen = strlen(rpath);
1286 dout(" path %.*s\n", *pathlen, rpath);
1293 * called under mdsc->mutex
1295 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1296 struct ceph_mds_request *req,
1299 struct ceph_msg *msg;
1300 struct ceph_mds_request_head *head;
1301 const char *path1 = NULL;
1302 const char *path2 = NULL;
1303 u64 ino1 = 0, ino2 = 0;
1304 int pathlen1 = 0, pathlen2 = 0;
1305 int freepath1 = 0, freepath2 = 0;
1311 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1312 req->r_path1, req->r_ino1.ino,
1313 &path1, &pathlen1, &ino1, &freepath1);
1319 ret = set_request_path_attr(NULL, req->r_old_dentry,
1320 req->r_path2, req->r_ino2.ino,
1321 &path2, &pathlen2, &ino2, &freepath2);
1327 len = sizeof(*head) +
1328 pathlen1 + pathlen2 + 2*(sizeof(u32) + sizeof(u64));
1330 /* calculate (max) length for cap releases */
1331 len += sizeof(struct ceph_mds_request_release) *
1332 (!!req->r_inode_drop + !!req->r_dentry_drop +
1333 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1334 if (req->r_dentry_drop)
1335 len += req->r_dentry->d_name.len;
1336 if (req->r_old_dentry_drop)
1337 len += req->r_old_dentry->d_name.len;
1339 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, 0, 0, NULL);
1343 msg->hdr.tid = cpu_to_le64(req->r_tid);
1345 head = msg->front.iov_base;
1346 p = msg->front.iov_base + sizeof(*head);
1347 end = msg->front.iov_base + msg->front.iov_len;
1349 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1350 head->op = cpu_to_le32(req->r_op);
1351 head->caller_uid = cpu_to_le32(current_fsuid());
1352 head->caller_gid = cpu_to_le32(current_fsgid());
1353 head->args = req->r_args;
1355 ceph_encode_filepath(&p, end, ino1, path1);
1356 ceph_encode_filepath(&p, end, ino2, path2);
1360 if (req->r_inode_drop)
1361 releases += ceph_encode_inode_release(&p,
1362 req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1363 mds, req->r_inode_drop, req->r_inode_unless, 0);
1364 if (req->r_dentry_drop)
1365 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1366 mds, req->r_dentry_drop, req->r_dentry_unless);
1367 if (req->r_old_dentry_drop)
1368 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1369 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1370 if (req->r_old_inode_drop)
1371 releases += ceph_encode_inode_release(&p,
1372 req->r_old_dentry->d_inode,
1373 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1374 head->num_releases = cpu_to_le16(releases);
1377 msg->front.iov_len = p - msg->front.iov_base;
1378 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1380 msg->pages = req->r_pages;
1381 msg->nr_pages = req->r_num_pages;
1382 msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1383 msg->hdr.data_off = cpu_to_le16(0);
1387 kfree((char *)path2);
1390 kfree((char *)path1);
1396 * called under mdsc->mutex if error, under no mutex if
1399 static void complete_request(struct ceph_mds_client *mdsc,
1400 struct ceph_mds_request *req)
1402 if (req->r_callback)
1403 req->r_callback(mdsc, req);
1405 complete(&req->r_completion);
1409 * called under mdsc->mutex
1411 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1412 struct ceph_mds_request *req,
1415 struct ceph_mds_request_head *rhead;
1416 struct ceph_msg *msg;
1421 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1422 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1424 if (req->r_request) {
1425 ceph_msg_put(req->r_request);
1426 req->r_request = NULL;
1428 msg = create_request_message(mdsc, req, mds);
1430 req->r_reply = ERR_PTR(PTR_ERR(msg));
1431 complete_request(mdsc, req);
1432 return -PTR_ERR(msg);
1434 req->r_request = msg;
1436 rhead = msg->front.iov_base;
1437 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1438 if (req->r_got_unsafe)
1439 flags |= CEPH_MDS_FLAG_REPLAY;
1440 if (req->r_locked_dir)
1441 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1442 rhead->flags = cpu_to_le32(flags);
1443 rhead->num_fwd = req->r_num_fwd;
1444 rhead->num_retry = req->r_attempts - 1;
1446 dout(" r_locked_dir = %p\n", req->r_locked_dir);
1448 if (req->r_target_inode && req->r_got_unsafe)
1449 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1456 * send request, or put it on the appropriate wait list.
1458 static int __do_request(struct ceph_mds_client *mdsc,
1459 struct ceph_mds_request *req)
1461 struct ceph_mds_session *session = NULL;
1468 if (req->r_timeout &&
1469 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1470 dout("do_request timed out\n");
1475 mds = __choose_mds(mdsc, req);
1477 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1478 dout("do_request no mds or not active, waiting for map\n");
1479 list_add(&req->r_wait, &mdsc->waiting_for_map);
1483 /* get, open session */
1484 session = __ceph_lookup_mds_session(mdsc, mds);
1486 session = register_session(mdsc, mds);
1487 dout("do_request mds%d session %p state %s\n", mds, session,
1488 session_state_name(session->s_state));
1489 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1490 session->s_state != CEPH_MDS_SESSION_HUNG) {
1491 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1492 session->s_state == CEPH_MDS_SESSION_CLOSING)
1493 __open_session(mdsc, session);
1494 list_add(&req->r_wait, &session->s_waiting);
1499 req->r_session = get_session(session);
1500 req->r_resend_mds = -1; /* forget any previous mds hint */
1502 if (req->r_request_started == 0) /* note request start time */
1503 req->r_request_started = jiffies;
1505 err = __prepare_send_request(mdsc, req, mds);
1507 ceph_msg_get(req->r_request);
1508 ceph_con_send(&session->s_con, req->r_request);
1512 ceph_put_mds_session(session);
1517 req->r_reply = ERR_PTR(err);
1518 complete_request(mdsc, req);
1523 * called under mdsc->mutex
1525 static void __wake_requests(struct ceph_mds_client *mdsc,
1526 struct list_head *head)
1528 struct ceph_mds_request *req, *nreq;
1530 list_for_each_entry_safe(req, nreq, head, r_wait) {
1531 list_del_init(&req->r_wait);
1532 __do_request(mdsc, req);
1537 * Wake up threads with requests pending for @mds, so that they can
1538 * resubmit their requests to a possibly different mds. If @all is set,
1539 * wake up if their requests has been forwarded to @mds, too.
1541 static void kick_requests(struct ceph_mds_client *mdsc, int mds, int all)
1543 struct ceph_mds_request *reqs[10];
1547 dout("kick_requests mds%d\n", mds);
1548 while (nexttid <= mdsc->last_tid) {
1549 got = radix_tree_gang_lookup(&mdsc->request_tree,
1550 (void **)&reqs, nexttid, 10);
1553 nexttid = reqs[got-1]->r_tid + 1;
1554 for (i = 0; i < got; i++) {
1555 if (reqs[i]->r_got_unsafe)
1557 if (reqs[i]->r_session &&
1558 reqs[i]->r_session->s_mds == mds) {
1559 dout(" kicking tid %llu\n", reqs[i]->r_tid);
1560 put_request_session(reqs[i]);
1561 __do_request(mdsc, reqs[i]);
1567 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1568 struct ceph_mds_request *req)
1570 dout("submit_request on %p\n", req);
1571 mutex_lock(&mdsc->mutex);
1572 __register_request(mdsc, req, NULL);
1573 __do_request(mdsc, req);
1574 mutex_unlock(&mdsc->mutex);
1578 * Synchrously perform an mds request. Take care of all of the
1579 * session setup, forwarding, retry details.
1581 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1583 struct ceph_mds_request *req)
1587 dout("do_request on %p\n", req);
1589 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1591 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1592 if (req->r_locked_dir)
1593 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1594 if (req->r_old_dentry)
1596 ceph_inode(req->r_old_dentry->d_parent->d_inode),
1600 mutex_lock(&mdsc->mutex);
1601 __register_request(mdsc, req, dir);
1602 __do_request(mdsc, req);
1605 if (!req->r_reply) {
1606 mutex_unlock(&mdsc->mutex);
1607 if (req->r_timeout) {
1608 err = (long)wait_for_completion_interruptible_timeout(
1609 &req->r_completion, req->r_timeout);
1611 req->r_reply = ERR_PTR(-EIO);
1613 req->r_reply = ERR_PTR(err);
1615 err = wait_for_completion_interruptible(
1616 &req->r_completion);
1618 req->r_reply = ERR_PTR(err);
1620 mutex_lock(&mdsc->mutex);
1623 if (IS_ERR(req->r_reply)) {
1624 err = PTR_ERR(req->r_reply);
1625 req->r_reply = NULL;
1627 if (err == -ERESTARTSYS) {
1629 req->r_aborted = true;
1631 if (req->r_locked_dir &&
1632 (req->r_op & CEPH_MDS_OP_WRITE)) {
1633 struct ceph_inode_info *ci =
1634 ceph_inode(req->r_locked_dir);
1636 dout("aborted, clearing I_COMPLETE on %p\n",
1638 spin_lock(&req->r_locked_dir->i_lock);
1639 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1640 ci->i_release_count++;
1641 spin_unlock(&req->r_locked_dir->i_lock);
1644 /* clean up this request */
1645 __unregister_request(mdsc, req);
1646 if (!list_empty(&req->r_unsafe_item))
1647 list_del_init(&req->r_unsafe_item);
1648 complete(&req->r_safe_completion);
1650 } else if (req->r_err) {
1653 err = le32_to_cpu(req->r_reply_info.head->result);
1655 mutex_unlock(&mdsc->mutex);
1657 dout("do_request %p done, result %d\n", req, err);
1664 * We take the session mutex and parse and process the reply immediately.
1665 * This preserves the logical ordering of replies, capabilities, etc., sent
1666 * by the MDS as they are applied to our local cache.
1668 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1670 struct ceph_mds_client *mdsc = session->s_mdsc;
1671 struct ceph_mds_request *req;
1672 struct ceph_mds_reply_head *head = msg->front.iov_base;
1673 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
1678 if (msg->hdr.src.name.type != CEPH_ENTITY_TYPE_MDS)
1680 if (msg->front.iov_len < sizeof(*head)) {
1681 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1686 /* get request, session */
1687 tid = le64_to_cpu(msg->hdr.tid);
1688 mutex_lock(&mdsc->mutex);
1689 req = __lookup_request(mdsc, tid);
1691 dout("handle_reply on unknown tid %llu\n", tid);
1692 mutex_unlock(&mdsc->mutex);
1695 dout("handle_reply %p\n", req);
1696 mds = le64_to_cpu(msg->hdr.src.name.num);
1698 /* correct session? */
1699 if (!req->r_session && req->r_session != session) {
1700 pr_err("mdsc_handle_reply got %llu on session mds%d"
1701 " not mds%d\n", tid, session->s_mds,
1702 req->r_session ? req->r_session->s_mds : -1);
1703 mutex_unlock(&mdsc->mutex);
1708 if ((req->r_got_unsafe && !head->safe) ||
1709 (req->r_got_safe && head->safe)) {
1710 pr_warning("got a dup %s reply on %llu from mds%d\n",
1711 head->safe ? "safe" : "unsafe", tid, mds);
1712 mutex_unlock(&mdsc->mutex);
1716 result = le32_to_cpu(head->result);
1719 * Tolerate 2 consecutive ESTALEs from the same mds.
1720 * FIXME: we should be looking at the cap migrate_seq.
1722 if (result == -ESTALE) {
1723 req->r_direct_mode = USE_AUTH_MDS;
1725 if (req->r_num_stale <= 2) {
1726 __do_request(mdsc, req);
1727 mutex_unlock(&mdsc->mutex);
1731 req->r_num_stale = 0;
1735 req->r_got_safe = true;
1736 __unregister_request(mdsc, req);
1737 complete(&req->r_safe_completion);
1739 if (req->r_got_unsafe) {
1741 * We already handled the unsafe response, now do the
1742 * cleanup. No need to examine the response; the MDS
1743 * doesn't include any result info in the safe
1744 * response. And even if it did, there is nothing
1745 * useful we could do with a revised return value.
1747 dout("got safe reply %llu, mds%d\n", tid, mds);
1748 list_del_init(&req->r_unsafe_item);
1750 /* last unsafe request during umount? */
1751 if (mdsc->stopping && !__get_oldest_tid(mdsc))
1752 complete(&mdsc->safe_umount_waiters);
1753 mutex_unlock(&mdsc->mutex);
1758 BUG_ON(req->r_reply);
1761 req->r_got_unsafe = true;
1762 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1765 dout("handle_reply tid %lld result %d\n", tid, result);
1766 rinfo = &req->r_reply_info;
1767 err = parse_reply_info(msg, rinfo);
1768 mutex_unlock(&mdsc->mutex);
1770 mutex_lock(&session->s_mutex);
1772 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1778 if (rinfo->snapblob_len) {
1779 down_write(&mdsc->snap_rwsem);
1780 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1781 rinfo->snapblob + rinfo->snapblob_len,
1782 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1783 downgrade_write(&mdsc->snap_rwsem);
1785 down_read(&mdsc->snap_rwsem);
1788 /* insert trace into our cache */
1789 err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1791 if (result == 0 && rinfo->dir_nr)
1792 ceph_readdir_prepopulate(req, req->r_session);
1793 ceph_unreserve_caps(&req->r_caps_reservation);
1796 up_read(&mdsc->snap_rwsem);
1805 add_cap_releases(mdsc, req->r_session, -1);
1806 mutex_unlock(&session->s_mutex);
1808 /* kick calling process */
1809 complete_request(mdsc, req);
1811 ceph_mdsc_put_request(req);
1818 * handle mds notification that our request has been forwarded.
1820 static void handle_forward(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
1822 struct ceph_mds_request *req;
1828 void *p = msg->front.iov_base;
1829 void *end = p + msg->front.iov_len;
1830 int from_mds, state;
1832 if (msg->hdr.src.name.type != CEPH_ENTITY_TYPE_MDS)
1834 from_mds = le64_to_cpu(msg->hdr.src.name.num);
1836 ceph_decode_need(&p, end, sizeof(u64)+2*sizeof(u32), bad);
1837 tid = ceph_decode_64(&p);
1838 next_mds = ceph_decode_32(&p);
1839 fwd_seq = ceph_decode_32(&p);
1840 must_resend = ceph_decode_8(&p);
1842 WARN_ON(must_resend); /* shouldn't happen. */
1844 mutex_lock(&mdsc->mutex);
1845 req = __lookup_request(mdsc, tid);
1847 dout("forward %llu dne\n", tid);
1848 goto out; /* dup reply? */
1851 state = mdsc->sessions[next_mds]->s_state;
1852 if (fwd_seq <= req->r_num_fwd) {
1853 dout("forward %llu to mds%d - old seq %d <= %d\n",
1854 tid, next_mds, req->r_num_fwd, fwd_seq);
1856 /* resend. forward race not possible; mds would drop */
1857 dout("forward %llu to mds%d (we resend)\n", tid, next_mds);
1858 req->r_num_fwd = fwd_seq;
1859 req->r_resend_mds = next_mds;
1860 put_request_session(req);
1861 __do_request(mdsc, req);
1863 ceph_mdsc_put_request(req);
1865 mutex_unlock(&mdsc->mutex);
1869 pr_err("mdsc_handle_forward decode error err=%d\n", err);
1873 * handle a mds session control message
1875 static void handle_session(struct ceph_mds_session *session,
1876 struct ceph_msg *msg)
1878 struct ceph_mds_client *mdsc = session->s_mdsc;
1882 struct ceph_mds_session_head *h = msg->front.iov_base;
1885 if (msg->hdr.src.name.type != CEPH_ENTITY_TYPE_MDS)
1887 mds = le64_to_cpu(msg->hdr.src.name.num);
1890 if (msg->front.iov_len != sizeof(*h))
1892 op = le32_to_cpu(h->op);
1893 seq = le64_to_cpu(h->seq);
1895 mutex_lock(&mdsc->mutex);
1896 /* FIXME: this ttl calculation is generous */
1897 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
1898 mutex_unlock(&mdsc->mutex);
1900 mutex_lock(&session->s_mutex);
1902 dout("handle_session mds%d %s %p state %s seq %llu\n",
1903 mds, ceph_session_op_name(op), session,
1904 session_state_name(session->s_state), seq);
1906 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
1907 session->s_state = CEPH_MDS_SESSION_OPEN;
1908 pr_info("mds%d came back\n", session->s_mds);
1912 case CEPH_SESSION_OPEN:
1913 session->s_state = CEPH_MDS_SESSION_OPEN;
1914 renewed_caps(mdsc, session, 0);
1917 __close_session(mdsc, session);
1920 case CEPH_SESSION_RENEWCAPS:
1921 if (session->s_renew_seq == seq)
1922 renewed_caps(mdsc, session, 1);
1925 case CEPH_SESSION_CLOSE:
1926 unregister_session(mdsc, session);
1927 remove_session_caps(session);
1928 wake = 1; /* for good measure */
1929 complete(&mdsc->session_close_waiters);
1930 kick_requests(mdsc, mds, 0); /* cur only */
1933 case CEPH_SESSION_STALE:
1934 pr_info("mds%d caps went stale, renewing\n",
1936 spin_lock(&session->s_cap_lock);
1937 session->s_cap_gen++;
1938 session->s_cap_ttl = 0;
1939 spin_unlock(&session->s_cap_lock);
1940 send_renew_caps(mdsc, session);
1943 case CEPH_SESSION_RECALL_STATE:
1944 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
1948 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
1952 mutex_unlock(&session->s_mutex);
1954 mutex_lock(&mdsc->mutex);
1955 __wake_requests(mdsc, &session->s_waiting);
1956 mutex_unlock(&mdsc->mutex);
1961 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
1962 (int)msg->front.iov_len);
1969 * called under session->mutex.
1971 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
1972 struct ceph_mds_session *session)
1974 struct ceph_mds_request *req, *nreq;
1977 dout("replay_unsafe_requests mds%d\n", session->s_mds);
1979 mutex_lock(&mdsc->mutex);
1980 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
1981 err = __prepare_send_request(mdsc, req, session->s_mds);
1983 ceph_msg_get(req->r_request);
1984 ceph_con_send(&session->s_con, req->r_request);
1987 mutex_unlock(&mdsc->mutex);
1991 * Encode information about a cap for a reconnect with the MDS.
1993 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
1996 struct ceph_mds_cap_reconnect rec;
1997 struct ceph_inode_info *ci;
1998 struct ceph_pagelist *pagelist = arg;
2002 struct dentry *dentry;
2006 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2007 inode, ceph_vinop(inode), cap, cap->cap_id,
2008 ceph_cap_string(cap->issued));
2009 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2013 dentry = d_find_alias(inode);
2015 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2017 err = PTR_ERR(path);
2024 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2028 spin_lock(&inode->i_lock);
2029 cap->seq = 0; /* reset cap seq */
2030 cap->issue_seq = 0; /* and issue_seq */
2031 rec.cap_id = cpu_to_le64(cap->cap_id);
2032 rec.pathbase = cpu_to_le64(pathbase);
2033 rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2034 rec.issued = cpu_to_le32(cap->issued);
2035 rec.size = cpu_to_le64(inode->i_size);
2036 ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2037 ceph_encode_timespec(&rec.atime, &inode->i_atime);
2038 rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2039 spin_unlock(&inode->i_lock);
2041 err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2051 * If an MDS fails and recovers, clients need to reconnect in order to
2052 * reestablish shared state. This includes all caps issued through
2053 * this session _and_ the snap_realm hierarchy. Because it's not
2054 * clear which snap realms the mds cares about, we send everything we
2055 * know about.. that ensures we'll then get any new info the
2056 * recovering MDS might have.
2058 * This is a relatively heavyweight operation, but it's rare.
2060 * called with mdsc->mutex held.
2062 static void send_mds_reconnect(struct ceph_mds_client *mdsc, int mds)
2064 struct ceph_mds_session *session = NULL;
2065 struct ceph_msg *reply;
2068 u64 next_snap_ino = 0;
2069 struct ceph_pagelist *pagelist;
2071 pr_info("reconnect to recovering mds%d\n", mds);
2073 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2075 goto fail_nopagelist;
2076 ceph_pagelist_init(pagelist);
2078 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, 0, 0, NULL);
2079 if (IS_ERR(reply)) {
2080 err = PTR_ERR(reply);
2085 session = __ceph_lookup_mds_session(mdsc, mds);
2086 mutex_unlock(&mdsc->mutex); /* drop lock for duration */
2089 mutex_lock(&session->s_mutex);
2091 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2094 ceph_con_open(&session->s_con,
2095 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2097 /* replay unsafe requests */
2098 replay_unsafe_requests(mdsc, session);
2100 dout("no session for mds%d, will send short reconnect\n",
2104 down_read(&mdsc->snap_rwsem);
2108 dout("session %p state %s\n", session,
2109 session_state_name(session->s_state));
2111 /* traverse this session's caps */
2112 err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2115 err = iterate_session_caps(session, encode_caps_cb, pagelist);
2120 * snaprealms. we provide mds with the ino, seq (version), and
2121 * parent for all of our realms. If the mds has any newer info,
2126 struct ceph_snap_realm *realm;
2127 struct ceph_mds_snaprealm_reconnect sr_rec;
2128 got = radix_tree_gang_lookup(&mdsc->snap_realms,
2129 (void **)&realm, next_snap_ino, 1);
2133 dout(" adding snap realm %llx seq %lld parent %llx\n",
2134 realm->ino, realm->seq, realm->parent_ino);
2135 sr_rec.ino = cpu_to_le64(realm->ino);
2136 sr_rec.seq = cpu_to_le64(realm->seq);
2137 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2138 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2141 next_snap_ino = realm->ino + 1;
2145 reply->pagelist = pagelist;
2146 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2147 reply->nr_pages = calc_pages_for(0, pagelist->length);
2148 ceph_con_send(&session->s_con, reply);
2151 session->s_state = CEPH_MDS_SESSION_OPEN;
2152 __wake_requests(mdsc, &session->s_waiting);
2156 up_read(&mdsc->snap_rwsem);
2158 mutex_unlock(&session->s_mutex);
2159 ceph_put_mds_session(session);
2161 mutex_lock(&mdsc->mutex);
2165 ceph_msg_put(reply);
2167 ceph_pagelist_release(pagelist);
2170 pr_err("ENOMEM preparing reconnect for mds%d\n", mds);
2176 * compare old and new mdsmaps, kicking requests
2177 * and closing out old connections as necessary
2179 * called under mdsc->mutex.
2181 static void check_new_map(struct ceph_mds_client *mdsc,
2182 struct ceph_mdsmap *newmap,
2183 struct ceph_mdsmap *oldmap)
2186 int oldstate, newstate;
2187 struct ceph_mds_session *s;
2189 dout("check_new_map new %u old %u\n",
2190 newmap->m_epoch, oldmap->m_epoch);
2192 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2193 if (mdsc->sessions[i] == NULL)
2195 s = mdsc->sessions[i];
2196 oldstate = ceph_mdsmap_get_state(oldmap, i);
2197 newstate = ceph_mdsmap_get_state(newmap, i);
2199 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2200 i, ceph_mds_state_name(oldstate),
2201 ceph_mds_state_name(newstate),
2202 session_state_name(s->s_state));
2204 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2205 ceph_mdsmap_get_addr(newmap, i),
2206 sizeof(struct ceph_entity_addr))) {
2207 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2208 /* the session never opened, just close it
2210 __wake_requests(mdsc, &s->s_waiting);
2211 unregister_session(mdsc, s);
2214 mutex_unlock(&mdsc->mutex);
2215 mutex_lock(&s->s_mutex);
2216 mutex_lock(&mdsc->mutex);
2217 ceph_con_close(&s->s_con);
2218 mutex_unlock(&s->s_mutex);
2219 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2222 /* kick any requests waiting on the recovering mds */
2223 kick_requests(mdsc, i, 1);
2224 } else if (oldstate == newstate) {
2225 continue; /* nothing new with this mds */
2231 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2232 newstate >= CEPH_MDS_STATE_RECONNECT)
2233 send_mds_reconnect(mdsc, i);
2236 * kick requests on any mds that has gone active.
2238 * kick requests on cur or forwarder: we may have sent
2239 * the request to mds1, mds1 told us it forwarded it
2240 * to mds2, but then we learn mds1 failed and can't be
2241 * sure it successfully forwarded our request before
2244 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2245 newstate >= CEPH_MDS_STATE_ACTIVE) {
2246 pr_info("mds%d reconnect completed\n", s->s_mds);
2247 kick_requests(mdsc, i, 1);
2248 ceph_kick_flushing_caps(mdsc, s);
2249 wake_up_session_caps(s, 1);
2261 * caller must hold session s_mutex, dentry->d_lock
2263 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2265 struct ceph_dentry_info *di = ceph_dentry(dentry);
2267 ceph_put_mds_session(di->lease_session);
2268 di->lease_session = NULL;
2271 static void handle_lease(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2273 struct super_block *sb = mdsc->client->sb;
2274 struct inode *inode;
2275 struct ceph_mds_session *session;
2276 struct ceph_inode_info *ci;
2277 struct dentry *parent, *dentry;
2278 struct ceph_dentry_info *di;
2280 struct ceph_mds_lease *h = msg->front.iov_base;
2281 struct ceph_vino vino;
2286 if (msg->hdr.src.name.type != CEPH_ENTITY_TYPE_MDS)
2288 mds = le64_to_cpu(msg->hdr.src.name.num);
2289 dout("handle_lease from mds%d\n", mds);
2292 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2294 vino.ino = le64_to_cpu(h->ino);
2295 vino.snap = CEPH_NOSNAP;
2296 mask = le16_to_cpu(h->mask);
2297 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2298 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2299 if (dname.len != get_unaligned_le32(h+1))
2303 mutex_lock(&mdsc->mutex);
2304 session = __ceph_lookup_mds_session(mdsc, mds);
2305 mutex_unlock(&mdsc->mutex);
2307 pr_err("handle_lease got lease but no session mds%d\n", mds);
2311 mutex_lock(&session->s_mutex);
2315 inode = ceph_find_inode(sb, vino);
2316 dout("handle_lease '%s', mask %d, ino %llx %p\n",
2317 ceph_lease_op_name(h->action), mask, vino.ino, inode);
2318 if (inode == NULL) {
2319 dout("handle_lease no inode %llx\n", vino.ino);
2322 ci = ceph_inode(inode);
2325 parent = d_find_alias(inode);
2327 dout("no parent dentry on inode %p\n", inode);
2329 goto release; /* hrm... */
2331 dname.hash = full_name_hash(dname.name, dname.len);
2332 dentry = d_lookup(parent, &dname);
2337 spin_lock(&dentry->d_lock);
2338 di = ceph_dentry(dentry);
2339 switch (h->action) {
2340 case CEPH_MDS_LEASE_REVOKE:
2341 if (di && di->lease_session == session) {
2342 h->seq = cpu_to_le32(di->lease_seq);
2343 __ceph_mdsc_drop_dentry_lease(dentry);
2348 case CEPH_MDS_LEASE_RENEW:
2349 if (di && di->lease_session == session &&
2350 di->lease_gen == session->s_cap_gen &&
2351 di->lease_renew_from &&
2352 di->lease_renew_after == 0) {
2353 unsigned long duration =
2354 le32_to_cpu(h->duration_ms) * HZ / 1000;
2356 di->lease_seq = le32_to_cpu(h->seq);
2357 dentry->d_time = di->lease_renew_from + duration;
2358 di->lease_renew_after = di->lease_renew_from +
2360 di->lease_renew_from = 0;
2364 spin_unlock(&dentry->d_lock);
2371 /* let's just reuse the same message */
2372 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2374 ceph_con_send(&session->s_con, msg);
2378 mutex_unlock(&session->s_mutex);
2379 ceph_put_mds_session(session);
2383 pr_err("corrupt lease message\n");
2387 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2388 struct inode *inode,
2389 struct dentry *dentry, char action,
2392 struct ceph_msg *msg;
2393 struct ceph_mds_lease *lease;
2394 int len = sizeof(*lease) + sizeof(u32);
2397 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2398 inode, dentry, ceph_lease_op_name(action), session->s_mds);
2399 dnamelen = dentry->d_name.len;
2402 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, 0, 0, NULL);
2405 lease = msg->front.iov_base;
2406 lease->action = action;
2407 lease->mask = cpu_to_le16(CEPH_LOCK_DN);
2408 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2409 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2410 lease->seq = cpu_to_le32(seq);
2411 put_unaligned_le32(dnamelen, lease + 1);
2412 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2415 * if this is a preemptive lease RELEASE, no need to
2416 * flush request stream, since the actual request will
2419 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2421 ceph_con_send(&session->s_con, msg);
2425 * Preemptively release a lease we expect to invalidate anyway.
2426 * Pass @inode always, @dentry is optional.
2428 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2429 struct dentry *dentry, int mask)
2431 struct ceph_dentry_info *di;
2432 struct ceph_mds_session *session;
2435 BUG_ON(inode == NULL);
2436 BUG_ON(dentry == NULL);
2437 BUG_ON(mask != CEPH_LOCK_DN);
2439 /* is dentry lease valid? */
2440 spin_lock(&dentry->d_lock);
2441 di = ceph_dentry(dentry);
2442 if (!di || !di->lease_session ||
2443 di->lease_session->s_mds < 0 ||
2444 di->lease_gen != di->lease_session->s_cap_gen ||
2445 !time_before(jiffies, dentry->d_time)) {
2446 dout("lease_release inode %p dentry %p -- "
2448 inode, dentry, mask);
2449 spin_unlock(&dentry->d_lock);
2453 /* we do have a lease on this dentry; note mds and seq */
2454 session = ceph_get_mds_session(di->lease_session);
2455 seq = di->lease_seq;
2456 __ceph_mdsc_drop_dentry_lease(dentry);
2457 spin_unlock(&dentry->d_lock);
2459 dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2460 inode, dentry, mask, session->s_mds);
2461 ceph_mdsc_lease_send_msg(session, inode, dentry,
2462 CEPH_MDS_LEASE_RELEASE, seq);
2463 ceph_put_mds_session(session);
2467 * drop all leases (and dentry refs) in preparation for umount
2469 static void drop_leases(struct ceph_mds_client *mdsc)
2473 dout("drop_leases\n");
2474 mutex_lock(&mdsc->mutex);
2475 for (i = 0; i < mdsc->max_sessions; i++) {
2476 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2479 mutex_unlock(&mdsc->mutex);
2480 mutex_lock(&s->s_mutex);
2481 mutex_unlock(&s->s_mutex);
2482 ceph_put_mds_session(s);
2483 mutex_lock(&mdsc->mutex);
2485 mutex_unlock(&mdsc->mutex);
2491 * delayed work -- periodically trim expired leases, renew caps with mds
2493 static void schedule_delayed(struct ceph_mds_client *mdsc)
2496 unsigned hz = round_jiffies_relative(HZ * delay);
2497 schedule_delayed_work(&mdsc->delayed_work, hz);
2500 static void delayed_work(struct work_struct *work)
2503 struct ceph_mds_client *mdsc =
2504 container_of(work, struct ceph_mds_client, delayed_work.work);
2508 dout("mdsc delayed_work\n");
2509 ceph_check_delayed_caps(mdsc);
2511 mutex_lock(&mdsc->mutex);
2512 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2513 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2514 mdsc->last_renew_caps);
2516 mdsc->last_renew_caps = jiffies;
2518 for (i = 0; i < mdsc->max_sessions; i++) {
2519 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2522 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2523 dout("resending session close request for mds%d\n",
2525 request_close_session(mdsc, s);
2526 ceph_put_mds_session(s);
2529 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2530 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2531 s->s_state = CEPH_MDS_SESSION_HUNG;
2532 pr_info("mds%d hung\n", s->s_mds);
2535 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2536 /* this mds is failed or recovering, just wait */
2537 ceph_put_mds_session(s);
2540 mutex_unlock(&mdsc->mutex);
2542 mutex_lock(&s->s_mutex);
2544 send_renew_caps(mdsc, s);
2546 ceph_con_keepalive(&s->s_con);
2547 add_cap_releases(mdsc, s, -1);
2548 send_cap_releases(mdsc, s);
2549 mutex_unlock(&s->s_mutex);
2550 ceph_put_mds_session(s);
2552 mutex_lock(&mdsc->mutex);
2554 mutex_unlock(&mdsc->mutex);
2556 schedule_delayed(mdsc);
2560 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2562 mdsc->client = client;
2563 mutex_init(&mdsc->mutex);
2564 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2565 init_completion(&mdsc->safe_umount_waiters);
2566 init_completion(&mdsc->session_close_waiters);
2567 INIT_LIST_HEAD(&mdsc->waiting_for_map);
2568 mdsc->sessions = NULL;
2569 mdsc->max_sessions = 0;
2571 init_rwsem(&mdsc->snap_rwsem);
2572 INIT_RADIX_TREE(&mdsc->snap_realms, GFP_NOFS);
2573 INIT_LIST_HEAD(&mdsc->snap_empty);
2574 spin_lock_init(&mdsc->snap_empty_lock);
2576 INIT_RADIX_TREE(&mdsc->request_tree, GFP_NOFS);
2577 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2578 mdsc->last_renew_caps = jiffies;
2579 INIT_LIST_HEAD(&mdsc->cap_delay_list);
2580 spin_lock_init(&mdsc->cap_delay_lock);
2581 INIT_LIST_HEAD(&mdsc->snap_flush_list);
2582 spin_lock_init(&mdsc->snap_flush_lock);
2583 mdsc->cap_flush_seq = 0;
2584 INIT_LIST_HEAD(&mdsc->cap_dirty);
2585 mdsc->num_cap_flushing = 0;
2586 spin_lock_init(&mdsc->cap_dirty_lock);
2587 init_waitqueue_head(&mdsc->cap_flushing_wq);
2588 spin_lock_init(&mdsc->dentry_lru_lock);
2589 INIT_LIST_HEAD(&mdsc->dentry_lru);
2594 * Wait for safe replies on open mds requests. If we time out, drop
2595 * all requests from the tree to avoid dangling dentry refs.
2597 static void wait_requests(struct ceph_mds_client *mdsc)
2599 struct ceph_mds_request *req;
2600 struct ceph_client *client = mdsc->client;
2602 mutex_lock(&mdsc->mutex);
2603 if (__get_oldest_tid(mdsc)) {
2604 mutex_unlock(&mdsc->mutex);
2605 dout("wait_requests waiting for requests\n");
2606 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2607 client->mount_args->mount_timeout * HZ);
2608 mutex_lock(&mdsc->mutex);
2610 /* tear down remaining requests */
2611 while (radix_tree_gang_lookup(&mdsc->request_tree,
2612 (void **)&req, 0, 1)) {
2613 dout("wait_requests timed out on tid %llu\n",
2615 radix_tree_delete(&mdsc->request_tree, req->r_tid);
2616 ceph_mdsc_put_request(req);
2619 mutex_unlock(&mdsc->mutex);
2620 dout("wait_requests done\n");
2624 * called before mount is ro, and before dentries are torn down.
2625 * (hmm, does this still race with new lookups?)
2627 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2629 dout("pre_umount\n");
2633 ceph_flush_dirty_caps(mdsc);
2634 wait_requests(mdsc);
2638 * wait for all write mds requests to flush.
2640 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2642 struct ceph_mds_request *req;
2646 mutex_lock(&mdsc->mutex);
2647 dout("wait_unsafe_requests want %lld\n", want_tid);
2649 got = radix_tree_gang_lookup(&mdsc->request_tree, (void **)&req,
2653 if (req->r_tid > want_tid)
2656 next_tid = req->r_tid + 1;
2657 if ((req->r_op & CEPH_MDS_OP_WRITE) == 0)
2658 continue; /* not a write op */
2660 ceph_mdsc_get_request(req);
2661 mutex_unlock(&mdsc->mutex);
2662 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
2663 req->r_tid, want_tid);
2664 wait_for_completion(&req->r_safe_completion);
2665 mutex_lock(&mdsc->mutex);
2666 ceph_mdsc_put_request(req);
2668 mutex_unlock(&mdsc->mutex);
2669 dout("wait_unsafe_requests done\n");
2672 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2674 u64 want_tid, want_flush;
2677 mutex_lock(&mdsc->mutex);
2678 want_tid = mdsc->last_tid;
2679 want_flush = mdsc->cap_flush_seq;
2680 mutex_unlock(&mdsc->mutex);
2681 dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2683 ceph_flush_dirty_caps(mdsc);
2685 wait_unsafe_requests(mdsc, want_tid);
2686 wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2691 * called after sb is ro.
2693 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2695 struct ceph_mds_session *session;
2698 struct ceph_client *client = mdsc->client;
2699 unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2701 dout("close_sessions\n");
2703 mutex_lock(&mdsc->mutex);
2705 /* close sessions */
2707 while (time_before(jiffies, started + timeout)) {
2708 dout("closing sessions\n");
2710 for (i = 0; i < mdsc->max_sessions; i++) {
2711 session = __ceph_lookup_mds_session(mdsc, i);
2714 mutex_unlock(&mdsc->mutex);
2715 mutex_lock(&session->s_mutex);
2716 __close_session(mdsc, session);
2717 mutex_unlock(&session->s_mutex);
2718 ceph_put_mds_session(session);
2719 mutex_lock(&mdsc->mutex);
2725 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2728 dout("waiting for sessions to close\n");
2729 mutex_unlock(&mdsc->mutex);
2730 wait_for_completion_timeout(&mdsc->session_close_waiters,
2732 mutex_lock(&mdsc->mutex);
2735 /* tear down remaining sessions */
2736 for (i = 0; i < mdsc->max_sessions; i++) {
2737 if (mdsc->sessions[i]) {
2738 session = get_session(mdsc->sessions[i]);
2739 unregister_session(mdsc, session);
2740 mutex_unlock(&mdsc->mutex);
2741 mutex_lock(&session->s_mutex);
2742 remove_session_caps(session);
2743 mutex_unlock(&session->s_mutex);
2744 ceph_put_mds_session(session);
2745 mutex_lock(&mdsc->mutex);
2749 WARN_ON(!list_empty(&mdsc->cap_delay_list));
2751 mutex_unlock(&mdsc->mutex);
2753 ceph_cleanup_empty_realms(mdsc);
2755 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2760 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2763 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2765 ceph_mdsmap_destroy(mdsc->mdsmap);
2766 kfree(mdsc->sessions);
2771 * handle mds map update.
2773 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2777 void *p = msg->front.iov_base;
2778 void *end = p + msg->front.iov_len;
2779 struct ceph_mdsmap *newmap, *oldmap;
2780 struct ceph_fsid fsid;
2783 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2784 ceph_decode_copy(&p, &fsid, sizeof(fsid));
2785 if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2787 epoch = ceph_decode_32(&p);
2788 maplen = ceph_decode_32(&p);
2789 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2791 /* do we need it? */
2792 ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2793 mutex_lock(&mdsc->mutex);
2794 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2795 dout("handle_map epoch %u <= our %u\n",
2796 epoch, mdsc->mdsmap->m_epoch);
2797 mutex_unlock(&mdsc->mutex);
2801 newmap = ceph_mdsmap_decode(&p, end);
2802 if (IS_ERR(newmap)) {
2803 err = PTR_ERR(newmap);
2807 /* swap into place */
2809 oldmap = mdsc->mdsmap;
2810 mdsc->mdsmap = newmap;
2811 check_new_map(mdsc, newmap, oldmap);
2812 ceph_mdsmap_destroy(oldmap);
2814 mdsc->mdsmap = newmap; /* first mds map */
2816 mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2818 __wake_requests(mdsc, &mdsc->waiting_for_map);
2820 mutex_unlock(&mdsc->mutex);
2821 schedule_delayed(mdsc);
2825 mutex_unlock(&mdsc->mutex);
2827 pr_err("error decoding mdsmap %d\n", err);
2831 static struct ceph_connection *con_get(struct ceph_connection *con)
2833 struct ceph_mds_session *s = con->private;
2835 if (get_session(s)) {
2836 dout("mdsc con_get %p %d -> %d\n", s,
2837 atomic_read(&s->s_ref) - 1, atomic_read(&s->s_ref));
2840 dout("mdsc con_get %p FAIL\n", s);
2844 static void con_put(struct ceph_connection *con)
2846 struct ceph_mds_session *s = con->private;
2848 dout("mdsc con_put %p %d -> %d\n", s, atomic_read(&s->s_ref),
2849 atomic_read(&s->s_ref) - 1);
2850 ceph_put_mds_session(s);
2854 * if the client is unresponsive for long enough, the mds will kill
2855 * the session entirely.
2857 static void peer_reset(struct ceph_connection *con)
2859 struct ceph_mds_session *s = con->private;
2861 pr_err("mds%d gave us the boot. IMPLEMENT RECONNECT.\n",
2865 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2867 struct ceph_mds_session *s = con->private;
2868 struct ceph_mds_client *mdsc = s->s_mdsc;
2869 int type = le16_to_cpu(msg->hdr.type);
2872 case CEPH_MSG_MDS_MAP:
2873 ceph_mdsc_handle_map(mdsc, msg);
2875 case CEPH_MSG_CLIENT_SESSION:
2876 handle_session(s, msg);
2878 case CEPH_MSG_CLIENT_REPLY:
2879 handle_reply(s, msg);
2881 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
2882 handle_forward(mdsc, msg);
2884 case CEPH_MSG_CLIENT_CAPS:
2885 ceph_handle_caps(s, msg);
2887 case CEPH_MSG_CLIENT_SNAP:
2888 ceph_handle_snap(mdsc, msg);
2890 case CEPH_MSG_CLIENT_LEASE:
2891 handle_lease(mdsc, msg);
2895 pr_err("received unknown message type %d %s\n", type,
2896 ceph_msg_type_name(type));
2904 static int get_authorizer(struct ceph_connection *con,
2905 void **buf, int *len, int *proto,
2906 void **reply_buf, int *reply_len, int force_new)
2908 struct ceph_mds_session *s = con->private;
2909 struct ceph_mds_client *mdsc = s->s_mdsc;
2910 struct ceph_auth_client *ac = mdsc->client->monc.auth;
2913 if (force_new && s->s_authorizer) {
2914 ac->ops->destroy_authorizer(ac, s->s_authorizer);
2915 s->s_authorizer = NULL;
2917 if (s->s_authorizer == NULL) {
2918 if (ac->ops->create_authorizer) {
2919 ret = ac->ops->create_authorizer(
2920 ac, CEPH_ENTITY_TYPE_MDS,
2922 &s->s_authorizer_buf,
2923 &s->s_authorizer_buf_len,
2924 &s->s_authorizer_reply_buf,
2925 &s->s_authorizer_reply_buf_len);
2931 *proto = ac->protocol;
2932 *buf = s->s_authorizer_buf;
2933 *len = s->s_authorizer_buf_len;
2934 *reply_buf = s->s_authorizer_reply_buf;
2935 *reply_len = s->s_authorizer_reply_buf_len;
2940 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2942 struct ceph_mds_session *s = con->private;
2943 struct ceph_mds_client *mdsc = s->s_mdsc;
2944 struct ceph_auth_client *ac = mdsc->client->monc.auth;
2946 return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
2949 const static struct ceph_connection_operations mds_con_ops = {
2952 .dispatch = dispatch,
2953 .get_authorizer = get_authorizer,
2954 .verify_authorizer_reply = verify_authorizer_reply,
2955 .peer_reset = peer_reset,
2956 .alloc_msg = ceph_alloc_msg,
2957 .alloc_middle = ceph_alloc_middle,