1 #include "ceph_debug.h"
3 #include <linux/wait.h>
4 #include <linux/slab.h>
5 #include <linux/sched.h>
7 #include "mds_client.h"
8 #include "mon_client.h"
10 #include "messenger.h"
16 * A cluster of MDS (metadata server) daemons is responsible for
17 * managing the file system namespace (the directory hierarchy and
18 * inodes) and for coordinating shared access to storage. Metadata is
19 * partitioning hierarchically across a number of servers, and that
20 * partition varies over time as the cluster adjusts the distribution
21 * in order to balance load.
23 * The MDS client is primarily responsible to managing synchronous
24 * metadata requests for operations like open, unlink, and so forth.
25 * If there is a MDS failure, we find out about it when we (possibly
26 * request and) receive a new MDS map, and can resubmit affected
29 * For the most part, though, we take advantage of a lossless
30 * communications channel to the MDS, and do not need to worry about
31 * timing out or resubmitting requests.
33 * We maintain a stateful "session" with each MDS we interact with.
34 * Within each session, we sent periodic heartbeat messages to ensure
35 * any capabilities or leases we have been issues remain valid. If
36 * the session times out and goes stale, our leases and capabilities
37 * are no longer valid.
40 static void __wake_requests(struct ceph_mds_client *mdsc,
41 struct list_head *head);
43 static const struct ceph_connection_operations mds_con_ops;
51 * parse individual inode info
53 static int parse_reply_info_in(void **p, void *end,
54 struct ceph_mds_reply_info_in *info)
59 *p += sizeof(struct ceph_mds_reply_inode) +
60 sizeof(*info->in->fragtree.splits) *
61 le32_to_cpu(info->in->fragtree.nsplits);
63 ceph_decode_32_safe(p, end, info->symlink_len, bad);
64 ceph_decode_need(p, end, info->symlink_len, bad);
66 *p += info->symlink_len;
68 ceph_decode_32_safe(p, end, info->xattr_len, bad);
69 ceph_decode_need(p, end, info->xattr_len, bad);
70 info->xattr_data = *p;
71 *p += info->xattr_len;
78 * parse a normal reply, which may contain a (dir+)dentry and/or a
81 static int parse_reply_info_trace(void **p, void *end,
82 struct ceph_mds_reply_info_parsed *info)
86 if (info->head->is_dentry) {
87 err = parse_reply_info_in(p, end, &info->diri);
91 if (unlikely(*p + sizeof(*info->dirfrag) > end))
94 *p += sizeof(*info->dirfrag) +
95 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
96 if (unlikely(*p > end))
99 ceph_decode_32_safe(p, end, info->dname_len, bad);
100 ceph_decode_need(p, end, info->dname_len, bad);
102 *p += info->dname_len;
104 *p += sizeof(*info->dlease);
107 if (info->head->is_target) {
108 err = parse_reply_info_in(p, end, &info->targeti);
113 if (unlikely(*p != end))
120 pr_err("problem parsing mds trace %d\n", err);
125 * parse readdir results
127 static int parse_reply_info_dir(void **p, void *end,
128 struct ceph_mds_reply_info_parsed *info)
134 if (*p + sizeof(*info->dir_dir) > end)
136 *p += sizeof(*info->dir_dir) +
137 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
141 ceph_decode_need(p, end, sizeof(num) + 2, bad);
142 num = ceph_decode_32(p);
143 info->dir_end = ceph_decode_8(p);
144 info->dir_complete = ceph_decode_8(p);
148 /* alloc large array */
150 info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
151 sizeof(*info->dir_dname) +
152 sizeof(*info->dir_dname_len) +
153 sizeof(*info->dir_dlease),
155 if (info->dir_in == NULL) {
159 info->dir_dname = (void *)(info->dir_in + num);
160 info->dir_dname_len = (void *)(info->dir_dname + num);
161 info->dir_dlease = (void *)(info->dir_dname_len + num);
165 ceph_decode_need(p, end, sizeof(u32)*2, bad);
166 info->dir_dname_len[i] = ceph_decode_32(p);
167 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
168 info->dir_dname[i] = *p;
169 *p += info->dir_dname_len[i];
170 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
172 info->dir_dlease[i] = *p;
173 *p += sizeof(struct ceph_mds_reply_lease);
176 err = parse_reply_info_in(p, end, &info->dir_in[i]);
191 pr_err("problem parsing dir contents %d\n", err);
196 * parse entire mds reply
198 static int parse_reply_info(struct ceph_msg *msg,
199 struct ceph_mds_reply_info_parsed *info)
205 info->head = msg->front.iov_base;
206 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
207 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
210 ceph_decode_32_safe(&p, end, len, bad);
212 err = parse_reply_info_trace(&p, p+len, info);
218 ceph_decode_32_safe(&p, end, len, bad);
220 err = parse_reply_info_dir(&p, p+len, info);
226 ceph_decode_32_safe(&p, end, len, bad);
227 info->snapblob_len = len;
238 pr_err("mds parse_reply err %d\n", err);
242 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
251 static const char *session_state_name(int s)
254 case CEPH_MDS_SESSION_NEW: return "new";
255 case CEPH_MDS_SESSION_OPENING: return "opening";
256 case CEPH_MDS_SESSION_OPEN: return "open";
257 case CEPH_MDS_SESSION_HUNG: return "hung";
258 case CEPH_MDS_SESSION_CLOSING: return "closing";
259 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
260 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
261 default: return "???";
265 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
267 if (atomic_inc_not_zero(&s->s_ref)) {
268 dout("mdsc get_session %p %d -> %d\n", s,
269 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
272 dout("mdsc get_session %p 0 -- FAIL", s);
277 void ceph_put_mds_session(struct ceph_mds_session *s)
279 dout("mdsc put_session %p %d -> %d\n", s,
280 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
281 if (atomic_dec_and_test(&s->s_ref)) {
283 s->s_mdsc->client->monc.auth->ops->destroy_authorizer(
284 s->s_mdsc->client->monc.auth, s->s_authorizer);
290 * called under mdsc->mutex
292 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
295 struct ceph_mds_session *session;
297 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
299 session = mdsc->sessions[mds];
300 dout("lookup_mds_session %p %d\n", session,
301 atomic_read(&session->s_ref));
302 get_session(session);
306 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
308 if (mds >= mdsc->max_sessions)
310 return mdsc->sessions[mds];
313 static int __verify_registered_session(struct ceph_mds_client *mdsc,
314 struct ceph_mds_session *s)
316 if (s->s_mds >= mdsc->max_sessions ||
317 mdsc->sessions[s->s_mds] != s)
323 * create+register a new session for given mds.
324 * called under mdsc->mutex.
326 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
329 struct ceph_mds_session *s;
331 s = kzalloc(sizeof(*s), GFP_NOFS);
333 return ERR_PTR(-ENOMEM);
336 s->s_state = CEPH_MDS_SESSION_NEW;
339 mutex_init(&s->s_mutex);
341 ceph_con_init(mdsc->client->msgr, &s->s_con);
342 s->s_con.private = s;
343 s->s_con.ops = &mds_con_ops;
344 s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
345 s->s_con.peer_name.num = cpu_to_le64(mds);
347 spin_lock_init(&s->s_cap_lock);
350 s->s_renew_requested = 0;
352 INIT_LIST_HEAD(&s->s_caps);
355 atomic_set(&s->s_ref, 1);
356 INIT_LIST_HEAD(&s->s_waiting);
357 INIT_LIST_HEAD(&s->s_unsafe);
358 s->s_num_cap_releases = 0;
359 s->s_cap_iterator = NULL;
360 INIT_LIST_HEAD(&s->s_cap_releases);
361 INIT_LIST_HEAD(&s->s_cap_releases_done);
362 INIT_LIST_HEAD(&s->s_cap_flushing);
363 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
365 dout("register_session mds%d\n", mds);
366 if (mds >= mdsc->max_sessions) {
367 int newmax = 1 << get_count_order(mds+1);
368 struct ceph_mds_session **sa;
370 dout("register_session realloc to %d\n", newmax);
371 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
374 if (mdsc->sessions) {
375 memcpy(sa, mdsc->sessions,
376 mdsc->max_sessions * sizeof(void *));
377 kfree(mdsc->sessions);
380 mdsc->max_sessions = newmax;
382 mdsc->sessions[mds] = s;
383 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
385 ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
391 return ERR_PTR(-ENOMEM);
395 * called under mdsc->mutex
397 static void __unregister_session(struct ceph_mds_client *mdsc,
398 struct ceph_mds_session *s)
400 dout("__unregister_session mds%d %p\n", s->s_mds, s);
401 BUG_ON(mdsc->sessions[s->s_mds] != s);
402 mdsc->sessions[s->s_mds] = NULL;
403 ceph_con_close(&s->s_con);
404 ceph_put_mds_session(s);
408 * drop session refs in request.
410 * should be last request ref, or hold mdsc->mutex
412 static void put_request_session(struct ceph_mds_request *req)
414 if (req->r_session) {
415 ceph_put_mds_session(req->r_session);
416 req->r_session = NULL;
420 void ceph_mdsc_release_request(struct kref *kref)
422 struct ceph_mds_request *req = container_of(kref,
423 struct ceph_mds_request,
426 ceph_msg_put(req->r_request);
428 ceph_msg_put(req->r_reply);
429 destroy_reply_info(&req->r_reply_info);
432 ceph_put_cap_refs(ceph_inode(req->r_inode),
436 if (req->r_locked_dir)
437 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
439 if (req->r_target_inode)
440 iput(req->r_target_inode);
443 if (req->r_old_dentry) {
445 ceph_inode(req->r_old_dentry->d_parent->d_inode),
447 dput(req->r_old_dentry);
451 put_request_session(req);
452 ceph_unreserve_caps(&req->r_caps_reservation);
457 * lookup session, bump ref if found.
459 * called under mdsc->mutex.
461 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
464 struct ceph_mds_request *req;
465 struct rb_node *n = mdsc->request_tree.rb_node;
468 req = rb_entry(n, struct ceph_mds_request, r_node);
469 if (tid < req->r_tid)
471 else if (tid > req->r_tid)
474 ceph_mdsc_get_request(req);
481 static void __insert_request(struct ceph_mds_client *mdsc,
482 struct ceph_mds_request *new)
484 struct rb_node **p = &mdsc->request_tree.rb_node;
485 struct rb_node *parent = NULL;
486 struct ceph_mds_request *req = NULL;
490 req = rb_entry(parent, struct ceph_mds_request, r_node);
491 if (new->r_tid < req->r_tid)
493 else if (new->r_tid > req->r_tid)
499 rb_link_node(&new->r_node, parent, p);
500 rb_insert_color(&new->r_node, &mdsc->request_tree);
504 * Register an in-flight request, and assign a tid. Link to directory
505 * are modifying (if any).
507 * Called under mdsc->mutex.
509 static void __register_request(struct ceph_mds_client *mdsc,
510 struct ceph_mds_request *req,
513 req->r_tid = ++mdsc->last_tid;
515 ceph_reserve_caps(&req->r_caps_reservation, req->r_num_caps);
516 dout("__register_request %p tid %lld\n", req, req->r_tid);
517 ceph_mdsc_get_request(req);
518 __insert_request(mdsc, req);
521 struct ceph_inode_info *ci = ceph_inode(dir);
523 spin_lock(&ci->i_unsafe_lock);
524 req->r_unsafe_dir = dir;
525 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
526 spin_unlock(&ci->i_unsafe_lock);
530 static void __unregister_request(struct ceph_mds_client *mdsc,
531 struct ceph_mds_request *req)
533 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
534 rb_erase(&req->r_node, &mdsc->request_tree);
535 RB_CLEAR_NODE(&req->r_node);
537 if (req->r_unsafe_dir) {
538 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
540 spin_lock(&ci->i_unsafe_lock);
541 list_del_init(&req->r_unsafe_dir_item);
542 spin_unlock(&ci->i_unsafe_lock);
545 ceph_mdsc_put_request(req);
549 * Choose mds to send request to next. If there is a hint set in the
550 * request (e.g., due to a prior forward hint from the mds), use that.
551 * Otherwise, consult frag tree and/or caps to identify the
552 * appropriate mds. If all else fails, choose randomly.
554 * Called under mdsc->mutex.
556 static int __choose_mds(struct ceph_mds_client *mdsc,
557 struct ceph_mds_request *req)
560 struct ceph_inode_info *ci;
561 struct ceph_cap *cap;
562 int mode = req->r_direct_mode;
564 u32 hash = req->r_direct_hash;
565 bool is_hash = req->r_direct_is_hash;
568 * is there a specific mds we should try? ignore hint if we have
569 * no session and the mds is not up (active or recovering).
571 if (req->r_resend_mds >= 0 &&
572 (__have_session(mdsc, req->r_resend_mds) ||
573 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
574 dout("choose_mds using resend_mds mds%d\n",
576 return req->r_resend_mds;
579 if (mode == USE_RANDOM_MDS)
584 inode = req->r_inode;
585 } else if (req->r_dentry) {
586 if (req->r_dentry->d_inode) {
587 inode = req->r_dentry->d_inode;
589 inode = req->r_dentry->d_parent->d_inode;
590 hash = req->r_dentry->d_name.hash;
594 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
598 ci = ceph_inode(inode);
600 if (is_hash && S_ISDIR(inode->i_mode)) {
601 struct ceph_inode_frag frag;
604 ceph_choose_frag(ci, hash, &frag, &found);
606 if (mode == USE_ANY_MDS && frag.ndist > 0) {
609 /* choose a random replica */
610 get_random_bytes(&r, 1);
613 dout("choose_mds %p %llx.%llx "
614 "frag %u mds%d (%d/%d)\n",
615 inode, ceph_vinop(inode),
621 /* since this file/dir wasn't known to be
622 * replicated, then we want to look for the
623 * authoritative mds. */
626 /* choose auth mds */
628 dout("choose_mds %p %llx.%llx "
629 "frag %u mds%d (auth)\n",
630 inode, ceph_vinop(inode), frag.frag, mds);
636 spin_lock(&inode->i_lock);
638 if (mode == USE_AUTH_MDS)
639 cap = ci->i_auth_cap;
640 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
641 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
643 spin_unlock(&inode->i_lock);
646 mds = cap->session->s_mds;
647 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
648 inode, ceph_vinop(inode), mds,
649 cap == ci->i_auth_cap ? "auth " : "", cap);
650 spin_unlock(&inode->i_lock);
654 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
655 dout("choose_mds chose random mds%d\n", mds);
663 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
665 struct ceph_msg *msg;
666 struct ceph_mds_session_head *h;
668 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS);
670 pr_err("create_session_msg ENOMEM creating msg\n");
673 h = msg->front.iov_base;
674 h->op = cpu_to_le32(op);
675 h->seq = cpu_to_le64(seq);
680 * send session open request.
682 * called under mdsc->mutex
684 static int __open_session(struct ceph_mds_client *mdsc,
685 struct ceph_mds_session *session)
687 struct ceph_msg *msg;
689 int mds = session->s_mds;
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);
702 ceph_con_send(&session->s_con, msg);
711 * Free preallocated cap messages assigned to this session
713 static void cleanup_cap_releases(struct ceph_mds_session *session)
715 struct ceph_msg *msg;
717 spin_lock(&session->s_cap_lock);
718 while (!list_empty(&session->s_cap_releases)) {
719 msg = list_first_entry(&session->s_cap_releases,
720 struct ceph_msg, list_head);
721 list_del_init(&msg->list_head);
724 while (!list_empty(&session->s_cap_releases_done)) {
725 msg = list_first_entry(&session->s_cap_releases_done,
726 struct ceph_msg, list_head);
727 list_del_init(&msg->list_head);
730 spin_unlock(&session->s_cap_lock);
734 * Helper to safely iterate over all caps associated with a session, with
735 * special care taken to handle a racing __ceph_remove_cap().
737 * Caller must hold session s_mutex.
739 static int iterate_session_caps(struct ceph_mds_session *session,
740 int (*cb)(struct inode *, struct ceph_cap *,
744 struct ceph_cap *cap;
745 struct inode *inode, *last_inode = NULL;
746 struct ceph_cap *old_cap = NULL;
749 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
750 spin_lock(&session->s_cap_lock);
751 p = session->s_caps.next;
752 while (p != &session->s_caps) {
753 cap = list_entry(p, struct ceph_cap, session_caps);
754 inode = igrab(&cap->ci->vfs_inode);
759 session->s_cap_iterator = cap;
760 spin_unlock(&session->s_cap_lock);
767 ceph_put_cap(old_cap);
771 ret = cb(inode, cap, arg);
774 spin_lock(&session->s_cap_lock);
776 if (cap->ci == NULL) {
777 dout("iterate_session_caps finishing cap %p removal\n",
779 BUG_ON(cap->session != session);
780 list_del_init(&cap->session_caps);
781 session->s_nr_caps--;
783 old_cap = cap; /* put_cap it w/o locks held */
790 session->s_cap_iterator = NULL;
791 spin_unlock(&session->s_cap_lock);
796 ceph_put_cap(old_cap);
801 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
804 struct ceph_inode_info *ci = ceph_inode(inode);
807 dout("removing cap %p, ci is %p, inode is %p\n",
808 cap, ci, &ci->vfs_inode);
809 spin_lock(&inode->i_lock);
810 __ceph_remove_cap(cap);
811 if (!__ceph_is_any_real_caps(ci)) {
812 struct ceph_mds_client *mdsc =
813 &ceph_sb_to_client(inode->i_sb)->mdsc;
815 spin_lock(&mdsc->cap_dirty_lock);
816 if (!list_empty(&ci->i_dirty_item)) {
817 pr_info(" dropping dirty %s state for %p %lld\n",
818 ceph_cap_string(ci->i_dirty_caps),
819 inode, ceph_ino(inode));
820 ci->i_dirty_caps = 0;
821 list_del_init(&ci->i_dirty_item);
824 if (!list_empty(&ci->i_flushing_item)) {
825 pr_info(" dropping dirty+flushing %s state for %p %lld\n",
826 ceph_cap_string(ci->i_flushing_caps),
827 inode, ceph_ino(inode));
828 ci->i_flushing_caps = 0;
829 list_del_init(&ci->i_flushing_item);
830 mdsc->num_cap_flushing--;
833 if (drop && ci->i_wrbuffer_ref) {
834 pr_info(" dropping dirty data for %p %lld\n",
835 inode, ceph_ino(inode));
836 ci->i_wrbuffer_ref = 0;
837 ci->i_wrbuffer_ref_head = 0;
840 spin_unlock(&mdsc->cap_dirty_lock);
842 spin_unlock(&inode->i_lock);
849 * caller must hold session s_mutex
851 static void remove_session_caps(struct ceph_mds_session *session)
853 dout("remove_session_caps on %p\n", session);
854 iterate_session_caps(session, remove_session_caps_cb, NULL);
855 BUG_ON(session->s_nr_caps > 0);
856 BUG_ON(!list_empty(&session->s_cap_flushing));
857 cleanup_cap_releases(session);
861 * wake up any threads waiting on this session's caps. if the cap is
862 * old (didn't get renewed on the client reconnect), remove it now.
864 * caller must hold s_mutex.
866 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
869 struct ceph_inode_info *ci = ceph_inode(inode);
871 wake_up_all(&ci->i_cap_wq);
873 spin_lock(&inode->i_lock);
874 ci->i_wanted_max_size = 0;
875 ci->i_requested_max_size = 0;
876 spin_unlock(&inode->i_lock);
881 static void wake_up_session_caps(struct ceph_mds_session *session,
884 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
885 iterate_session_caps(session, wake_up_session_cb,
886 (void *)(unsigned long)reconnect);
890 * Send periodic message to MDS renewing all currently held caps. The
891 * ack will reset the expiration for all caps from this session.
893 * caller holds s_mutex
895 static int send_renew_caps(struct ceph_mds_client *mdsc,
896 struct ceph_mds_session *session)
898 struct ceph_msg *msg;
901 if (time_after_eq(jiffies, session->s_cap_ttl) &&
902 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
903 pr_info("mds%d caps stale\n", session->s_mds);
904 session->s_renew_requested = jiffies;
906 /* do not try to renew caps until a recovering mds has reconnected
907 * with its clients. */
908 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
909 if (state < CEPH_MDS_STATE_RECONNECT) {
910 dout("send_renew_caps ignoring mds%d (%s)\n",
911 session->s_mds, ceph_mds_state_name(state));
915 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
916 ceph_mds_state_name(state));
917 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
918 ++session->s_renew_seq);
921 ceph_con_send(&session->s_con, msg);
926 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
928 * Called under session->s_mutex
930 static void renewed_caps(struct ceph_mds_client *mdsc,
931 struct ceph_mds_session *session, int is_renew)
936 spin_lock(&session->s_cap_lock);
937 was_stale = is_renew && (session->s_cap_ttl == 0 ||
938 time_after_eq(jiffies, session->s_cap_ttl));
940 session->s_cap_ttl = session->s_renew_requested +
941 mdsc->mdsmap->m_session_timeout*HZ;
944 if (time_before(jiffies, session->s_cap_ttl)) {
945 pr_info("mds%d caps renewed\n", session->s_mds);
948 pr_info("mds%d caps still stale\n", session->s_mds);
951 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
952 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
953 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
954 spin_unlock(&session->s_cap_lock);
957 wake_up_session_caps(session, 0);
961 * send a session close request
963 static int request_close_session(struct ceph_mds_client *mdsc,
964 struct ceph_mds_session *session)
966 struct ceph_msg *msg;
968 dout("request_close_session mds%d state %s seq %lld\n",
969 session->s_mds, session_state_name(session->s_state),
971 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
974 ceph_con_send(&session->s_con, msg);
979 * Called with s_mutex held.
981 static int __close_session(struct ceph_mds_client *mdsc,
982 struct ceph_mds_session *session)
984 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
986 session->s_state = CEPH_MDS_SESSION_CLOSING;
987 return request_close_session(mdsc, session);
993 * Because we can't cache an inode without one or more caps, we do
994 * this indirectly: if a cap is unused, we prune its aliases, at which
995 * point the inode will hopefully get dropped to.
997 * Yes, this is a bit sloppy. Our only real goal here is to respond to
998 * memory pressure from the MDS, though, so it needn't be perfect.
1000 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1002 struct ceph_mds_session *session = arg;
1003 struct ceph_inode_info *ci = ceph_inode(inode);
1004 int used, oissued, mine;
1006 if (session->s_trim_caps <= 0)
1009 spin_lock(&inode->i_lock);
1010 mine = cap->issued | cap->implemented;
1011 used = __ceph_caps_used(ci);
1012 oissued = __ceph_caps_issued_other(ci, cap);
1014 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1015 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1016 ceph_cap_string(used));
1017 if (ci->i_dirty_caps)
1018 goto out; /* dirty caps */
1019 if ((used & ~oissued) & mine)
1020 goto out; /* we need these caps */
1022 session->s_trim_caps--;
1024 /* we aren't the only cap.. just remove us */
1025 __ceph_remove_cap(cap);
1027 /* try to drop referring dentries */
1028 spin_unlock(&inode->i_lock);
1029 d_prune_aliases(inode);
1030 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1031 inode, cap, atomic_read(&inode->i_count));
1036 spin_unlock(&inode->i_lock);
1041 * Trim session cap count down to some max number.
1043 static int trim_caps(struct ceph_mds_client *mdsc,
1044 struct ceph_mds_session *session,
1047 int trim_caps = session->s_nr_caps - max_caps;
1049 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1050 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1051 if (trim_caps > 0) {
1052 session->s_trim_caps = trim_caps;
1053 iterate_session_caps(session, trim_caps_cb, session);
1054 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1055 session->s_mds, session->s_nr_caps, max_caps,
1056 trim_caps - session->s_trim_caps);
1057 session->s_trim_caps = 0;
1063 * Allocate cap_release messages. If there is a partially full message
1064 * in the queue, try to allocate enough to cover it's remainder, so that
1065 * we can send it immediately.
1067 * Called under s_mutex.
1069 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1070 struct ceph_mds_session *session)
1072 struct ceph_msg *msg, *partial = NULL;
1073 struct ceph_mds_cap_release *head;
1075 int extra = mdsc->client->mount_args->cap_release_safety;
1078 dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1081 spin_lock(&session->s_cap_lock);
1083 if (!list_empty(&session->s_cap_releases)) {
1084 msg = list_first_entry(&session->s_cap_releases,
1087 head = msg->front.iov_base;
1088 num = le32_to_cpu(head->num);
1090 dout(" partial %p with (%d/%d)\n", msg, num,
1091 (int)CEPH_CAPS_PER_RELEASE);
1092 extra += CEPH_CAPS_PER_RELEASE - num;
1096 while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1097 spin_unlock(&session->s_cap_lock);
1098 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1102 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1103 (int)msg->front.iov_len);
1104 head = msg->front.iov_base;
1105 head->num = cpu_to_le32(0);
1106 msg->front.iov_len = sizeof(*head);
1107 spin_lock(&session->s_cap_lock);
1108 list_add(&msg->list_head, &session->s_cap_releases);
1109 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1113 head = partial->front.iov_base;
1114 num = le32_to_cpu(head->num);
1115 dout(" queueing partial %p with %d/%d\n", partial, num,
1116 (int)CEPH_CAPS_PER_RELEASE);
1117 list_move_tail(&partial->list_head,
1118 &session->s_cap_releases_done);
1119 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1122 spin_unlock(&session->s_cap_lock);
1128 * flush all dirty inode data to disk.
1130 * returns true if we've flushed through want_flush_seq
1132 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1136 dout("check_cap_flush want %lld\n", want_flush_seq);
1137 mutex_lock(&mdsc->mutex);
1138 for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1139 struct ceph_mds_session *session = mdsc->sessions[mds];
1143 get_session(session);
1144 mutex_unlock(&mdsc->mutex);
1146 mutex_lock(&session->s_mutex);
1147 if (!list_empty(&session->s_cap_flushing)) {
1148 struct ceph_inode_info *ci =
1149 list_entry(session->s_cap_flushing.next,
1150 struct ceph_inode_info,
1152 struct inode *inode = &ci->vfs_inode;
1154 spin_lock(&inode->i_lock);
1155 if (ci->i_cap_flush_seq <= want_flush_seq) {
1156 dout("check_cap_flush still flushing %p "
1157 "seq %lld <= %lld to mds%d\n", inode,
1158 ci->i_cap_flush_seq, want_flush_seq,
1162 spin_unlock(&inode->i_lock);
1164 mutex_unlock(&session->s_mutex);
1165 ceph_put_mds_session(session);
1169 mutex_lock(&mdsc->mutex);
1172 mutex_unlock(&mdsc->mutex);
1173 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1178 * called under s_mutex
1180 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1181 struct ceph_mds_session *session)
1183 struct ceph_msg *msg;
1185 dout("send_cap_releases mds%d\n", session->s_mds);
1186 spin_lock(&session->s_cap_lock);
1187 while (!list_empty(&session->s_cap_releases_done)) {
1188 msg = list_first_entry(&session->s_cap_releases_done,
1189 struct ceph_msg, list_head);
1190 list_del_init(&msg->list_head);
1191 spin_unlock(&session->s_cap_lock);
1192 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1193 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1194 ceph_con_send(&session->s_con, msg);
1195 spin_lock(&session->s_cap_lock);
1197 spin_unlock(&session->s_cap_lock);
1200 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1201 struct ceph_mds_session *session)
1203 struct ceph_msg *msg;
1204 struct ceph_mds_cap_release *head;
1207 dout("discard_cap_releases mds%d\n", session->s_mds);
1208 spin_lock(&session->s_cap_lock);
1210 /* zero out the in-progress message */
1211 msg = list_first_entry(&session->s_cap_releases,
1212 struct ceph_msg, list_head);
1213 head = msg->front.iov_base;
1214 num = le32_to_cpu(head->num);
1215 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1216 head->num = cpu_to_le32(0);
1217 session->s_num_cap_releases += num;
1219 /* requeue completed messages */
1220 while (!list_empty(&session->s_cap_releases_done)) {
1221 msg = list_first_entry(&session->s_cap_releases_done,
1222 struct ceph_msg, list_head);
1223 list_del_init(&msg->list_head);
1225 head = msg->front.iov_base;
1226 num = le32_to_cpu(head->num);
1227 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1229 session->s_num_cap_releases += num;
1230 head->num = cpu_to_le32(0);
1231 msg->front.iov_len = sizeof(*head);
1232 list_add(&msg->list_head, &session->s_cap_releases);
1235 spin_unlock(&session->s_cap_lock);
1243 * Create an mds request.
1245 struct ceph_mds_request *
1246 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1248 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1251 return ERR_PTR(-ENOMEM);
1253 mutex_init(&req->r_fill_mutex);
1254 req->r_started = jiffies;
1255 req->r_resend_mds = -1;
1256 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1258 kref_init(&req->r_kref);
1259 INIT_LIST_HEAD(&req->r_wait);
1260 init_completion(&req->r_completion);
1261 init_completion(&req->r_safe_completion);
1262 INIT_LIST_HEAD(&req->r_unsafe_item);
1265 req->r_direct_mode = mode;
1270 * return oldest (lowest) request, tid in request tree, 0 if none.
1272 * called under mdsc->mutex.
1274 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1276 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1278 return rb_entry(rb_first(&mdsc->request_tree),
1279 struct ceph_mds_request, r_node);
1282 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1284 struct ceph_mds_request *req = __get_oldest_req(mdsc);
1292 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1293 * on build_path_from_dentry in fs/cifs/dir.c.
1295 * If @stop_on_nosnap, generate path relative to the first non-snapped
1298 * Encode hidden .snap dirs as a double /, i.e.
1299 * foo/.snap/bar -> foo//bar
1301 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1304 struct dentry *temp;
1309 return ERR_PTR(-EINVAL);
1313 for (temp = dentry; !IS_ROOT(temp);) {
1314 struct inode *inode = temp->d_inode;
1315 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1316 len++; /* slash only */
1317 else if (stop_on_nosnap && inode &&
1318 ceph_snap(inode) == CEPH_NOSNAP)
1321 len += 1 + temp->d_name.len;
1322 temp = temp->d_parent;
1324 pr_err("build_path corrupt dentry %p\n", dentry);
1325 return ERR_PTR(-EINVAL);
1329 len--; /* no leading '/' */
1331 path = kmalloc(len+1, GFP_NOFS);
1333 return ERR_PTR(-ENOMEM);
1335 path[pos] = 0; /* trailing null */
1336 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1337 struct inode *inode = temp->d_inode;
1339 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1340 dout("build_path path+%d: %p SNAPDIR\n",
1342 } else if (stop_on_nosnap && inode &&
1343 ceph_snap(inode) == CEPH_NOSNAP) {
1346 pos -= temp->d_name.len;
1349 strncpy(path + pos, temp->d_name.name,
1354 temp = temp->d_parent;
1356 pr_err("build_path corrupt dentry\n");
1358 return ERR_PTR(-EINVAL);
1362 pr_err("build_path did not end path lookup where "
1363 "expected, namelen is %d, pos is %d\n", len, pos);
1364 /* presumably this is only possible if racing with a
1365 rename of one of the parent directories (we can not
1366 lock the dentries above us to prevent this, but
1367 retrying should be harmless) */
1372 *base = ceph_ino(temp->d_inode);
1374 dout("build_path on %p %d built %llx '%.*s'\n",
1375 dentry, atomic_read(&dentry->d_count), *base, len, path);
1379 static int build_dentry_path(struct dentry *dentry,
1380 const char **ppath, int *ppathlen, u64 *pino,
1385 if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1386 *pino = ceph_ino(dentry->d_parent->d_inode);
1387 *ppath = dentry->d_name.name;
1388 *ppathlen = dentry->d_name.len;
1391 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1393 return PTR_ERR(path);
1399 static int build_inode_path(struct inode *inode,
1400 const char **ppath, int *ppathlen, u64 *pino,
1403 struct dentry *dentry;
1406 if (ceph_snap(inode) == CEPH_NOSNAP) {
1407 *pino = ceph_ino(inode);
1411 dentry = d_find_alias(inode);
1412 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1415 return PTR_ERR(path);
1422 * request arguments may be specified via an inode *, a dentry *, or
1423 * an explicit ino+path.
1425 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1426 const char *rpath, u64 rino,
1427 const char **ppath, int *pathlen,
1428 u64 *ino, int *freepath)
1433 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1434 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1436 } else if (rdentry) {
1437 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1438 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1443 *pathlen = strlen(rpath);
1444 dout(" path %.*s\n", *pathlen, rpath);
1451 * called under mdsc->mutex
1453 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1454 struct ceph_mds_request *req,
1457 struct ceph_msg *msg;
1458 struct ceph_mds_request_head *head;
1459 const char *path1 = NULL;
1460 const char *path2 = NULL;
1461 u64 ino1 = 0, ino2 = 0;
1462 int pathlen1 = 0, pathlen2 = 0;
1463 int freepath1 = 0, freepath2 = 0;
1469 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1470 req->r_path1, req->r_ino1.ino,
1471 &path1, &pathlen1, &ino1, &freepath1);
1477 ret = set_request_path_attr(NULL, req->r_old_dentry,
1478 req->r_path2, req->r_ino2.ino,
1479 &path2, &pathlen2, &ino2, &freepath2);
1485 len = sizeof(*head) +
1486 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1488 /* calculate (max) length for cap releases */
1489 len += sizeof(struct ceph_mds_request_release) *
1490 (!!req->r_inode_drop + !!req->r_dentry_drop +
1491 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1492 if (req->r_dentry_drop)
1493 len += req->r_dentry->d_name.len;
1494 if (req->r_old_dentry_drop)
1495 len += req->r_old_dentry->d_name.len;
1497 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS);
1499 msg = ERR_PTR(-ENOMEM);
1503 msg->hdr.tid = cpu_to_le64(req->r_tid);
1505 head = msg->front.iov_base;
1506 p = msg->front.iov_base + sizeof(*head);
1507 end = msg->front.iov_base + msg->front.iov_len;
1509 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1510 head->op = cpu_to_le32(req->r_op);
1511 head->caller_uid = cpu_to_le32(current_fsuid());
1512 head->caller_gid = cpu_to_le32(current_fsgid());
1513 head->args = req->r_args;
1515 ceph_encode_filepath(&p, end, ino1, path1);
1516 ceph_encode_filepath(&p, end, ino2, path2);
1518 /* make note of release offset, in case we need to replay */
1519 req->r_request_release_offset = p - msg->front.iov_base;
1523 if (req->r_inode_drop)
1524 releases += ceph_encode_inode_release(&p,
1525 req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1526 mds, req->r_inode_drop, req->r_inode_unless, 0);
1527 if (req->r_dentry_drop)
1528 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1529 mds, req->r_dentry_drop, req->r_dentry_unless);
1530 if (req->r_old_dentry_drop)
1531 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1532 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1533 if (req->r_old_inode_drop)
1534 releases += ceph_encode_inode_release(&p,
1535 req->r_old_dentry->d_inode,
1536 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1537 head->num_releases = cpu_to_le16(releases);
1540 msg->front.iov_len = p - msg->front.iov_base;
1541 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1543 msg->pages = req->r_pages;
1544 msg->nr_pages = req->r_num_pages;
1545 msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1546 msg->hdr.data_off = cpu_to_le16(0);
1550 kfree((char *)path2);
1553 kfree((char *)path1);
1559 * called under mdsc->mutex if error, under no mutex if
1562 static void complete_request(struct ceph_mds_client *mdsc,
1563 struct ceph_mds_request *req)
1565 if (req->r_callback)
1566 req->r_callback(mdsc, req);
1568 complete_all(&req->r_completion);
1572 * called under mdsc->mutex
1574 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1575 struct ceph_mds_request *req,
1578 struct ceph_mds_request_head *rhead;
1579 struct ceph_msg *msg;
1584 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1585 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1587 if (req->r_got_unsafe) {
1589 * Replay. Do not regenerate message (and rebuild
1590 * paths, etc.); just use the original message.
1591 * Rebuilding paths will break for renames because
1592 * d_move mangles the src name.
1594 msg = req->r_request;
1595 rhead = msg->front.iov_base;
1597 flags = le32_to_cpu(rhead->flags);
1598 flags |= CEPH_MDS_FLAG_REPLAY;
1599 rhead->flags = cpu_to_le32(flags);
1601 if (req->r_target_inode)
1602 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1604 rhead->num_retry = req->r_attempts - 1;
1606 /* remove cap/dentry releases from message */
1607 rhead->num_releases = 0;
1608 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1609 msg->front.iov_len = req->r_request_release_offset;
1613 if (req->r_request) {
1614 ceph_msg_put(req->r_request);
1615 req->r_request = NULL;
1617 msg = create_request_message(mdsc, req, mds);
1619 req->r_err = PTR_ERR(msg);
1620 complete_request(mdsc, req);
1621 return PTR_ERR(msg);
1623 req->r_request = msg;
1625 rhead = msg->front.iov_base;
1626 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1627 if (req->r_got_unsafe)
1628 flags |= CEPH_MDS_FLAG_REPLAY;
1629 if (req->r_locked_dir)
1630 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1631 rhead->flags = cpu_to_le32(flags);
1632 rhead->num_fwd = req->r_num_fwd;
1633 rhead->num_retry = req->r_attempts - 1;
1636 dout(" r_locked_dir = %p\n", req->r_locked_dir);
1641 * send request, or put it on the appropriate wait list.
1643 static int __do_request(struct ceph_mds_client *mdsc,
1644 struct ceph_mds_request *req)
1646 struct ceph_mds_session *session = NULL;
1650 if (req->r_err || req->r_got_result)
1653 if (req->r_timeout &&
1654 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1655 dout("do_request timed out\n");
1660 mds = __choose_mds(mdsc, req);
1662 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1663 dout("do_request no mds or not active, waiting for map\n");
1664 list_add(&req->r_wait, &mdsc->waiting_for_map);
1668 /* get, open session */
1669 session = __ceph_lookup_mds_session(mdsc, mds);
1671 session = register_session(mdsc, mds);
1672 if (IS_ERR(session)) {
1673 err = PTR_ERR(session);
1677 dout("do_request mds%d session %p state %s\n", mds, session,
1678 session_state_name(session->s_state));
1679 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1680 session->s_state != CEPH_MDS_SESSION_HUNG) {
1681 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1682 session->s_state == CEPH_MDS_SESSION_CLOSING)
1683 __open_session(mdsc, session);
1684 list_add(&req->r_wait, &session->s_waiting);
1689 req->r_session = get_session(session);
1690 req->r_resend_mds = -1; /* forget any previous mds hint */
1692 if (req->r_request_started == 0) /* note request start time */
1693 req->r_request_started = jiffies;
1695 err = __prepare_send_request(mdsc, req, mds);
1697 ceph_msg_get(req->r_request);
1698 ceph_con_send(&session->s_con, req->r_request);
1702 ceph_put_mds_session(session);
1708 complete_request(mdsc, req);
1713 * called under mdsc->mutex
1715 static void __wake_requests(struct ceph_mds_client *mdsc,
1716 struct list_head *head)
1718 struct ceph_mds_request *req, *nreq;
1720 list_for_each_entry_safe(req, nreq, head, r_wait) {
1721 list_del_init(&req->r_wait);
1722 __do_request(mdsc, req);
1727 * Wake up threads with requests pending for @mds, so that they can
1728 * resubmit their requests to a possibly different mds.
1730 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1732 struct ceph_mds_request *req;
1735 dout("kick_requests mds%d\n", mds);
1736 for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1737 req = rb_entry(p, struct ceph_mds_request, r_node);
1738 if (req->r_got_unsafe)
1740 if (req->r_session &&
1741 req->r_session->s_mds == mds) {
1742 dout(" kicking tid %llu\n", req->r_tid);
1743 put_request_session(req);
1744 __do_request(mdsc, req);
1749 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1750 struct ceph_mds_request *req)
1752 dout("submit_request on %p\n", req);
1753 mutex_lock(&mdsc->mutex);
1754 __register_request(mdsc, req, NULL);
1755 __do_request(mdsc, req);
1756 mutex_unlock(&mdsc->mutex);
1760 * Synchrously perform an mds request. Take care of all of the
1761 * session setup, forwarding, retry details.
1763 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1765 struct ceph_mds_request *req)
1769 dout("do_request on %p\n", req);
1771 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1773 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1774 if (req->r_locked_dir)
1775 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1776 if (req->r_old_dentry)
1778 ceph_inode(req->r_old_dentry->d_parent->d_inode),
1782 mutex_lock(&mdsc->mutex);
1783 __register_request(mdsc, req, dir);
1784 __do_request(mdsc, req);
1788 __unregister_request(mdsc, req);
1789 dout("do_request early error %d\n", err);
1794 mutex_unlock(&mdsc->mutex);
1795 dout("do_request waiting\n");
1796 if (req->r_timeout) {
1797 err = (long)wait_for_completion_killable_timeout(
1798 &req->r_completion, req->r_timeout);
1802 err = wait_for_completion_killable(&req->r_completion);
1804 dout("do_request waited, got %d\n", err);
1805 mutex_lock(&mdsc->mutex);
1807 /* only abort if we didn't race with a real reply */
1808 if (req->r_got_result) {
1809 err = le32_to_cpu(req->r_reply_info.head->result);
1810 } else if (err < 0) {
1811 dout("aborted request %lld with %d\n", req->r_tid, err);
1814 * ensure we aren't running concurrently with
1815 * ceph_fill_trace or ceph_readdir_prepopulate, which
1816 * rely on locks (dir mutex) held by our caller.
1818 mutex_lock(&req->r_fill_mutex);
1820 req->r_aborted = true;
1821 mutex_unlock(&req->r_fill_mutex);
1823 if (req->r_locked_dir &&
1824 (req->r_op & CEPH_MDS_OP_WRITE))
1825 ceph_invalidate_dir_request(req);
1831 mutex_unlock(&mdsc->mutex);
1832 dout("do_request %p done, result %d\n", req, err);
1837 * Invalidate dir I_COMPLETE, dentry lease state on an aborted MDS
1838 * namespace request.
1840 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
1842 struct inode *inode = req->r_locked_dir;
1843 struct ceph_inode_info *ci = ceph_inode(inode);
1845 dout("invalidate_dir_request %p (I_COMPLETE, lease(s))\n", inode);
1846 spin_lock(&inode->i_lock);
1847 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1848 ci->i_release_count++;
1849 spin_unlock(&inode->i_lock);
1852 ceph_invalidate_dentry_lease(req->r_dentry);
1853 if (req->r_old_dentry)
1854 ceph_invalidate_dentry_lease(req->r_old_dentry);
1860 * We take the session mutex and parse and process the reply immediately.
1861 * This preserves the logical ordering of replies, capabilities, etc., sent
1862 * by the MDS as they are applied to our local cache.
1864 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1866 struct ceph_mds_client *mdsc = session->s_mdsc;
1867 struct ceph_mds_request *req;
1868 struct ceph_mds_reply_head *head = msg->front.iov_base;
1869 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
1872 int mds = session->s_mds;
1874 if (msg->front.iov_len < sizeof(*head)) {
1875 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1880 /* get request, session */
1881 tid = le64_to_cpu(msg->hdr.tid);
1882 mutex_lock(&mdsc->mutex);
1883 req = __lookup_request(mdsc, tid);
1885 dout("handle_reply on unknown tid %llu\n", tid);
1886 mutex_unlock(&mdsc->mutex);
1889 dout("handle_reply %p\n", req);
1891 /* correct session? */
1892 if (req->r_session != session) {
1893 pr_err("mdsc_handle_reply got %llu on session mds%d"
1894 " not mds%d\n", tid, session->s_mds,
1895 req->r_session ? req->r_session->s_mds : -1);
1896 mutex_unlock(&mdsc->mutex);
1901 if ((req->r_got_unsafe && !head->safe) ||
1902 (req->r_got_safe && head->safe)) {
1903 pr_warning("got a dup %s reply on %llu from mds%d\n",
1904 head->safe ? "safe" : "unsafe", tid, mds);
1905 mutex_unlock(&mdsc->mutex);
1908 if (req->r_got_safe && !head->safe) {
1909 pr_warning("got unsafe after safe on %llu from mds%d\n",
1911 mutex_unlock(&mdsc->mutex);
1915 result = le32_to_cpu(head->result);
1918 * Tolerate 2 consecutive ESTALEs from the same mds.
1919 * FIXME: we should be looking at the cap migrate_seq.
1921 if (result == -ESTALE) {
1922 req->r_direct_mode = USE_AUTH_MDS;
1924 if (req->r_num_stale <= 2) {
1925 __do_request(mdsc, req);
1926 mutex_unlock(&mdsc->mutex);
1930 req->r_num_stale = 0;
1934 req->r_got_safe = true;
1935 __unregister_request(mdsc, req);
1936 complete_all(&req->r_safe_completion);
1938 if (req->r_got_unsafe) {
1940 * We already handled the unsafe response, now do the
1941 * cleanup. No need to examine the response; the MDS
1942 * doesn't include any result info in the safe
1943 * response. And even if it did, there is nothing
1944 * useful we could do with a revised return value.
1946 dout("got safe reply %llu, mds%d\n", tid, mds);
1947 list_del_init(&req->r_unsafe_item);
1949 /* last unsafe request during umount? */
1950 if (mdsc->stopping && !__get_oldest_req(mdsc))
1951 complete_all(&mdsc->safe_umount_waiters);
1952 mutex_unlock(&mdsc->mutex);
1956 req->r_got_unsafe = true;
1957 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1960 dout("handle_reply tid %lld result %d\n", tid, result);
1961 rinfo = &req->r_reply_info;
1962 err = parse_reply_info(msg, rinfo);
1963 mutex_unlock(&mdsc->mutex);
1965 mutex_lock(&session->s_mutex);
1967 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1973 if (rinfo->snapblob_len) {
1974 down_write(&mdsc->snap_rwsem);
1975 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1976 rinfo->snapblob + rinfo->snapblob_len,
1977 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1978 downgrade_write(&mdsc->snap_rwsem);
1980 down_read(&mdsc->snap_rwsem);
1983 /* insert trace into our cache */
1984 mutex_lock(&req->r_fill_mutex);
1985 err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1987 if (result == 0 && rinfo->dir_nr)
1988 ceph_readdir_prepopulate(req, req->r_session);
1989 ceph_unreserve_caps(&req->r_caps_reservation);
1991 mutex_unlock(&req->r_fill_mutex);
1993 up_read(&mdsc->snap_rwsem);
1995 mutex_lock(&mdsc->mutex);
1996 if (!req->r_aborted) {
2002 req->r_got_result = true;
2005 dout("reply arrived after request %lld was aborted\n", tid);
2007 mutex_unlock(&mdsc->mutex);
2009 ceph_add_cap_releases(mdsc, req->r_session);
2010 mutex_unlock(&session->s_mutex);
2012 /* kick calling process */
2013 complete_request(mdsc, req);
2015 ceph_mdsc_put_request(req);
2022 * handle mds notification that our request has been forwarded.
2024 static void handle_forward(struct ceph_mds_client *mdsc,
2025 struct ceph_mds_session *session,
2026 struct ceph_msg *msg)
2028 struct ceph_mds_request *req;
2029 u64 tid = le64_to_cpu(msg->hdr.tid);
2033 void *p = msg->front.iov_base;
2034 void *end = p + msg->front.iov_len;
2036 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2037 next_mds = ceph_decode_32(&p);
2038 fwd_seq = ceph_decode_32(&p);
2040 mutex_lock(&mdsc->mutex);
2041 req = __lookup_request(mdsc, tid);
2043 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2044 goto out; /* dup reply? */
2047 if (req->r_aborted) {
2048 dout("forward tid %llu aborted, unregistering\n", tid);
2049 __unregister_request(mdsc, req);
2050 } else if (fwd_seq <= req->r_num_fwd) {
2051 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2052 tid, next_mds, req->r_num_fwd, fwd_seq);
2054 /* resend. forward race not possible; mds would drop */
2055 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2057 BUG_ON(req->r_got_result);
2058 req->r_num_fwd = fwd_seq;
2059 req->r_resend_mds = next_mds;
2060 put_request_session(req);
2061 __do_request(mdsc, req);
2063 ceph_mdsc_put_request(req);
2065 mutex_unlock(&mdsc->mutex);
2069 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2073 * handle a mds session control message
2075 static void handle_session(struct ceph_mds_session *session,
2076 struct ceph_msg *msg)
2078 struct ceph_mds_client *mdsc = session->s_mdsc;
2081 int mds = session->s_mds;
2082 struct ceph_mds_session_head *h = msg->front.iov_base;
2086 if (msg->front.iov_len != sizeof(*h))
2088 op = le32_to_cpu(h->op);
2089 seq = le64_to_cpu(h->seq);
2091 mutex_lock(&mdsc->mutex);
2092 if (op == CEPH_SESSION_CLOSE)
2093 __unregister_session(mdsc, session);
2094 /* FIXME: this ttl calculation is generous */
2095 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2096 mutex_unlock(&mdsc->mutex);
2098 mutex_lock(&session->s_mutex);
2100 dout("handle_session mds%d %s %p state %s seq %llu\n",
2101 mds, ceph_session_op_name(op), session,
2102 session_state_name(session->s_state), seq);
2104 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2105 session->s_state = CEPH_MDS_SESSION_OPEN;
2106 pr_info("mds%d came back\n", session->s_mds);
2110 case CEPH_SESSION_OPEN:
2111 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2112 pr_info("mds%d reconnect success\n", session->s_mds);
2113 session->s_state = CEPH_MDS_SESSION_OPEN;
2114 renewed_caps(mdsc, session, 0);
2117 __close_session(mdsc, session);
2120 case CEPH_SESSION_RENEWCAPS:
2121 if (session->s_renew_seq == seq)
2122 renewed_caps(mdsc, session, 1);
2125 case CEPH_SESSION_CLOSE:
2126 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2127 pr_info("mds%d reconnect denied\n", session->s_mds);
2128 remove_session_caps(session);
2129 wake = 1; /* for good measure */
2130 complete_all(&mdsc->session_close_waiters);
2131 kick_requests(mdsc, mds);
2134 case CEPH_SESSION_STALE:
2135 pr_info("mds%d caps went stale, renewing\n",
2137 spin_lock(&session->s_cap_lock);
2138 session->s_cap_gen++;
2139 session->s_cap_ttl = 0;
2140 spin_unlock(&session->s_cap_lock);
2141 send_renew_caps(mdsc, session);
2144 case CEPH_SESSION_RECALL_STATE:
2145 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2149 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2153 mutex_unlock(&session->s_mutex);
2155 mutex_lock(&mdsc->mutex);
2156 __wake_requests(mdsc, &session->s_waiting);
2157 mutex_unlock(&mdsc->mutex);
2162 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2163 (int)msg->front.iov_len);
2170 * called under session->mutex.
2172 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2173 struct ceph_mds_session *session)
2175 struct ceph_mds_request *req, *nreq;
2178 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2180 mutex_lock(&mdsc->mutex);
2181 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2182 err = __prepare_send_request(mdsc, req, session->s_mds);
2184 ceph_msg_get(req->r_request);
2185 ceph_con_send(&session->s_con, req->r_request);
2188 mutex_unlock(&mdsc->mutex);
2192 * Encode information about a cap for a reconnect with the MDS.
2194 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2197 struct ceph_mds_cap_reconnect rec;
2198 struct ceph_inode_info *ci;
2199 struct ceph_pagelist *pagelist = arg;
2203 struct dentry *dentry;
2207 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2208 inode, ceph_vinop(inode), cap, cap->cap_id,
2209 ceph_cap_string(cap->issued));
2210 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2214 dentry = d_find_alias(inode);
2216 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2218 err = PTR_ERR(path);
2225 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2229 spin_lock(&inode->i_lock);
2230 cap->seq = 0; /* reset cap seq */
2231 cap->issue_seq = 0; /* and issue_seq */
2232 rec.cap_id = cpu_to_le64(cap->cap_id);
2233 rec.pathbase = cpu_to_le64(pathbase);
2234 rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2235 rec.issued = cpu_to_le32(cap->issued);
2236 rec.size = cpu_to_le64(inode->i_size);
2237 ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2238 ceph_encode_timespec(&rec.atime, &inode->i_atime);
2239 rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2240 spin_unlock(&inode->i_lock);
2242 err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2252 * If an MDS fails and recovers, clients need to reconnect in order to
2253 * reestablish shared state. This includes all caps issued through
2254 * this session _and_ the snap_realm hierarchy. Because it's not
2255 * clear which snap realms the mds cares about, we send everything we
2256 * know about.. that ensures we'll then get any new info the
2257 * recovering MDS might have.
2259 * This is a relatively heavyweight operation, but it's rare.
2261 * called with mdsc->mutex held.
2263 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2264 struct ceph_mds_session *session)
2266 struct ceph_msg *reply;
2268 int mds = session->s_mds;
2270 struct ceph_pagelist *pagelist;
2272 pr_info("mds%d reconnect start\n", mds);
2274 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2276 goto fail_nopagelist;
2277 ceph_pagelist_init(pagelist);
2279 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS);
2283 mutex_lock(&session->s_mutex);
2284 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2287 ceph_con_open(&session->s_con,
2288 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2290 /* replay unsafe requests */
2291 replay_unsafe_requests(mdsc, session);
2293 down_read(&mdsc->snap_rwsem);
2295 dout("session %p state %s\n", session,
2296 session_state_name(session->s_state));
2298 /* drop old cap expires; we're about to reestablish that state */
2299 discard_cap_releases(mdsc, session);
2301 /* traverse this session's caps */
2302 err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2305 err = iterate_session_caps(session, encode_caps_cb, pagelist);
2310 * snaprealms. we provide mds with the ino, seq (version), and
2311 * parent for all of our realms. If the mds has any newer info,
2314 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2315 struct ceph_snap_realm *realm =
2316 rb_entry(p, struct ceph_snap_realm, node);
2317 struct ceph_mds_snaprealm_reconnect sr_rec;
2319 dout(" adding snap realm %llx seq %lld parent %llx\n",
2320 realm->ino, realm->seq, realm->parent_ino);
2321 sr_rec.ino = cpu_to_le64(realm->ino);
2322 sr_rec.seq = cpu_to_le64(realm->seq);
2323 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2324 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2329 reply->pagelist = pagelist;
2330 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2331 reply->nr_pages = calc_pages_for(0, pagelist->length);
2332 ceph_con_send(&session->s_con, reply);
2334 mutex_unlock(&session->s_mutex);
2336 mutex_lock(&mdsc->mutex);
2337 __wake_requests(mdsc, &session->s_waiting);
2338 mutex_unlock(&mdsc->mutex);
2340 up_read(&mdsc->snap_rwsem);
2344 ceph_msg_put(reply);
2345 up_read(&mdsc->snap_rwsem);
2346 mutex_unlock(&session->s_mutex);
2348 ceph_pagelist_release(pagelist);
2351 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2357 * compare old and new mdsmaps, kicking requests
2358 * and closing out old connections as necessary
2360 * called under mdsc->mutex.
2362 static void check_new_map(struct ceph_mds_client *mdsc,
2363 struct ceph_mdsmap *newmap,
2364 struct ceph_mdsmap *oldmap)
2367 int oldstate, newstate;
2368 struct ceph_mds_session *s;
2370 dout("check_new_map new %u old %u\n",
2371 newmap->m_epoch, oldmap->m_epoch);
2373 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2374 if (mdsc->sessions[i] == NULL)
2376 s = mdsc->sessions[i];
2377 oldstate = ceph_mdsmap_get_state(oldmap, i);
2378 newstate = ceph_mdsmap_get_state(newmap, i);
2380 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2381 i, ceph_mds_state_name(oldstate),
2382 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2383 ceph_mds_state_name(newstate),
2384 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2385 session_state_name(s->s_state));
2387 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2388 ceph_mdsmap_get_addr(newmap, i),
2389 sizeof(struct ceph_entity_addr))) {
2390 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2391 /* the session never opened, just close it
2393 __wake_requests(mdsc, &s->s_waiting);
2394 __unregister_session(mdsc, s);
2397 mutex_unlock(&mdsc->mutex);
2398 mutex_lock(&s->s_mutex);
2399 mutex_lock(&mdsc->mutex);
2400 ceph_con_close(&s->s_con);
2401 mutex_unlock(&s->s_mutex);
2402 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2405 /* kick any requests waiting on the recovering mds */
2406 kick_requests(mdsc, i);
2407 } else if (oldstate == newstate) {
2408 continue; /* nothing new with this mds */
2414 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2415 newstate >= CEPH_MDS_STATE_RECONNECT) {
2416 mutex_unlock(&mdsc->mutex);
2417 send_mds_reconnect(mdsc, s);
2418 mutex_lock(&mdsc->mutex);
2422 * kick request on any mds that has gone active.
2424 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2425 newstate >= CEPH_MDS_STATE_ACTIVE) {
2426 if (oldstate != CEPH_MDS_STATE_CREATING &&
2427 oldstate != CEPH_MDS_STATE_STARTING)
2428 pr_info("mds%d recovery completed\n", s->s_mds);
2429 kick_requests(mdsc, i);
2430 ceph_kick_flushing_caps(mdsc, s);
2431 wake_up_session_caps(s, 1);
2443 * caller must hold session s_mutex, dentry->d_lock
2445 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2447 struct ceph_dentry_info *di = ceph_dentry(dentry);
2449 ceph_put_mds_session(di->lease_session);
2450 di->lease_session = NULL;
2453 static void handle_lease(struct ceph_mds_client *mdsc,
2454 struct ceph_mds_session *session,
2455 struct ceph_msg *msg)
2457 struct super_block *sb = mdsc->client->sb;
2458 struct inode *inode;
2459 struct ceph_inode_info *ci;
2460 struct dentry *parent, *dentry;
2461 struct ceph_dentry_info *di;
2462 int mds = session->s_mds;
2463 struct ceph_mds_lease *h = msg->front.iov_base;
2465 struct ceph_vino vino;
2470 dout("handle_lease from mds%d\n", mds);
2473 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2475 vino.ino = le64_to_cpu(h->ino);
2476 vino.snap = CEPH_NOSNAP;
2477 mask = le16_to_cpu(h->mask);
2478 seq = le32_to_cpu(h->seq);
2479 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2480 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2481 if (dname.len != get_unaligned_le32(h+1))
2484 mutex_lock(&session->s_mutex);
2488 inode = ceph_find_inode(sb, vino);
2489 dout("handle_lease %s, mask %d, ino %llx %p %.*s\n",
2490 ceph_lease_op_name(h->action), mask, vino.ino, inode,
2491 dname.len, dname.name);
2492 if (inode == NULL) {
2493 dout("handle_lease no inode %llx\n", vino.ino);
2496 ci = ceph_inode(inode);
2499 parent = d_find_alias(inode);
2501 dout("no parent dentry on inode %p\n", inode);
2503 goto release; /* hrm... */
2505 dname.hash = full_name_hash(dname.name, dname.len);
2506 dentry = d_lookup(parent, &dname);
2511 spin_lock(&dentry->d_lock);
2512 di = ceph_dentry(dentry);
2513 switch (h->action) {
2514 case CEPH_MDS_LEASE_REVOKE:
2515 if (di && di->lease_session == session) {
2516 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2517 h->seq = cpu_to_le32(di->lease_seq);
2518 __ceph_mdsc_drop_dentry_lease(dentry);
2523 case CEPH_MDS_LEASE_RENEW:
2524 if (di && di->lease_session == session &&
2525 di->lease_gen == session->s_cap_gen &&
2526 di->lease_renew_from &&
2527 di->lease_renew_after == 0) {
2528 unsigned long duration =
2529 le32_to_cpu(h->duration_ms) * HZ / 1000;
2531 di->lease_seq = seq;
2532 dentry->d_time = di->lease_renew_from + duration;
2533 di->lease_renew_after = di->lease_renew_from +
2535 di->lease_renew_from = 0;
2539 spin_unlock(&dentry->d_lock);
2546 /* let's just reuse the same message */
2547 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2549 ceph_con_send(&session->s_con, msg);
2553 mutex_unlock(&session->s_mutex);
2557 pr_err("corrupt lease message\n");
2561 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2562 struct inode *inode,
2563 struct dentry *dentry, char action,
2566 struct ceph_msg *msg;
2567 struct ceph_mds_lease *lease;
2568 int len = sizeof(*lease) + sizeof(u32);
2571 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2572 inode, dentry, ceph_lease_op_name(action), session->s_mds);
2573 dnamelen = dentry->d_name.len;
2576 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS);
2579 lease = msg->front.iov_base;
2580 lease->action = action;
2581 lease->mask = cpu_to_le16(1);
2582 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2583 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2584 lease->seq = cpu_to_le32(seq);
2585 put_unaligned_le32(dnamelen, lease + 1);
2586 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2589 * if this is a preemptive lease RELEASE, no need to
2590 * flush request stream, since the actual request will
2593 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2595 ceph_con_send(&session->s_con, msg);
2599 * Preemptively release a lease we expect to invalidate anyway.
2600 * Pass @inode always, @dentry is optional.
2602 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2603 struct dentry *dentry, int mask)
2605 struct ceph_dentry_info *di;
2606 struct ceph_mds_session *session;
2609 BUG_ON(inode == NULL);
2610 BUG_ON(dentry == NULL);
2613 /* is dentry lease valid? */
2614 spin_lock(&dentry->d_lock);
2615 di = ceph_dentry(dentry);
2616 if (!di || !di->lease_session ||
2617 di->lease_session->s_mds < 0 ||
2618 di->lease_gen != di->lease_session->s_cap_gen ||
2619 !time_before(jiffies, dentry->d_time)) {
2620 dout("lease_release inode %p dentry %p -- "
2622 inode, dentry, mask);
2623 spin_unlock(&dentry->d_lock);
2627 /* we do have a lease on this dentry; note mds and seq */
2628 session = ceph_get_mds_session(di->lease_session);
2629 seq = di->lease_seq;
2630 __ceph_mdsc_drop_dentry_lease(dentry);
2631 spin_unlock(&dentry->d_lock);
2633 dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2634 inode, dentry, mask, session->s_mds);
2635 ceph_mdsc_lease_send_msg(session, inode, dentry,
2636 CEPH_MDS_LEASE_RELEASE, seq);
2637 ceph_put_mds_session(session);
2641 * drop all leases (and dentry refs) in preparation for umount
2643 static void drop_leases(struct ceph_mds_client *mdsc)
2647 dout("drop_leases\n");
2648 mutex_lock(&mdsc->mutex);
2649 for (i = 0; i < mdsc->max_sessions; i++) {
2650 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2653 mutex_unlock(&mdsc->mutex);
2654 mutex_lock(&s->s_mutex);
2655 mutex_unlock(&s->s_mutex);
2656 ceph_put_mds_session(s);
2657 mutex_lock(&mdsc->mutex);
2659 mutex_unlock(&mdsc->mutex);
2665 * delayed work -- periodically trim expired leases, renew caps with mds
2667 static void schedule_delayed(struct ceph_mds_client *mdsc)
2670 unsigned hz = round_jiffies_relative(HZ * delay);
2671 schedule_delayed_work(&mdsc->delayed_work, hz);
2674 static void delayed_work(struct work_struct *work)
2677 struct ceph_mds_client *mdsc =
2678 container_of(work, struct ceph_mds_client, delayed_work.work);
2682 dout("mdsc delayed_work\n");
2683 ceph_check_delayed_caps(mdsc);
2685 mutex_lock(&mdsc->mutex);
2686 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2687 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2688 mdsc->last_renew_caps);
2690 mdsc->last_renew_caps = jiffies;
2692 for (i = 0; i < mdsc->max_sessions; i++) {
2693 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2696 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2697 dout("resending session close request for mds%d\n",
2699 request_close_session(mdsc, s);
2700 ceph_put_mds_session(s);
2703 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2704 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2705 s->s_state = CEPH_MDS_SESSION_HUNG;
2706 pr_info("mds%d hung\n", s->s_mds);
2709 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2710 /* this mds is failed or recovering, just wait */
2711 ceph_put_mds_session(s);
2714 mutex_unlock(&mdsc->mutex);
2716 mutex_lock(&s->s_mutex);
2718 send_renew_caps(mdsc, s);
2720 ceph_con_keepalive(&s->s_con);
2721 ceph_add_cap_releases(mdsc, s);
2722 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2723 s->s_state == CEPH_MDS_SESSION_HUNG)
2724 ceph_send_cap_releases(mdsc, s);
2725 mutex_unlock(&s->s_mutex);
2726 ceph_put_mds_session(s);
2728 mutex_lock(&mdsc->mutex);
2730 mutex_unlock(&mdsc->mutex);
2732 schedule_delayed(mdsc);
2736 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2738 mdsc->client = client;
2739 mutex_init(&mdsc->mutex);
2740 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2741 if (mdsc->mdsmap == NULL)
2744 init_completion(&mdsc->safe_umount_waiters);
2745 init_completion(&mdsc->session_close_waiters);
2746 INIT_LIST_HEAD(&mdsc->waiting_for_map);
2747 mdsc->sessions = NULL;
2748 mdsc->max_sessions = 0;
2750 init_rwsem(&mdsc->snap_rwsem);
2751 mdsc->snap_realms = RB_ROOT;
2752 INIT_LIST_HEAD(&mdsc->snap_empty);
2753 spin_lock_init(&mdsc->snap_empty_lock);
2755 mdsc->request_tree = RB_ROOT;
2756 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2757 mdsc->last_renew_caps = jiffies;
2758 INIT_LIST_HEAD(&mdsc->cap_delay_list);
2759 spin_lock_init(&mdsc->cap_delay_lock);
2760 INIT_LIST_HEAD(&mdsc->snap_flush_list);
2761 spin_lock_init(&mdsc->snap_flush_lock);
2762 mdsc->cap_flush_seq = 0;
2763 INIT_LIST_HEAD(&mdsc->cap_dirty);
2764 mdsc->num_cap_flushing = 0;
2765 spin_lock_init(&mdsc->cap_dirty_lock);
2766 init_waitqueue_head(&mdsc->cap_flushing_wq);
2767 spin_lock_init(&mdsc->dentry_lru_lock);
2768 INIT_LIST_HEAD(&mdsc->dentry_lru);
2774 * Wait for safe replies on open mds requests. If we time out, drop
2775 * all requests from the tree to avoid dangling dentry refs.
2777 static void wait_requests(struct ceph_mds_client *mdsc)
2779 struct ceph_mds_request *req;
2780 struct ceph_client *client = mdsc->client;
2782 mutex_lock(&mdsc->mutex);
2783 if (__get_oldest_req(mdsc)) {
2784 mutex_unlock(&mdsc->mutex);
2786 dout("wait_requests waiting for requests\n");
2787 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2788 client->mount_args->mount_timeout * HZ);
2790 /* tear down remaining requests */
2791 mutex_lock(&mdsc->mutex);
2792 while ((req = __get_oldest_req(mdsc))) {
2793 dout("wait_requests timed out on tid %llu\n",
2795 __unregister_request(mdsc, req);
2798 mutex_unlock(&mdsc->mutex);
2799 dout("wait_requests done\n");
2803 * called before mount is ro, and before dentries are torn down.
2804 * (hmm, does this still race with new lookups?)
2806 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2808 dout("pre_umount\n");
2812 ceph_flush_dirty_caps(mdsc);
2813 wait_requests(mdsc);
2816 * wait for reply handlers to drop their request refs and
2817 * their inode/dcache refs
2823 * wait for all write mds requests to flush.
2825 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2827 struct ceph_mds_request *req = NULL, *nextreq;
2830 mutex_lock(&mdsc->mutex);
2831 dout("wait_unsafe_requests want %lld\n", want_tid);
2833 req = __get_oldest_req(mdsc);
2834 while (req && req->r_tid <= want_tid) {
2835 /* find next request */
2836 n = rb_next(&req->r_node);
2838 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2841 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2843 ceph_mdsc_get_request(req);
2845 ceph_mdsc_get_request(nextreq);
2846 mutex_unlock(&mdsc->mutex);
2847 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
2848 req->r_tid, want_tid);
2849 wait_for_completion(&req->r_safe_completion);
2850 mutex_lock(&mdsc->mutex);
2851 ceph_mdsc_put_request(req);
2853 break; /* next dne before, so we're done! */
2854 if (RB_EMPTY_NODE(&nextreq->r_node)) {
2855 /* next request was removed from tree */
2856 ceph_mdsc_put_request(nextreq);
2859 ceph_mdsc_put_request(nextreq); /* won't go away */
2863 mutex_unlock(&mdsc->mutex);
2864 dout("wait_unsafe_requests done\n");
2867 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2869 u64 want_tid, want_flush;
2871 if (mdsc->client->mount_state == CEPH_MOUNT_SHUTDOWN)
2875 mutex_lock(&mdsc->mutex);
2876 want_tid = mdsc->last_tid;
2877 want_flush = mdsc->cap_flush_seq;
2878 mutex_unlock(&mdsc->mutex);
2879 dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2881 ceph_flush_dirty_caps(mdsc);
2883 wait_unsafe_requests(mdsc, want_tid);
2884 wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2889 * called after sb is ro.
2891 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2893 struct ceph_mds_session *session;
2896 struct ceph_client *client = mdsc->client;
2897 unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2899 dout("close_sessions\n");
2901 mutex_lock(&mdsc->mutex);
2903 /* close sessions */
2905 while (time_before(jiffies, started + timeout)) {
2906 dout("closing sessions\n");
2908 for (i = 0; i < mdsc->max_sessions; i++) {
2909 session = __ceph_lookup_mds_session(mdsc, i);
2912 mutex_unlock(&mdsc->mutex);
2913 mutex_lock(&session->s_mutex);
2914 __close_session(mdsc, session);
2915 mutex_unlock(&session->s_mutex);
2916 ceph_put_mds_session(session);
2917 mutex_lock(&mdsc->mutex);
2923 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2926 dout("waiting for sessions to close\n");
2927 mutex_unlock(&mdsc->mutex);
2928 wait_for_completion_timeout(&mdsc->session_close_waiters,
2930 mutex_lock(&mdsc->mutex);
2933 /* tear down remaining sessions */
2934 for (i = 0; i < mdsc->max_sessions; i++) {
2935 if (mdsc->sessions[i]) {
2936 session = get_session(mdsc->sessions[i]);
2937 __unregister_session(mdsc, session);
2938 mutex_unlock(&mdsc->mutex);
2939 mutex_lock(&session->s_mutex);
2940 remove_session_caps(session);
2941 mutex_unlock(&session->s_mutex);
2942 ceph_put_mds_session(session);
2943 mutex_lock(&mdsc->mutex);
2947 WARN_ON(!list_empty(&mdsc->cap_delay_list));
2949 mutex_unlock(&mdsc->mutex);
2951 ceph_cleanup_empty_realms(mdsc);
2953 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2958 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2961 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2963 ceph_mdsmap_destroy(mdsc->mdsmap);
2964 kfree(mdsc->sessions);
2969 * handle mds map update.
2971 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2975 void *p = msg->front.iov_base;
2976 void *end = p + msg->front.iov_len;
2977 struct ceph_mdsmap *newmap, *oldmap;
2978 struct ceph_fsid fsid;
2981 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2982 ceph_decode_copy(&p, &fsid, sizeof(fsid));
2983 if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2985 epoch = ceph_decode_32(&p);
2986 maplen = ceph_decode_32(&p);
2987 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2989 /* do we need it? */
2990 ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2991 mutex_lock(&mdsc->mutex);
2992 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2993 dout("handle_map epoch %u <= our %u\n",
2994 epoch, mdsc->mdsmap->m_epoch);
2995 mutex_unlock(&mdsc->mutex);
2999 newmap = ceph_mdsmap_decode(&p, end);
3000 if (IS_ERR(newmap)) {
3001 err = PTR_ERR(newmap);
3005 /* swap into place */
3007 oldmap = mdsc->mdsmap;
3008 mdsc->mdsmap = newmap;
3009 check_new_map(mdsc, newmap, oldmap);
3010 ceph_mdsmap_destroy(oldmap);
3012 mdsc->mdsmap = newmap; /* first mds map */
3014 mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3016 __wake_requests(mdsc, &mdsc->waiting_for_map);
3018 mutex_unlock(&mdsc->mutex);
3019 schedule_delayed(mdsc);
3023 mutex_unlock(&mdsc->mutex);
3025 pr_err("error decoding mdsmap %d\n", err);
3029 static struct ceph_connection *con_get(struct ceph_connection *con)
3031 struct ceph_mds_session *s = con->private;
3033 if (get_session(s)) {
3034 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3037 dout("mdsc con_get %p FAIL\n", s);
3041 static void con_put(struct ceph_connection *con)
3043 struct ceph_mds_session *s = con->private;
3045 ceph_put_mds_session(s);
3046 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
3050 * if the client is unresponsive for long enough, the mds will kill
3051 * the session entirely.
3053 static void peer_reset(struct ceph_connection *con)
3055 struct ceph_mds_session *s = con->private;
3056 struct ceph_mds_client *mdsc = s->s_mdsc;
3058 pr_warning("mds%d closed our session\n", s->s_mds);
3059 send_mds_reconnect(mdsc, s);
3062 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3064 struct ceph_mds_session *s = con->private;
3065 struct ceph_mds_client *mdsc = s->s_mdsc;
3066 int type = le16_to_cpu(msg->hdr.type);
3068 mutex_lock(&mdsc->mutex);
3069 if (__verify_registered_session(mdsc, s) < 0) {
3070 mutex_unlock(&mdsc->mutex);
3073 mutex_unlock(&mdsc->mutex);
3076 case CEPH_MSG_MDS_MAP:
3077 ceph_mdsc_handle_map(mdsc, msg);
3079 case CEPH_MSG_CLIENT_SESSION:
3080 handle_session(s, msg);
3082 case CEPH_MSG_CLIENT_REPLY:
3083 handle_reply(s, msg);
3085 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3086 handle_forward(mdsc, s, msg);
3088 case CEPH_MSG_CLIENT_CAPS:
3089 ceph_handle_caps(s, msg);
3091 case CEPH_MSG_CLIENT_SNAP:
3092 ceph_handle_snap(mdsc, s, msg);
3094 case CEPH_MSG_CLIENT_LEASE:
3095 handle_lease(mdsc, s, msg);
3099 pr_err("received unknown message type %d %s\n", type,
3100 ceph_msg_type_name(type));
3109 static int get_authorizer(struct ceph_connection *con,
3110 void **buf, int *len, int *proto,
3111 void **reply_buf, int *reply_len, int force_new)
3113 struct ceph_mds_session *s = con->private;
3114 struct ceph_mds_client *mdsc = s->s_mdsc;
3115 struct ceph_auth_client *ac = mdsc->client->monc.auth;
3118 if (force_new && s->s_authorizer) {
3119 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3120 s->s_authorizer = NULL;
3122 if (s->s_authorizer == NULL) {
3123 if (ac->ops->create_authorizer) {
3124 ret = ac->ops->create_authorizer(
3125 ac, CEPH_ENTITY_TYPE_MDS,
3127 &s->s_authorizer_buf,
3128 &s->s_authorizer_buf_len,
3129 &s->s_authorizer_reply_buf,
3130 &s->s_authorizer_reply_buf_len);
3136 *proto = ac->protocol;
3137 *buf = s->s_authorizer_buf;
3138 *len = s->s_authorizer_buf_len;
3139 *reply_buf = s->s_authorizer_reply_buf;
3140 *reply_len = s->s_authorizer_reply_buf_len;
3145 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3147 struct ceph_mds_session *s = con->private;
3148 struct ceph_mds_client *mdsc = s->s_mdsc;
3149 struct ceph_auth_client *ac = mdsc->client->monc.auth;
3151 return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3154 static int invalidate_authorizer(struct ceph_connection *con)
3156 struct ceph_mds_session *s = con->private;
3157 struct ceph_mds_client *mdsc = s->s_mdsc;
3158 struct ceph_auth_client *ac = mdsc->client->monc.auth;
3160 if (ac->ops->invalidate_authorizer)
3161 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3163 return ceph_monc_validate_auth(&mdsc->client->monc);
3166 static const struct ceph_connection_operations mds_con_ops = {
3169 .dispatch = dispatch,
3170 .get_authorizer = get_authorizer,
3171 .verify_authorizer_reply = verify_authorizer_reply,
3172 .invalidate_authorizer = invalidate_authorizer,
3173 .peer_reset = peer_reset,