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37 * are no longer valid.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(&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,
1073 struct ceph_msg *msg;
1074 struct ceph_mds_cap_release *head;
1078 extra = mdsc->client->mount_args->cap_release_safety;
1080 spin_lock(&session->s_cap_lock);
1082 if (!list_empty(&session->s_cap_releases)) {
1083 msg = list_first_entry(&session->s_cap_releases,
1086 head = msg->front.iov_base;
1087 extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1090 while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1091 spin_unlock(&session->s_cap_lock);
1092 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1096 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1097 (int)msg->front.iov_len);
1098 head = msg->front.iov_base;
1099 head->num = cpu_to_le32(0);
1100 msg->front.iov_len = sizeof(*head);
1101 spin_lock(&session->s_cap_lock);
1102 list_add(&msg->list_head, &session->s_cap_releases);
1103 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1106 if (!list_empty(&session->s_cap_releases)) {
1107 msg = list_first_entry(&session->s_cap_releases,
1110 head = msg->front.iov_base;
1112 dout(" queueing non-full %p (%d)\n", msg,
1113 le32_to_cpu(head->num));
1114 list_move_tail(&msg->list_head,
1115 &session->s_cap_releases_done);
1116 session->s_num_cap_releases -=
1117 CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1121 spin_unlock(&session->s_cap_lock);
1127 * flush all dirty inode data to disk.
1129 * returns true if we've flushed through want_flush_seq
1131 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1135 dout("check_cap_flush want %lld\n", want_flush_seq);
1136 mutex_lock(&mdsc->mutex);
1137 for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1138 struct ceph_mds_session *session = mdsc->sessions[mds];
1142 get_session(session);
1143 mutex_unlock(&mdsc->mutex);
1145 mutex_lock(&session->s_mutex);
1146 if (!list_empty(&session->s_cap_flushing)) {
1147 struct ceph_inode_info *ci =
1148 list_entry(session->s_cap_flushing.next,
1149 struct ceph_inode_info,
1151 struct inode *inode = &ci->vfs_inode;
1153 spin_lock(&inode->i_lock);
1154 if (ci->i_cap_flush_seq <= want_flush_seq) {
1155 dout("check_cap_flush still flushing %p "
1156 "seq %lld <= %lld to mds%d\n", inode,
1157 ci->i_cap_flush_seq, want_flush_seq,
1161 spin_unlock(&inode->i_lock);
1163 mutex_unlock(&session->s_mutex);
1164 ceph_put_mds_session(session);
1168 mutex_lock(&mdsc->mutex);
1171 mutex_unlock(&mdsc->mutex);
1172 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1177 * called under s_mutex
1179 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1180 struct ceph_mds_session *session)
1182 struct ceph_msg *msg;
1184 dout("send_cap_releases mds%d\n", session->s_mds);
1185 spin_lock(&session->s_cap_lock);
1186 while (!list_empty(&session->s_cap_releases_done)) {
1187 msg = list_first_entry(&session->s_cap_releases_done,
1188 struct ceph_msg, list_head);
1189 list_del_init(&msg->list_head);
1190 spin_unlock(&session->s_cap_lock);
1191 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1192 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1193 ceph_con_send(&session->s_con, msg);
1194 spin_lock(&session->s_cap_lock);
1196 spin_unlock(&session->s_cap_lock);
1199 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1200 struct ceph_mds_session *session)
1202 struct ceph_msg *msg;
1203 struct ceph_mds_cap_release *head;
1206 dout("discard_cap_releases mds%d\n", session->s_mds);
1207 spin_lock(&session->s_cap_lock);
1209 /* zero out the in-progress message */
1210 msg = list_first_entry(&session->s_cap_releases,
1211 struct ceph_msg, list_head);
1212 head = msg->front.iov_base;
1213 num = le32_to_cpu(head->num);
1214 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1215 head->num = cpu_to_le32(0);
1216 session->s_num_cap_releases += num;
1218 /* requeue completed messages */
1219 while (!list_empty(&session->s_cap_releases_done)) {
1220 msg = list_first_entry(&session->s_cap_releases_done,
1221 struct ceph_msg, list_head);
1222 list_del_init(&msg->list_head);
1224 head = msg->front.iov_base;
1225 num = le32_to_cpu(head->num);
1226 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1228 session->s_num_cap_releases += num;
1229 head->num = cpu_to_le32(0);
1230 msg->front.iov_len = sizeof(*head);
1231 list_add(&msg->list_head, &session->s_cap_releases);
1234 spin_unlock(&session->s_cap_lock);
1242 * Create an mds request.
1244 struct ceph_mds_request *
1245 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1247 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1250 return ERR_PTR(-ENOMEM);
1252 mutex_init(&req->r_fill_mutex);
1253 req->r_started = jiffies;
1254 req->r_resend_mds = -1;
1255 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1257 kref_init(&req->r_kref);
1258 INIT_LIST_HEAD(&req->r_wait);
1259 init_completion(&req->r_completion);
1260 init_completion(&req->r_safe_completion);
1261 INIT_LIST_HEAD(&req->r_unsafe_item);
1264 req->r_direct_mode = mode;
1269 * return oldest (lowest) request, tid in request tree, 0 if none.
1271 * called under mdsc->mutex.
1273 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1275 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1277 return rb_entry(rb_first(&mdsc->request_tree),
1278 struct ceph_mds_request, r_node);
1281 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1283 struct ceph_mds_request *req = __get_oldest_req(mdsc);
1291 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1292 * on build_path_from_dentry in fs/cifs/dir.c.
1294 * If @stop_on_nosnap, generate path relative to the first non-snapped
1297 * Encode hidden .snap dirs as a double /, i.e.
1298 * foo/.snap/bar -> foo//bar
1300 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1303 struct dentry *temp;
1308 return ERR_PTR(-EINVAL);
1312 for (temp = dentry; !IS_ROOT(temp);) {
1313 struct inode *inode = temp->d_inode;
1314 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1315 len++; /* slash only */
1316 else if (stop_on_nosnap && inode &&
1317 ceph_snap(inode) == CEPH_NOSNAP)
1320 len += 1 + temp->d_name.len;
1321 temp = temp->d_parent;
1323 pr_err("build_path corrupt dentry %p\n", dentry);
1324 return ERR_PTR(-EINVAL);
1328 len--; /* no leading '/' */
1330 path = kmalloc(len+1, GFP_NOFS);
1332 return ERR_PTR(-ENOMEM);
1334 path[pos] = 0; /* trailing null */
1335 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1336 struct inode *inode = temp->d_inode;
1338 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1339 dout("build_path path+%d: %p SNAPDIR\n",
1341 } else if (stop_on_nosnap && inode &&
1342 ceph_snap(inode) == CEPH_NOSNAP) {
1345 pos -= temp->d_name.len;
1348 strncpy(path + pos, temp->d_name.name,
1353 temp = temp->d_parent;
1355 pr_err("build_path corrupt dentry\n");
1357 return ERR_PTR(-EINVAL);
1361 pr_err("build_path did not end path lookup where "
1362 "expected, namelen is %d, pos is %d\n", len, pos);
1363 /* presumably this is only possible if racing with a
1364 rename of one of the parent directories (we can not
1365 lock the dentries above us to prevent this, but
1366 retrying should be harmless) */
1371 *base = ceph_ino(temp->d_inode);
1373 dout("build_path on %p %d built %llx '%.*s'\n",
1374 dentry, atomic_read(&dentry->d_count), *base, len, path);
1378 static int build_dentry_path(struct dentry *dentry,
1379 const char **ppath, int *ppathlen, u64 *pino,
1384 if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1385 *pino = ceph_ino(dentry->d_parent->d_inode);
1386 *ppath = dentry->d_name.name;
1387 *ppathlen = dentry->d_name.len;
1390 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1392 return PTR_ERR(path);
1398 static int build_inode_path(struct inode *inode,
1399 const char **ppath, int *ppathlen, u64 *pino,
1402 struct dentry *dentry;
1405 if (ceph_snap(inode) == CEPH_NOSNAP) {
1406 *pino = ceph_ino(inode);
1410 dentry = d_find_alias(inode);
1411 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1414 return PTR_ERR(path);
1421 * request arguments may be specified via an inode *, a dentry *, or
1422 * an explicit ino+path.
1424 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1425 const char *rpath, u64 rino,
1426 const char **ppath, int *pathlen,
1427 u64 *ino, int *freepath)
1432 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1433 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1435 } else if (rdentry) {
1436 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1437 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1442 *pathlen = strlen(rpath);
1443 dout(" path %.*s\n", *pathlen, rpath);
1450 * called under mdsc->mutex
1452 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1453 struct ceph_mds_request *req,
1456 struct ceph_msg *msg;
1457 struct ceph_mds_request_head *head;
1458 const char *path1 = NULL;
1459 const char *path2 = NULL;
1460 u64 ino1 = 0, ino2 = 0;
1461 int pathlen1 = 0, pathlen2 = 0;
1462 int freepath1 = 0, freepath2 = 0;
1468 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1469 req->r_path1, req->r_ino1.ino,
1470 &path1, &pathlen1, &ino1, &freepath1);
1476 ret = set_request_path_attr(NULL, req->r_old_dentry,
1477 req->r_path2, req->r_ino2.ino,
1478 &path2, &pathlen2, &ino2, &freepath2);
1484 len = sizeof(*head) +
1485 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1487 /* calculate (max) length for cap releases */
1488 len += sizeof(struct ceph_mds_request_release) *
1489 (!!req->r_inode_drop + !!req->r_dentry_drop +
1490 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1491 if (req->r_dentry_drop)
1492 len += req->r_dentry->d_name.len;
1493 if (req->r_old_dentry_drop)
1494 len += req->r_old_dentry->d_name.len;
1496 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS);
1498 msg = ERR_PTR(-ENOMEM);
1502 msg->hdr.tid = cpu_to_le64(req->r_tid);
1504 head = msg->front.iov_base;
1505 p = msg->front.iov_base + sizeof(*head);
1506 end = msg->front.iov_base + msg->front.iov_len;
1508 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1509 head->op = cpu_to_le32(req->r_op);
1510 head->caller_uid = cpu_to_le32(current_fsuid());
1511 head->caller_gid = cpu_to_le32(current_fsgid());
1512 head->args = req->r_args;
1514 ceph_encode_filepath(&p, end, ino1, path1);
1515 ceph_encode_filepath(&p, end, ino2, path2);
1519 if (req->r_inode_drop)
1520 releases += ceph_encode_inode_release(&p,
1521 req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1522 mds, req->r_inode_drop, req->r_inode_unless, 0);
1523 if (req->r_dentry_drop)
1524 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1525 mds, req->r_dentry_drop, req->r_dentry_unless);
1526 if (req->r_old_dentry_drop)
1527 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1528 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1529 if (req->r_old_inode_drop)
1530 releases += ceph_encode_inode_release(&p,
1531 req->r_old_dentry->d_inode,
1532 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1533 head->num_releases = cpu_to_le16(releases);
1536 msg->front.iov_len = p - msg->front.iov_base;
1537 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1539 msg->pages = req->r_pages;
1540 msg->nr_pages = req->r_num_pages;
1541 msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1542 msg->hdr.data_off = cpu_to_le16(0);
1546 kfree((char *)path2);
1549 kfree((char *)path1);
1555 * called under mdsc->mutex if error, under no mutex if
1558 static void complete_request(struct ceph_mds_client *mdsc,
1559 struct ceph_mds_request *req)
1561 if (req->r_callback)
1562 req->r_callback(mdsc, req);
1564 complete(&req->r_completion);
1568 * called under mdsc->mutex
1570 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1571 struct ceph_mds_request *req,
1574 struct ceph_mds_request_head *rhead;
1575 struct ceph_msg *msg;
1580 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1581 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1583 if (req->r_request) {
1584 ceph_msg_put(req->r_request);
1585 req->r_request = NULL;
1587 msg = create_request_message(mdsc, req, mds);
1589 req->r_err = PTR_ERR(msg);
1590 complete_request(mdsc, req);
1591 return PTR_ERR(msg);
1593 req->r_request = msg;
1595 rhead = msg->front.iov_base;
1596 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1597 if (req->r_got_unsafe)
1598 flags |= CEPH_MDS_FLAG_REPLAY;
1599 if (req->r_locked_dir)
1600 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1601 rhead->flags = cpu_to_le32(flags);
1602 rhead->num_fwd = req->r_num_fwd;
1603 rhead->num_retry = req->r_attempts - 1;
1605 dout(" r_locked_dir = %p\n", req->r_locked_dir);
1607 if (req->r_target_inode && req->r_got_unsafe)
1608 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1615 * send request, or put it on the appropriate wait list.
1617 static int __do_request(struct ceph_mds_client *mdsc,
1618 struct ceph_mds_request *req)
1620 struct ceph_mds_session *session = NULL;
1624 if (req->r_err || req->r_got_result)
1627 if (req->r_timeout &&
1628 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1629 dout("do_request timed out\n");
1634 mds = __choose_mds(mdsc, req);
1636 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1637 dout("do_request no mds or not active, waiting for map\n");
1638 list_add(&req->r_wait, &mdsc->waiting_for_map);
1642 /* get, open session */
1643 session = __ceph_lookup_mds_session(mdsc, mds);
1645 session = register_session(mdsc, mds);
1646 if (IS_ERR(session)) {
1647 err = PTR_ERR(session);
1651 dout("do_request mds%d session %p state %s\n", mds, session,
1652 session_state_name(session->s_state));
1653 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1654 session->s_state != CEPH_MDS_SESSION_HUNG) {
1655 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1656 session->s_state == CEPH_MDS_SESSION_CLOSING)
1657 __open_session(mdsc, session);
1658 list_add(&req->r_wait, &session->s_waiting);
1663 req->r_session = get_session(session);
1664 req->r_resend_mds = -1; /* forget any previous mds hint */
1666 if (req->r_request_started == 0) /* note request start time */
1667 req->r_request_started = jiffies;
1669 err = __prepare_send_request(mdsc, req, mds);
1671 ceph_msg_get(req->r_request);
1672 ceph_con_send(&session->s_con, req->r_request);
1676 ceph_put_mds_session(session);
1682 complete_request(mdsc, req);
1687 * called under mdsc->mutex
1689 static void __wake_requests(struct ceph_mds_client *mdsc,
1690 struct list_head *head)
1692 struct ceph_mds_request *req, *nreq;
1694 list_for_each_entry_safe(req, nreq, head, r_wait) {
1695 list_del_init(&req->r_wait);
1696 __do_request(mdsc, req);
1701 * Wake up threads with requests pending for @mds, so that they can
1702 * resubmit their requests to a possibly different mds.
1704 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1706 struct ceph_mds_request *req;
1709 dout("kick_requests mds%d\n", mds);
1710 for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1711 req = rb_entry(p, struct ceph_mds_request, r_node);
1712 if (req->r_got_unsafe)
1714 if (req->r_session &&
1715 req->r_session->s_mds == mds) {
1716 dout(" kicking tid %llu\n", req->r_tid);
1717 put_request_session(req);
1718 __do_request(mdsc, req);
1723 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1724 struct ceph_mds_request *req)
1726 dout("submit_request on %p\n", req);
1727 mutex_lock(&mdsc->mutex);
1728 __register_request(mdsc, req, NULL);
1729 __do_request(mdsc, req);
1730 mutex_unlock(&mdsc->mutex);
1734 * Synchrously perform an mds request. Take care of all of the
1735 * session setup, forwarding, retry details.
1737 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1739 struct ceph_mds_request *req)
1743 dout("do_request on %p\n", req);
1745 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1747 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1748 if (req->r_locked_dir)
1749 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1750 if (req->r_old_dentry)
1752 ceph_inode(req->r_old_dentry->d_parent->d_inode),
1756 mutex_lock(&mdsc->mutex);
1757 __register_request(mdsc, req, dir);
1758 __do_request(mdsc, req);
1762 __unregister_request(mdsc, req);
1763 dout("do_request early error %d\n", err);
1768 mutex_unlock(&mdsc->mutex);
1769 dout("do_request waiting\n");
1770 if (req->r_timeout) {
1771 err = (long)wait_for_completion_killable_timeout(
1772 &req->r_completion, req->r_timeout);
1776 err = wait_for_completion_killable(&req->r_completion);
1778 dout("do_request waited, got %d\n", err);
1779 mutex_lock(&mdsc->mutex);
1781 /* only abort if we didn't race with a real reply */
1782 if (req->r_got_result) {
1783 err = le32_to_cpu(req->r_reply_info.head->result);
1784 } else if (err < 0) {
1785 dout("aborted request %lld with %d\n", req->r_tid, err);
1788 * ensure we aren't running concurrently with
1789 * ceph_fill_trace or ceph_readdir_prepopulate, which
1790 * rely on locks (dir mutex) held by our caller.
1792 mutex_lock(&req->r_fill_mutex);
1794 req->r_aborted = true;
1795 mutex_unlock(&req->r_fill_mutex);
1797 if (req->r_locked_dir &&
1798 (req->r_op & CEPH_MDS_OP_WRITE))
1799 ceph_invalidate_dir_request(req);
1805 mutex_unlock(&mdsc->mutex);
1806 dout("do_request %p done, result %d\n", req, err);
1811 * Invalidate dir I_COMPLETE, dentry lease state on an aborted MDS
1812 * namespace request.
1814 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
1816 struct inode *inode = req->r_locked_dir;
1817 struct ceph_inode_info *ci = ceph_inode(inode);
1819 dout("invalidate_dir_request %p (I_COMPLETE, lease(s))\n", inode);
1820 spin_lock(&inode->i_lock);
1821 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1822 ci->i_release_count++;
1823 spin_unlock(&inode->i_lock);
1826 ceph_invalidate_dentry_lease(req->r_dentry);
1827 if (req->r_old_dentry)
1828 ceph_invalidate_dentry_lease(req->r_old_dentry);
1834 * We take the session mutex and parse and process the reply immediately.
1835 * This preserves the logical ordering of replies, capabilities, etc., sent
1836 * by the MDS as they are applied to our local cache.
1838 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1840 struct ceph_mds_client *mdsc = session->s_mdsc;
1841 struct ceph_mds_request *req;
1842 struct ceph_mds_reply_head *head = msg->front.iov_base;
1843 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
1846 int mds = session->s_mds;
1848 if (msg->front.iov_len < sizeof(*head)) {
1849 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1854 /* get request, session */
1855 tid = le64_to_cpu(msg->hdr.tid);
1856 mutex_lock(&mdsc->mutex);
1857 req = __lookup_request(mdsc, tid);
1859 dout("handle_reply on unknown tid %llu\n", tid);
1860 mutex_unlock(&mdsc->mutex);
1863 dout("handle_reply %p\n", req);
1865 /* correct session? */
1866 if (req->r_session != session) {
1867 pr_err("mdsc_handle_reply got %llu on session mds%d"
1868 " not mds%d\n", tid, session->s_mds,
1869 req->r_session ? req->r_session->s_mds : -1);
1870 mutex_unlock(&mdsc->mutex);
1875 if ((req->r_got_unsafe && !head->safe) ||
1876 (req->r_got_safe && head->safe)) {
1877 pr_warning("got a dup %s reply on %llu from mds%d\n",
1878 head->safe ? "safe" : "unsafe", tid, mds);
1879 mutex_unlock(&mdsc->mutex);
1882 if (req->r_got_safe && !head->safe) {
1883 pr_warning("got unsafe after safe on %llu from mds%d\n",
1885 mutex_unlock(&mdsc->mutex);
1889 result = le32_to_cpu(head->result);
1892 * Tolerate 2 consecutive ESTALEs from the same mds.
1893 * FIXME: we should be looking at the cap migrate_seq.
1895 if (result == -ESTALE) {
1896 req->r_direct_mode = USE_AUTH_MDS;
1898 if (req->r_num_stale <= 2) {
1899 __do_request(mdsc, req);
1900 mutex_unlock(&mdsc->mutex);
1904 req->r_num_stale = 0;
1908 req->r_got_safe = true;
1909 __unregister_request(mdsc, req);
1910 complete(&req->r_safe_completion);
1912 if (req->r_got_unsafe) {
1914 * We already handled the unsafe response, now do the
1915 * cleanup. No need to examine the response; the MDS
1916 * doesn't include any result info in the safe
1917 * response. And even if it did, there is nothing
1918 * useful we could do with a revised return value.
1920 dout("got safe reply %llu, mds%d\n", tid, mds);
1921 list_del_init(&req->r_unsafe_item);
1923 /* last unsafe request during umount? */
1924 if (mdsc->stopping && !__get_oldest_req(mdsc))
1925 complete(&mdsc->safe_umount_waiters);
1926 mutex_unlock(&mdsc->mutex);
1930 req->r_got_unsafe = true;
1931 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1934 dout("handle_reply tid %lld result %d\n", tid, result);
1935 rinfo = &req->r_reply_info;
1936 err = parse_reply_info(msg, rinfo);
1937 mutex_unlock(&mdsc->mutex);
1939 mutex_lock(&session->s_mutex);
1941 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1947 if (rinfo->snapblob_len) {
1948 down_write(&mdsc->snap_rwsem);
1949 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1950 rinfo->snapblob + rinfo->snapblob_len,
1951 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1952 downgrade_write(&mdsc->snap_rwsem);
1954 down_read(&mdsc->snap_rwsem);
1957 /* insert trace into our cache */
1958 mutex_lock(&req->r_fill_mutex);
1959 err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1961 if (result == 0 && rinfo->dir_nr)
1962 ceph_readdir_prepopulate(req, req->r_session);
1963 ceph_unreserve_caps(&req->r_caps_reservation);
1965 mutex_unlock(&req->r_fill_mutex);
1967 up_read(&mdsc->snap_rwsem);
1969 mutex_lock(&mdsc->mutex);
1970 if (!req->r_aborted) {
1976 req->r_got_result = true;
1979 dout("reply arrived after request %lld was aborted\n", tid);
1981 mutex_unlock(&mdsc->mutex);
1983 ceph_add_cap_releases(mdsc, req->r_session, -1);
1984 mutex_unlock(&session->s_mutex);
1986 /* kick calling process */
1987 complete_request(mdsc, req);
1989 ceph_mdsc_put_request(req);
1996 * handle mds notification that our request has been forwarded.
1998 static void handle_forward(struct ceph_mds_client *mdsc,
1999 struct ceph_mds_session *session,
2000 struct ceph_msg *msg)
2002 struct ceph_mds_request *req;
2003 u64 tid = le64_to_cpu(msg->hdr.tid);
2007 void *p = msg->front.iov_base;
2008 void *end = p + msg->front.iov_len;
2010 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2011 next_mds = ceph_decode_32(&p);
2012 fwd_seq = ceph_decode_32(&p);
2014 mutex_lock(&mdsc->mutex);
2015 req = __lookup_request(mdsc, tid);
2017 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2018 goto out; /* dup reply? */
2021 if (req->r_aborted) {
2022 dout("forward tid %llu aborted, unregistering\n", tid);
2023 __unregister_request(mdsc, req);
2024 } else if (fwd_seq <= req->r_num_fwd) {
2025 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2026 tid, next_mds, req->r_num_fwd, fwd_seq);
2028 /* resend. forward race not possible; mds would drop */
2029 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2031 BUG_ON(req->r_got_result);
2032 req->r_num_fwd = fwd_seq;
2033 req->r_resend_mds = next_mds;
2034 put_request_session(req);
2035 __do_request(mdsc, req);
2037 ceph_mdsc_put_request(req);
2039 mutex_unlock(&mdsc->mutex);
2043 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2047 * handle a mds session control message
2049 static void handle_session(struct ceph_mds_session *session,
2050 struct ceph_msg *msg)
2052 struct ceph_mds_client *mdsc = session->s_mdsc;
2055 int mds = session->s_mds;
2056 struct ceph_mds_session_head *h = msg->front.iov_base;
2060 if (msg->front.iov_len != sizeof(*h))
2062 op = le32_to_cpu(h->op);
2063 seq = le64_to_cpu(h->seq);
2065 mutex_lock(&mdsc->mutex);
2066 if (op == CEPH_SESSION_CLOSE)
2067 __unregister_session(mdsc, session);
2068 /* FIXME: this ttl calculation is generous */
2069 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2070 mutex_unlock(&mdsc->mutex);
2072 mutex_lock(&session->s_mutex);
2074 dout("handle_session mds%d %s %p state %s seq %llu\n",
2075 mds, ceph_session_op_name(op), session,
2076 session_state_name(session->s_state), seq);
2078 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2079 session->s_state = CEPH_MDS_SESSION_OPEN;
2080 pr_info("mds%d came back\n", session->s_mds);
2084 case CEPH_SESSION_OPEN:
2085 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2086 pr_info("mds%d reconnect success\n", session->s_mds);
2087 session->s_state = CEPH_MDS_SESSION_OPEN;
2088 renewed_caps(mdsc, session, 0);
2091 __close_session(mdsc, session);
2094 case CEPH_SESSION_RENEWCAPS:
2095 if (session->s_renew_seq == seq)
2096 renewed_caps(mdsc, session, 1);
2099 case CEPH_SESSION_CLOSE:
2100 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2101 pr_info("mds%d reconnect denied\n", session->s_mds);
2102 remove_session_caps(session);
2103 wake = 1; /* for good measure */
2104 complete(&mdsc->session_close_waiters);
2105 kick_requests(mdsc, mds);
2108 case CEPH_SESSION_STALE:
2109 pr_info("mds%d caps went stale, renewing\n",
2111 spin_lock(&session->s_cap_lock);
2112 session->s_cap_gen++;
2113 session->s_cap_ttl = 0;
2114 spin_unlock(&session->s_cap_lock);
2115 send_renew_caps(mdsc, session);
2118 case CEPH_SESSION_RECALL_STATE:
2119 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2123 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2127 mutex_unlock(&session->s_mutex);
2129 mutex_lock(&mdsc->mutex);
2130 __wake_requests(mdsc, &session->s_waiting);
2131 mutex_unlock(&mdsc->mutex);
2136 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2137 (int)msg->front.iov_len);
2144 * called under session->mutex.
2146 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2147 struct ceph_mds_session *session)
2149 struct ceph_mds_request *req, *nreq;
2152 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2154 mutex_lock(&mdsc->mutex);
2155 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2156 err = __prepare_send_request(mdsc, req, session->s_mds);
2158 ceph_msg_get(req->r_request);
2159 ceph_con_send(&session->s_con, req->r_request);
2162 mutex_unlock(&mdsc->mutex);
2166 * Encode information about a cap for a reconnect with the MDS.
2168 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2171 struct ceph_mds_cap_reconnect rec;
2172 struct ceph_inode_info *ci;
2173 struct ceph_pagelist *pagelist = arg;
2177 struct dentry *dentry;
2181 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2182 inode, ceph_vinop(inode), cap, cap->cap_id,
2183 ceph_cap_string(cap->issued));
2184 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2188 dentry = d_find_alias(inode);
2190 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2192 err = PTR_ERR(path);
2199 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2203 spin_lock(&inode->i_lock);
2204 cap->seq = 0; /* reset cap seq */
2205 cap->issue_seq = 0; /* and issue_seq */
2206 rec.cap_id = cpu_to_le64(cap->cap_id);
2207 rec.pathbase = cpu_to_le64(pathbase);
2208 rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2209 rec.issued = cpu_to_le32(cap->issued);
2210 rec.size = cpu_to_le64(inode->i_size);
2211 ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2212 ceph_encode_timespec(&rec.atime, &inode->i_atime);
2213 rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2214 spin_unlock(&inode->i_lock);
2216 err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2226 * If an MDS fails and recovers, clients need to reconnect in order to
2227 * reestablish shared state. This includes all caps issued through
2228 * this session _and_ the snap_realm hierarchy. Because it's not
2229 * clear which snap realms the mds cares about, we send everything we
2230 * know about.. that ensures we'll then get any new info the
2231 * recovering MDS might have.
2233 * This is a relatively heavyweight operation, but it's rare.
2235 * called with mdsc->mutex held.
2237 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2238 struct ceph_mds_session *session)
2240 struct ceph_msg *reply;
2242 int mds = session->s_mds;
2244 struct ceph_pagelist *pagelist;
2246 pr_info("mds%d reconnect start\n", mds);
2248 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2250 goto fail_nopagelist;
2251 ceph_pagelist_init(pagelist);
2253 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS);
2257 mutex_lock(&session->s_mutex);
2258 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2261 ceph_con_open(&session->s_con,
2262 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2264 /* replay unsafe requests */
2265 replay_unsafe_requests(mdsc, session);
2267 down_read(&mdsc->snap_rwsem);
2269 dout("session %p state %s\n", session,
2270 session_state_name(session->s_state));
2272 /* drop old cap expires; we're about to reestablish that state */
2273 discard_cap_releases(mdsc, session);
2275 /* traverse this session's caps */
2276 err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2279 err = iterate_session_caps(session, encode_caps_cb, pagelist);
2284 * snaprealms. we provide mds with the ino, seq (version), and
2285 * parent for all of our realms. If the mds has any newer info,
2288 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2289 struct ceph_snap_realm *realm =
2290 rb_entry(p, struct ceph_snap_realm, node);
2291 struct ceph_mds_snaprealm_reconnect sr_rec;
2293 dout(" adding snap realm %llx seq %lld parent %llx\n",
2294 realm->ino, realm->seq, realm->parent_ino);
2295 sr_rec.ino = cpu_to_le64(realm->ino);
2296 sr_rec.seq = cpu_to_le64(realm->seq);
2297 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2298 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2303 reply->pagelist = pagelist;
2304 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2305 reply->nr_pages = calc_pages_for(0, pagelist->length);
2306 ceph_con_send(&session->s_con, reply);
2308 mutex_unlock(&session->s_mutex);
2310 mutex_lock(&mdsc->mutex);
2311 __wake_requests(mdsc, &session->s_waiting);
2312 mutex_unlock(&mdsc->mutex);
2314 up_read(&mdsc->snap_rwsem);
2318 ceph_msg_put(reply);
2319 up_read(&mdsc->snap_rwsem);
2320 mutex_unlock(&session->s_mutex);
2322 ceph_pagelist_release(pagelist);
2325 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2331 * compare old and new mdsmaps, kicking requests
2332 * and closing out old connections as necessary
2334 * called under mdsc->mutex.
2336 static void check_new_map(struct ceph_mds_client *mdsc,
2337 struct ceph_mdsmap *newmap,
2338 struct ceph_mdsmap *oldmap)
2341 int oldstate, newstate;
2342 struct ceph_mds_session *s;
2344 dout("check_new_map new %u old %u\n",
2345 newmap->m_epoch, oldmap->m_epoch);
2347 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2348 if (mdsc->sessions[i] == NULL)
2350 s = mdsc->sessions[i];
2351 oldstate = ceph_mdsmap_get_state(oldmap, i);
2352 newstate = ceph_mdsmap_get_state(newmap, i);
2354 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2355 i, ceph_mds_state_name(oldstate),
2356 ceph_mds_state_name(newstate),
2357 session_state_name(s->s_state));
2359 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2360 ceph_mdsmap_get_addr(newmap, i),
2361 sizeof(struct ceph_entity_addr))) {
2362 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2363 /* the session never opened, just close it
2365 __wake_requests(mdsc, &s->s_waiting);
2366 __unregister_session(mdsc, s);
2369 mutex_unlock(&mdsc->mutex);
2370 mutex_lock(&s->s_mutex);
2371 mutex_lock(&mdsc->mutex);
2372 ceph_con_close(&s->s_con);
2373 mutex_unlock(&s->s_mutex);
2374 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2377 /* kick any requests waiting on the recovering mds */
2378 kick_requests(mdsc, i);
2379 } else if (oldstate == newstate) {
2380 continue; /* nothing new with this mds */
2386 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2387 newstate >= CEPH_MDS_STATE_RECONNECT) {
2388 mutex_unlock(&mdsc->mutex);
2389 send_mds_reconnect(mdsc, s);
2390 mutex_lock(&mdsc->mutex);
2394 * kick request on any mds that has gone active.
2396 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2397 newstate >= CEPH_MDS_STATE_ACTIVE) {
2398 if (oldstate != CEPH_MDS_STATE_CREATING &&
2399 oldstate != CEPH_MDS_STATE_STARTING)
2400 pr_info("mds%d recovery completed\n", s->s_mds);
2401 kick_requests(mdsc, i);
2402 ceph_kick_flushing_caps(mdsc, s);
2403 wake_up_session_caps(s, 1);
2415 * caller must hold session s_mutex, dentry->d_lock
2417 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2419 struct ceph_dentry_info *di = ceph_dentry(dentry);
2421 ceph_put_mds_session(di->lease_session);
2422 di->lease_session = NULL;
2425 static void handle_lease(struct ceph_mds_client *mdsc,
2426 struct ceph_mds_session *session,
2427 struct ceph_msg *msg)
2429 struct super_block *sb = mdsc->client->sb;
2430 struct inode *inode;
2431 struct ceph_inode_info *ci;
2432 struct dentry *parent, *dentry;
2433 struct ceph_dentry_info *di;
2434 int mds = session->s_mds;
2435 struct ceph_mds_lease *h = msg->front.iov_base;
2437 struct ceph_vino vino;
2442 dout("handle_lease from mds%d\n", mds);
2445 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2447 vino.ino = le64_to_cpu(h->ino);
2448 vino.snap = CEPH_NOSNAP;
2449 mask = le16_to_cpu(h->mask);
2450 seq = le32_to_cpu(h->seq);
2451 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2452 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2453 if (dname.len != get_unaligned_le32(h+1))
2456 mutex_lock(&session->s_mutex);
2460 inode = ceph_find_inode(sb, vino);
2461 dout("handle_lease %s, mask %d, ino %llx %p %.*s\n",
2462 ceph_lease_op_name(h->action), mask, vino.ino, inode,
2463 dname.len, dname.name);
2464 if (inode == NULL) {
2465 dout("handle_lease no inode %llx\n", vino.ino);
2468 ci = ceph_inode(inode);
2471 parent = d_find_alias(inode);
2473 dout("no parent dentry on inode %p\n", inode);
2475 goto release; /* hrm... */
2477 dname.hash = full_name_hash(dname.name, dname.len);
2478 dentry = d_lookup(parent, &dname);
2483 spin_lock(&dentry->d_lock);
2484 di = ceph_dentry(dentry);
2485 switch (h->action) {
2486 case CEPH_MDS_LEASE_REVOKE:
2487 if (di && di->lease_session == session) {
2488 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2489 h->seq = cpu_to_le32(di->lease_seq);
2490 __ceph_mdsc_drop_dentry_lease(dentry);
2495 case CEPH_MDS_LEASE_RENEW:
2496 if (di && di->lease_session == session &&
2497 di->lease_gen == session->s_cap_gen &&
2498 di->lease_renew_from &&
2499 di->lease_renew_after == 0) {
2500 unsigned long duration =
2501 le32_to_cpu(h->duration_ms) * HZ / 1000;
2503 di->lease_seq = seq;
2504 dentry->d_time = di->lease_renew_from + duration;
2505 di->lease_renew_after = di->lease_renew_from +
2507 di->lease_renew_from = 0;
2511 spin_unlock(&dentry->d_lock);
2518 /* let's just reuse the same message */
2519 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2521 ceph_con_send(&session->s_con, msg);
2525 mutex_unlock(&session->s_mutex);
2529 pr_err("corrupt lease message\n");
2533 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2534 struct inode *inode,
2535 struct dentry *dentry, char action,
2538 struct ceph_msg *msg;
2539 struct ceph_mds_lease *lease;
2540 int len = sizeof(*lease) + sizeof(u32);
2543 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2544 inode, dentry, ceph_lease_op_name(action), session->s_mds);
2545 dnamelen = dentry->d_name.len;
2548 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS);
2551 lease = msg->front.iov_base;
2552 lease->action = action;
2553 lease->mask = cpu_to_le16(1);
2554 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2555 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2556 lease->seq = cpu_to_le32(seq);
2557 put_unaligned_le32(dnamelen, lease + 1);
2558 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2561 * if this is a preemptive lease RELEASE, no need to
2562 * flush request stream, since the actual request will
2565 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2567 ceph_con_send(&session->s_con, msg);
2571 * Preemptively release a lease we expect to invalidate anyway.
2572 * Pass @inode always, @dentry is optional.
2574 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2575 struct dentry *dentry, int mask)
2577 struct ceph_dentry_info *di;
2578 struct ceph_mds_session *session;
2581 BUG_ON(inode == NULL);
2582 BUG_ON(dentry == NULL);
2585 /* is dentry lease valid? */
2586 spin_lock(&dentry->d_lock);
2587 di = ceph_dentry(dentry);
2588 if (!di || !di->lease_session ||
2589 di->lease_session->s_mds < 0 ||
2590 di->lease_gen != di->lease_session->s_cap_gen ||
2591 !time_before(jiffies, dentry->d_time)) {
2592 dout("lease_release inode %p dentry %p -- "
2594 inode, dentry, mask);
2595 spin_unlock(&dentry->d_lock);
2599 /* we do have a lease on this dentry; note mds and seq */
2600 session = ceph_get_mds_session(di->lease_session);
2601 seq = di->lease_seq;
2602 __ceph_mdsc_drop_dentry_lease(dentry);
2603 spin_unlock(&dentry->d_lock);
2605 dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2606 inode, dentry, mask, session->s_mds);
2607 ceph_mdsc_lease_send_msg(session, inode, dentry,
2608 CEPH_MDS_LEASE_RELEASE, seq);
2609 ceph_put_mds_session(session);
2613 * drop all leases (and dentry refs) in preparation for umount
2615 static void drop_leases(struct ceph_mds_client *mdsc)
2619 dout("drop_leases\n");
2620 mutex_lock(&mdsc->mutex);
2621 for (i = 0; i < mdsc->max_sessions; i++) {
2622 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2625 mutex_unlock(&mdsc->mutex);
2626 mutex_lock(&s->s_mutex);
2627 mutex_unlock(&s->s_mutex);
2628 ceph_put_mds_session(s);
2629 mutex_lock(&mdsc->mutex);
2631 mutex_unlock(&mdsc->mutex);
2637 * delayed work -- periodically trim expired leases, renew caps with mds
2639 static void schedule_delayed(struct ceph_mds_client *mdsc)
2642 unsigned hz = round_jiffies_relative(HZ * delay);
2643 schedule_delayed_work(&mdsc->delayed_work, hz);
2646 static void delayed_work(struct work_struct *work)
2649 struct ceph_mds_client *mdsc =
2650 container_of(work, struct ceph_mds_client, delayed_work.work);
2654 dout("mdsc delayed_work\n");
2655 ceph_check_delayed_caps(mdsc);
2657 mutex_lock(&mdsc->mutex);
2658 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2659 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2660 mdsc->last_renew_caps);
2662 mdsc->last_renew_caps = jiffies;
2664 for (i = 0; i < mdsc->max_sessions; i++) {
2665 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2668 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2669 dout("resending session close request for mds%d\n",
2671 request_close_session(mdsc, s);
2672 ceph_put_mds_session(s);
2675 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2676 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2677 s->s_state = CEPH_MDS_SESSION_HUNG;
2678 pr_info("mds%d hung\n", s->s_mds);
2681 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2682 /* this mds is failed or recovering, just wait */
2683 ceph_put_mds_session(s);
2686 mutex_unlock(&mdsc->mutex);
2688 mutex_lock(&s->s_mutex);
2690 send_renew_caps(mdsc, s);
2692 ceph_con_keepalive(&s->s_con);
2693 ceph_add_cap_releases(mdsc, s, -1);
2694 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2695 s->s_state == CEPH_MDS_SESSION_HUNG)
2696 ceph_send_cap_releases(mdsc, s);
2697 mutex_unlock(&s->s_mutex);
2698 ceph_put_mds_session(s);
2700 mutex_lock(&mdsc->mutex);
2702 mutex_unlock(&mdsc->mutex);
2704 schedule_delayed(mdsc);
2708 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2710 mdsc->client = client;
2711 mutex_init(&mdsc->mutex);
2712 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2713 if (mdsc->mdsmap == NULL)
2716 init_completion(&mdsc->safe_umount_waiters);
2717 init_completion(&mdsc->session_close_waiters);
2718 INIT_LIST_HEAD(&mdsc->waiting_for_map);
2719 mdsc->sessions = NULL;
2720 mdsc->max_sessions = 0;
2722 init_rwsem(&mdsc->snap_rwsem);
2723 mdsc->snap_realms = RB_ROOT;
2724 INIT_LIST_HEAD(&mdsc->snap_empty);
2725 spin_lock_init(&mdsc->snap_empty_lock);
2727 mdsc->request_tree = RB_ROOT;
2728 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2729 mdsc->last_renew_caps = jiffies;
2730 INIT_LIST_HEAD(&mdsc->cap_delay_list);
2731 spin_lock_init(&mdsc->cap_delay_lock);
2732 INIT_LIST_HEAD(&mdsc->snap_flush_list);
2733 spin_lock_init(&mdsc->snap_flush_lock);
2734 mdsc->cap_flush_seq = 0;
2735 INIT_LIST_HEAD(&mdsc->cap_dirty);
2736 mdsc->num_cap_flushing = 0;
2737 spin_lock_init(&mdsc->cap_dirty_lock);
2738 init_waitqueue_head(&mdsc->cap_flushing_wq);
2739 spin_lock_init(&mdsc->dentry_lru_lock);
2740 INIT_LIST_HEAD(&mdsc->dentry_lru);
2746 * Wait for safe replies on open mds requests. If we time out, drop
2747 * all requests from the tree to avoid dangling dentry refs.
2749 static void wait_requests(struct ceph_mds_client *mdsc)
2751 struct ceph_mds_request *req;
2752 struct ceph_client *client = mdsc->client;
2754 mutex_lock(&mdsc->mutex);
2755 if (__get_oldest_req(mdsc)) {
2756 mutex_unlock(&mdsc->mutex);
2758 dout("wait_requests waiting for requests\n");
2759 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2760 client->mount_args->mount_timeout * HZ);
2762 /* tear down remaining requests */
2763 mutex_lock(&mdsc->mutex);
2764 while ((req = __get_oldest_req(mdsc))) {
2765 dout("wait_requests timed out on tid %llu\n",
2767 __unregister_request(mdsc, req);
2770 mutex_unlock(&mdsc->mutex);
2771 dout("wait_requests done\n");
2775 * called before mount is ro, and before dentries are torn down.
2776 * (hmm, does this still race with new lookups?)
2778 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2780 dout("pre_umount\n");
2784 ceph_flush_dirty_caps(mdsc);
2785 wait_requests(mdsc);
2789 * wait for all write mds requests to flush.
2791 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2793 struct ceph_mds_request *req = NULL, *nextreq;
2796 mutex_lock(&mdsc->mutex);
2797 dout("wait_unsafe_requests want %lld\n", want_tid);
2799 req = __get_oldest_req(mdsc);
2800 while (req && req->r_tid <= want_tid) {
2801 /* find next request */
2802 n = rb_next(&req->r_node);
2804 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2807 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2809 ceph_mdsc_get_request(req);
2811 ceph_mdsc_get_request(nextreq);
2812 mutex_unlock(&mdsc->mutex);
2813 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
2814 req->r_tid, want_tid);
2815 wait_for_completion(&req->r_safe_completion);
2816 mutex_lock(&mdsc->mutex);
2817 ceph_mdsc_put_request(req);
2819 break; /* next dne before, so we're done! */
2820 if (RB_EMPTY_NODE(&nextreq->r_node)) {
2821 /* next request was removed from tree */
2822 ceph_mdsc_put_request(nextreq);
2825 ceph_mdsc_put_request(nextreq); /* won't go away */
2829 mutex_unlock(&mdsc->mutex);
2830 dout("wait_unsafe_requests done\n");
2833 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2835 u64 want_tid, want_flush;
2837 if (mdsc->client->mount_state == CEPH_MOUNT_SHUTDOWN)
2841 mutex_lock(&mdsc->mutex);
2842 want_tid = mdsc->last_tid;
2843 want_flush = mdsc->cap_flush_seq;
2844 mutex_unlock(&mdsc->mutex);
2845 dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2847 ceph_flush_dirty_caps(mdsc);
2849 wait_unsafe_requests(mdsc, want_tid);
2850 wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2855 * called after sb is ro.
2857 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2859 struct ceph_mds_session *session;
2862 struct ceph_client *client = mdsc->client;
2863 unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2865 dout("close_sessions\n");
2867 mutex_lock(&mdsc->mutex);
2869 /* close sessions */
2871 while (time_before(jiffies, started + timeout)) {
2872 dout("closing sessions\n");
2874 for (i = 0; i < mdsc->max_sessions; i++) {
2875 session = __ceph_lookup_mds_session(mdsc, i);
2878 mutex_unlock(&mdsc->mutex);
2879 mutex_lock(&session->s_mutex);
2880 __close_session(mdsc, session);
2881 mutex_unlock(&session->s_mutex);
2882 ceph_put_mds_session(session);
2883 mutex_lock(&mdsc->mutex);
2889 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2892 dout("waiting for sessions to close\n");
2893 mutex_unlock(&mdsc->mutex);
2894 wait_for_completion_timeout(&mdsc->session_close_waiters,
2896 mutex_lock(&mdsc->mutex);
2899 /* tear down remaining sessions */
2900 for (i = 0; i < mdsc->max_sessions; i++) {
2901 if (mdsc->sessions[i]) {
2902 session = get_session(mdsc->sessions[i]);
2903 __unregister_session(mdsc, session);
2904 mutex_unlock(&mdsc->mutex);
2905 mutex_lock(&session->s_mutex);
2906 remove_session_caps(session);
2907 mutex_unlock(&session->s_mutex);
2908 ceph_put_mds_session(session);
2909 mutex_lock(&mdsc->mutex);
2913 WARN_ON(!list_empty(&mdsc->cap_delay_list));
2915 mutex_unlock(&mdsc->mutex);
2917 ceph_cleanup_empty_realms(mdsc);
2919 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2924 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2927 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2929 ceph_mdsmap_destroy(mdsc->mdsmap);
2930 kfree(mdsc->sessions);
2935 * handle mds map update.
2937 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2941 void *p = msg->front.iov_base;
2942 void *end = p + msg->front.iov_len;
2943 struct ceph_mdsmap *newmap, *oldmap;
2944 struct ceph_fsid fsid;
2947 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2948 ceph_decode_copy(&p, &fsid, sizeof(fsid));
2949 if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2951 epoch = ceph_decode_32(&p);
2952 maplen = ceph_decode_32(&p);
2953 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2955 /* do we need it? */
2956 ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2957 mutex_lock(&mdsc->mutex);
2958 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2959 dout("handle_map epoch %u <= our %u\n",
2960 epoch, mdsc->mdsmap->m_epoch);
2961 mutex_unlock(&mdsc->mutex);
2965 newmap = ceph_mdsmap_decode(&p, end);
2966 if (IS_ERR(newmap)) {
2967 err = PTR_ERR(newmap);
2971 /* swap into place */
2973 oldmap = mdsc->mdsmap;
2974 mdsc->mdsmap = newmap;
2975 check_new_map(mdsc, newmap, oldmap);
2976 ceph_mdsmap_destroy(oldmap);
2978 mdsc->mdsmap = newmap; /* first mds map */
2980 mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2982 __wake_requests(mdsc, &mdsc->waiting_for_map);
2984 mutex_unlock(&mdsc->mutex);
2985 schedule_delayed(mdsc);
2989 mutex_unlock(&mdsc->mutex);
2991 pr_err("error decoding mdsmap %d\n", err);
2995 static struct ceph_connection *con_get(struct ceph_connection *con)
2997 struct ceph_mds_session *s = con->private;
2999 if (get_session(s)) {
3000 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3003 dout("mdsc con_get %p FAIL\n", s);
3007 static void con_put(struct ceph_connection *con)
3009 struct ceph_mds_session *s = con->private;
3011 ceph_put_mds_session(s);
3012 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
3016 * if the client is unresponsive for long enough, the mds will kill
3017 * the session entirely.
3019 static void peer_reset(struct ceph_connection *con)
3021 struct ceph_mds_session *s = con->private;
3022 struct ceph_mds_client *mdsc = s->s_mdsc;
3024 pr_warning("mds%d closed our session\n", s->s_mds);
3025 send_mds_reconnect(mdsc, s);
3028 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3030 struct ceph_mds_session *s = con->private;
3031 struct ceph_mds_client *mdsc = s->s_mdsc;
3032 int type = le16_to_cpu(msg->hdr.type);
3034 mutex_lock(&mdsc->mutex);
3035 if (__verify_registered_session(mdsc, s) < 0) {
3036 mutex_unlock(&mdsc->mutex);
3039 mutex_unlock(&mdsc->mutex);
3042 case CEPH_MSG_MDS_MAP:
3043 ceph_mdsc_handle_map(mdsc, msg);
3045 case CEPH_MSG_CLIENT_SESSION:
3046 handle_session(s, msg);
3048 case CEPH_MSG_CLIENT_REPLY:
3049 handle_reply(s, msg);
3051 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3052 handle_forward(mdsc, s, msg);
3054 case CEPH_MSG_CLIENT_CAPS:
3055 ceph_handle_caps(s, msg);
3057 case CEPH_MSG_CLIENT_SNAP:
3058 ceph_handle_snap(mdsc, s, msg);
3060 case CEPH_MSG_CLIENT_LEASE:
3061 handle_lease(mdsc, s, msg);
3065 pr_err("received unknown message type %d %s\n", type,
3066 ceph_msg_type_name(type));
3075 static int get_authorizer(struct ceph_connection *con,
3076 void **buf, int *len, int *proto,
3077 void **reply_buf, int *reply_len, int force_new)
3079 struct ceph_mds_session *s = con->private;
3080 struct ceph_mds_client *mdsc = s->s_mdsc;
3081 struct ceph_auth_client *ac = mdsc->client->monc.auth;
3084 if (force_new && s->s_authorizer) {
3085 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3086 s->s_authorizer = NULL;
3088 if (s->s_authorizer == NULL) {
3089 if (ac->ops->create_authorizer) {
3090 ret = ac->ops->create_authorizer(
3091 ac, CEPH_ENTITY_TYPE_MDS,
3093 &s->s_authorizer_buf,
3094 &s->s_authorizer_buf_len,
3095 &s->s_authorizer_reply_buf,
3096 &s->s_authorizer_reply_buf_len);
3102 *proto = ac->protocol;
3103 *buf = s->s_authorizer_buf;
3104 *len = s->s_authorizer_buf_len;
3105 *reply_buf = s->s_authorizer_reply_buf;
3106 *reply_len = s->s_authorizer_reply_buf_len;
3111 static int verify_authorizer_reply(struct ceph_connection *con, int len)
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;
3117 return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3120 static int invalidate_authorizer(struct ceph_connection *con)
3122 struct ceph_mds_session *s = con->private;
3123 struct ceph_mds_client *mdsc = s->s_mdsc;
3124 struct ceph_auth_client *ac = mdsc->client->monc.auth;
3126 if (ac->ops->invalidate_authorizer)
3127 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3129 return ceph_monc_validate_auth(&mdsc->client->monc);
3132 static const struct ceph_connection_operations mds_con_ops = {
3135 .dispatch = dispatch,
3136 .get_authorizer = get_authorizer,
3137 .verify_authorizer_reply = verify_authorizer_reply,
3138 .invalidate_authorizer = invalidate_authorizer,
3139 .peer_reset = peer_reset,