1 #include "ceph_debug.h"
3 #include <linux/wait.h>
4 #include <linux/sched.h>
6 #include "mds_client.h"
7 #include "mon_client.h"
15 * A cluster of MDS (metadata server) daemons is responsible for
16 * managing the file system namespace (the directory hierarchy and
17 * inodes) and for coordinating shared access to storage. Metadata is
18 * partitioning hierarchically across a number of servers, and that
19 * partition varies over time as the cluster adjusts the distribution
20 * in order to balance load.
22 * The MDS client is primarily responsible to managing synchronous
23 * metadata requests for operations like open, unlink, and so forth.
24 * If there is a MDS failure, we find out about it when we (possibly
25 * request and) receive a new MDS map, and can resubmit affected
28 * For the most part, though, we take advantage of a lossless
29 * communications channel to the MDS, and do not need to worry about
30 * timing out or resubmitting requests.
32 * We maintain a stateful "session" with each MDS we interact with.
33 * Within each session, we sent periodic heartbeat messages to ensure
34 * any capabilities or leases we have been issues remain valid. If
35 * the session times out and goes stale, our leases and capabilities
36 * are no longer valid.
39 static void __wake_requests(struct ceph_mds_client *mdsc,
40 struct list_head *head);
42 const static struct ceph_connection_operations mds_con_ops;
50 * parse individual inode info
52 static int parse_reply_info_in(void **p, void *end,
53 struct ceph_mds_reply_info_in *info)
58 *p += sizeof(struct ceph_mds_reply_inode) +
59 sizeof(*info->in->fragtree.splits) *
60 le32_to_cpu(info->in->fragtree.nsplits);
62 ceph_decode_32_safe(p, end, info->symlink_len, bad);
63 ceph_decode_need(p, end, info->symlink_len, bad);
65 *p += info->symlink_len;
67 ceph_decode_32_safe(p, end, info->xattr_len, bad);
68 ceph_decode_need(p, end, info->xattr_len, bad);
69 info->xattr_data = *p;
70 *p += info->xattr_len;
77 * parse a normal reply, which may contain a (dir+)dentry and/or a
80 static int parse_reply_info_trace(void **p, void *end,
81 struct ceph_mds_reply_info_parsed *info)
85 if (info->head->is_dentry) {
86 err = parse_reply_info_in(p, end, &info->diri);
90 if (unlikely(*p + sizeof(*info->dirfrag) > end))
93 *p += sizeof(*info->dirfrag) +
94 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
95 if (unlikely(*p > end))
98 ceph_decode_32_safe(p, end, info->dname_len, bad);
99 ceph_decode_need(p, end, info->dname_len, bad);
101 *p += info->dname_len;
103 *p += sizeof(*info->dlease);
106 if (info->head->is_target) {
107 err = parse_reply_info_in(p, end, &info->targeti);
112 if (unlikely(*p != end))
119 pr_err("problem parsing mds trace %d\n", err);
124 * parse readdir results
126 static int parse_reply_info_dir(void **p, void *end,
127 struct ceph_mds_reply_info_parsed *info)
133 if (*p + sizeof(*info->dir_dir) > end)
135 *p += sizeof(*info->dir_dir) +
136 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
140 ceph_decode_need(p, end, sizeof(num) + 2, bad);
141 num = ceph_decode_32(p);
142 info->dir_end = ceph_decode_8(p);
143 info->dir_complete = ceph_decode_8(p);
147 /* alloc large array */
149 info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
150 sizeof(*info->dir_dname) +
151 sizeof(*info->dir_dname_len) +
152 sizeof(*info->dir_dlease),
154 if (info->dir_in == NULL) {
158 info->dir_dname = (void *)(info->dir_in + num);
159 info->dir_dname_len = (void *)(info->dir_dname + num);
160 info->dir_dlease = (void *)(info->dir_dname_len + num);
164 ceph_decode_need(p, end, sizeof(u32)*2, bad);
165 info->dir_dname_len[i] = ceph_decode_32(p);
166 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
167 info->dir_dname[i] = *p;
168 *p += info->dir_dname_len[i];
169 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
171 info->dir_dlease[i] = *p;
172 *p += sizeof(struct ceph_mds_reply_lease);
175 err = parse_reply_info_in(p, end, &info->dir_in[i]);
190 pr_err("problem parsing dir contents %d\n", err);
195 * parse entire mds reply
197 static int parse_reply_info(struct ceph_msg *msg,
198 struct ceph_mds_reply_info_parsed *info)
204 info->head = msg->front.iov_base;
205 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
206 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
209 ceph_decode_32_safe(&p, end, len, bad);
211 err = parse_reply_info_trace(&p, p+len, info);
217 ceph_decode_32_safe(&p, end, len, bad);
219 err = parse_reply_info_dir(&p, p+len, info);
225 ceph_decode_32_safe(&p, end, len, bad);
226 info->snapblob_len = len;
237 pr_err("mds parse_reply err %d\n", err);
241 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
250 static const char *session_state_name(int s)
253 case CEPH_MDS_SESSION_NEW: return "new";
254 case CEPH_MDS_SESSION_OPENING: return "opening";
255 case CEPH_MDS_SESSION_OPEN: return "open";
256 case CEPH_MDS_SESSION_HUNG: return "hung";
257 case CEPH_MDS_SESSION_CLOSING: return "closing";
258 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
259 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
260 default: return "???";
264 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
266 if (atomic_inc_not_zero(&s->s_ref)) {
267 dout("mdsc get_session %p %d -> %d\n", s,
268 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
271 dout("mdsc get_session %p 0 -- FAIL", s);
276 void ceph_put_mds_session(struct ceph_mds_session *s)
278 dout("mdsc put_session %p %d -> %d\n", s,
279 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
280 if (atomic_dec_and_test(&s->s_ref)) {
282 s->s_mdsc->client->monc.auth->ops->destroy_authorizer(
283 s->s_mdsc->client->monc.auth, s->s_authorizer);
289 * called under mdsc->mutex
291 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
294 struct ceph_mds_session *session;
296 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
298 session = mdsc->sessions[mds];
299 dout("lookup_mds_session %p %d\n", session,
300 atomic_read(&session->s_ref));
301 get_session(session);
305 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
307 if (mds >= mdsc->max_sessions)
309 return mdsc->sessions[mds];
312 static int __verify_registered_session(struct ceph_mds_client *mdsc,
313 struct ceph_mds_session *s)
315 if (s->s_mds >= mdsc->max_sessions ||
316 mdsc->sessions[s->s_mds] != s)
322 * create+register a new session for given mds.
323 * called under mdsc->mutex.
325 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
328 struct ceph_mds_session *s;
330 s = kzalloc(sizeof(*s), GFP_NOFS);
333 s->s_state = CEPH_MDS_SESSION_NEW;
336 mutex_init(&s->s_mutex);
338 ceph_con_init(mdsc->client->msgr, &s->s_con);
339 s->s_con.private = s;
340 s->s_con.ops = &mds_con_ops;
341 s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
342 s->s_con.peer_name.num = cpu_to_le64(mds);
344 spin_lock_init(&s->s_cap_lock);
347 s->s_renew_requested = 0;
349 INIT_LIST_HEAD(&s->s_caps);
352 atomic_set(&s->s_ref, 1);
353 INIT_LIST_HEAD(&s->s_waiting);
354 INIT_LIST_HEAD(&s->s_unsafe);
355 s->s_num_cap_releases = 0;
356 s->s_cap_iterator = NULL;
357 INIT_LIST_HEAD(&s->s_cap_releases);
358 INIT_LIST_HEAD(&s->s_cap_releases_done);
359 INIT_LIST_HEAD(&s->s_cap_flushing);
360 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
362 dout("register_session mds%d\n", mds);
363 if (mds >= mdsc->max_sessions) {
364 int newmax = 1 << get_count_order(mds+1);
365 struct ceph_mds_session **sa;
367 dout("register_session realloc to %d\n", newmax);
368 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
371 if (mdsc->sessions) {
372 memcpy(sa, mdsc->sessions,
373 mdsc->max_sessions * sizeof(void *));
374 kfree(mdsc->sessions);
377 mdsc->max_sessions = newmax;
379 mdsc->sessions[mds] = s;
380 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
382 ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
388 return ERR_PTR(-ENOMEM);
392 * called under mdsc->mutex
394 static void __unregister_session(struct ceph_mds_client *mdsc,
395 struct ceph_mds_session *s)
397 dout("__unregister_session mds%d %p\n", s->s_mds, s);
398 BUG_ON(mdsc->sessions[s->s_mds] != s);
399 mdsc->sessions[s->s_mds] = NULL;
400 ceph_con_close(&s->s_con);
401 ceph_put_mds_session(s);
405 * drop session refs in request.
407 * should be last request ref, or hold mdsc->mutex
409 static void put_request_session(struct ceph_mds_request *req)
411 if (req->r_session) {
412 ceph_put_mds_session(req->r_session);
413 req->r_session = NULL;
417 void ceph_mdsc_release_request(struct kref *kref)
419 struct ceph_mds_request *req = container_of(kref,
420 struct ceph_mds_request,
423 ceph_msg_put(req->r_request);
425 ceph_msg_put(req->r_reply);
426 destroy_reply_info(&req->r_reply_info);
429 ceph_put_cap_refs(ceph_inode(req->r_inode),
433 if (req->r_locked_dir)
434 ceph_put_cap_refs(ceph_inode(req->r_locked_dir),
436 if (req->r_target_inode)
437 iput(req->r_target_inode);
440 if (req->r_old_dentry) {
442 ceph_inode(req->r_old_dentry->d_parent->d_inode),
444 dput(req->r_old_dentry);
448 put_request_session(req);
449 ceph_unreserve_caps(&req->r_caps_reservation);
454 * lookup session, bump ref if found.
456 * called under mdsc->mutex.
458 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
461 struct ceph_mds_request *req;
462 struct rb_node *n = mdsc->request_tree.rb_node;
465 req = rb_entry(n, struct ceph_mds_request, r_node);
466 if (tid < req->r_tid)
468 else if (tid > req->r_tid)
471 ceph_mdsc_get_request(req);
478 static void __insert_request(struct ceph_mds_client *mdsc,
479 struct ceph_mds_request *new)
481 struct rb_node **p = &mdsc->request_tree.rb_node;
482 struct rb_node *parent = NULL;
483 struct ceph_mds_request *req = NULL;
487 req = rb_entry(parent, struct ceph_mds_request, r_node);
488 if (new->r_tid < req->r_tid)
490 else if (new->r_tid > req->r_tid)
496 rb_link_node(&new->r_node, parent, p);
497 rb_insert_color(&new->r_node, &mdsc->request_tree);
501 * Register an in-flight request, and assign a tid. Link to directory
502 * are modifying (if any).
504 * Called under mdsc->mutex.
506 static void __register_request(struct ceph_mds_client *mdsc,
507 struct ceph_mds_request *req,
510 req->r_tid = ++mdsc->last_tid;
512 ceph_reserve_caps(&req->r_caps_reservation, req->r_num_caps);
513 dout("__register_request %p tid %lld\n", req, req->r_tid);
514 ceph_mdsc_get_request(req);
515 __insert_request(mdsc, req);
518 struct ceph_inode_info *ci = ceph_inode(dir);
520 spin_lock(&ci->i_unsafe_lock);
521 req->r_unsafe_dir = dir;
522 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
523 spin_unlock(&ci->i_unsafe_lock);
527 static void __unregister_request(struct ceph_mds_client *mdsc,
528 struct ceph_mds_request *req)
530 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
531 rb_erase(&req->r_node, &mdsc->request_tree);
532 ceph_mdsc_put_request(req);
534 if (req->r_unsafe_dir) {
535 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
537 spin_lock(&ci->i_unsafe_lock);
538 list_del_init(&req->r_unsafe_dir_item);
539 spin_unlock(&ci->i_unsafe_lock);
544 * Choose mds to send request to next. If there is a hint set in the
545 * request (e.g., due to a prior forward hint from the mds), use that.
546 * Otherwise, consult frag tree and/or caps to identify the
547 * appropriate mds. If all else fails, choose randomly.
549 * Called under mdsc->mutex.
551 static int __choose_mds(struct ceph_mds_client *mdsc,
552 struct ceph_mds_request *req)
555 struct ceph_inode_info *ci;
556 struct ceph_cap *cap;
557 int mode = req->r_direct_mode;
559 u32 hash = req->r_direct_hash;
560 bool is_hash = req->r_direct_is_hash;
563 * is there a specific mds we should try? ignore hint if we have
564 * no session and the mds is not up (active or recovering).
566 if (req->r_resend_mds >= 0 &&
567 (__have_session(mdsc, req->r_resend_mds) ||
568 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
569 dout("choose_mds using resend_mds mds%d\n",
571 return req->r_resend_mds;
574 if (mode == USE_RANDOM_MDS)
579 inode = req->r_inode;
580 } else if (req->r_dentry) {
581 if (req->r_dentry->d_inode) {
582 inode = req->r_dentry->d_inode;
584 inode = req->r_dentry->d_parent->d_inode;
585 hash = req->r_dentry->d_name.hash;
589 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
593 ci = ceph_inode(inode);
595 if (is_hash && S_ISDIR(inode->i_mode)) {
596 struct ceph_inode_frag frag;
599 ceph_choose_frag(ci, hash, &frag, &found);
601 if (mode == USE_ANY_MDS && frag.ndist > 0) {
604 /* choose a random replica */
605 get_random_bytes(&r, 1);
608 dout("choose_mds %p %llx.%llx "
609 "frag %u mds%d (%d/%d)\n",
610 inode, ceph_vinop(inode),
616 /* since this file/dir wasn't known to be
617 * replicated, then we want to look for the
618 * authoritative mds. */
621 /* choose auth mds */
623 dout("choose_mds %p %llx.%llx "
624 "frag %u mds%d (auth)\n",
625 inode, ceph_vinop(inode), frag.frag, mds);
631 spin_lock(&inode->i_lock);
633 if (mode == USE_AUTH_MDS)
634 cap = ci->i_auth_cap;
635 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
636 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
638 spin_unlock(&inode->i_lock);
641 mds = cap->session->s_mds;
642 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
643 inode, ceph_vinop(inode), mds,
644 cap == ci->i_auth_cap ? "auth " : "", cap);
645 spin_unlock(&inode->i_lock);
649 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
650 dout("choose_mds chose random mds%d\n", mds);
658 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
660 struct ceph_msg *msg;
661 struct ceph_mds_session_head *h;
663 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), 0, 0, NULL);
665 pr_err("create_session_msg ENOMEM creating msg\n");
666 return ERR_PTR(PTR_ERR(msg));
668 h = msg->front.iov_base;
669 h->op = cpu_to_le32(op);
670 h->seq = cpu_to_le64(seq);
675 * send session open request.
677 * called under mdsc->mutex
679 static int __open_session(struct ceph_mds_client *mdsc,
680 struct ceph_mds_session *session)
682 struct ceph_msg *msg;
684 int mds = session->s_mds;
687 /* wait for mds to go active? */
688 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
689 dout("open_session to mds%d (%s)\n", mds,
690 ceph_mds_state_name(mstate));
691 session->s_state = CEPH_MDS_SESSION_OPENING;
692 session->s_renew_requested = jiffies;
694 /* send connect message */
695 msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
700 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.
736 * caller must hold session s_mutex
738 static int iterate_session_caps(struct ceph_mds_session *session,
739 int (*cb)(struct inode *, struct ceph_cap *,
743 struct ceph_cap *cap;
744 struct inode *inode, *last_inode = NULL;
745 struct ceph_cap *old_cap = NULL;
748 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
749 spin_lock(&session->s_cap_lock);
750 p = session->s_caps.next;
751 while (p != &session->s_caps) {
752 cap = list_entry(p, struct ceph_cap, session_caps);
753 inode = igrab(&cap->ci->vfs_inode);
758 session->s_cap_iterator = cap;
759 spin_unlock(&session->s_cap_lock);
766 ceph_put_cap(old_cap);
770 ret = cb(inode, cap, arg);
773 spin_lock(&session->s_cap_lock);
775 if (cap->ci == NULL) {
776 dout("iterate_session_caps finishing cap %p removal\n",
778 BUG_ON(cap->session != session);
779 list_del_init(&cap->session_caps);
780 session->s_nr_caps--;
782 old_cap = cap; /* put_cap it w/o locks held */
789 session->s_cap_iterator = NULL;
790 spin_unlock(&session->s_cap_lock);
795 ceph_put_cap(old_cap);
800 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
803 struct ceph_inode_info *ci = ceph_inode(inode);
804 dout("removing cap %p, ci is %p, inode is %p\n",
805 cap, ci, &ci->vfs_inode);
806 ceph_remove_cap(cap);
811 * caller must hold session s_mutex
813 static void remove_session_caps(struct ceph_mds_session *session)
815 dout("remove_session_caps on %p\n", session);
816 iterate_session_caps(session, remove_session_caps_cb, NULL);
817 BUG_ON(session->s_nr_caps > 0);
818 cleanup_cap_releases(session);
822 * wake up any threads waiting on this session's caps. if the cap is
823 * old (didn't get renewed on the client reconnect), remove it now.
825 * caller must hold s_mutex.
827 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
830 struct ceph_inode_info *ci = ceph_inode(inode);
832 wake_up(&ci->i_cap_wq);
834 spin_lock(&inode->i_lock);
835 ci->i_wanted_max_size = 0;
836 ci->i_requested_max_size = 0;
837 spin_unlock(&inode->i_lock);
842 static void wake_up_session_caps(struct ceph_mds_session *session,
845 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
846 iterate_session_caps(session, wake_up_session_cb,
847 (void *)(unsigned long)reconnect);
851 * Send periodic message to MDS renewing all currently held caps. The
852 * ack will reset the expiration for all caps from this session.
854 * caller holds s_mutex
856 static int send_renew_caps(struct ceph_mds_client *mdsc,
857 struct ceph_mds_session *session)
859 struct ceph_msg *msg;
862 if (time_after_eq(jiffies, session->s_cap_ttl) &&
863 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
864 pr_info("mds%d caps stale\n", session->s_mds);
866 /* do not try to renew caps until a recovering mds has reconnected
867 * with its clients. */
868 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
869 if (state < CEPH_MDS_STATE_RECONNECT) {
870 dout("send_renew_caps ignoring mds%d (%s)\n",
871 session->s_mds, ceph_mds_state_name(state));
875 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
876 ceph_mds_state_name(state));
877 session->s_renew_requested = jiffies;
878 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
879 ++session->s_renew_seq);
882 ceph_con_send(&session->s_con, msg);
887 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
889 * Called under session->s_mutex
891 static void renewed_caps(struct ceph_mds_client *mdsc,
892 struct ceph_mds_session *session, int is_renew)
897 spin_lock(&session->s_cap_lock);
898 was_stale = is_renew && (session->s_cap_ttl == 0 ||
899 time_after_eq(jiffies, session->s_cap_ttl));
901 session->s_cap_ttl = session->s_renew_requested +
902 mdsc->mdsmap->m_session_timeout*HZ;
905 if (time_before(jiffies, session->s_cap_ttl)) {
906 pr_info("mds%d caps renewed\n", session->s_mds);
909 pr_info("mds%d caps still stale\n", session->s_mds);
912 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
913 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
914 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
915 spin_unlock(&session->s_cap_lock);
918 wake_up_session_caps(session, 0);
922 * send a session close request
924 static int request_close_session(struct ceph_mds_client *mdsc,
925 struct ceph_mds_session *session)
927 struct ceph_msg *msg;
930 dout("request_close_session mds%d state %s seq %lld\n",
931 session->s_mds, session_state_name(session->s_state),
933 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
937 ceph_con_send(&session->s_con, msg);
942 * Called with s_mutex held.
944 static int __close_session(struct ceph_mds_client *mdsc,
945 struct ceph_mds_session *session)
947 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
949 session->s_state = CEPH_MDS_SESSION_CLOSING;
950 return request_close_session(mdsc, session);
956 * Because we can't cache an inode without one or more caps, we do
957 * this indirectly: if a cap is unused, we prune its aliases, at which
958 * point the inode will hopefully get dropped to.
960 * Yes, this is a bit sloppy. Our only real goal here is to respond to
961 * memory pressure from the MDS, though, so it needn't be perfect.
963 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
965 struct ceph_mds_session *session = arg;
966 struct ceph_inode_info *ci = ceph_inode(inode);
967 int used, oissued, mine;
969 if (session->s_trim_caps <= 0)
972 spin_lock(&inode->i_lock);
973 mine = cap->issued | cap->implemented;
974 used = __ceph_caps_used(ci);
975 oissued = __ceph_caps_issued_other(ci, cap);
977 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
978 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
979 ceph_cap_string(used));
980 if (ci->i_dirty_caps)
981 goto out; /* dirty caps */
982 if ((used & ~oissued) & mine)
983 goto out; /* we need these caps */
985 session->s_trim_caps--;
987 /* we aren't the only cap.. just remove us */
988 __ceph_remove_cap(cap);
990 /* try to drop referring dentries */
991 spin_unlock(&inode->i_lock);
992 d_prune_aliases(inode);
993 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
994 inode, cap, atomic_read(&inode->i_count));
999 spin_unlock(&inode->i_lock);
1004 * Trim session cap count down to some max number.
1006 static int trim_caps(struct ceph_mds_client *mdsc,
1007 struct ceph_mds_session *session,
1010 int trim_caps = session->s_nr_caps - max_caps;
1012 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1013 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1014 if (trim_caps > 0) {
1015 session->s_trim_caps = trim_caps;
1016 iterate_session_caps(session, trim_caps_cb, session);
1017 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1018 session->s_mds, session->s_nr_caps, max_caps,
1019 trim_caps - session->s_trim_caps);
1020 session->s_trim_caps = 0;
1026 * Allocate cap_release messages. If there is a partially full message
1027 * in the queue, try to allocate enough to cover it's remainder, so that
1028 * we can send it immediately.
1030 * Called under s_mutex.
1032 static int add_cap_releases(struct ceph_mds_client *mdsc,
1033 struct ceph_mds_session *session,
1036 struct ceph_msg *msg;
1037 struct ceph_mds_cap_release *head;
1041 extra = mdsc->client->mount_args->cap_release_safety;
1043 spin_lock(&session->s_cap_lock);
1045 if (!list_empty(&session->s_cap_releases)) {
1046 msg = list_first_entry(&session->s_cap_releases,
1049 head = msg->front.iov_base;
1050 extra += CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1053 while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1054 spin_unlock(&session->s_cap_lock);
1055 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1059 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1060 (int)msg->front.iov_len);
1061 head = msg->front.iov_base;
1062 head->num = cpu_to_le32(0);
1063 msg->front.iov_len = sizeof(*head);
1064 spin_lock(&session->s_cap_lock);
1065 list_add(&msg->list_head, &session->s_cap_releases);
1066 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1069 if (!list_empty(&session->s_cap_releases)) {
1070 msg = list_first_entry(&session->s_cap_releases,
1073 head = msg->front.iov_base;
1075 dout(" queueing non-full %p (%d)\n", msg,
1076 le32_to_cpu(head->num));
1077 list_move_tail(&msg->list_head,
1078 &session->s_cap_releases_done);
1079 session->s_num_cap_releases -=
1080 CEPH_CAPS_PER_RELEASE - le32_to_cpu(head->num);
1084 spin_unlock(&session->s_cap_lock);
1090 * flush all dirty inode data to disk.
1092 * returns true if we've flushed through want_flush_seq
1094 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1098 dout("check_cap_flush want %lld\n", want_flush_seq);
1099 mutex_lock(&mdsc->mutex);
1100 for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1101 struct ceph_mds_session *session = mdsc->sessions[mds];
1105 get_session(session);
1106 mutex_unlock(&mdsc->mutex);
1108 mutex_lock(&session->s_mutex);
1109 if (!list_empty(&session->s_cap_flushing)) {
1110 struct ceph_inode_info *ci =
1111 list_entry(session->s_cap_flushing.next,
1112 struct ceph_inode_info,
1114 struct inode *inode = &ci->vfs_inode;
1116 spin_lock(&inode->i_lock);
1117 if (ci->i_cap_flush_seq <= want_flush_seq) {
1118 dout("check_cap_flush still flushing %p "
1119 "seq %lld <= %lld to mds%d\n", inode,
1120 ci->i_cap_flush_seq, want_flush_seq,
1124 spin_unlock(&inode->i_lock);
1126 mutex_unlock(&session->s_mutex);
1127 ceph_put_mds_session(session);
1131 mutex_lock(&mdsc->mutex);
1134 mutex_unlock(&mdsc->mutex);
1135 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1140 * called under s_mutex
1142 static void send_cap_releases(struct ceph_mds_client *mdsc,
1143 struct ceph_mds_session *session)
1145 struct ceph_msg *msg;
1147 dout("send_cap_releases mds%d\n", session->s_mds);
1149 spin_lock(&session->s_cap_lock);
1150 if (list_empty(&session->s_cap_releases_done))
1152 msg = list_first_entry(&session->s_cap_releases_done,
1153 struct ceph_msg, list_head);
1154 list_del_init(&msg->list_head);
1155 spin_unlock(&session->s_cap_lock);
1156 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1157 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1158 ceph_con_send(&session->s_con, msg);
1160 spin_unlock(&session->s_cap_lock);
1168 * Create an mds request.
1170 struct ceph_mds_request *
1171 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1173 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1176 return ERR_PTR(-ENOMEM);
1178 req->r_started = jiffies;
1179 req->r_resend_mds = -1;
1180 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1182 kref_init(&req->r_kref);
1183 INIT_LIST_HEAD(&req->r_wait);
1184 init_completion(&req->r_completion);
1185 init_completion(&req->r_safe_completion);
1186 INIT_LIST_HEAD(&req->r_unsafe_item);
1189 req->r_direct_mode = mode;
1194 * return oldest (lowest) request, tid in request tree, 0 if none.
1196 * called under mdsc->mutex.
1198 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1200 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1202 return rb_entry(rb_first(&mdsc->request_tree),
1203 struct ceph_mds_request, r_node);
1206 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1208 struct ceph_mds_request *req = __get_oldest_req(mdsc);
1216 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1217 * on build_path_from_dentry in fs/cifs/dir.c.
1219 * If @stop_on_nosnap, generate path relative to the first non-snapped
1222 * Encode hidden .snap dirs as a double /, i.e.
1223 * foo/.snap/bar -> foo//bar
1225 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1228 struct dentry *temp;
1233 return ERR_PTR(-EINVAL);
1237 for (temp = dentry; !IS_ROOT(temp);) {
1238 struct inode *inode = temp->d_inode;
1239 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1240 len++; /* slash only */
1241 else if (stop_on_nosnap && inode &&
1242 ceph_snap(inode) == CEPH_NOSNAP)
1245 len += 1 + temp->d_name.len;
1246 temp = temp->d_parent;
1248 pr_err("build_path_dentry corrupt dentry %p\n", dentry);
1249 return ERR_PTR(-EINVAL);
1253 len--; /* no leading '/' */
1255 path = kmalloc(len+1, GFP_NOFS);
1257 return ERR_PTR(-ENOMEM);
1259 path[pos] = 0; /* trailing null */
1260 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1261 struct inode *inode = temp->d_inode;
1263 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1264 dout("build_path_dentry path+%d: %p SNAPDIR\n",
1266 } else if (stop_on_nosnap && inode &&
1267 ceph_snap(inode) == CEPH_NOSNAP) {
1270 pos -= temp->d_name.len;
1273 strncpy(path + pos, temp->d_name.name,
1275 dout("build_path_dentry path+%d: %p '%.*s'\n",
1276 pos, temp, temp->d_name.len, path + pos);
1280 temp = temp->d_parent;
1282 pr_err("build_path_dentry corrupt dentry\n");
1284 return ERR_PTR(-EINVAL);
1288 pr_err("build_path_dentry did not end path lookup where "
1289 "expected, namelen is %d, pos is %d\n", len, pos);
1290 /* presumably this is only possible if racing with a
1291 rename of one of the parent directories (we can not
1292 lock the dentries above us to prevent this, but
1293 retrying should be harmless) */
1298 *base = ceph_ino(temp->d_inode);
1300 dout("build_path_dentry on %p %d built %llx '%.*s'\n",
1301 dentry, atomic_read(&dentry->d_count), *base, len, path);
1305 static int build_dentry_path(struct dentry *dentry,
1306 const char **ppath, int *ppathlen, u64 *pino,
1311 if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1312 *pino = ceph_ino(dentry->d_parent->d_inode);
1313 *ppath = dentry->d_name.name;
1314 *ppathlen = dentry->d_name.len;
1317 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1319 return PTR_ERR(path);
1325 static int build_inode_path(struct inode *inode,
1326 const char **ppath, int *ppathlen, u64 *pino,
1329 struct dentry *dentry;
1332 if (ceph_snap(inode) == CEPH_NOSNAP) {
1333 *pino = ceph_ino(inode);
1337 dentry = d_find_alias(inode);
1338 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1341 return PTR_ERR(path);
1348 * request arguments may be specified via an inode *, a dentry *, or
1349 * an explicit ino+path.
1351 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1352 const char *rpath, u64 rino,
1353 const char **ppath, int *pathlen,
1354 u64 *ino, int *freepath)
1359 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1360 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1362 } else if (rdentry) {
1363 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1364 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1369 *pathlen = strlen(rpath);
1370 dout(" path %.*s\n", *pathlen, rpath);
1377 * called under mdsc->mutex
1379 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1380 struct ceph_mds_request *req,
1383 struct ceph_msg *msg;
1384 struct ceph_mds_request_head *head;
1385 const char *path1 = NULL;
1386 const char *path2 = NULL;
1387 u64 ino1 = 0, ino2 = 0;
1388 int pathlen1 = 0, pathlen2 = 0;
1389 int freepath1 = 0, freepath2 = 0;
1395 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1396 req->r_path1, req->r_ino1.ino,
1397 &path1, &pathlen1, &ino1, &freepath1);
1403 ret = set_request_path_attr(NULL, req->r_old_dentry,
1404 req->r_path2, req->r_ino2.ino,
1405 &path2, &pathlen2, &ino2, &freepath2);
1411 len = sizeof(*head) +
1412 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1414 /* calculate (max) length for cap releases */
1415 len += sizeof(struct ceph_mds_request_release) *
1416 (!!req->r_inode_drop + !!req->r_dentry_drop +
1417 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1418 if (req->r_dentry_drop)
1419 len += req->r_dentry->d_name.len;
1420 if (req->r_old_dentry_drop)
1421 len += req->r_old_dentry->d_name.len;
1423 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, 0, 0, NULL);
1427 msg->hdr.tid = cpu_to_le64(req->r_tid);
1429 head = msg->front.iov_base;
1430 p = msg->front.iov_base + sizeof(*head);
1431 end = msg->front.iov_base + msg->front.iov_len;
1433 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1434 head->op = cpu_to_le32(req->r_op);
1435 head->caller_uid = cpu_to_le32(current_fsuid());
1436 head->caller_gid = cpu_to_le32(current_fsgid());
1437 head->args = req->r_args;
1439 ceph_encode_filepath(&p, end, ino1, path1);
1440 ceph_encode_filepath(&p, end, ino2, path2);
1444 if (req->r_inode_drop)
1445 releases += ceph_encode_inode_release(&p,
1446 req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1447 mds, req->r_inode_drop, req->r_inode_unless, 0);
1448 if (req->r_dentry_drop)
1449 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1450 mds, req->r_dentry_drop, req->r_dentry_unless);
1451 if (req->r_old_dentry_drop)
1452 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1453 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1454 if (req->r_old_inode_drop)
1455 releases += ceph_encode_inode_release(&p,
1456 req->r_old_dentry->d_inode,
1457 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1458 head->num_releases = cpu_to_le16(releases);
1461 msg->front.iov_len = p - msg->front.iov_base;
1462 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1464 msg->pages = req->r_pages;
1465 msg->nr_pages = req->r_num_pages;
1466 msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1467 msg->hdr.data_off = cpu_to_le16(0);
1471 kfree((char *)path2);
1474 kfree((char *)path1);
1480 * called under mdsc->mutex if error, under no mutex if
1483 static void complete_request(struct ceph_mds_client *mdsc,
1484 struct ceph_mds_request *req)
1486 if (req->r_callback)
1487 req->r_callback(mdsc, req);
1489 complete(&req->r_completion);
1493 * called under mdsc->mutex
1495 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1496 struct ceph_mds_request *req,
1499 struct ceph_mds_request_head *rhead;
1500 struct ceph_msg *msg;
1505 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1506 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1508 if (req->r_request) {
1509 ceph_msg_put(req->r_request);
1510 req->r_request = NULL;
1512 msg = create_request_message(mdsc, req, mds);
1514 req->r_reply = ERR_PTR(PTR_ERR(msg));
1515 complete_request(mdsc, req);
1516 return -PTR_ERR(msg);
1518 req->r_request = msg;
1520 rhead = msg->front.iov_base;
1521 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1522 if (req->r_got_unsafe)
1523 flags |= CEPH_MDS_FLAG_REPLAY;
1524 if (req->r_locked_dir)
1525 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1526 rhead->flags = cpu_to_le32(flags);
1527 rhead->num_fwd = req->r_num_fwd;
1528 rhead->num_retry = req->r_attempts - 1;
1530 dout(" r_locked_dir = %p\n", req->r_locked_dir);
1532 if (req->r_target_inode && req->r_got_unsafe)
1533 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1540 * send request, or put it on the appropriate wait list.
1542 static int __do_request(struct ceph_mds_client *mdsc,
1543 struct ceph_mds_request *req)
1545 struct ceph_mds_session *session = NULL;
1552 if (req->r_timeout &&
1553 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1554 dout("do_request timed out\n");
1559 mds = __choose_mds(mdsc, req);
1561 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1562 dout("do_request no mds or not active, waiting for map\n");
1563 list_add(&req->r_wait, &mdsc->waiting_for_map);
1567 /* get, open session */
1568 session = __ceph_lookup_mds_session(mdsc, mds);
1570 session = register_session(mdsc, mds);
1571 dout("do_request mds%d session %p state %s\n", mds, session,
1572 session_state_name(session->s_state));
1573 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1574 session->s_state != CEPH_MDS_SESSION_HUNG) {
1575 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1576 session->s_state == CEPH_MDS_SESSION_CLOSING)
1577 __open_session(mdsc, session);
1578 list_add(&req->r_wait, &session->s_waiting);
1583 req->r_session = get_session(session);
1584 req->r_resend_mds = -1; /* forget any previous mds hint */
1586 if (req->r_request_started == 0) /* note request start time */
1587 req->r_request_started = jiffies;
1589 err = __prepare_send_request(mdsc, req, mds);
1591 ceph_msg_get(req->r_request);
1592 ceph_con_send(&session->s_con, req->r_request);
1596 ceph_put_mds_session(session);
1601 req->r_reply = ERR_PTR(err);
1602 complete_request(mdsc, req);
1607 * called under mdsc->mutex
1609 static void __wake_requests(struct ceph_mds_client *mdsc,
1610 struct list_head *head)
1612 struct ceph_mds_request *req, *nreq;
1614 list_for_each_entry_safe(req, nreq, head, r_wait) {
1615 list_del_init(&req->r_wait);
1616 __do_request(mdsc, req);
1621 * Wake up threads with requests pending for @mds, so that they can
1622 * resubmit their requests to a possibly different mds. If @all is set,
1623 * wake up if their requests has been forwarded to @mds, too.
1625 static void kick_requests(struct ceph_mds_client *mdsc, int mds, int all)
1627 struct ceph_mds_request *req;
1630 dout("kick_requests mds%d\n", mds);
1631 for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1632 req = rb_entry(p, struct ceph_mds_request, r_node);
1633 if (req->r_got_unsafe)
1635 if (req->r_session &&
1636 req->r_session->s_mds == mds) {
1637 dout(" kicking tid %llu\n", req->r_tid);
1638 put_request_session(req);
1639 __do_request(mdsc, req);
1644 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1645 struct ceph_mds_request *req)
1647 dout("submit_request on %p\n", req);
1648 mutex_lock(&mdsc->mutex);
1649 __register_request(mdsc, req, NULL);
1650 __do_request(mdsc, req);
1651 mutex_unlock(&mdsc->mutex);
1655 * Synchrously perform an mds request. Take care of all of the
1656 * session setup, forwarding, retry details.
1658 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1660 struct ceph_mds_request *req)
1664 dout("do_request on %p\n", req);
1666 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1668 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1669 if (req->r_locked_dir)
1670 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1671 if (req->r_old_dentry)
1673 ceph_inode(req->r_old_dentry->d_parent->d_inode),
1677 mutex_lock(&mdsc->mutex);
1678 __register_request(mdsc, req, dir);
1679 __do_request(mdsc, req);
1682 if (!req->r_reply) {
1683 mutex_unlock(&mdsc->mutex);
1684 if (req->r_timeout) {
1685 err = (long)wait_for_completion_interruptible_timeout(
1686 &req->r_completion, req->r_timeout);
1688 req->r_reply = ERR_PTR(-EIO);
1690 req->r_reply = ERR_PTR(err);
1692 err = wait_for_completion_interruptible(
1693 &req->r_completion);
1695 req->r_reply = ERR_PTR(err);
1697 mutex_lock(&mdsc->mutex);
1700 if (IS_ERR(req->r_reply)) {
1701 err = PTR_ERR(req->r_reply);
1702 req->r_reply = NULL;
1704 if (err == -ERESTARTSYS) {
1706 req->r_aborted = true;
1708 if (req->r_locked_dir &&
1709 (req->r_op & CEPH_MDS_OP_WRITE)) {
1710 struct ceph_inode_info *ci =
1711 ceph_inode(req->r_locked_dir);
1713 dout("aborted, clearing I_COMPLETE on %p\n",
1715 spin_lock(&req->r_locked_dir->i_lock);
1716 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1717 ci->i_release_count++;
1718 spin_unlock(&req->r_locked_dir->i_lock);
1721 /* clean up this request */
1722 __unregister_request(mdsc, req);
1723 if (!list_empty(&req->r_unsafe_item))
1724 list_del_init(&req->r_unsafe_item);
1725 complete(&req->r_safe_completion);
1727 } else if (req->r_err) {
1730 err = le32_to_cpu(req->r_reply_info.head->result);
1732 mutex_unlock(&mdsc->mutex);
1734 dout("do_request %p done, result %d\n", req, err);
1741 * We take the session mutex and parse and process the reply immediately.
1742 * This preserves the logical ordering of replies, capabilities, etc., sent
1743 * by the MDS as they are applied to our local cache.
1745 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1747 struct ceph_mds_client *mdsc = session->s_mdsc;
1748 struct ceph_mds_request *req;
1749 struct ceph_mds_reply_head *head = msg->front.iov_base;
1750 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
1753 int mds = session->s_mds;
1755 if (msg->front.iov_len < sizeof(*head)) {
1756 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1761 /* get request, session */
1762 tid = le64_to_cpu(msg->hdr.tid);
1763 mutex_lock(&mdsc->mutex);
1764 req = __lookup_request(mdsc, tid);
1766 dout("handle_reply on unknown tid %llu\n", tid);
1767 mutex_unlock(&mdsc->mutex);
1770 dout("handle_reply %p\n", req);
1772 /* correct session? */
1773 if (!req->r_session && req->r_session != session) {
1774 pr_err("mdsc_handle_reply got %llu on session mds%d"
1775 " not mds%d\n", tid, session->s_mds,
1776 req->r_session ? req->r_session->s_mds : -1);
1777 mutex_unlock(&mdsc->mutex);
1782 if ((req->r_got_unsafe && !head->safe) ||
1783 (req->r_got_safe && head->safe)) {
1784 pr_warning("got a dup %s reply on %llu from mds%d\n",
1785 head->safe ? "safe" : "unsafe", tid, mds);
1786 mutex_unlock(&mdsc->mutex);
1790 result = le32_to_cpu(head->result);
1793 * Tolerate 2 consecutive ESTALEs from the same mds.
1794 * FIXME: we should be looking at the cap migrate_seq.
1796 if (result == -ESTALE) {
1797 req->r_direct_mode = USE_AUTH_MDS;
1799 if (req->r_num_stale <= 2) {
1800 __do_request(mdsc, req);
1801 mutex_unlock(&mdsc->mutex);
1805 req->r_num_stale = 0;
1809 req->r_got_safe = true;
1810 __unregister_request(mdsc, req);
1811 complete(&req->r_safe_completion);
1813 if (req->r_got_unsafe) {
1815 * We already handled the unsafe response, now do the
1816 * cleanup. No need to examine the response; the MDS
1817 * doesn't include any result info in the safe
1818 * response. And even if it did, there is nothing
1819 * useful we could do with a revised return value.
1821 dout("got safe reply %llu, mds%d\n", tid, mds);
1822 list_del_init(&req->r_unsafe_item);
1824 /* last unsafe request during umount? */
1825 if (mdsc->stopping && !__get_oldest_req(mdsc))
1826 complete(&mdsc->safe_umount_waiters);
1827 mutex_unlock(&mdsc->mutex);
1832 BUG_ON(req->r_reply);
1835 req->r_got_unsafe = true;
1836 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1839 dout("handle_reply tid %lld result %d\n", tid, result);
1840 rinfo = &req->r_reply_info;
1841 err = parse_reply_info(msg, rinfo);
1842 mutex_unlock(&mdsc->mutex);
1844 mutex_lock(&session->s_mutex);
1846 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
1852 if (rinfo->snapblob_len) {
1853 down_write(&mdsc->snap_rwsem);
1854 ceph_update_snap_trace(mdsc, rinfo->snapblob,
1855 rinfo->snapblob + rinfo->snapblob_len,
1856 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
1857 downgrade_write(&mdsc->snap_rwsem);
1859 down_read(&mdsc->snap_rwsem);
1862 /* insert trace into our cache */
1863 err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
1865 if (result == 0 && rinfo->dir_nr)
1866 ceph_readdir_prepopulate(req, req->r_session);
1867 ceph_unreserve_caps(&req->r_caps_reservation);
1870 up_read(&mdsc->snap_rwsem);
1879 add_cap_releases(mdsc, req->r_session, -1);
1880 mutex_unlock(&session->s_mutex);
1882 /* kick calling process */
1883 complete_request(mdsc, req);
1885 ceph_mdsc_put_request(req);
1892 * handle mds notification that our request has been forwarded.
1894 static void handle_forward(struct ceph_mds_client *mdsc,
1895 struct ceph_mds_session *session,
1896 struct ceph_msg *msg)
1898 struct ceph_mds_request *req;
1899 u64 tid = le64_to_cpu(msg->hdr.tid);
1904 void *p = msg->front.iov_base;
1905 void *end = p + msg->front.iov_len;
1907 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1908 next_mds = ceph_decode_32(&p);
1909 fwd_seq = ceph_decode_32(&p);
1910 must_resend = ceph_decode_8(&p);
1912 WARN_ON(must_resend); /* shouldn't happen. */
1914 mutex_lock(&mdsc->mutex);
1915 req = __lookup_request(mdsc, tid);
1917 dout("forward %llu dne\n", tid);
1918 goto out; /* dup reply? */
1921 if (fwd_seq <= req->r_num_fwd) {
1922 dout("forward %llu to mds%d - old seq %d <= %d\n",
1923 tid, next_mds, req->r_num_fwd, fwd_seq);
1925 /* resend. forward race not possible; mds would drop */
1926 dout("forward %llu to mds%d (we resend)\n", tid, next_mds);
1927 req->r_num_fwd = fwd_seq;
1928 req->r_resend_mds = next_mds;
1929 put_request_session(req);
1930 __do_request(mdsc, req);
1932 ceph_mdsc_put_request(req);
1934 mutex_unlock(&mdsc->mutex);
1938 pr_err("mdsc_handle_forward decode error err=%d\n", err);
1942 * handle a mds session control message
1944 static void handle_session(struct ceph_mds_session *session,
1945 struct ceph_msg *msg)
1947 struct ceph_mds_client *mdsc = session->s_mdsc;
1950 int mds = session->s_mds;
1951 struct ceph_mds_session_head *h = msg->front.iov_base;
1955 if (msg->front.iov_len != sizeof(*h))
1957 op = le32_to_cpu(h->op);
1958 seq = le64_to_cpu(h->seq);
1960 mutex_lock(&mdsc->mutex);
1961 if (op == CEPH_SESSION_CLOSE)
1962 __unregister_session(mdsc, session);
1963 /* FIXME: this ttl calculation is generous */
1964 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
1965 mutex_unlock(&mdsc->mutex);
1967 mutex_lock(&session->s_mutex);
1969 dout("handle_session mds%d %s %p state %s seq %llu\n",
1970 mds, ceph_session_op_name(op), session,
1971 session_state_name(session->s_state), seq);
1973 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
1974 session->s_state = CEPH_MDS_SESSION_OPEN;
1975 pr_info("mds%d came back\n", session->s_mds);
1979 case CEPH_SESSION_OPEN:
1980 session->s_state = CEPH_MDS_SESSION_OPEN;
1981 renewed_caps(mdsc, session, 0);
1984 __close_session(mdsc, session);
1987 case CEPH_SESSION_RENEWCAPS:
1988 if (session->s_renew_seq == seq)
1989 renewed_caps(mdsc, session, 1);
1992 case CEPH_SESSION_CLOSE:
1993 remove_session_caps(session);
1994 wake = 1; /* for good measure */
1995 complete(&mdsc->session_close_waiters);
1996 kick_requests(mdsc, mds, 0); /* cur only */
1999 case CEPH_SESSION_STALE:
2000 pr_info("mds%d caps went stale, renewing\n",
2002 spin_lock(&session->s_cap_lock);
2003 session->s_cap_gen++;
2004 session->s_cap_ttl = 0;
2005 spin_unlock(&session->s_cap_lock);
2006 send_renew_caps(mdsc, session);
2009 case CEPH_SESSION_RECALL_STATE:
2010 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2014 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2018 mutex_unlock(&session->s_mutex);
2020 mutex_lock(&mdsc->mutex);
2021 __wake_requests(mdsc, &session->s_waiting);
2022 mutex_unlock(&mdsc->mutex);
2027 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2028 (int)msg->front.iov_len);
2035 * called under session->mutex.
2037 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2038 struct ceph_mds_session *session)
2040 struct ceph_mds_request *req, *nreq;
2043 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2045 mutex_lock(&mdsc->mutex);
2046 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2047 err = __prepare_send_request(mdsc, req, session->s_mds);
2049 ceph_msg_get(req->r_request);
2050 ceph_con_send(&session->s_con, req->r_request);
2053 mutex_unlock(&mdsc->mutex);
2057 * Encode information about a cap for a reconnect with the MDS.
2059 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2062 struct ceph_mds_cap_reconnect rec;
2063 struct ceph_inode_info *ci;
2064 struct ceph_pagelist *pagelist = arg;
2068 struct dentry *dentry;
2072 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2073 inode, ceph_vinop(inode), cap, cap->cap_id,
2074 ceph_cap_string(cap->issued));
2075 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2079 dentry = d_find_alias(inode);
2081 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2083 err = PTR_ERR(path);
2090 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2094 spin_lock(&inode->i_lock);
2095 cap->seq = 0; /* reset cap seq */
2096 cap->issue_seq = 0; /* and issue_seq */
2097 rec.cap_id = cpu_to_le64(cap->cap_id);
2098 rec.pathbase = cpu_to_le64(pathbase);
2099 rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2100 rec.issued = cpu_to_le32(cap->issued);
2101 rec.size = cpu_to_le64(inode->i_size);
2102 ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2103 ceph_encode_timespec(&rec.atime, &inode->i_atime);
2104 rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2105 spin_unlock(&inode->i_lock);
2107 err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2117 * If an MDS fails and recovers, clients need to reconnect in order to
2118 * reestablish shared state. This includes all caps issued through
2119 * this session _and_ the snap_realm hierarchy. Because it's not
2120 * clear which snap realms the mds cares about, we send everything we
2121 * know about.. that ensures we'll then get any new info the
2122 * recovering MDS might have.
2124 * This is a relatively heavyweight operation, but it's rare.
2126 * called with mdsc->mutex held.
2128 static void send_mds_reconnect(struct ceph_mds_client *mdsc, int mds)
2130 struct ceph_mds_session *session = NULL;
2131 struct ceph_msg *reply;
2134 struct ceph_pagelist *pagelist;
2136 pr_info("reconnect to recovering mds%d\n", mds);
2138 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2140 goto fail_nopagelist;
2141 ceph_pagelist_init(pagelist);
2143 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, 0, 0, NULL);
2144 if (IS_ERR(reply)) {
2145 err = PTR_ERR(reply);
2150 session = __ceph_lookup_mds_session(mdsc, mds);
2151 mutex_unlock(&mdsc->mutex); /* drop lock for duration */
2154 mutex_lock(&session->s_mutex);
2156 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2159 ceph_con_open(&session->s_con,
2160 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2162 /* replay unsafe requests */
2163 replay_unsafe_requests(mdsc, session);
2165 dout("no session for mds%d, will send short reconnect\n",
2169 down_read(&mdsc->snap_rwsem);
2173 dout("session %p state %s\n", session,
2174 session_state_name(session->s_state));
2176 /* traverse this session's caps */
2177 err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2180 err = iterate_session_caps(session, encode_caps_cb, pagelist);
2185 * snaprealms. we provide mds with the ino, seq (version), and
2186 * parent for all of our realms. If the mds has any newer info,
2189 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2190 struct ceph_snap_realm *realm =
2191 rb_entry(p, struct ceph_snap_realm, node);
2192 struct ceph_mds_snaprealm_reconnect sr_rec;
2194 dout(" adding snap realm %llx seq %lld parent %llx\n",
2195 realm->ino, realm->seq, realm->parent_ino);
2196 sr_rec.ino = cpu_to_le64(realm->ino);
2197 sr_rec.seq = cpu_to_le64(realm->seq);
2198 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2199 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2205 reply->pagelist = pagelist;
2206 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2207 reply->nr_pages = calc_pages_for(0, pagelist->length);
2208 ceph_con_send(&session->s_con, reply);
2211 session->s_state = CEPH_MDS_SESSION_OPEN;
2212 __wake_requests(mdsc, &session->s_waiting);
2216 up_read(&mdsc->snap_rwsem);
2218 mutex_unlock(&session->s_mutex);
2219 ceph_put_mds_session(session);
2221 mutex_lock(&mdsc->mutex);
2225 ceph_msg_put(reply);
2227 ceph_pagelist_release(pagelist);
2230 pr_err("ENOMEM preparing reconnect for mds%d\n", mds);
2236 * compare old and new mdsmaps, kicking requests
2237 * and closing out old connections as necessary
2239 * called under mdsc->mutex.
2241 static void check_new_map(struct ceph_mds_client *mdsc,
2242 struct ceph_mdsmap *newmap,
2243 struct ceph_mdsmap *oldmap)
2246 int oldstate, newstate;
2247 struct ceph_mds_session *s;
2249 dout("check_new_map new %u old %u\n",
2250 newmap->m_epoch, oldmap->m_epoch);
2252 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2253 if (mdsc->sessions[i] == NULL)
2255 s = mdsc->sessions[i];
2256 oldstate = ceph_mdsmap_get_state(oldmap, i);
2257 newstate = ceph_mdsmap_get_state(newmap, i);
2259 dout("check_new_map mds%d state %s -> %s (session %s)\n",
2260 i, ceph_mds_state_name(oldstate),
2261 ceph_mds_state_name(newstate),
2262 session_state_name(s->s_state));
2264 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2265 ceph_mdsmap_get_addr(newmap, i),
2266 sizeof(struct ceph_entity_addr))) {
2267 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2268 /* the session never opened, just close it
2270 __wake_requests(mdsc, &s->s_waiting);
2271 __unregister_session(mdsc, s);
2274 mutex_unlock(&mdsc->mutex);
2275 mutex_lock(&s->s_mutex);
2276 mutex_lock(&mdsc->mutex);
2277 ceph_con_close(&s->s_con);
2278 mutex_unlock(&s->s_mutex);
2279 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2282 /* kick any requests waiting on the recovering mds */
2283 kick_requests(mdsc, i, 1);
2284 } else if (oldstate == newstate) {
2285 continue; /* nothing new with this mds */
2291 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2292 newstate >= CEPH_MDS_STATE_RECONNECT)
2293 send_mds_reconnect(mdsc, i);
2296 * kick requests on any mds that has gone active.
2298 * kick requests on cur or forwarder: we may have sent
2299 * the request to mds1, mds1 told us it forwarded it
2300 * to mds2, but then we learn mds1 failed and can't be
2301 * sure it successfully forwarded our request before
2304 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2305 newstate >= CEPH_MDS_STATE_ACTIVE) {
2306 pr_info("mds%d reconnect completed\n", s->s_mds);
2307 kick_requests(mdsc, i, 1);
2308 ceph_kick_flushing_caps(mdsc, s);
2309 wake_up_session_caps(s, 1);
2321 * caller must hold session s_mutex, dentry->d_lock
2323 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2325 struct ceph_dentry_info *di = ceph_dentry(dentry);
2327 ceph_put_mds_session(di->lease_session);
2328 di->lease_session = NULL;
2331 static void handle_lease(struct ceph_mds_client *mdsc,
2332 struct ceph_mds_session *session,
2333 struct ceph_msg *msg)
2335 struct super_block *sb = mdsc->client->sb;
2336 struct inode *inode;
2337 struct ceph_inode_info *ci;
2338 struct dentry *parent, *dentry;
2339 struct ceph_dentry_info *di;
2340 int mds = session->s_mds;
2341 struct ceph_mds_lease *h = msg->front.iov_base;
2342 struct ceph_vino vino;
2347 dout("handle_lease from mds%d\n", mds);
2350 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2352 vino.ino = le64_to_cpu(h->ino);
2353 vino.snap = CEPH_NOSNAP;
2354 mask = le16_to_cpu(h->mask);
2355 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2356 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2357 if (dname.len != get_unaligned_le32(h+1))
2360 mutex_lock(&session->s_mutex);
2364 inode = ceph_find_inode(sb, vino);
2365 dout("handle_lease '%s', mask %d, ino %llx %p\n",
2366 ceph_lease_op_name(h->action), mask, vino.ino, inode);
2367 if (inode == NULL) {
2368 dout("handle_lease no inode %llx\n", vino.ino);
2371 ci = ceph_inode(inode);
2374 parent = d_find_alias(inode);
2376 dout("no parent dentry on inode %p\n", inode);
2378 goto release; /* hrm... */
2380 dname.hash = full_name_hash(dname.name, dname.len);
2381 dentry = d_lookup(parent, &dname);
2386 spin_lock(&dentry->d_lock);
2387 di = ceph_dentry(dentry);
2388 switch (h->action) {
2389 case CEPH_MDS_LEASE_REVOKE:
2390 if (di && di->lease_session == session) {
2391 h->seq = cpu_to_le32(di->lease_seq);
2392 __ceph_mdsc_drop_dentry_lease(dentry);
2397 case CEPH_MDS_LEASE_RENEW:
2398 if (di && di->lease_session == session &&
2399 di->lease_gen == session->s_cap_gen &&
2400 di->lease_renew_from &&
2401 di->lease_renew_after == 0) {
2402 unsigned long duration =
2403 le32_to_cpu(h->duration_ms) * HZ / 1000;
2405 di->lease_seq = le32_to_cpu(h->seq);
2406 dentry->d_time = di->lease_renew_from + duration;
2407 di->lease_renew_after = di->lease_renew_from +
2409 di->lease_renew_from = 0;
2413 spin_unlock(&dentry->d_lock);
2420 /* let's just reuse the same message */
2421 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2423 ceph_con_send(&session->s_con, msg);
2427 mutex_unlock(&session->s_mutex);
2431 pr_err("corrupt lease message\n");
2435 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2436 struct inode *inode,
2437 struct dentry *dentry, char action,
2440 struct ceph_msg *msg;
2441 struct ceph_mds_lease *lease;
2442 int len = sizeof(*lease) + sizeof(u32);
2445 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2446 inode, dentry, ceph_lease_op_name(action), session->s_mds);
2447 dnamelen = dentry->d_name.len;
2450 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, 0, 0, NULL);
2453 lease = msg->front.iov_base;
2454 lease->action = action;
2455 lease->mask = cpu_to_le16(CEPH_LOCK_DN);
2456 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2457 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2458 lease->seq = cpu_to_le32(seq);
2459 put_unaligned_le32(dnamelen, lease + 1);
2460 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2463 * if this is a preemptive lease RELEASE, no need to
2464 * flush request stream, since the actual request will
2467 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2469 ceph_con_send(&session->s_con, msg);
2473 * Preemptively release a lease we expect to invalidate anyway.
2474 * Pass @inode always, @dentry is optional.
2476 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2477 struct dentry *dentry, int mask)
2479 struct ceph_dentry_info *di;
2480 struct ceph_mds_session *session;
2483 BUG_ON(inode == NULL);
2484 BUG_ON(dentry == NULL);
2485 BUG_ON(mask != CEPH_LOCK_DN);
2487 /* is dentry lease valid? */
2488 spin_lock(&dentry->d_lock);
2489 di = ceph_dentry(dentry);
2490 if (!di || !di->lease_session ||
2491 di->lease_session->s_mds < 0 ||
2492 di->lease_gen != di->lease_session->s_cap_gen ||
2493 !time_before(jiffies, dentry->d_time)) {
2494 dout("lease_release inode %p dentry %p -- "
2496 inode, dentry, mask);
2497 spin_unlock(&dentry->d_lock);
2501 /* we do have a lease on this dentry; note mds and seq */
2502 session = ceph_get_mds_session(di->lease_session);
2503 seq = di->lease_seq;
2504 __ceph_mdsc_drop_dentry_lease(dentry);
2505 spin_unlock(&dentry->d_lock);
2507 dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2508 inode, dentry, mask, session->s_mds);
2509 ceph_mdsc_lease_send_msg(session, inode, dentry,
2510 CEPH_MDS_LEASE_RELEASE, seq);
2511 ceph_put_mds_session(session);
2515 * drop all leases (and dentry refs) in preparation for umount
2517 static void drop_leases(struct ceph_mds_client *mdsc)
2521 dout("drop_leases\n");
2522 mutex_lock(&mdsc->mutex);
2523 for (i = 0; i < mdsc->max_sessions; i++) {
2524 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2527 mutex_unlock(&mdsc->mutex);
2528 mutex_lock(&s->s_mutex);
2529 mutex_unlock(&s->s_mutex);
2530 ceph_put_mds_session(s);
2531 mutex_lock(&mdsc->mutex);
2533 mutex_unlock(&mdsc->mutex);
2539 * delayed work -- periodically trim expired leases, renew caps with mds
2541 static void schedule_delayed(struct ceph_mds_client *mdsc)
2544 unsigned hz = round_jiffies_relative(HZ * delay);
2545 schedule_delayed_work(&mdsc->delayed_work, hz);
2548 static void delayed_work(struct work_struct *work)
2551 struct ceph_mds_client *mdsc =
2552 container_of(work, struct ceph_mds_client, delayed_work.work);
2556 dout("mdsc delayed_work\n");
2557 ceph_check_delayed_caps(mdsc);
2559 mutex_lock(&mdsc->mutex);
2560 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2561 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2562 mdsc->last_renew_caps);
2564 mdsc->last_renew_caps = jiffies;
2566 for (i = 0; i < mdsc->max_sessions; i++) {
2567 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2570 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2571 dout("resending session close request for mds%d\n",
2573 request_close_session(mdsc, s);
2574 ceph_put_mds_session(s);
2577 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2578 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2579 s->s_state = CEPH_MDS_SESSION_HUNG;
2580 pr_info("mds%d hung\n", s->s_mds);
2583 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2584 /* this mds is failed or recovering, just wait */
2585 ceph_put_mds_session(s);
2588 mutex_unlock(&mdsc->mutex);
2590 mutex_lock(&s->s_mutex);
2592 send_renew_caps(mdsc, s);
2594 ceph_con_keepalive(&s->s_con);
2595 add_cap_releases(mdsc, s, -1);
2596 send_cap_releases(mdsc, s);
2597 mutex_unlock(&s->s_mutex);
2598 ceph_put_mds_session(s);
2600 mutex_lock(&mdsc->mutex);
2602 mutex_unlock(&mdsc->mutex);
2604 schedule_delayed(mdsc);
2608 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2610 mdsc->client = client;
2611 mutex_init(&mdsc->mutex);
2612 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2613 init_completion(&mdsc->safe_umount_waiters);
2614 init_completion(&mdsc->session_close_waiters);
2615 INIT_LIST_HEAD(&mdsc->waiting_for_map);
2616 mdsc->sessions = NULL;
2617 mdsc->max_sessions = 0;
2619 init_rwsem(&mdsc->snap_rwsem);
2620 mdsc->snap_realms = RB_ROOT;
2621 INIT_LIST_HEAD(&mdsc->snap_empty);
2622 spin_lock_init(&mdsc->snap_empty_lock);
2624 mdsc->request_tree = RB_ROOT;
2625 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2626 mdsc->last_renew_caps = jiffies;
2627 INIT_LIST_HEAD(&mdsc->cap_delay_list);
2628 spin_lock_init(&mdsc->cap_delay_lock);
2629 INIT_LIST_HEAD(&mdsc->snap_flush_list);
2630 spin_lock_init(&mdsc->snap_flush_lock);
2631 mdsc->cap_flush_seq = 0;
2632 INIT_LIST_HEAD(&mdsc->cap_dirty);
2633 mdsc->num_cap_flushing = 0;
2634 spin_lock_init(&mdsc->cap_dirty_lock);
2635 init_waitqueue_head(&mdsc->cap_flushing_wq);
2636 spin_lock_init(&mdsc->dentry_lru_lock);
2637 INIT_LIST_HEAD(&mdsc->dentry_lru);
2642 * Wait for safe replies on open mds requests. If we time out, drop
2643 * all requests from the tree to avoid dangling dentry refs.
2645 static void wait_requests(struct ceph_mds_client *mdsc)
2647 struct ceph_mds_request *req;
2648 struct ceph_client *client = mdsc->client;
2650 mutex_lock(&mdsc->mutex);
2651 if (__get_oldest_req(mdsc)) {
2652 mutex_unlock(&mdsc->mutex);
2654 dout("wait_requests waiting for requests\n");
2655 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2656 client->mount_args->mount_timeout * HZ);
2658 /* tear down remaining requests */
2659 mutex_lock(&mdsc->mutex);
2660 while ((req = __get_oldest_req(mdsc))) {
2661 dout("wait_requests timed out on tid %llu\n",
2663 __unregister_request(mdsc, req);
2666 mutex_unlock(&mdsc->mutex);
2667 dout("wait_requests done\n");
2671 * called before mount is ro, and before dentries are torn down.
2672 * (hmm, does this still race with new lookups?)
2674 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2676 dout("pre_umount\n");
2680 ceph_flush_dirty_caps(mdsc);
2681 wait_requests(mdsc);
2685 * wait for all write mds requests to flush.
2687 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2689 struct ceph_mds_request *req = NULL;
2692 mutex_lock(&mdsc->mutex);
2693 dout("wait_unsafe_requests want %lld\n", want_tid);
2694 req = __get_oldest_req(mdsc);
2695 while (req && req->r_tid <= want_tid) {
2696 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2698 ceph_mdsc_get_request(req);
2699 mutex_unlock(&mdsc->mutex);
2700 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
2701 req->r_tid, want_tid);
2702 wait_for_completion(&req->r_safe_completion);
2703 mutex_lock(&mdsc->mutex);
2704 n = rb_next(&req->r_node);
2705 ceph_mdsc_put_request(req);
2707 n = rb_next(&req->r_node);
2711 req = rb_entry(n, struct ceph_mds_request, r_node);
2713 mutex_unlock(&mdsc->mutex);
2714 dout("wait_unsafe_requests done\n");
2717 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2719 u64 want_tid, want_flush;
2722 mutex_lock(&mdsc->mutex);
2723 want_tid = mdsc->last_tid;
2724 want_flush = mdsc->cap_flush_seq;
2725 mutex_unlock(&mdsc->mutex);
2726 dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2728 ceph_flush_dirty_caps(mdsc);
2730 wait_unsafe_requests(mdsc, want_tid);
2731 wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2736 * called after sb is ro.
2738 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2740 struct ceph_mds_session *session;
2743 struct ceph_client *client = mdsc->client;
2744 unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2746 dout("close_sessions\n");
2748 mutex_lock(&mdsc->mutex);
2750 /* close sessions */
2752 while (time_before(jiffies, started + timeout)) {
2753 dout("closing sessions\n");
2755 for (i = 0; i < mdsc->max_sessions; i++) {
2756 session = __ceph_lookup_mds_session(mdsc, i);
2759 mutex_unlock(&mdsc->mutex);
2760 mutex_lock(&session->s_mutex);
2761 __close_session(mdsc, session);
2762 mutex_unlock(&session->s_mutex);
2763 ceph_put_mds_session(session);
2764 mutex_lock(&mdsc->mutex);
2770 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2773 dout("waiting for sessions to close\n");
2774 mutex_unlock(&mdsc->mutex);
2775 wait_for_completion_timeout(&mdsc->session_close_waiters,
2777 mutex_lock(&mdsc->mutex);
2780 /* tear down remaining sessions */
2781 for (i = 0; i < mdsc->max_sessions; i++) {
2782 if (mdsc->sessions[i]) {
2783 session = get_session(mdsc->sessions[i]);
2784 __unregister_session(mdsc, session);
2785 mutex_unlock(&mdsc->mutex);
2786 mutex_lock(&session->s_mutex);
2787 remove_session_caps(session);
2788 mutex_unlock(&session->s_mutex);
2789 ceph_put_mds_session(session);
2790 mutex_lock(&mdsc->mutex);
2794 WARN_ON(!list_empty(&mdsc->cap_delay_list));
2796 mutex_unlock(&mdsc->mutex);
2798 ceph_cleanup_empty_realms(mdsc);
2800 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2805 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
2808 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
2810 ceph_mdsmap_destroy(mdsc->mdsmap);
2811 kfree(mdsc->sessions);
2816 * handle mds map update.
2818 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
2822 void *p = msg->front.iov_base;
2823 void *end = p + msg->front.iov_len;
2824 struct ceph_mdsmap *newmap, *oldmap;
2825 struct ceph_fsid fsid;
2828 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
2829 ceph_decode_copy(&p, &fsid, sizeof(fsid));
2830 if (ceph_check_fsid(mdsc->client, &fsid) < 0)
2832 epoch = ceph_decode_32(&p);
2833 maplen = ceph_decode_32(&p);
2834 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
2836 /* do we need it? */
2837 ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
2838 mutex_lock(&mdsc->mutex);
2839 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
2840 dout("handle_map epoch %u <= our %u\n",
2841 epoch, mdsc->mdsmap->m_epoch);
2842 mutex_unlock(&mdsc->mutex);
2846 newmap = ceph_mdsmap_decode(&p, end);
2847 if (IS_ERR(newmap)) {
2848 err = PTR_ERR(newmap);
2852 /* swap into place */
2854 oldmap = mdsc->mdsmap;
2855 mdsc->mdsmap = newmap;
2856 check_new_map(mdsc, newmap, oldmap);
2857 ceph_mdsmap_destroy(oldmap);
2859 mdsc->mdsmap = newmap; /* first mds map */
2861 mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
2863 __wake_requests(mdsc, &mdsc->waiting_for_map);
2865 mutex_unlock(&mdsc->mutex);
2866 schedule_delayed(mdsc);
2870 mutex_unlock(&mdsc->mutex);
2872 pr_err("error decoding mdsmap %d\n", err);
2876 static struct ceph_connection *con_get(struct ceph_connection *con)
2878 struct ceph_mds_session *s = con->private;
2880 if (get_session(s)) {
2881 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
2884 dout("mdsc con_get %p FAIL\n", s);
2888 static void con_put(struct ceph_connection *con)
2890 struct ceph_mds_session *s = con->private;
2892 ceph_put_mds_session(s);
2893 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
2897 * if the client is unresponsive for long enough, the mds will kill
2898 * the session entirely.
2900 static void peer_reset(struct ceph_connection *con)
2902 struct ceph_mds_session *s = con->private;
2904 pr_err("mds%d gave us the boot. IMPLEMENT RECONNECT.\n",
2908 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2910 struct ceph_mds_session *s = con->private;
2911 struct ceph_mds_client *mdsc = s->s_mdsc;
2912 int type = le16_to_cpu(msg->hdr.type);
2914 mutex_lock(&mdsc->mutex);
2915 if (__verify_registered_session(mdsc, s) < 0) {
2916 mutex_unlock(&mdsc->mutex);
2919 mutex_unlock(&mdsc->mutex);
2922 case CEPH_MSG_MDS_MAP:
2923 ceph_mdsc_handle_map(mdsc, msg);
2925 case CEPH_MSG_CLIENT_SESSION:
2926 handle_session(s, msg);
2928 case CEPH_MSG_CLIENT_REPLY:
2929 handle_reply(s, msg);
2931 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
2932 handle_forward(mdsc, s, msg);
2934 case CEPH_MSG_CLIENT_CAPS:
2935 ceph_handle_caps(s, msg);
2937 case CEPH_MSG_CLIENT_SNAP:
2938 ceph_handle_snap(mdsc, s, msg);
2940 case CEPH_MSG_CLIENT_LEASE:
2941 handle_lease(mdsc, s, msg);
2945 pr_err("received unknown message type %d %s\n", type,
2946 ceph_msg_type_name(type));
2955 static int get_authorizer(struct ceph_connection *con,
2956 void **buf, int *len, int *proto,
2957 void **reply_buf, int *reply_len, int force_new)
2959 struct ceph_mds_session *s = con->private;
2960 struct ceph_mds_client *mdsc = s->s_mdsc;
2961 struct ceph_auth_client *ac = mdsc->client->monc.auth;
2964 if (force_new && s->s_authorizer) {
2965 ac->ops->destroy_authorizer(ac, s->s_authorizer);
2966 s->s_authorizer = NULL;
2968 if (s->s_authorizer == NULL) {
2969 if (ac->ops->create_authorizer) {
2970 ret = ac->ops->create_authorizer(
2971 ac, CEPH_ENTITY_TYPE_MDS,
2973 &s->s_authorizer_buf,
2974 &s->s_authorizer_buf_len,
2975 &s->s_authorizer_reply_buf,
2976 &s->s_authorizer_reply_buf_len);
2982 *proto = ac->protocol;
2983 *buf = s->s_authorizer_buf;
2984 *len = s->s_authorizer_buf_len;
2985 *reply_buf = s->s_authorizer_reply_buf;
2986 *reply_len = s->s_authorizer_reply_buf_len;
2991 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2993 struct ceph_mds_session *s = con->private;
2994 struct ceph_mds_client *mdsc = s->s_mdsc;
2995 struct ceph_auth_client *ac = mdsc->client->monc.auth;
2997 return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3000 static int invalidate_authorizer(struct ceph_connection *con)
3002 struct ceph_mds_session *s = con->private;
3003 struct ceph_mds_client *mdsc = s->s_mdsc;
3004 struct ceph_auth_client *ac = mdsc->client->monc.auth;
3006 if (ac->ops->invalidate_authorizer)
3007 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3009 return ceph_monc_validate_auth(&mdsc->client->monc);
3012 const static struct ceph_connection_operations mds_con_ops = {
3015 .dispatch = dispatch,
3016 .get_authorizer = get_authorizer,
3017 .verify_authorizer_reply = verify_authorizer_reply,
3018 .invalidate_authorizer = invalidate_authorizer,
3019 .peer_reset = peer_reset,