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37 * are no longer valid.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_mdsc, &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(mdsc, &req->r_caps_reservation,
517 dout("__register_request %p tid %lld\n", req, req->r_tid);
518 ceph_mdsc_get_request(req);
519 __insert_request(mdsc, req);
522 struct ceph_inode_info *ci = ceph_inode(dir);
524 spin_lock(&ci->i_unsafe_lock);
525 req->r_unsafe_dir = dir;
526 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
527 spin_unlock(&ci->i_unsafe_lock);
531 static void __unregister_request(struct ceph_mds_client *mdsc,
532 struct ceph_mds_request *req)
534 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
535 rb_erase(&req->r_node, &mdsc->request_tree);
536 RB_CLEAR_NODE(&req->r_node);
538 if (req->r_unsafe_dir) {
539 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
541 spin_lock(&ci->i_unsafe_lock);
542 list_del_init(&req->r_unsafe_dir_item);
543 spin_unlock(&ci->i_unsafe_lock);
546 ceph_mdsc_put_request(req);
550 * Choose mds to send request to next. If there is a hint set in the
551 * request (e.g., due to a prior forward hint from the mds), use that.
552 * Otherwise, consult frag tree and/or caps to identify the
553 * appropriate mds. If all else fails, choose randomly.
555 * Called under mdsc->mutex.
557 static int __choose_mds(struct ceph_mds_client *mdsc,
558 struct ceph_mds_request *req)
561 struct ceph_inode_info *ci;
562 struct ceph_cap *cap;
563 int mode = req->r_direct_mode;
565 u32 hash = req->r_direct_hash;
566 bool is_hash = req->r_direct_is_hash;
569 * is there a specific mds we should try? ignore hint if we have
570 * no session and the mds is not up (active or recovering).
572 if (req->r_resend_mds >= 0 &&
573 (__have_session(mdsc, req->r_resend_mds) ||
574 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
575 dout("choose_mds using resend_mds mds%d\n",
577 return req->r_resend_mds;
580 if (mode == USE_RANDOM_MDS)
585 inode = req->r_inode;
586 } else if (req->r_dentry) {
587 if (req->r_dentry->d_inode) {
588 inode = req->r_dentry->d_inode;
590 inode = req->r_dentry->d_parent->d_inode;
591 hash = req->r_dentry->d_name.hash;
595 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
599 ci = ceph_inode(inode);
601 if (is_hash && S_ISDIR(inode->i_mode)) {
602 struct ceph_inode_frag frag;
605 ceph_choose_frag(ci, hash, &frag, &found);
607 if (mode == USE_ANY_MDS && frag.ndist > 0) {
610 /* choose a random replica */
611 get_random_bytes(&r, 1);
614 dout("choose_mds %p %llx.%llx "
615 "frag %u mds%d (%d/%d)\n",
616 inode, ceph_vinop(inode),
622 /* since this file/dir wasn't known to be
623 * replicated, then we want to look for the
624 * authoritative mds. */
627 /* choose auth mds */
629 dout("choose_mds %p %llx.%llx "
630 "frag %u mds%d (auth)\n",
631 inode, ceph_vinop(inode), frag.frag, mds);
637 spin_lock(&inode->i_lock);
639 if (mode == USE_AUTH_MDS)
640 cap = ci->i_auth_cap;
641 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
642 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
644 spin_unlock(&inode->i_lock);
647 mds = cap->session->s_mds;
648 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
649 inode, ceph_vinop(inode), mds,
650 cap == ci->i_auth_cap ? "auth " : "", cap);
651 spin_unlock(&inode->i_lock);
655 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
656 dout("choose_mds chose random mds%d\n", mds);
664 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
666 struct ceph_msg *msg;
667 struct ceph_mds_session_head *h;
669 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS);
671 pr_err("create_session_msg ENOMEM creating msg\n");
674 h = msg->front.iov_base;
675 h->op = cpu_to_le32(op);
676 h->seq = cpu_to_le64(seq);
681 * send session open request.
683 * called under mdsc->mutex
685 static int __open_session(struct ceph_mds_client *mdsc,
686 struct ceph_mds_session *session)
688 struct ceph_msg *msg;
690 int mds = session->s_mds;
692 /* wait for mds to go active? */
693 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
694 dout("open_session to mds%d (%s)\n", mds,
695 ceph_mds_state_name(mstate));
696 session->s_state = CEPH_MDS_SESSION_OPENING;
697 session->s_renew_requested = jiffies;
699 /* send connect message */
700 msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
703 ceph_con_send(&session->s_con, msg);
708 * open sessions for any export targets for the given mds
710 * called under mdsc->mutex
712 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
713 struct ceph_mds_session *session)
715 struct ceph_mds_info *mi;
716 struct ceph_mds_session *ts;
717 int i, mds = session->s_mds;
720 if (mds >= mdsc->mdsmap->m_max_mds)
722 mi = &mdsc->mdsmap->m_info[mds];
723 dout("open_export_target_sessions for mds%d (%d targets)\n",
724 session->s_mds, mi->num_export_targets);
726 for (i = 0; i < mi->num_export_targets; i++) {
727 target = mi->export_targets[i];
728 ts = __ceph_lookup_mds_session(mdsc, target);
730 ts = register_session(mdsc, target);
734 if (session->s_state == CEPH_MDS_SESSION_NEW ||
735 session->s_state == CEPH_MDS_SESSION_CLOSING)
736 __open_session(mdsc, session);
738 dout(" mds%d target mds%d %p is %s\n", session->s_mds,
739 i, ts, session_state_name(ts->s_state));
740 ceph_put_mds_session(ts);
749 * Free preallocated cap messages assigned to this session
751 static void cleanup_cap_releases(struct ceph_mds_session *session)
753 struct ceph_msg *msg;
755 spin_lock(&session->s_cap_lock);
756 while (!list_empty(&session->s_cap_releases)) {
757 msg = list_first_entry(&session->s_cap_releases,
758 struct ceph_msg, list_head);
759 list_del_init(&msg->list_head);
762 while (!list_empty(&session->s_cap_releases_done)) {
763 msg = list_first_entry(&session->s_cap_releases_done,
764 struct ceph_msg, list_head);
765 list_del_init(&msg->list_head);
768 spin_unlock(&session->s_cap_lock);
772 * Helper to safely iterate over all caps associated with a session, with
773 * special care taken to handle a racing __ceph_remove_cap().
775 * Caller must hold session s_mutex.
777 static int iterate_session_caps(struct ceph_mds_session *session,
778 int (*cb)(struct inode *, struct ceph_cap *,
782 struct ceph_cap *cap;
783 struct inode *inode, *last_inode = NULL;
784 struct ceph_cap *old_cap = NULL;
787 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
788 spin_lock(&session->s_cap_lock);
789 p = session->s_caps.next;
790 while (p != &session->s_caps) {
791 cap = list_entry(p, struct ceph_cap, session_caps);
792 inode = igrab(&cap->ci->vfs_inode);
797 session->s_cap_iterator = cap;
798 spin_unlock(&session->s_cap_lock);
805 ceph_put_cap(session->s_mdsc, old_cap);
809 ret = cb(inode, cap, arg);
812 spin_lock(&session->s_cap_lock);
814 if (cap->ci == NULL) {
815 dout("iterate_session_caps finishing cap %p removal\n",
817 BUG_ON(cap->session != session);
818 list_del_init(&cap->session_caps);
819 session->s_nr_caps--;
821 old_cap = cap; /* put_cap it w/o locks held */
828 session->s_cap_iterator = NULL;
829 spin_unlock(&session->s_cap_lock);
834 ceph_put_cap(session->s_mdsc, old_cap);
839 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
842 struct ceph_inode_info *ci = ceph_inode(inode);
845 dout("removing cap %p, ci is %p, inode is %p\n",
846 cap, ci, &ci->vfs_inode);
847 spin_lock(&inode->i_lock);
848 __ceph_remove_cap(cap);
849 if (!__ceph_is_any_real_caps(ci)) {
850 struct ceph_mds_client *mdsc =
851 &ceph_sb_to_client(inode->i_sb)->mdsc;
853 spin_lock(&mdsc->cap_dirty_lock);
854 if (!list_empty(&ci->i_dirty_item)) {
855 pr_info(" dropping dirty %s state for %p %lld\n",
856 ceph_cap_string(ci->i_dirty_caps),
857 inode, ceph_ino(inode));
858 ci->i_dirty_caps = 0;
859 list_del_init(&ci->i_dirty_item);
862 if (!list_empty(&ci->i_flushing_item)) {
863 pr_info(" dropping dirty+flushing %s state for %p %lld\n",
864 ceph_cap_string(ci->i_flushing_caps),
865 inode, ceph_ino(inode));
866 ci->i_flushing_caps = 0;
867 list_del_init(&ci->i_flushing_item);
868 mdsc->num_cap_flushing--;
871 if (drop && ci->i_wrbuffer_ref) {
872 pr_info(" dropping dirty data for %p %lld\n",
873 inode, ceph_ino(inode));
874 ci->i_wrbuffer_ref = 0;
875 ci->i_wrbuffer_ref_head = 0;
878 spin_unlock(&mdsc->cap_dirty_lock);
880 spin_unlock(&inode->i_lock);
887 * caller must hold session s_mutex
889 static void remove_session_caps(struct ceph_mds_session *session)
891 dout("remove_session_caps on %p\n", session);
892 iterate_session_caps(session, remove_session_caps_cb, NULL);
893 BUG_ON(session->s_nr_caps > 0);
894 BUG_ON(!list_empty(&session->s_cap_flushing));
895 cleanup_cap_releases(session);
899 * wake up any threads waiting on this session's caps. if the cap is
900 * old (didn't get renewed on the client reconnect), remove it now.
902 * caller must hold s_mutex.
904 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
907 struct ceph_inode_info *ci = ceph_inode(inode);
909 wake_up_all(&ci->i_cap_wq);
911 spin_lock(&inode->i_lock);
912 ci->i_wanted_max_size = 0;
913 ci->i_requested_max_size = 0;
914 spin_unlock(&inode->i_lock);
919 static void wake_up_session_caps(struct ceph_mds_session *session,
922 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
923 iterate_session_caps(session, wake_up_session_cb,
924 (void *)(unsigned long)reconnect);
928 * Send periodic message to MDS renewing all currently held caps. The
929 * ack will reset the expiration for all caps from this session.
931 * caller holds s_mutex
933 static int send_renew_caps(struct ceph_mds_client *mdsc,
934 struct ceph_mds_session *session)
936 struct ceph_msg *msg;
939 if (time_after_eq(jiffies, session->s_cap_ttl) &&
940 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
941 pr_info("mds%d caps stale\n", session->s_mds);
942 session->s_renew_requested = jiffies;
944 /* do not try to renew caps until a recovering mds has reconnected
945 * with its clients. */
946 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
947 if (state < CEPH_MDS_STATE_RECONNECT) {
948 dout("send_renew_caps ignoring mds%d (%s)\n",
949 session->s_mds, ceph_mds_state_name(state));
953 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
954 ceph_mds_state_name(state));
955 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
956 ++session->s_renew_seq);
959 ceph_con_send(&session->s_con, msg);
964 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
966 * Called under session->s_mutex
968 static void renewed_caps(struct ceph_mds_client *mdsc,
969 struct ceph_mds_session *session, int is_renew)
974 spin_lock(&session->s_cap_lock);
975 was_stale = is_renew && (session->s_cap_ttl == 0 ||
976 time_after_eq(jiffies, session->s_cap_ttl));
978 session->s_cap_ttl = session->s_renew_requested +
979 mdsc->mdsmap->m_session_timeout*HZ;
982 if (time_before(jiffies, session->s_cap_ttl)) {
983 pr_info("mds%d caps renewed\n", session->s_mds);
986 pr_info("mds%d caps still stale\n", session->s_mds);
989 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
990 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
991 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
992 spin_unlock(&session->s_cap_lock);
995 wake_up_session_caps(session, 0);
999 * send a session close request
1001 static int request_close_session(struct ceph_mds_client *mdsc,
1002 struct ceph_mds_session *session)
1004 struct ceph_msg *msg;
1006 dout("request_close_session mds%d state %s seq %lld\n",
1007 session->s_mds, session_state_name(session->s_state),
1009 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1012 ceph_con_send(&session->s_con, msg);
1017 * Called with s_mutex held.
1019 static int __close_session(struct ceph_mds_client *mdsc,
1020 struct ceph_mds_session *session)
1022 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1024 session->s_state = CEPH_MDS_SESSION_CLOSING;
1025 return request_close_session(mdsc, session);
1029 * Trim old(er) caps.
1031 * Because we can't cache an inode without one or more caps, we do
1032 * this indirectly: if a cap is unused, we prune its aliases, at which
1033 * point the inode will hopefully get dropped to.
1035 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1036 * memory pressure from the MDS, though, so it needn't be perfect.
1038 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1040 struct ceph_mds_session *session = arg;
1041 struct ceph_inode_info *ci = ceph_inode(inode);
1042 int used, oissued, mine;
1044 if (session->s_trim_caps <= 0)
1047 spin_lock(&inode->i_lock);
1048 mine = cap->issued | cap->implemented;
1049 used = __ceph_caps_used(ci);
1050 oissued = __ceph_caps_issued_other(ci, cap);
1052 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1053 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1054 ceph_cap_string(used));
1055 if (ci->i_dirty_caps)
1056 goto out; /* dirty caps */
1057 if ((used & ~oissued) & mine)
1058 goto out; /* we need these caps */
1060 session->s_trim_caps--;
1062 /* we aren't the only cap.. just remove us */
1063 __ceph_remove_cap(cap);
1065 /* try to drop referring dentries */
1066 spin_unlock(&inode->i_lock);
1067 d_prune_aliases(inode);
1068 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1069 inode, cap, atomic_read(&inode->i_count));
1074 spin_unlock(&inode->i_lock);
1079 * Trim session cap count down to some max number.
1081 static int trim_caps(struct ceph_mds_client *mdsc,
1082 struct ceph_mds_session *session,
1085 int trim_caps = session->s_nr_caps - max_caps;
1087 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1088 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1089 if (trim_caps > 0) {
1090 session->s_trim_caps = trim_caps;
1091 iterate_session_caps(session, trim_caps_cb, session);
1092 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1093 session->s_mds, session->s_nr_caps, max_caps,
1094 trim_caps - session->s_trim_caps);
1095 session->s_trim_caps = 0;
1101 * Allocate cap_release messages. If there is a partially full message
1102 * in the queue, try to allocate enough to cover it's remainder, so that
1103 * we can send it immediately.
1105 * Called under s_mutex.
1107 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1108 struct ceph_mds_session *session)
1110 struct ceph_msg *msg, *partial = NULL;
1111 struct ceph_mds_cap_release *head;
1113 int extra = mdsc->client->mount_args->cap_release_safety;
1116 dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1119 spin_lock(&session->s_cap_lock);
1121 if (!list_empty(&session->s_cap_releases)) {
1122 msg = list_first_entry(&session->s_cap_releases,
1125 head = msg->front.iov_base;
1126 num = le32_to_cpu(head->num);
1128 dout(" partial %p with (%d/%d)\n", msg, num,
1129 (int)CEPH_CAPS_PER_RELEASE);
1130 extra += CEPH_CAPS_PER_RELEASE - num;
1134 while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1135 spin_unlock(&session->s_cap_lock);
1136 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1140 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1141 (int)msg->front.iov_len);
1142 head = msg->front.iov_base;
1143 head->num = cpu_to_le32(0);
1144 msg->front.iov_len = sizeof(*head);
1145 spin_lock(&session->s_cap_lock);
1146 list_add(&msg->list_head, &session->s_cap_releases);
1147 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1151 head = partial->front.iov_base;
1152 num = le32_to_cpu(head->num);
1153 dout(" queueing partial %p with %d/%d\n", partial, num,
1154 (int)CEPH_CAPS_PER_RELEASE);
1155 list_move_tail(&partial->list_head,
1156 &session->s_cap_releases_done);
1157 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1160 spin_unlock(&session->s_cap_lock);
1166 * flush all dirty inode data to disk.
1168 * returns true if we've flushed through want_flush_seq
1170 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1174 dout("check_cap_flush want %lld\n", want_flush_seq);
1175 mutex_lock(&mdsc->mutex);
1176 for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1177 struct ceph_mds_session *session = mdsc->sessions[mds];
1181 get_session(session);
1182 mutex_unlock(&mdsc->mutex);
1184 mutex_lock(&session->s_mutex);
1185 if (!list_empty(&session->s_cap_flushing)) {
1186 struct ceph_inode_info *ci =
1187 list_entry(session->s_cap_flushing.next,
1188 struct ceph_inode_info,
1190 struct inode *inode = &ci->vfs_inode;
1192 spin_lock(&inode->i_lock);
1193 if (ci->i_cap_flush_seq <= want_flush_seq) {
1194 dout("check_cap_flush still flushing %p "
1195 "seq %lld <= %lld to mds%d\n", inode,
1196 ci->i_cap_flush_seq, want_flush_seq,
1200 spin_unlock(&inode->i_lock);
1202 mutex_unlock(&session->s_mutex);
1203 ceph_put_mds_session(session);
1207 mutex_lock(&mdsc->mutex);
1210 mutex_unlock(&mdsc->mutex);
1211 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1216 * called under s_mutex
1218 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1219 struct ceph_mds_session *session)
1221 struct ceph_msg *msg;
1223 dout("send_cap_releases mds%d\n", session->s_mds);
1224 spin_lock(&session->s_cap_lock);
1225 while (!list_empty(&session->s_cap_releases_done)) {
1226 msg = list_first_entry(&session->s_cap_releases_done,
1227 struct ceph_msg, list_head);
1228 list_del_init(&msg->list_head);
1229 spin_unlock(&session->s_cap_lock);
1230 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1231 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1232 ceph_con_send(&session->s_con, msg);
1233 spin_lock(&session->s_cap_lock);
1235 spin_unlock(&session->s_cap_lock);
1238 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1239 struct ceph_mds_session *session)
1241 struct ceph_msg *msg;
1242 struct ceph_mds_cap_release *head;
1245 dout("discard_cap_releases mds%d\n", session->s_mds);
1246 spin_lock(&session->s_cap_lock);
1248 /* zero out the in-progress message */
1249 msg = list_first_entry(&session->s_cap_releases,
1250 struct ceph_msg, list_head);
1251 head = msg->front.iov_base;
1252 num = le32_to_cpu(head->num);
1253 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1254 head->num = cpu_to_le32(0);
1255 session->s_num_cap_releases += num;
1257 /* requeue completed messages */
1258 while (!list_empty(&session->s_cap_releases_done)) {
1259 msg = list_first_entry(&session->s_cap_releases_done,
1260 struct ceph_msg, list_head);
1261 list_del_init(&msg->list_head);
1263 head = msg->front.iov_base;
1264 num = le32_to_cpu(head->num);
1265 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1267 session->s_num_cap_releases += num;
1268 head->num = cpu_to_le32(0);
1269 msg->front.iov_len = sizeof(*head);
1270 list_add(&msg->list_head, &session->s_cap_releases);
1273 spin_unlock(&session->s_cap_lock);
1281 * Create an mds request.
1283 struct ceph_mds_request *
1284 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1286 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1289 return ERR_PTR(-ENOMEM);
1291 mutex_init(&req->r_fill_mutex);
1293 req->r_started = jiffies;
1294 req->r_resend_mds = -1;
1295 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1297 kref_init(&req->r_kref);
1298 INIT_LIST_HEAD(&req->r_wait);
1299 init_completion(&req->r_completion);
1300 init_completion(&req->r_safe_completion);
1301 INIT_LIST_HEAD(&req->r_unsafe_item);
1304 req->r_direct_mode = mode;
1309 * return oldest (lowest) request, tid in request tree, 0 if none.
1311 * called under mdsc->mutex.
1313 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1315 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1317 return rb_entry(rb_first(&mdsc->request_tree),
1318 struct ceph_mds_request, r_node);
1321 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1323 struct ceph_mds_request *req = __get_oldest_req(mdsc);
1331 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1332 * on build_path_from_dentry in fs/cifs/dir.c.
1334 * If @stop_on_nosnap, generate path relative to the first non-snapped
1337 * Encode hidden .snap dirs as a double /, i.e.
1338 * foo/.snap/bar -> foo//bar
1340 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1343 struct dentry *temp;
1348 return ERR_PTR(-EINVAL);
1352 for (temp = dentry; !IS_ROOT(temp);) {
1353 struct inode *inode = temp->d_inode;
1354 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1355 len++; /* slash only */
1356 else if (stop_on_nosnap && inode &&
1357 ceph_snap(inode) == CEPH_NOSNAP)
1360 len += 1 + temp->d_name.len;
1361 temp = temp->d_parent;
1363 pr_err("build_path corrupt dentry %p\n", dentry);
1364 return ERR_PTR(-EINVAL);
1368 len--; /* no leading '/' */
1370 path = kmalloc(len+1, GFP_NOFS);
1372 return ERR_PTR(-ENOMEM);
1374 path[pos] = 0; /* trailing null */
1375 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1376 struct inode *inode = temp->d_inode;
1378 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1379 dout("build_path path+%d: %p SNAPDIR\n",
1381 } else if (stop_on_nosnap && inode &&
1382 ceph_snap(inode) == CEPH_NOSNAP) {
1385 pos -= temp->d_name.len;
1388 strncpy(path + pos, temp->d_name.name,
1393 temp = temp->d_parent;
1395 pr_err("build_path corrupt dentry\n");
1397 return ERR_PTR(-EINVAL);
1401 pr_err("build_path did not end path lookup where "
1402 "expected, namelen is %d, pos is %d\n", len, pos);
1403 /* presumably this is only possible if racing with a
1404 rename of one of the parent directories (we can not
1405 lock the dentries above us to prevent this, but
1406 retrying should be harmless) */
1411 *base = ceph_ino(temp->d_inode);
1413 dout("build_path on %p %d built %llx '%.*s'\n",
1414 dentry, atomic_read(&dentry->d_count), *base, len, path);
1418 static int build_dentry_path(struct dentry *dentry,
1419 const char **ppath, int *ppathlen, u64 *pino,
1424 if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1425 *pino = ceph_ino(dentry->d_parent->d_inode);
1426 *ppath = dentry->d_name.name;
1427 *ppathlen = dentry->d_name.len;
1430 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1432 return PTR_ERR(path);
1438 static int build_inode_path(struct inode *inode,
1439 const char **ppath, int *ppathlen, u64 *pino,
1442 struct dentry *dentry;
1445 if (ceph_snap(inode) == CEPH_NOSNAP) {
1446 *pino = ceph_ino(inode);
1450 dentry = d_find_alias(inode);
1451 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1454 return PTR_ERR(path);
1461 * request arguments may be specified via an inode *, a dentry *, or
1462 * an explicit ino+path.
1464 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1465 const char *rpath, u64 rino,
1466 const char **ppath, int *pathlen,
1467 u64 *ino, int *freepath)
1472 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1473 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1475 } else if (rdentry) {
1476 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1477 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1482 *pathlen = strlen(rpath);
1483 dout(" path %.*s\n", *pathlen, rpath);
1490 * called under mdsc->mutex
1492 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1493 struct ceph_mds_request *req,
1496 struct ceph_msg *msg;
1497 struct ceph_mds_request_head *head;
1498 const char *path1 = NULL;
1499 const char *path2 = NULL;
1500 u64 ino1 = 0, ino2 = 0;
1501 int pathlen1 = 0, pathlen2 = 0;
1502 int freepath1 = 0, freepath2 = 0;
1508 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1509 req->r_path1, req->r_ino1.ino,
1510 &path1, &pathlen1, &ino1, &freepath1);
1516 ret = set_request_path_attr(NULL, req->r_old_dentry,
1517 req->r_path2, req->r_ino2.ino,
1518 &path2, &pathlen2, &ino2, &freepath2);
1524 len = sizeof(*head) +
1525 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1527 /* calculate (max) length for cap releases */
1528 len += sizeof(struct ceph_mds_request_release) *
1529 (!!req->r_inode_drop + !!req->r_dentry_drop +
1530 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1531 if (req->r_dentry_drop)
1532 len += req->r_dentry->d_name.len;
1533 if (req->r_old_dentry_drop)
1534 len += req->r_old_dentry->d_name.len;
1536 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS);
1538 msg = ERR_PTR(-ENOMEM);
1542 msg->hdr.tid = cpu_to_le64(req->r_tid);
1544 head = msg->front.iov_base;
1545 p = msg->front.iov_base + sizeof(*head);
1546 end = msg->front.iov_base + msg->front.iov_len;
1548 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1549 head->op = cpu_to_le32(req->r_op);
1550 head->caller_uid = cpu_to_le32(current_fsuid());
1551 head->caller_gid = cpu_to_le32(current_fsgid());
1552 head->args = req->r_args;
1554 ceph_encode_filepath(&p, end, ino1, path1);
1555 ceph_encode_filepath(&p, end, ino2, path2);
1557 /* make note of release offset, in case we need to replay */
1558 req->r_request_release_offset = p - msg->front.iov_base;
1562 if (req->r_inode_drop)
1563 releases += ceph_encode_inode_release(&p,
1564 req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1565 mds, req->r_inode_drop, req->r_inode_unless, 0);
1566 if (req->r_dentry_drop)
1567 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1568 mds, req->r_dentry_drop, req->r_dentry_unless);
1569 if (req->r_old_dentry_drop)
1570 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1571 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1572 if (req->r_old_inode_drop)
1573 releases += ceph_encode_inode_release(&p,
1574 req->r_old_dentry->d_inode,
1575 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1576 head->num_releases = cpu_to_le16(releases);
1579 msg->front.iov_len = p - msg->front.iov_base;
1580 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1582 msg->pages = req->r_pages;
1583 msg->nr_pages = req->r_num_pages;
1584 msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1585 msg->hdr.data_off = cpu_to_le16(0);
1589 kfree((char *)path2);
1592 kfree((char *)path1);
1598 * called under mdsc->mutex if error, under no mutex if
1601 static void complete_request(struct ceph_mds_client *mdsc,
1602 struct ceph_mds_request *req)
1604 if (req->r_callback)
1605 req->r_callback(mdsc, req);
1607 complete_all(&req->r_completion);
1611 * called under mdsc->mutex
1613 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1614 struct ceph_mds_request *req,
1617 struct ceph_mds_request_head *rhead;
1618 struct ceph_msg *msg;
1623 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1624 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1626 if (req->r_got_unsafe) {
1628 * Replay. Do not regenerate message (and rebuild
1629 * paths, etc.); just use the original message.
1630 * Rebuilding paths will break for renames because
1631 * d_move mangles the src name.
1633 msg = req->r_request;
1634 rhead = msg->front.iov_base;
1636 flags = le32_to_cpu(rhead->flags);
1637 flags |= CEPH_MDS_FLAG_REPLAY;
1638 rhead->flags = cpu_to_le32(flags);
1640 if (req->r_target_inode)
1641 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1643 rhead->num_retry = req->r_attempts - 1;
1645 /* remove cap/dentry releases from message */
1646 rhead->num_releases = 0;
1647 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1648 msg->front.iov_len = req->r_request_release_offset;
1652 if (req->r_request) {
1653 ceph_msg_put(req->r_request);
1654 req->r_request = NULL;
1656 msg = create_request_message(mdsc, req, mds);
1658 req->r_err = PTR_ERR(msg);
1659 complete_request(mdsc, req);
1660 return PTR_ERR(msg);
1662 req->r_request = msg;
1664 rhead = msg->front.iov_base;
1665 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1666 if (req->r_got_unsafe)
1667 flags |= CEPH_MDS_FLAG_REPLAY;
1668 if (req->r_locked_dir)
1669 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1670 rhead->flags = cpu_to_le32(flags);
1671 rhead->num_fwd = req->r_num_fwd;
1672 rhead->num_retry = req->r_attempts - 1;
1675 dout(" r_locked_dir = %p\n", req->r_locked_dir);
1680 * send request, or put it on the appropriate wait list.
1682 static int __do_request(struct ceph_mds_client *mdsc,
1683 struct ceph_mds_request *req)
1685 struct ceph_mds_session *session = NULL;
1689 if (req->r_err || req->r_got_result)
1692 if (req->r_timeout &&
1693 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1694 dout("do_request timed out\n");
1699 mds = __choose_mds(mdsc, req);
1701 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1702 dout("do_request no mds or not active, waiting for map\n");
1703 list_add(&req->r_wait, &mdsc->waiting_for_map);
1707 /* get, open session */
1708 session = __ceph_lookup_mds_session(mdsc, mds);
1710 session = register_session(mdsc, mds);
1711 if (IS_ERR(session)) {
1712 err = PTR_ERR(session);
1716 dout("do_request mds%d session %p state %s\n", mds, session,
1717 session_state_name(session->s_state));
1718 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1719 session->s_state != CEPH_MDS_SESSION_HUNG) {
1720 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1721 session->s_state == CEPH_MDS_SESSION_CLOSING)
1722 __open_session(mdsc, session);
1723 list_add(&req->r_wait, &session->s_waiting);
1728 req->r_session = get_session(session);
1729 req->r_resend_mds = -1; /* forget any previous mds hint */
1731 if (req->r_request_started == 0) /* note request start time */
1732 req->r_request_started = jiffies;
1734 err = __prepare_send_request(mdsc, req, mds);
1736 ceph_msg_get(req->r_request);
1737 ceph_con_send(&session->s_con, req->r_request);
1741 ceph_put_mds_session(session);
1747 complete_request(mdsc, req);
1752 * called under mdsc->mutex
1754 static void __wake_requests(struct ceph_mds_client *mdsc,
1755 struct list_head *head)
1757 struct ceph_mds_request *req, *nreq;
1759 list_for_each_entry_safe(req, nreq, head, r_wait) {
1760 list_del_init(&req->r_wait);
1761 __do_request(mdsc, req);
1766 * Wake up threads with requests pending for @mds, so that they can
1767 * resubmit their requests to a possibly different mds.
1769 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1771 struct ceph_mds_request *req;
1774 dout("kick_requests mds%d\n", mds);
1775 for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1776 req = rb_entry(p, struct ceph_mds_request, r_node);
1777 if (req->r_got_unsafe)
1779 if (req->r_session &&
1780 req->r_session->s_mds == mds) {
1781 dout(" kicking tid %llu\n", req->r_tid);
1782 put_request_session(req);
1783 __do_request(mdsc, req);
1788 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1789 struct ceph_mds_request *req)
1791 dout("submit_request on %p\n", req);
1792 mutex_lock(&mdsc->mutex);
1793 __register_request(mdsc, req, NULL);
1794 __do_request(mdsc, req);
1795 mutex_unlock(&mdsc->mutex);
1799 * Synchrously perform an mds request. Take care of all of the
1800 * session setup, forwarding, retry details.
1802 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1804 struct ceph_mds_request *req)
1808 dout("do_request on %p\n", req);
1810 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1812 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1813 if (req->r_locked_dir)
1814 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1815 if (req->r_old_dentry)
1817 ceph_inode(req->r_old_dentry->d_parent->d_inode),
1821 mutex_lock(&mdsc->mutex);
1822 __register_request(mdsc, req, dir);
1823 __do_request(mdsc, req);
1827 __unregister_request(mdsc, req);
1828 dout("do_request early error %d\n", err);
1833 mutex_unlock(&mdsc->mutex);
1834 dout("do_request waiting\n");
1835 if (req->r_timeout) {
1836 err = (long)wait_for_completion_killable_timeout(
1837 &req->r_completion, req->r_timeout);
1841 err = wait_for_completion_killable(&req->r_completion);
1843 dout("do_request waited, got %d\n", err);
1844 mutex_lock(&mdsc->mutex);
1846 /* only abort if we didn't race with a real reply */
1847 if (req->r_got_result) {
1848 err = le32_to_cpu(req->r_reply_info.head->result);
1849 } else if (err < 0) {
1850 dout("aborted request %lld with %d\n", req->r_tid, err);
1853 * ensure we aren't running concurrently with
1854 * ceph_fill_trace or ceph_readdir_prepopulate, which
1855 * rely on locks (dir mutex) held by our caller.
1857 mutex_lock(&req->r_fill_mutex);
1859 req->r_aborted = true;
1860 mutex_unlock(&req->r_fill_mutex);
1862 if (req->r_locked_dir &&
1863 (req->r_op & CEPH_MDS_OP_WRITE))
1864 ceph_invalidate_dir_request(req);
1870 mutex_unlock(&mdsc->mutex);
1871 dout("do_request %p done, result %d\n", req, err);
1876 * Invalidate dir I_COMPLETE, dentry lease state on an aborted MDS
1877 * namespace request.
1879 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
1881 struct inode *inode = req->r_locked_dir;
1882 struct ceph_inode_info *ci = ceph_inode(inode);
1884 dout("invalidate_dir_request %p (I_COMPLETE, lease(s))\n", inode);
1885 spin_lock(&inode->i_lock);
1886 ci->i_ceph_flags &= ~CEPH_I_COMPLETE;
1887 ci->i_release_count++;
1888 spin_unlock(&inode->i_lock);
1891 ceph_invalidate_dentry_lease(req->r_dentry);
1892 if (req->r_old_dentry)
1893 ceph_invalidate_dentry_lease(req->r_old_dentry);
1899 * We take the session mutex and parse and process the reply immediately.
1900 * This preserves the logical ordering of replies, capabilities, etc., sent
1901 * by the MDS as they are applied to our local cache.
1903 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
1905 struct ceph_mds_client *mdsc = session->s_mdsc;
1906 struct ceph_mds_request *req;
1907 struct ceph_mds_reply_head *head = msg->front.iov_base;
1908 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
1911 int mds = session->s_mds;
1913 if (msg->front.iov_len < sizeof(*head)) {
1914 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
1919 /* get request, session */
1920 tid = le64_to_cpu(msg->hdr.tid);
1921 mutex_lock(&mdsc->mutex);
1922 req = __lookup_request(mdsc, tid);
1924 dout("handle_reply on unknown tid %llu\n", tid);
1925 mutex_unlock(&mdsc->mutex);
1928 dout("handle_reply %p\n", req);
1930 /* correct session? */
1931 if (req->r_session != session) {
1932 pr_err("mdsc_handle_reply got %llu on session mds%d"
1933 " not mds%d\n", tid, session->s_mds,
1934 req->r_session ? req->r_session->s_mds : -1);
1935 mutex_unlock(&mdsc->mutex);
1940 if ((req->r_got_unsafe && !head->safe) ||
1941 (req->r_got_safe && head->safe)) {
1942 pr_warning("got a dup %s reply on %llu from mds%d\n",
1943 head->safe ? "safe" : "unsafe", tid, mds);
1944 mutex_unlock(&mdsc->mutex);
1947 if (req->r_got_safe && !head->safe) {
1948 pr_warning("got unsafe after safe on %llu from mds%d\n",
1950 mutex_unlock(&mdsc->mutex);
1954 result = le32_to_cpu(head->result);
1957 * Tolerate 2 consecutive ESTALEs from the same mds.
1958 * FIXME: we should be looking at the cap migrate_seq.
1960 if (result == -ESTALE) {
1961 req->r_direct_mode = USE_AUTH_MDS;
1963 if (req->r_num_stale <= 2) {
1964 __do_request(mdsc, req);
1965 mutex_unlock(&mdsc->mutex);
1969 req->r_num_stale = 0;
1973 req->r_got_safe = true;
1974 __unregister_request(mdsc, req);
1975 complete_all(&req->r_safe_completion);
1977 if (req->r_got_unsafe) {
1979 * We already handled the unsafe response, now do the
1980 * cleanup. No need to examine the response; the MDS
1981 * doesn't include any result info in the safe
1982 * response. And even if it did, there is nothing
1983 * useful we could do with a revised return value.
1985 dout("got safe reply %llu, mds%d\n", tid, mds);
1986 list_del_init(&req->r_unsafe_item);
1988 /* last unsafe request during umount? */
1989 if (mdsc->stopping && !__get_oldest_req(mdsc))
1990 complete_all(&mdsc->safe_umount_waiters);
1991 mutex_unlock(&mdsc->mutex);
1995 req->r_got_unsafe = true;
1996 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
1999 dout("handle_reply tid %lld result %d\n", tid, result);
2000 rinfo = &req->r_reply_info;
2001 err = parse_reply_info(msg, rinfo);
2002 mutex_unlock(&mdsc->mutex);
2004 mutex_lock(&session->s_mutex);
2006 pr_err("mdsc_handle_reply got corrupt reply mds%d\n", mds);
2012 if (rinfo->snapblob_len) {
2013 down_write(&mdsc->snap_rwsem);
2014 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2015 rinfo->snapblob + rinfo->snapblob_len,
2016 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2017 downgrade_write(&mdsc->snap_rwsem);
2019 down_read(&mdsc->snap_rwsem);
2022 /* insert trace into our cache */
2023 mutex_lock(&req->r_fill_mutex);
2024 err = ceph_fill_trace(mdsc->client->sb, req, req->r_session);
2026 if (result == 0 && rinfo->dir_nr)
2027 ceph_readdir_prepopulate(req, req->r_session);
2028 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2030 mutex_unlock(&req->r_fill_mutex);
2032 up_read(&mdsc->snap_rwsem);
2034 mutex_lock(&mdsc->mutex);
2035 if (!req->r_aborted) {
2041 req->r_got_result = true;
2044 dout("reply arrived after request %lld was aborted\n", tid);
2046 mutex_unlock(&mdsc->mutex);
2048 ceph_add_cap_releases(mdsc, req->r_session);
2049 mutex_unlock(&session->s_mutex);
2051 /* kick calling process */
2052 complete_request(mdsc, req);
2054 ceph_mdsc_put_request(req);
2061 * handle mds notification that our request has been forwarded.
2063 static void handle_forward(struct ceph_mds_client *mdsc,
2064 struct ceph_mds_session *session,
2065 struct ceph_msg *msg)
2067 struct ceph_mds_request *req;
2068 u64 tid = le64_to_cpu(msg->hdr.tid);
2072 void *p = msg->front.iov_base;
2073 void *end = p + msg->front.iov_len;
2075 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2076 next_mds = ceph_decode_32(&p);
2077 fwd_seq = ceph_decode_32(&p);
2079 mutex_lock(&mdsc->mutex);
2080 req = __lookup_request(mdsc, tid);
2082 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2083 goto out; /* dup reply? */
2086 if (req->r_aborted) {
2087 dout("forward tid %llu aborted, unregistering\n", tid);
2088 __unregister_request(mdsc, req);
2089 } else if (fwd_seq <= req->r_num_fwd) {
2090 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2091 tid, next_mds, req->r_num_fwd, fwd_seq);
2093 /* resend. forward race not possible; mds would drop */
2094 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2096 BUG_ON(req->r_got_result);
2097 req->r_num_fwd = fwd_seq;
2098 req->r_resend_mds = next_mds;
2099 put_request_session(req);
2100 __do_request(mdsc, req);
2102 ceph_mdsc_put_request(req);
2104 mutex_unlock(&mdsc->mutex);
2108 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2112 * handle a mds session control message
2114 static void handle_session(struct ceph_mds_session *session,
2115 struct ceph_msg *msg)
2117 struct ceph_mds_client *mdsc = session->s_mdsc;
2120 int mds = session->s_mds;
2121 struct ceph_mds_session_head *h = msg->front.iov_base;
2125 if (msg->front.iov_len != sizeof(*h))
2127 op = le32_to_cpu(h->op);
2128 seq = le64_to_cpu(h->seq);
2130 mutex_lock(&mdsc->mutex);
2131 if (op == CEPH_SESSION_CLOSE)
2132 __unregister_session(mdsc, session);
2133 /* FIXME: this ttl calculation is generous */
2134 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2135 mutex_unlock(&mdsc->mutex);
2137 mutex_lock(&session->s_mutex);
2139 dout("handle_session mds%d %s %p state %s seq %llu\n",
2140 mds, ceph_session_op_name(op), session,
2141 session_state_name(session->s_state), seq);
2143 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2144 session->s_state = CEPH_MDS_SESSION_OPEN;
2145 pr_info("mds%d came back\n", session->s_mds);
2149 case CEPH_SESSION_OPEN:
2150 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2151 pr_info("mds%d reconnect success\n", session->s_mds);
2152 session->s_state = CEPH_MDS_SESSION_OPEN;
2153 renewed_caps(mdsc, session, 0);
2156 __close_session(mdsc, session);
2159 case CEPH_SESSION_RENEWCAPS:
2160 if (session->s_renew_seq == seq)
2161 renewed_caps(mdsc, session, 1);
2164 case CEPH_SESSION_CLOSE:
2165 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2166 pr_info("mds%d reconnect denied\n", session->s_mds);
2167 remove_session_caps(session);
2168 wake = 1; /* for good measure */
2169 complete_all(&mdsc->session_close_waiters);
2170 kick_requests(mdsc, mds);
2173 case CEPH_SESSION_STALE:
2174 pr_info("mds%d caps went stale, renewing\n",
2176 spin_lock(&session->s_cap_lock);
2177 session->s_cap_gen++;
2178 session->s_cap_ttl = 0;
2179 spin_unlock(&session->s_cap_lock);
2180 send_renew_caps(mdsc, session);
2183 case CEPH_SESSION_RECALL_STATE:
2184 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2188 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2192 mutex_unlock(&session->s_mutex);
2194 mutex_lock(&mdsc->mutex);
2195 __wake_requests(mdsc, &session->s_waiting);
2196 mutex_unlock(&mdsc->mutex);
2201 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2202 (int)msg->front.iov_len);
2209 * called under session->mutex.
2211 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2212 struct ceph_mds_session *session)
2214 struct ceph_mds_request *req, *nreq;
2217 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2219 mutex_lock(&mdsc->mutex);
2220 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2221 err = __prepare_send_request(mdsc, req, session->s_mds);
2223 ceph_msg_get(req->r_request);
2224 ceph_con_send(&session->s_con, req->r_request);
2227 mutex_unlock(&mdsc->mutex);
2231 * Encode information about a cap for a reconnect with the MDS.
2233 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2236 struct ceph_mds_cap_reconnect rec;
2237 struct ceph_inode_info *ci;
2238 struct ceph_pagelist *pagelist = arg;
2242 struct dentry *dentry;
2246 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2247 inode, ceph_vinop(inode), cap, cap->cap_id,
2248 ceph_cap_string(cap->issued));
2249 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2253 dentry = d_find_alias(inode);
2255 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2257 err = PTR_ERR(path);
2264 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2268 spin_lock(&inode->i_lock);
2269 cap->seq = 0; /* reset cap seq */
2270 cap->issue_seq = 0; /* and issue_seq */
2271 rec.cap_id = cpu_to_le64(cap->cap_id);
2272 rec.pathbase = cpu_to_le64(pathbase);
2273 rec.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2274 rec.issued = cpu_to_le32(cap->issued);
2275 rec.size = cpu_to_le64(inode->i_size);
2276 ceph_encode_timespec(&rec.mtime, &inode->i_mtime);
2277 ceph_encode_timespec(&rec.atime, &inode->i_atime);
2278 rec.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2279 spin_unlock(&inode->i_lock);
2281 err = ceph_pagelist_append(pagelist, &rec, sizeof(rec));
2291 * If an MDS fails and recovers, clients need to reconnect in order to
2292 * reestablish shared state. This includes all caps issued through
2293 * this session _and_ the snap_realm hierarchy. Because it's not
2294 * clear which snap realms the mds cares about, we send everything we
2295 * know about.. that ensures we'll then get any new info the
2296 * recovering MDS might have.
2298 * This is a relatively heavyweight operation, but it's rare.
2300 * called with mdsc->mutex held.
2302 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2303 struct ceph_mds_session *session)
2305 struct ceph_msg *reply;
2307 int mds = session->s_mds;
2309 struct ceph_pagelist *pagelist;
2311 pr_info("mds%d reconnect start\n", mds);
2313 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2315 goto fail_nopagelist;
2316 ceph_pagelist_init(pagelist);
2318 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS);
2322 mutex_lock(&session->s_mutex);
2323 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2326 ceph_con_open(&session->s_con,
2327 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2329 /* replay unsafe requests */
2330 replay_unsafe_requests(mdsc, session);
2332 down_read(&mdsc->snap_rwsem);
2334 dout("session %p state %s\n", session,
2335 session_state_name(session->s_state));
2337 /* drop old cap expires; we're about to reestablish that state */
2338 discard_cap_releases(mdsc, session);
2340 /* traverse this session's caps */
2341 err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2344 err = iterate_session_caps(session, encode_caps_cb, pagelist);
2349 * snaprealms. we provide mds with the ino, seq (version), and
2350 * parent for all of our realms. If the mds has any newer info,
2353 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2354 struct ceph_snap_realm *realm =
2355 rb_entry(p, struct ceph_snap_realm, node);
2356 struct ceph_mds_snaprealm_reconnect sr_rec;
2358 dout(" adding snap realm %llx seq %lld parent %llx\n",
2359 realm->ino, realm->seq, realm->parent_ino);
2360 sr_rec.ino = cpu_to_le64(realm->ino);
2361 sr_rec.seq = cpu_to_le64(realm->seq);
2362 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2363 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2368 reply->pagelist = pagelist;
2369 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2370 reply->nr_pages = calc_pages_for(0, pagelist->length);
2371 ceph_con_send(&session->s_con, reply);
2373 mutex_unlock(&session->s_mutex);
2375 mutex_lock(&mdsc->mutex);
2376 __wake_requests(mdsc, &session->s_waiting);
2377 mutex_unlock(&mdsc->mutex);
2379 up_read(&mdsc->snap_rwsem);
2383 ceph_msg_put(reply);
2384 up_read(&mdsc->snap_rwsem);
2385 mutex_unlock(&session->s_mutex);
2387 ceph_pagelist_release(pagelist);
2390 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2396 * compare old and new mdsmaps, kicking requests
2397 * and closing out old connections as necessary
2399 * called under mdsc->mutex.
2401 static void check_new_map(struct ceph_mds_client *mdsc,
2402 struct ceph_mdsmap *newmap,
2403 struct ceph_mdsmap *oldmap)
2406 int oldstate, newstate;
2407 struct ceph_mds_session *s;
2409 dout("check_new_map new %u old %u\n",
2410 newmap->m_epoch, oldmap->m_epoch);
2412 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2413 if (mdsc->sessions[i] == NULL)
2415 s = mdsc->sessions[i];
2416 oldstate = ceph_mdsmap_get_state(oldmap, i);
2417 newstate = ceph_mdsmap_get_state(newmap, i);
2419 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2420 i, ceph_mds_state_name(oldstate),
2421 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2422 ceph_mds_state_name(newstate),
2423 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2424 session_state_name(s->s_state));
2426 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2427 ceph_mdsmap_get_addr(newmap, i),
2428 sizeof(struct ceph_entity_addr))) {
2429 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2430 /* the session never opened, just close it
2432 __wake_requests(mdsc, &s->s_waiting);
2433 __unregister_session(mdsc, s);
2436 mutex_unlock(&mdsc->mutex);
2437 mutex_lock(&s->s_mutex);
2438 mutex_lock(&mdsc->mutex);
2439 ceph_con_close(&s->s_con);
2440 mutex_unlock(&s->s_mutex);
2441 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2444 /* kick any requests waiting on the recovering mds */
2445 kick_requests(mdsc, i);
2446 } else if (oldstate == newstate) {
2447 continue; /* nothing new with this mds */
2453 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2454 newstate >= CEPH_MDS_STATE_RECONNECT) {
2455 mutex_unlock(&mdsc->mutex);
2456 send_mds_reconnect(mdsc, s);
2457 mutex_lock(&mdsc->mutex);
2461 * kick request on any mds that has gone active.
2463 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2464 newstate >= CEPH_MDS_STATE_ACTIVE) {
2465 if (oldstate != CEPH_MDS_STATE_CREATING &&
2466 oldstate != CEPH_MDS_STATE_STARTING)
2467 pr_info("mds%d recovery completed\n", s->s_mds);
2468 kick_requests(mdsc, i);
2469 ceph_kick_flushing_caps(mdsc, s);
2470 wake_up_session_caps(s, 1);
2474 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2475 s = mdsc->sessions[i];
2478 if (!ceph_mdsmap_is_laggy(newmap, i))
2480 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2481 s->s_state == CEPH_MDS_SESSION_HUNG ||
2482 s->s_state == CEPH_MDS_SESSION_CLOSING) {
2483 dout(" connecting to export targets of laggy mds%d\n",
2485 __open_export_target_sessions(mdsc, s);
2497 * caller must hold session s_mutex, dentry->d_lock
2499 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2501 struct ceph_dentry_info *di = ceph_dentry(dentry);
2503 ceph_put_mds_session(di->lease_session);
2504 di->lease_session = NULL;
2507 static void handle_lease(struct ceph_mds_client *mdsc,
2508 struct ceph_mds_session *session,
2509 struct ceph_msg *msg)
2511 struct super_block *sb = mdsc->client->sb;
2512 struct inode *inode;
2513 struct ceph_inode_info *ci;
2514 struct dentry *parent, *dentry;
2515 struct ceph_dentry_info *di;
2516 int mds = session->s_mds;
2517 struct ceph_mds_lease *h = msg->front.iov_base;
2519 struct ceph_vino vino;
2524 dout("handle_lease from mds%d\n", mds);
2527 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2529 vino.ino = le64_to_cpu(h->ino);
2530 vino.snap = CEPH_NOSNAP;
2531 mask = le16_to_cpu(h->mask);
2532 seq = le32_to_cpu(h->seq);
2533 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2534 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2535 if (dname.len != get_unaligned_le32(h+1))
2538 mutex_lock(&session->s_mutex);
2542 inode = ceph_find_inode(sb, vino);
2543 dout("handle_lease %s, mask %d, ino %llx %p %.*s\n",
2544 ceph_lease_op_name(h->action), mask, vino.ino, inode,
2545 dname.len, dname.name);
2546 if (inode == NULL) {
2547 dout("handle_lease no inode %llx\n", vino.ino);
2550 ci = ceph_inode(inode);
2553 parent = d_find_alias(inode);
2555 dout("no parent dentry on inode %p\n", inode);
2557 goto release; /* hrm... */
2559 dname.hash = full_name_hash(dname.name, dname.len);
2560 dentry = d_lookup(parent, &dname);
2565 spin_lock(&dentry->d_lock);
2566 di = ceph_dentry(dentry);
2567 switch (h->action) {
2568 case CEPH_MDS_LEASE_REVOKE:
2569 if (di && di->lease_session == session) {
2570 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2571 h->seq = cpu_to_le32(di->lease_seq);
2572 __ceph_mdsc_drop_dentry_lease(dentry);
2577 case CEPH_MDS_LEASE_RENEW:
2578 if (di && di->lease_session == session &&
2579 di->lease_gen == session->s_cap_gen &&
2580 di->lease_renew_from &&
2581 di->lease_renew_after == 0) {
2582 unsigned long duration =
2583 le32_to_cpu(h->duration_ms) * HZ / 1000;
2585 di->lease_seq = seq;
2586 dentry->d_time = di->lease_renew_from + duration;
2587 di->lease_renew_after = di->lease_renew_from +
2589 di->lease_renew_from = 0;
2593 spin_unlock(&dentry->d_lock);
2600 /* let's just reuse the same message */
2601 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2603 ceph_con_send(&session->s_con, msg);
2607 mutex_unlock(&session->s_mutex);
2611 pr_err("corrupt lease message\n");
2615 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2616 struct inode *inode,
2617 struct dentry *dentry, char action,
2620 struct ceph_msg *msg;
2621 struct ceph_mds_lease *lease;
2622 int len = sizeof(*lease) + sizeof(u32);
2625 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2626 inode, dentry, ceph_lease_op_name(action), session->s_mds);
2627 dnamelen = dentry->d_name.len;
2630 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS);
2633 lease = msg->front.iov_base;
2634 lease->action = action;
2635 lease->mask = cpu_to_le16(1);
2636 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2637 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2638 lease->seq = cpu_to_le32(seq);
2639 put_unaligned_le32(dnamelen, lease + 1);
2640 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2643 * if this is a preemptive lease RELEASE, no need to
2644 * flush request stream, since the actual request will
2647 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2649 ceph_con_send(&session->s_con, msg);
2653 * Preemptively release a lease we expect to invalidate anyway.
2654 * Pass @inode always, @dentry is optional.
2656 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2657 struct dentry *dentry, int mask)
2659 struct ceph_dentry_info *di;
2660 struct ceph_mds_session *session;
2663 BUG_ON(inode == NULL);
2664 BUG_ON(dentry == NULL);
2667 /* is dentry lease valid? */
2668 spin_lock(&dentry->d_lock);
2669 di = ceph_dentry(dentry);
2670 if (!di || !di->lease_session ||
2671 di->lease_session->s_mds < 0 ||
2672 di->lease_gen != di->lease_session->s_cap_gen ||
2673 !time_before(jiffies, dentry->d_time)) {
2674 dout("lease_release inode %p dentry %p -- "
2676 inode, dentry, mask);
2677 spin_unlock(&dentry->d_lock);
2681 /* we do have a lease on this dentry; note mds and seq */
2682 session = ceph_get_mds_session(di->lease_session);
2683 seq = di->lease_seq;
2684 __ceph_mdsc_drop_dentry_lease(dentry);
2685 spin_unlock(&dentry->d_lock);
2687 dout("lease_release inode %p dentry %p mask %d to mds%d\n",
2688 inode, dentry, mask, session->s_mds);
2689 ceph_mdsc_lease_send_msg(session, inode, dentry,
2690 CEPH_MDS_LEASE_RELEASE, seq);
2691 ceph_put_mds_session(session);
2695 * drop all leases (and dentry refs) in preparation for umount
2697 static void drop_leases(struct ceph_mds_client *mdsc)
2701 dout("drop_leases\n");
2702 mutex_lock(&mdsc->mutex);
2703 for (i = 0; i < mdsc->max_sessions; i++) {
2704 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2707 mutex_unlock(&mdsc->mutex);
2708 mutex_lock(&s->s_mutex);
2709 mutex_unlock(&s->s_mutex);
2710 ceph_put_mds_session(s);
2711 mutex_lock(&mdsc->mutex);
2713 mutex_unlock(&mdsc->mutex);
2719 * delayed work -- periodically trim expired leases, renew caps with mds
2721 static void schedule_delayed(struct ceph_mds_client *mdsc)
2724 unsigned hz = round_jiffies_relative(HZ * delay);
2725 schedule_delayed_work(&mdsc->delayed_work, hz);
2728 static void delayed_work(struct work_struct *work)
2731 struct ceph_mds_client *mdsc =
2732 container_of(work, struct ceph_mds_client, delayed_work.work);
2736 dout("mdsc delayed_work\n");
2737 ceph_check_delayed_caps(mdsc);
2739 mutex_lock(&mdsc->mutex);
2740 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2741 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2742 mdsc->last_renew_caps);
2744 mdsc->last_renew_caps = jiffies;
2746 for (i = 0; i < mdsc->max_sessions; i++) {
2747 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2750 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2751 dout("resending session close request for mds%d\n",
2753 request_close_session(mdsc, s);
2754 ceph_put_mds_session(s);
2757 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2758 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2759 s->s_state = CEPH_MDS_SESSION_HUNG;
2760 pr_info("mds%d hung\n", s->s_mds);
2763 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2764 /* this mds is failed or recovering, just wait */
2765 ceph_put_mds_session(s);
2768 mutex_unlock(&mdsc->mutex);
2770 mutex_lock(&s->s_mutex);
2772 send_renew_caps(mdsc, s);
2774 ceph_con_keepalive(&s->s_con);
2775 ceph_add_cap_releases(mdsc, s);
2776 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2777 s->s_state == CEPH_MDS_SESSION_HUNG)
2778 ceph_send_cap_releases(mdsc, s);
2779 mutex_unlock(&s->s_mutex);
2780 ceph_put_mds_session(s);
2782 mutex_lock(&mdsc->mutex);
2784 mutex_unlock(&mdsc->mutex);
2786 schedule_delayed(mdsc);
2790 int ceph_mdsc_init(struct ceph_mds_client *mdsc, struct ceph_client *client)
2792 mdsc->client = client;
2793 mutex_init(&mdsc->mutex);
2794 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
2795 if (mdsc->mdsmap == NULL)
2798 init_completion(&mdsc->safe_umount_waiters);
2799 init_completion(&mdsc->session_close_waiters);
2800 INIT_LIST_HEAD(&mdsc->waiting_for_map);
2801 mdsc->sessions = NULL;
2802 mdsc->max_sessions = 0;
2804 init_rwsem(&mdsc->snap_rwsem);
2805 mdsc->snap_realms = RB_ROOT;
2806 INIT_LIST_HEAD(&mdsc->snap_empty);
2807 spin_lock_init(&mdsc->snap_empty_lock);
2809 mdsc->request_tree = RB_ROOT;
2810 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
2811 mdsc->last_renew_caps = jiffies;
2812 INIT_LIST_HEAD(&mdsc->cap_delay_list);
2813 spin_lock_init(&mdsc->cap_delay_lock);
2814 INIT_LIST_HEAD(&mdsc->snap_flush_list);
2815 spin_lock_init(&mdsc->snap_flush_lock);
2816 mdsc->cap_flush_seq = 0;
2817 INIT_LIST_HEAD(&mdsc->cap_dirty);
2818 mdsc->num_cap_flushing = 0;
2819 spin_lock_init(&mdsc->cap_dirty_lock);
2820 init_waitqueue_head(&mdsc->cap_flushing_wq);
2821 spin_lock_init(&mdsc->dentry_lru_lock);
2822 INIT_LIST_HEAD(&mdsc->dentry_lru);
2824 ceph_caps_init(mdsc);
2825 ceph_adjust_min_caps(mdsc, client->min_caps);
2831 * Wait for safe replies on open mds requests. If we time out, drop
2832 * all requests from the tree to avoid dangling dentry refs.
2834 static void wait_requests(struct ceph_mds_client *mdsc)
2836 struct ceph_mds_request *req;
2837 struct ceph_client *client = mdsc->client;
2839 mutex_lock(&mdsc->mutex);
2840 if (__get_oldest_req(mdsc)) {
2841 mutex_unlock(&mdsc->mutex);
2843 dout("wait_requests waiting for requests\n");
2844 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
2845 client->mount_args->mount_timeout * HZ);
2847 /* tear down remaining requests */
2848 mutex_lock(&mdsc->mutex);
2849 while ((req = __get_oldest_req(mdsc))) {
2850 dout("wait_requests timed out on tid %llu\n",
2852 __unregister_request(mdsc, req);
2855 mutex_unlock(&mdsc->mutex);
2856 dout("wait_requests done\n");
2860 * called before mount is ro, and before dentries are torn down.
2861 * (hmm, does this still race with new lookups?)
2863 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
2865 dout("pre_umount\n");
2869 ceph_flush_dirty_caps(mdsc);
2870 wait_requests(mdsc);
2873 * wait for reply handlers to drop their request refs and
2874 * their inode/dcache refs
2880 * wait for all write mds requests to flush.
2882 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
2884 struct ceph_mds_request *req = NULL, *nextreq;
2887 mutex_lock(&mdsc->mutex);
2888 dout("wait_unsafe_requests want %lld\n", want_tid);
2890 req = __get_oldest_req(mdsc);
2891 while (req && req->r_tid <= want_tid) {
2892 /* find next request */
2893 n = rb_next(&req->r_node);
2895 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
2898 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
2900 ceph_mdsc_get_request(req);
2902 ceph_mdsc_get_request(nextreq);
2903 mutex_unlock(&mdsc->mutex);
2904 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
2905 req->r_tid, want_tid);
2906 wait_for_completion(&req->r_safe_completion);
2907 mutex_lock(&mdsc->mutex);
2908 ceph_mdsc_put_request(req);
2910 break; /* next dne before, so we're done! */
2911 if (RB_EMPTY_NODE(&nextreq->r_node)) {
2912 /* next request was removed from tree */
2913 ceph_mdsc_put_request(nextreq);
2916 ceph_mdsc_put_request(nextreq); /* won't go away */
2920 mutex_unlock(&mdsc->mutex);
2921 dout("wait_unsafe_requests done\n");
2924 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
2926 u64 want_tid, want_flush;
2928 if (mdsc->client->mount_state == CEPH_MOUNT_SHUTDOWN)
2932 mutex_lock(&mdsc->mutex);
2933 want_tid = mdsc->last_tid;
2934 want_flush = mdsc->cap_flush_seq;
2935 mutex_unlock(&mdsc->mutex);
2936 dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
2938 ceph_flush_dirty_caps(mdsc);
2940 wait_unsafe_requests(mdsc, want_tid);
2941 wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
2946 * called after sb is ro.
2948 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
2950 struct ceph_mds_session *session;
2953 struct ceph_client *client = mdsc->client;
2954 unsigned long started, timeout = client->mount_args->mount_timeout * HZ;
2956 dout("close_sessions\n");
2958 mutex_lock(&mdsc->mutex);
2960 /* close sessions */
2962 while (time_before(jiffies, started + timeout)) {
2963 dout("closing sessions\n");
2965 for (i = 0; i < mdsc->max_sessions; i++) {
2966 session = __ceph_lookup_mds_session(mdsc, i);
2969 mutex_unlock(&mdsc->mutex);
2970 mutex_lock(&session->s_mutex);
2971 __close_session(mdsc, session);
2972 mutex_unlock(&session->s_mutex);
2973 ceph_put_mds_session(session);
2974 mutex_lock(&mdsc->mutex);
2980 if (client->mount_state == CEPH_MOUNT_SHUTDOWN)
2983 dout("waiting for sessions to close\n");
2984 mutex_unlock(&mdsc->mutex);
2985 wait_for_completion_timeout(&mdsc->session_close_waiters,
2987 mutex_lock(&mdsc->mutex);
2990 /* tear down remaining sessions */
2991 for (i = 0; i < mdsc->max_sessions; i++) {
2992 if (mdsc->sessions[i]) {
2993 session = get_session(mdsc->sessions[i]);
2994 __unregister_session(mdsc, session);
2995 mutex_unlock(&mdsc->mutex);
2996 mutex_lock(&session->s_mutex);
2997 remove_session_caps(session);
2998 mutex_unlock(&session->s_mutex);
2999 ceph_put_mds_session(session);
3000 mutex_lock(&mdsc->mutex);
3004 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3006 mutex_unlock(&mdsc->mutex);
3008 ceph_cleanup_empty_realms(mdsc);
3010 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3015 void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3018 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3020 ceph_mdsmap_destroy(mdsc->mdsmap);
3021 kfree(mdsc->sessions);
3022 ceph_caps_finalize(mdsc);
3027 * handle mds map update.
3029 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3033 void *p = msg->front.iov_base;
3034 void *end = p + msg->front.iov_len;
3035 struct ceph_mdsmap *newmap, *oldmap;
3036 struct ceph_fsid fsid;
3039 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3040 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3041 if (ceph_check_fsid(mdsc->client, &fsid) < 0)
3043 epoch = ceph_decode_32(&p);
3044 maplen = ceph_decode_32(&p);
3045 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3047 /* do we need it? */
3048 ceph_monc_got_mdsmap(&mdsc->client->monc, epoch);
3049 mutex_lock(&mdsc->mutex);
3050 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3051 dout("handle_map epoch %u <= our %u\n",
3052 epoch, mdsc->mdsmap->m_epoch);
3053 mutex_unlock(&mdsc->mutex);
3057 newmap = ceph_mdsmap_decode(&p, end);
3058 if (IS_ERR(newmap)) {
3059 err = PTR_ERR(newmap);
3063 /* swap into place */
3065 oldmap = mdsc->mdsmap;
3066 mdsc->mdsmap = newmap;
3067 check_new_map(mdsc, newmap, oldmap);
3068 ceph_mdsmap_destroy(oldmap);
3070 mdsc->mdsmap = newmap; /* first mds map */
3072 mdsc->client->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3074 __wake_requests(mdsc, &mdsc->waiting_for_map);
3076 mutex_unlock(&mdsc->mutex);
3077 schedule_delayed(mdsc);
3081 mutex_unlock(&mdsc->mutex);
3083 pr_err("error decoding mdsmap %d\n", err);
3087 static struct ceph_connection *con_get(struct ceph_connection *con)
3089 struct ceph_mds_session *s = con->private;
3091 if (get_session(s)) {
3092 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3095 dout("mdsc con_get %p FAIL\n", s);
3099 static void con_put(struct ceph_connection *con)
3101 struct ceph_mds_session *s = con->private;
3103 ceph_put_mds_session(s);
3104 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref));
3108 * if the client is unresponsive for long enough, the mds will kill
3109 * the session entirely.
3111 static void peer_reset(struct ceph_connection *con)
3113 struct ceph_mds_session *s = con->private;
3114 struct ceph_mds_client *mdsc = s->s_mdsc;
3116 pr_warning("mds%d closed our session\n", s->s_mds);
3117 send_mds_reconnect(mdsc, s);
3120 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3122 struct ceph_mds_session *s = con->private;
3123 struct ceph_mds_client *mdsc = s->s_mdsc;
3124 int type = le16_to_cpu(msg->hdr.type);
3126 mutex_lock(&mdsc->mutex);
3127 if (__verify_registered_session(mdsc, s) < 0) {
3128 mutex_unlock(&mdsc->mutex);
3131 mutex_unlock(&mdsc->mutex);
3134 case CEPH_MSG_MDS_MAP:
3135 ceph_mdsc_handle_map(mdsc, msg);
3137 case CEPH_MSG_CLIENT_SESSION:
3138 handle_session(s, msg);
3140 case CEPH_MSG_CLIENT_REPLY:
3141 handle_reply(s, msg);
3143 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3144 handle_forward(mdsc, s, msg);
3146 case CEPH_MSG_CLIENT_CAPS:
3147 ceph_handle_caps(s, msg);
3149 case CEPH_MSG_CLIENT_SNAP:
3150 ceph_handle_snap(mdsc, s, msg);
3152 case CEPH_MSG_CLIENT_LEASE:
3153 handle_lease(mdsc, s, msg);
3157 pr_err("received unknown message type %d %s\n", type,
3158 ceph_msg_type_name(type));
3167 static int get_authorizer(struct ceph_connection *con,
3168 void **buf, int *len, int *proto,
3169 void **reply_buf, int *reply_len, int force_new)
3171 struct ceph_mds_session *s = con->private;
3172 struct ceph_mds_client *mdsc = s->s_mdsc;
3173 struct ceph_auth_client *ac = mdsc->client->monc.auth;
3176 if (force_new && s->s_authorizer) {
3177 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3178 s->s_authorizer = NULL;
3180 if (s->s_authorizer == NULL) {
3181 if (ac->ops->create_authorizer) {
3182 ret = ac->ops->create_authorizer(
3183 ac, CEPH_ENTITY_TYPE_MDS,
3185 &s->s_authorizer_buf,
3186 &s->s_authorizer_buf_len,
3187 &s->s_authorizer_reply_buf,
3188 &s->s_authorizer_reply_buf_len);
3194 *proto = ac->protocol;
3195 *buf = s->s_authorizer_buf;
3196 *len = s->s_authorizer_buf_len;
3197 *reply_buf = s->s_authorizer_reply_buf;
3198 *reply_len = s->s_authorizer_reply_buf_len;
3203 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3205 struct ceph_mds_session *s = con->private;
3206 struct ceph_mds_client *mdsc = s->s_mdsc;
3207 struct ceph_auth_client *ac = mdsc->client->monc.auth;
3209 return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3212 static int invalidate_authorizer(struct ceph_connection *con)
3214 struct ceph_mds_session *s = con->private;
3215 struct ceph_mds_client *mdsc = s->s_mdsc;
3216 struct ceph_auth_client *ac = mdsc->client->monc.auth;
3218 if (ac->ops->invalidate_authorizer)
3219 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3221 return ceph_monc_validate_auth(&mdsc->client->monc);
3224 static const struct ceph_connection_operations mds_con_ops = {
3227 .dispatch = dispatch,
3228 .get_authorizer = get_authorizer,
3229 .verify_authorizer_reply = verify_authorizer_reply,
3230 .invalidate_authorizer = invalidate_authorizer,
3231 .peer_reset = peer_reset,