1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
6 #include <linux/sched.h>
7 #include <linux/debugfs.h>
8 #include <linux/seq_file.h>
11 #include "mds_client.h"
13 #include <linux/ceph/messenger.h>
14 #include <linux/ceph/decode.h>
15 #include <linux/ceph/pagelist.h>
16 #include <linux/ceph/auth.h>
17 #include <linux/ceph/debugfs.h>
20 * A cluster of MDS (metadata server) daemons is responsible for
21 * managing the file system namespace (the directory hierarchy and
22 * inodes) and for coordinating shared access to storage. Metadata is
23 * partitioning hierarchically across a number of servers, and that
24 * partition varies over time as the cluster adjusts the distribution
25 * in order to balance load.
27 * The MDS client is primarily responsible to managing synchronous
28 * metadata requests for operations like open, unlink, and so forth.
29 * If there is a MDS failure, we find out about it when we (possibly
30 * request and) receive a new MDS map, and can resubmit affected
33 * For the most part, though, we take advantage of a lossless
34 * communications channel to the MDS, and do not need to worry about
35 * timing out or resubmitting requests.
37 * We maintain a stateful "session" with each MDS we interact with.
38 * Within each session, we sent periodic heartbeat messages to ensure
39 * any capabilities or leases we have been issues remain valid. If
40 * the session times out and goes stale, our leases and capabilities
41 * are no longer valid.
44 struct ceph_reconnect_state {
45 struct ceph_pagelist *pagelist;
49 static void __wake_requests(struct ceph_mds_client *mdsc,
50 struct list_head *head);
52 static const struct ceph_connection_operations mds_con_ops;
60 * parse individual inode info
62 static int parse_reply_info_in(void **p, void *end,
63 struct ceph_mds_reply_info_in *info,
69 *p += sizeof(struct ceph_mds_reply_inode) +
70 sizeof(*info->in->fragtree.splits) *
71 le32_to_cpu(info->in->fragtree.nsplits);
73 ceph_decode_32_safe(p, end, info->symlink_len, bad);
74 ceph_decode_need(p, end, info->symlink_len, bad);
76 *p += info->symlink_len;
78 if (features & CEPH_FEATURE_DIRLAYOUTHASH)
79 ceph_decode_copy_safe(p, end, &info->dir_layout,
80 sizeof(info->dir_layout), bad);
82 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
84 ceph_decode_32_safe(p, end, info->xattr_len, bad);
85 ceph_decode_need(p, end, info->xattr_len, bad);
86 info->xattr_data = *p;
87 *p += info->xattr_len;
94 * parse a normal reply, which may contain a (dir+)dentry and/or a
97 static int parse_reply_info_trace(void **p, void *end,
98 struct ceph_mds_reply_info_parsed *info,
103 if (info->head->is_dentry) {
104 err = parse_reply_info_in(p, end, &info->diri, features);
108 if (unlikely(*p + sizeof(*info->dirfrag) > end))
111 *p += sizeof(*info->dirfrag) +
112 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
113 if (unlikely(*p > end))
116 ceph_decode_32_safe(p, end, info->dname_len, bad);
117 ceph_decode_need(p, end, info->dname_len, bad);
119 *p += info->dname_len;
121 *p += sizeof(*info->dlease);
124 if (info->head->is_target) {
125 err = parse_reply_info_in(p, end, &info->targeti, features);
130 if (unlikely(*p != end))
137 pr_err("problem parsing mds trace %d\n", err);
142 * parse readdir results
144 static int parse_reply_info_dir(void **p, void *end,
145 struct ceph_mds_reply_info_parsed *info,
152 if (*p + sizeof(*info->dir_dir) > end)
154 *p += sizeof(*info->dir_dir) +
155 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
159 ceph_decode_need(p, end, sizeof(num) + 2, bad);
160 num = ceph_decode_32(p);
161 info->dir_end = ceph_decode_8(p);
162 info->dir_complete = ceph_decode_8(p);
166 /* alloc large array */
168 info->dir_in = kcalloc(num, sizeof(*info->dir_in) +
169 sizeof(*info->dir_dname) +
170 sizeof(*info->dir_dname_len) +
171 sizeof(*info->dir_dlease),
173 if (info->dir_in == NULL) {
177 info->dir_dname = (void *)(info->dir_in + num);
178 info->dir_dname_len = (void *)(info->dir_dname + num);
179 info->dir_dlease = (void *)(info->dir_dname_len + num);
183 ceph_decode_need(p, end, sizeof(u32)*2, bad);
184 info->dir_dname_len[i] = ceph_decode_32(p);
185 ceph_decode_need(p, end, info->dir_dname_len[i], bad);
186 info->dir_dname[i] = *p;
187 *p += info->dir_dname_len[i];
188 dout("parsed dir dname '%.*s'\n", info->dir_dname_len[i],
190 info->dir_dlease[i] = *p;
191 *p += sizeof(struct ceph_mds_reply_lease);
194 err = parse_reply_info_in(p, end, &info->dir_in[i], features);
209 pr_err("problem parsing dir contents %d\n", err);
214 * parse fcntl F_GETLK results
216 static int parse_reply_info_filelock(void **p, void *end,
217 struct ceph_mds_reply_info_parsed *info,
220 if (*p + sizeof(*info->filelock_reply) > end)
223 info->filelock_reply = *p;
224 *p += sizeof(*info->filelock_reply);
226 if (unlikely(*p != end))
235 * parse extra results
237 static int parse_reply_info_extra(void **p, void *end,
238 struct ceph_mds_reply_info_parsed *info,
241 u32 op = le32_to_cpu(info->head->op);
243 if (op == CEPH_MDS_OP_GETFILELOCK)
244 return parse_reply_info_filelock(p, end, info, features);
246 return parse_reply_info_dir(p, end, info, features);
250 * parse entire mds reply
252 static int parse_reply_info(struct ceph_msg *msg,
253 struct ceph_mds_reply_info_parsed *info,
260 info->head = msg->front.iov_base;
261 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
262 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
265 ceph_decode_32_safe(&p, end, len, bad);
267 err = parse_reply_info_trace(&p, p+len, info, features);
273 ceph_decode_32_safe(&p, end, len, bad);
275 err = parse_reply_info_extra(&p, p+len, info, features);
281 ceph_decode_32_safe(&p, end, len, bad);
282 info->snapblob_len = len;
293 pr_err("mds parse_reply err %d\n", err);
297 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
306 static const char *session_state_name(int s)
309 case CEPH_MDS_SESSION_NEW: return "new";
310 case CEPH_MDS_SESSION_OPENING: return "opening";
311 case CEPH_MDS_SESSION_OPEN: return "open";
312 case CEPH_MDS_SESSION_HUNG: return "hung";
313 case CEPH_MDS_SESSION_CLOSING: return "closing";
314 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
315 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
316 default: return "???";
320 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
322 if (atomic_inc_not_zero(&s->s_ref)) {
323 dout("mdsc get_session %p %d -> %d\n", s,
324 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
327 dout("mdsc get_session %p 0 -- FAIL", s);
332 void ceph_put_mds_session(struct ceph_mds_session *s)
334 dout("mdsc put_session %p %d -> %d\n", s,
335 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
336 if (atomic_dec_and_test(&s->s_ref)) {
338 s->s_mdsc->fsc->client->monc.auth->ops->destroy_authorizer(
339 s->s_mdsc->fsc->client->monc.auth,
346 * called under mdsc->mutex
348 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
351 struct ceph_mds_session *session;
353 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
355 session = mdsc->sessions[mds];
356 dout("lookup_mds_session %p %d\n", session,
357 atomic_read(&session->s_ref));
358 get_session(session);
362 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
364 if (mds >= mdsc->max_sessions)
366 return mdsc->sessions[mds];
369 static int __verify_registered_session(struct ceph_mds_client *mdsc,
370 struct ceph_mds_session *s)
372 if (s->s_mds >= mdsc->max_sessions ||
373 mdsc->sessions[s->s_mds] != s)
379 * create+register a new session for given mds.
380 * called under mdsc->mutex.
382 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
385 struct ceph_mds_session *s;
387 s = kzalloc(sizeof(*s), GFP_NOFS);
389 return ERR_PTR(-ENOMEM);
392 s->s_state = CEPH_MDS_SESSION_NEW;
395 mutex_init(&s->s_mutex);
397 ceph_con_init(mdsc->fsc->client->msgr, &s->s_con);
398 s->s_con.private = s;
399 s->s_con.ops = &mds_con_ops;
400 s->s_con.peer_name.type = CEPH_ENTITY_TYPE_MDS;
401 s->s_con.peer_name.num = cpu_to_le64(mds);
403 spin_lock_init(&s->s_cap_lock);
406 s->s_renew_requested = 0;
408 INIT_LIST_HEAD(&s->s_caps);
411 atomic_set(&s->s_ref, 1);
412 INIT_LIST_HEAD(&s->s_waiting);
413 INIT_LIST_HEAD(&s->s_unsafe);
414 s->s_num_cap_releases = 0;
415 s->s_cap_iterator = NULL;
416 INIT_LIST_HEAD(&s->s_cap_releases);
417 INIT_LIST_HEAD(&s->s_cap_releases_done);
418 INIT_LIST_HEAD(&s->s_cap_flushing);
419 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
421 dout("register_session mds%d\n", mds);
422 if (mds >= mdsc->max_sessions) {
423 int newmax = 1 << get_count_order(mds+1);
424 struct ceph_mds_session **sa;
426 dout("register_session realloc to %d\n", newmax);
427 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
430 if (mdsc->sessions) {
431 memcpy(sa, mdsc->sessions,
432 mdsc->max_sessions * sizeof(void *));
433 kfree(mdsc->sessions);
436 mdsc->max_sessions = newmax;
438 mdsc->sessions[mds] = s;
439 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
441 ceph_con_open(&s->s_con, ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
447 return ERR_PTR(-ENOMEM);
451 * called under mdsc->mutex
453 static void __unregister_session(struct ceph_mds_client *mdsc,
454 struct ceph_mds_session *s)
456 dout("__unregister_session mds%d %p\n", s->s_mds, s);
457 BUG_ON(mdsc->sessions[s->s_mds] != s);
458 mdsc->sessions[s->s_mds] = NULL;
459 ceph_con_close(&s->s_con);
460 ceph_put_mds_session(s);
464 * drop session refs in request.
466 * should be last request ref, or hold mdsc->mutex
468 static void put_request_session(struct ceph_mds_request *req)
470 if (req->r_session) {
471 ceph_put_mds_session(req->r_session);
472 req->r_session = NULL;
476 void ceph_mdsc_release_request(struct kref *kref)
478 struct ceph_mds_request *req = container_of(kref,
479 struct ceph_mds_request,
482 ceph_msg_put(req->r_request);
484 ceph_msg_put(req->r_reply);
485 destroy_reply_info(&req->r_reply_info);
488 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
491 if (req->r_locked_dir)
492 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
493 if (req->r_target_inode)
494 iput(req->r_target_inode);
497 if (req->r_old_dentry) {
499 * track (and drop pins for) r_old_dentry_dir
500 * separately, since r_old_dentry's d_parent may have
501 * changed between the dir mutex being dropped and
502 * this request being freed.
504 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
506 dput(req->r_old_dentry);
507 iput(req->r_old_dentry_dir);
511 put_request_session(req);
512 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
517 * lookup session, bump ref if found.
519 * called under mdsc->mutex.
521 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
524 struct ceph_mds_request *req;
525 struct rb_node *n = mdsc->request_tree.rb_node;
528 req = rb_entry(n, struct ceph_mds_request, r_node);
529 if (tid < req->r_tid)
531 else if (tid > req->r_tid)
534 ceph_mdsc_get_request(req);
541 static void __insert_request(struct ceph_mds_client *mdsc,
542 struct ceph_mds_request *new)
544 struct rb_node **p = &mdsc->request_tree.rb_node;
545 struct rb_node *parent = NULL;
546 struct ceph_mds_request *req = NULL;
550 req = rb_entry(parent, struct ceph_mds_request, r_node);
551 if (new->r_tid < req->r_tid)
553 else if (new->r_tid > req->r_tid)
559 rb_link_node(&new->r_node, parent, p);
560 rb_insert_color(&new->r_node, &mdsc->request_tree);
564 * Register an in-flight request, and assign a tid. Link to directory
565 * are modifying (if any).
567 * Called under mdsc->mutex.
569 static void __register_request(struct ceph_mds_client *mdsc,
570 struct ceph_mds_request *req,
573 req->r_tid = ++mdsc->last_tid;
575 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
577 dout("__register_request %p tid %lld\n", req, req->r_tid);
578 ceph_mdsc_get_request(req);
579 __insert_request(mdsc, req);
581 req->r_uid = current_fsuid();
582 req->r_gid = current_fsgid();
585 struct ceph_inode_info *ci = ceph_inode(dir);
588 spin_lock(&ci->i_unsafe_lock);
589 req->r_unsafe_dir = dir;
590 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
591 spin_unlock(&ci->i_unsafe_lock);
595 static void __unregister_request(struct ceph_mds_client *mdsc,
596 struct ceph_mds_request *req)
598 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
599 rb_erase(&req->r_node, &mdsc->request_tree);
600 RB_CLEAR_NODE(&req->r_node);
602 if (req->r_unsafe_dir) {
603 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
605 spin_lock(&ci->i_unsafe_lock);
606 list_del_init(&req->r_unsafe_dir_item);
607 spin_unlock(&ci->i_unsafe_lock);
609 iput(req->r_unsafe_dir);
610 req->r_unsafe_dir = NULL;
613 complete_all(&req->r_safe_completion);
615 ceph_mdsc_put_request(req);
619 * Choose mds to send request to next. If there is a hint set in the
620 * request (e.g., due to a prior forward hint from the mds), use that.
621 * Otherwise, consult frag tree and/or caps to identify the
622 * appropriate mds. If all else fails, choose randomly.
624 * Called under mdsc->mutex.
626 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
629 * we don't need to worry about protecting the d_parent access
630 * here because we never renaming inside the snapped namespace
631 * except to resplice to another snapdir, and either the old or new
632 * result is a valid result.
634 while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
635 dentry = dentry->d_parent;
639 static int __choose_mds(struct ceph_mds_client *mdsc,
640 struct ceph_mds_request *req)
643 struct ceph_inode_info *ci;
644 struct ceph_cap *cap;
645 int mode = req->r_direct_mode;
647 u32 hash = req->r_direct_hash;
648 bool is_hash = req->r_direct_is_hash;
651 * is there a specific mds we should try? ignore hint if we have
652 * no session and the mds is not up (active or recovering).
654 if (req->r_resend_mds >= 0 &&
655 (__have_session(mdsc, req->r_resend_mds) ||
656 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
657 dout("choose_mds using resend_mds mds%d\n",
659 return req->r_resend_mds;
662 if (mode == USE_RANDOM_MDS)
667 inode = req->r_inode;
668 } else if (req->r_dentry) {
669 /* ignore race with rename; old or new d_parent is okay */
670 struct dentry *parent = req->r_dentry->d_parent;
671 struct inode *dir = parent->d_inode;
673 if (dir->i_sb != mdsc->fsc->sb) {
675 inode = req->r_dentry->d_inode;
676 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
677 /* direct snapped/virtual snapdir requests
678 * based on parent dir inode */
679 struct dentry *dn = get_nonsnap_parent(parent);
681 dout("__choose_mds using nonsnap parent %p\n", inode);
682 } else if (req->r_dentry->d_inode) {
684 inode = req->r_dentry->d_inode;
688 hash = ceph_dentry_hash(dir, req->r_dentry);
693 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
697 ci = ceph_inode(inode);
699 if (is_hash && S_ISDIR(inode->i_mode)) {
700 struct ceph_inode_frag frag;
703 ceph_choose_frag(ci, hash, &frag, &found);
705 if (mode == USE_ANY_MDS && frag.ndist > 0) {
708 /* choose a random replica */
709 get_random_bytes(&r, 1);
712 dout("choose_mds %p %llx.%llx "
713 "frag %u mds%d (%d/%d)\n",
714 inode, ceph_vinop(inode),
717 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
718 CEPH_MDS_STATE_ACTIVE)
722 /* since this file/dir wasn't known to be
723 * replicated, then we want to look for the
724 * authoritative mds. */
727 /* choose auth mds */
729 dout("choose_mds %p %llx.%llx "
730 "frag %u mds%d (auth)\n",
731 inode, ceph_vinop(inode), frag.frag, mds);
732 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
733 CEPH_MDS_STATE_ACTIVE)
739 spin_lock(&ci->i_ceph_lock);
741 if (mode == USE_AUTH_MDS)
742 cap = ci->i_auth_cap;
743 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
744 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
746 spin_unlock(&ci->i_ceph_lock);
749 mds = cap->session->s_mds;
750 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
751 inode, ceph_vinop(inode), mds,
752 cap == ci->i_auth_cap ? "auth " : "", cap);
753 spin_unlock(&ci->i_ceph_lock);
757 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
758 dout("choose_mds chose random mds%d\n", mds);
766 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
768 struct ceph_msg *msg;
769 struct ceph_mds_session_head *h;
771 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
774 pr_err("create_session_msg ENOMEM creating msg\n");
777 h = msg->front.iov_base;
778 h->op = cpu_to_le32(op);
779 h->seq = cpu_to_le64(seq);
784 * send session open request.
786 * called under mdsc->mutex
788 static int __open_session(struct ceph_mds_client *mdsc,
789 struct ceph_mds_session *session)
791 struct ceph_msg *msg;
793 int mds = session->s_mds;
795 /* wait for mds to go active? */
796 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
797 dout("open_session to mds%d (%s)\n", mds,
798 ceph_mds_state_name(mstate));
799 session->s_state = CEPH_MDS_SESSION_OPENING;
800 session->s_renew_requested = jiffies;
802 /* send connect message */
803 msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
806 ceph_con_send(&session->s_con, msg);
811 * open sessions for any export targets for the given mds
813 * called under mdsc->mutex
815 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
816 struct ceph_mds_session *session)
818 struct ceph_mds_info *mi;
819 struct ceph_mds_session *ts;
820 int i, mds = session->s_mds;
823 if (mds >= mdsc->mdsmap->m_max_mds)
825 mi = &mdsc->mdsmap->m_info[mds];
826 dout("open_export_target_sessions for mds%d (%d targets)\n",
827 session->s_mds, mi->num_export_targets);
829 for (i = 0; i < mi->num_export_targets; i++) {
830 target = mi->export_targets[i];
831 ts = __ceph_lookup_mds_session(mdsc, target);
833 ts = register_session(mdsc, target);
837 if (session->s_state == CEPH_MDS_SESSION_NEW ||
838 session->s_state == CEPH_MDS_SESSION_CLOSING)
839 __open_session(mdsc, session);
841 dout(" mds%d target mds%d %p is %s\n", session->s_mds,
842 i, ts, session_state_name(ts->s_state));
843 ceph_put_mds_session(ts);
847 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
848 struct ceph_mds_session *session)
850 mutex_lock(&mdsc->mutex);
851 __open_export_target_sessions(mdsc, session);
852 mutex_unlock(&mdsc->mutex);
860 * Free preallocated cap messages assigned to this session
862 static void cleanup_cap_releases(struct ceph_mds_session *session)
864 struct ceph_msg *msg;
866 spin_lock(&session->s_cap_lock);
867 while (!list_empty(&session->s_cap_releases)) {
868 msg = list_first_entry(&session->s_cap_releases,
869 struct ceph_msg, list_head);
870 list_del_init(&msg->list_head);
873 while (!list_empty(&session->s_cap_releases_done)) {
874 msg = list_first_entry(&session->s_cap_releases_done,
875 struct ceph_msg, list_head);
876 list_del_init(&msg->list_head);
879 spin_unlock(&session->s_cap_lock);
883 * Helper to safely iterate over all caps associated with a session, with
884 * special care taken to handle a racing __ceph_remove_cap().
886 * Caller must hold session s_mutex.
888 static int iterate_session_caps(struct ceph_mds_session *session,
889 int (*cb)(struct inode *, struct ceph_cap *,
893 struct ceph_cap *cap;
894 struct inode *inode, *last_inode = NULL;
895 struct ceph_cap *old_cap = NULL;
898 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
899 spin_lock(&session->s_cap_lock);
900 p = session->s_caps.next;
901 while (p != &session->s_caps) {
902 cap = list_entry(p, struct ceph_cap, session_caps);
903 inode = igrab(&cap->ci->vfs_inode);
908 session->s_cap_iterator = cap;
909 spin_unlock(&session->s_cap_lock);
916 ceph_put_cap(session->s_mdsc, old_cap);
920 ret = cb(inode, cap, arg);
923 spin_lock(&session->s_cap_lock);
925 if (cap->ci == NULL) {
926 dout("iterate_session_caps finishing cap %p removal\n",
928 BUG_ON(cap->session != session);
929 list_del_init(&cap->session_caps);
930 session->s_nr_caps--;
932 old_cap = cap; /* put_cap it w/o locks held */
939 session->s_cap_iterator = NULL;
940 spin_unlock(&session->s_cap_lock);
945 ceph_put_cap(session->s_mdsc, old_cap);
950 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
953 struct ceph_inode_info *ci = ceph_inode(inode);
956 dout("removing cap %p, ci is %p, inode is %p\n",
957 cap, ci, &ci->vfs_inode);
958 spin_lock(&ci->i_ceph_lock);
959 __ceph_remove_cap(cap);
960 if (!__ceph_is_any_real_caps(ci)) {
961 struct ceph_mds_client *mdsc =
962 ceph_sb_to_client(inode->i_sb)->mdsc;
964 spin_lock(&mdsc->cap_dirty_lock);
965 if (!list_empty(&ci->i_dirty_item)) {
966 pr_info(" dropping dirty %s state for %p %lld\n",
967 ceph_cap_string(ci->i_dirty_caps),
968 inode, ceph_ino(inode));
969 ci->i_dirty_caps = 0;
970 list_del_init(&ci->i_dirty_item);
973 if (!list_empty(&ci->i_flushing_item)) {
974 pr_info(" dropping dirty+flushing %s state for %p %lld\n",
975 ceph_cap_string(ci->i_flushing_caps),
976 inode, ceph_ino(inode));
977 ci->i_flushing_caps = 0;
978 list_del_init(&ci->i_flushing_item);
979 mdsc->num_cap_flushing--;
982 if (drop && ci->i_wrbuffer_ref) {
983 pr_info(" dropping dirty data for %p %lld\n",
984 inode, ceph_ino(inode));
985 ci->i_wrbuffer_ref = 0;
986 ci->i_wrbuffer_ref_head = 0;
989 spin_unlock(&mdsc->cap_dirty_lock);
991 spin_unlock(&ci->i_ceph_lock);
998 * caller must hold session s_mutex
1000 static void remove_session_caps(struct ceph_mds_session *session)
1002 dout("remove_session_caps on %p\n", session);
1003 iterate_session_caps(session, remove_session_caps_cb, NULL);
1004 BUG_ON(session->s_nr_caps > 0);
1005 BUG_ON(!list_empty(&session->s_cap_flushing));
1006 cleanup_cap_releases(session);
1010 * wake up any threads waiting on this session's caps. if the cap is
1011 * old (didn't get renewed on the client reconnect), remove it now.
1013 * caller must hold s_mutex.
1015 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1018 struct ceph_inode_info *ci = ceph_inode(inode);
1020 wake_up_all(&ci->i_cap_wq);
1022 spin_lock(&ci->i_ceph_lock);
1023 ci->i_wanted_max_size = 0;
1024 ci->i_requested_max_size = 0;
1025 spin_unlock(&ci->i_ceph_lock);
1030 static void wake_up_session_caps(struct ceph_mds_session *session,
1033 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1034 iterate_session_caps(session, wake_up_session_cb,
1035 (void *)(unsigned long)reconnect);
1039 * Send periodic message to MDS renewing all currently held caps. The
1040 * ack will reset the expiration for all caps from this session.
1042 * caller holds s_mutex
1044 static int send_renew_caps(struct ceph_mds_client *mdsc,
1045 struct ceph_mds_session *session)
1047 struct ceph_msg *msg;
1050 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1051 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1052 pr_info("mds%d caps stale\n", session->s_mds);
1053 session->s_renew_requested = jiffies;
1055 /* do not try to renew caps until a recovering mds has reconnected
1056 * with its clients. */
1057 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1058 if (state < CEPH_MDS_STATE_RECONNECT) {
1059 dout("send_renew_caps ignoring mds%d (%s)\n",
1060 session->s_mds, ceph_mds_state_name(state));
1064 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1065 ceph_mds_state_name(state));
1066 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1067 ++session->s_renew_seq);
1070 ceph_con_send(&session->s_con, msg);
1075 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1077 * Called under session->s_mutex
1079 static void renewed_caps(struct ceph_mds_client *mdsc,
1080 struct ceph_mds_session *session, int is_renew)
1085 spin_lock(&session->s_cap_lock);
1086 was_stale = is_renew && (session->s_cap_ttl == 0 ||
1087 time_after_eq(jiffies, session->s_cap_ttl));
1089 session->s_cap_ttl = session->s_renew_requested +
1090 mdsc->mdsmap->m_session_timeout*HZ;
1093 if (time_before(jiffies, session->s_cap_ttl)) {
1094 pr_info("mds%d caps renewed\n", session->s_mds);
1097 pr_info("mds%d caps still stale\n", session->s_mds);
1100 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1101 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1102 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1103 spin_unlock(&session->s_cap_lock);
1106 wake_up_session_caps(session, 0);
1110 * send a session close request
1112 static int request_close_session(struct ceph_mds_client *mdsc,
1113 struct ceph_mds_session *session)
1115 struct ceph_msg *msg;
1117 dout("request_close_session mds%d state %s seq %lld\n",
1118 session->s_mds, session_state_name(session->s_state),
1120 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1123 ceph_con_send(&session->s_con, msg);
1128 * Called with s_mutex held.
1130 static int __close_session(struct ceph_mds_client *mdsc,
1131 struct ceph_mds_session *session)
1133 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1135 session->s_state = CEPH_MDS_SESSION_CLOSING;
1136 return request_close_session(mdsc, session);
1140 * Trim old(er) caps.
1142 * Because we can't cache an inode without one or more caps, we do
1143 * this indirectly: if a cap is unused, we prune its aliases, at which
1144 * point the inode will hopefully get dropped to.
1146 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1147 * memory pressure from the MDS, though, so it needn't be perfect.
1149 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1151 struct ceph_mds_session *session = arg;
1152 struct ceph_inode_info *ci = ceph_inode(inode);
1153 int used, oissued, mine;
1155 if (session->s_trim_caps <= 0)
1158 spin_lock(&ci->i_ceph_lock);
1159 mine = cap->issued | cap->implemented;
1160 used = __ceph_caps_used(ci);
1161 oissued = __ceph_caps_issued_other(ci, cap);
1163 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s\n",
1164 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1165 ceph_cap_string(used));
1166 if (ci->i_dirty_caps)
1167 goto out; /* dirty caps */
1168 if ((used & ~oissued) & mine)
1169 goto out; /* we need these caps */
1171 session->s_trim_caps--;
1173 /* we aren't the only cap.. just remove us */
1174 __ceph_remove_cap(cap);
1176 /* try to drop referring dentries */
1177 spin_unlock(&ci->i_ceph_lock);
1178 d_prune_aliases(inode);
1179 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1180 inode, cap, atomic_read(&inode->i_count));
1185 spin_unlock(&ci->i_ceph_lock);
1190 * Trim session cap count down to some max number.
1192 static int trim_caps(struct ceph_mds_client *mdsc,
1193 struct ceph_mds_session *session,
1196 int trim_caps = session->s_nr_caps - max_caps;
1198 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1199 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1200 if (trim_caps > 0) {
1201 session->s_trim_caps = trim_caps;
1202 iterate_session_caps(session, trim_caps_cb, session);
1203 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1204 session->s_mds, session->s_nr_caps, max_caps,
1205 trim_caps - session->s_trim_caps);
1206 session->s_trim_caps = 0;
1212 * Allocate cap_release messages. If there is a partially full message
1213 * in the queue, try to allocate enough to cover it's remainder, so that
1214 * we can send it immediately.
1216 * Called under s_mutex.
1218 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1219 struct ceph_mds_session *session)
1221 struct ceph_msg *msg, *partial = NULL;
1222 struct ceph_mds_cap_release *head;
1224 int extra = mdsc->fsc->mount_options->cap_release_safety;
1227 dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1230 spin_lock(&session->s_cap_lock);
1232 if (!list_empty(&session->s_cap_releases)) {
1233 msg = list_first_entry(&session->s_cap_releases,
1236 head = msg->front.iov_base;
1237 num = le32_to_cpu(head->num);
1239 dout(" partial %p with (%d/%d)\n", msg, num,
1240 (int)CEPH_CAPS_PER_RELEASE);
1241 extra += CEPH_CAPS_PER_RELEASE - num;
1245 while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1246 spin_unlock(&session->s_cap_lock);
1247 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1251 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1252 (int)msg->front.iov_len);
1253 head = msg->front.iov_base;
1254 head->num = cpu_to_le32(0);
1255 msg->front.iov_len = sizeof(*head);
1256 spin_lock(&session->s_cap_lock);
1257 list_add(&msg->list_head, &session->s_cap_releases);
1258 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1262 head = partial->front.iov_base;
1263 num = le32_to_cpu(head->num);
1264 dout(" queueing partial %p with %d/%d\n", partial, num,
1265 (int)CEPH_CAPS_PER_RELEASE);
1266 list_move_tail(&partial->list_head,
1267 &session->s_cap_releases_done);
1268 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1271 spin_unlock(&session->s_cap_lock);
1277 * flush all dirty inode data to disk.
1279 * returns true if we've flushed through want_flush_seq
1281 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1285 dout("check_cap_flush want %lld\n", want_flush_seq);
1286 mutex_lock(&mdsc->mutex);
1287 for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1288 struct ceph_mds_session *session = mdsc->sessions[mds];
1292 get_session(session);
1293 mutex_unlock(&mdsc->mutex);
1295 mutex_lock(&session->s_mutex);
1296 if (!list_empty(&session->s_cap_flushing)) {
1297 struct ceph_inode_info *ci =
1298 list_entry(session->s_cap_flushing.next,
1299 struct ceph_inode_info,
1301 struct inode *inode = &ci->vfs_inode;
1303 spin_lock(&ci->i_ceph_lock);
1304 if (ci->i_cap_flush_seq <= want_flush_seq) {
1305 dout("check_cap_flush still flushing %p "
1306 "seq %lld <= %lld to mds%d\n", inode,
1307 ci->i_cap_flush_seq, want_flush_seq,
1311 spin_unlock(&ci->i_ceph_lock);
1313 mutex_unlock(&session->s_mutex);
1314 ceph_put_mds_session(session);
1318 mutex_lock(&mdsc->mutex);
1321 mutex_unlock(&mdsc->mutex);
1322 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1327 * called under s_mutex
1329 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1330 struct ceph_mds_session *session)
1332 struct ceph_msg *msg;
1334 dout("send_cap_releases mds%d\n", session->s_mds);
1335 spin_lock(&session->s_cap_lock);
1336 while (!list_empty(&session->s_cap_releases_done)) {
1337 msg = list_first_entry(&session->s_cap_releases_done,
1338 struct ceph_msg, list_head);
1339 list_del_init(&msg->list_head);
1340 spin_unlock(&session->s_cap_lock);
1341 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1342 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1343 ceph_con_send(&session->s_con, msg);
1344 spin_lock(&session->s_cap_lock);
1346 spin_unlock(&session->s_cap_lock);
1349 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1350 struct ceph_mds_session *session)
1352 struct ceph_msg *msg;
1353 struct ceph_mds_cap_release *head;
1356 dout("discard_cap_releases mds%d\n", session->s_mds);
1357 spin_lock(&session->s_cap_lock);
1359 /* zero out the in-progress message */
1360 msg = list_first_entry(&session->s_cap_releases,
1361 struct ceph_msg, list_head);
1362 head = msg->front.iov_base;
1363 num = le32_to_cpu(head->num);
1364 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, num);
1365 head->num = cpu_to_le32(0);
1366 session->s_num_cap_releases += num;
1368 /* requeue completed messages */
1369 while (!list_empty(&session->s_cap_releases_done)) {
1370 msg = list_first_entry(&session->s_cap_releases_done,
1371 struct ceph_msg, list_head);
1372 list_del_init(&msg->list_head);
1374 head = msg->front.iov_base;
1375 num = le32_to_cpu(head->num);
1376 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1378 session->s_num_cap_releases += num;
1379 head->num = cpu_to_le32(0);
1380 msg->front.iov_len = sizeof(*head);
1381 list_add(&msg->list_head, &session->s_cap_releases);
1384 spin_unlock(&session->s_cap_lock);
1392 * Create an mds request.
1394 struct ceph_mds_request *
1395 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1397 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1400 return ERR_PTR(-ENOMEM);
1402 mutex_init(&req->r_fill_mutex);
1404 req->r_started = jiffies;
1405 req->r_resend_mds = -1;
1406 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1408 kref_init(&req->r_kref);
1409 INIT_LIST_HEAD(&req->r_wait);
1410 init_completion(&req->r_completion);
1411 init_completion(&req->r_safe_completion);
1412 INIT_LIST_HEAD(&req->r_unsafe_item);
1415 req->r_direct_mode = mode;
1420 * return oldest (lowest) request, tid in request tree, 0 if none.
1422 * called under mdsc->mutex.
1424 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1426 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1428 return rb_entry(rb_first(&mdsc->request_tree),
1429 struct ceph_mds_request, r_node);
1432 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1434 struct ceph_mds_request *req = __get_oldest_req(mdsc);
1442 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1443 * on build_path_from_dentry in fs/cifs/dir.c.
1445 * If @stop_on_nosnap, generate path relative to the first non-snapped
1448 * Encode hidden .snap dirs as a double /, i.e.
1449 * foo/.snap/bar -> foo//bar
1451 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1454 struct dentry *temp;
1460 return ERR_PTR(-EINVAL);
1464 seq = read_seqbegin(&rename_lock);
1466 for (temp = dentry; !IS_ROOT(temp);) {
1467 struct inode *inode = temp->d_inode;
1468 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1469 len++; /* slash only */
1470 else if (stop_on_nosnap && inode &&
1471 ceph_snap(inode) == CEPH_NOSNAP)
1474 len += 1 + temp->d_name.len;
1475 temp = temp->d_parent;
1478 pr_err("build_path corrupt dentry %p\n", dentry);
1479 return ERR_PTR(-EINVAL);
1484 len--; /* no leading '/' */
1486 path = kmalloc(len+1, GFP_NOFS);
1488 return ERR_PTR(-ENOMEM);
1490 path[pos] = 0; /* trailing null */
1492 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1493 struct inode *inode;
1495 spin_lock(&temp->d_lock);
1496 inode = temp->d_inode;
1497 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1498 dout("build_path path+%d: %p SNAPDIR\n",
1500 } else if (stop_on_nosnap && inode &&
1501 ceph_snap(inode) == CEPH_NOSNAP) {
1502 spin_unlock(&temp->d_lock);
1505 pos -= temp->d_name.len;
1507 spin_unlock(&temp->d_lock);
1510 strncpy(path + pos, temp->d_name.name,
1513 spin_unlock(&temp->d_lock);
1516 temp = temp->d_parent;
1519 pr_err("build_path corrupt dentry\n");
1521 return ERR_PTR(-EINVAL);
1525 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1526 pr_err("build_path did not end path lookup where "
1527 "expected, namelen is %d, pos is %d\n", len, pos);
1528 /* presumably this is only possible if racing with a
1529 rename of one of the parent directories (we can not
1530 lock the dentries above us to prevent this, but
1531 retrying should be harmless) */
1536 *base = ceph_ino(temp->d_inode);
1538 dout("build_path on %p %d built %llx '%.*s'\n",
1539 dentry, dentry->d_count, *base, len, path);
1543 static int build_dentry_path(struct dentry *dentry,
1544 const char **ppath, int *ppathlen, u64 *pino,
1549 if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1550 *pino = ceph_ino(dentry->d_parent->d_inode);
1551 *ppath = dentry->d_name.name;
1552 *ppathlen = dentry->d_name.len;
1555 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1557 return PTR_ERR(path);
1563 static int build_inode_path(struct inode *inode,
1564 const char **ppath, int *ppathlen, u64 *pino,
1567 struct dentry *dentry;
1570 if (ceph_snap(inode) == CEPH_NOSNAP) {
1571 *pino = ceph_ino(inode);
1575 dentry = d_find_alias(inode);
1576 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1579 return PTR_ERR(path);
1586 * request arguments may be specified via an inode *, a dentry *, or
1587 * an explicit ino+path.
1589 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1590 const char *rpath, u64 rino,
1591 const char **ppath, int *pathlen,
1592 u64 *ino, int *freepath)
1597 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1598 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1600 } else if (rdentry) {
1601 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1602 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1604 } else if (rpath || rino) {
1607 *pathlen = strlen(rpath);
1608 dout(" path %.*s\n", *pathlen, rpath);
1615 * called under mdsc->mutex
1617 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1618 struct ceph_mds_request *req,
1621 struct ceph_msg *msg;
1622 struct ceph_mds_request_head *head;
1623 const char *path1 = NULL;
1624 const char *path2 = NULL;
1625 u64 ino1 = 0, ino2 = 0;
1626 int pathlen1 = 0, pathlen2 = 0;
1627 int freepath1 = 0, freepath2 = 0;
1633 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1634 req->r_path1, req->r_ino1.ino,
1635 &path1, &pathlen1, &ino1, &freepath1);
1641 ret = set_request_path_attr(NULL, req->r_old_dentry,
1642 req->r_path2, req->r_ino2.ino,
1643 &path2, &pathlen2, &ino2, &freepath2);
1649 len = sizeof(*head) +
1650 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1652 /* calculate (max) length for cap releases */
1653 len += sizeof(struct ceph_mds_request_release) *
1654 (!!req->r_inode_drop + !!req->r_dentry_drop +
1655 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1656 if (req->r_dentry_drop)
1657 len += req->r_dentry->d_name.len;
1658 if (req->r_old_dentry_drop)
1659 len += req->r_old_dentry->d_name.len;
1661 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1663 msg = ERR_PTR(-ENOMEM);
1667 msg->hdr.tid = cpu_to_le64(req->r_tid);
1669 head = msg->front.iov_base;
1670 p = msg->front.iov_base + sizeof(*head);
1671 end = msg->front.iov_base + msg->front.iov_len;
1673 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1674 head->op = cpu_to_le32(req->r_op);
1675 head->caller_uid = cpu_to_le32(req->r_uid);
1676 head->caller_gid = cpu_to_le32(req->r_gid);
1677 head->args = req->r_args;
1679 ceph_encode_filepath(&p, end, ino1, path1);
1680 ceph_encode_filepath(&p, end, ino2, path2);
1682 /* make note of release offset, in case we need to replay */
1683 req->r_request_release_offset = p - msg->front.iov_base;
1687 if (req->r_inode_drop)
1688 releases += ceph_encode_inode_release(&p,
1689 req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1690 mds, req->r_inode_drop, req->r_inode_unless, 0);
1691 if (req->r_dentry_drop)
1692 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1693 mds, req->r_dentry_drop, req->r_dentry_unless);
1694 if (req->r_old_dentry_drop)
1695 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1696 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1697 if (req->r_old_inode_drop)
1698 releases += ceph_encode_inode_release(&p,
1699 req->r_old_dentry->d_inode,
1700 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1701 head->num_releases = cpu_to_le16(releases);
1704 msg->front.iov_len = p - msg->front.iov_base;
1705 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1707 msg->pages = req->r_pages;
1708 msg->nr_pages = req->r_num_pages;
1709 msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1710 msg->hdr.data_off = cpu_to_le16(0);
1714 kfree((char *)path2);
1717 kfree((char *)path1);
1723 * called under mdsc->mutex if error, under no mutex if
1726 static void complete_request(struct ceph_mds_client *mdsc,
1727 struct ceph_mds_request *req)
1729 if (req->r_callback)
1730 req->r_callback(mdsc, req);
1732 complete_all(&req->r_completion);
1736 * called under mdsc->mutex
1738 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1739 struct ceph_mds_request *req,
1742 struct ceph_mds_request_head *rhead;
1743 struct ceph_msg *msg;
1748 struct ceph_cap *cap =
1749 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1752 req->r_sent_on_mseq = cap->mseq;
1754 req->r_sent_on_mseq = -1;
1756 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1757 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1759 if (req->r_got_unsafe) {
1761 * Replay. Do not regenerate message (and rebuild
1762 * paths, etc.); just use the original message.
1763 * Rebuilding paths will break for renames because
1764 * d_move mangles the src name.
1766 msg = req->r_request;
1767 rhead = msg->front.iov_base;
1769 flags = le32_to_cpu(rhead->flags);
1770 flags |= CEPH_MDS_FLAG_REPLAY;
1771 rhead->flags = cpu_to_le32(flags);
1773 if (req->r_target_inode)
1774 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1776 rhead->num_retry = req->r_attempts - 1;
1778 /* remove cap/dentry releases from message */
1779 rhead->num_releases = 0;
1780 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1781 msg->front.iov_len = req->r_request_release_offset;
1785 if (req->r_request) {
1786 ceph_msg_put(req->r_request);
1787 req->r_request = NULL;
1789 msg = create_request_message(mdsc, req, mds);
1791 req->r_err = PTR_ERR(msg);
1792 complete_request(mdsc, req);
1793 return PTR_ERR(msg);
1795 req->r_request = msg;
1797 rhead = msg->front.iov_base;
1798 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1799 if (req->r_got_unsafe)
1800 flags |= CEPH_MDS_FLAG_REPLAY;
1801 if (req->r_locked_dir)
1802 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1803 rhead->flags = cpu_to_le32(flags);
1804 rhead->num_fwd = req->r_num_fwd;
1805 rhead->num_retry = req->r_attempts - 1;
1808 dout(" r_locked_dir = %p\n", req->r_locked_dir);
1813 * send request, or put it on the appropriate wait list.
1815 static int __do_request(struct ceph_mds_client *mdsc,
1816 struct ceph_mds_request *req)
1818 struct ceph_mds_session *session = NULL;
1822 if (req->r_err || req->r_got_result) {
1824 __unregister_request(mdsc, req);
1828 if (req->r_timeout &&
1829 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1830 dout("do_request timed out\n");
1835 put_request_session(req);
1837 mds = __choose_mds(mdsc, req);
1839 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1840 dout("do_request no mds or not active, waiting for map\n");
1841 list_add(&req->r_wait, &mdsc->waiting_for_map);
1845 /* get, open session */
1846 session = __ceph_lookup_mds_session(mdsc, mds);
1848 session = register_session(mdsc, mds);
1849 if (IS_ERR(session)) {
1850 err = PTR_ERR(session);
1854 req->r_session = get_session(session);
1856 dout("do_request mds%d session %p state %s\n", mds, session,
1857 session_state_name(session->s_state));
1858 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1859 session->s_state != CEPH_MDS_SESSION_HUNG) {
1860 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1861 session->s_state == CEPH_MDS_SESSION_CLOSING)
1862 __open_session(mdsc, session);
1863 list_add(&req->r_wait, &session->s_waiting);
1868 req->r_resend_mds = -1; /* forget any previous mds hint */
1870 if (req->r_request_started == 0) /* note request start time */
1871 req->r_request_started = jiffies;
1873 err = __prepare_send_request(mdsc, req, mds);
1875 ceph_msg_get(req->r_request);
1876 ceph_con_send(&session->s_con, req->r_request);
1880 ceph_put_mds_session(session);
1886 complete_request(mdsc, req);
1891 * called under mdsc->mutex
1893 static void __wake_requests(struct ceph_mds_client *mdsc,
1894 struct list_head *head)
1896 struct ceph_mds_request *req, *nreq;
1898 list_for_each_entry_safe(req, nreq, head, r_wait) {
1899 list_del_init(&req->r_wait);
1900 __do_request(mdsc, req);
1905 * Wake up threads with requests pending for @mds, so that they can
1906 * resubmit their requests to a possibly different mds.
1908 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
1910 struct ceph_mds_request *req;
1913 dout("kick_requests mds%d\n", mds);
1914 for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
1915 req = rb_entry(p, struct ceph_mds_request, r_node);
1916 if (req->r_got_unsafe)
1918 if (req->r_session &&
1919 req->r_session->s_mds == mds) {
1920 dout(" kicking tid %llu\n", req->r_tid);
1921 __do_request(mdsc, req);
1926 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
1927 struct ceph_mds_request *req)
1929 dout("submit_request on %p\n", req);
1930 mutex_lock(&mdsc->mutex);
1931 __register_request(mdsc, req, NULL);
1932 __do_request(mdsc, req);
1933 mutex_unlock(&mdsc->mutex);
1937 * Synchrously perform an mds request. Take care of all of the
1938 * session setup, forwarding, retry details.
1940 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
1942 struct ceph_mds_request *req)
1946 dout("do_request on %p\n", req);
1948 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
1950 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
1951 if (req->r_locked_dir)
1952 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
1953 if (req->r_old_dentry)
1954 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
1958 mutex_lock(&mdsc->mutex);
1959 __register_request(mdsc, req, dir);
1960 __do_request(mdsc, req);
1964 __unregister_request(mdsc, req);
1965 dout("do_request early error %d\n", err);
1970 mutex_unlock(&mdsc->mutex);
1971 dout("do_request waiting\n");
1972 if (req->r_timeout) {
1973 err = (long)wait_for_completion_killable_timeout(
1974 &req->r_completion, req->r_timeout);
1978 err = wait_for_completion_killable(&req->r_completion);
1980 dout("do_request waited, got %d\n", err);
1981 mutex_lock(&mdsc->mutex);
1983 /* only abort if we didn't race with a real reply */
1984 if (req->r_got_result) {
1985 err = le32_to_cpu(req->r_reply_info.head->result);
1986 } else if (err < 0) {
1987 dout("aborted request %lld with %d\n", req->r_tid, err);
1990 * ensure we aren't running concurrently with
1991 * ceph_fill_trace or ceph_readdir_prepopulate, which
1992 * rely on locks (dir mutex) held by our caller.
1994 mutex_lock(&req->r_fill_mutex);
1996 req->r_aborted = true;
1997 mutex_unlock(&req->r_fill_mutex);
1999 if (req->r_locked_dir &&
2000 (req->r_op & CEPH_MDS_OP_WRITE))
2001 ceph_invalidate_dir_request(req);
2007 mutex_unlock(&mdsc->mutex);
2008 dout("do_request %p done, result %d\n", req, err);
2013 * Invalidate dir D_COMPLETE, dentry lease state on an aborted MDS
2014 * namespace request.
2016 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2018 struct inode *inode = req->r_locked_dir;
2019 struct ceph_inode_info *ci = ceph_inode(inode);
2021 dout("invalidate_dir_request %p (D_COMPLETE, lease(s))\n", inode);
2022 spin_lock(&ci->i_ceph_lock);
2023 ceph_dir_clear_complete(inode);
2024 ci->i_release_count++;
2025 spin_unlock(&ci->i_ceph_lock);
2028 ceph_invalidate_dentry_lease(req->r_dentry);
2029 if (req->r_old_dentry)
2030 ceph_invalidate_dentry_lease(req->r_old_dentry);
2036 * We take the session mutex and parse and process the reply immediately.
2037 * This preserves the logical ordering of replies, capabilities, etc., sent
2038 * by the MDS as they are applied to our local cache.
2040 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2042 struct ceph_mds_client *mdsc = session->s_mdsc;
2043 struct ceph_mds_request *req;
2044 struct ceph_mds_reply_head *head = msg->front.iov_base;
2045 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2048 int mds = session->s_mds;
2050 if (msg->front.iov_len < sizeof(*head)) {
2051 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2056 /* get request, session */
2057 tid = le64_to_cpu(msg->hdr.tid);
2058 mutex_lock(&mdsc->mutex);
2059 req = __lookup_request(mdsc, tid);
2061 dout("handle_reply on unknown tid %llu\n", tid);
2062 mutex_unlock(&mdsc->mutex);
2065 dout("handle_reply %p\n", req);
2067 /* correct session? */
2068 if (req->r_session != session) {
2069 pr_err("mdsc_handle_reply got %llu on session mds%d"
2070 " not mds%d\n", tid, session->s_mds,
2071 req->r_session ? req->r_session->s_mds : -1);
2072 mutex_unlock(&mdsc->mutex);
2077 if ((req->r_got_unsafe && !head->safe) ||
2078 (req->r_got_safe && head->safe)) {
2079 pr_warning("got a dup %s reply on %llu from mds%d\n",
2080 head->safe ? "safe" : "unsafe", tid, mds);
2081 mutex_unlock(&mdsc->mutex);
2084 if (req->r_got_safe && !head->safe) {
2085 pr_warning("got unsafe after safe on %llu from mds%d\n",
2087 mutex_unlock(&mdsc->mutex);
2091 result = le32_to_cpu(head->result);
2095 * if we're not talking to the authority, send to them
2096 * if the authority has changed while we weren't looking,
2097 * send to new authority
2098 * Otherwise we just have to return an ESTALE
2100 if (result == -ESTALE) {
2101 dout("got ESTALE on request %llu", req->r_tid);
2102 if (!req->r_inode) {
2103 /* do nothing; not an authority problem */
2104 } else if (req->r_direct_mode != USE_AUTH_MDS) {
2105 dout("not using auth, setting for that now");
2106 req->r_direct_mode = USE_AUTH_MDS;
2107 __do_request(mdsc, req);
2108 mutex_unlock(&mdsc->mutex);
2111 struct ceph_inode_info *ci = ceph_inode(req->r_inode);
2112 struct ceph_cap *cap = NULL;
2115 cap = ceph_get_cap_for_mds(ci,
2116 req->r_session->s_mds);
2118 dout("already using auth");
2119 if ((!cap || cap != ci->i_auth_cap) ||
2120 (cap->mseq != req->r_sent_on_mseq)) {
2121 dout("but cap changed, so resending");
2122 __do_request(mdsc, req);
2123 mutex_unlock(&mdsc->mutex);
2127 dout("have to return ESTALE on request %llu", req->r_tid);
2132 req->r_got_safe = true;
2133 __unregister_request(mdsc, req);
2135 if (req->r_got_unsafe) {
2137 * We already handled the unsafe response, now do the
2138 * cleanup. No need to examine the response; the MDS
2139 * doesn't include any result info in the safe
2140 * response. And even if it did, there is nothing
2141 * useful we could do with a revised return value.
2143 dout("got safe reply %llu, mds%d\n", tid, mds);
2144 list_del_init(&req->r_unsafe_item);
2146 /* last unsafe request during umount? */
2147 if (mdsc->stopping && !__get_oldest_req(mdsc))
2148 complete_all(&mdsc->safe_umount_waiters);
2149 mutex_unlock(&mdsc->mutex);
2153 req->r_got_unsafe = true;
2154 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2157 dout("handle_reply tid %lld result %d\n", tid, result);
2158 rinfo = &req->r_reply_info;
2159 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2160 mutex_unlock(&mdsc->mutex);
2162 mutex_lock(&session->s_mutex);
2164 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2170 if (rinfo->snapblob_len) {
2171 down_write(&mdsc->snap_rwsem);
2172 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2173 rinfo->snapblob + rinfo->snapblob_len,
2174 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2175 downgrade_write(&mdsc->snap_rwsem);
2177 down_read(&mdsc->snap_rwsem);
2180 /* insert trace into our cache */
2181 mutex_lock(&req->r_fill_mutex);
2182 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2184 if (result == 0 && req->r_op != CEPH_MDS_OP_GETFILELOCK &&
2186 ceph_readdir_prepopulate(req, req->r_session);
2187 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2189 mutex_unlock(&req->r_fill_mutex);
2191 up_read(&mdsc->snap_rwsem);
2193 mutex_lock(&mdsc->mutex);
2194 if (!req->r_aborted) {
2200 req->r_got_result = true;
2203 dout("reply arrived after request %lld was aborted\n", tid);
2205 mutex_unlock(&mdsc->mutex);
2207 ceph_add_cap_releases(mdsc, req->r_session);
2208 mutex_unlock(&session->s_mutex);
2210 /* kick calling process */
2211 complete_request(mdsc, req);
2213 ceph_mdsc_put_request(req);
2220 * handle mds notification that our request has been forwarded.
2222 static void handle_forward(struct ceph_mds_client *mdsc,
2223 struct ceph_mds_session *session,
2224 struct ceph_msg *msg)
2226 struct ceph_mds_request *req;
2227 u64 tid = le64_to_cpu(msg->hdr.tid);
2231 void *p = msg->front.iov_base;
2232 void *end = p + msg->front.iov_len;
2234 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2235 next_mds = ceph_decode_32(&p);
2236 fwd_seq = ceph_decode_32(&p);
2238 mutex_lock(&mdsc->mutex);
2239 req = __lookup_request(mdsc, tid);
2241 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2242 goto out; /* dup reply? */
2245 if (req->r_aborted) {
2246 dout("forward tid %llu aborted, unregistering\n", tid);
2247 __unregister_request(mdsc, req);
2248 } else if (fwd_seq <= req->r_num_fwd) {
2249 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2250 tid, next_mds, req->r_num_fwd, fwd_seq);
2252 /* resend. forward race not possible; mds would drop */
2253 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2255 BUG_ON(req->r_got_result);
2256 req->r_num_fwd = fwd_seq;
2257 req->r_resend_mds = next_mds;
2258 put_request_session(req);
2259 __do_request(mdsc, req);
2261 ceph_mdsc_put_request(req);
2263 mutex_unlock(&mdsc->mutex);
2267 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2271 * handle a mds session control message
2273 static void handle_session(struct ceph_mds_session *session,
2274 struct ceph_msg *msg)
2276 struct ceph_mds_client *mdsc = session->s_mdsc;
2279 int mds = session->s_mds;
2280 struct ceph_mds_session_head *h = msg->front.iov_base;
2284 if (msg->front.iov_len != sizeof(*h))
2286 op = le32_to_cpu(h->op);
2287 seq = le64_to_cpu(h->seq);
2289 mutex_lock(&mdsc->mutex);
2290 if (op == CEPH_SESSION_CLOSE)
2291 __unregister_session(mdsc, session);
2292 /* FIXME: this ttl calculation is generous */
2293 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2294 mutex_unlock(&mdsc->mutex);
2296 mutex_lock(&session->s_mutex);
2298 dout("handle_session mds%d %s %p state %s seq %llu\n",
2299 mds, ceph_session_op_name(op), session,
2300 session_state_name(session->s_state), seq);
2302 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2303 session->s_state = CEPH_MDS_SESSION_OPEN;
2304 pr_info("mds%d came back\n", session->s_mds);
2308 case CEPH_SESSION_OPEN:
2309 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2310 pr_info("mds%d reconnect success\n", session->s_mds);
2311 session->s_state = CEPH_MDS_SESSION_OPEN;
2312 renewed_caps(mdsc, session, 0);
2315 __close_session(mdsc, session);
2318 case CEPH_SESSION_RENEWCAPS:
2319 if (session->s_renew_seq == seq)
2320 renewed_caps(mdsc, session, 1);
2323 case CEPH_SESSION_CLOSE:
2324 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2325 pr_info("mds%d reconnect denied\n", session->s_mds);
2326 remove_session_caps(session);
2327 wake = 1; /* for good measure */
2328 wake_up_all(&mdsc->session_close_wq);
2329 kick_requests(mdsc, mds);
2332 case CEPH_SESSION_STALE:
2333 pr_info("mds%d caps went stale, renewing\n",
2335 spin_lock(&session->s_cap_lock);
2336 session->s_cap_gen++;
2337 session->s_cap_ttl = 0;
2338 spin_unlock(&session->s_cap_lock);
2339 send_renew_caps(mdsc, session);
2342 case CEPH_SESSION_RECALL_STATE:
2343 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2347 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2351 mutex_unlock(&session->s_mutex);
2353 mutex_lock(&mdsc->mutex);
2354 __wake_requests(mdsc, &session->s_waiting);
2355 mutex_unlock(&mdsc->mutex);
2360 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2361 (int)msg->front.iov_len);
2368 * called under session->mutex.
2370 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2371 struct ceph_mds_session *session)
2373 struct ceph_mds_request *req, *nreq;
2376 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2378 mutex_lock(&mdsc->mutex);
2379 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2380 err = __prepare_send_request(mdsc, req, session->s_mds);
2382 ceph_msg_get(req->r_request);
2383 ceph_con_send(&session->s_con, req->r_request);
2386 mutex_unlock(&mdsc->mutex);
2390 * Encode information about a cap for a reconnect with the MDS.
2392 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2396 struct ceph_mds_cap_reconnect v2;
2397 struct ceph_mds_cap_reconnect_v1 v1;
2400 struct ceph_inode_info *ci;
2401 struct ceph_reconnect_state *recon_state = arg;
2402 struct ceph_pagelist *pagelist = recon_state->pagelist;
2406 struct dentry *dentry;
2410 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2411 inode, ceph_vinop(inode), cap, cap->cap_id,
2412 ceph_cap_string(cap->issued));
2413 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2417 dentry = d_find_alias(inode);
2419 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2421 err = PTR_ERR(path);
2428 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2432 spin_lock(&ci->i_ceph_lock);
2433 cap->seq = 0; /* reset cap seq */
2434 cap->issue_seq = 0; /* and issue_seq */
2436 if (recon_state->flock) {
2437 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2438 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2439 rec.v2.issued = cpu_to_le32(cap->issued);
2440 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2441 rec.v2.pathbase = cpu_to_le64(pathbase);
2442 rec.v2.flock_len = 0;
2443 reclen = sizeof(rec.v2);
2445 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2446 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2447 rec.v1.issued = cpu_to_le32(cap->issued);
2448 rec.v1.size = cpu_to_le64(inode->i_size);
2449 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2450 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2451 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2452 rec.v1.pathbase = cpu_to_le64(pathbase);
2453 reclen = sizeof(rec.v1);
2455 spin_unlock(&ci->i_ceph_lock);
2457 if (recon_state->flock) {
2458 int num_fcntl_locks, num_flock_locks;
2459 struct ceph_pagelist_cursor trunc_point;
2461 ceph_pagelist_set_cursor(pagelist, &trunc_point);
2464 ceph_count_locks(inode, &num_fcntl_locks,
2466 rec.v2.flock_len = (2*sizeof(u32) +
2467 (num_fcntl_locks+num_flock_locks) *
2468 sizeof(struct ceph_filelock));
2471 /* pre-alloc pagelist */
2472 ceph_pagelist_truncate(pagelist, &trunc_point);
2473 err = ceph_pagelist_append(pagelist, &rec, reclen);
2475 err = ceph_pagelist_reserve(pagelist,
2481 err = ceph_encode_locks(inode,
2487 } while (err == -ENOSPC);
2489 err = ceph_pagelist_append(pagelist, &rec, reclen);
2501 * If an MDS fails and recovers, clients need to reconnect in order to
2502 * reestablish shared state. This includes all caps issued through
2503 * this session _and_ the snap_realm hierarchy. Because it's not
2504 * clear which snap realms the mds cares about, we send everything we
2505 * know about.. that ensures we'll then get any new info the
2506 * recovering MDS might have.
2508 * This is a relatively heavyweight operation, but it's rare.
2510 * called with mdsc->mutex held.
2512 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2513 struct ceph_mds_session *session)
2515 struct ceph_msg *reply;
2517 int mds = session->s_mds;
2519 struct ceph_pagelist *pagelist;
2520 struct ceph_reconnect_state recon_state;
2522 pr_info("mds%d reconnect start\n", mds);
2524 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2526 goto fail_nopagelist;
2527 ceph_pagelist_init(pagelist);
2529 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2533 mutex_lock(&session->s_mutex);
2534 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2537 ceph_con_open(&session->s_con,
2538 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2540 /* replay unsafe requests */
2541 replay_unsafe_requests(mdsc, session);
2543 down_read(&mdsc->snap_rwsem);
2545 dout("session %p state %s\n", session,
2546 session_state_name(session->s_state));
2548 /* drop old cap expires; we're about to reestablish that state */
2549 discard_cap_releases(mdsc, session);
2551 /* traverse this session's caps */
2552 err = ceph_pagelist_encode_32(pagelist, session->s_nr_caps);
2556 recon_state.pagelist = pagelist;
2557 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2558 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2563 * snaprealms. we provide mds with the ino, seq (version), and
2564 * parent for all of our realms. If the mds has any newer info,
2567 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2568 struct ceph_snap_realm *realm =
2569 rb_entry(p, struct ceph_snap_realm, node);
2570 struct ceph_mds_snaprealm_reconnect sr_rec;
2572 dout(" adding snap realm %llx seq %lld parent %llx\n",
2573 realm->ino, realm->seq, realm->parent_ino);
2574 sr_rec.ino = cpu_to_le64(realm->ino);
2575 sr_rec.seq = cpu_to_le64(realm->seq);
2576 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2577 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2582 reply->pagelist = pagelist;
2583 if (recon_state.flock)
2584 reply->hdr.version = cpu_to_le16(2);
2585 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2586 reply->nr_pages = calc_pages_for(0, pagelist->length);
2587 ceph_con_send(&session->s_con, reply);
2589 mutex_unlock(&session->s_mutex);
2591 mutex_lock(&mdsc->mutex);
2592 __wake_requests(mdsc, &session->s_waiting);
2593 mutex_unlock(&mdsc->mutex);
2595 up_read(&mdsc->snap_rwsem);
2599 ceph_msg_put(reply);
2600 up_read(&mdsc->snap_rwsem);
2601 mutex_unlock(&session->s_mutex);
2603 ceph_pagelist_release(pagelist);
2606 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2612 * compare old and new mdsmaps, kicking requests
2613 * and closing out old connections as necessary
2615 * called under mdsc->mutex.
2617 static void check_new_map(struct ceph_mds_client *mdsc,
2618 struct ceph_mdsmap *newmap,
2619 struct ceph_mdsmap *oldmap)
2622 int oldstate, newstate;
2623 struct ceph_mds_session *s;
2625 dout("check_new_map new %u old %u\n",
2626 newmap->m_epoch, oldmap->m_epoch);
2628 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2629 if (mdsc->sessions[i] == NULL)
2631 s = mdsc->sessions[i];
2632 oldstate = ceph_mdsmap_get_state(oldmap, i);
2633 newstate = ceph_mdsmap_get_state(newmap, i);
2635 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2636 i, ceph_mds_state_name(oldstate),
2637 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2638 ceph_mds_state_name(newstate),
2639 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2640 session_state_name(s->s_state));
2642 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
2643 ceph_mdsmap_get_addr(newmap, i),
2644 sizeof(struct ceph_entity_addr))) {
2645 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2646 /* the session never opened, just close it
2648 __wake_requests(mdsc, &s->s_waiting);
2649 __unregister_session(mdsc, s);
2652 mutex_unlock(&mdsc->mutex);
2653 mutex_lock(&s->s_mutex);
2654 mutex_lock(&mdsc->mutex);
2655 ceph_con_close(&s->s_con);
2656 mutex_unlock(&s->s_mutex);
2657 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2660 /* kick any requests waiting on the recovering mds */
2661 kick_requests(mdsc, i);
2662 } else if (oldstate == newstate) {
2663 continue; /* nothing new with this mds */
2669 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2670 newstate >= CEPH_MDS_STATE_RECONNECT) {
2671 mutex_unlock(&mdsc->mutex);
2672 send_mds_reconnect(mdsc, s);
2673 mutex_lock(&mdsc->mutex);
2677 * kick request on any mds that has gone active.
2679 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2680 newstate >= CEPH_MDS_STATE_ACTIVE) {
2681 if (oldstate != CEPH_MDS_STATE_CREATING &&
2682 oldstate != CEPH_MDS_STATE_STARTING)
2683 pr_info("mds%d recovery completed\n", s->s_mds);
2684 kick_requests(mdsc, i);
2685 ceph_kick_flushing_caps(mdsc, s);
2686 wake_up_session_caps(s, 1);
2690 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2691 s = mdsc->sessions[i];
2694 if (!ceph_mdsmap_is_laggy(newmap, i))
2696 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2697 s->s_state == CEPH_MDS_SESSION_HUNG ||
2698 s->s_state == CEPH_MDS_SESSION_CLOSING) {
2699 dout(" connecting to export targets of laggy mds%d\n",
2701 __open_export_target_sessions(mdsc, s);
2713 * caller must hold session s_mutex, dentry->d_lock
2715 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2717 struct ceph_dentry_info *di = ceph_dentry(dentry);
2719 ceph_put_mds_session(di->lease_session);
2720 di->lease_session = NULL;
2723 static void handle_lease(struct ceph_mds_client *mdsc,
2724 struct ceph_mds_session *session,
2725 struct ceph_msg *msg)
2727 struct super_block *sb = mdsc->fsc->sb;
2728 struct inode *inode;
2729 struct dentry *parent, *dentry;
2730 struct ceph_dentry_info *di;
2731 int mds = session->s_mds;
2732 struct ceph_mds_lease *h = msg->front.iov_base;
2734 struct ceph_vino vino;
2738 dout("handle_lease from mds%d\n", mds);
2741 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2743 vino.ino = le64_to_cpu(h->ino);
2744 vino.snap = CEPH_NOSNAP;
2745 seq = le32_to_cpu(h->seq);
2746 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2747 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2748 if (dname.len != get_unaligned_le32(h+1))
2751 mutex_lock(&session->s_mutex);
2755 inode = ceph_find_inode(sb, vino);
2756 dout("handle_lease %s, ino %llx %p %.*s\n",
2757 ceph_lease_op_name(h->action), vino.ino, inode,
2758 dname.len, dname.name);
2759 if (inode == NULL) {
2760 dout("handle_lease no inode %llx\n", vino.ino);
2765 parent = d_find_alias(inode);
2767 dout("no parent dentry on inode %p\n", inode);
2769 goto release; /* hrm... */
2771 dname.hash = full_name_hash(dname.name, dname.len);
2772 dentry = d_lookup(parent, &dname);
2777 spin_lock(&dentry->d_lock);
2778 di = ceph_dentry(dentry);
2779 switch (h->action) {
2780 case CEPH_MDS_LEASE_REVOKE:
2781 if (di && di->lease_session == session) {
2782 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2783 h->seq = cpu_to_le32(di->lease_seq);
2784 __ceph_mdsc_drop_dentry_lease(dentry);
2789 case CEPH_MDS_LEASE_RENEW:
2790 if (di && di->lease_session == session &&
2791 di->lease_gen == session->s_cap_gen &&
2792 di->lease_renew_from &&
2793 di->lease_renew_after == 0) {
2794 unsigned long duration =
2795 le32_to_cpu(h->duration_ms) * HZ / 1000;
2797 di->lease_seq = seq;
2798 dentry->d_time = di->lease_renew_from + duration;
2799 di->lease_renew_after = di->lease_renew_from +
2801 di->lease_renew_from = 0;
2805 spin_unlock(&dentry->d_lock);
2812 /* let's just reuse the same message */
2813 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2815 ceph_con_send(&session->s_con, msg);
2819 mutex_unlock(&session->s_mutex);
2823 pr_err("corrupt lease message\n");
2827 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
2828 struct inode *inode,
2829 struct dentry *dentry, char action,
2832 struct ceph_msg *msg;
2833 struct ceph_mds_lease *lease;
2834 int len = sizeof(*lease) + sizeof(u32);
2837 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
2838 inode, dentry, ceph_lease_op_name(action), session->s_mds);
2839 dnamelen = dentry->d_name.len;
2842 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
2845 lease = msg->front.iov_base;
2846 lease->action = action;
2847 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
2848 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
2849 lease->seq = cpu_to_le32(seq);
2850 put_unaligned_le32(dnamelen, lease + 1);
2851 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
2854 * if this is a preemptive lease RELEASE, no need to
2855 * flush request stream, since the actual request will
2858 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
2860 ceph_con_send(&session->s_con, msg);
2864 * Preemptively release a lease we expect to invalidate anyway.
2865 * Pass @inode always, @dentry is optional.
2867 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
2868 struct dentry *dentry)
2870 struct ceph_dentry_info *di;
2871 struct ceph_mds_session *session;
2874 BUG_ON(inode == NULL);
2875 BUG_ON(dentry == NULL);
2877 /* is dentry lease valid? */
2878 spin_lock(&dentry->d_lock);
2879 di = ceph_dentry(dentry);
2880 if (!di || !di->lease_session ||
2881 di->lease_session->s_mds < 0 ||
2882 di->lease_gen != di->lease_session->s_cap_gen ||
2883 !time_before(jiffies, dentry->d_time)) {
2884 dout("lease_release inode %p dentry %p -- "
2887 spin_unlock(&dentry->d_lock);
2891 /* we do have a lease on this dentry; note mds and seq */
2892 session = ceph_get_mds_session(di->lease_session);
2893 seq = di->lease_seq;
2894 __ceph_mdsc_drop_dentry_lease(dentry);
2895 spin_unlock(&dentry->d_lock);
2897 dout("lease_release inode %p dentry %p to mds%d\n",
2898 inode, dentry, session->s_mds);
2899 ceph_mdsc_lease_send_msg(session, inode, dentry,
2900 CEPH_MDS_LEASE_RELEASE, seq);
2901 ceph_put_mds_session(session);
2905 * drop all leases (and dentry refs) in preparation for umount
2907 static void drop_leases(struct ceph_mds_client *mdsc)
2911 dout("drop_leases\n");
2912 mutex_lock(&mdsc->mutex);
2913 for (i = 0; i < mdsc->max_sessions; i++) {
2914 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2917 mutex_unlock(&mdsc->mutex);
2918 mutex_lock(&s->s_mutex);
2919 mutex_unlock(&s->s_mutex);
2920 ceph_put_mds_session(s);
2921 mutex_lock(&mdsc->mutex);
2923 mutex_unlock(&mdsc->mutex);
2929 * delayed work -- periodically trim expired leases, renew caps with mds
2931 static void schedule_delayed(struct ceph_mds_client *mdsc)
2934 unsigned hz = round_jiffies_relative(HZ * delay);
2935 schedule_delayed_work(&mdsc->delayed_work, hz);
2938 static void delayed_work(struct work_struct *work)
2941 struct ceph_mds_client *mdsc =
2942 container_of(work, struct ceph_mds_client, delayed_work.work);
2946 dout("mdsc delayed_work\n");
2947 ceph_check_delayed_caps(mdsc);
2949 mutex_lock(&mdsc->mutex);
2950 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
2951 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
2952 mdsc->last_renew_caps);
2954 mdsc->last_renew_caps = jiffies;
2956 for (i = 0; i < mdsc->max_sessions; i++) {
2957 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
2960 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
2961 dout("resending session close request for mds%d\n",
2963 request_close_session(mdsc, s);
2964 ceph_put_mds_session(s);
2967 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
2968 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
2969 s->s_state = CEPH_MDS_SESSION_HUNG;
2970 pr_info("mds%d hung\n", s->s_mds);
2973 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
2974 /* this mds is failed or recovering, just wait */
2975 ceph_put_mds_session(s);
2978 mutex_unlock(&mdsc->mutex);
2980 mutex_lock(&s->s_mutex);
2982 send_renew_caps(mdsc, s);
2984 ceph_con_keepalive(&s->s_con);
2985 ceph_add_cap_releases(mdsc, s);
2986 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2987 s->s_state == CEPH_MDS_SESSION_HUNG)
2988 ceph_send_cap_releases(mdsc, s);
2989 mutex_unlock(&s->s_mutex);
2990 ceph_put_mds_session(s);
2992 mutex_lock(&mdsc->mutex);
2994 mutex_unlock(&mdsc->mutex);
2996 schedule_delayed(mdsc);
2999 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3002 struct ceph_mds_client *mdsc;
3004 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3009 mutex_init(&mdsc->mutex);
3010 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3011 if (mdsc->mdsmap == NULL)
3014 init_completion(&mdsc->safe_umount_waiters);
3015 init_waitqueue_head(&mdsc->session_close_wq);
3016 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3017 mdsc->sessions = NULL;
3018 mdsc->max_sessions = 0;
3020 init_rwsem(&mdsc->snap_rwsem);
3021 mdsc->snap_realms = RB_ROOT;
3022 INIT_LIST_HEAD(&mdsc->snap_empty);
3023 spin_lock_init(&mdsc->snap_empty_lock);
3025 mdsc->request_tree = RB_ROOT;
3026 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3027 mdsc->last_renew_caps = jiffies;
3028 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3029 spin_lock_init(&mdsc->cap_delay_lock);
3030 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3031 spin_lock_init(&mdsc->snap_flush_lock);
3032 mdsc->cap_flush_seq = 0;
3033 INIT_LIST_HEAD(&mdsc->cap_dirty);
3034 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3035 mdsc->num_cap_flushing = 0;
3036 spin_lock_init(&mdsc->cap_dirty_lock);
3037 init_waitqueue_head(&mdsc->cap_flushing_wq);
3038 spin_lock_init(&mdsc->dentry_lru_lock);
3039 INIT_LIST_HEAD(&mdsc->dentry_lru);
3041 ceph_caps_init(mdsc);
3042 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3048 * Wait for safe replies on open mds requests. If we time out, drop
3049 * all requests from the tree to avoid dangling dentry refs.
3051 static void wait_requests(struct ceph_mds_client *mdsc)
3053 struct ceph_mds_request *req;
3054 struct ceph_fs_client *fsc = mdsc->fsc;
3056 mutex_lock(&mdsc->mutex);
3057 if (__get_oldest_req(mdsc)) {
3058 mutex_unlock(&mdsc->mutex);
3060 dout("wait_requests waiting for requests\n");
3061 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3062 fsc->client->options->mount_timeout * HZ);
3064 /* tear down remaining requests */
3065 mutex_lock(&mdsc->mutex);
3066 while ((req = __get_oldest_req(mdsc))) {
3067 dout("wait_requests timed out on tid %llu\n",
3069 __unregister_request(mdsc, req);
3072 mutex_unlock(&mdsc->mutex);
3073 dout("wait_requests done\n");
3077 * called before mount is ro, and before dentries are torn down.
3078 * (hmm, does this still race with new lookups?)
3080 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3082 dout("pre_umount\n");
3086 ceph_flush_dirty_caps(mdsc);
3087 wait_requests(mdsc);
3090 * wait for reply handlers to drop their request refs and
3091 * their inode/dcache refs
3097 * wait for all write mds requests to flush.
3099 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3101 struct ceph_mds_request *req = NULL, *nextreq;
3104 mutex_lock(&mdsc->mutex);
3105 dout("wait_unsafe_requests want %lld\n", want_tid);
3107 req = __get_oldest_req(mdsc);
3108 while (req && req->r_tid <= want_tid) {
3109 /* find next request */
3110 n = rb_next(&req->r_node);
3112 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3115 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3117 ceph_mdsc_get_request(req);
3119 ceph_mdsc_get_request(nextreq);
3120 mutex_unlock(&mdsc->mutex);
3121 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3122 req->r_tid, want_tid);
3123 wait_for_completion(&req->r_safe_completion);
3124 mutex_lock(&mdsc->mutex);
3125 ceph_mdsc_put_request(req);
3127 break; /* next dne before, so we're done! */
3128 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3129 /* next request was removed from tree */
3130 ceph_mdsc_put_request(nextreq);
3133 ceph_mdsc_put_request(nextreq); /* won't go away */
3137 mutex_unlock(&mdsc->mutex);
3138 dout("wait_unsafe_requests done\n");
3141 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3143 u64 want_tid, want_flush;
3145 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3149 mutex_lock(&mdsc->mutex);
3150 want_tid = mdsc->last_tid;
3151 want_flush = mdsc->cap_flush_seq;
3152 mutex_unlock(&mdsc->mutex);
3153 dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3155 ceph_flush_dirty_caps(mdsc);
3157 wait_unsafe_requests(mdsc, want_tid);
3158 wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3162 * true if all sessions are closed, or we force unmount
3164 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3168 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3171 mutex_lock(&mdsc->mutex);
3172 for (i = 0; i < mdsc->max_sessions; i++)
3173 if (mdsc->sessions[i])
3175 mutex_unlock(&mdsc->mutex);
3180 * called after sb is ro.
3182 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3184 struct ceph_mds_session *session;
3186 struct ceph_fs_client *fsc = mdsc->fsc;
3187 unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3189 dout("close_sessions\n");
3191 /* close sessions */
3192 mutex_lock(&mdsc->mutex);
3193 for (i = 0; i < mdsc->max_sessions; i++) {
3194 session = __ceph_lookup_mds_session(mdsc, i);
3197 mutex_unlock(&mdsc->mutex);
3198 mutex_lock(&session->s_mutex);
3199 __close_session(mdsc, session);
3200 mutex_unlock(&session->s_mutex);
3201 ceph_put_mds_session(session);
3202 mutex_lock(&mdsc->mutex);
3204 mutex_unlock(&mdsc->mutex);
3206 dout("waiting for sessions to close\n");
3207 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3210 /* tear down remaining sessions */
3211 mutex_lock(&mdsc->mutex);
3212 for (i = 0; i < mdsc->max_sessions; i++) {
3213 if (mdsc->sessions[i]) {
3214 session = get_session(mdsc->sessions[i]);
3215 __unregister_session(mdsc, session);
3216 mutex_unlock(&mdsc->mutex);
3217 mutex_lock(&session->s_mutex);
3218 remove_session_caps(session);
3219 mutex_unlock(&session->s_mutex);
3220 ceph_put_mds_session(session);
3221 mutex_lock(&mdsc->mutex);
3224 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3225 mutex_unlock(&mdsc->mutex);
3227 ceph_cleanup_empty_realms(mdsc);
3229 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3234 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3237 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3239 ceph_mdsmap_destroy(mdsc->mdsmap);
3240 kfree(mdsc->sessions);
3241 ceph_caps_finalize(mdsc);
3244 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3246 struct ceph_mds_client *mdsc = fsc->mdsc;
3248 dout("mdsc_destroy %p\n", mdsc);
3249 ceph_mdsc_stop(mdsc);
3251 /* flush out any connection work with references to us */
3256 dout("mdsc_destroy %p done\n", mdsc);
3261 * handle mds map update.
3263 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3267 void *p = msg->front.iov_base;
3268 void *end = p + msg->front.iov_len;
3269 struct ceph_mdsmap *newmap, *oldmap;
3270 struct ceph_fsid fsid;
3273 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3274 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3275 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3277 epoch = ceph_decode_32(&p);
3278 maplen = ceph_decode_32(&p);
3279 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3281 /* do we need it? */
3282 ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3283 mutex_lock(&mdsc->mutex);
3284 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3285 dout("handle_map epoch %u <= our %u\n",
3286 epoch, mdsc->mdsmap->m_epoch);
3287 mutex_unlock(&mdsc->mutex);
3291 newmap = ceph_mdsmap_decode(&p, end);
3292 if (IS_ERR(newmap)) {
3293 err = PTR_ERR(newmap);
3297 /* swap into place */
3299 oldmap = mdsc->mdsmap;
3300 mdsc->mdsmap = newmap;
3301 check_new_map(mdsc, newmap, oldmap);
3302 ceph_mdsmap_destroy(oldmap);
3304 mdsc->mdsmap = newmap; /* first mds map */
3306 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3308 __wake_requests(mdsc, &mdsc->waiting_for_map);
3310 mutex_unlock(&mdsc->mutex);
3311 schedule_delayed(mdsc);
3315 mutex_unlock(&mdsc->mutex);
3317 pr_err("error decoding mdsmap %d\n", err);
3321 static struct ceph_connection *con_get(struct ceph_connection *con)
3323 struct ceph_mds_session *s = con->private;
3325 if (get_session(s)) {
3326 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3329 dout("mdsc con_get %p FAIL\n", s);
3333 static void con_put(struct ceph_connection *con)
3335 struct ceph_mds_session *s = con->private;
3337 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3338 ceph_put_mds_session(s);
3342 * if the client is unresponsive for long enough, the mds will kill
3343 * the session entirely.
3345 static void peer_reset(struct ceph_connection *con)
3347 struct ceph_mds_session *s = con->private;
3348 struct ceph_mds_client *mdsc = s->s_mdsc;
3350 pr_warning("mds%d closed our session\n", s->s_mds);
3351 send_mds_reconnect(mdsc, s);
3354 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3356 struct ceph_mds_session *s = con->private;
3357 struct ceph_mds_client *mdsc = s->s_mdsc;
3358 int type = le16_to_cpu(msg->hdr.type);
3360 mutex_lock(&mdsc->mutex);
3361 if (__verify_registered_session(mdsc, s) < 0) {
3362 mutex_unlock(&mdsc->mutex);
3365 mutex_unlock(&mdsc->mutex);
3368 case CEPH_MSG_MDS_MAP:
3369 ceph_mdsc_handle_map(mdsc, msg);
3371 case CEPH_MSG_CLIENT_SESSION:
3372 handle_session(s, msg);
3374 case CEPH_MSG_CLIENT_REPLY:
3375 handle_reply(s, msg);
3377 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3378 handle_forward(mdsc, s, msg);
3380 case CEPH_MSG_CLIENT_CAPS:
3381 ceph_handle_caps(s, msg);
3383 case CEPH_MSG_CLIENT_SNAP:
3384 ceph_handle_snap(mdsc, s, msg);
3386 case CEPH_MSG_CLIENT_LEASE:
3387 handle_lease(mdsc, s, msg);
3391 pr_err("received unknown message type %d %s\n", type,
3392 ceph_msg_type_name(type));
3401 static int get_authorizer(struct ceph_connection *con,
3402 void **buf, int *len, int *proto,
3403 void **reply_buf, int *reply_len, int force_new)
3405 struct ceph_mds_session *s = con->private;
3406 struct ceph_mds_client *mdsc = s->s_mdsc;
3407 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3410 if (force_new && s->s_authorizer) {
3411 ac->ops->destroy_authorizer(ac, s->s_authorizer);
3412 s->s_authorizer = NULL;
3414 if (s->s_authorizer == NULL) {
3415 if (ac->ops->create_authorizer) {
3416 ret = ac->ops->create_authorizer(
3417 ac, CEPH_ENTITY_TYPE_MDS,
3419 &s->s_authorizer_buf,
3420 &s->s_authorizer_buf_len,
3421 &s->s_authorizer_reply_buf,
3422 &s->s_authorizer_reply_buf_len);
3428 *proto = ac->protocol;
3429 *buf = s->s_authorizer_buf;
3430 *len = s->s_authorizer_buf_len;
3431 *reply_buf = s->s_authorizer_reply_buf;
3432 *reply_len = s->s_authorizer_reply_buf_len;
3437 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3439 struct ceph_mds_session *s = con->private;
3440 struct ceph_mds_client *mdsc = s->s_mdsc;
3441 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3443 return ac->ops->verify_authorizer_reply(ac, s->s_authorizer, len);
3446 static int invalidate_authorizer(struct ceph_connection *con)
3448 struct ceph_mds_session *s = con->private;
3449 struct ceph_mds_client *mdsc = s->s_mdsc;
3450 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3452 if (ac->ops->invalidate_authorizer)
3453 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3455 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3458 static const struct ceph_connection_operations mds_con_ops = {
3461 .dispatch = dispatch,
3462 .get_authorizer = get_authorizer,
3463 .verify_authorizer_reply = verify_authorizer_reply,
3464 .invalidate_authorizer = invalidate_authorizer,
3465 .peer_reset = peer_reset,