1 #include <linux/ceph/ceph_debug.h>
4 #include <linux/wait.h>
5 #include <linux/slab.h>
7 #include <linux/sched.h>
8 #include <linux/debugfs.h>
9 #include <linux/seq_file.h>
12 #include "mds_client.h"
14 #include <linux/ceph/ceph_features.h>
15 #include <linux/ceph/messenger.h>
16 #include <linux/ceph/decode.h>
17 #include <linux/ceph/pagelist.h>
18 #include <linux/ceph/auth.h>
19 #include <linux/ceph/debugfs.h>
22 * A cluster of MDS (metadata server) daemons is responsible for
23 * managing the file system namespace (the directory hierarchy and
24 * inodes) and for coordinating shared access to storage. Metadata is
25 * partitioning hierarchically across a number of servers, and that
26 * partition varies over time as the cluster adjusts the distribution
27 * in order to balance load.
29 * The MDS client is primarily responsible to managing synchronous
30 * metadata requests for operations like open, unlink, and so forth.
31 * If there is a MDS failure, we find out about it when we (possibly
32 * request and) receive a new MDS map, and can resubmit affected
35 * For the most part, though, we take advantage of a lossless
36 * communications channel to the MDS, and do not need to worry about
37 * timing out or resubmitting requests.
39 * We maintain a stateful "session" with each MDS we interact with.
40 * Within each session, we sent periodic heartbeat messages to ensure
41 * any capabilities or leases we have been issues remain valid. If
42 * the session times out and goes stale, our leases and capabilities
43 * are no longer valid.
46 struct ceph_reconnect_state {
48 struct ceph_pagelist *pagelist;
52 static void __wake_requests(struct ceph_mds_client *mdsc,
53 struct list_head *head);
55 static const struct ceph_connection_operations mds_con_ops;
63 * parse individual inode info
65 static int parse_reply_info_in(void **p, void *end,
66 struct ceph_mds_reply_info_in *info,
72 *p += sizeof(struct ceph_mds_reply_inode) +
73 sizeof(*info->in->fragtree.splits) *
74 le32_to_cpu(info->in->fragtree.nsplits);
76 ceph_decode_32_safe(p, end, info->symlink_len, bad);
77 ceph_decode_need(p, end, info->symlink_len, bad);
79 *p += info->symlink_len;
81 if (features & CEPH_FEATURE_DIRLAYOUTHASH)
82 ceph_decode_copy_safe(p, end, &info->dir_layout,
83 sizeof(info->dir_layout), bad);
85 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
87 ceph_decode_32_safe(p, end, info->xattr_len, bad);
88 ceph_decode_need(p, end, info->xattr_len, bad);
89 info->xattr_data = *p;
90 *p += info->xattr_len;
97 * parse a normal reply, which may contain a (dir+)dentry and/or a
100 static int parse_reply_info_trace(void **p, void *end,
101 struct ceph_mds_reply_info_parsed *info,
106 if (info->head->is_dentry) {
107 err = parse_reply_info_in(p, end, &info->diri, features);
111 if (unlikely(*p + sizeof(*info->dirfrag) > end))
114 *p += sizeof(*info->dirfrag) +
115 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
116 if (unlikely(*p > end))
119 ceph_decode_32_safe(p, end, info->dname_len, bad);
120 ceph_decode_need(p, end, info->dname_len, bad);
122 *p += info->dname_len;
124 *p += sizeof(*info->dlease);
127 if (info->head->is_target) {
128 err = parse_reply_info_in(p, end, &info->targeti, features);
133 if (unlikely(*p != end))
140 pr_err("problem parsing mds trace %d\n", err);
145 * parse readdir results
147 static int parse_reply_info_dir(void **p, void *end,
148 struct ceph_mds_reply_info_parsed *info,
155 if (*p + sizeof(*info->dir_dir) > end)
157 *p += sizeof(*info->dir_dir) +
158 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
162 ceph_decode_need(p, end, sizeof(num) + 2, bad);
163 num = ceph_decode_32(p);
164 info->dir_end = ceph_decode_8(p);
165 info->dir_complete = ceph_decode_8(p);
169 BUG_ON(!info->dir_in);
170 info->dir_dname = (void *)(info->dir_in + num);
171 info->dir_dname_len = (void *)(info->dir_dname + num);
172 info->dir_dlease = (void *)(info->dir_dname_len + num);
173 if ((unsigned long)(info->dir_dlease + num) >
174 (unsigned long)info->dir_in + info->dir_buf_size) {
175 pr_err("dir contents are larger than expected\n");
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 create results
237 static int parse_reply_info_create(void **p, void *end,
238 struct ceph_mds_reply_info_parsed *info,
241 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
243 info->has_create_ino = false;
245 info->has_create_ino = true;
246 info->ino = ceph_decode_64(p);
250 if (unlikely(*p != end))
259 * parse extra results
261 static int parse_reply_info_extra(void **p, void *end,
262 struct ceph_mds_reply_info_parsed *info,
265 if (info->head->op == CEPH_MDS_OP_GETFILELOCK)
266 return parse_reply_info_filelock(p, end, info, features);
267 else if (info->head->op == CEPH_MDS_OP_READDIR ||
268 info->head->op == CEPH_MDS_OP_LSSNAP)
269 return parse_reply_info_dir(p, end, info, features);
270 else if (info->head->op == CEPH_MDS_OP_CREATE)
271 return parse_reply_info_create(p, end, info, features);
277 * parse entire mds reply
279 static int parse_reply_info(struct ceph_msg *msg,
280 struct ceph_mds_reply_info_parsed *info,
287 info->head = msg->front.iov_base;
288 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
289 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
292 ceph_decode_32_safe(&p, end, len, bad);
294 ceph_decode_need(&p, end, len, bad);
295 err = parse_reply_info_trace(&p, p+len, info, features);
301 ceph_decode_32_safe(&p, end, len, bad);
303 ceph_decode_need(&p, end, len, bad);
304 err = parse_reply_info_extra(&p, p+len, info, features);
310 ceph_decode_32_safe(&p, end, len, bad);
311 info->snapblob_len = len;
322 pr_err("mds parse_reply err %d\n", err);
326 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
330 free_pages((unsigned long)info->dir_in, get_order(info->dir_buf_size));
337 static const char *session_state_name(int s)
340 case CEPH_MDS_SESSION_NEW: return "new";
341 case CEPH_MDS_SESSION_OPENING: return "opening";
342 case CEPH_MDS_SESSION_OPEN: return "open";
343 case CEPH_MDS_SESSION_HUNG: return "hung";
344 case CEPH_MDS_SESSION_CLOSING: return "closing";
345 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
346 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
347 default: return "???";
351 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
353 if (atomic_inc_not_zero(&s->s_ref)) {
354 dout("mdsc get_session %p %d -> %d\n", s,
355 atomic_read(&s->s_ref)-1, atomic_read(&s->s_ref));
358 dout("mdsc get_session %p 0 -- FAIL", s);
363 void ceph_put_mds_session(struct ceph_mds_session *s)
365 dout("mdsc put_session %p %d -> %d\n", s,
366 atomic_read(&s->s_ref), atomic_read(&s->s_ref)-1);
367 if (atomic_dec_and_test(&s->s_ref)) {
368 if (s->s_auth.authorizer)
369 ceph_auth_destroy_authorizer(
370 s->s_mdsc->fsc->client->monc.auth,
371 s->s_auth.authorizer);
377 * called under mdsc->mutex
379 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
382 struct ceph_mds_session *session;
384 if (mds >= mdsc->max_sessions || mdsc->sessions[mds] == NULL)
386 session = mdsc->sessions[mds];
387 dout("lookup_mds_session %p %d\n", session,
388 atomic_read(&session->s_ref));
389 get_session(session);
393 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
395 if (mds >= mdsc->max_sessions)
397 return mdsc->sessions[mds];
400 static int __verify_registered_session(struct ceph_mds_client *mdsc,
401 struct ceph_mds_session *s)
403 if (s->s_mds >= mdsc->max_sessions ||
404 mdsc->sessions[s->s_mds] != s)
410 * create+register a new session for given mds.
411 * called under mdsc->mutex.
413 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
416 struct ceph_mds_session *s;
418 if (mds >= mdsc->mdsmap->m_max_mds)
419 return ERR_PTR(-EINVAL);
421 s = kzalloc(sizeof(*s), GFP_NOFS);
423 return ERR_PTR(-ENOMEM);
426 s->s_state = CEPH_MDS_SESSION_NEW;
429 mutex_init(&s->s_mutex);
431 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
433 spin_lock_init(&s->s_gen_ttl_lock);
435 s->s_cap_ttl = jiffies - 1;
437 spin_lock_init(&s->s_cap_lock);
438 s->s_renew_requested = 0;
440 INIT_LIST_HEAD(&s->s_caps);
443 atomic_set(&s->s_ref, 1);
444 INIT_LIST_HEAD(&s->s_waiting);
445 INIT_LIST_HEAD(&s->s_unsafe);
446 s->s_num_cap_releases = 0;
447 s->s_cap_reconnect = 0;
448 s->s_cap_iterator = NULL;
449 INIT_LIST_HEAD(&s->s_cap_releases);
450 INIT_LIST_HEAD(&s->s_cap_releases_done);
451 INIT_LIST_HEAD(&s->s_cap_flushing);
452 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
454 dout("register_session mds%d\n", mds);
455 if (mds >= mdsc->max_sessions) {
456 int newmax = 1 << get_count_order(mds+1);
457 struct ceph_mds_session **sa;
459 dout("register_session realloc to %d\n", newmax);
460 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
463 if (mdsc->sessions) {
464 memcpy(sa, mdsc->sessions,
465 mdsc->max_sessions * sizeof(void *));
466 kfree(mdsc->sessions);
469 mdsc->max_sessions = newmax;
471 mdsc->sessions[mds] = s;
472 atomic_inc(&s->s_ref); /* one ref to sessions[], one to caller */
474 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
475 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
481 return ERR_PTR(-ENOMEM);
485 * called under mdsc->mutex
487 static void __unregister_session(struct ceph_mds_client *mdsc,
488 struct ceph_mds_session *s)
490 dout("__unregister_session mds%d %p\n", s->s_mds, s);
491 BUG_ON(mdsc->sessions[s->s_mds] != s);
492 mdsc->sessions[s->s_mds] = NULL;
493 ceph_con_close(&s->s_con);
494 ceph_put_mds_session(s);
498 * drop session refs in request.
500 * should be last request ref, or hold mdsc->mutex
502 static void put_request_session(struct ceph_mds_request *req)
504 if (req->r_session) {
505 ceph_put_mds_session(req->r_session);
506 req->r_session = NULL;
510 void ceph_mdsc_release_request(struct kref *kref)
512 struct ceph_mds_request *req = container_of(kref,
513 struct ceph_mds_request,
515 destroy_reply_info(&req->r_reply_info);
517 ceph_msg_put(req->r_request);
519 ceph_msg_put(req->r_reply);
521 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
524 if (req->r_locked_dir)
525 ceph_put_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
526 if (req->r_target_inode)
527 iput(req->r_target_inode);
530 if (req->r_old_dentry)
531 dput(req->r_old_dentry);
532 if (req->r_old_dentry_dir) {
534 * track (and drop pins for) r_old_dentry_dir
535 * separately, since r_old_dentry's d_parent may have
536 * changed between the dir mutex being dropped and
537 * this request being freed.
539 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
541 iput(req->r_old_dentry_dir);
545 put_request_session(req);
546 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
551 * lookup session, bump ref if found.
553 * called under mdsc->mutex.
555 static struct ceph_mds_request *__lookup_request(struct ceph_mds_client *mdsc,
558 struct ceph_mds_request *req;
559 struct rb_node *n = mdsc->request_tree.rb_node;
562 req = rb_entry(n, struct ceph_mds_request, r_node);
563 if (tid < req->r_tid)
565 else if (tid > req->r_tid)
568 ceph_mdsc_get_request(req);
575 static void __insert_request(struct ceph_mds_client *mdsc,
576 struct ceph_mds_request *new)
578 struct rb_node **p = &mdsc->request_tree.rb_node;
579 struct rb_node *parent = NULL;
580 struct ceph_mds_request *req = NULL;
584 req = rb_entry(parent, struct ceph_mds_request, r_node);
585 if (new->r_tid < req->r_tid)
587 else if (new->r_tid > req->r_tid)
593 rb_link_node(&new->r_node, parent, p);
594 rb_insert_color(&new->r_node, &mdsc->request_tree);
598 * Register an in-flight request, and assign a tid. Link to directory
599 * are modifying (if any).
601 * Called under mdsc->mutex.
603 static void __register_request(struct ceph_mds_client *mdsc,
604 struct ceph_mds_request *req,
607 req->r_tid = ++mdsc->last_tid;
609 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
611 dout("__register_request %p tid %lld\n", req, req->r_tid);
612 ceph_mdsc_get_request(req);
613 __insert_request(mdsc, req);
615 req->r_uid = current_fsuid();
616 req->r_gid = current_fsgid();
619 struct ceph_inode_info *ci = ceph_inode(dir);
622 spin_lock(&ci->i_unsafe_lock);
623 req->r_unsafe_dir = dir;
624 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
625 spin_unlock(&ci->i_unsafe_lock);
629 static void __unregister_request(struct ceph_mds_client *mdsc,
630 struct ceph_mds_request *req)
632 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
633 rb_erase(&req->r_node, &mdsc->request_tree);
634 RB_CLEAR_NODE(&req->r_node);
636 if (req->r_unsafe_dir) {
637 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
639 spin_lock(&ci->i_unsafe_lock);
640 list_del_init(&req->r_unsafe_dir_item);
641 spin_unlock(&ci->i_unsafe_lock);
643 iput(req->r_unsafe_dir);
644 req->r_unsafe_dir = NULL;
647 complete_all(&req->r_safe_completion);
649 ceph_mdsc_put_request(req);
653 * Choose mds to send request to next. If there is a hint set in the
654 * request (e.g., due to a prior forward hint from the mds), use that.
655 * Otherwise, consult frag tree and/or caps to identify the
656 * appropriate mds. If all else fails, choose randomly.
658 * Called under mdsc->mutex.
660 static struct dentry *get_nonsnap_parent(struct dentry *dentry)
663 * we don't need to worry about protecting the d_parent access
664 * here because we never renaming inside the snapped namespace
665 * except to resplice to another snapdir, and either the old or new
666 * result is a valid result.
668 while (!IS_ROOT(dentry) && ceph_snap(dentry->d_inode) != CEPH_NOSNAP)
669 dentry = dentry->d_parent;
673 static int __choose_mds(struct ceph_mds_client *mdsc,
674 struct ceph_mds_request *req)
677 struct ceph_inode_info *ci;
678 struct ceph_cap *cap;
679 int mode = req->r_direct_mode;
681 u32 hash = req->r_direct_hash;
682 bool is_hash = req->r_direct_is_hash;
685 * is there a specific mds we should try? ignore hint if we have
686 * no session and the mds is not up (active or recovering).
688 if (req->r_resend_mds >= 0 &&
689 (__have_session(mdsc, req->r_resend_mds) ||
690 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
691 dout("choose_mds using resend_mds mds%d\n",
693 return req->r_resend_mds;
696 if (mode == USE_RANDOM_MDS)
701 inode = req->r_inode;
702 } else if (req->r_dentry) {
703 /* ignore race with rename; old or new d_parent is okay */
704 struct dentry *parent = req->r_dentry->d_parent;
705 struct inode *dir = parent->d_inode;
707 if (dir->i_sb != mdsc->fsc->sb) {
709 inode = req->r_dentry->d_inode;
710 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
711 /* direct snapped/virtual snapdir requests
712 * based on parent dir inode */
713 struct dentry *dn = get_nonsnap_parent(parent);
715 dout("__choose_mds using nonsnap parent %p\n", inode);
718 inode = req->r_dentry->d_inode;
719 if (!inode || mode == USE_AUTH_MDS) {
722 hash = ceph_dentry_hash(dir, req->r_dentry);
728 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
732 ci = ceph_inode(inode);
734 if (is_hash && S_ISDIR(inode->i_mode)) {
735 struct ceph_inode_frag frag;
738 ceph_choose_frag(ci, hash, &frag, &found);
740 if (mode == USE_ANY_MDS && frag.ndist > 0) {
743 /* choose a random replica */
744 get_random_bytes(&r, 1);
747 dout("choose_mds %p %llx.%llx "
748 "frag %u mds%d (%d/%d)\n",
749 inode, ceph_vinop(inode),
752 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
753 CEPH_MDS_STATE_ACTIVE)
757 /* since this file/dir wasn't known to be
758 * replicated, then we want to look for the
759 * authoritative mds. */
762 /* choose auth mds */
764 dout("choose_mds %p %llx.%llx "
765 "frag %u mds%d (auth)\n",
766 inode, ceph_vinop(inode), frag.frag, mds);
767 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
768 CEPH_MDS_STATE_ACTIVE)
774 spin_lock(&ci->i_ceph_lock);
776 if (mode == USE_AUTH_MDS)
777 cap = ci->i_auth_cap;
778 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
779 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
781 spin_unlock(&ci->i_ceph_lock);
784 mds = cap->session->s_mds;
785 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
786 inode, ceph_vinop(inode), mds,
787 cap == ci->i_auth_cap ? "auth " : "", cap);
788 spin_unlock(&ci->i_ceph_lock);
792 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
793 dout("choose_mds chose random mds%d\n", mds);
801 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
803 struct ceph_msg *msg;
804 struct ceph_mds_session_head *h;
806 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
809 pr_err("create_session_msg ENOMEM creating msg\n");
812 h = msg->front.iov_base;
813 h->op = cpu_to_le32(op);
814 h->seq = cpu_to_le64(seq);
819 * send session open request.
821 * called under mdsc->mutex
823 static int __open_session(struct ceph_mds_client *mdsc,
824 struct ceph_mds_session *session)
826 struct ceph_msg *msg;
828 int mds = session->s_mds;
830 /* wait for mds to go active? */
831 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
832 dout("open_session to mds%d (%s)\n", mds,
833 ceph_mds_state_name(mstate));
834 session->s_state = CEPH_MDS_SESSION_OPENING;
835 session->s_renew_requested = jiffies;
837 /* send connect message */
838 msg = create_session_msg(CEPH_SESSION_REQUEST_OPEN, session->s_seq);
841 ceph_con_send(&session->s_con, msg);
846 * open sessions for any export targets for the given mds
848 * called under mdsc->mutex
850 static struct ceph_mds_session *
851 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
853 struct ceph_mds_session *session;
855 session = __ceph_lookup_mds_session(mdsc, target);
857 session = register_session(mdsc, target);
861 if (session->s_state == CEPH_MDS_SESSION_NEW ||
862 session->s_state == CEPH_MDS_SESSION_CLOSING)
863 __open_session(mdsc, session);
868 struct ceph_mds_session *
869 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
871 struct ceph_mds_session *session;
873 dout("open_export_target_session to mds%d\n", target);
875 mutex_lock(&mdsc->mutex);
876 session = __open_export_target_session(mdsc, target);
877 mutex_unlock(&mdsc->mutex);
882 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
883 struct ceph_mds_session *session)
885 struct ceph_mds_info *mi;
886 struct ceph_mds_session *ts;
887 int i, mds = session->s_mds;
889 if (mds >= mdsc->mdsmap->m_max_mds)
892 mi = &mdsc->mdsmap->m_info[mds];
893 dout("open_export_target_sessions for mds%d (%d targets)\n",
894 session->s_mds, mi->num_export_targets);
896 for (i = 0; i < mi->num_export_targets; i++) {
897 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
899 ceph_put_mds_session(ts);
903 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
904 struct ceph_mds_session *session)
906 mutex_lock(&mdsc->mutex);
907 __open_export_target_sessions(mdsc, session);
908 mutex_unlock(&mdsc->mutex);
916 * Free preallocated cap messages assigned to this session
918 static void cleanup_cap_releases(struct ceph_mds_session *session)
920 struct ceph_msg *msg;
922 spin_lock(&session->s_cap_lock);
923 while (!list_empty(&session->s_cap_releases)) {
924 msg = list_first_entry(&session->s_cap_releases,
925 struct ceph_msg, list_head);
926 list_del_init(&msg->list_head);
929 while (!list_empty(&session->s_cap_releases_done)) {
930 msg = list_first_entry(&session->s_cap_releases_done,
931 struct ceph_msg, list_head);
932 list_del_init(&msg->list_head);
935 spin_unlock(&session->s_cap_lock);
939 * Helper to safely iterate over all caps associated with a session, with
940 * special care taken to handle a racing __ceph_remove_cap().
942 * Caller must hold session s_mutex.
944 static int iterate_session_caps(struct ceph_mds_session *session,
945 int (*cb)(struct inode *, struct ceph_cap *,
949 struct ceph_cap *cap;
950 struct inode *inode, *last_inode = NULL;
951 struct ceph_cap *old_cap = NULL;
954 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
955 spin_lock(&session->s_cap_lock);
956 p = session->s_caps.next;
957 while (p != &session->s_caps) {
958 cap = list_entry(p, struct ceph_cap, session_caps);
959 inode = igrab(&cap->ci->vfs_inode);
964 session->s_cap_iterator = cap;
965 spin_unlock(&session->s_cap_lock);
972 ceph_put_cap(session->s_mdsc, old_cap);
976 ret = cb(inode, cap, arg);
979 spin_lock(&session->s_cap_lock);
981 if (cap->ci == NULL) {
982 dout("iterate_session_caps finishing cap %p removal\n",
984 BUG_ON(cap->session != session);
985 list_del_init(&cap->session_caps);
986 session->s_nr_caps--;
988 old_cap = cap; /* put_cap it w/o locks held */
995 session->s_cap_iterator = NULL;
996 spin_unlock(&session->s_cap_lock);
1001 ceph_put_cap(session->s_mdsc, old_cap);
1006 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1009 struct ceph_inode_info *ci = ceph_inode(inode);
1012 dout("removing cap %p, ci is %p, inode is %p\n",
1013 cap, ci, &ci->vfs_inode);
1014 spin_lock(&ci->i_ceph_lock);
1015 __ceph_remove_cap(cap, false);
1016 if (!__ceph_is_any_real_caps(ci)) {
1017 struct ceph_mds_client *mdsc =
1018 ceph_sb_to_client(inode->i_sb)->mdsc;
1020 spin_lock(&mdsc->cap_dirty_lock);
1021 if (!list_empty(&ci->i_dirty_item)) {
1022 pr_info(" dropping dirty %s state for %p %lld\n",
1023 ceph_cap_string(ci->i_dirty_caps),
1024 inode, ceph_ino(inode));
1025 ci->i_dirty_caps = 0;
1026 list_del_init(&ci->i_dirty_item);
1029 if (!list_empty(&ci->i_flushing_item)) {
1030 pr_info(" dropping dirty+flushing %s state for %p %lld\n",
1031 ceph_cap_string(ci->i_flushing_caps),
1032 inode, ceph_ino(inode));
1033 ci->i_flushing_caps = 0;
1034 list_del_init(&ci->i_flushing_item);
1035 mdsc->num_cap_flushing--;
1038 if (drop && ci->i_wrbuffer_ref) {
1039 pr_info(" dropping dirty data for %p %lld\n",
1040 inode, ceph_ino(inode));
1041 ci->i_wrbuffer_ref = 0;
1042 ci->i_wrbuffer_ref_head = 0;
1045 spin_unlock(&mdsc->cap_dirty_lock);
1047 spin_unlock(&ci->i_ceph_lock);
1054 * caller must hold session s_mutex
1056 static void remove_session_caps(struct ceph_mds_session *session)
1058 dout("remove_session_caps on %p\n", session);
1059 iterate_session_caps(session, remove_session_caps_cb, NULL);
1061 spin_lock(&session->s_cap_lock);
1062 if (session->s_nr_caps > 0) {
1063 struct super_block *sb = session->s_mdsc->fsc->sb;
1064 struct inode *inode;
1065 struct ceph_cap *cap, *prev = NULL;
1066 struct ceph_vino vino;
1068 * iterate_session_caps() skips inodes that are being
1069 * deleted, we need to wait until deletions are complete.
1070 * __wait_on_freeing_inode() is designed for the job,
1071 * but it is not exported, so use lookup inode function
1074 while (!list_empty(&session->s_caps)) {
1075 cap = list_entry(session->s_caps.next,
1076 struct ceph_cap, session_caps);
1080 vino = cap->ci->i_vino;
1081 spin_unlock(&session->s_cap_lock);
1083 inode = ceph_find_inode(sb, vino);
1086 spin_lock(&session->s_cap_lock);
1089 spin_unlock(&session->s_cap_lock);
1091 BUG_ON(session->s_nr_caps > 0);
1092 BUG_ON(!list_empty(&session->s_cap_flushing));
1093 cleanup_cap_releases(session);
1097 * wake up any threads waiting on this session's caps. if the cap is
1098 * old (didn't get renewed on the client reconnect), remove it now.
1100 * caller must hold s_mutex.
1102 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1105 struct ceph_inode_info *ci = ceph_inode(inode);
1107 wake_up_all(&ci->i_cap_wq);
1109 spin_lock(&ci->i_ceph_lock);
1110 ci->i_wanted_max_size = 0;
1111 ci->i_requested_max_size = 0;
1112 spin_unlock(&ci->i_ceph_lock);
1117 static void wake_up_session_caps(struct ceph_mds_session *session,
1120 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1121 iterate_session_caps(session, wake_up_session_cb,
1122 (void *)(unsigned long)reconnect);
1126 * Send periodic message to MDS renewing all currently held caps. The
1127 * ack will reset the expiration for all caps from this session.
1129 * caller holds s_mutex
1131 static int send_renew_caps(struct ceph_mds_client *mdsc,
1132 struct ceph_mds_session *session)
1134 struct ceph_msg *msg;
1137 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1138 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1139 pr_info("mds%d caps stale\n", session->s_mds);
1140 session->s_renew_requested = jiffies;
1142 /* do not try to renew caps until a recovering mds has reconnected
1143 * with its clients. */
1144 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1145 if (state < CEPH_MDS_STATE_RECONNECT) {
1146 dout("send_renew_caps ignoring mds%d (%s)\n",
1147 session->s_mds, ceph_mds_state_name(state));
1151 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1152 ceph_mds_state_name(state));
1153 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1154 ++session->s_renew_seq);
1157 ceph_con_send(&session->s_con, msg);
1161 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1162 struct ceph_mds_session *session, u64 seq)
1164 struct ceph_msg *msg;
1166 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1167 session->s_mds, session_state_name(session->s_state), seq);
1168 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1171 ceph_con_send(&session->s_con, msg);
1177 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1179 * Called under session->s_mutex
1181 static void renewed_caps(struct ceph_mds_client *mdsc,
1182 struct ceph_mds_session *session, int is_renew)
1187 spin_lock(&session->s_cap_lock);
1188 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1190 session->s_cap_ttl = session->s_renew_requested +
1191 mdsc->mdsmap->m_session_timeout*HZ;
1194 if (time_before(jiffies, session->s_cap_ttl)) {
1195 pr_info("mds%d caps renewed\n", session->s_mds);
1198 pr_info("mds%d caps still stale\n", session->s_mds);
1201 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1202 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1203 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1204 spin_unlock(&session->s_cap_lock);
1207 wake_up_session_caps(session, 0);
1211 * send a session close request
1213 static int request_close_session(struct ceph_mds_client *mdsc,
1214 struct ceph_mds_session *session)
1216 struct ceph_msg *msg;
1218 dout("request_close_session mds%d state %s seq %lld\n",
1219 session->s_mds, session_state_name(session->s_state),
1221 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1224 ceph_con_send(&session->s_con, msg);
1229 * Called with s_mutex held.
1231 static int __close_session(struct ceph_mds_client *mdsc,
1232 struct ceph_mds_session *session)
1234 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1236 session->s_state = CEPH_MDS_SESSION_CLOSING;
1237 return request_close_session(mdsc, session);
1241 * Trim old(er) caps.
1243 * Because we can't cache an inode without one or more caps, we do
1244 * this indirectly: if a cap is unused, we prune its aliases, at which
1245 * point the inode will hopefully get dropped to.
1247 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1248 * memory pressure from the MDS, though, so it needn't be perfect.
1250 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1252 struct ceph_mds_session *session = arg;
1253 struct ceph_inode_info *ci = ceph_inode(inode);
1254 int used, wanted, oissued, mine;
1256 if (session->s_trim_caps <= 0)
1259 spin_lock(&ci->i_ceph_lock);
1260 mine = cap->issued | cap->implemented;
1261 used = __ceph_caps_used(ci);
1262 wanted = __ceph_caps_file_wanted(ci);
1263 oissued = __ceph_caps_issued_other(ci, cap);
1265 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1266 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1267 ceph_cap_string(used), ceph_cap_string(wanted));
1268 if (cap == ci->i_auth_cap) {
1269 if (ci->i_dirty_caps | ci->i_flushing_caps)
1271 if ((used | wanted) & CEPH_CAP_ANY_WR)
1274 if ((used | wanted) & ~oissued & mine)
1275 goto out; /* we need these caps */
1277 session->s_trim_caps--;
1279 /* we aren't the only cap.. just remove us */
1280 __ceph_remove_cap(cap, true);
1282 /* try to drop referring dentries */
1283 spin_unlock(&ci->i_ceph_lock);
1284 d_prune_aliases(inode);
1285 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1286 inode, cap, atomic_read(&inode->i_count));
1291 spin_unlock(&ci->i_ceph_lock);
1296 * Trim session cap count down to some max number.
1298 static int trim_caps(struct ceph_mds_client *mdsc,
1299 struct ceph_mds_session *session,
1302 int trim_caps = session->s_nr_caps - max_caps;
1304 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1305 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1306 if (trim_caps > 0) {
1307 session->s_trim_caps = trim_caps;
1308 iterate_session_caps(session, trim_caps_cb, session);
1309 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1310 session->s_mds, session->s_nr_caps, max_caps,
1311 trim_caps - session->s_trim_caps);
1312 session->s_trim_caps = 0;
1315 ceph_add_cap_releases(mdsc, session);
1316 ceph_send_cap_releases(mdsc, session);
1321 * Allocate cap_release messages. If there is a partially full message
1322 * in the queue, try to allocate enough to cover it's remainder, so that
1323 * we can send it immediately.
1325 * Called under s_mutex.
1327 int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1328 struct ceph_mds_session *session)
1330 struct ceph_msg *msg, *partial = NULL;
1331 struct ceph_mds_cap_release *head;
1333 int extra = mdsc->fsc->mount_options->cap_release_safety;
1336 dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds,
1339 spin_lock(&session->s_cap_lock);
1341 if (!list_empty(&session->s_cap_releases)) {
1342 msg = list_first_entry(&session->s_cap_releases,
1345 head = msg->front.iov_base;
1346 num = le32_to_cpu(head->num);
1348 dout(" partial %p with (%d/%d)\n", msg, num,
1349 (int)CEPH_CAPS_PER_RELEASE);
1350 extra += CEPH_CAPS_PER_RELEASE - num;
1354 while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1355 spin_unlock(&session->s_cap_lock);
1356 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1360 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1361 (int)msg->front.iov_len);
1362 head = msg->front.iov_base;
1363 head->num = cpu_to_le32(0);
1364 msg->front.iov_len = sizeof(*head);
1365 spin_lock(&session->s_cap_lock);
1366 list_add(&msg->list_head, &session->s_cap_releases);
1367 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1371 head = partial->front.iov_base;
1372 num = le32_to_cpu(head->num);
1373 dout(" queueing partial %p with %d/%d\n", partial, num,
1374 (int)CEPH_CAPS_PER_RELEASE);
1375 list_move_tail(&partial->list_head,
1376 &session->s_cap_releases_done);
1377 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1380 spin_unlock(&session->s_cap_lock);
1386 * flush all dirty inode data to disk.
1388 * returns true if we've flushed through want_flush_seq
1390 static int check_cap_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1394 dout("check_cap_flush want %lld\n", want_flush_seq);
1395 mutex_lock(&mdsc->mutex);
1396 for (mds = 0; ret && mds < mdsc->max_sessions; mds++) {
1397 struct ceph_mds_session *session = mdsc->sessions[mds];
1401 get_session(session);
1402 mutex_unlock(&mdsc->mutex);
1404 mutex_lock(&session->s_mutex);
1405 if (!list_empty(&session->s_cap_flushing)) {
1406 struct ceph_inode_info *ci =
1407 list_entry(session->s_cap_flushing.next,
1408 struct ceph_inode_info,
1410 struct inode *inode = &ci->vfs_inode;
1412 spin_lock(&ci->i_ceph_lock);
1413 if (ci->i_cap_flush_seq <= want_flush_seq) {
1414 dout("check_cap_flush still flushing %p "
1415 "seq %lld <= %lld to mds%d\n", inode,
1416 ci->i_cap_flush_seq, want_flush_seq,
1420 spin_unlock(&ci->i_ceph_lock);
1422 mutex_unlock(&session->s_mutex);
1423 ceph_put_mds_session(session);
1427 mutex_lock(&mdsc->mutex);
1430 mutex_unlock(&mdsc->mutex);
1431 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq);
1436 * called under s_mutex
1438 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1439 struct ceph_mds_session *session)
1441 struct ceph_msg *msg;
1443 dout("send_cap_releases mds%d\n", session->s_mds);
1444 spin_lock(&session->s_cap_lock);
1445 while (!list_empty(&session->s_cap_releases_done)) {
1446 msg = list_first_entry(&session->s_cap_releases_done,
1447 struct ceph_msg, list_head);
1448 list_del_init(&msg->list_head);
1449 spin_unlock(&session->s_cap_lock);
1450 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1451 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1452 ceph_con_send(&session->s_con, msg);
1453 spin_lock(&session->s_cap_lock);
1455 spin_unlock(&session->s_cap_lock);
1458 static void discard_cap_releases(struct ceph_mds_client *mdsc,
1459 struct ceph_mds_session *session)
1461 struct ceph_msg *msg;
1462 struct ceph_mds_cap_release *head;
1465 dout("discard_cap_releases mds%d\n", session->s_mds);
1467 if (!list_empty(&session->s_cap_releases)) {
1468 /* zero out the in-progress message */
1469 msg = list_first_entry(&session->s_cap_releases,
1470 struct ceph_msg, list_head);
1471 head = msg->front.iov_base;
1472 num = le32_to_cpu(head->num);
1473 dout("discard_cap_releases mds%d %p %u\n",
1474 session->s_mds, msg, num);
1475 head->num = cpu_to_le32(0);
1476 msg->front.iov_len = sizeof(*head);
1477 session->s_num_cap_releases += num;
1480 /* requeue completed messages */
1481 while (!list_empty(&session->s_cap_releases_done)) {
1482 msg = list_first_entry(&session->s_cap_releases_done,
1483 struct ceph_msg, list_head);
1484 list_del_init(&msg->list_head);
1486 head = msg->front.iov_base;
1487 num = le32_to_cpu(head->num);
1488 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg,
1490 session->s_num_cap_releases += num;
1491 head->num = cpu_to_le32(0);
1492 msg->front.iov_len = sizeof(*head);
1493 list_add(&msg->list_head, &session->s_cap_releases);
1501 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1504 struct ceph_inode_info *ci = ceph_inode(dir);
1505 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1506 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1507 size_t size = sizeof(*rinfo->dir_in) + sizeof(*rinfo->dir_dname_len) +
1508 sizeof(*rinfo->dir_dname) + sizeof(*rinfo->dir_dlease);
1509 int order, num_entries;
1511 spin_lock(&ci->i_ceph_lock);
1512 num_entries = ci->i_files + ci->i_subdirs;
1513 spin_unlock(&ci->i_ceph_lock);
1514 num_entries = max(num_entries, 1);
1515 num_entries = min(num_entries, opt->max_readdir);
1517 order = get_order(size * num_entries);
1518 while (order >= 0) {
1519 rinfo->dir_in = (void*)__get_free_pages(GFP_NOFS | __GFP_NOWARN,
1528 num_entries = (PAGE_SIZE << order) / size;
1529 num_entries = min(num_entries, opt->max_readdir);
1531 rinfo->dir_buf_size = PAGE_SIZE << order;
1532 req->r_num_caps = num_entries + 1;
1533 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1534 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1539 * Create an mds request.
1541 struct ceph_mds_request *
1542 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1544 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1547 return ERR_PTR(-ENOMEM);
1549 mutex_init(&req->r_fill_mutex);
1551 req->r_started = jiffies;
1552 req->r_resend_mds = -1;
1553 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1555 kref_init(&req->r_kref);
1556 INIT_LIST_HEAD(&req->r_wait);
1557 init_completion(&req->r_completion);
1558 init_completion(&req->r_safe_completion);
1559 INIT_LIST_HEAD(&req->r_unsafe_item);
1562 req->r_direct_mode = mode;
1567 * return oldest (lowest) request, tid in request tree, 0 if none.
1569 * called under mdsc->mutex.
1571 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1573 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1575 return rb_entry(rb_first(&mdsc->request_tree),
1576 struct ceph_mds_request, r_node);
1579 static u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1581 struct ceph_mds_request *req = __get_oldest_req(mdsc);
1589 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1590 * on build_path_from_dentry in fs/cifs/dir.c.
1592 * If @stop_on_nosnap, generate path relative to the first non-snapped
1595 * Encode hidden .snap dirs as a double /, i.e.
1596 * foo/.snap/bar -> foo//bar
1598 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1601 struct dentry *temp;
1607 return ERR_PTR(-EINVAL);
1611 seq = read_seqbegin(&rename_lock);
1613 for (temp = dentry; !IS_ROOT(temp);) {
1614 struct inode *inode = temp->d_inode;
1615 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1616 len++; /* slash only */
1617 else if (stop_on_nosnap && inode &&
1618 ceph_snap(inode) == CEPH_NOSNAP)
1621 len += 1 + temp->d_name.len;
1622 temp = temp->d_parent;
1626 len--; /* no leading '/' */
1628 path = kmalloc(len+1, GFP_NOFS);
1630 return ERR_PTR(-ENOMEM);
1632 path[pos] = 0; /* trailing null */
1634 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1635 struct inode *inode;
1637 spin_lock(&temp->d_lock);
1638 inode = temp->d_inode;
1639 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1640 dout("build_path path+%d: %p SNAPDIR\n",
1642 } else if (stop_on_nosnap && inode &&
1643 ceph_snap(inode) == CEPH_NOSNAP) {
1644 spin_unlock(&temp->d_lock);
1647 pos -= temp->d_name.len;
1649 spin_unlock(&temp->d_lock);
1652 strncpy(path + pos, temp->d_name.name,
1655 spin_unlock(&temp->d_lock);
1658 temp = temp->d_parent;
1661 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1662 pr_err("build_path did not end path lookup where "
1663 "expected, namelen is %d, pos is %d\n", len, pos);
1664 /* presumably this is only possible if racing with a
1665 rename of one of the parent directories (we can not
1666 lock the dentries above us to prevent this, but
1667 retrying should be harmless) */
1672 *base = ceph_ino(temp->d_inode);
1674 dout("build_path on %p %d built %llx '%.*s'\n",
1675 dentry, d_count(dentry), *base, len, path);
1679 static int build_dentry_path(struct dentry *dentry,
1680 const char **ppath, int *ppathlen, u64 *pino,
1685 if (ceph_snap(dentry->d_parent->d_inode) == CEPH_NOSNAP) {
1686 *pino = ceph_ino(dentry->d_parent->d_inode);
1687 *ppath = dentry->d_name.name;
1688 *ppathlen = dentry->d_name.len;
1691 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1693 return PTR_ERR(path);
1699 static int build_inode_path(struct inode *inode,
1700 const char **ppath, int *ppathlen, u64 *pino,
1703 struct dentry *dentry;
1706 if (ceph_snap(inode) == CEPH_NOSNAP) {
1707 *pino = ceph_ino(inode);
1711 dentry = d_find_alias(inode);
1712 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1715 return PTR_ERR(path);
1722 * request arguments may be specified via an inode *, a dentry *, or
1723 * an explicit ino+path.
1725 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1726 const char *rpath, u64 rino,
1727 const char **ppath, int *pathlen,
1728 u64 *ino, int *freepath)
1733 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1734 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1736 } else if (rdentry) {
1737 r = build_dentry_path(rdentry, ppath, pathlen, ino, freepath);
1738 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1740 } else if (rpath || rino) {
1743 *pathlen = rpath ? strlen(rpath) : 0;
1744 dout(" path %.*s\n", *pathlen, rpath);
1751 * called under mdsc->mutex
1753 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1754 struct ceph_mds_request *req,
1757 struct ceph_msg *msg;
1758 struct ceph_mds_request_head *head;
1759 const char *path1 = NULL;
1760 const char *path2 = NULL;
1761 u64 ino1 = 0, ino2 = 0;
1762 int pathlen1 = 0, pathlen2 = 0;
1763 int freepath1 = 0, freepath2 = 0;
1769 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1770 req->r_path1, req->r_ino1.ino,
1771 &path1, &pathlen1, &ino1, &freepath1);
1777 ret = set_request_path_attr(NULL, req->r_old_dentry,
1778 req->r_path2, req->r_ino2.ino,
1779 &path2, &pathlen2, &ino2, &freepath2);
1785 len = sizeof(*head) +
1786 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64));
1788 /* calculate (max) length for cap releases */
1789 len += sizeof(struct ceph_mds_request_release) *
1790 (!!req->r_inode_drop + !!req->r_dentry_drop +
1791 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1792 if (req->r_dentry_drop)
1793 len += req->r_dentry->d_name.len;
1794 if (req->r_old_dentry_drop)
1795 len += req->r_old_dentry->d_name.len;
1797 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1799 msg = ERR_PTR(-ENOMEM);
1803 msg->hdr.tid = cpu_to_le64(req->r_tid);
1805 head = msg->front.iov_base;
1806 p = msg->front.iov_base + sizeof(*head);
1807 end = msg->front.iov_base + msg->front.iov_len;
1809 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1810 head->op = cpu_to_le32(req->r_op);
1811 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1812 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1813 head->args = req->r_args;
1815 ceph_encode_filepath(&p, end, ino1, path1);
1816 ceph_encode_filepath(&p, end, ino2, path2);
1818 /* make note of release offset, in case we need to replay */
1819 req->r_request_release_offset = p - msg->front.iov_base;
1823 if (req->r_inode_drop)
1824 releases += ceph_encode_inode_release(&p,
1825 req->r_inode ? req->r_inode : req->r_dentry->d_inode,
1826 mds, req->r_inode_drop, req->r_inode_unless, 0);
1827 if (req->r_dentry_drop)
1828 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1829 mds, req->r_dentry_drop, req->r_dentry_unless);
1830 if (req->r_old_dentry_drop)
1831 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1832 mds, req->r_old_dentry_drop, req->r_old_dentry_unless);
1833 if (req->r_old_inode_drop)
1834 releases += ceph_encode_inode_release(&p,
1835 req->r_old_dentry->d_inode,
1836 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
1837 head->num_releases = cpu_to_le16(releases);
1840 msg->front.iov_len = p - msg->front.iov_base;
1841 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1843 if (req->r_data_len) {
1844 /* outbound data set only by ceph_sync_setxattr() */
1845 BUG_ON(!req->r_pages);
1846 ceph_msg_data_add_pages(msg, req->r_pages, req->r_data_len, 0);
1849 msg->hdr.data_len = cpu_to_le32(req->r_data_len);
1850 msg->hdr.data_off = cpu_to_le16(0);
1854 kfree((char *)path2);
1857 kfree((char *)path1);
1863 * called under mdsc->mutex if error, under no mutex if
1866 static void complete_request(struct ceph_mds_client *mdsc,
1867 struct ceph_mds_request *req)
1869 if (req->r_callback)
1870 req->r_callback(mdsc, req);
1872 complete_all(&req->r_completion);
1876 * called under mdsc->mutex
1878 static int __prepare_send_request(struct ceph_mds_client *mdsc,
1879 struct ceph_mds_request *req,
1882 struct ceph_mds_request_head *rhead;
1883 struct ceph_msg *msg;
1888 struct ceph_cap *cap =
1889 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
1892 req->r_sent_on_mseq = cap->mseq;
1894 req->r_sent_on_mseq = -1;
1896 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
1897 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
1899 if (req->r_got_unsafe) {
1901 * Replay. Do not regenerate message (and rebuild
1902 * paths, etc.); just use the original message.
1903 * Rebuilding paths will break for renames because
1904 * d_move mangles the src name.
1906 msg = req->r_request;
1907 rhead = msg->front.iov_base;
1909 flags = le32_to_cpu(rhead->flags);
1910 flags |= CEPH_MDS_FLAG_REPLAY;
1911 rhead->flags = cpu_to_le32(flags);
1913 if (req->r_target_inode)
1914 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
1916 rhead->num_retry = req->r_attempts - 1;
1918 /* remove cap/dentry releases from message */
1919 rhead->num_releases = 0;
1920 msg->hdr.front_len = cpu_to_le32(req->r_request_release_offset);
1921 msg->front.iov_len = req->r_request_release_offset;
1925 if (req->r_request) {
1926 ceph_msg_put(req->r_request);
1927 req->r_request = NULL;
1929 msg = create_request_message(mdsc, req, mds);
1931 req->r_err = PTR_ERR(msg);
1932 complete_request(mdsc, req);
1933 return PTR_ERR(msg);
1935 req->r_request = msg;
1937 rhead = msg->front.iov_base;
1938 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
1939 if (req->r_got_unsafe)
1940 flags |= CEPH_MDS_FLAG_REPLAY;
1941 if (req->r_locked_dir)
1942 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
1943 rhead->flags = cpu_to_le32(flags);
1944 rhead->num_fwd = req->r_num_fwd;
1945 rhead->num_retry = req->r_attempts - 1;
1948 dout(" r_locked_dir = %p\n", req->r_locked_dir);
1953 * send request, or put it on the appropriate wait list.
1955 static int __do_request(struct ceph_mds_client *mdsc,
1956 struct ceph_mds_request *req)
1958 struct ceph_mds_session *session = NULL;
1962 if (req->r_err || req->r_got_result) {
1964 __unregister_request(mdsc, req);
1968 if (req->r_timeout &&
1969 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
1970 dout("do_request timed out\n");
1975 put_request_session(req);
1977 mds = __choose_mds(mdsc, req);
1979 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
1980 dout("do_request no mds or not active, waiting for map\n");
1981 list_add(&req->r_wait, &mdsc->waiting_for_map);
1985 /* get, open session */
1986 session = __ceph_lookup_mds_session(mdsc, mds);
1988 session = register_session(mdsc, mds);
1989 if (IS_ERR(session)) {
1990 err = PTR_ERR(session);
1994 req->r_session = get_session(session);
1996 dout("do_request mds%d session %p state %s\n", mds, session,
1997 session_state_name(session->s_state));
1998 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
1999 session->s_state != CEPH_MDS_SESSION_HUNG) {
2000 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2001 session->s_state == CEPH_MDS_SESSION_CLOSING)
2002 __open_session(mdsc, session);
2003 list_add(&req->r_wait, &session->s_waiting);
2008 req->r_resend_mds = -1; /* forget any previous mds hint */
2010 if (req->r_request_started == 0) /* note request start time */
2011 req->r_request_started = jiffies;
2013 err = __prepare_send_request(mdsc, req, mds);
2015 ceph_msg_get(req->r_request);
2016 ceph_con_send(&session->s_con, req->r_request);
2020 ceph_put_mds_session(session);
2026 complete_request(mdsc, req);
2031 * called under mdsc->mutex
2033 static void __wake_requests(struct ceph_mds_client *mdsc,
2034 struct list_head *head)
2036 struct ceph_mds_request *req;
2037 LIST_HEAD(tmp_list);
2039 list_splice_init(head, &tmp_list);
2041 while (!list_empty(&tmp_list)) {
2042 req = list_entry(tmp_list.next,
2043 struct ceph_mds_request, r_wait);
2044 list_del_init(&req->r_wait);
2045 dout(" wake request %p tid %llu\n", req, req->r_tid);
2046 __do_request(mdsc, req);
2051 * Wake up threads with requests pending for @mds, so that they can
2052 * resubmit their requests to a possibly different mds.
2054 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2056 struct ceph_mds_request *req;
2059 dout("kick_requests mds%d\n", mds);
2060 for (p = rb_first(&mdsc->request_tree); p; p = rb_next(p)) {
2061 req = rb_entry(p, struct ceph_mds_request, r_node);
2062 if (req->r_got_unsafe)
2064 if (req->r_session &&
2065 req->r_session->s_mds == mds) {
2066 dout(" kicking tid %llu\n", req->r_tid);
2067 __do_request(mdsc, req);
2072 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2073 struct ceph_mds_request *req)
2075 dout("submit_request on %p\n", req);
2076 mutex_lock(&mdsc->mutex);
2077 __register_request(mdsc, req, NULL);
2078 __do_request(mdsc, req);
2079 mutex_unlock(&mdsc->mutex);
2083 * Synchrously perform an mds request. Take care of all of the
2084 * session setup, forwarding, retry details.
2086 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2088 struct ceph_mds_request *req)
2092 dout("do_request on %p\n", req);
2094 /* take CAP_PIN refs for r_inode, r_locked_dir, r_old_dentry */
2096 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2097 if (req->r_locked_dir)
2098 ceph_get_cap_refs(ceph_inode(req->r_locked_dir), CEPH_CAP_PIN);
2099 if (req->r_old_dentry_dir)
2100 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2104 mutex_lock(&mdsc->mutex);
2105 __register_request(mdsc, req, dir);
2106 __do_request(mdsc, req);
2110 __unregister_request(mdsc, req);
2111 dout("do_request early error %d\n", err);
2116 mutex_unlock(&mdsc->mutex);
2117 dout("do_request waiting\n");
2118 if (req->r_timeout) {
2119 err = (long)wait_for_completion_killable_timeout(
2120 &req->r_completion, req->r_timeout);
2124 err = wait_for_completion_killable(&req->r_completion);
2126 dout("do_request waited, got %d\n", err);
2127 mutex_lock(&mdsc->mutex);
2129 /* only abort if we didn't race with a real reply */
2130 if (req->r_got_result) {
2131 err = le32_to_cpu(req->r_reply_info.head->result);
2132 } else if (err < 0) {
2133 dout("aborted request %lld with %d\n", req->r_tid, err);
2136 * ensure we aren't running concurrently with
2137 * ceph_fill_trace or ceph_readdir_prepopulate, which
2138 * rely on locks (dir mutex) held by our caller.
2140 mutex_lock(&req->r_fill_mutex);
2142 req->r_aborted = true;
2143 mutex_unlock(&req->r_fill_mutex);
2145 if (req->r_locked_dir &&
2146 (req->r_op & CEPH_MDS_OP_WRITE))
2147 ceph_invalidate_dir_request(req);
2153 mutex_unlock(&mdsc->mutex);
2154 dout("do_request %p done, result %d\n", req, err);
2159 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2160 * namespace request.
2162 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2164 struct inode *inode = req->r_locked_dir;
2166 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2168 ceph_dir_clear_complete(inode);
2170 ceph_invalidate_dentry_lease(req->r_dentry);
2171 if (req->r_old_dentry)
2172 ceph_invalidate_dentry_lease(req->r_old_dentry);
2178 * We take the session mutex and parse and process the reply immediately.
2179 * This preserves the logical ordering of replies, capabilities, etc., sent
2180 * by the MDS as they are applied to our local cache.
2182 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2184 struct ceph_mds_client *mdsc = session->s_mdsc;
2185 struct ceph_mds_request *req;
2186 struct ceph_mds_reply_head *head = msg->front.iov_base;
2187 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2190 int mds = session->s_mds;
2192 if (msg->front.iov_len < sizeof(*head)) {
2193 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2198 /* get request, session */
2199 tid = le64_to_cpu(msg->hdr.tid);
2200 mutex_lock(&mdsc->mutex);
2201 req = __lookup_request(mdsc, tid);
2203 dout("handle_reply on unknown tid %llu\n", tid);
2204 mutex_unlock(&mdsc->mutex);
2207 dout("handle_reply %p\n", req);
2209 /* correct session? */
2210 if (req->r_session != session) {
2211 pr_err("mdsc_handle_reply got %llu on session mds%d"
2212 " not mds%d\n", tid, session->s_mds,
2213 req->r_session ? req->r_session->s_mds : -1);
2214 mutex_unlock(&mdsc->mutex);
2219 if ((req->r_got_unsafe && !head->safe) ||
2220 (req->r_got_safe && head->safe)) {
2221 pr_warning("got a dup %s reply on %llu from mds%d\n",
2222 head->safe ? "safe" : "unsafe", tid, mds);
2223 mutex_unlock(&mdsc->mutex);
2226 if (req->r_got_safe && !head->safe) {
2227 pr_warning("got unsafe after safe on %llu from mds%d\n",
2229 mutex_unlock(&mdsc->mutex);
2233 result = le32_to_cpu(head->result);
2237 * if we're not talking to the authority, send to them
2238 * if the authority has changed while we weren't looking,
2239 * send to new authority
2240 * Otherwise we just have to return an ESTALE
2242 if (result == -ESTALE) {
2243 dout("got ESTALE on request %llu", req->r_tid);
2244 if (req->r_direct_mode != USE_AUTH_MDS) {
2245 dout("not using auth, setting for that now");
2246 req->r_direct_mode = USE_AUTH_MDS;
2247 __do_request(mdsc, req);
2248 mutex_unlock(&mdsc->mutex);
2251 int mds = __choose_mds(mdsc, req);
2252 if (mds >= 0 && mds != req->r_session->s_mds) {
2253 dout("but auth changed, so resending");
2254 __do_request(mdsc, req);
2255 mutex_unlock(&mdsc->mutex);
2259 dout("have to return ESTALE on request %llu", req->r_tid);
2264 req->r_got_safe = true;
2265 __unregister_request(mdsc, req);
2267 if (req->r_got_unsafe) {
2269 * We already handled the unsafe response, now do the
2270 * cleanup. No need to examine the response; the MDS
2271 * doesn't include any result info in the safe
2272 * response. And even if it did, there is nothing
2273 * useful we could do with a revised return value.
2275 dout("got safe reply %llu, mds%d\n", tid, mds);
2276 list_del_init(&req->r_unsafe_item);
2278 /* last unsafe request during umount? */
2279 if (mdsc->stopping && !__get_oldest_req(mdsc))
2280 complete_all(&mdsc->safe_umount_waiters);
2281 mutex_unlock(&mdsc->mutex);
2285 req->r_got_unsafe = true;
2286 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2289 dout("handle_reply tid %lld result %d\n", tid, result);
2290 rinfo = &req->r_reply_info;
2291 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2292 mutex_unlock(&mdsc->mutex);
2294 mutex_lock(&session->s_mutex);
2296 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2302 if (rinfo->snapblob_len) {
2303 down_write(&mdsc->snap_rwsem);
2304 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2305 rinfo->snapblob + rinfo->snapblob_len,
2306 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP);
2307 downgrade_write(&mdsc->snap_rwsem);
2309 down_read(&mdsc->snap_rwsem);
2312 /* insert trace into our cache */
2313 mutex_lock(&req->r_fill_mutex);
2314 err = ceph_fill_trace(mdsc->fsc->sb, req, req->r_session);
2316 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2317 req->r_op == CEPH_MDS_OP_LSSNAP))
2318 ceph_readdir_prepopulate(req, req->r_session);
2319 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2321 mutex_unlock(&req->r_fill_mutex);
2323 up_read(&mdsc->snap_rwsem);
2325 mutex_lock(&mdsc->mutex);
2326 if (!req->r_aborted) {
2332 req->r_got_result = true;
2335 dout("reply arrived after request %lld was aborted\n", tid);
2337 mutex_unlock(&mdsc->mutex);
2339 ceph_add_cap_releases(mdsc, req->r_session);
2340 mutex_unlock(&session->s_mutex);
2342 /* kick calling process */
2343 complete_request(mdsc, req);
2345 ceph_mdsc_put_request(req);
2352 * handle mds notification that our request has been forwarded.
2354 static void handle_forward(struct ceph_mds_client *mdsc,
2355 struct ceph_mds_session *session,
2356 struct ceph_msg *msg)
2358 struct ceph_mds_request *req;
2359 u64 tid = le64_to_cpu(msg->hdr.tid);
2363 void *p = msg->front.iov_base;
2364 void *end = p + msg->front.iov_len;
2366 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2367 next_mds = ceph_decode_32(&p);
2368 fwd_seq = ceph_decode_32(&p);
2370 mutex_lock(&mdsc->mutex);
2371 req = __lookup_request(mdsc, tid);
2373 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2374 goto out; /* dup reply? */
2377 if (req->r_aborted) {
2378 dout("forward tid %llu aborted, unregistering\n", tid);
2379 __unregister_request(mdsc, req);
2380 } else if (fwd_seq <= req->r_num_fwd) {
2381 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2382 tid, next_mds, req->r_num_fwd, fwd_seq);
2384 /* resend. forward race not possible; mds would drop */
2385 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2387 BUG_ON(req->r_got_result);
2388 req->r_num_fwd = fwd_seq;
2389 req->r_resend_mds = next_mds;
2390 put_request_session(req);
2391 __do_request(mdsc, req);
2393 ceph_mdsc_put_request(req);
2395 mutex_unlock(&mdsc->mutex);
2399 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2403 * handle a mds session control message
2405 static void handle_session(struct ceph_mds_session *session,
2406 struct ceph_msg *msg)
2408 struct ceph_mds_client *mdsc = session->s_mdsc;
2411 int mds = session->s_mds;
2412 struct ceph_mds_session_head *h = msg->front.iov_base;
2416 if (msg->front.iov_len != sizeof(*h))
2418 op = le32_to_cpu(h->op);
2419 seq = le64_to_cpu(h->seq);
2421 mutex_lock(&mdsc->mutex);
2422 if (op == CEPH_SESSION_CLOSE)
2423 __unregister_session(mdsc, session);
2424 /* FIXME: this ttl calculation is generous */
2425 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2426 mutex_unlock(&mdsc->mutex);
2428 mutex_lock(&session->s_mutex);
2430 dout("handle_session mds%d %s %p state %s seq %llu\n",
2431 mds, ceph_session_op_name(op), session,
2432 session_state_name(session->s_state), seq);
2434 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2435 session->s_state = CEPH_MDS_SESSION_OPEN;
2436 pr_info("mds%d came back\n", session->s_mds);
2440 case CEPH_SESSION_OPEN:
2441 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2442 pr_info("mds%d reconnect success\n", session->s_mds);
2443 session->s_state = CEPH_MDS_SESSION_OPEN;
2444 renewed_caps(mdsc, session, 0);
2447 __close_session(mdsc, session);
2450 case CEPH_SESSION_RENEWCAPS:
2451 if (session->s_renew_seq == seq)
2452 renewed_caps(mdsc, session, 1);
2455 case CEPH_SESSION_CLOSE:
2456 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2457 pr_info("mds%d reconnect denied\n", session->s_mds);
2458 remove_session_caps(session);
2459 wake = 1; /* for good measure */
2460 wake_up_all(&mdsc->session_close_wq);
2461 kick_requests(mdsc, mds);
2464 case CEPH_SESSION_STALE:
2465 pr_info("mds%d caps went stale, renewing\n",
2467 spin_lock(&session->s_gen_ttl_lock);
2468 session->s_cap_gen++;
2469 session->s_cap_ttl = jiffies - 1;
2470 spin_unlock(&session->s_gen_ttl_lock);
2471 send_renew_caps(mdsc, session);
2474 case CEPH_SESSION_RECALL_STATE:
2475 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2478 case CEPH_SESSION_FLUSHMSG:
2479 send_flushmsg_ack(mdsc, session, seq);
2483 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2487 mutex_unlock(&session->s_mutex);
2489 mutex_lock(&mdsc->mutex);
2490 __wake_requests(mdsc, &session->s_waiting);
2491 mutex_unlock(&mdsc->mutex);
2496 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2497 (int)msg->front.iov_len);
2504 * called under session->mutex.
2506 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2507 struct ceph_mds_session *session)
2509 struct ceph_mds_request *req, *nreq;
2512 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2514 mutex_lock(&mdsc->mutex);
2515 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2516 err = __prepare_send_request(mdsc, req, session->s_mds);
2518 ceph_msg_get(req->r_request);
2519 ceph_con_send(&session->s_con, req->r_request);
2522 mutex_unlock(&mdsc->mutex);
2526 * Encode information about a cap for a reconnect with the MDS.
2528 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2532 struct ceph_mds_cap_reconnect v2;
2533 struct ceph_mds_cap_reconnect_v1 v1;
2536 struct ceph_inode_info *ci;
2537 struct ceph_reconnect_state *recon_state = arg;
2538 struct ceph_pagelist *pagelist = recon_state->pagelist;
2542 struct dentry *dentry;
2546 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2547 inode, ceph_vinop(inode), cap, cap->cap_id,
2548 ceph_cap_string(cap->issued));
2549 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2553 dentry = d_find_alias(inode);
2555 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2557 err = PTR_ERR(path);
2564 err = ceph_pagelist_encode_string(pagelist, path, pathlen);
2568 spin_lock(&ci->i_ceph_lock);
2569 cap->seq = 0; /* reset cap seq */
2570 cap->issue_seq = 0; /* and issue_seq */
2571 cap->mseq = 0; /* and migrate_seq */
2572 cap->cap_gen = cap->session->s_cap_gen;
2574 if (recon_state->flock) {
2575 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2576 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2577 rec.v2.issued = cpu_to_le32(cap->issued);
2578 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2579 rec.v2.pathbase = cpu_to_le64(pathbase);
2580 rec.v2.flock_len = 0;
2581 reclen = sizeof(rec.v2);
2583 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2584 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2585 rec.v1.issued = cpu_to_le32(cap->issued);
2586 rec.v1.size = cpu_to_le64(inode->i_size);
2587 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2588 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2589 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2590 rec.v1.pathbase = cpu_to_le64(pathbase);
2591 reclen = sizeof(rec.v1);
2593 spin_unlock(&ci->i_ceph_lock);
2595 if (recon_state->flock) {
2596 int num_fcntl_locks, num_flock_locks;
2597 struct ceph_filelock *flocks;
2600 spin_lock(&inode->i_lock);
2601 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2602 spin_unlock(&inode->i_lock);
2603 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2604 sizeof(struct ceph_filelock), GFP_NOFS);
2609 spin_lock(&inode->i_lock);
2610 err = ceph_encode_locks_to_buffer(inode, flocks,
2613 spin_unlock(&inode->i_lock);
2621 * number of encoded locks is stable, so copy to pagelist
2623 rec.v2.flock_len = cpu_to_le32(2*sizeof(u32) +
2624 (num_fcntl_locks+num_flock_locks) *
2625 sizeof(struct ceph_filelock));
2626 err = ceph_pagelist_append(pagelist, &rec, reclen);
2628 err = ceph_locks_to_pagelist(flocks, pagelist,
2633 err = ceph_pagelist_append(pagelist, &rec, reclen);
2636 recon_state->nr_caps++;
2646 * If an MDS fails and recovers, clients need to reconnect in order to
2647 * reestablish shared state. This includes all caps issued through
2648 * this session _and_ the snap_realm hierarchy. Because it's not
2649 * clear which snap realms the mds cares about, we send everything we
2650 * know about.. that ensures we'll then get any new info the
2651 * recovering MDS might have.
2653 * This is a relatively heavyweight operation, but it's rare.
2655 * called with mdsc->mutex held.
2657 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2658 struct ceph_mds_session *session)
2660 struct ceph_msg *reply;
2662 int mds = session->s_mds;
2665 struct ceph_pagelist *pagelist;
2666 struct ceph_reconnect_state recon_state;
2668 pr_info("mds%d reconnect start\n", mds);
2670 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2672 goto fail_nopagelist;
2673 ceph_pagelist_init(pagelist);
2675 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
2679 mutex_lock(&session->s_mutex);
2680 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
2683 ceph_con_close(&session->s_con);
2684 ceph_con_open(&session->s_con,
2685 CEPH_ENTITY_TYPE_MDS, mds,
2686 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
2688 /* replay unsafe requests */
2689 replay_unsafe_requests(mdsc, session);
2691 down_read(&mdsc->snap_rwsem);
2693 dout("session %p state %s\n", session,
2694 session_state_name(session->s_state));
2696 spin_lock(&session->s_gen_ttl_lock);
2697 session->s_cap_gen++;
2698 spin_unlock(&session->s_gen_ttl_lock);
2700 spin_lock(&session->s_cap_lock);
2702 * notify __ceph_remove_cap() that we are composing cap reconnect.
2703 * If a cap get released before being added to the cap reconnect,
2704 * __ceph_remove_cap() should skip queuing cap release.
2706 session->s_cap_reconnect = 1;
2707 /* drop old cap expires; we're about to reestablish that state */
2708 discard_cap_releases(mdsc, session);
2709 spin_unlock(&session->s_cap_lock);
2711 /* traverse this session's caps */
2712 s_nr_caps = session->s_nr_caps;
2713 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
2717 recon_state.nr_caps = 0;
2718 recon_state.pagelist = pagelist;
2719 recon_state.flock = session->s_con.peer_features & CEPH_FEATURE_FLOCK;
2720 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
2724 spin_lock(&session->s_cap_lock);
2725 session->s_cap_reconnect = 0;
2726 spin_unlock(&session->s_cap_lock);
2729 * snaprealms. we provide mds with the ino, seq (version), and
2730 * parent for all of our realms. If the mds has any newer info,
2733 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
2734 struct ceph_snap_realm *realm =
2735 rb_entry(p, struct ceph_snap_realm, node);
2736 struct ceph_mds_snaprealm_reconnect sr_rec;
2738 dout(" adding snap realm %llx seq %lld parent %llx\n",
2739 realm->ino, realm->seq, realm->parent_ino);
2740 sr_rec.ino = cpu_to_le64(realm->ino);
2741 sr_rec.seq = cpu_to_le64(realm->seq);
2742 sr_rec.parent = cpu_to_le64(realm->parent_ino);
2743 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
2748 if (recon_state.flock)
2749 reply->hdr.version = cpu_to_le16(2);
2751 /* raced with cap release? */
2752 if (s_nr_caps != recon_state.nr_caps) {
2753 struct page *page = list_first_entry(&pagelist->head,
2755 __le32 *addr = kmap_atomic(page);
2756 *addr = cpu_to_le32(recon_state.nr_caps);
2757 kunmap_atomic(addr);
2760 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2761 ceph_msg_data_add_pagelist(reply, pagelist);
2762 ceph_con_send(&session->s_con, reply);
2764 mutex_unlock(&session->s_mutex);
2766 mutex_lock(&mdsc->mutex);
2767 __wake_requests(mdsc, &session->s_waiting);
2768 mutex_unlock(&mdsc->mutex);
2770 up_read(&mdsc->snap_rwsem);
2774 ceph_msg_put(reply);
2775 up_read(&mdsc->snap_rwsem);
2776 mutex_unlock(&session->s_mutex);
2778 ceph_pagelist_release(pagelist);
2781 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
2787 * compare old and new mdsmaps, kicking requests
2788 * and closing out old connections as necessary
2790 * called under mdsc->mutex.
2792 static void check_new_map(struct ceph_mds_client *mdsc,
2793 struct ceph_mdsmap *newmap,
2794 struct ceph_mdsmap *oldmap)
2797 int oldstate, newstate;
2798 struct ceph_mds_session *s;
2800 dout("check_new_map new %u old %u\n",
2801 newmap->m_epoch, oldmap->m_epoch);
2803 for (i = 0; i < oldmap->m_max_mds && i < mdsc->max_sessions; i++) {
2804 if (mdsc->sessions[i] == NULL)
2806 s = mdsc->sessions[i];
2807 oldstate = ceph_mdsmap_get_state(oldmap, i);
2808 newstate = ceph_mdsmap_get_state(newmap, i);
2810 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
2811 i, ceph_mds_state_name(oldstate),
2812 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
2813 ceph_mds_state_name(newstate),
2814 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
2815 session_state_name(s->s_state));
2817 if (i >= newmap->m_max_mds ||
2818 memcmp(ceph_mdsmap_get_addr(oldmap, i),
2819 ceph_mdsmap_get_addr(newmap, i),
2820 sizeof(struct ceph_entity_addr))) {
2821 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
2822 /* the session never opened, just close it
2824 __wake_requests(mdsc, &s->s_waiting);
2825 __unregister_session(mdsc, s);
2828 mutex_unlock(&mdsc->mutex);
2829 mutex_lock(&s->s_mutex);
2830 mutex_lock(&mdsc->mutex);
2831 ceph_con_close(&s->s_con);
2832 mutex_unlock(&s->s_mutex);
2833 s->s_state = CEPH_MDS_SESSION_RESTARTING;
2836 /* kick any requests waiting on the recovering mds */
2837 kick_requests(mdsc, i);
2838 } else if (oldstate == newstate) {
2839 continue; /* nothing new with this mds */
2845 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
2846 newstate >= CEPH_MDS_STATE_RECONNECT) {
2847 mutex_unlock(&mdsc->mutex);
2848 send_mds_reconnect(mdsc, s);
2849 mutex_lock(&mdsc->mutex);
2853 * kick request on any mds that has gone active.
2855 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
2856 newstate >= CEPH_MDS_STATE_ACTIVE) {
2857 if (oldstate != CEPH_MDS_STATE_CREATING &&
2858 oldstate != CEPH_MDS_STATE_STARTING)
2859 pr_info("mds%d recovery completed\n", s->s_mds);
2860 kick_requests(mdsc, i);
2861 ceph_kick_flushing_caps(mdsc, s);
2862 wake_up_session_caps(s, 1);
2866 for (i = 0; i < newmap->m_max_mds && i < mdsc->max_sessions; i++) {
2867 s = mdsc->sessions[i];
2870 if (!ceph_mdsmap_is_laggy(newmap, i))
2872 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
2873 s->s_state == CEPH_MDS_SESSION_HUNG ||
2874 s->s_state == CEPH_MDS_SESSION_CLOSING) {
2875 dout(" connecting to export targets of laggy mds%d\n",
2877 __open_export_target_sessions(mdsc, s);
2889 * caller must hold session s_mutex, dentry->d_lock
2891 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
2893 struct ceph_dentry_info *di = ceph_dentry(dentry);
2895 ceph_put_mds_session(di->lease_session);
2896 di->lease_session = NULL;
2899 static void handle_lease(struct ceph_mds_client *mdsc,
2900 struct ceph_mds_session *session,
2901 struct ceph_msg *msg)
2903 struct super_block *sb = mdsc->fsc->sb;
2904 struct inode *inode;
2905 struct dentry *parent, *dentry;
2906 struct ceph_dentry_info *di;
2907 int mds = session->s_mds;
2908 struct ceph_mds_lease *h = msg->front.iov_base;
2910 struct ceph_vino vino;
2914 dout("handle_lease from mds%d\n", mds);
2917 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
2919 vino.ino = le64_to_cpu(h->ino);
2920 vino.snap = CEPH_NOSNAP;
2921 seq = le32_to_cpu(h->seq);
2922 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
2923 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
2924 if (dname.len != get_unaligned_le32(h+1))
2927 mutex_lock(&session->s_mutex);
2931 inode = ceph_find_inode(sb, vino);
2932 dout("handle_lease %s, ino %llx %p %.*s\n",
2933 ceph_lease_op_name(h->action), vino.ino, inode,
2934 dname.len, dname.name);
2935 if (inode == NULL) {
2936 dout("handle_lease no inode %llx\n", vino.ino);
2941 parent = d_find_alias(inode);
2943 dout("no parent dentry on inode %p\n", inode);
2945 goto release; /* hrm... */
2947 dname.hash = full_name_hash(dname.name, dname.len);
2948 dentry = d_lookup(parent, &dname);
2953 spin_lock(&dentry->d_lock);
2954 di = ceph_dentry(dentry);
2955 switch (h->action) {
2956 case CEPH_MDS_LEASE_REVOKE:
2957 if (di->lease_session == session) {
2958 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
2959 h->seq = cpu_to_le32(di->lease_seq);
2960 __ceph_mdsc_drop_dentry_lease(dentry);
2965 case CEPH_MDS_LEASE_RENEW:
2966 if (di->lease_session == session &&
2967 di->lease_gen == session->s_cap_gen &&
2968 di->lease_renew_from &&
2969 di->lease_renew_after == 0) {
2970 unsigned long duration =
2971 le32_to_cpu(h->duration_ms) * HZ / 1000;
2973 di->lease_seq = seq;
2974 dentry->d_time = di->lease_renew_from + duration;
2975 di->lease_renew_after = di->lease_renew_from +
2977 di->lease_renew_from = 0;
2981 spin_unlock(&dentry->d_lock);
2988 /* let's just reuse the same message */
2989 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
2991 ceph_con_send(&session->s_con, msg);
2995 mutex_unlock(&session->s_mutex);
2999 pr_err("corrupt lease message\n");
3003 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3004 struct inode *inode,
3005 struct dentry *dentry, char action,
3008 struct ceph_msg *msg;
3009 struct ceph_mds_lease *lease;
3010 int len = sizeof(*lease) + sizeof(u32);
3013 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3014 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3015 dnamelen = dentry->d_name.len;
3018 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3021 lease = msg->front.iov_base;
3022 lease->action = action;
3023 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3024 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3025 lease->seq = cpu_to_le32(seq);
3026 put_unaligned_le32(dnamelen, lease + 1);
3027 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3030 * if this is a preemptive lease RELEASE, no need to
3031 * flush request stream, since the actual request will
3034 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3036 ceph_con_send(&session->s_con, msg);
3040 * Preemptively release a lease we expect to invalidate anyway.
3041 * Pass @inode always, @dentry is optional.
3043 void ceph_mdsc_lease_release(struct ceph_mds_client *mdsc, struct inode *inode,
3044 struct dentry *dentry)
3046 struct ceph_dentry_info *di;
3047 struct ceph_mds_session *session;
3050 BUG_ON(inode == NULL);
3051 BUG_ON(dentry == NULL);
3053 /* is dentry lease valid? */
3054 spin_lock(&dentry->d_lock);
3055 di = ceph_dentry(dentry);
3056 if (!di || !di->lease_session ||
3057 di->lease_session->s_mds < 0 ||
3058 di->lease_gen != di->lease_session->s_cap_gen ||
3059 !time_before(jiffies, dentry->d_time)) {
3060 dout("lease_release inode %p dentry %p -- "
3063 spin_unlock(&dentry->d_lock);
3067 /* we do have a lease on this dentry; note mds and seq */
3068 session = ceph_get_mds_session(di->lease_session);
3069 seq = di->lease_seq;
3070 __ceph_mdsc_drop_dentry_lease(dentry);
3071 spin_unlock(&dentry->d_lock);
3073 dout("lease_release inode %p dentry %p to mds%d\n",
3074 inode, dentry, session->s_mds);
3075 ceph_mdsc_lease_send_msg(session, inode, dentry,
3076 CEPH_MDS_LEASE_RELEASE, seq);
3077 ceph_put_mds_session(session);
3081 * drop all leases (and dentry refs) in preparation for umount
3083 static void drop_leases(struct ceph_mds_client *mdsc)
3087 dout("drop_leases\n");
3088 mutex_lock(&mdsc->mutex);
3089 for (i = 0; i < mdsc->max_sessions; i++) {
3090 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3093 mutex_unlock(&mdsc->mutex);
3094 mutex_lock(&s->s_mutex);
3095 mutex_unlock(&s->s_mutex);
3096 ceph_put_mds_session(s);
3097 mutex_lock(&mdsc->mutex);
3099 mutex_unlock(&mdsc->mutex);
3105 * delayed work -- periodically trim expired leases, renew caps with mds
3107 static void schedule_delayed(struct ceph_mds_client *mdsc)
3110 unsigned hz = round_jiffies_relative(HZ * delay);
3111 schedule_delayed_work(&mdsc->delayed_work, hz);
3114 static void delayed_work(struct work_struct *work)
3117 struct ceph_mds_client *mdsc =
3118 container_of(work, struct ceph_mds_client, delayed_work.work);
3122 dout("mdsc delayed_work\n");
3123 ceph_check_delayed_caps(mdsc);
3125 mutex_lock(&mdsc->mutex);
3126 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3127 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3128 mdsc->last_renew_caps);
3130 mdsc->last_renew_caps = jiffies;
3132 for (i = 0; i < mdsc->max_sessions; i++) {
3133 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3136 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3137 dout("resending session close request for mds%d\n",
3139 request_close_session(mdsc, s);
3140 ceph_put_mds_session(s);
3143 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3144 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3145 s->s_state = CEPH_MDS_SESSION_HUNG;
3146 pr_info("mds%d hung\n", s->s_mds);
3149 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3150 /* this mds is failed or recovering, just wait */
3151 ceph_put_mds_session(s);
3154 mutex_unlock(&mdsc->mutex);
3156 mutex_lock(&s->s_mutex);
3158 send_renew_caps(mdsc, s);
3160 ceph_con_keepalive(&s->s_con);
3161 ceph_add_cap_releases(mdsc, s);
3162 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3163 s->s_state == CEPH_MDS_SESSION_HUNG)
3164 ceph_send_cap_releases(mdsc, s);
3165 mutex_unlock(&s->s_mutex);
3166 ceph_put_mds_session(s);
3168 mutex_lock(&mdsc->mutex);
3170 mutex_unlock(&mdsc->mutex);
3172 schedule_delayed(mdsc);
3175 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3178 struct ceph_mds_client *mdsc;
3180 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3185 mutex_init(&mdsc->mutex);
3186 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3187 if (mdsc->mdsmap == NULL) {
3192 init_completion(&mdsc->safe_umount_waiters);
3193 init_waitqueue_head(&mdsc->session_close_wq);
3194 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3195 mdsc->sessions = NULL;
3196 mdsc->max_sessions = 0;
3198 init_rwsem(&mdsc->snap_rwsem);
3199 mdsc->snap_realms = RB_ROOT;
3200 INIT_LIST_HEAD(&mdsc->snap_empty);
3201 spin_lock_init(&mdsc->snap_empty_lock);
3203 mdsc->request_tree = RB_ROOT;
3204 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3205 mdsc->last_renew_caps = jiffies;
3206 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3207 spin_lock_init(&mdsc->cap_delay_lock);
3208 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3209 spin_lock_init(&mdsc->snap_flush_lock);
3210 mdsc->cap_flush_seq = 0;
3211 INIT_LIST_HEAD(&mdsc->cap_dirty);
3212 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3213 mdsc->num_cap_flushing = 0;
3214 spin_lock_init(&mdsc->cap_dirty_lock);
3215 init_waitqueue_head(&mdsc->cap_flushing_wq);
3216 spin_lock_init(&mdsc->dentry_lru_lock);
3217 INIT_LIST_HEAD(&mdsc->dentry_lru);
3219 ceph_caps_init(mdsc);
3220 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3226 * Wait for safe replies on open mds requests. If we time out, drop
3227 * all requests from the tree to avoid dangling dentry refs.
3229 static void wait_requests(struct ceph_mds_client *mdsc)
3231 struct ceph_mds_request *req;
3232 struct ceph_fs_client *fsc = mdsc->fsc;
3234 mutex_lock(&mdsc->mutex);
3235 if (__get_oldest_req(mdsc)) {
3236 mutex_unlock(&mdsc->mutex);
3238 dout("wait_requests waiting for requests\n");
3239 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3240 fsc->client->options->mount_timeout * HZ);
3242 /* tear down remaining requests */
3243 mutex_lock(&mdsc->mutex);
3244 while ((req = __get_oldest_req(mdsc))) {
3245 dout("wait_requests timed out on tid %llu\n",
3247 __unregister_request(mdsc, req);
3250 mutex_unlock(&mdsc->mutex);
3251 dout("wait_requests done\n");
3255 * called before mount is ro, and before dentries are torn down.
3256 * (hmm, does this still race with new lookups?)
3258 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3260 dout("pre_umount\n");
3264 ceph_flush_dirty_caps(mdsc);
3265 wait_requests(mdsc);
3268 * wait for reply handlers to drop their request refs and
3269 * their inode/dcache refs
3275 * wait for all write mds requests to flush.
3277 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3279 struct ceph_mds_request *req = NULL, *nextreq;
3282 mutex_lock(&mdsc->mutex);
3283 dout("wait_unsafe_requests want %lld\n", want_tid);
3285 req = __get_oldest_req(mdsc);
3286 while (req && req->r_tid <= want_tid) {
3287 /* find next request */
3288 n = rb_next(&req->r_node);
3290 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3293 if ((req->r_op & CEPH_MDS_OP_WRITE)) {
3295 ceph_mdsc_get_request(req);
3297 ceph_mdsc_get_request(nextreq);
3298 mutex_unlock(&mdsc->mutex);
3299 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3300 req->r_tid, want_tid);
3301 wait_for_completion(&req->r_safe_completion);
3302 mutex_lock(&mdsc->mutex);
3303 ceph_mdsc_put_request(req);
3305 break; /* next dne before, so we're done! */
3306 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3307 /* next request was removed from tree */
3308 ceph_mdsc_put_request(nextreq);
3311 ceph_mdsc_put_request(nextreq); /* won't go away */
3315 mutex_unlock(&mdsc->mutex);
3316 dout("wait_unsafe_requests done\n");
3319 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3321 u64 want_tid, want_flush;
3323 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3327 mutex_lock(&mdsc->mutex);
3328 want_tid = mdsc->last_tid;
3329 want_flush = mdsc->cap_flush_seq;
3330 mutex_unlock(&mdsc->mutex);
3331 dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush);
3333 ceph_flush_dirty_caps(mdsc);
3335 wait_unsafe_requests(mdsc, want_tid);
3336 wait_event(mdsc->cap_flushing_wq, check_cap_flush(mdsc, want_flush));
3340 * true if all sessions are closed, or we force unmount
3342 static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3346 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3349 mutex_lock(&mdsc->mutex);
3350 for (i = 0; i < mdsc->max_sessions; i++)
3351 if (mdsc->sessions[i])
3353 mutex_unlock(&mdsc->mutex);
3358 * called after sb is ro.
3360 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3362 struct ceph_mds_session *session;
3364 struct ceph_fs_client *fsc = mdsc->fsc;
3365 unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3367 dout("close_sessions\n");
3369 /* close sessions */
3370 mutex_lock(&mdsc->mutex);
3371 for (i = 0; i < mdsc->max_sessions; i++) {
3372 session = __ceph_lookup_mds_session(mdsc, i);
3375 mutex_unlock(&mdsc->mutex);
3376 mutex_lock(&session->s_mutex);
3377 __close_session(mdsc, session);
3378 mutex_unlock(&session->s_mutex);
3379 ceph_put_mds_session(session);
3380 mutex_lock(&mdsc->mutex);
3382 mutex_unlock(&mdsc->mutex);
3384 dout("waiting for sessions to close\n");
3385 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3388 /* tear down remaining sessions */
3389 mutex_lock(&mdsc->mutex);
3390 for (i = 0; i < mdsc->max_sessions; i++) {
3391 if (mdsc->sessions[i]) {
3392 session = get_session(mdsc->sessions[i]);
3393 __unregister_session(mdsc, session);
3394 mutex_unlock(&mdsc->mutex);
3395 mutex_lock(&session->s_mutex);
3396 remove_session_caps(session);
3397 mutex_unlock(&session->s_mutex);
3398 ceph_put_mds_session(session);
3399 mutex_lock(&mdsc->mutex);
3402 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3403 mutex_unlock(&mdsc->mutex);
3405 ceph_cleanup_empty_realms(mdsc);
3407 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3412 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3415 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3417 ceph_mdsmap_destroy(mdsc->mdsmap);
3418 kfree(mdsc->sessions);
3419 ceph_caps_finalize(mdsc);
3422 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3424 struct ceph_mds_client *mdsc = fsc->mdsc;
3426 dout("mdsc_destroy %p\n", mdsc);
3427 ceph_mdsc_stop(mdsc);
3429 /* flush out any connection work with references to us */
3434 dout("mdsc_destroy %p done\n", mdsc);
3439 * handle mds map update.
3441 void ceph_mdsc_handle_map(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3445 void *p = msg->front.iov_base;
3446 void *end = p + msg->front.iov_len;
3447 struct ceph_mdsmap *newmap, *oldmap;
3448 struct ceph_fsid fsid;
3451 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3452 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3453 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3455 epoch = ceph_decode_32(&p);
3456 maplen = ceph_decode_32(&p);
3457 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3459 /* do we need it? */
3460 ceph_monc_got_mdsmap(&mdsc->fsc->client->monc, epoch);
3461 mutex_lock(&mdsc->mutex);
3462 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3463 dout("handle_map epoch %u <= our %u\n",
3464 epoch, mdsc->mdsmap->m_epoch);
3465 mutex_unlock(&mdsc->mutex);
3469 newmap = ceph_mdsmap_decode(&p, end);
3470 if (IS_ERR(newmap)) {
3471 err = PTR_ERR(newmap);
3475 /* swap into place */
3477 oldmap = mdsc->mdsmap;
3478 mdsc->mdsmap = newmap;
3479 check_new_map(mdsc, newmap, oldmap);
3480 ceph_mdsmap_destroy(oldmap);
3482 mdsc->mdsmap = newmap; /* first mds map */
3484 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3486 __wake_requests(mdsc, &mdsc->waiting_for_map);
3488 mutex_unlock(&mdsc->mutex);
3489 schedule_delayed(mdsc);
3493 mutex_unlock(&mdsc->mutex);
3495 pr_err("error decoding mdsmap %d\n", err);
3499 static struct ceph_connection *con_get(struct ceph_connection *con)
3501 struct ceph_mds_session *s = con->private;
3503 if (get_session(s)) {
3504 dout("mdsc con_get %p ok (%d)\n", s, atomic_read(&s->s_ref));
3507 dout("mdsc con_get %p FAIL\n", s);
3511 static void con_put(struct ceph_connection *con)
3513 struct ceph_mds_session *s = con->private;
3515 dout("mdsc con_put %p (%d)\n", s, atomic_read(&s->s_ref) - 1);
3516 ceph_put_mds_session(s);
3520 * if the client is unresponsive for long enough, the mds will kill
3521 * the session entirely.
3523 static void peer_reset(struct ceph_connection *con)
3525 struct ceph_mds_session *s = con->private;
3526 struct ceph_mds_client *mdsc = s->s_mdsc;
3528 pr_warning("mds%d closed our session\n", s->s_mds);
3529 send_mds_reconnect(mdsc, s);
3532 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3534 struct ceph_mds_session *s = con->private;
3535 struct ceph_mds_client *mdsc = s->s_mdsc;
3536 int type = le16_to_cpu(msg->hdr.type);
3538 mutex_lock(&mdsc->mutex);
3539 if (__verify_registered_session(mdsc, s) < 0) {
3540 mutex_unlock(&mdsc->mutex);
3543 mutex_unlock(&mdsc->mutex);
3546 case CEPH_MSG_MDS_MAP:
3547 ceph_mdsc_handle_map(mdsc, msg);
3549 case CEPH_MSG_CLIENT_SESSION:
3550 handle_session(s, msg);
3552 case CEPH_MSG_CLIENT_REPLY:
3553 handle_reply(s, msg);
3555 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3556 handle_forward(mdsc, s, msg);
3558 case CEPH_MSG_CLIENT_CAPS:
3559 ceph_handle_caps(s, msg);
3561 case CEPH_MSG_CLIENT_SNAP:
3562 ceph_handle_snap(mdsc, s, msg);
3564 case CEPH_MSG_CLIENT_LEASE:
3565 handle_lease(mdsc, s, msg);
3569 pr_err("received unknown message type %d %s\n", type,
3570 ceph_msg_type_name(type));
3581 * Note: returned pointer is the address of a structure that's
3582 * managed separately. Caller must *not* attempt to free it.
3584 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
3585 int *proto, int force_new)
3587 struct ceph_mds_session *s = con->private;
3588 struct ceph_mds_client *mdsc = s->s_mdsc;
3589 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3590 struct ceph_auth_handshake *auth = &s->s_auth;
3592 if (force_new && auth->authorizer) {
3593 ceph_auth_destroy_authorizer(ac, auth->authorizer);
3594 auth->authorizer = NULL;
3596 if (!auth->authorizer) {
3597 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3600 return ERR_PTR(ret);
3602 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
3605 return ERR_PTR(ret);
3607 *proto = ac->protocol;
3613 static int verify_authorizer_reply(struct ceph_connection *con, int len)
3615 struct ceph_mds_session *s = con->private;
3616 struct ceph_mds_client *mdsc = s->s_mdsc;
3617 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3619 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer, len);
3622 static int invalidate_authorizer(struct ceph_connection *con)
3624 struct ceph_mds_session *s = con->private;
3625 struct ceph_mds_client *mdsc = s->s_mdsc;
3626 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
3628 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
3630 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
3633 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
3634 struct ceph_msg_header *hdr, int *skip)
3636 struct ceph_msg *msg;
3637 int type = (int) le16_to_cpu(hdr->type);
3638 int front_len = (int) le32_to_cpu(hdr->front_len);
3644 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
3646 pr_err("unable to allocate msg type %d len %d\n",
3654 static const struct ceph_connection_operations mds_con_ops = {
3657 .dispatch = dispatch,
3658 .get_authorizer = get_authorizer,
3659 .verify_authorizer_reply = verify_authorizer_reply,
3660 .invalidate_authorizer = invalidate_authorizer,
3661 .peer_reset = peer_reset,
3662 .alloc_msg = mds_alloc_msg,