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>
10 #include <linux/ratelimit.h>
13 #include "mds_client.h"
15 #include <linux/ceph/ceph_features.h>
16 #include <linux/ceph/messenger.h>
17 #include <linux/ceph/decode.h>
18 #include <linux/ceph/pagelist.h>
19 #include <linux/ceph/auth.h>
20 #include <linux/ceph/debugfs.h>
23 * A cluster of MDS (metadata server) daemons is responsible for
24 * managing the file system namespace (the directory hierarchy and
25 * inodes) and for coordinating shared access to storage. Metadata is
26 * partitioning hierarchically across a number of servers, and that
27 * partition varies over time as the cluster adjusts the distribution
28 * in order to balance load.
30 * The MDS client is primarily responsible to managing synchronous
31 * metadata requests for operations like open, unlink, and so forth.
32 * If there is a MDS failure, we find out about it when we (possibly
33 * request and) receive a new MDS map, and can resubmit affected
36 * For the most part, though, we take advantage of a lossless
37 * communications channel to the MDS, and do not need to worry about
38 * timing out or resubmitting requests.
40 * We maintain a stateful "session" with each MDS we interact with.
41 * Within each session, we sent periodic heartbeat messages to ensure
42 * any capabilities or leases we have been issues remain valid. If
43 * the session times out and goes stale, our leases and capabilities
44 * are no longer valid.
47 struct ceph_reconnect_state {
49 struct ceph_pagelist *pagelist;
53 static void __wake_requests(struct ceph_mds_client *mdsc,
54 struct list_head *head);
56 static const struct ceph_connection_operations mds_con_ops;
64 * parse individual inode info
66 static int parse_reply_info_in(void **p, void *end,
67 struct ceph_mds_reply_info_in *info,
73 *p += sizeof(struct ceph_mds_reply_inode) +
74 sizeof(*info->in->fragtree.splits) *
75 le32_to_cpu(info->in->fragtree.nsplits);
77 ceph_decode_32_safe(p, end, info->symlink_len, bad);
78 ceph_decode_need(p, end, info->symlink_len, bad);
80 *p += info->symlink_len;
82 if (features & CEPH_FEATURE_DIRLAYOUTHASH)
83 ceph_decode_copy_safe(p, end, &info->dir_layout,
84 sizeof(info->dir_layout), bad);
86 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
88 ceph_decode_32_safe(p, end, info->xattr_len, bad);
89 ceph_decode_need(p, end, info->xattr_len, bad);
90 info->xattr_data = *p;
91 *p += info->xattr_len;
93 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
94 ceph_decode_64_safe(p, end, info->inline_version, bad);
95 ceph_decode_32_safe(p, end, info->inline_len, bad);
96 ceph_decode_need(p, end, info->inline_len, bad);
97 info->inline_data = *p;
98 *p += info->inline_len;
100 info->inline_version = CEPH_INLINE_NONE;
102 info->pool_ns_len = 0;
103 info->pool_ns_data = NULL;
104 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
105 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
106 if (info->pool_ns_len > 0) {
107 ceph_decode_need(p, end, info->pool_ns_len, bad);
108 info->pool_ns_data = *p;
109 *p += info->pool_ns_len;
119 * parse a normal reply, which may contain a (dir+)dentry and/or a
122 static int parse_reply_info_trace(void **p, void *end,
123 struct ceph_mds_reply_info_parsed *info,
128 if (info->head->is_dentry) {
129 err = parse_reply_info_in(p, end, &info->diri, features);
133 if (unlikely(*p + sizeof(*info->dirfrag) > end))
136 *p += sizeof(*info->dirfrag) +
137 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
138 if (unlikely(*p > end))
141 ceph_decode_32_safe(p, end, info->dname_len, bad);
142 ceph_decode_need(p, end, info->dname_len, bad);
144 *p += info->dname_len;
146 *p += sizeof(*info->dlease);
149 if (info->head->is_target) {
150 err = parse_reply_info_in(p, end, &info->targeti, features);
155 if (unlikely(*p != end))
162 pr_err("problem parsing mds trace %d\n", err);
167 * parse readdir results
169 static int parse_reply_info_dir(void **p, void *end,
170 struct ceph_mds_reply_info_parsed *info,
177 if (*p + sizeof(*info->dir_dir) > end)
179 *p += sizeof(*info->dir_dir) +
180 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
184 ceph_decode_need(p, end, sizeof(num) + 2, bad);
185 num = ceph_decode_32(p);
187 u16 flags = ceph_decode_16(p);
188 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
189 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
190 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
191 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
196 BUG_ON(!info->dir_entries);
197 if ((unsigned long)(info->dir_entries + num) >
198 (unsigned long)info->dir_entries + info->dir_buf_size) {
199 pr_err("dir contents are larger than expected\n");
206 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
208 ceph_decode_need(p, end, sizeof(u32)*2, bad);
209 rde->name_len = ceph_decode_32(p);
210 ceph_decode_need(p, end, rde->name_len, bad);
213 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
215 *p += sizeof(struct ceph_mds_reply_lease);
218 err = parse_reply_info_in(p, end, &rde->inode, features);
221 /* ceph_readdir_prepopulate() will update it */
235 pr_err("problem parsing dir contents %d\n", err);
240 * parse fcntl F_GETLK results
242 static int parse_reply_info_filelock(void **p, void *end,
243 struct ceph_mds_reply_info_parsed *info,
246 if (*p + sizeof(*info->filelock_reply) > end)
249 info->filelock_reply = *p;
250 *p += sizeof(*info->filelock_reply);
252 if (unlikely(*p != end))
261 * parse create results
263 static int parse_reply_info_create(void **p, void *end,
264 struct ceph_mds_reply_info_parsed *info,
267 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
269 info->has_create_ino = false;
271 info->has_create_ino = true;
272 info->ino = ceph_decode_64(p);
276 if (unlikely(*p != end))
285 * parse extra results
287 static int parse_reply_info_extra(void **p, void *end,
288 struct ceph_mds_reply_info_parsed *info,
291 u32 op = le32_to_cpu(info->head->op);
293 if (op == CEPH_MDS_OP_GETFILELOCK)
294 return parse_reply_info_filelock(p, end, info, features);
295 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
296 return parse_reply_info_dir(p, end, info, features);
297 else if (op == CEPH_MDS_OP_CREATE)
298 return parse_reply_info_create(p, end, info, features);
304 * parse entire mds reply
306 static int parse_reply_info(struct ceph_msg *msg,
307 struct ceph_mds_reply_info_parsed *info,
314 info->head = msg->front.iov_base;
315 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
316 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
319 ceph_decode_32_safe(&p, end, len, bad);
321 ceph_decode_need(&p, end, len, bad);
322 err = parse_reply_info_trace(&p, p+len, info, features);
328 ceph_decode_32_safe(&p, end, len, bad);
330 ceph_decode_need(&p, end, len, bad);
331 err = parse_reply_info_extra(&p, p+len, info, features);
337 ceph_decode_32_safe(&p, end, len, bad);
338 info->snapblob_len = len;
349 pr_err("mds parse_reply err %d\n", err);
353 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
355 if (!info->dir_entries)
357 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
364 const char *ceph_session_state_name(int s)
367 case CEPH_MDS_SESSION_NEW: return "new";
368 case CEPH_MDS_SESSION_OPENING: return "opening";
369 case CEPH_MDS_SESSION_OPEN: return "open";
370 case CEPH_MDS_SESSION_HUNG: return "hung";
371 case CEPH_MDS_SESSION_CLOSING: return "closing";
372 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
373 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
374 case CEPH_MDS_SESSION_REJECTED: return "rejected";
375 default: return "???";
379 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
381 if (refcount_inc_not_zero(&s->s_ref)) {
382 dout("mdsc get_session %p %d -> %d\n", s,
383 refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
386 dout("mdsc get_session %p 0 -- FAIL", s);
391 void ceph_put_mds_session(struct ceph_mds_session *s)
393 dout("mdsc put_session %p %d -> %d\n", s,
394 refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
395 if (refcount_dec_and_test(&s->s_ref)) {
396 if (s->s_auth.authorizer)
397 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
403 * called under mdsc->mutex
405 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
408 struct ceph_mds_session *session;
410 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
412 session = mdsc->sessions[mds];
413 dout("lookup_mds_session %p %d\n", session,
414 refcount_read(&session->s_ref));
415 get_session(session);
419 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
421 if (mds >= mdsc->max_sessions)
423 return mdsc->sessions[mds];
426 static int __verify_registered_session(struct ceph_mds_client *mdsc,
427 struct ceph_mds_session *s)
429 if (s->s_mds >= mdsc->max_sessions ||
430 mdsc->sessions[s->s_mds] != s)
436 * create+register a new session for given mds.
437 * called under mdsc->mutex.
439 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
442 struct ceph_mds_session *s;
444 if (mds >= mdsc->mdsmap->m_num_mds)
445 return ERR_PTR(-EINVAL);
447 s = kzalloc(sizeof(*s), GFP_NOFS);
449 return ERR_PTR(-ENOMEM);
452 s->s_state = CEPH_MDS_SESSION_NEW;
455 mutex_init(&s->s_mutex);
457 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
459 spin_lock_init(&s->s_gen_ttl_lock);
461 s->s_cap_ttl = jiffies - 1;
463 spin_lock_init(&s->s_cap_lock);
464 s->s_renew_requested = 0;
466 INIT_LIST_HEAD(&s->s_caps);
469 refcount_set(&s->s_ref, 1);
470 INIT_LIST_HEAD(&s->s_waiting);
471 INIT_LIST_HEAD(&s->s_unsafe);
472 s->s_num_cap_releases = 0;
473 s->s_cap_reconnect = 0;
474 s->s_cap_iterator = NULL;
475 INIT_LIST_HEAD(&s->s_cap_releases);
476 INIT_LIST_HEAD(&s->s_cap_flushing);
478 dout("register_session mds%d\n", mds);
479 if (mds >= mdsc->max_sessions) {
480 int newmax = 1 << get_count_order(mds+1);
481 struct ceph_mds_session **sa;
483 dout("register_session realloc to %d\n", newmax);
484 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
487 if (mdsc->sessions) {
488 memcpy(sa, mdsc->sessions,
489 mdsc->max_sessions * sizeof(void *));
490 kfree(mdsc->sessions);
493 mdsc->max_sessions = newmax;
495 mdsc->sessions[mds] = s;
496 atomic_inc(&mdsc->num_sessions);
497 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
499 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
500 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
506 return ERR_PTR(-ENOMEM);
510 * called under mdsc->mutex
512 static void __unregister_session(struct ceph_mds_client *mdsc,
513 struct ceph_mds_session *s)
515 dout("__unregister_session mds%d %p\n", s->s_mds, s);
516 BUG_ON(mdsc->sessions[s->s_mds] != s);
517 mdsc->sessions[s->s_mds] = NULL;
518 ceph_con_close(&s->s_con);
519 ceph_put_mds_session(s);
520 atomic_dec(&mdsc->num_sessions);
524 * drop session refs in request.
526 * should be last request ref, or hold mdsc->mutex
528 static void put_request_session(struct ceph_mds_request *req)
530 if (req->r_session) {
531 ceph_put_mds_session(req->r_session);
532 req->r_session = NULL;
536 void ceph_mdsc_release_request(struct kref *kref)
538 struct ceph_mds_request *req = container_of(kref,
539 struct ceph_mds_request,
541 destroy_reply_info(&req->r_reply_info);
543 ceph_msg_put(req->r_request);
545 ceph_msg_put(req->r_reply);
547 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
551 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
552 iput(req->r_target_inode);
555 if (req->r_old_dentry)
556 dput(req->r_old_dentry);
557 if (req->r_old_dentry_dir) {
559 * track (and drop pins for) r_old_dentry_dir
560 * separately, since r_old_dentry's d_parent may have
561 * changed between the dir mutex being dropped and
562 * this request being freed.
564 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
566 iput(req->r_old_dentry_dir);
571 ceph_pagelist_release(req->r_pagelist);
572 put_request_session(req);
573 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
577 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
580 * lookup session, bump ref if found.
582 * called under mdsc->mutex.
584 static struct ceph_mds_request *
585 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
587 struct ceph_mds_request *req;
589 req = lookup_request(&mdsc->request_tree, tid);
591 ceph_mdsc_get_request(req);
597 * Register an in-flight request, and assign a tid. Link to directory
598 * are modifying (if any).
600 * Called under mdsc->mutex.
602 static void __register_request(struct ceph_mds_client *mdsc,
603 struct ceph_mds_request *req,
606 req->r_tid = ++mdsc->last_tid;
608 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
610 dout("__register_request %p tid %lld\n", req, req->r_tid);
611 ceph_mdsc_get_request(req);
612 insert_request(&mdsc->request_tree, req);
614 req->r_uid = current_fsuid();
615 req->r_gid = current_fsgid();
617 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
618 mdsc->oldest_tid = req->r_tid;
622 req->r_unsafe_dir = dir;
626 static void __unregister_request(struct ceph_mds_client *mdsc,
627 struct ceph_mds_request *req)
629 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
631 /* Never leave an unregistered request on an unsafe list! */
632 list_del_init(&req->r_unsafe_item);
634 if (req->r_tid == mdsc->oldest_tid) {
635 struct rb_node *p = rb_next(&req->r_node);
636 mdsc->oldest_tid = 0;
638 struct ceph_mds_request *next_req =
639 rb_entry(p, struct ceph_mds_request, r_node);
640 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
641 mdsc->oldest_tid = next_req->r_tid;
648 erase_request(&mdsc->request_tree, req);
650 if (req->r_unsafe_dir &&
651 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
652 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
653 spin_lock(&ci->i_unsafe_lock);
654 list_del_init(&req->r_unsafe_dir_item);
655 spin_unlock(&ci->i_unsafe_lock);
657 if (req->r_target_inode &&
658 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
659 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
660 spin_lock(&ci->i_unsafe_lock);
661 list_del_init(&req->r_unsafe_target_item);
662 spin_unlock(&ci->i_unsafe_lock);
665 if (req->r_unsafe_dir) {
666 iput(req->r_unsafe_dir);
667 req->r_unsafe_dir = NULL;
670 complete_all(&req->r_safe_completion);
672 ceph_mdsc_put_request(req);
676 * Walk back up the dentry tree until we hit a dentry representing a
677 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
678 * when calling this) to ensure that the objects won't disappear while we're
679 * working with them. Once we hit a candidate dentry, we attempt to take a
680 * reference to it, and return that as the result.
682 static struct inode *get_nonsnap_parent(struct dentry *dentry)
684 struct inode *inode = NULL;
686 while (dentry && !IS_ROOT(dentry)) {
687 inode = d_inode_rcu(dentry);
688 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
690 dentry = dentry->d_parent;
693 inode = igrab(inode);
698 * Choose mds to send request to next. If there is a hint set in the
699 * request (e.g., due to a prior forward hint from the mds), use that.
700 * Otherwise, consult frag tree and/or caps to identify the
701 * appropriate mds. If all else fails, choose randomly.
703 * Called under mdsc->mutex.
705 static int __choose_mds(struct ceph_mds_client *mdsc,
706 struct ceph_mds_request *req)
709 struct ceph_inode_info *ci;
710 struct ceph_cap *cap;
711 int mode = req->r_direct_mode;
713 u32 hash = req->r_direct_hash;
714 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
717 * is there a specific mds we should try? ignore hint if we have
718 * no session and the mds is not up (active or recovering).
720 if (req->r_resend_mds >= 0 &&
721 (__have_session(mdsc, req->r_resend_mds) ||
722 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
723 dout("choose_mds using resend_mds mds%d\n",
725 return req->r_resend_mds;
728 if (mode == USE_RANDOM_MDS)
733 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
734 inode = req->r_inode;
737 /* req->r_dentry is non-null for LSSNAP request.
739 WARN_ON_ONCE(!req->r_dentry);
742 if (!inode && req->r_dentry) {
743 /* ignore race with rename; old or new d_parent is okay */
744 struct dentry *parent;
748 parent = req->r_dentry->d_parent;
749 dir = req->r_parent ? : d_inode_rcu(parent);
751 if (!dir || dir->i_sb != mdsc->fsc->sb) {
752 /* not this fs or parent went negative */
753 inode = d_inode(req->r_dentry);
756 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
757 /* direct snapped/virtual snapdir requests
758 * based on parent dir inode */
759 inode = get_nonsnap_parent(parent);
760 dout("__choose_mds using nonsnap parent %p\n", inode);
763 inode = d_inode(req->r_dentry);
764 if (!inode || mode == USE_AUTH_MDS) {
767 hash = ceph_dentry_hash(dir, req->r_dentry);
776 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
780 ci = ceph_inode(inode);
782 if (is_hash && S_ISDIR(inode->i_mode)) {
783 struct ceph_inode_frag frag;
786 ceph_choose_frag(ci, hash, &frag, &found);
788 if (mode == USE_ANY_MDS && frag.ndist > 0) {
791 /* choose a random replica */
792 get_random_bytes(&r, 1);
795 dout("choose_mds %p %llx.%llx "
796 "frag %u mds%d (%d/%d)\n",
797 inode, ceph_vinop(inode),
800 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
801 CEPH_MDS_STATE_ACTIVE)
805 /* since this file/dir wasn't known to be
806 * replicated, then we want to look for the
807 * authoritative mds. */
810 /* choose auth mds */
812 dout("choose_mds %p %llx.%llx "
813 "frag %u mds%d (auth)\n",
814 inode, ceph_vinop(inode), frag.frag, mds);
815 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
816 CEPH_MDS_STATE_ACTIVE)
822 spin_lock(&ci->i_ceph_lock);
824 if (mode == USE_AUTH_MDS)
825 cap = ci->i_auth_cap;
826 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
827 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
829 spin_unlock(&ci->i_ceph_lock);
833 mds = cap->session->s_mds;
834 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
835 inode, ceph_vinop(inode), mds,
836 cap == ci->i_auth_cap ? "auth " : "", cap);
837 spin_unlock(&ci->i_ceph_lock);
843 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
844 dout("choose_mds chose random mds%d\n", mds);
852 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
854 struct ceph_msg *msg;
855 struct ceph_mds_session_head *h;
857 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
860 pr_err("create_session_msg ENOMEM creating msg\n");
863 h = msg->front.iov_base;
864 h->op = cpu_to_le32(op);
865 h->seq = cpu_to_le64(seq);
871 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
872 * to include additional client metadata fields.
874 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
876 struct ceph_msg *msg;
877 struct ceph_mds_session_head *h;
879 int metadata_bytes = 0;
880 int metadata_key_count = 0;
881 struct ceph_options *opt = mdsc->fsc->client->options;
882 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
885 const char* metadata[][2] = {
886 {"hostname", mdsc->nodename},
887 {"kernel_version", init_utsname()->release},
888 {"entity_id", opt->name ? : ""},
889 {"root", fsopt->server_path ? : "/"},
893 /* Calculate serialized length of metadata */
894 metadata_bytes = 4; /* map length */
895 for (i = 0; metadata[i][0]; ++i) {
896 metadata_bytes += 8 + strlen(metadata[i][0]) +
897 strlen(metadata[i][1]);
898 metadata_key_count++;
901 /* Allocate the message */
902 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
905 pr_err("create_session_msg ENOMEM creating msg\n");
908 h = msg->front.iov_base;
909 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
910 h->seq = cpu_to_le64(seq);
913 * Serialize client metadata into waiting buffer space, using
914 * the format that userspace expects for map<string, string>
916 * ClientSession messages with metadata are v2
918 msg->hdr.version = cpu_to_le16(2);
919 msg->hdr.compat_version = cpu_to_le16(1);
921 /* The write pointer, following the session_head structure */
922 p = msg->front.iov_base + sizeof(*h);
924 /* Number of entries in the map */
925 ceph_encode_32(&p, metadata_key_count);
927 /* Two length-prefixed strings for each entry in the map */
928 for (i = 0; metadata[i][0]; ++i) {
929 size_t const key_len = strlen(metadata[i][0]);
930 size_t const val_len = strlen(metadata[i][1]);
932 ceph_encode_32(&p, key_len);
933 memcpy(p, metadata[i][0], key_len);
935 ceph_encode_32(&p, val_len);
936 memcpy(p, metadata[i][1], val_len);
944 * send session open request.
946 * called under mdsc->mutex
948 static int __open_session(struct ceph_mds_client *mdsc,
949 struct ceph_mds_session *session)
951 struct ceph_msg *msg;
953 int mds = session->s_mds;
955 /* wait for mds to go active? */
956 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
957 dout("open_session to mds%d (%s)\n", mds,
958 ceph_mds_state_name(mstate));
959 session->s_state = CEPH_MDS_SESSION_OPENING;
960 session->s_renew_requested = jiffies;
962 /* send connect message */
963 msg = create_session_open_msg(mdsc, session->s_seq);
966 ceph_con_send(&session->s_con, msg);
971 * open sessions for any export targets for the given mds
973 * called under mdsc->mutex
975 static struct ceph_mds_session *
976 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
978 struct ceph_mds_session *session;
980 session = __ceph_lookup_mds_session(mdsc, target);
982 session = register_session(mdsc, target);
986 if (session->s_state == CEPH_MDS_SESSION_NEW ||
987 session->s_state == CEPH_MDS_SESSION_CLOSING)
988 __open_session(mdsc, session);
993 struct ceph_mds_session *
994 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
996 struct ceph_mds_session *session;
998 dout("open_export_target_session to mds%d\n", target);
1000 mutex_lock(&mdsc->mutex);
1001 session = __open_export_target_session(mdsc, target);
1002 mutex_unlock(&mdsc->mutex);
1007 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1008 struct ceph_mds_session *session)
1010 struct ceph_mds_info *mi;
1011 struct ceph_mds_session *ts;
1012 int i, mds = session->s_mds;
1014 if (mds >= mdsc->mdsmap->m_num_mds)
1017 mi = &mdsc->mdsmap->m_info[mds];
1018 dout("open_export_target_sessions for mds%d (%d targets)\n",
1019 session->s_mds, mi->num_export_targets);
1021 for (i = 0; i < mi->num_export_targets; i++) {
1022 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1024 ceph_put_mds_session(ts);
1028 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1029 struct ceph_mds_session *session)
1031 mutex_lock(&mdsc->mutex);
1032 __open_export_target_sessions(mdsc, session);
1033 mutex_unlock(&mdsc->mutex);
1040 /* caller holds s_cap_lock, we drop it */
1041 static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1042 struct ceph_mds_session *session)
1043 __releases(session->s_cap_lock)
1045 LIST_HEAD(tmp_list);
1046 list_splice_init(&session->s_cap_releases, &tmp_list);
1047 session->s_num_cap_releases = 0;
1048 spin_unlock(&session->s_cap_lock);
1050 dout("cleanup_cap_releases mds%d\n", session->s_mds);
1051 while (!list_empty(&tmp_list)) {
1052 struct ceph_cap *cap;
1053 /* zero out the in-progress message */
1054 cap = list_first_entry(&tmp_list,
1055 struct ceph_cap, session_caps);
1056 list_del(&cap->session_caps);
1057 ceph_put_cap(mdsc, cap);
1061 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1062 struct ceph_mds_session *session)
1064 struct ceph_mds_request *req;
1067 dout("cleanup_session_requests mds%d\n", session->s_mds);
1068 mutex_lock(&mdsc->mutex);
1069 while (!list_empty(&session->s_unsafe)) {
1070 req = list_first_entry(&session->s_unsafe,
1071 struct ceph_mds_request, r_unsafe_item);
1072 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1074 __unregister_request(mdsc, req);
1076 /* zero r_attempts, so kick_requests() will re-send requests */
1077 p = rb_first(&mdsc->request_tree);
1079 req = rb_entry(p, struct ceph_mds_request, r_node);
1081 if (req->r_session &&
1082 req->r_session->s_mds == session->s_mds)
1083 req->r_attempts = 0;
1085 mutex_unlock(&mdsc->mutex);
1089 * Helper to safely iterate over all caps associated with a session, with
1090 * special care taken to handle a racing __ceph_remove_cap().
1092 * Caller must hold session s_mutex.
1094 static int iterate_session_caps(struct ceph_mds_session *session,
1095 int (*cb)(struct inode *, struct ceph_cap *,
1098 struct list_head *p;
1099 struct ceph_cap *cap;
1100 struct inode *inode, *last_inode = NULL;
1101 struct ceph_cap *old_cap = NULL;
1104 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1105 spin_lock(&session->s_cap_lock);
1106 p = session->s_caps.next;
1107 while (p != &session->s_caps) {
1108 cap = list_entry(p, struct ceph_cap, session_caps);
1109 inode = igrab(&cap->ci->vfs_inode);
1114 session->s_cap_iterator = cap;
1115 spin_unlock(&session->s_cap_lock);
1122 ceph_put_cap(session->s_mdsc, old_cap);
1126 ret = cb(inode, cap, arg);
1129 spin_lock(&session->s_cap_lock);
1132 dout("iterate_session_caps finishing cap %p removal\n",
1134 BUG_ON(cap->session != session);
1135 cap->session = NULL;
1136 list_del_init(&cap->session_caps);
1137 session->s_nr_caps--;
1138 if (cap->queue_release) {
1139 list_add_tail(&cap->session_caps,
1140 &session->s_cap_releases);
1141 session->s_num_cap_releases++;
1143 old_cap = cap; /* put_cap it w/o locks held */
1151 session->s_cap_iterator = NULL;
1152 spin_unlock(&session->s_cap_lock);
1156 ceph_put_cap(session->s_mdsc, old_cap);
1161 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1164 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1165 struct ceph_inode_info *ci = ceph_inode(inode);
1166 LIST_HEAD(to_remove);
1168 bool invalidate = false;
1170 dout("removing cap %p, ci is %p, inode is %p\n",
1171 cap, ci, &ci->vfs_inode);
1172 spin_lock(&ci->i_ceph_lock);
1173 __ceph_remove_cap(cap, false);
1174 if (!ci->i_auth_cap) {
1175 struct ceph_cap_flush *cf;
1176 struct ceph_mds_client *mdsc = fsc->mdsc;
1178 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1180 if (ci->i_wrbuffer_ref > 0 &&
1181 READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1184 while (!list_empty(&ci->i_cap_flush_list)) {
1185 cf = list_first_entry(&ci->i_cap_flush_list,
1186 struct ceph_cap_flush, i_list);
1187 list_move(&cf->i_list, &to_remove);
1190 spin_lock(&mdsc->cap_dirty_lock);
1192 list_for_each_entry(cf, &to_remove, i_list)
1193 list_del(&cf->g_list);
1195 if (!list_empty(&ci->i_dirty_item)) {
1196 pr_warn_ratelimited(
1197 " dropping dirty %s state for %p %lld\n",
1198 ceph_cap_string(ci->i_dirty_caps),
1199 inode, ceph_ino(inode));
1200 ci->i_dirty_caps = 0;
1201 list_del_init(&ci->i_dirty_item);
1204 if (!list_empty(&ci->i_flushing_item)) {
1205 pr_warn_ratelimited(
1206 " dropping dirty+flushing %s state for %p %lld\n",
1207 ceph_cap_string(ci->i_flushing_caps),
1208 inode, ceph_ino(inode));
1209 ci->i_flushing_caps = 0;
1210 list_del_init(&ci->i_flushing_item);
1211 mdsc->num_cap_flushing--;
1214 spin_unlock(&mdsc->cap_dirty_lock);
1216 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1217 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1218 ci->i_prealloc_cap_flush = NULL;
1221 spin_unlock(&ci->i_ceph_lock);
1222 while (!list_empty(&to_remove)) {
1223 struct ceph_cap_flush *cf;
1224 cf = list_first_entry(&to_remove,
1225 struct ceph_cap_flush, i_list);
1226 list_del(&cf->i_list);
1227 ceph_free_cap_flush(cf);
1230 wake_up_all(&ci->i_cap_wq);
1232 ceph_queue_invalidate(inode);
1239 * caller must hold session s_mutex
1241 static void remove_session_caps(struct ceph_mds_session *session)
1243 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1244 struct super_block *sb = fsc->sb;
1245 dout("remove_session_caps on %p\n", session);
1246 iterate_session_caps(session, remove_session_caps_cb, fsc);
1248 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1250 spin_lock(&session->s_cap_lock);
1251 if (session->s_nr_caps > 0) {
1252 struct inode *inode;
1253 struct ceph_cap *cap, *prev = NULL;
1254 struct ceph_vino vino;
1256 * iterate_session_caps() skips inodes that are being
1257 * deleted, we need to wait until deletions are complete.
1258 * __wait_on_freeing_inode() is designed for the job,
1259 * but it is not exported, so use lookup inode function
1262 while (!list_empty(&session->s_caps)) {
1263 cap = list_entry(session->s_caps.next,
1264 struct ceph_cap, session_caps);
1268 vino = cap->ci->i_vino;
1269 spin_unlock(&session->s_cap_lock);
1271 inode = ceph_find_inode(sb, vino);
1274 spin_lock(&session->s_cap_lock);
1278 // drop cap expires and unlock s_cap_lock
1279 cleanup_cap_releases(session->s_mdsc, session);
1281 BUG_ON(session->s_nr_caps > 0);
1282 BUG_ON(!list_empty(&session->s_cap_flushing));
1286 * wake up any threads waiting on this session's caps. if the cap is
1287 * old (didn't get renewed on the client reconnect), remove it now.
1289 * caller must hold s_mutex.
1291 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1294 struct ceph_inode_info *ci = ceph_inode(inode);
1297 spin_lock(&ci->i_ceph_lock);
1298 ci->i_wanted_max_size = 0;
1299 ci->i_requested_max_size = 0;
1300 spin_unlock(&ci->i_ceph_lock);
1302 wake_up_all(&ci->i_cap_wq);
1306 static void wake_up_session_caps(struct ceph_mds_session *session,
1309 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1310 iterate_session_caps(session, wake_up_session_cb,
1311 (void *)(unsigned long)reconnect);
1315 * Send periodic message to MDS renewing all currently held caps. The
1316 * ack will reset the expiration for all caps from this session.
1318 * caller holds s_mutex
1320 static int send_renew_caps(struct ceph_mds_client *mdsc,
1321 struct ceph_mds_session *session)
1323 struct ceph_msg *msg;
1326 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1327 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1328 pr_info("mds%d caps stale\n", session->s_mds);
1329 session->s_renew_requested = jiffies;
1331 /* do not try to renew caps until a recovering mds has reconnected
1332 * with its clients. */
1333 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1334 if (state < CEPH_MDS_STATE_RECONNECT) {
1335 dout("send_renew_caps ignoring mds%d (%s)\n",
1336 session->s_mds, ceph_mds_state_name(state));
1340 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1341 ceph_mds_state_name(state));
1342 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1343 ++session->s_renew_seq);
1346 ceph_con_send(&session->s_con, msg);
1350 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1351 struct ceph_mds_session *session, u64 seq)
1353 struct ceph_msg *msg;
1355 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1356 session->s_mds, ceph_session_state_name(session->s_state), seq);
1357 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1360 ceph_con_send(&session->s_con, msg);
1366 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1368 * Called under session->s_mutex
1370 static void renewed_caps(struct ceph_mds_client *mdsc,
1371 struct ceph_mds_session *session, int is_renew)
1376 spin_lock(&session->s_cap_lock);
1377 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1379 session->s_cap_ttl = session->s_renew_requested +
1380 mdsc->mdsmap->m_session_timeout*HZ;
1383 if (time_before(jiffies, session->s_cap_ttl)) {
1384 pr_info("mds%d caps renewed\n", session->s_mds);
1387 pr_info("mds%d caps still stale\n", session->s_mds);
1390 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1391 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1392 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1393 spin_unlock(&session->s_cap_lock);
1396 wake_up_session_caps(session, 0);
1400 * send a session close request
1402 static int request_close_session(struct ceph_mds_client *mdsc,
1403 struct ceph_mds_session *session)
1405 struct ceph_msg *msg;
1407 dout("request_close_session mds%d state %s seq %lld\n",
1408 session->s_mds, ceph_session_state_name(session->s_state),
1410 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1413 ceph_con_send(&session->s_con, msg);
1418 * Called with s_mutex held.
1420 static int __close_session(struct ceph_mds_client *mdsc,
1421 struct ceph_mds_session *session)
1423 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1425 session->s_state = CEPH_MDS_SESSION_CLOSING;
1426 return request_close_session(mdsc, session);
1430 * Trim old(er) caps.
1432 * Because we can't cache an inode without one or more caps, we do
1433 * this indirectly: if a cap is unused, we prune its aliases, at which
1434 * point the inode will hopefully get dropped to.
1436 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1437 * memory pressure from the MDS, though, so it needn't be perfect.
1439 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1441 struct ceph_mds_session *session = arg;
1442 struct ceph_inode_info *ci = ceph_inode(inode);
1443 int used, wanted, oissued, mine;
1445 if (session->s_trim_caps <= 0)
1448 spin_lock(&ci->i_ceph_lock);
1449 mine = cap->issued | cap->implemented;
1450 used = __ceph_caps_used(ci);
1451 wanted = __ceph_caps_file_wanted(ci);
1452 oissued = __ceph_caps_issued_other(ci, cap);
1454 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1455 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1456 ceph_cap_string(used), ceph_cap_string(wanted));
1457 if (cap == ci->i_auth_cap) {
1458 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1459 !list_empty(&ci->i_cap_snaps))
1461 if ((used | wanted) & CEPH_CAP_ANY_WR)
1464 /* The inode has cached pages, but it's no longer used.
1465 * we can safely drop it */
1466 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1467 !(oissued & CEPH_CAP_FILE_CACHE)) {
1471 if ((used | wanted) & ~oissued & mine)
1472 goto out; /* we need these caps */
1474 session->s_trim_caps--;
1476 /* we aren't the only cap.. just remove us */
1477 __ceph_remove_cap(cap, true);
1479 /* try dropping referring dentries */
1480 spin_unlock(&ci->i_ceph_lock);
1481 d_prune_aliases(inode);
1482 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1483 inode, cap, atomic_read(&inode->i_count));
1488 spin_unlock(&ci->i_ceph_lock);
1493 * Trim session cap count down to some max number.
1495 static int trim_caps(struct ceph_mds_client *mdsc,
1496 struct ceph_mds_session *session,
1499 int trim_caps = session->s_nr_caps - max_caps;
1501 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1502 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1503 if (trim_caps > 0) {
1504 session->s_trim_caps = trim_caps;
1505 iterate_session_caps(session, trim_caps_cb, session);
1506 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1507 session->s_mds, session->s_nr_caps, max_caps,
1508 trim_caps - session->s_trim_caps);
1509 session->s_trim_caps = 0;
1512 ceph_send_cap_releases(mdsc, session);
1516 static int check_caps_flush(struct ceph_mds_client *mdsc,
1521 spin_lock(&mdsc->cap_dirty_lock);
1522 if (!list_empty(&mdsc->cap_flush_list)) {
1523 struct ceph_cap_flush *cf =
1524 list_first_entry(&mdsc->cap_flush_list,
1525 struct ceph_cap_flush, g_list);
1526 if (cf->tid <= want_flush_tid) {
1527 dout("check_caps_flush still flushing tid "
1528 "%llu <= %llu\n", cf->tid, want_flush_tid);
1532 spin_unlock(&mdsc->cap_dirty_lock);
1537 * flush all dirty inode data to disk.
1539 * returns true if we've flushed through want_flush_tid
1541 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1544 dout("check_caps_flush want %llu\n", want_flush_tid);
1546 wait_event(mdsc->cap_flushing_wq,
1547 check_caps_flush(mdsc, want_flush_tid));
1549 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1553 * called under s_mutex
1555 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1556 struct ceph_mds_session *session)
1558 struct ceph_msg *msg = NULL;
1559 struct ceph_mds_cap_release *head;
1560 struct ceph_mds_cap_item *item;
1561 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1562 struct ceph_cap *cap;
1563 LIST_HEAD(tmp_list);
1564 int num_cap_releases;
1565 __le32 barrier, *cap_barrier;
1567 down_read(&osdc->lock);
1568 barrier = cpu_to_le32(osdc->epoch_barrier);
1569 up_read(&osdc->lock);
1571 spin_lock(&session->s_cap_lock);
1573 list_splice_init(&session->s_cap_releases, &tmp_list);
1574 num_cap_releases = session->s_num_cap_releases;
1575 session->s_num_cap_releases = 0;
1576 spin_unlock(&session->s_cap_lock);
1578 while (!list_empty(&tmp_list)) {
1580 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1581 PAGE_SIZE, GFP_NOFS, false);
1584 head = msg->front.iov_base;
1585 head->num = cpu_to_le32(0);
1586 msg->front.iov_len = sizeof(*head);
1588 msg->hdr.version = cpu_to_le16(2);
1589 msg->hdr.compat_version = cpu_to_le16(1);
1592 cap = list_first_entry(&tmp_list, struct ceph_cap,
1594 list_del(&cap->session_caps);
1597 head = msg->front.iov_base;
1598 le32_add_cpu(&head->num, 1);
1599 item = msg->front.iov_base + msg->front.iov_len;
1600 item->ino = cpu_to_le64(cap->cap_ino);
1601 item->cap_id = cpu_to_le64(cap->cap_id);
1602 item->migrate_seq = cpu_to_le32(cap->mseq);
1603 item->seq = cpu_to_le32(cap->issue_seq);
1604 msg->front.iov_len += sizeof(*item);
1606 ceph_put_cap(mdsc, cap);
1608 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1609 // Append cap_barrier field
1610 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1611 *cap_barrier = barrier;
1612 msg->front.iov_len += sizeof(*cap_barrier);
1614 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1615 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1616 ceph_con_send(&session->s_con, msg);
1621 BUG_ON(num_cap_releases != 0);
1623 spin_lock(&session->s_cap_lock);
1624 if (!list_empty(&session->s_cap_releases))
1626 spin_unlock(&session->s_cap_lock);
1629 // Append cap_barrier field
1630 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1631 *cap_barrier = barrier;
1632 msg->front.iov_len += sizeof(*cap_barrier);
1634 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1635 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1636 ceph_con_send(&session->s_con, msg);
1640 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1642 spin_lock(&session->s_cap_lock);
1643 list_splice(&tmp_list, &session->s_cap_releases);
1644 session->s_num_cap_releases += num_cap_releases;
1645 spin_unlock(&session->s_cap_lock);
1652 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1655 struct ceph_inode_info *ci = ceph_inode(dir);
1656 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1657 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1658 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1659 int order, num_entries;
1661 spin_lock(&ci->i_ceph_lock);
1662 num_entries = ci->i_files + ci->i_subdirs;
1663 spin_unlock(&ci->i_ceph_lock);
1664 num_entries = max(num_entries, 1);
1665 num_entries = min(num_entries, opt->max_readdir);
1667 order = get_order(size * num_entries);
1668 while (order >= 0) {
1669 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1672 if (rinfo->dir_entries)
1676 if (!rinfo->dir_entries)
1679 num_entries = (PAGE_SIZE << order) / size;
1680 num_entries = min(num_entries, opt->max_readdir);
1682 rinfo->dir_buf_size = PAGE_SIZE << order;
1683 req->r_num_caps = num_entries + 1;
1684 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1685 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1690 * Create an mds request.
1692 struct ceph_mds_request *
1693 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1695 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1698 return ERR_PTR(-ENOMEM);
1700 mutex_init(&req->r_fill_mutex);
1702 req->r_started = jiffies;
1703 req->r_resend_mds = -1;
1704 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1705 INIT_LIST_HEAD(&req->r_unsafe_target_item);
1707 kref_init(&req->r_kref);
1708 RB_CLEAR_NODE(&req->r_node);
1709 INIT_LIST_HEAD(&req->r_wait);
1710 init_completion(&req->r_completion);
1711 init_completion(&req->r_safe_completion);
1712 INIT_LIST_HEAD(&req->r_unsafe_item);
1714 req->r_stamp = timespec_trunc(current_kernel_time(), mdsc->fsc->sb->s_time_gran);
1717 req->r_direct_mode = mode;
1722 * return oldest (lowest) request, tid in request tree, 0 if none.
1724 * called under mdsc->mutex.
1726 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1728 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1730 return rb_entry(rb_first(&mdsc->request_tree),
1731 struct ceph_mds_request, r_node);
1734 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1736 return mdsc->oldest_tid;
1740 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1741 * on build_path_from_dentry in fs/cifs/dir.c.
1743 * If @stop_on_nosnap, generate path relative to the first non-snapped
1746 * Encode hidden .snap dirs as a double /, i.e.
1747 * foo/.snap/bar -> foo//bar
1749 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1752 struct dentry *temp;
1758 return ERR_PTR(-EINVAL);
1762 seq = read_seqbegin(&rename_lock);
1764 for (temp = dentry; !IS_ROOT(temp);) {
1765 struct inode *inode = d_inode(temp);
1766 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1767 len++; /* slash only */
1768 else if (stop_on_nosnap && inode &&
1769 ceph_snap(inode) == CEPH_NOSNAP)
1772 len += 1 + temp->d_name.len;
1773 temp = temp->d_parent;
1777 len--; /* no leading '/' */
1779 path = kmalloc(len+1, GFP_NOFS);
1781 return ERR_PTR(-ENOMEM);
1783 path[pos] = 0; /* trailing null */
1785 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1786 struct inode *inode;
1788 spin_lock(&temp->d_lock);
1789 inode = d_inode(temp);
1790 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1791 dout("build_path path+%d: %p SNAPDIR\n",
1793 } else if (stop_on_nosnap && inode &&
1794 ceph_snap(inode) == CEPH_NOSNAP) {
1795 spin_unlock(&temp->d_lock);
1798 pos -= temp->d_name.len;
1800 spin_unlock(&temp->d_lock);
1803 strncpy(path + pos, temp->d_name.name,
1806 spin_unlock(&temp->d_lock);
1809 temp = temp->d_parent;
1812 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1813 pr_err("build_path did not end path lookup where "
1814 "expected, namelen is %d, pos is %d\n", len, pos);
1815 /* presumably this is only possible if racing with a
1816 rename of one of the parent directories (we can not
1817 lock the dentries above us to prevent this, but
1818 retrying should be harmless) */
1823 *base = ceph_ino(d_inode(temp));
1825 dout("build_path on %p %d built %llx '%.*s'\n",
1826 dentry, d_count(dentry), *base, len, path);
1830 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1831 const char **ppath, int *ppathlen, u64 *pino,
1838 dir = d_inode_rcu(dentry->d_parent);
1839 if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1840 *pino = ceph_ino(dir);
1842 *ppath = dentry->d_name.name;
1843 *ppathlen = dentry->d_name.len;
1847 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1849 return PTR_ERR(path);
1855 static int build_inode_path(struct inode *inode,
1856 const char **ppath, int *ppathlen, u64 *pino,
1859 struct dentry *dentry;
1862 if (ceph_snap(inode) == CEPH_NOSNAP) {
1863 *pino = ceph_ino(inode);
1867 dentry = d_find_alias(inode);
1868 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1871 return PTR_ERR(path);
1878 * request arguments may be specified via an inode *, a dentry *, or
1879 * an explicit ino+path.
1881 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1882 struct inode *rdiri, const char *rpath,
1883 u64 rino, const char **ppath, int *pathlen,
1884 u64 *ino, int *freepath)
1889 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1890 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1892 } else if (rdentry) {
1893 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
1895 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1897 } else if (rpath || rino) {
1900 *pathlen = rpath ? strlen(rpath) : 0;
1901 dout(" path %.*s\n", *pathlen, rpath);
1908 * called under mdsc->mutex
1910 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1911 struct ceph_mds_request *req,
1912 int mds, bool drop_cap_releases)
1914 struct ceph_msg *msg;
1915 struct ceph_mds_request_head *head;
1916 const char *path1 = NULL;
1917 const char *path2 = NULL;
1918 u64 ino1 = 0, ino2 = 0;
1919 int pathlen1 = 0, pathlen2 = 0;
1920 int freepath1 = 0, freepath2 = 0;
1926 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1927 req->r_parent, req->r_path1, req->r_ino1.ino,
1928 &path1, &pathlen1, &ino1, &freepath1);
1934 ret = set_request_path_attr(NULL, req->r_old_dentry,
1935 req->r_old_dentry_dir,
1936 req->r_path2, req->r_ino2.ino,
1937 &path2, &pathlen2, &ino2, &freepath2);
1943 len = sizeof(*head) +
1944 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1945 sizeof(struct ceph_timespec);
1947 /* calculate (max) length for cap releases */
1948 len += sizeof(struct ceph_mds_request_release) *
1949 (!!req->r_inode_drop + !!req->r_dentry_drop +
1950 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1951 if (req->r_dentry_drop)
1952 len += req->r_dentry->d_name.len;
1953 if (req->r_old_dentry_drop)
1954 len += req->r_old_dentry->d_name.len;
1956 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1958 msg = ERR_PTR(-ENOMEM);
1962 msg->hdr.version = cpu_to_le16(2);
1963 msg->hdr.tid = cpu_to_le64(req->r_tid);
1965 head = msg->front.iov_base;
1966 p = msg->front.iov_base + sizeof(*head);
1967 end = msg->front.iov_base + msg->front.iov_len;
1969 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1970 head->op = cpu_to_le32(req->r_op);
1971 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1972 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1973 head->args = req->r_args;
1975 ceph_encode_filepath(&p, end, ino1, path1);
1976 ceph_encode_filepath(&p, end, ino2, path2);
1978 /* make note of release offset, in case we need to replay */
1979 req->r_request_release_offset = p - msg->front.iov_base;
1983 if (req->r_inode_drop)
1984 releases += ceph_encode_inode_release(&p,
1985 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
1986 mds, req->r_inode_drop, req->r_inode_unless, 0);
1987 if (req->r_dentry_drop)
1988 releases += ceph_encode_dentry_release(&p, req->r_dentry,
1989 req->r_parent, mds, req->r_dentry_drop,
1990 req->r_dentry_unless);
1991 if (req->r_old_dentry_drop)
1992 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
1993 req->r_old_dentry_dir, mds,
1994 req->r_old_dentry_drop,
1995 req->r_old_dentry_unless);
1996 if (req->r_old_inode_drop)
1997 releases += ceph_encode_inode_release(&p,
1998 d_inode(req->r_old_dentry),
1999 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2001 if (drop_cap_releases) {
2003 p = msg->front.iov_base + req->r_request_release_offset;
2006 head->num_releases = cpu_to_le16(releases);
2010 struct ceph_timespec ts;
2011 ceph_encode_timespec(&ts, &req->r_stamp);
2012 ceph_encode_copy(&p, &ts, sizeof(ts));
2016 msg->front.iov_len = p - msg->front.iov_base;
2017 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2019 if (req->r_pagelist) {
2020 struct ceph_pagelist *pagelist = req->r_pagelist;
2021 refcount_inc(&pagelist->refcnt);
2022 ceph_msg_data_add_pagelist(msg, pagelist);
2023 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2025 msg->hdr.data_len = 0;
2028 msg->hdr.data_off = cpu_to_le16(0);
2032 kfree((char *)path2);
2035 kfree((char *)path1);
2041 * called under mdsc->mutex if error, under no mutex if
2044 static void complete_request(struct ceph_mds_client *mdsc,
2045 struct ceph_mds_request *req)
2047 if (req->r_callback)
2048 req->r_callback(mdsc, req);
2050 complete_all(&req->r_completion);
2054 * called under mdsc->mutex
2056 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2057 struct ceph_mds_request *req,
2058 int mds, bool drop_cap_releases)
2060 struct ceph_mds_request_head *rhead;
2061 struct ceph_msg *msg;
2066 struct ceph_cap *cap =
2067 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2070 req->r_sent_on_mseq = cap->mseq;
2072 req->r_sent_on_mseq = -1;
2074 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2075 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2077 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2080 * Replay. Do not regenerate message (and rebuild
2081 * paths, etc.); just use the original message.
2082 * Rebuilding paths will break for renames because
2083 * d_move mangles the src name.
2085 msg = req->r_request;
2086 rhead = msg->front.iov_base;
2088 flags = le32_to_cpu(rhead->flags);
2089 flags |= CEPH_MDS_FLAG_REPLAY;
2090 rhead->flags = cpu_to_le32(flags);
2092 if (req->r_target_inode)
2093 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2095 rhead->num_retry = req->r_attempts - 1;
2097 /* remove cap/dentry releases from message */
2098 rhead->num_releases = 0;
2101 p = msg->front.iov_base + req->r_request_release_offset;
2103 struct ceph_timespec ts;
2104 ceph_encode_timespec(&ts, &req->r_stamp);
2105 ceph_encode_copy(&p, &ts, sizeof(ts));
2108 msg->front.iov_len = p - msg->front.iov_base;
2109 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2113 if (req->r_request) {
2114 ceph_msg_put(req->r_request);
2115 req->r_request = NULL;
2117 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2119 req->r_err = PTR_ERR(msg);
2120 return PTR_ERR(msg);
2122 req->r_request = msg;
2124 rhead = msg->front.iov_base;
2125 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2126 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2127 flags |= CEPH_MDS_FLAG_REPLAY;
2129 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2130 rhead->flags = cpu_to_le32(flags);
2131 rhead->num_fwd = req->r_num_fwd;
2132 rhead->num_retry = req->r_attempts - 1;
2135 dout(" r_parent = %p\n", req->r_parent);
2140 * send request, or put it on the appropriate wait list.
2142 static int __do_request(struct ceph_mds_client *mdsc,
2143 struct ceph_mds_request *req)
2145 struct ceph_mds_session *session = NULL;
2149 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2150 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2151 __unregister_request(mdsc, req);
2155 if (req->r_timeout &&
2156 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2157 dout("do_request timed out\n");
2161 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2162 dout("do_request forced umount\n");
2166 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2167 if (mdsc->mdsmap_err) {
2168 err = mdsc->mdsmap_err;
2169 dout("do_request mdsmap err %d\n", err);
2172 if (mdsc->mdsmap->m_epoch == 0) {
2173 dout("do_request no mdsmap, waiting for map\n");
2174 list_add(&req->r_wait, &mdsc->waiting_for_map);
2177 if (!(mdsc->fsc->mount_options->flags &
2178 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2179 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2181 pr_info("probably no mds server is up\n");
2186 put_request_session(req);
2188 mds = __choose_mds(mdsc, req);
2190 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2191 dout("do_request no mds or not active, waiting for map\n");
2192 list_add(&req->r_wait, &mdsc->waiting_for_map);
2196 /* get, open session */
2197 session = __ceph_lookup_mds_session(mdsc, mds);
2199 session = register_session(mdsc, mds);
2200 if (IS_ERR(session)) {
2201 err = PTR_ERR(session);
2205 req->r_session = get_session(session);
2207 dout("do_request mds%d session %p state %s\n", mds, session,
2208 ceph_session_state_name(session->s_state));
2209 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2210 session->s_state != CEPH_MDS_SESSION_HUNG) {
2211 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2215 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2216 session->s_state == CEPH_MDS_SESSION_CLOSING)
2217 __open_session(mdsc, session);
2218 list_add(&req->r_wait, &session->s_waiting);
2223 req->r_resend_mds = -1; /* forget any previous mds hint */
2225 if (req->r_request_started == 0) /* note request start time */
2226 req->r_request_started = jiffies;
2228 err = __prepare_send_request(mdsc, req, mds, false);
2230 ceph_msg_get(req->r_request);
2231 ceph_con_send(&session->s_con, req->r_request);
2235 ceph_put_mds_session(session);
2238 dout("__do_request early error %d\n", err);
2240 complete_request(mdsc, req);
2241 __unregister_request(mdsc, req);
2248 * called under mdsc->mutex
2250 static void __wake_requests(struct ceph_mds_client *mdsc,
2251 struct list_head *head)
2253 struct ceph_mds_request *req;
2254 LIST_HEAD(tmp_list);
2256 list_splice_init(head, &tmp_list);
2258 while (!list_empty(&tmp_list)) {
2259 req = list_entry(tmp_list.next,
2260 struct ceph_mds_request, r_wait);
2261 list_del_init(&req->r_wait);
2262 dout(" wake request %p tid %llu\n", req, req->r_tid);
2263 __do_request(mdsc, req);
2268 * Wake up threads with requests pending for @mds, so that they can
2269 * resubmit their requests to a possibly different mds.
2271 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2273 struct ceph_mds_request *req;
2274 struct rb_node *p = rb_first(&mdsc->request_tree);
2276 dout("kick_requests mds%d\n", mds);
2278 req = rb_entry(p, struct ceph_mds_request, r_node);
2280 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2282 if (req->r_attempts > 0)
2283 continue; /* only new requests */
2284 if (req->r_session &&
2285 req->r_session->s_mds == mds) {
2286 dout(" kicking tid %llu\n", req->r_tid);
2287 list_del_init(&req->r_wait);
2288 __do_request(mdsc, req);
2293 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2294 struct ceph_mds_request *req)
2296 dout("submit_request on %p\n", req);
2297 mutex_lock(&mdsc->mutex);
2298 __register_request(mdsc, req, NULL);
2299 __do_request(mdsc, req);
2300 mutex_unlock(&mdsc->mutex);
2304 * Synchrously perform an mds request. Take care of all of the
2305 * session setup, forwarding, retry details.
2307 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2309 struct ceph_mds_request *req)
2313 dout("do_request on %p\n", req);
2315 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2317 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2319 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2320 if (req->r_old_dentry_dir)
2321 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2325 mutex_lock(&mdsc->mutex);
2326 __register_request(mdsc, req, dir);
2327 __do_request(mdsc, req);
2335 mutex_unlock(&mdsc->mutex);
2336 dout("do_request waiting\n");
2337 if (!req->r_timeout && req->r_wait_for_completion) {
2338 err = req->r_wait_for_completion(mdsc, req);
2340 long timeleft = wait_for_completion_killable_timeout(
2342 ceph_timeout_jiffies(req->r_timeout));
2346 err = -EIO; /* timed out */
2348 err = timeleft; /* killed */
2350 dout("do_request waited, got %d\n", err);
2351 mutex_lock(&mdsc->mutex);
2353 /* only abort if we didn't race with a real reply */
2354 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2355 err = le32_to_cpu(req->r_reply_info.head->result);
2356 } else if (err < 0) {
2357 dout("aborted request %lld with %d\n", req->r_tid, err);
2360 * ensure we aren't running concurrently with
2361 * ceph_fill_trace or ceph_readdir_prepopulate, which
2362 * rely on locks (dir mutex) held by our caller.
2364 mutex_lock(&req->r_fill_mutex);
2366 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2367 mutex_unlock(&req->r_fill_mutex);
2369 if (req->r_parent &&
2370 (req->r_op & CEPH_MDS_OP_WRITE))
2371 ceph_invalidate_dir_request(req);
2377 mutex_unlock(&mdsc->mutex);
2378 dout("do_request %p done, result %d\n", req, err);
2383 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2384 * namespace request.
2386 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2388 struct inode *inode = req->r_parent;
2390 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2392 ceph_dir_clear_complete(inode);
2394 ceph_invalidate_dentry_lease(req->r_dentry);
2395 if (req->r_old_dentry)
2396 ceph_invalidate_dentry_lease(req->r_old_dentry);
2402 * We take the session mutex and parse and process the reply immediately.
2403 * This preserves the logical ordering of replies, capabilities, etc., sent
2404 * by the MDS as they are applied to our local cache.
2406 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2408 struct ceph_mds_client *mdsc = session->s_mdsc;
2409 struct ceph_mds_request *req;
2410 struct ceph_mds_reply_head *head = msg->front.iov_base;
2411 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2412 struct ceph_snap_realm *realm;
2415 int mds = session->s_mds;
2417 if (msg->front.iov_len < sizeof(*head)) {
2418 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2423 /* get request, session */
2424 tid = le64_to_cpu(msg->hdr.tid);
2425 mutex_lock(&mdsc->mutex);
2426 req = lookup_get_request(mdsc, tid);
2428 dout("handle_reply on unknown tid %llu\n", tid);
2429 mutex_unlock(&mdsc->mutex);
2432 dout("handle_reply %p\n", req);
2434 /* correct session? */
2435 if (req->r_session != session) {
2436 pr_err("mdsc_handle_reply got %llu on session mds%d"
2437 " not mds%d\n", tid, session->s_mds,
2438 req->r_session ? req->r_session->s_mds : -1);
2439 mutex_unlock(&mdsc->mutex);
2444 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2445 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2446 pr_warn("got a dup %s reply on %llu from mds%d\n",
2447 head->safe ? "safe" : "unsafe", tid, mds);
2448 mutex_unlock(&mdsc->mutex);
2451 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2452 pr_warn("got unsafe after safe on %llu from mds%d\n",
2454 mutex_unlock(&mdsc->mutex);
2458 result = le32_to_cpu(head->result);
2462 * if we're not talking to the authority, send to them
2463 * if the authority has changed while we weren't looking,
2464 * send to new authority
2465 * Otherwise we just have to return an ESTALE
2467 if (result == -ESTALE) {
2468 dout("got ESTALE on request %llu", req->r_tid);
2469 req->r_resend_mds = -1;
2470 if (req->r_direct_mode != USE_AUTH_MDS) {
2471 dout("not using auth, setting for that now");
2472 req->r_direct_mode = USE_AUTH_MDS;
2473 __do_request(mdsc, req);
2474 mutex_unlock(&mdsc->mutex);
2477 int mds = __choose_mds(mdsc, req);
2478 if (mds >= 0 && mds != req->r_session->s_mds) {
2479 dout("but auth changed, so resending");
2480 __do_request(mdsc, req);
2481 mutex_unlock(&mdsc->mutex);
2485 dout("have to return ESTALE on request %llu", req->r_tid);
2490 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2491 __unregister_request(mdsc, req);
2493 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2495 * We already handled the unsafe response, now do the
2496 * cleanup. No need to examine the response; the MDS
2497 * doesn't include any result info in the safe
2498 * response. And even if it did, there is nothing
2499 * useful we could do with a revised return value.
2501 dout("got safe reply %llu, mds%d\n", tid, mds);
2503 /* last unsafe request during umount? */
2504 if (mdsc->stopping && !__get_oldest_req(mdsc))
2505 complete_all(&mdsc->safe_umount_waiters);
2506 mutex_unlock(&mdsc->mutex);
2510 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2511 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2512 if (req->r_unsafe_dir) {
2513 struct ceph_inode_info *ci =
2514 ceph_inode(req->r_unsafe_dir);
2515 spin_lock(&ci->i_unsafe_lock);
2516 list_add_tail(&req->r_unsafe_dir_item,
2517 &ci->i_unsafe_dirops);
2518 spin_unlock(&ci->i_unsafe_lock);
2522 dout("handle_reply tid %lld result %d\n", tid, result);
2523 rinfo = &req->r_reply_info;
2524 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2525 mutex_unlock(&mdsc->mutex);
2527 mutex_lock(&session->s_mutex);
2529 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2536 if (rinfo->snapblob_len) {
2537 down_write(&mdsc->snap_rwsem);
2538 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2539 rinfo->snapblob + rinfo->snapblob_len,
2540 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2542 downgrade_write(&mdsc->snap_rwsem);
2544 down_read(&mdsc->snap_rwsem);
2547 /* insert trace into our cache */
2548 mutex_lock(&req->r_fill_mutex);
2549 current->journal_info = req;
2550 err = ceph_fill_trace(mdsc->fsc->sb, req);
2552 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2553 req->r_op == CEPH_MDS_OP_LSSNAP))
2554 ceph_readdir_prepopulate(req, req->r_session);
2555 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2557 current->journal_info = NULL;
2558 mutex_unlock(&req->r_fill_mutex);
2560 up_read(&mdsc->snap_rwsem);
2562 ceph_put_snap_realm(mdsc, realm);
2564 if (err == 0 && req->r_target_inode &&
2565 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2566 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2567 spin_lock(&ci->i_unsafe_lock);
2568 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2569 spin_unlock(&ci->i_unsafe_lock);
2572 mutex_lock(&mdsc->mutex);
2573 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2577 req->r_reply = ceph_msg_get(msg);
2578 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2581 dout("reply arrived after request %lld was aborted\n", tid);
2583 mutex_unlock(&mdsc->mutex);
2585 mutex_unlock(&session->s_mutex);
2587 /* kick calling process */
2588 complete_request(mdsc, req);
2590 ceph_mdsc_put_request(req);
2597 * handle mds notification that our request has been forwarded.
2599 static void handle_forward(struct ceph_mds_client *mdsc,
2600 struct ceph_mds_session *session,
2601 struct ceph_msg *msg)
2603 struct ceph_mds_request *req;
2604 u64 tid = le64_to_cpu(msg->hdr.tid);
2608 void *p = msg->front.iov_base;
2609 void *end = p + msg->front.iov_len;
2611 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2612 next_mds = ceph_decode_32(&p);
2613 fwd_seq = ceph_decode_32(&p);
2615 mutex_lock(&mdsc->mutex);
2616 req = lookup_get_request(mdsc, tid);
2618 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2619 goto out; /* dup reply? */
2622 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2623 dout("forward tid %llu aborted, unregistering\n", tid);
2624 __unregister_request(mdsc, req);
2625 } else if (fwd_seq <= req->r_num_fwd) {
2626 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2627 tid, next_mds, req->r_num_fwd, fwd_seq);
2629 /* resend. forward race not possible; mds would drop */
2630 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2632 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2633 req->r_attempts = 0;
2634 req->r_num_fwd = fwd_seq;
2635 req->r_resend_mds = next_mds;
2636 put_request_session(req);
2637 __do_request(mdsc, req);
2639 ceph_mdsc_put_request(req);
2641 mutex_unlock(&mdsc->mutex);
2645 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2649 * handle a mds session control message
2651 static void handle_session(struct ceph_mds_session *session,
2652 struct ceph_msg *msg)
2654 struct ceph_mds_client *mdsc = session->s_mdsc;
2657 int mds = session->s_mds;
2658 struct ceph_mds_session_head *h = msg->front.iov_base;
2662 if (msg->front.iov_len != sizeof(*h))
2664 op = le32_to_cpu(h->op);
2665 seq = le64_to_cpu(h->seq);
2667 mutex_lock(&mdsc->mutex);
2668 if (op == CEPH_SESSION_CLOSE) {
2669 get_session(session);
2670 __unregister_session(mdsc, session);
2672 /* FIXME: this ttl calculation is generous */
2673 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2674 mutex_unlock(&mdsc->mutex);
2676 mutex_lock(&session->s_mutex);
2678 dout("handle_session mds%d %s %p state %s seq %llu\n",
2679 mds, ceph_session_op_name(op), session,
2680 ceph_session_state_name(session->s_state), seq);
2682 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2683 session->s_state = CEPH_MDS_SESSION_OPEN;
2684 pr_info("mds%d came back\n", session->s_mds);
2688 case CEPH_SESSION_OPEN:
2689 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2690 pr_info("mds%d reconnect success\n", session->s_mds);
2691 session->s_state = CEPH_MDS_SESSION_OPEN;
2692 renewed_caps(mdsc, session, 0);
2695 __close_session(mdsc, session);
2698 case CEPH_SESSION_RENEWCAPS:
2699 if (session->s_renew_seq == seq)
2700 renewed_caps(mdsc, session, 1);
2703 case CEPH_SESSION_CLOSE:
2704 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2705 pr_info("mds%d reconnect denied\n", session->s_mds);
2706 cleanup_session_requests(mdsc, session);
2707 remove_session_caps(session);
2708 wake = 2; /* for good measure */
2709 wake_up_all(&mdsc->session_close_wq);
2712 case CEPH_SESSION_STALE:
2713 pr_info("mds%d caps went stale, renewing\n",
2715 spin_lock(&session->s_gen_ttl_lock);
2716 session->s_cap_gen++;
2717 session->s_cap_ttl = jiffies - 1;
2718 spin_unlock(&session->s_gen_ttl_lock);
2719 send_renew_caps(mdsc, session);
2722 case CEPH_SESSION_RECALL_STATE:
2723 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2726 case CEPH_SESSION_FLUSHMSG:
2727 send_flushmsg_ack(mdsc, session, seq);
2730 case CEPH_SESSION_FORCE_RO:
2731 dout("force_session_readonly %p\n", session);
2732 spin_lock(&session->s_cap_lock);
2733 session->s_readonly = true;
2734 spin_unlock(&session->s_cap_lock);
2735 wake_up_session_caps(session, 0);
2738 case CEPH_SESSION_REJECT:
2739 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2740 pr_info("mds%d rejected session\n", session->s_mds);
2741 session->s_state = CEPH_MDS_SESSION_REJECTED;
2742 cleanup_session_requests(mdsc, session);
2743 remove_session_caps(session);
2744 wake = 2; /* for good measure */
2748 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2752 mutex_unlock(&session->s_mutex);
2754 mutex_lock(&mdsc->mutex);
2755 __wake_requests(mdsc, &session->s_waiting);
2757 kick_requests(mdsc, mds);
2758 mutex_unlock(&mdsc->mutex);
2760 if (op == CEPH_SESSION_CLOSE)
2761 ceph_put_mds_session(session);
2765 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2766 (int)msg->front.iov_len);
2773 * called under session->mutex.
2775 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2776 struct ceph_mds_session *session)
2778 struct ceph_mds_request *req, *nreq;
2782 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2784 mutex_lock(&mdsc->mutex);
2785 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2786 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2788 ceph_msg_get(req->r_request);
2789 ceph_con_send(&session->s_con, req->r_request);
2794 * also re-send old requests when MDS enters reconnect stage. So that MDS
2795 * can process completed request in clientreplay stage.
2797 p = rb_first(&mdsc->request_tree);
2799 req = rb_entry(p, struct ceph_mds_request, r_node);
2801 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2803 if (req->r_attempts == 0)
2804 continue; /* only old requests */
2805 if (req->r_session &&
2806 req->r_session->s_mds == session->s_mds) {
2807 err = __prepare_send_request(mdsc, req,
2808 session->s_mds, true);
2810 ceph_msg_get(req->r_request);
2811 ceph_con_send(&session->s_con, req->r_request);
2815 mutex_unlock(&mdsc->mutex);
2819 * Encode information about a cap for a reconnect with the MDS.
2821 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2825 struct ceph_mds_cap_reconnect v2;
2826 struct ceph_mds_cap_reconnect_v1 v1;
2828 struct ceph_inode_info *ci;
2829 struct ceph_reconnect_state *recon_state = arg;
2830 struct ceph_pagelist *pagelist = recon_state->pagelist;
2835 struct dentry *dentry;
2839 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2840 inode, ceph_vinop(inode), cap, cap->cap_id,
2841 ceph_cap_string(cap->issued));
2842 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2846 dentry = d_find_alias(inode);
2848 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2850 err = PTR_ERR(path);
2859 spin_lock(&ci->i_ceph_lock);
2860 cap->seq = 0; /* reset cap seq */
2861 cap->issue_seq = 0; /* and issue_seq */
2862 cap->mseq = 0; /* and migrate_seq */
2863 cap->cap_gen = cap->session->s_cap_gen;
2865 if (recon_state->msg_version >= 2) {
2866 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2867 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2868 rec.v2.issued = cpu_to_le32(cap->issued);
2869 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2870 rec.v2.pathbase = cpu_to_le64(pathbase);
2871 rec.v2.flock_len = 0;
2873 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2874 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2875 rec.v1.issued = cpu_to_le32(cap->issued);
2876 rec.v1.size = cpu_to_le64(inode->i_size);
2877 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2878 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2879 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2880 rec.v1.pathbase = cpu_to_le64(pathbase);
2883 if (list_empty(&ci->i_cap_snaps)) {
2884 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
2886 struct ceph_cap_snap *capsnap =
2887 list_first_entry(&ci->i_cap_snaps,
2888 struct ceph_cap_snap, ci_item);
2889 snap_follows = capsnap->follows;
2891 spin_unlock(&ci->i_ceph_lock);
2893 if (recon_state->msg_version >= 2) {
2894 int num_fcntl_locks, num_flock_locks;
2895 struct ceph_filelock *flocks;
2896 size_t struct_len, total_len = 0;
2900 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2901 flocks = kmalloc((num_fcntl_locks+num_flock_locks) *
2902 sizeof(struct ceph_filelock), GFP_NOFS);
2907 err = ceph_encode_locks_to_buffer(inode, flocks,
2917 if (recon_state->msg_version >= 3) {
2918 /* version, compat_version and struct_len */
2919 total_len = 2 * sizeof(u8) + sizeof(u32);
2923 * number of encoded locks is stable, so copy to pagelist
2925 struct_len = 2 * sizeof(u32) +
2926 (num_fcntl_locks + num_flock_locks) *
2927 sizeof(struct ceph_filelock);
2928 rec.v2.flock_len = cpu_to_le32(struct_len);
2930 struct_len += sizeof(rec.v2);
2931 struct_len += sizeof(u32) + pathlen;
2934 struct_len += sizeof(u64); /* snap_follows */
2936 total_len += struct_len;
2937 err = ceph_pagelist_reserve(pagelist, total_len);
2940 if (recon_state->msg_version >= 3) {
2941 ceph_pagelist_encode_8(pagelist, struct_v);
2942 ceph_pagelist_encode_8(pagelist, 1);
2943 ceph_pagelist_encode_32(pagelist, struct_len);
2945 ceph_pagelist_encode_string(pagelist, path, pathlen);
2946 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2947 ceph_locks_to_pagelist(flocks, pagelist,
2951 ceph_pagelist_encode_64(pagelist, snap_follows);
2955 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2956 err = ceph_pagelist_reserve(pagelist, size);
2958 ceph_pagelist_encode_string(pagelist, path, pathlen);
2959 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2963 recon_state->nr_caps++;
2973 * If an MDS fails and recovers, clients need to reconnect in order to
2974 * reestablish shared state. This includes all caps issued through
2975 * this session _and_ the snap_realm hierarchy. Because it's not
2976 * clear which snap realms the mds cares about, we send everything we
2977 * know about.. that ensures we'll then get any new info the
2978 * recovering MDS might have.
2980 * This is a relatively heavyweight operation, but it's rare.
2982 * called with mdsc->mutex held.
2984 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2985 struct ceph_mds_session *session)
2987 struct ceph_msg *reply;
2989 int mds = session->s_mds;
2992 struct ceph_pagelist *pagelist;
2993 struct ceph_reconnect_state recon_state;
2995 pr_info("mds%d reconnect start\n", mds);
2997 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
2999 goto fail_nopagelist;
3000 ceph_pagelist_init(pagelist);
3002 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
3006 mutex_lock(&session->s_mutex);
3007 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3010 dout("session %p state %s\n", session,
3011 ceph_session_state_name(session->s_state));
3013 spin_lock(&session->s_gen_ttl_lock);
3014 session->s_cap_gen++;
3015 spin_unlock(&session->s_gen_ttl_lock);
3017 spin_lock(&session->s_cap_lock);
3018 /* don't know if session is readonly */
3019 session->s_readonly = 0;
3021 * notify __ceph_remove_cap() that we are composing cap reconnect.
3022 * If a cap get released before being added to the cap reconnect,
3023 * __ceph_remove_cap() should skip queuing cap release.
3025 session->s_cap_reconnect = 1;
3026 /* drop old cap expires; we're about to reestablish that state */
3027 cleanup_cap_releases(mdsc, session);
3029 /* trim unused caps to reduce MDS's cache rejoin time */
3030 if (mdsc->fsc->sb->s_root)
3031 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3033 ceph_con_close(&session->s_con);
3034 ceph_con_open(&session->s_con,
3035 CEPH_ENTITY_TYPE_MDS, mds,
3036 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3038 /* replay unsafe requests */
3039 replay_unsafe_requests(mdsc, session);
3041 down_read(&mdsc->snap_rwsem);
3043 /* traverse this session's caps */
3044 s_nr_caps = session->s_nr_caps;
3045 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3049 recon_state.nr_caps = 0;
3050 recon_state.pagelist = pagelist;
3051 if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3052 recon_state.msg_version = 3;
3053 else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3054 recon_state.msg_version = 2;
3056 recon_state.msg_version = 1;
3057 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3061 spin_lock(&session->s_cap_lock);
3062 session->s_cap_reconnect = 0;
3063 spin_unlock(&session->s_cap_lock);
3066 * snaprealms. we provide mds with the ino, seq (version), and
3067 * parent for all of our realms. If the mds has any newer info,
3070 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3071 struct ceph_snap_realm *realm =
3072 rb_entry(p, struct ceph_snap_realm, node);
3073 struct ceph_mds_snaprealm_reconnect sr_rec;
3075 dout(" adding snap realm %llx seq %lld parent %llx\n",
3076 realm->ino, realm->seq, realm->parent_ino);
3077 sr_rec.ino = cpu_to_le64(realm->ino);
3078 sr_rec.seq = cpu_to_le64(realm->seq);
3079 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3080 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3085 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3087 /* raced with cap release? */
3088 if (s_nr_caps != recon_state.nr_caps) {
3089 struct page *page = list_first_entry(&pagelist->head,
3091 __le32 *addr = kmap_atomic(page);
3092 *addr = cpu_to_le32(recon_state.nr_caps);
3093 kunmap_atomic(addr);
3096 reply->hdr.data_len = cpu_to_le32(pagelist->length);
3097 ceph_msg_data_add_pagelist(reply, pagelist);
3099 ceph_early_kick_flushing_caps(mdsc, session);
3101 ceph_con_send(&session->s_con, reply);
3103 mutex_unlock(&session->s_mutex);
3105 mutex_lock(&mdsc->mutex);
3106 __wake_requests(mdsc, &session->s_waiting);
3107 mutex_unlock(&mdsc->mutex);
3109 up_read(&mdsc->snap_rwsem);
3113 ceph_msg_put(reply);
3114 up_read(&mdsc->snap_rwsem);
3115 mutex_unlock(&session->s_mutex);
3117 ceph_pagelist_release(pagelist);
3119 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3125 * compare old and new mdsmaps, kicking requests
3126 * and closing out old connections as necessary
3128 * called under mdsc->mutex.
3130 static void check_new_map(struct ceph_mds_client *mdsc,
3131 struct ceph_mdsmap *newmap,
3132 struct ceph_mdsmap *oldmap)
3135 int oldstate, newstate;
3136 struct ceph_mds_session *s;
3138 dout("check_new_map new %u old %u\n",
3139 newmap->m_epoch, oldmap->m_epoch);
3141 for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3142 if (!mdsc->sessions[i])
3144 s = mdsc->sessions[i];
3145 oldstate = ceph_mdsmap_get_state(oldmap, i);
3146 newstate = ceph_mdsmap_get_state(newmap, i);
3148 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3149 i, ceph_mds_state_name(oldstate),
3150 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3151 ceph_mds_state_name(newstate),
3152 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3153 ceph_session_state_name(s->s_state));
3155 if (i >= newmap->m_num_mds ||
3156 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3157 ceph_mdsmap_get_addr(newmap, i),
3158 sizeof(struct ceph_entity_addr))) {
3159 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3160 /* the session never opened, just close it
3163 __unregister_session(mdsc, s);
3164 __wake_requests(mdsc, &s->s_waiting);
3165 ceph_put_mds_session(s);
3166 } else if (i >= newmap->m_num_mds) {
3167 /* force close session for stopped mds */
3169 __unregister_session(mdsc, s);
3170 __wake_requests(mdsc, &s->s_waiting);
3171 kick_requests(mdsc, i);
3172 mutex_unlock(&mdsc->mutex);
3174 mutex_lock(&s->s_mutex);
3175 cleanup_session_requests(mdsc, s);
3176 remove_session_caps(s);
3177 mutex_unlock(&s->s_mutex);
3179 ceph_put_mds_session(s);
3181 mutex_lock(&mdsc->mutex);
3184 mutex_unlock(&mdsc->mutex);
3185 mutex_lock(&s->s_mutex);
3186 mutex_lock(&mdsc->mutex);
3187 ceph_con_close(&s->s_con);
3188 mutex_unlock(&s->s_mutex);
3189 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3191 } else if (oldstate == newstate) {
3192 continue; /* nothing new with this mds */
3198 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3199 newstate >= CEPH_MDS_STATE_RECONNECT) {
3200 mutex_unlock(&mdsc->mutex);
3201 send_mds_reconnect(mdsc, s);
3202 mutex_lock(&mdsc->mutex);
3206 * kick request on any mds that has gone active.
3208 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3209 newstate >= CEPH_MDS_STATE_ACTIVE) {
3210 if (oldstate != CEPH_MDS_STATE_CREATING &&
3211 oldstate != CEPH_MDS_STATE_STARTING)
3212 pr_info("mds%d recovery completed\n", s->s_mds);
3213 kick_requests(mdsc, i);
3214 ceph_kick_flushing_caps(mdsc, s);
3215 wake_up_session_caps(s, 1);
3219 for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3220 s = mdsc->sessions[i];
3223 if (!ceph_mdsmap_is_laggy(newmap, i))
3225 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3226 s->s_state == CEPH_MDS_SESSION_HUNG ||
3227 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3228 dout(" connecting to export targets of laggy mds%d\n",
3230 __open_export_target_sessions(mdsc, s);
3242 * caller must hold session s_mutex, dentry->d_lock
3244 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3246 struct ceph_dentry_info *di = ceph_dentry(dentry);
3248 ceph_put_mds_session(di->lease_session);
3249 di->lease_session = NULL;
3252 static void handle_lease(struct ceph_mds_client *mdsc,
3253 struct ceph_mds_session *session,
3254 struct ceph_msg *msg)
3256 struct super_block *sb = mdsc->fsc->sb;
3257 struct inode *inode;
3258 struct dentry *parent, *dentry;
3259 struct ceph_dentry_info *di;
3260 int mds = session->s_mds;
3261 struct ceph_mds_lease *h = msg->front.iov_base;
3263 struct ceph_vino vino;
3267 dout("handle_lease from mds%d\n", mds);
3270 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3272 vino.ino = le64_to_cpu(h->ino);
3273 vino.snap = CEPH_NOSNAP;
3274 seq = le32_to_cpu(h->seq);
3275 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3276 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3277 if (dname.len != get_unaligned_le32(h+1))
3281 inode = ceph_find_inode(sb, vino);
3282 dout("handle_lease %s, ino %llx %p %.*s\n",
3283 ceph_lease_op_name(h->action), vino.ino, inode,
3284 dname.len, dname.name);
3286 mutex_lock(&session->s_mutex);
3290 dout("handle_lease no inode %llx\n", vino.ino);
3295 parent = d_find_alias(inode);
3297 dout("no parent dentry on inode %p\n", inode);
3299 goto release; /* hrm... */
3301 dname.hash = full_name_hash(parent, dname.name, dname.len);
3302 dentry = d_lookup(parent, &dname);
3307 spin_lock(&dentry->d_lock);
3308 di = ceph_dentry(dentry);
3309 switch (h->action) {
3310 case CEPH_MDS_LEASE_REVOKE:
3311 if (di->lease_session == session) {
3312 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3313 h->seq = cpu_to_le32(di->lease_seq);
3314 __ceph_mdsc_drop_dentry_lease(dentry);
3319 case CEPH_MDS_LEASE_RENEW:
3320 if (di->lease_session == session &&
3321 di->lease_gen == session->s_cap_gen &&
3322 di->lease_renew_from &&
3323 di->lease_renew_after == 0) {
3324 unsigned long duration =
3325 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3327 di->lease_seq = seq;
3328 di->time = di->lease_renew_from + duration;
3329 di->lease_renew_after = di->lease_renew_from +
3331 di->lease_renew_from = 0;
3335 spin_unlock(&dentry->d_lock);
3342 /* let's just reuse the same message */
3343 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3345 ceph_con_send(&session->s_con, msg);
3349 mutex_unlock(&session->s_mutex);
3353 pr_err("corrupt lease message\n");
3357 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3358 struct inode *inode,
3359 struct dentry *dentry, char action,
3362 struct ceph_msg *msg;
3363 struct ceph_mds_lease *lease;
3364 int len = sizeof(*lease) + sizeof(u32);
3367 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3368 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3369 dnamelen = dentry->d_name.len;
3372 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3375 lease = msg->front.iov_base;
3376 lease->action = action;
3377 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3378 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3379 lease->seq = cpu_to_le32(seq);
3380 put_unaligned_le32(dnamelen, lease + 1);
3381 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3384 * if this is a preemptive lease RELEASE, no need to
3385 * flush request stream, since the actual request will
3388 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3390 ceph_con_send(&session->s_con, msg);
3394 * drop all leases (and dentry refs) in preparation for umount
3396 static void drop_leases(struct ceph_mds_client *mdsc)
3400 dout("drop_leases\n");
3401 mutex_lock(&mdsc->mutex);
3402 for (i = 0; i < mdsc->max_sessions; i++) {
3403 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3406 mutex_unlock(&mdsc->mutex);
3407 mutex_lock(&s->s_mutex);
3408 mutex_unlock(&s->s_mutex);
3409 ceph_put_mds_session(s);
3410 mutex_lock(&mdsc->mutex);
3412 mutex_unlock(&mdsc->mutex);
3418 * delayed work -- periodically trim expired leases, renew caps with mds
3420 static void schedule_delayed(struct ceph_mds_client *mdsc)
3423 unsigned hz = round_jiffies_relative(HZ * delay);
3424 schedule_delayed_work(&mdsc->delayed_work, hz);
3427 static void delayed_work(struct work_struct *work)
3430 struct ceph_mds_client *mdsc =
3431 container_of(work, struct ceph_mds_client, delayed_work.work);
3435 dout("mdsc delayed_work\n");
3436 ceph_check_delayed_caps(mdsc);
3438 mutex_lock(&mdsc->mutex);
3439 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3440 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3441 mdsc->last_renew_caps);
3443 mdsc->last_renew_caps = jiffies;
3445 for (i = 0; i < mdsc->max_sessions; i++) {
3446 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3449 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3450 dout("resending session close request for mds%d\n",
3452 request_close_session(mdsc, s);
3453 ceph_put_mds_session(s);
3456 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3457 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3458 s->s_state = CEPH_MDS_SESSION_HUNG;
3459 pr_info("mds%d hung\n", s->s_mds);
3462 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3463 /* this mds is failed or recovering, just wait */
3464 ceph_put_mds_session(s);
3467 mutex_unlock(&mdsc->mutex);
3469 mutex_lock(&s->s_mutex);
3471 send_renew_caps(mdsc, s);
3473 ceph_con_keepalive(&s->s_con);
3474 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3475 s->s_state == CEPH_MDS_SESSION_HUNG)
3476 ceph_send_cap_releases(mdsc, s);
3477 mutex_unlock(&s->s_mutex);
3478 ceph_put_mds_session(s);
3480 mutex_lock(&mdsc->mutex);
3482 mutex_unlock(&mdsc->mutex);
3484 schedule_delayed(mdsc);
3487 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3490 struct ceph_mds_client *mdsc;
3492 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3497 mutex_init(&mdsc->mutex);
3498 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3499 if (!mdsc->mdsmap) {
3504 init_completion(&mdsc->safe_umount_waiters);
3505 init_waitqueue_head(&mdsc->session_close_wq);
3506 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3507 mdsc->sessions = NULL;
3508 atomic_set(&mdsc->num_sessions, 0);
3509 mdsc->max_sessions = 0;
3511 mdsc->last_snap_seq = 0;
3512 init_rwsem(&mdsc->snap_rwsem);
3513 mdsc->snap_realms = RB_ROOT;
3514 INIT_LIST_HEAD(&mdsc->snap_empty);
3515 spin_lock_init(&mdsc->snap_empty_lock);
3517 mdsc->oldest_tid = 0;
3518 mdsc->request_tree = RB_ROOT;
3519 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3520 mdsc->last_renew_caps = jiffies;
3521 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3522 spin_lock_init(&mdsc->cap_delay_lock);
3523 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3524 spin_lock_init(&mdsc->snap_flush_lock);
3525 mdsc->last_cap_flush_tid = 1;
3526 INIT_LIST_HEAD(&mdsc->cap_flush_list);
3527 INIT_LIST_HEAD(&mdsc->cap_dirty);
3528 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3529 mdsc->num_cap_flushing = 0;
3530 spin_lock_init(&mdsc->cap_dirty_lock);
3531 init_waitqueue_head(&mdsc->cap_flushing_wq);
3532 spin_lock_init(&mdsc->dentry_lru_lock);
3533 INIT_LIST_HEAD(&mdsc->dentry_lru);
3535 ceph_caps_init(mdsc);
3536 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3538 init_rwsem(&mdsc->pool_perm_rwsem);
3539 mdsc->pool_perm_tree = RB_ROOT;
3541 strncpy(mdsc->nodename, utsname()->nodename,
3542 sizeof(mdsc->nodename) - 1);
3547 * Wait for safe replies on open mds requests. If we time out, drop
3548 * all requests from the tree to avoid dangling dentry refs.
3550 static void wait_requests(struct ceph_mds_client *mdsc)
3552 struct ceph_options *opts = mdsc->fsc->client->options;
3553 struct ceph_mds_request *req;
3555 mutex_lock(&mdsc->mutex);
3556 if (__get_oldest_req(mdsc)) {
3557 mutex_unlock(&mdsc->mutex);
3559 dout("wait_requests waiting for requests\n");
3560 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3561 ceph_timeout_jiffies(opts->mount_timeout));
3563 /* tear down remaining requests */
3564 mutex_lock(&mdsc->mutex);
3565 while ((req = __get_oldest_req(mdsc))) {
3566 dout("wait_requests timed out on tid %llu\n",
3568 __unregister_request(mdsc, req);
3571 mutex_unlock(&mdsc->mutex);
3572 dout("wait_requests done\n");
3576 * called before mount is ro, and before dentries are torn down.
3577 * (hmm, does this still race with new lookups?)
3579 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3581 dout("pre_umount\n");
3585 ceph_flush_dirty_caps(mdsc);
3586 wait_requests(mdsc);
3589 * wait for reply handlers to drop their request refs and
3590 * their inode/dcache refs
3596 * wait for all write mds requests to flush.
3598 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3600 struct ceph_mds_request *req = NULL, *nextreq;
3603 mutex_lock(&mdsc->mutex);
3604 dout("wait_unsafe_requests want %lld\n", want_tid);
3606 req = __get_oldest_req(mdsc);
3607 while (req && req->r_tid <= want_tid) {
3608 /* find next request */
3609 n = rb_next(&req->r_node);
3611 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3614 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3615 (req->r_op & CEPH_MDS_OP_WRITE)) {
3617 ceph_mdsc_get_request(req);
3619 ceph_mdsc_get_request(nextreq);
3620 mutex_unlock(&mdsc->mutex);
3621 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3622 req->r_tid, want_tid);
3623 wait_for_completion(&req->r_safe_completion);
3624 mutex_lock(&mdsc->mutex);
3625 ceph_mdsc_put_request(req);
3627 break; /* next dne before, so we're done! */
3628 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3629 /* next request was removed from tree */
3630 ceph_mdsc_put_request(nextreq);
3633 ceph_mdsc_put_request(nextreq); /* won't go away */
3637 mutex_unlock(&mdsc->mutex);
3638 dout("wait_unsafe_requests done\n");
3641 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3643 u64 want_tid, want_flush;
3645 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3649 mutex_lock(&mdsc->mutex);
3650 want_tid = mdsc->last_tid;
3651 mutex_unlock(&mdsc->mutex);
3653 ceph_flush_dirty_caps(mdsc);
3654 spin_lock(&mdsc->cap_dirty_lock);
3655 want_flush = mdsc->last_cap_flush_tid;
3656 if (!list_empty(&mdsc->cap_flush_list)) {
3657 struct ceph_cap_flush *cf =
3658 list_last_entry(&mdsc->cap_flush_list,
3659 struct ceph_cap_flush, g_list);
3662 spin_unlock(&mdsc->cap_dirty_lock);
3664 dout("sync want tid %lld flush_seq %lld\n",
3665 want_tid, want_flush);
3667 wait_unsafe_requests(mdsc, want_tid);
3668 wait_caps_flush(mdsc, want_flush);
3672 * true if all sessions are closed, or we force unmount
3674 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3676 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3678 return atomic_read(&mdsc->num_sessions) <= skipped;
3682 * called after sb is ro.
3684 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3686 struct ceph_options *opts = mdsc->fsc->client->options;
3687 struct ceph_mds_session *session;
3691 dout("close_sessions\n");
3693 /* close sessions */
3694 mutex_lock(&mdsc->mutex);
3695 for (i = 0; i < mdsc->max_sessions; i++) {
3696 session = __ceph_lookup_mds_session(mdsc, i);
3699 mutex_unlock(&mdsc->mutex);
3700 mutex_lock(&session->s_mutex);
3701 if (__close_session(mdsc, session) <= 0)
3703 mutex_unlock(&session->s_mutex);
3704 ceph_put_mds_session(session);
3705 mutex_lock(&mdsc->mutex);
3707 mutex_unlock(&mdsc->mutex);
3709 dout("waiting for sessions to close\n");
3710 wait_event_timeout(mdsc->session_close_wq,
3711 done_closing_sessions(mdsc, skipped),
3712 ceph_timeout_jiffies(opts->mount_timeout));
3714 /* tear down remaining sessions */
3715 mutex_lock(&mdsc->mutex);
3716 for (i = 0; i < mdsc->max_sessions; i++) {
3717 if (mdsc->sessions[i]) {
3718 session = get_session(mdsc->sessions[i]);
3719 __unregister_session(mdsc, session);
3720 mutex_unlock(&mdsc->mutex);
3721 mutex_lock(&session->s_mutex);
3722 remove_session_caps(session);
3723 mutex_unlock(&session->s_mutex);
3724 ceph_put_mds_session(session);
3725 mutex_lock(&mdsc->mutex);
3728 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3729 mutex_unlock(&mdsc->mutex);
3731 ceph_cleanup_empty_realms(mdsc);
3733 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3738 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3740 struct ceph_mds_session *session;
3743 dout("force umount\n");
3745 mutex_lock(&mdsc->mutex);
3746 for (mds = 0; mds < mdsc->max_sessions; mds++) {
3747 session = __ceph_lookup_mds_session(mdsc, mds);
3750 mutex_unlock(&mdsc->mutex);
3751 mutex_lock(&session->s_mutex);
3752 __close_session(mdsc, session);
3753 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3754 cleanup_session_requests(mdsc, session);
3755 remove_session_caps(session);
3757 mutex_unlock(&session->s_mutex);
3758 ceph_put_mds_session(session);
3759 mutex_lock(&mdsc->mutex);
3760 kick_requests(mdsc, mds);
3762 __wake_requests(mdsc, &mdsc->waiting_for_map);
3763 mutex_unlock(&mdsc->mutex);
3766 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3769 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3771 ceph_mdsmap_destroy(mdsc->mdsmap);
3772 kfree(mdsc->sessions);
3773 ceph_caps_finalize(mdsc);
3774 ceph_pool_perm_destroy(mdsc);
3777 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3779 struct ceph_mds_client *mdsc = fsc->mdsc;
3780 dout("mdsc_destroy %p\n", mdsc);
3782 /* flush out any connection work with references to us */
3785 ceph_mdsc_stop(mdsc);
3789 dout("mdsc_destroy %p done\n", mdsc);
3792 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3794 struct ceph_fs_client *fsc = mdsc->fsc;
3795 const char *mds_namespace = fsc->mount_options->mds_namespace;
3796 void *p = msg->front.iov_base;
3797 void *end = p + msg->front.iov_len;
3801 u32 mount_fscid = (u32)-1;
3802 u8 struct_v, struct_cv;
3805 ceph_decode_need(&p, end, sizeof(u32), bad);
3806 epoch = ceph_decode_32(&p);
3808 dout("handle_fsmap epoch %u\n", epoch);
3810 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3811 struct_v = ceph_decode_8(&p);
3812 struct_cv = ceph_decode_8(&p);
3813 map_len = ceph_decode_32(&p);
3815 ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3816 p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3818 num_fs = ceph_decode_32(&p);
3819 while (num_fs-- > 0) {
3820 void *info_p, *info_end;
3825 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3826 info_v = ceph_decode_8(&p);
3827 info_cv = ceph_decode_8(&p);
3828 info_len = ceph_decode_32(&p);
3829 ceph_decode_need(&p, end, info_len, bad);
3831 info_end = p + info_len;
3834 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3835 fscid = ceph_decode_32(&info_p);
3836 namelen = ceph_decode_32(&info_p);
3837 ceph_decode_need(&info_p, info_end, namelen, bad);
3839 if (mds_namespace &&
3840 strlen(mds_namespace) == namelen &&
3841 !strncmp(mds_namespace, (char *)info_p, namelen)) {
3842 mount_fscid = fscid;
3847 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3848 if (mount_fscid != (u32)-1) {
3849 fsc->client->monc.fs_cluster_id = mount_fscid;
3850 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3852 ceph_monc_renew_subs(&fsc->client->monc);
3859 pr_err("error decoding fsmap\n");
3861 mutex_lock(&mdsc->mutex);
3862 mdsc->mdsmap_err = -ENOENT;
3863 __wake_requests(mdsc, &mdsc->waiting_for_map);
3864 mutex_unlock(&mdsc->mutex);
3869 * handle mds map update.
3871 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3875 void *p = msg->front.iov_base;
3876 void *end = p + msg->front.iov_len;
3877 struct ceph_mdsmap *newmap, *oldmap;
3878 struct ceph_fsid fsid;
3881 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3882 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3883 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3885 epoch = ceph_decode_32(&p);
3886 maplen = ceph_decode_32(&p);
3887 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3889 /* do we need it? */
3890 mutex_lock(&mdsc->mutex);
3891 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3892 dout("handle_map epoch %u <= our %u\n",
3893 epoch, mdsc->mdsmap->m_epoch);
3894 mutex_unlock(&mdsc->mutex);
3898 newmap = ceph_mdsmap_decode(&p, end);
3899 if (IS_ERR(newmap)) {
3900 err = PTR_ERR(newmap);
3904 /* swap into place */
3906 oldmap = mdsc->mdsmap;
3907 mdsc->mdsmap = newmap;
3908 check_new_map(mdsc, newmap, oldmap);
3909 ceph_mdsmap_destroy(oldmap);
3911 mdsc->mdsmap = newmap; /* first mds map */
3913 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3915 __wake_requests(mdsc, &mdsc->waiting_for_map);
3916 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3917 mdsc->mdsmap->m_epoch);
3919 mutex_unlock(&mdsc->mutex);
3920 schedule_delayed(mdsc);
3924 mutex_unlock(&mdsc->mutex);
3926 pr_err("error decoding mdsmap %d\n", err);
3930 static struct ceph_connection *con_get(struct ceph_connection *con)
3932 struct ceph_mds_session *s = con->private;
3934 if (get_session(s)) {
3935 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
3938 dout("mdsc con_get %p FAIL\n", s);
3942 static void con_put(struct ceph_connection *con)
3944 struct ceph_mds_session *s = con->private;
3946 dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
3947 ceph_put_mds_session(s);
3951 * if the client is unresponsive for long enough, the mds will kill
3952 * the session entirely.
3954 static void peer_reset(struct ceph_connection *con)
3956 struct ceph_mds_session *s = con->private;
3957 struct ceph_mds_client *mdsc = s->s_mdsc;
3959 pr_warn("mds%d closed our session\n", s->s_mds);
3960 send_mds_reconnect(mdsc, s);
3963 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3965 struct ceph_mds_session *s = con->private;
3966 struct ceph_mds_client *mdsc = s->s_mdsc;
3967 int type = le16_to_cpu(msg->hdr.type);
3969 mutex_lock(&mdsc->mutex);
3970 if (__verify_registered_session(mdsc, s) < 0) {
3971 mutex_unlock(&mdsc->mutex);
3974 mutex_unlock(&mdsc->mutex);
3977 case CEPH_MSG_MDS_MAP:
3978 ceph_mdsc_handle_mdsmap(mdsc, msg);
3980 case CEPH_MSG_FS_MAP_USER:
3981 ceph_mdsc_handle_fsmap(mdsc, msg);
3983 case CEPH_MSG_CLIENT_SESSION:
3984 handle_session(s, msg);
3986 case CEPH_MSG_CLIENT_REPLY:
3987 handle_reply(s, msg);
3989 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
3990 handle_forward(mdsc, s, msg);
3992 case CEPH_MSG_CLIENT_CAPS:
3993 ceph_handle_caps(s, msg);
3995 case CEPH_MSG_CLIENT_SNAP:
3996 ceph_handle_snap(mdsc, s, msg);
3998 case CEPH_MSG_CLIENT_LEASE:
3999 handle_lease(mdsc, s, msg);
4003 pr_err("received unknown message type %d %s\n", type,
4004 ceph_msg_type_name(type));
4015 * Note: returned pointer is the address of a structure that's
4016 * managed separately. Caller must *not* attempt to free it.
4018 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4019 int *proto, int force_new)
4021 struct ceph_mds_session *s = con->private;
4022 struct ceph_mds_client *mdsc = s->s_mdsc;
4023 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4024 struct ceph_auth_handshake *auth = &s->s_auth;
4026 if (force_new && auth->authorizer) {
4027 ceph_auth_destroy_authorizer(auth->authorizer);
4028 auth->authorizer = NULL;
4030 if (!auth->authorizer) {
4031 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4034 return ERR_PTR(ret);
4036 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4039 return ERR_PTR(ret);
4041 *proto = ac->protocol;
4047 static int verify_authorizer_reply(struct ceph_connection *con)
4049 struct ceph_mds_session *s = con->private;
4050 struct ceph_mds_client *mdsc = s->s_mdsc;
4051 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4053 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4056 static int invalidate_authorizer(struct ceph_connection *con)
4058 struct ceph_mds_session *s = con->private;
4059 struct ceph_mds_client *mdsc = s->s_mdsc;
4060 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4062 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4064 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4067 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4068 struct ceph_msg_header *hdr, int *skip)
4070 struct ceph_msg *msg;
4071 int type = (int) le16_to_cpu(hdr->type);
4072 int front_len = (int) le32_to_cpu(hdr->front_len);
4078 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4080 pr_err("unable to allocate msg type %d len %d\n",
4088 static int mds_sign_message(struct ceph_msg *msg)
4090 struct ceph_mds_session *s = msg->con->private;
4091 struct ceph_auth_handshake *auth = &s->s_auth;
4093 return ceph_auth_sign_message(auth, msg);
4096 static int mds_check_message_signature(struct ceph_msg *msg)
4098 struct ceph_mds_session *s = msg->con->private;
4099 struct ceph_auth_handshake *auth = &s->s_auth;
4101 return ceph_auth_check_message_signature(auth, msg);
4104 static const struct ceph_connection_operations mds_con_ops = {
4107 .dispatch = dispatch,
4108 .get_authorizer = get_authorizer,
4109 .verify_authorizer_reply = verify_authorizer_reply,
4110 .invalidate_authorizer = invalidate_authorizer,
4111 .peer_reset = peer_reset,
4112 .alloc_msg = mds_alloc_msg,
4113 .sign_message = mds_sign_message,
4114 .check_message_signature = mds_check_message_signature,