1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
5 #include <linux/wait.h>
6 #include <linux/slab.h>
8 #include <linux/sched.h>
9 #include <linux/debugfs.h>
10 #include <linux/seq_file.h>
11 #include <linux/ratelimit.h>
14 #include "mds_client.h"
16 #include <linux/ceph/ceph_features.h>
17 #include <linux/ceph/messenger.h>
18 #include <linux/ceph/decode.h>
19 #include <linux/ceph/pagelist.h>
20 #include <linux/ceph/auth.h>
21 #include <linux/ceph/debugfs.h>
24 * A cluster of MDS (metadata server) daemons is responsible for
25 * managing the file system namespace (the directory hierarchy and
26 * inodes) and for coordinating shared access to storage. Metadata is
27 * partitioning hierarchically across a number of servers, and that
28 * partition varies over time as the cluster adjusts the distribution
29 * in order to balance load.
31 * The MDS client is primarily responsible to managing synchronous
32 * metadata requests for operations like open, unlink, and so forth.
33 * If there is a MDS failure, we find out about it when we (possibly
34 * request and) receive a new MDS map, and can resubmit affected
37 * For the most part, though, we take advantage of a lossless
38 * communications channel to the MDS, and do not need to worry about
39 * timing out or resubmitting requests.
41 * We maintain a stateful "session" with each MDS we interact with.
42 * Within each session, we sent periodic heartbeat messages to ensure
43 * any capabilities or leases we have been issues remain valid. If
44 * the session times out and goes stale, our leases and capabilities
45 * are no longer valid.
48 struct ceph_reconnect_state {
50 struct ceph_pagelist *pagelist;
54 static void __wake_requests(struct ceph_mds_client *mdsc,
55 struct list_head *head);
57 static const struct ceph_connection_operations mds_con_ops;
65 * parse individual inode info
67 static int parse_reply_info_in(void **p, void *end,
68 struct ceph_mds_reply_info_in *info,
74 *p += sizeof(struct ceph_mds_reply_inode) +
75 sizeof(*info->in->fragtree.splits) *
76 le32_to_cpu(info->in->fragtree.nsplits);
78 ceph_decode_32_safe(p, end, info->symlink_len, bad);
79 ceph_decode_need(p, end, info->symlink_len, bad);
81 *p += info->symlink_len;
83 if (features & CEPH_FEATURE_DIRLAYOUTHASH)
84 ceph_decode_copy_safe(p, end, &info->dir_layout,
85 sizeof(info->dir_layout), bad);
87 memset(&info->dir_layout, 0, sizeof(info->dir_layout));
89 ceph_decode_32_safe(p, end, info->xattr_len, bad);
90 ceph_decode_need(p, end, info->xattr_len, bad);
91 info->xattr_data = *p;
92 *p += info->xattr_len;
94 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
95 ceph_decode_64_safe(p, end, info->inline_version, bad);
96 ceph_decode_32_safe(p, end, info->inline_len, bad);
97 ceph_decode_need(p, end, info->inline_len, bad);
98 info->inline_data = *p;
99 *p += info->inline_len;
101 info->inline_version = CEPH_INLINE_NONE;
103 info->pool_ns_len = 0;
104 info->pool_ns_data = NULL;
105 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
106 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
107 if (info->pool_ns_len > 0) {
108 ceph_decode_need(p, end, info->pool_ns_len, bad);
109 info->pool_ns_data = *p;
110 *p += info->pool_ns_len;
120 * parse a normal reply, which may contain a (dir+)dentry and/or a
123 static int parse_reply_info_trace(void **p, void *end,
124 struct ceph_mds_reply_info_parsed *info,
129 if (info->head->is_dentry) {
130 err = parse_reply_info_in(p, end, &info->diri, features);
134 if (unlikely(*p + sizeof(*info->dirfrag) > end))
137 *p += sizeof(*info->dirfrag) +
138 sizeof(u32)*le32_to_cpu(info->dirfrag->ndist);
139 if (unlikely(*p > end))
142 ceph_decode_32_safe(p, end, info->dname_len, bad);
143 ceph_decode_need(p, end, info->dname_len, bad);
145 *p += info->dname_len;
147 *p += sizeof(*info->dlease);
150 if (info->head->is_target) {
151 err = parse_reply_info_in(p, end, &info->targeti, features);
156 if (unlikely(*p != end))
163 pr_err("problem parsing mds trace %d\n", err);
168 * parse readdir results
170 static int parse_reply_info_dir(void **p, void *end,
171 struct ceph_mds_reply_info_parsed *info,
178 if (*p + sizeof(*info->dir_dir) > end)
180 *p += sizeof(*info->dir_dir) +
181 sizeof(u32)*le32_to_cpu(info->dir_dir->ndist);
185 ceph_decode_need(p, end, sizeof(num) + 2, bad);
186 num = ceph_decode_32(p);
188 u16 flags = ceph_decode_16(p);
189 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
190 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
191 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
192 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
197 BUG_ON(!info->dir_entries);
198 if ((unsigned long)(info->dir_entries + num) >
199 (unsigned long)info->dir_entries + info->dir_buf_size) {
200 pr_err("dir contents are larger than expected\n");
207 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
209 ceph_decode_need(p, end, sizeof(u32)*2, bad);
210 rde->name_len = ceph_decode_32(p);
211 ceph_decode_need(p, end, rde->name_len, bad);
214 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
216 *p += sizeof(struct ceph_mds_reply_lease);
219 err = parse_reply_info_in(p, end, &rde->inode, features);
222 /* ceph_readdir_prepopulate() will update it */
236 pr_err("problem parsing dir contents %d\n", err);
241 * parse fcntl F_GETLK results
243 static int parse_reply_info_filelock(void **p, void *end,
244 struct ceph_mds_reply_info_parsed *info,
247 if (*p + sizeof(*info->filelock_reply) > end)
250 info->filelock_reply = *p;
251 *p += sizeof(*info->filelock_reply);
253 if (unlikely(*p != end))
262 * parse create results
264 static int parse_reply_info_create(void **p, void *end,
265 struct ceph_mds_reply_info_parsed *info,
268 if (features & CEPH_FEATURE_REPLY_CREATE_INODE) {
270 info->has_create_ino = false;
272 info->has_create_ino = true;
273 info->ino = ceph_decode_64(p);
277 if (unlikely(*p != end))
286 * parse extra results
288 static int parse_reply_info_extra(void **p, void *end,
289 struct ceph_mds_reply_info_parsed *info,
292 u32 op = le32_to_cpu(info->head->op);
294 if (op == CEPH_MDS_OP_GETFILELOCK)
295 return parse_reply_info_filelock(p, end, info, features);
296 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
297 return parse_reply_info_dir(p, end, info, features);
298 else if (op == CEPH_MDS_OP_CREATE)
299 return parse_reply_info_create(p, end, info, features);
305 * parse entire mds reply
307 static int parse_reply_info(struct ceph_msg *msg,
308 struct ceph_mds_reply_info_parsed *info,
315 info->head = msg->front.iov_base;
316 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
317 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
320 ceph_decode_32_safe(&p, end, len, bad);
322 ceph_decode_need(&p, end, len, bad);
323 err = parse_reply_info_trace(&p, p+len, info, features);
329 ceph_decode_32_safe(&p, end, len, bad);
331 ceph_decode_need(&p, end, len, bad);
332 err = parse_reply_info_extra(&p, p+len, info, features);
338 ceph_decode_32_safe(&p, end, len, bad);
339 info->snapblob_len = len;
350 pr_err("mds parse_reply err %d\n", err);
354 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
356 if (!info->dir_entries)
358 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
365 const char *ceph_session_state_name(int s)
368 case CEPH_MDS_SESSION_NEW: return "new";
369 case CEPH_MDS_SESSION_OPENING: return "opening";
370 case CEPH_MDS_SESSION_OPEN: return "open";
371 case CEPH_MDS_SESSION_HUNG: return "hung";
372 case CEPH_MDS_SESSION_CLOSING: return "closing";
373 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
374 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
375 case CEPH_MDS_SESSION_REJECTED: return "rejected";
376 default: return "???";
380 static struct ceph_mds_session *get_session(struct ceph_mds_session *s)
382 if (refcount_inc_not_zero(&s->s_ref)) {
383 dout("mdsc get_session %p %d -> %d\n", s,
384 refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
387 dout("mdsc get_session %p 0 -- FAIL", s);
392 void ceph_put_mds_session(struct ceph_mds_session *s)
394 dout("mdsc put_session %p %d -> %d\n", s,
395 refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
396 if (refcount_dec_and_test(&s->s_ref)) {
397 if (s->s_auth.authorizer)
398 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
404 * called under mdsc->mutex
406 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
409 struct ceph_mds_session *session;
411 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
413 session = mdsc->sessions[mds];
414 dout("lookup_mds_session %p %d\n", session,
415 refcount_read(&session->s_ref));
416 get_session(session);
420 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
422 if (mds >= mdsc->max_sessions)
424 return mdsc->sessions[mds];
427 static int __verify_registered_session(struct ceph_mds_client *mdsc,
428 struct ceph_mds_session *s)
430 if (s->s_mds >= mdsc->max_sessions ||
431 mdsc->sessions[s->s_mds] != s)
437 * create+register a new session for given mds.
438 * called under mdsc->mutex.
440 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
443 struct ceph_mds_session *s;
445 if (mds >= mdsc->mdsmap->m_num_mds)
446 return ERR_PTR(-EINVAL);
448 s = kzalloc(sizeof(*s), GFP_NOFS);
450 return ERR_PTR(-ENOMEM);
453 s->s_state = CEPH_MDS_SESSION_NEW;
456 mutex_init(&s->s_mutex);
458 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
460 spin_lock_init(&s->s_gen_ttl_lock);
462 s->s_cap_ttl = jiffies - 1;
464 spin_lock_init(&s->s_cap_lock);
465 s->s_renew_requested = 0;
467 INIT_LIST_HEAD(&s->s_caps);
470 refcount_set(&s->s_ref, 1);
471 INIT_LIST_HEAD(&s->s_waiting);
472 INIT_LIST_HEAD(&s->s_unsafe);
473 s->s_num_cap_releases = 0;
474 s->s_cap_reconnect = 0;
475 s->s_cap_iterator = NULL;
476 INIT_LIST_HEAD(&s->s_cap_releases);
477 INIT_LIST_HEAD(&s->s_cap_flushing);
479 dout("register_session mds%d\n", mds);
480 if (mds >= mdsc->max_sessions) {
481 int newmax = 1 << get_count_order(mds+1);
482 struct ceph_mds_session **sa;
484 dout("register_session realloc to %d\n", newmax);
485 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
488 if (mdsc->sessions) {
489 memcpy(sa, mdsc->sessions,
490 mdsc->max_sessions * sizeof(void *));
491 kfree(mdsc->sessions);
494 mdsc->max_sessions = newmax;
496 mdsc->sessions[mds] = s;
497 atomic_inc(&mdsc->num_sessions);
498 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
500 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
501 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
507 return ERR_PTR(-ENOMEM);
511 * called under mdsc->mutex
513 static void __unregister_session(struct ceph_mds_client *mdsc,
514 struct ceph_mds_session *s)
516 dout("__unregister_session mds%d %p\n", s->s_mds, s);
517 BUG_ON(mdsc->sessions[s->s_mds] != s);
518 mdsc->sessions[s->s_mds] = NULL;
519 ceph_con_close(&s->s_con);
520 ceph_put_mds_session(s);
521 atomic_dec(&mdsc->num_sessions);
525 * drop session refs in request.
527 * should be last request ref, or hold mdsc->mutex
529 static void put_request_session(struct ceph_mds_request *req)
531 if (req->r_session) {
532 ceph_put_mds_session(req->r_session);
533 req->r_session = NULL;
537 void ceph_mdsc_release_request(struct kref *kref)
539 struct ceph_mds_request *req = container_of(kref,
540 struct ceph_mds_request,
542 destroy_reply_info(&req->r_reply_info);
544 ceph_msg_put(req->r_request);
546 ceph_msg_put(req->r_reply);
548 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
552 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
553 iput(req->r_target_inode);
556 if (req->r_old_dentry)
557 dput(req->r_old_dentry);
558 if (req->r_old_dentry_dir) {
560 * track (and drop pins for) r_old_dentry_dir
561 * separately, since r_old_dentry's d_parent may have
562 * changed between the dir mutex being dropped and
563 * this request being freed.
565 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
567 iput(req->r_old_dentry_dir);
572 ceph_pagelist_release(req->r_pagelist);
573 put_request_session(req);
574 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
578 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
581 * lookup session, bump ref if found.
583 * called under mdsc->mutex.
585 static struct ceph_mds_request *
586 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
588 struct ceph_mds_request *req;
590 req = lookup_request(&mdsc->request_tree, tid);
592 ceph_mdsc_get_request(req);
598 * Register an in-flight request, and assign a tid. Link to directory
599 * are modifying (if any).
601 * Called under mdsc->mutex.
603 static void __register_request(struct ceph_mds_client *mdsc,
604 struct ceph_mds_request *req,
607 req->r_tid = ++mdsc->last_tid;
609 ceph_reserve_caps(mdsc, &req->r_caps_reservation,
611 dout("__register_request %p tid %lld\n", req, req->r_tid);
612 ceph_mdsc_get_request(req);
613 insert_request(&mdsc->request_tree, req);
615 req->r_uid = current_fsuid();
616 req->r_gid = current_fsgid();
618 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
619 mdsc->oldest_tid = req->r_tid;
623 req->r_unsafe_dir = dir;
627 static void __unregister_request(struct ceph_mds_client *mdsc,
628 struct ceph_mds_request *req)
630 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
632 /* Never leave an unregistered request on an unsafe list! */
633 list_del_init(&req->r_unsafe_item);
635 if (req->r_tid == mdsc->oldest_tid) {
636 struct rb_node *p = rb_next(&req->r_node);
637 mdsc->oldest_tid = 0;
639 struct ceph_mds_request *next_req =
640 rb_entry(p, struct ceph_mds_request, r_node);
641 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
642 mdsc->oldest_tid = next_req->r_tid;
649 erase_request(&mdsc->request_tree, req);
651 if (req->r_unsafe_dir &&
652 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
653 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
654 spin_lock(&ci->i_unsafe_lock);
655 list_del_init(&req->r_unsafe_dir_item);
656 spin_unlock(&ci->i_unsafe_lock);
658 if (req->r_target_inode &&
659 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
660 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
661 spin_lock(&ci->i_unsafe_lock);
662 list_del_init(&req->r_unsafe_target_item);
663 spin_unlock(&ci->i_unsafe_lock);
666 if (req->r_unsafe_dir) {
667 iput(req->r_unsafe_dir);
668 req->r_unsafe_dir = NULL;
671 complete_all(&req->r_safe_completion);
673 ceph_mdsc_put_request(req);
677 * Walk back up the dentry tree until we hit a dentry representing a
678 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
679 * when calling this) to ensure that the objects won't disappear while we're
680 * working with them. Once we hit a candidate dentry, we attempt to take a
681 * reference to it, and return that as the result.
683 static struct inode *get_nonsnap_parent(struct dentry *dentry)
685 struct inode *inode = NULL;
687 while (dentry && !IS_ROOT(dentry)) {
688 inode = d_inode_rcu(dentry);
689 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
691 dentry = dentry->d_parent;
694 inode = igrab(inode);
699 * Choose mds to send request to next. If there is a hint set in the
700 * request (e.g., due to a prior forward hint from the mds), use that.
701 * Otherwise, consult frag tree and/or caps to identify the
702 * appropriate mds. If all else fails, choose randomly.
704 * Called under mdsc->mutex.
706 static int __choose_mds(struct ceph_mds_client *mdsc,
707 struct ceph_mds_request *req)
710 struct ceph_inode_info *ci;
711 struct ceph_cap *cap;
712 int mode = req->r_direct_mode;
714 u32 hash = req->r_direct_hash;
715 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
718 * is there a specific mds we should try? ignore hint if we have
719 * no session and the mds is not up (active or recovering).
721 if (req->r_resend_mds >= 0 &&
722 (__have_session(mdsc, req->r_resend_mds) ||
723 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
724 dout("choose_mds using resend_mds mds%d\n",
726 return req->r_resend_mds;
729 if (mode == USE_RANDOM_MDS)
734 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
735 inode = req->r_inode;
738 /* req->r_dentry is non-null for LSSNAP request */
740 inode = get_nonsnap_parent(req->r_dentry);
742 dout("__choose_mds using snapdir's parent %p\n", inode);
744 } else if (req->r_dentry) {
745 /* ignore race with rename; old or new d_parent is okay */
746 struct dentry *parent;
750 parent = req->r_dentry->d_parent;
751 dir = req->r_parent ? : d_inode_rcu(parent);
753 if (!dir || dir->i_sb != mdsc->fsc->sb) {
754 /* not this fs or parent went negative */
755 inode = d_inode(req->r_dentry);
758 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
759 /* direct snapped/virtual snapdir requests
760 * based on parent dir inode */
761 inode = get_nonsnap_parent(parent);
762 dout("__choose_mds using nonsnap parent %p\n", inode);
765 inode = d_inode(req->r_dentry);
766 if (!inode || mode == USE_AUTH_MDS) {
769 hash = ceph_dentry_hash(dir, req->r_dentry);
778 dout("__choose_mds %p is_hash=%d (%d) mode %d\n", inode, (int)is_hash,
782 ci = ceph_inode(inode);
784 if (is_hash && S_ISDIR(inode->i_mode)) {
785 struct ceph_inode_frag frag;
788 ceph_choose_frag(ci, hash, &frag, &found);
790 if (mode == USE_ANY_MDS && frag.ndist > 0) {
793 /* choose a random replica */
794 get_random_bytes(&r, 1);
797 dout("choose_mds %p %llx.%llx "
798 "frag %u mds%d (%d/%d)\n",
799 inode, ceph_vinop(inode),
802 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
803 CEPH_MDS_STATE_ACTIVE)
807 /* since this file/dir wasn't known to be
808 * replicated, then we want to look for the
809 * authoritative mds. */
812 /* choose auth mds */
814 dout("choose_mds %p %llx.%llx "
815 "frag %u mds%d (auth)\n",
816 inode, ceph_vinop(inode), frag.frag, mds);
817 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
818 CEPH_MDS_STATE_ACTIVE)
824 spin_lock(&ci->i_ceph_lock);
826 if (mode == USE_AUTH_MDS)
827 cap = ci->i_auth_cap;
828 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
829 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
831 spin_unlock(&ci->i_ceph_lock);
835 mds = cap->session->s_mds;
836 dout("choose_mds %p %llx.%llx mds%d (%scap %p)\n",
837 inode, ceph_vinop(inode), mds,
838 cap == ci->i_auth_cap ? "auth " : "", cap);
839 spin_unlock(&ci->i_ceph_lock);
845 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
846 dout("choose_mds chose random mds%d\n", mds);
854 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
856 struct ceph_msg *msg;
857 struct ceph_mds_session_head *h;
859 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
862 pr_err("create_session_msg ENOMEM creating msg\n");
865 h = msg->front.iov_base;
866 h->op = cpu_to_le32(op);
867 h->seq = cpu_to_le64(seq);
873 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
874 * to include additional client metadata fields.
876 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
878 struct ceph_msg *msg;
879 struct ceph_mds_session_head *h;
881 int metadata_bytes = 0;
882 int metadata_key_count = 0;
883 struct ceph_options *opt = mdsc->fsc->client->options;
884 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
887 const char* metadata[][2] = {
888 {"hostname", mdsc->nodename},
889 {"kernel_version", init_utsname()->release},
890 {"entity_id", opt->name ? : ""},
891 {"root", fsopt->server_path ? : "/"},
895 /* Calculate serialized length of metadata */
896 metadata_bytes = 4; /* map length */
897 for (i = 0; metadata[i][0]; ++i) {
898 metadata_bytes += 8 + strlen(metadata[i][0]) +
899 strlen(metadata[i][1]);
900 metadata_key_count++;
903 /* Allocate the message */
904 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + metadata_bytes,
907 pr_err("create_session_msg ENOMEM creating msg\n");
910 h = msg->front.iov_base;
911 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
912 h->seq = cpu_to_le64(seq);
915 * Serialize client metadata into waiting buffer space, using
916 * the format that userspace expects for map<string, string>
918 * ClientSession messages with metadata are v2
920 msg->hdr.version = cpu_to_le16(2);
921 msg->hdr.compat_version = cpu_to_le16(1);
923 /* The write pointer, following the session_head structure */
924 p = msg->front.iov_base + sizeof(*h);
926 /* Number of entries in the map */
927 ceph_encode_32(&p, metadata_key_count);
929 /* Two length-prefixed strings for each entry in the map */
930 for (i = 0; metadata[i][0]; ++i) {
931 size_t const key_len = strlen(metadata[i][0]);
932 size_t const val_len = strlen(metadata[i][1]);
934 ceph_encode_32(&p, key_len);
935 memcpy(p, metadata[i][0], key_len);
937 ceph_encode_32(&p, val_len);
938 memcpy(p, metadata[i][1], val_len);
946 * send session open request.
948 * called under mdsc->mutex
950 static int __open_session(struct ceph_mds_client *mdsc,
951 struct ceph_mds_session *session)
953 struct ceph_msg *msg;
955 int mds = session->s_mds;
957 /* wait for mds to go active? */
958 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
959 dout("open_session to mds%d (%s)\n", mds,
960 ceph_mds_state_name(mstate));
961 session->s_state = CEPH_MDS_SESSION_OPENING;
962 session->s_renew_requested = jiffies;
964 /* send connect message */
965 msg = create_session_open_msg(mdsc, session->s_seq);
968 ceph_con_send(&session->s_con, msg);
973 * open sessions for any export targets for the given mds
975 * called under mdsc->mutex
977 static struct ceph_mds_session *
978 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
980 struct ceph_mds_session *session;
982 session = __ceph_lookup_mds_session(mdsc, target);
984 session = register_session(mdsc, target);
988 if (session->s_state == CEPH_MDS_SESSION_NEW ||
989 session->s_state == CEPH_MDS_SESSION_CLOSING)
990 __open_session(mdsc, session);
995 struct ceph_mds_session *
996 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
998 struct ceph_mds_session *session;
1000 dout("open_export_target_session to mds%d\n", target);
1002 mutex_lock(&mdsc->mutex);
1003 session = __open_export_target_session(mdsc, target);
1004 mutex_unlock(&mdsc->mutex);
1009 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1010 struct ceph_mds_session *session)
1012 struct ceph_mds_info *mi;
1013 struct ceph_mds_session *ts;
1014 int i, mds = session->s_mds;
1016 if (mds >= mdsc->mdsmap->m_num_mds)
1019 mi = &mdsc->mdsmap->m_info[mds];
1020 dout("open_export_target_sessions for mds%d (%d targets)\n",
1021 session->s_mds, mi->num_export_targets);
1023 for (i = 0; i < mi->num_export_targets; i++) {
1024 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1026 ceph_put_mds_session(ts);
1030 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1031 struct ceph_mds_session *session)
1033 mutex_lock(&mdsc->mutex);
1034 __open_export_target_sessions(mdsc, session);
1035 mutex_unlock(&mdsc->mutex);
1042 static void detach_cap_releases(struct ceph_mds_session *session,
1043 struct list_head *target)
1045 lockdep_assert_held(&session->s_cap_lock);
1047 list_splice_init(&session->s_cap_releases, target);
1048 session->s_num_cap_releases = 0;
1049 dout("dispose_cap_releases mds%d\n", session->s_mds);
1052 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1053 struct list_head *dispose)
1055 while (!list_empty(dispose)) {
1056 struct ceph_cap *cap;
1057 /* zero out the in-progress message */
1058 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1059 list_del(&cap->session_caps);
1060 ceph_put_cap(mdsc, cap);
1064 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1065 struct ceph_mds_session *session)
1067 struct ceph_mds_request *req;
1070 dout("cleanup_session_requests mds%d\n", session->s_mds);
1071 mutex_lock(&mdsc->mutex);
1072 while (!list_empty(&session->s_unsafe)) {
1073 req = list_first_entry(&session->s_unsafe,
1074 struct ceph_mds_request, r_unsafe_item);
1075 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1077 __unregister_request(mdsc, req);
1079 /* zero r_attempts, so kick_requests() will re-send requests */
1080 p = rb_first(&mdsc->request_tree);
1082 req = rb_entry(p, struct ceph_mds_request, r_node);
1084 if (req->r_session &&
1085 req->r_session->s_mds == session->s_mds)
1086 req->r_attempts = 0;
1088 mutex_unlock(&mdsc->mutex);
1092 * Helper to safely iterate over all caps associated with a session, with
1093 * special care taken to handle a racing __ceph_remove_cap().
1095 * Caller must hold session s_mutex.
1097 static int iterate_session_caps(struct ceph_mds_session *session,
1098 int (*cb)(struct inode *, struct ceph_cap *,
1101 struct list_head *p;
1102 struct ceph_cap *cap;
1103 struct inode *inode, *last_inode = NULL;
1104 struct ceph_cap *old_cap = NULL;
1107 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1108 spin_lock(&session->s_cap_lock);
1109 p = session->s_caps.next;
1110 while (p != &session->s_caps) {
1111 cap = list_entry(p, struct ceph_cap, session_caps);
1112 inode = igrab(&cap->ci->vfs_inode);
1117 session->s_cap_iterator = cap;
1118 spin_unlock(&session->s_cap_lock);
1125 ceph_put_cap(session->s_mdsc, old_cap);
1129 ret = cb(inode, cap, arg);
1132 spin_lock(&session->s_cap_lock);
1135 dout("iterate_session_caps finishing cap %p removal\n",
1137 BUG_ON(cap->session != session);
1138 cap->session = NULL;
1139 list_del_init(&cap->session_caps);
1140 session->s_nr_caps--;
1141 if (cap->queue_release) {
1142 list_add_tail(&cap->session_caps,
1143 &session->s_cap_releases);
1144 session->s_num_cap_releases++;
1146 old_cap = cap; /* put_cap it w/o locks held */
1154 session->s_cap_iterator = NULL;
1155 spin_unlock(&session->s_cap_lock);
1159 ceph_put_cap(session->s_mdsc, old_cap);
1164 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1167 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1168 struct ceph_inode_info *ci = ceph_inode(inode);
1169 LIST_HEAD(to_remove);
1171 bool invalidate = false;
1173 dout("removing cap %p, ci is %p, inode is %p\n",
1174 cap, ci, &ci->vfs_inode);
1175 spin_lock(&ci->i_ceph_lock);
1176 __ceph_remove_cap(cap, false);
1177 if (!ci->i_auth_cap) {
1178 struct ceph_cap_flush *cf;
1179 struct ceph_mds_client *mdsc = fsc->mdsc;
1181 ci->i_ceph_flags |= CEPH_I_CAP_DROPPED;
1183 if (ci->i_wrbuffer_ref > 0 &&
1184 READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
1187 while (!list_empty(&ci->i_cap_flush_list)) {
1188 cf = list_first_entry(&ci->i_cap_flush_list,
1189 struct ceph_cap_flush, i_list);
1190 list_move(&cf->i_list, &to_remove);
1193 spin_lock(&mdsc->cap_dirty_lock);
1195 list_for_each_entry(cf, &to_remove, i_list)
1196 list_del(&cf->g_list);
1198 if (!list_empty(&ci->i_dirty_item)) {
1199 pr_warn_ratelimited(
1200 " dropping dirty %s state for %p %lld\n",
1201 ceph_cap_string(ci->i_dirty_caps),
1202 inode, ceph_ino(inode));
1203 ci->i_dirty_caps = 0;
1204 list_del_init(&ci->i_dirty_item);
1207 if (!list_empty(&ci->i_flushing_item)) {
1208 pr_warn_ratelimited(
1209 " dropping dirty+flushing %s state for %p %lld\n",
1210 ceph_cap_string(ci->i_flushing_caps),
1211 inode, ceph_ino(inode));
1212 ci->i_flushing_caps = 0;
1213 list_del_init(&ci->i_flushing_item);
1214 mdsc->num_cap_flushing--;
1217 spin_unlock(&mdsc->cap_dirty_lock);
1219 if (atomic_read(&ci->i_filelock_ref) > 0) {
1220 /* make further file lock syscall return -EIO */
1221 ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1222 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1223 inode, ceph_ino(inode));
1226 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1227 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1228 ci->i_prealloc_cap_flush = NULL;
1231 spin_unlock(&ci->i_ceph_lock);
1232 while (!list_empty(&to_remove)) {
1233 struct ceph_cap_flush *cf;
1234 cf = list_first_entry(&to_remove,
1235 struct ceph_cap_flush, i_list);
1236 list_del(&cf->i_list);
1237 ceph_free_cap_flush(cf);
1240 wake_up_all(&ci->i_cap_wq);
1242 ceph_queue_invalidate(inode);
1249 * caller must hold session s_mutex
1251 static void remove_session_caps(struct ceph_mds_session *session)
1253 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1254 struct super_block *sb = fsc->sb;
1257 dout("remove_session_caps on %p\n", session);
1258 iterate_session_caps(session, remove_session_caps_cb, fsc);
1260 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1262 spin_lock(&session->s_cap_lock);
1263 if (session->s_nr_caps > 0) {
1264 struct inode *inode;
1265 struct ceph_cap *cap, *prev = NULL;
1266 struct ceph_vino vino;
1268 * iterate_session_caps() skips inodes that are being
1269 * deleted, we need to wait until deletions are complete.
1270 * __wait_on_freeing_inode() is designed for the job,
1271 * but it is not exported, so use lookup inode function
1274 while (!list_empty(&session->s_caps)) {
1275 cap = list_entry(session->s_caps.next,
1276 struct ceph_cap, session_caps);
1280 vino = cap->ci->i_vino;
1281 spin_unlock(&session->s_cap_lock);
1283 inode = ceph_find_inode(sb, vino);
1286 spin_lock(&session->s_cap_lock);
1290 // drop cap expires and unlock s_cap_lock
1291 detach_cap_releases(session, &dispose);
1293 BUG_ON(session->s_nr_caps > 0);
1294 BUG_ON(!list_empty(&session->s_cap_flushing));
1295 spin_unlock(&session->s_cap_lock);
1296 dispose_cap_releases(session->s_mdsc, &dispose);
1300 * wake up any threads waiting on this session's caps. if the cap is
1301 * old (didn't get renewed on the client reconnect), remove it now.
1303 * caller must hold s_mutex.
1305 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1308 struct ceph_inode_info *ci = ceph_inode(inode);
1311 spin_lock(&ci->i_ceph_lock);
1312 ci->i_wanted_max_size = 0;
1313 ci->i_requested_max_size = 0;
1314 spin_unlock(&ci->i_ceph_lock);
1316 wake_up_all(&ci->i_cap_wq);
1320 static void wake_up_session_caps(struct ceph_mds_session *session,
1323 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1324 iterate_session_caps(session, wake_up_session_cb,
1325 (void *)(unsigned long)reconnect);
1329 * Send periodic message to MDS renewing all currently held caps. The
1330 * ack will reset the expiration for all caps from this session.
1332 * caller holds s_mutex
1334 static int send_renew_caps(struct ceph_mds_client *mdsc,
1335 struct ceph_mds_session *session)
1337 struct ceph_msg *msg;
1340 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1341 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1342 pr_info("mds%d caps stale\n", session->s_mds);
1343 session->s_renew_requested = jiffies;
1345 /* do not try to renew caps until a recovering mds has reconnected
1346 * with its clients. */
1347 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1348 if (state < CEPH_MDS_STATE_RECONNECT) {
1349 dout("send_renew_caps ignoring mds%d (%s)\n",
1350 session->s_mds, ceph_mds_state_name(state));
1354 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1355 ceph_mds_state_name(state));
1356 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1357 ++session->s_renew_seq);
1360 ceph_con_send(&session->s_con, msg);
1364 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1365 struct ceph_mds_session *session, u64 seq)
1367 struct ceph_msg *msg;
1369 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1370 session->s_mds, ceph_session_state_name(session->s_state), seq);
1371 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1374 ceph_con_send(&session->s_con, msg);
1380 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1382 * Called under session->s_mutex
1384 static void renewed_caps(struct ceph_mds_client *mdsc,
1385 struct ceph_mds_session *session, int is_renew)
1390 spin_lock(&session->s_cap_lock);
1391 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1393 session->s_cap_ttl = session->s_renew_requested +
1394 mdsc->mdsmap->m_session_timeout*HZ;
1397 if (time_before(jiffies, session->s_cap_ttl)) {
1398 pr_info("mds%d caps renewed\n", session->s_mds);
1401 pr_info("mds%d caps still stale\n", session->s_mds);
1404 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1405 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1406 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1407 spin_unlock(&session->s_cap_lock);
1410 wake_up_session_caps(session, 0);
1414 * send a session close request
1416 static int request_close_session(struct ceph_mds_client *mdsc,
1417 struct ceph_mds_session *session)
1419 struct ceph_msg *msg;
1421 dout("request_close_session mds%d state %s seq %lld\n",
1422 session->s_mds, ceph_session_state_name(session->s_state),
1424 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1427 ceph_con_send(&session->s_con, msg);
1432 * Called with s_mutex held.
1434 static int __close_session(struct ceph_mds_client *mdsc,
1435 struct ceph_mds_session *session)
1437 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1439 session->s_state = CEPH_MDS_SESSION_CLOSING;
1440 return request_close_session(mdsc, session);
1444 * Trim old(er) caps.
1446 * Because we can't cache an inode without one or more caps, we do
1447 * this indirectly: if a cap is unused, we prune its aliases, at which
1448 * point the inode will hopefully get dropped to.
1450 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1451 * memory pressure from the MDS, though, so it needn't be perfect.
1453 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1455 struct ceph_mds_session *session = arg;
1456 struct ceph_inode_info *ci = ceph_inode(inode);
1457 int used, wanted, oissued, mine;
1459 if (session->s_trim_caps <= 0)
1462 spin_lock(&ci->i_ceph_lock);
1463 mine = cap->issued | cap->implemented;
1464 used = __ceph_caps_used(ci);
1465 wanted = __ceph_caps_file_wanted(ci);
1466 oissued = __ceph_caps_issued_other(ci, cap);
1468 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1469 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1470 ceph_cap_string(used), ceph_cap_string(wanted));
1471 if (cap == ci->i_auth_cap) {
1472 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1473 !list_empty(&ci->i_cap_snaps))
1475 if ((used | wanted) & CEPH_CAP_ANY_WR)
1477 /* Note: it's possible that i_filelock_ref becomes non-zero
1478 * after dropping auth caps. It doesn't hurt because reply
1479 * of lock mds request will re-add auth caps. */
1480 if (atomic_read(&ci->i_filelock_ref) > 0)
1483 /* The inode has cached pages, but it's no longer used.
1484 * we can safely drop it */
1485 if (wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1486 !(oissued & CEPH_CAP_FILE_CACHE)) {
1490 if ((used | wanted) & ~oissued & mine)
1491 goto out; /* we need these caps */
1493 session->s_trim_caps--;
1495 /* we aren't the only cap.. just remove us */
1496 __ceph_remove_cap(cap, true);
1498 /* try dropping referring dentries */
1499 spin_unlock(&ci->i_ceph_lock);
1500 d_prune_aliases(inode);
1501 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1502 inode, cap, atomic_read(&inode->i_count));
1507 spin_unlock(&ci->i_ceph_lock);
1512 * Trim session cap count down to some max number.
1514 static int trim_caps(struct ceph_mds_client *mdsc,
1515 struct ceph_mds_session *session,
1518 int trim_caps = session->s_nr_caps - max_caps;
1520 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1521 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1522 if (trim_caps > 0) {
1523 session->s_trim_caps = trim_caps;
1524 iterate_session_caps(session, trim_caps_cb, session);
1525 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1526 session->s_mds, session->s_nr_caps, max_caps,
1527 trim_caps - session->s_trim_caps);
1528 session->s_trim_caps = 0;
1531 ceph_send_cap_releases(mdsc, session);
1535 static int check_caps_flush(struct ceph_mds_client *mdsc,
1540 spin_lock(&mdsc->cap_dirty_lock);
1541 if (!list_empty(&mdsc->cap_flush_list)) {
1542 struct ceph_cap_flush *cf =
1543 list_first_entry(&mdsc->cap_flush_list,
1544 struct ceph_cap_flush, g_list);
1545 if (cf->tid <= want_flush_tid) {
1546 dout("check_caps_flush still flushing tid "
1547 "%llu <= %llu\n", cf->tid, want_flush_tid);
1551 spin_unlock(&mdsc->cap_dirty_lock);
1556 * flush all dirty inode data to disk.
1558 * returns true if we've flushed through want_flush_tid
1560 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1563 dout("check_caps_flush want %llu\n", want_flush_tid);
1565 wait_event(mdsc->cap_flushing_wq,
1566 check_caps_flush(mdsc, want_flush_tid));
1568 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1572 * called under s_mutex
1574 void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1575 struct ceph_mds_session *session)
1577 struct ceph_msg *msg = NULL;
1578 struct ceph_mds_cap_release *head;
1579 struct ceph_mds_cap_item *item;
1580 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1581 struct ceph_cap *cap;
1582 LIST_HEAD(tmp_list);
1583 int num_cap_releases;
1584 __le32 barrier, *cap_barrier;
1586 down_read(&osdc->lock);
1587 barrier = cpu_to_le32(osdc->epoch_barrier);
1588 up_read(&osdc->lock);
1590 spin_lock(&session->s_cap_lock);
1592 list_splice_init(&session->s_cap_releases, &tmp_list);
1593 num_cap_releases = session->s_num_cap_releases;
1594 session->s_num_cap_releases = 0;
1595 spin_unlock(&session->s_cap_lock);
1597 while (!list_empty(&tmp_list)) {
1599 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1600 PAGE_SIZE, GFP_NOFS, false);
1603 head = msg->front.iov_base;
1604 head->num = cpu_to_le32(0);
1605 msg->front.iov_len = sizeof(*head);
1607 msg->hdr.version = cpu_to_le16(2);
1608 msg->hdr.compat_version = cpu_to_le16(1);
1611 cap = list_first_entry(&tmp_list, struct ceph_cap,
1613 list_del(&cap->session_caps);
1616 head = msg->front.iov_base;
1617 le32_add_cpu(&head->num, 1);
1618 item = msg->front.iov_base + msg->front.iov_len;
1619 item->ino = cpu_to_le64(cap->cap_ino);
1620 item->cap_id = cpu_to_le64(cap->cap_id);
1621 item->migrate_seq = cpu_to_le32(cap->mseq);
1622 item->seq = cpu_to_le32(cap->issue_seq);
1623 msg->front.iov_len += sizeof(*item);
1625 ceph_put_cap(mdsc, cap);
1627 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1628 // Append cap_barrier field
1629 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1630 *cap_barrier = barrier;
1631 msg->front.iov_len += sizeof(*cap_barrier);
1633 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1634 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1635 ceph_con_send(&session->s_con, msg);
1640 BUG_ON(num_cap_releases != 0);
1642 spin_lock(&session->s_cap_lock);
1643 if (!list_empty(&session->s_cap_releases))
1645 spin_unlock(&session->s_cap_lock);
1648 // Append cap_barrier field
1649 cap_barrier = msg->front.iov_base + msg->front.iov_len;
1650 *cap_barrier = barrier;
1651 msg->front.iov_len += sizeof(*cap_barrier);
1653 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1654 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1655 ceph_con_send(&session->s_con, msg);
1659 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1661 spin_lock(&session->s_cap_lock);
1662 list_splice(&tmp_list, &session->s_cap_releases);
1663 session->s_num_cap_releases += num_cap_releases;
1664 spin_unlock(&session->s_cap_lock);
1671 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1674 struct ceph_inode_info *ci = ceph_inode(dir);
1675 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
1676 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
1677 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
1678 int order, num_entries;
1680 spin_lock(&ci->i_ceph_lock);
1681 num_entries = ci->i_files + ci->i_subdirs;
1682 spin_unlock(&ci->i_ceph_lock);
1683 num_entries = max(num_entries, 1);
1684 num_entries = min(num_entries, opt->max_readdir);
1686 order = get_order(size * num_entries);
1687 while (order >= 0) {
1688 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
1691 if (rinfo->dir_entries)
1695 if (!rinfo->dir_entries)
1698 num_entries = (PAGE_SIZE << order) / size;
1699 num_entries = min(num_entries, opt->max_readdir);
1701 rinfo->dir_buf_size = PAGE_SIZE << order;
1702 req->r_num_caps = num_entries + 1;
1703 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
1704 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
1709 * Create an mds request.
1711 struct ceph_mds_request *
1712 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
1714 struct ceph_mds_request *req = kzalloc(sizeof(*req), GFP_NOFS);
1717 return ERR_PTR(-ENOMEM);
1719 mutex_init(&req->r_fill_mutex);
1721 req->r_started = jiffies;
1722 req->r_resend_mds = -1;
1723 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
1724 INIT_LIST_HEAD(&req->r_unsafe_target_item);
1726 kref_init(&req->r_kref);
1727 RB_CLEAR_NODE(&req->r_node);
1728 INIT_LIST_HEAD(&req->r_wait);
1729 init_completion(&req->r_completion);
1730 init_completion(&req->r_safe_completion);
1731 INIT_LIST_HEAD(&req->r_unsafe_item);
1733 req->r_stamp = timespec_trunc(current_kernel_time(), mdsc->fsc->sb->s_time_gran);
1736 req->r_direct_mode = mode;
1741 * return oldest (lowest) request, tid in request tree, 0 if none.
1743 * called under mdsc->mutex.
1745 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1747 if (RB_EMPTY_ROOT(&mdsc->request_tree))
1749 return rb_entry(rb_first(&mdsc->request_tree),
1750 struct ceph_mds_request, r_node);
1753 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1755 return mdsc->oldest_tid;
1759 * Build a dentry's path. Allocate on heap; caller must kfree. Based
1760 * on build_path_from_dentry in fs/cifs/dir.c.
1762 * If @stop_on_nosnap, generate path relative to the first non-snapped
1765 * Encode hidden .snap dirs as a double /, i.e.
1766 * foo/.snap/bar -> foo//bar
1768 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *base,
1771 struct dentry *temp;
1777 return ERR_PTR(-EINVAL);
1781 seq = read_seqbegin(&rename_lock);
1783 for (temp = dentry; !IS_ROOT(temp);) {
1784 struct inode *inode = d_inode(temp);
1785 if (inode && ceph_snap(inode) == CEPH_SNAPDIR)
1786 len++; /* slash only */
1787 else if (stop_on_nosnap && inode &&
1788 ceph_snap(inode) == CEPH_NOSNAP)
1791 len += 1 + temp->d_name.len;
1792 temp = temp->d_parent;
1796 len--; /* no leading '/' */
1798 path = kmalloc(len+1, GFP_NOFS);
1800 return ERR_PTR(-ENOMEM);
1802 path[pos] = 0; /* trailing null */
1804 for (temp = dentry; !IS_ROOT(temp) && pos != 0; ) {
1805 struct inode *inode;
1807 spin_lock(&temp->d_lock);
1808 inode = d_inode(temp);
1809 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1810 dout("build_path path+%d: %p SNAPDIR\n",
1812 } else if (stop_on_nosnap && inode &&
1813 ceph_snap(inode) == CEPH_NOSNAP) {
1814 spin_unlock(&temp->d_lock);
1817 pos -= temp->d_name.len;
1819 spin_unlock(&temp->d_lock);
1822 strncpy(path + pos, temp->d_name.name,
1825 spin_unlock(&temp->d_lock);
1828 temp = temp->d_parent;
1831 if (pos != 0 || read_seqretry(&rename_lock, seq)) {
1832 pr_err("build_path did not end path lookup where "
1833 "expected, namelen is %d, pos is %d\n", len, pos);
1834 /* presumably this is only possible if racing with a
1835 rename of one of the parent directories (we can not
1836 lock the dentries above us to prevent this, but
1837 retrying should be harmless) */
1842 *base = ceph_ino(d_inode(temp));
1844 dout("build_path on %p %d built %llx '%.*s'\n",
1845 dentry, d_count(dentry), *base, len, path);
1849 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
1850 const char **ppath, int *ppathlen, u64 *pino,
1857 dir = d_inode_rcu(dentry->d_parent);
1858 if (dir && ceph_snap(dir) == CEPH_NOSNAP) {
1859 *pino = ceph_ino(dir);
1861 *ppath = dentry->d_name.name;
1862 *ppathlen = dentry->d_name.len;
1866 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1868 return PTR_ERR(path);
1874 static int build_inode_path(struct inode *inode,
1875 const char **ppath, int *ppathlen, u64 *pino,
1878 struct dentry *dentry;
1881 if (ceph_snap(inode) == CEPH_NOSNAP) {
1882 *pino = ceph_ino(inode);
1886 dentry = d_find_alias(inode);
1887 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
1890 return PTR_ERR(path);
1897 * request arguments may be specified via an inode *, a dentry *, or
1898 * an explicit ino+path.
1900 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
1901 struct inode *rdiri, const char *rpath,
1902 u64 rino, const char **ppath, int *pathlen,
1903 u64 *ino, int *freepath)
1908 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
1909 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
1911 } else if (rdentry) {
1912 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
1914 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
1916 } else if (rpath || rino) {
1919 *pathlen = rpath ? strlen(rpath) : 0;
1920 dout(" path %.*s\n", *pathlen, rpath);
1927 * called under mdsc->mutex
1929 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
1930 struct ceph_mds_request *req,
1931 int mds, bool drop_cap_releases)
1933 struct ceph_msg *msg;
1934 struct ceph_mds_request_head *head;
1935 const char *path1 = NULL;
1936 const char *path2 = NULL;
1937 u64 ino1 = 0, ino2 = 0;
1938 int pathlen1 = 0, pathlen2 = 0;
1939 int freepath1 = 0, freepath2 = 0;
1945 ret = set_request_path_attr(req->r_inode, req->r_dentry,
1946 req->r_parent, req->r_path1, req->r_ino1.ino,
1947 &path1, &pathlen1, &ino1, &freepath1);
1953 ret = set_request_path_attr(NULL, req->r_old_dentry,
1954 req->r_old_dentry_dir,
1955 req->r_path2, req->r_ino2.ino,
1956 &path2, &pathlen2, &ino2, &freepath2);
1962 len = sizeof(*head) +
1963 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
1964 sizeof(struct ceph_timespec);
1966 /* calculate (max) length for cap releases */
1967 len += sizeof(struct ceph_mds_request_release) *
1968 (!!req->r_inode_drop + !!req->r_dentry_drop +
1969 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
1970 if (req->r_dentry_drop)
1971 len += req->r_dentry->d_name.len;
1972 if (req->r_old_dentry_drop)
1973 len += req->r_old_dentry->d_name.len;
1975 msg = ceph_msg_new(CEPH_MSG_CLIENT_REQUEST, len, GFP_NOFS, false);
1977 msg = ERR_PTR(-ENOMEM);
1981 msg->hdr.version = cpu_to_le16(2);
1982 msg->hdr.tid = cpu_to_le64(req->r_tid);
1984 head = msg->front.iov_base;
1985 p = msg->front.iov_base + sizeof(*head);
1986 end = msg->front.iov_base + msg->front.iov_len;
1988 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
1989 head->op = cpu_to_le32(req->r_op);
1990 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
1991 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
1992 head->args = req->r_args;
1994 ceph_encode_filepath(&p, end, ino1, path1);
1995 ceph_encode_filepath(&p, end, ino2, path2);
1997 /* make note of release offset, in case we need to replay */
1998 req->r_request_release_offset = p - msg->front.iov_base;
2002 if (req->r_inode_drop)
2003 releases += ceph_encode_inode_release(&p,
2004 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2005 mds, req->r_inode_drop, req->r_inode_unless, 0);
2006 if (req->r_dentry_drop)
2007 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2008 req->r_parent, mds, req->r_dentry_drop,
2009 req->r_dentry_unless);
2010 if (req->r_old_dentry_drop)
2011 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2012 req->r_old_dentry_dir, mds,
2013 req->r_old_dentry_drop,
2014 req->r_old_dentry_unless);
2015 if (req->r_old_inode_drop)
2016 releases += ceph_encode_inode_release(&p,
2017 d_inode(req->r_old_dentry),
2018 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2020 if (drop_cap_releases) {
2022 p = msg->front.iov_base + req->r_request_release_offset;
2025 head->num_releases = cpu_to_le16(releases);
2029 struct ceph_timespec ts;
2030 ceph_encode_timespec(&ts, &req->r_stamp);
2031 ceph_encode_copy(&p, &ts, sizeof(ts));
2035 msg->front.iov_len = p - msg->front.iov_base;
2036 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2038 if (req->r_pagelist) {
2039 struct ceph_pagelist *pagelist = req->r_pagelist;
2040 refcount_inc(&pagelist->refcnt);
2041 ceph_msg_data_add_pagelist(msg, pagelist);
2042 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2044 msg->hdr.data_len = 0;
2047 msg->hdr.data_off = cpu_to_le16(0);
2051 kfree((char *)path2);
2054 kfree((char *)path1);
2060 * called under mdsc->mutex if error, under no mutex if
2063 static void complete_request(struct ceph_mds_client *mdsc,
2064 struct ceph_mds_request *req)
2066 if (req->r_callback)
2067 req->r_callback(mdsc, req);
2069 complete_all(&req->r_completion);
2073 * called under mdsc->mutex
2075 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2076 struct ceph_mds_request *req,
2077 int mds, bool drop_cap_releases)
2079 struct ceph_mds_request_head *rhead;
2080 struct ceph_msg *msg;
2085 struct ceph_cap *cap =
2086 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2089 req->r_sent_on_mseq = cap->mseq;
2091 req->r_sent_on_mseq = -1;
2093 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2094 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2096 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2099 * Replay. Do not regenerate message (and rebuild
2100 * paths, etc.); just use the original message.
2101 * Rebuilding paths will break for renames because
2102 * d_move mangles the src name.
2104 msg = req->r_request;
2105 rhead = msg->front.iov_base;
2107 flags = le32_to_cpu(rhead->flags);
2108 flags |= CEPH_MDS_FLAG_REPLAY;
2109 rhead->flags = cpu_to_le32(flags);
2111 if (req->r_target_inode)
2112 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2114 rhead->num_retry = req->r_attempts - 1;
2116 /* remove cap/dentry releases from message */
2117 rhead->num_releases = 0;
2120 p = msg->front.iov_base + req->r_request_release_offset;
2122 struct ceph_timespec ts;
2123 ceph_encode_timespec(&ts, &req->r_stamp);
2124 ceph_encode_copy(&p, &ts, sizeof(ts));
2127 msg->front.iov_len = p - msg->front.iov_base;
2128 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2132 if (req->r_request) {
2133 ceph_msg_put(req->r_request);
2134 req->r_request = NULL;
2136 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2138 req->r_err = PTR_ERR(msg);
2139 return PTR_ERR(msg);
2141 req->r_request = msg;
2143 rhead = msg->front.iov_base;
2144 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2145 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2146 flags |= CEPH_MDS_FLAG_REPLAY;
2148 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2149 rhead->flags = cpu_to_le32(flags);
2150 rhead->num_fwd = req->r_num_fwd;
2151 rhead->num_retry = req->r_attempts - 1;
2154 dout(" r_parent = %p\n", req->r_parent);
2159 * send request, or put it on the appropriate wait list.
2161 static int __do_request(struct ceph_mds_client *mdsc,
2162 struct ceph_mds_request *req)
2164 struct ceph_mds_session *session = NULL;
2168 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2169 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2170 __unregister_request(mdsc, req);
2174 if (req->r_timeout &&
2175 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2176 dout("do_request timed out\n");
2180 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2181 dout("do_request forced umount\n");
2185 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2186 if (mdsc->mdsmap_err) {
2187 err = mdsc->mdsmap_err;
2188 dout("do_request mdsmap err %d\n", err);
2191 if (mdsc->mdsmap->m_epoch == 0) {
2192 dout("do_request no mdsmap, waiting for map\n");
2193 list_add(&req->r_wait, &mdsc->waiting_for_map);
2196 if (!(mdsc->fsc->mount_options->flags &
2197 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2198 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2200 pr_info("probably no mds server is up\n");
2205 put_request_session(req);
2207 mds = __choose_mds(mdsc, req);
2209 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2210 dout("do_request no mds or not active, waiting for map\n");
2211 list_add(&req->r_wait, &mdsc->waiting_for_map);
2215 /* get, open session */
2216 session = __ceph_lookup_mds_session(mdsc, mds);
2218 session = register_session(mdsc, mds);
2219 if (IS_ERR(session)) {
2220 err = PTR_ERR(session);
2224 req->r_session = get_session(session);
2226 dout("do_request mds%d session %p state %s\n", mds, session,
2227 ceph_session_state_name(session->s_state));
2228 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2229 session->s_state != CEPH_MDS_SESSION_HUNG) {
2230 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2234 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2235 session->s_state == CEPH_MDS_SESSION_CLOSING)
2236 __open_session(mdsc, session);
2237 list_add(&req->r_wait, &session->s_waiting);
2242 req->r_resend_mds = -1; /* forget any previous mds hint */
2244 if (req->r_request_started == 0) /* note request start time */
2245 req->r_request_started = jiffies;
2247 err = __prepare_send_request(mdsc, req, mds, false);
2249 ceph_msg_get(req->r_request);
2250 ceph_con_send(&session->s_con, req->r_request);
2254 ceph_put_mds_session(session);
2257 dout("__do_request early error %d\n", err);
2259 complete_request(mdsc, req);
2260 __unregister_request(mdsc, req);
2267 * called under mdsc->mutex
2269 static void __wake_requests(struct ceph_mds_client *mdsc,
2270 struct list_head *head)
2272 struct ceph_mds_request *req;
2273 LIST_HEAD(tmp_list);
2275 list_splice_init(head, &tmp_list);
2277 while (!list_empty(&tmp_list)) {
2278 req = list_entry(tmp_list.next,
2279 struct ceph_mds_request, r_wait);
2280 list_del_init(&req->r_wait);
2281 dout(" wake request %p tid %llu\n", req, req->r_tid);
2282 __do_request(mdsc, req);
2287 * Wake up threads with requests pending for @mds, so that they can
2288 * resubmit their requests to a possibly different mds.
2290 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2292 struct ceph_mds_request *req;
2293 struct rb_node *p = rb_first(&mdsc->request_tree);
2295 dout("kick_requests mds%d\n", mds);
2297 req = rb_entry(p, struct ceph_mds_request, r_node);
2299 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2301 if (req->r_attempts > 0)
2302 continue; /* only new requests */
2303 if (req->r_session &&
2304 req->r_session->s_mds == mds) {
2305 dout(" kicking tid %llu\n", req->r_tid);
2306 list_del_init(&req->r_wait);
2307 __do_request(mdsc, req);
2312 void ceph_mdsc_submit_request(struct ceph_mds_client *mdsc,
2313 struct ceph_mds_request *req)
2315 dout("submit_request on %p\n", req);
2316 mutex_lock(&mdsc->mutex);
2317 __register_request(mdsc, req, NULL);
2318 __do_request(mdsc, req);
2319 mutex_unlock(&mdsc->mutex);
2323 * Synchrously perform an mds request. Take care of all of the
2324 * session setup, forwarding, retry details.
2326 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2328 struct ceph_mds_request *req)
2332 dout("do_request on %p\n", req);
2334 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2336 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2338 ceph_get_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
2339 if (req->r_old_dentry_dir)
2340 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2344 mutex_lock(&mdsc->mutex);
2345 __register_request(mdsc, req, dir);
2346 __do_request(mdsc, req);
2354 mutex_unlock(&mdsc->mutex);
2355 dout("do_request waiting\n");
2356 if (!req->r_timeout && req->r_wait_for_completion) {
2357 err = req->r_wait_for_completion(mdsc, req);
2359 long timeleft = wait_for_completion_killable_timeout(
2361 ceph_timeout_jiffies(req->r_timeout));
2365 err = -EIO; /* timed out */
2367 err = timeleft; /* killed */
2369 dout("do_request waited, got %d\n", err);
2370 mutex_lock(&mdsc->mutex);
2372 /* only abort if we didn't race with a real reply */
2373 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2374 err = le32_to_cpu(req->r_reply_info.head->result);
2375 } else if (err < 0) {
2376 dout("aborted request %lld with %d\n", req->r_tid, err);
2379 * ensure we aren't running concurrently with
2380 * ceph_fill_trace or ceph_readdir_prepopulate, which
2381 * rely on locks (dir mutex) held by our caller.
2383 mutex_lock(&req->r_fill_mutex);
2385 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2386 mutex_unlock(&req->r_fill_mutex);
2388 if (req->r_parent &&
2389 (req->r_op & CEPH_MDS_OP_WRITE))
2390 ceph_invalidate_dir_request(req);
2396 mutex_unlock(&mdsc->mutex);
2397 dout("do_request %p done, result %d\n", req, err);
2402 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2403 * namespace request.
2405 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2407 struct inode *inode = req->r_parent;
2409 dout("invalidate_dir_request %p (complete, lease(s))\n", inode);
2411 ceph_dir_clear_complete(inode);
2413 ceph_invalidate_dentry_lease(req->r_dentry);
2414 if (req->r_old_dentry)
2415 ceph_invalidate_dentry_lease(req->r_old_dentry);
2421 * We take the session mutex and parse and process the reply immediately.
2422 * This preserves the logical ordering of replies, capabilities, etc., sent
2423 * by the MDS as they are applied to our local cache.
2425 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2427 struct ceph_mds_client *mdsc = session->s_mdsc;
2428 struct ceph_mds_request *req;
2429 struct ceph_mds_reply_head *head = msg->front.iov_base;
2430 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2431 struct ceph_snap_realm *realm;
2434 int mds = session->s_mds;
2436 if (msg->front.iov_len < sizeof(*head)) {
2437 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2442 /* get request, session */
2443 tid = le64_to_cpu(msg->hdr.tid);
2444 mutex_lock(&mdsc->mutex);
2445 req = lookup_get_request(mdsc, tid);
2447 dout("handle_reply on unknown tid %llu\n", tid);
2448 mutex_unlock(&mdsc->mutex);
2451 dout("handle_reply %p\n", req);
2453 /* correct session? */
2454 if (req->r_session != session) {
2455 pr_err("mdsc_handle_reply got %llu on session mds%d"
2456 " not mds%d\n", tid, session->s_mds,
2457 req->r_session ? req->r_session->s_mds : -1);
2458 mutex_unlock(&mdsc->mutex);
2463 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
2464 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
2465 pr_warn("got a dup %s reply on %llu from mds%d\n",
2466 head->safe ? "safe" : "unsafe", tid, mds);
2467 mutex_unlock(&mdsc->mutex);
2470 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
2471 pr_warn("got unsafe after safe on %llu from mds%d\n",
2473 mutex_unlock(&mdsc->mutex);
2477 result = le32_to_cpu(head->result);
2481 * if we're not talking to the authority, send to them
2482 * if the authority has changed while we weren't looking,
2483 * send to new authority
2484 * Otherwise we just have to return an ESTALE
2486 if (result == -ESTALE) {
2487 dout("got ESTALE on request %llu", req->r_tid);
2488 req->r_resend_mds = -1;
2489 if (req->r_direct_mode != USE_AUTH_MDS) {
2490 dout("not using auth, setting for that now");
2491 req->r_direct_mode = USE_AUTH_MDS;
2492 __do_request(mdsc, req);
2493 mutex_unlock(&mdsc->mutex);
2496 int mds = __choose_mds(mdsc, req);
2497 if (mds >= 0 && mds != req->r_session->s_mds) {
2498 dout("but auth changed, so resending");
2499 __do_request(mdsc, req);
2500 mutex_unlock(&mdsc->mutex);
2504 dout("have to return ESTALE on request %llu", req->r_tid);
2509 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
2510 __unregister_request(mdsc, req);
2512 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2514 * We already handled the unsafe response, now do the
2515 * cleanup. No need to examine the response; the MDS
2516 * doesn't include any result info in the safe
2517 * response. And even if it did, there is nothing
2518 * useful we could do with a revised return value.
2520 dout("got safe reply %llu, mds%d\n", tid, mds);
2522 /* last unsafe request during umount? */
2523 if (mdsc->stopping && !__get_oldest_req(mdsc))
2524 complete_all(&mdsc->safe_umount_waiters);
2525 mutex_unlock(&mdsc->mutex);
2529 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
2530 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
2531 if (req->r_unsafe_dir) {
2532 struct ceph_inode_info *ci =
2533 ceph_inode(req->r_unsafe_dir);
2534 spin_lock(&ci->i_unsafe_lock);
2535 list_add_tail(&req->r_unsafe_dir_item,
2536 &ci->i_unsafe_dirops);
2537 spin_unlock(&ci->i_unsafe_lock);
2541 dout("handle_reply tid %lld result %d\n", tid, result);
2542 rinfo = &req->r_reply_info;
2543 err = parse_reply_info(msg, rinfo, session->s_con.peer_features);
2544 mutex_unlock(&mdsc->mutex);
2546 mutex_lock(&session->s_mutex);
2548 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
2555 if (rinfo->snapblob_len) {
2556 down_write(&mdsc->snap_rwsem);
2557 ceph_update_snap_trace(mdsc, rinfo->snapblob,
2558 rinfo->snapblob + rinfo->snapblob_len,
2559 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
2561 downgrade_write(&mdsc->snap_rwsem);
2563 down_read(&mdsc->snap_rwsem);
2566 /* insert trace into our cache */
2567 mutex_lock(&req->r_fill_mutex);
2568 current->journal_info = req;
2569 err = ceph_fill_trace(mdsc->fsc->sb, req);
2571 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
2572 req->r_op == CEPH_MDS_OP_LSSNAP))
2573 ceph_readdir_prepopulate(req, req->r_session);
2574 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
2576 current->journal_info = NULL;
2577 mutex_unlock(&req->r_fill_mutex);
2579 up_read(&mdsc->snap_rwsem);
2581 ceph_put_snap_realm(mdsc, realm);
2583 if (err == 0 && req->r_target_inode &&
2584 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2585 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
2586 spin_lock(&ci->i_unsafe_lock);
2587 list_add_tail(&req->r_unsafe_target_item, &ci->i_unsafe_iops);
2588 spin_unlock(&ci->i_unsafe_lock);
2591 mutex_lock(&mdsc->mutex);
2592 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2596 req->r_reply = ceph_msg_get(msg);
2597 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
2600 dout("reply arrived after request %lld was aborted\n", tid);
2602 mutex_unlock(&mdsc->mutex);
2604 mutex_unlock(&session->s_mutex);
2606 /* kick calling process */
2607 complete_request(mdsc, req);
2609 ceph_mdsc_put_request(req);
2616 * handle mds notification that our request has been forwarded.
2618 static void handle_forward(struct ceph_mds_client *mdsc,
2619 struct ceph_mds_session *session,
2620 struct ceph_msg *msg)
2622 struct ceph_mds_request *req;
2623 u64 tid = le64_to_cpu(msg->hdr.tid);
2627 void *p = msg->front.iov_base;
2628 void *end = p + msg->front.iov_len;
2630 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
2631 next_mds = ceph_decode_32(&p);
2632 fwd_seq = ceph_decode_32(&p);
2634 mutex_lock(&mdsc->mutex);
2635 req = lookup_get_request(mdsc, tid);
2637 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
2638 goto out; /* dup reply? */
2641 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
2642 dout("forward tid %llu aborted, unregistering\n", tid);
2643 __unregister_request(mdsc, req);
2644 } else if (fwd_seq <= req->r_num_fwd) {
2645 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
2646 tid, next_mds, req->r_num_fwd, fwd_seq);
2648 /* resend. forward race not possible; mds would drop */
2649 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
2651 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
2652 req->r_attempts = 0;
2653 req->r_num_fwd = fwd_seq;
2654 req->r_resend_mds = next_mds;
2655 put_request_session(req);
2656 __do_request(mdsc, req);
2658 ceph_mdsc_put_request(req);
2660 mutex_unlock(&mdsc->mutex);
2664 pr_err("mdsc_handle_forward decode error err=%d\n", err);
2668 * handle a mds session control message
2670 static void handle_session(struct ceph_mds_session *session,
2671 struct ceph_msg *msg)
2673 struct ceph_mds_client *mdsc = session->s_mdsc;
2676 int mds = session->s_mds;
2677 struct ceph_mds_session_head *h = msg->front.iov_base;
2681 if (msg->front.iov_len != sizeof(*h))
2683 op = le32_to_cpu(h->op);
2684 seq = le64_to_cpu(h->seq);
2686 mutex_lock(&mdsc->mutex);
2687 if (op == CEPH_SESSION_CLOSE) {
2688 get_session(session);
2689 __unregister_session(mdsc, session);
2691 /* FIXME: this ttl calculation is generous */
2692 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
2693 mutex_unlock(&mdsc->mutex);
2695 mutex_lock(&session->s_mutex);
2697 dout("handle_session mds%d %s %p state %s seq %llu\n",
2698 mds, ceph_session_op_name(op), session,
2699 ceph_session_state_name(session->s_state), seq);
2701 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
2702 session->s_state = CEPH_MDS_SESSION_OPEN;
2703 pr_info("mds%d came back\n", session->s_mds);
2707 case CEPH_SESSION_OPEN:
2708 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2709 pr_info("mds%d reconnect success\n", session->s_mds);
2710 session->s_state = CEPH_MDS_SESSION_OPEN;
2711 renewed_caps(mdsc, session, 0);
2714 __close_session(mdsc, session);
2717 case CEPH_SESSION_RENEWCAPS:
2718 if (session->s_renew_seq == seq)
2719 renewed_caps(mdsc, session, 1);
2722 case CEPH_SESSION_CLOSE:
2723 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
2724 pr_info("mds%d reconnect denied\n", session->s_mds);
2725 cleanup_session_requests(mdsc, session);
2726 remove_session_caps(session);
2727 wake = 2; /* for good measure */
2728 wake_up_all(&mdsc->session_close_wq);
2731 case CEPH_SESSION_STALE:
2732 pr_info("mds%d caps went stale, renewing\n",
2734 spin_lock(&session->s_gen_ttl_lock);
2735 session->s_cap_gen++;
2736 session->s_cap_ttl = jiffies - 1;
2737 spin_unlock(&session->s_gen_ttl_lock);
2738 send_renew_caps(mdsc, session);
2741 case CEPH_SESSION_RECALL_STATE:
2742 trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
2745 case CEPH_SESSION_FLUSHMSG:
2746 send_flushmsg_ack(mdsc, session, seq);
2749 case CEPH_SESSION_FORCE_RO:
2750 dout("force_session_readonly %p\n", session);
2751 spin_lock(&session->s_cap_lock);
2752 session->s_readonly = true;
2753 spin_unlock(&session->s_cap_lock);
2754 wake_up_session_caps(session, 0);
2757 case CEPH_SESSION_REJECT:
2758 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
2759 pr_info("mds%d rejected session\n", session->s_mds);
2760 session->s_state = CEPH_MDS_SESSION_REJECTED;
2761 cleanup_session_requests(mdsc, session);
2762 remove_session_caps(session);
2763 wake = 2; /* for good measure */
2767 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
2771 mutex_unlock(&session->s_mutex);
2773 mutex_lock(&mdsc->mutex);
2774 __wake_requests(mdsc, &session->s_waiting);
2776 kick_requests(mdsc, mds);
2777 mutex_unlock(&mdsc->mutex);
2779 if (op == CEPH_SESSION_CLOSE)
2780 ceph_put_mds_session(session);
2784 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
2785 (int)msg->front.iov_len);
2792 * called under session->mutex.
2794 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
2795 struct ceph_mds_session *session)
2797 struct ceph_mds_request *req, *nreq;
2801 dout("replay_unsafe_requests mds%d\n", session->s_mds);
2803 mutex_lock(&mdsc->mutex);
2804 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item) {
2805 err = __prepare_send_request(mdsc, req, session->s_mds, true);
2807 ceph_msg_get(req->r_request);
2808 ceph_con_send(&session->s_con, req->r_request);
2813 * also re-send old requests when MDS enters reconnect stage. So that MDS
2814 * can process completed request in clientreplay stage.
2816 p = rb_first(&mdsc->request_tree);
2818 req = rb_entry(p, struct ceph_mds_request, r_node);
2820 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2822 if (req->r_attempts == 0)
2823 continue; /* only old requests */
2824 if (req->r_session &&
2825 req->r_session->s_mds == session->s_mds) {
2826 err = __prepare_send_request(mdsc, req,
2827 session->s_mds, true);
2829 ceph_msg_get(req->r_request);
2830 ceph_con_send(&session->s_con, req->r_request);
2834 mutex_unlock(&mdsc->mutex);
2838 * Encode information about a cap for a reconnect with the MDS.
2840 static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
2844 struct ceph_mds_cap_reconnect v2;
2845 struct ceph_mds_cap_reconnect_v1 v1;
2847 struct ceph_inode_info *ci = cap->ci;
2848 struct ceph_reconnect_state *recon_state = arg;
2849 struct ceph_pagelist *pagelist = recon_state->pagelist;
2854 struct dentry *dentry;
2856 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
2857 inode, ceph_vinop(inode), cap, cap->cap_id,
2858 ceph_cap_string(cap->issued));
2859 err = ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
2863 dentry = d_find_alias(inode);
2865 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase, 0);
2867 err = PTR_ERR(path);
2876 spin_lock(&ci->i_ceph_lock);
2877 cap->seq = 0; /* reset cap seq */
2878 cap->issue_seq = 0; /* and issue_seq */
2879 cap->mseq = 0; /* and migrate_seq */
2880 cap->cap_gen = cap->session->s_cap_gen;
2882 if (recon_state->msg_version >= 2) {
2883 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
2884 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2885 rec.v2.issued = cpu_to_le32(cap->issued);
2886 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2887 rec.v2.pathbase = cpu_to_le64(pathbase);
2889 (ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1;
2891 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
2892 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
2893 rec.v1.issued = cpu_to_le32(cap->issued);
2894 rec.v1.size = cpu_to_le64(inode->i_size);
2895 ceph_encode_timespec(&rec.v1.mtime, &inode->i_mtime);
2896 ceph_encode_timespec(&rec.v1.atime, &inode->i_atime);
2897 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
2898 rec.v1.pathbase = cpu_to_le64(pathbase);
2901 if (list_empty(&ci->i_cap_snaps)) {
2902 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
2904 struct ceph_cap_snap *capsnap =
2905 list_first_entry(&ci->i_cap_snaps,
2906 struct ceph_cap_snap, ci_item);
2907 snap_follows = capsnap->follows;
2909 spin_unlock(&ci->i_ceph_lock);
2911 if (recon_state->msg_version >= 2) {
2912 int num_fcntl_locks, num_flock_locks;
2913 struct ceph_filelock *flocks = NULL;
2914 size_t struct_len, total_len = 0;
2918 if (rec.v2.flock_len) {
2919 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
2921 num_fcntl_locks = 0;
2922 num_flock_locks = 0;
2924 if (num_fcntl_locks + num_flock_locks > 0) {
2925 flocks = kmalloc((num_fcntl_locks + num_flock_locks) *
2926 sizeof(struct ceph_filelock), GFP_NOFS);
2931 err = ceph_encode_locks_to_buffer(inode, flocks,
2946 if (recon_state->msg_version >= 3) {
2947 /* version, compat_version and struct_len */
2948 total_len = 2 * sizeof(u8) + sizeof(u32);
2952 * number of encoded locks is stable, so copy to pagelist
2954 struct_len = 2 * sizeof(u32) +
2955 (num_fcntl_locks + num_flock_locks) *
2956 sizeof(struct ceph_filelock);
2957 rec.v2.flock_len = cpu_to_le32(struct_len);
2959 struct_len += sizeof(rec.v2);
2960 struct_len += sizeof(u32) + pathlen;
2963 struct_len += sizeof(u64); /* snap_follows */
2965 total_len += struct_len;
2966 err = ceph_pagelist_reserve(pagelist, total_len);
2969 if (recon_state->msg_version >= 3) {
2970 ceph_pagelist_encode_8(pagelist, struct_v);
2971 ceph_pagelist_encode_8(pagelist, 1);
2972 ceph_pagelist_encode_32(pagelist, struct_len);
2974 ceph_pagelist_encode_string(pagelist, path, pathlen);
2975 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
2976 ceph_locks_to_pagelist(flocks, pagelist,
2980 ceph_pagelist_encode_64(pagelist, snap_follows);
2984 size_t size = sizeof(u32) + pathlen + sizeof(rec.v1);
2985 err = ceph_pagelist_reserve(pagelist, size);
2987 ceph_pagelist_encode_string(pagelist, path, pathlen);
2988 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
2992 recon_state->nr_caps++;
3002 * If an MDS fails and recovers, clients need to reconnect in order to
3003 * reestablish shared state. This includes all caps issued through
3004 * this session _and_ the snap_realm hierarchy. Because it's not
3005 * clear which snap realms the mds cares about, we send everything we
3006 * know about.. that ensures we'll then get any new info the
3007 * recovering MDS might have.
3009 * This is a relatively heavyweight operation, but it's rare.
3011 * called with mdsc->mutex held.
3013 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3014 struct ceph_mds_session *session)
3016 struct ceph_msg *reply;
3018 int mds = session->s_mds;
3021 struct ceph_pagelist *pagelist;
3022 struct ceph_reconnect_state recon_state;
3025 pr_info("mds%d reconnect start\n", mds);
3027 pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS);
3029 goto fail_nopagelist;
3030 ceph_pagelist_init(pagelist);
3032 reply = ceph_msg_new(CEPH_MSG_CLIENT_RECONNECT, 0, GFP_NOFS, false);
3036 mutex_lock(&session->s_mutex);
3037 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3040 dout("session %p state %s\n", session,
3041 ceph_session_state_name(session->s_state));
3043 spin_lock(&session->s_gen_ttl_lock);
3044 session->s_cap_gen++;
3045 spin_unlock(&session->s_gen_ttl_lock);
3047 spin_lock(&session->s_cap_lock);
3048 /* don't know if session is readonly */
3049 session->s_readonly = 0;
3051 * notify __ceph_remove_cap() that we are composing cap reconnect.
3052 * If a cap get released before being added to the cap reconnect,
3053 * __ceph_remove_cap() should skip queuing cap release.
3055 session->s_cap_reconnect = 1;
3056 /* drop old cap expires; we're about to reestablish that state */
3057 detach_cap_releases(session, &dispose);
3058 spin_unlock(&session->s_cap_lock);
3059 dispose_cap_releases(mdsc, &dispose);
3061 /* trim unused caps to reduce MDS's cache rejoin time */
3062 if (mdsc->fsc->sb->s_root)
3063 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3065 ceph_con_close(&session->s_con);
3066 ceph_con_open(&session->s_con,
3067 CEPH_ENTITY_TYPE_MDS, mds,
3068 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3070 /* replay unsafe requests */
3071 replay_unsafe_requests(mdsc, session);
3073 down_read(&mdsc->snap_rwsem);
3075 /* traverse this session's caps */
3076 s_nr_caps = session->s_nr_caps;
3077 err = ceph_pagelist_encode_32(pagelist, s_nr_caps);
3081 recon_state.nr_caps = 0;
3082 recon_state.pagelist = pagelist;
3083 if (session->s_con.peer_features & CEPH_FEATURE_MDSENC)
3084 recon_state.msg_version = 3;
3085 else if (session->s_con.peer_features & CEPH_FEATURE_FLOCK)
3086 recon_state.msg_version = 2;
3088 recon_state.msg_version = 1;
3089 err = iterate_session_caps(session, encode_caps_cb, &recon_state);
3093 spin_lock(&session->s_cap_lock);
3094 session->s_cap_reconnect = 0;
3095 spin_unlock(&session->s_cap_lock);
3098 * snaprealms. we provide mds with the ino, seq (version), and
3099 * parent for all of our realms. If the mds has any newer info,
3102 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3103 struct ceph_snap_realm *realm =
3104 rb_entry(p, struct ceph_snap_realm, node);
3105 struct ceph_mds_snaprealm_reconnect sr_rec;
3107 dout(" adding snap realm %llx seq %lld parent %llx\n",
3108 realm->ino, realm->seq, realm->parent_ino);
3109 sr_rec.ino = cpu_to_le64(realm->ino);
3110 sr_rec.seq = cpu_to_le64(realm->seq);
3111 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3112 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3117 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3119 /* raced with cap release? */
3120 if (s_nr_caps != recon_state.nr_caps) {
3121 struct page *page = list_first_entry(&pagelist->head,
3123 __le32 *addr = kmap_atomic(page);
3124 *addr = cpu_to_le32(recon_state.nr_caps);
3125 kunmap_atomic(addr);
3128 reply->hdr.data_len = cpu_to_le32(pagelist->length);
3129 ceph_msg_data_add_pagelist(reply, pagelist);
3131 ceph_early_kick_flushing_caps(mdsc, session);
3133 ceph_con_send(&session->s_con, reply);
3135 mutex_unlock(&session->s_mutex);
3137 mutex_lock(&mdsc->mutex);
3138 __wake_requests(mdsc, &session->s_waiting);
3139 mutex_unlock(&mdsc->mutex);
3141 up_read(&mdsc->snap_rwsem);
3145 ceph_msg_put(reply);
3146 up_read(&mdsc->snap_rwsem);
3147 mutex_unlock(&session->s_mutex);
3149 ceph_pagelist_release(pagelist);
3151 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3157 * compare old and new mdsmaps, kicking requests
3158 * and closing out old connections as necessary
3160 * called under mdsc->mutex.
3162 static void check_new_map(struct ceph_mds_client *mdsc,
3163 struct ceph_mdsmap *newmap,
3164 struct ceph_mdsmap *oldmap)
3167 int oldstate, newstate;
3168 struct ceph_mds_session *s;
3170 dout("check_new_map new %u old %u\n",
3171 newmap->m_epoch, oldmap->m_epoch);
3173 for (i = 0; i < oldmap->m_num_mds && i < mdsc->max_sessions; i++) {
3174 if (!mdsc->sessions[i])
3176 s = mdsc->sessions[i];
3177 oldstate = ceph_mdsmap_get_state(oldmap, i);
3178 newstate = ceph_mdsmap_get_state(newmap, i);
3180 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3181 i, ceph_mds_state_name(oldstate),
3182 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3183 ceph_mds_state_name(newstate),
3184 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3185 ceph_session_state_name(s->s_state));
3187 if (i >= newmap->m_num_mds ||
3188 memcmp(ceph_mdsmap_get_addr(oldmap, i),
3189 ceph_mdsmap_get_addr(newmap, i),
3190 sizeof(struct ceph_entity_addr))) {
3191 if (s->s_state == CEPH_MDS_SESSION_OPENING) {
3192 /* the session never opened, just close it
3195 __unregister_session(mdsc, s);
3196 __wake_requests(mdsc, &s->s_waiting);
3197 ceph_put_mds_session(s);
3198 } else if (i >= newmap->m_num_mds) {
3199 /* force close session for stopped mds */
3201 __unregister_session(mdsc, s);
3202 __wake_requests(mdsc, &s->s_waiting);
3203 kick_requests(mdsc, i);
3204 mutex_unlock(&mdsc->mutex);
3206 mutex_lock(&s->s_mutex);
3207 cleanup_session_requests(mdsc, s);
3208 remove_session_caps(s);
3209 mutex_unlock(&s->s_mutex);
3211 ceph_put_mds_session(s);
3213 mutex_lock(&mdsc->mutex);
3216 mutex_unlock(&mdsc->mutex);
3217 mutex_lock(&s->s_mutex);
3218 mutex_lock(&mdsc->mutex);
3219 ceph_con_close(&s->s_con);
3220 mutex_unlock(&s->s_mutex);
3221 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3223 } else if (oldstate == newstate) {
3224 continue; /* nothing new with this mds */
3230 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
3231 newstate >= CEPH_MDS_STATE_RECONNECT) {
3232 mutex_unlock(&mdsc->mutex);
3233 send_mds_reconnect(mdsc, s);
3234 mutex_lock(&mdsc->mutex);
3238 * kick request on any mds that has gone active.
3240 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
3241 newstate >= CEPH_MDS_STATE_ACTIVE) {
3242 if (oldstate != CEPH_MDS_STATE_CREATING &&
3243 oldstate != CEPH_MDS_STATE_STARTING)
3244 pr_info("mds%d recovery completed\n", s->s_mds);
3245 kick_requests(mdsc, i);
3246 ceph_kick_flushing_caps(mdsc, s);
3247 wake_up_session_caps(s, 1);
3251 for (i = 0; i < newmap->m_num_mds && i < mdsc->max_sessions; i++) {
3252 s = mdsc->sessions[i];
3255 if (!ceph_mdsmap_is_laggy(newmap, i))
3257 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3258 s->s_state == CEPH_MDS_SESSION_HUNG ||
3259 s->s_state == CEPH_MDS_SESSION_CLOSING) {
3260 dout(" connecting to export targets of laggy mds%d\n",
3262 __open_export_target_sessions(mdsc, s);
3274 * caller must hold session s_mutex, dentry->d_lock
3276 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
3278 struct ceph_dentry_info *di = ceph_dentry(dentry);
3280 ceph_put_mds_session(di->lease_session);
3281 di->lease_session = NULL;
3284 static void handle_lease(struct ceph_mds_client *mdsc,
3285 struct ceph_mds_session *session,
3286 struct ceph_msg *msg)
3288 struct super_block *sb = mdsc->fsc->sb;
3289 struct inode *inode;
3290 struct dentry *parent, *dentry;
3291 struct ceph_dentry_info *di;
3292 int mds = session->s_mds;
3293 struct ceph_mds_lease *h = msg->front.iov_base;
3295 struct ceph_vino vino;
3299 dout("handle_lease from mds%d\n", mds);
3302 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
3304 vino.ino = le64_to_cpu(h->ino);
3305 vino.snap = CEPH_NOSNAP;
3306 seq = le32_to_cpu(h->seq);
3307 dname.name = (void *)h + sizeof(*h) + sizeof(u32);
3308 dname.len = msg->front.iov_len - sizeof(*h) - sizeof(u32);
3309 if (dname.len != get_unaligned_le32(h+1))
3313 inode = ceph_find_inode(sb, vino);
3314 dout("handle_lease %s, ino %llx %p %.*s\n",
3315 ceph_lease_op_name(h->action), vino.ino, inode,
3316 dname.len, dname.name);
3318 mutex_lock(&session->s_mutex);
3322 dout("handle_lease no inode %llx\n", vino.ino);
3327 parent = d_find_alias(inode);
3329 dout("no parent dentry on inode %p\n", inode);
3331 goto release; /* hrm... */
3333 dname.hash = full_name_hash(parent, dname.name, dname.len);
3334 dentry = d_lookup(parent, &dname);
3339 spin_lock(&dentry->d_lock);
3340 di = ceph_dentry(dentry);
3341 switch (h->action) {
3342 case CEPH_MDS_LEASE_REVOKE:
3343 if (di->lease_session == session) {
3344 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
3345 h->seq = cpu_to_le32(di->lease_seq);
3346 __ceph_mdsc_drop_dentry_lease(dentry);
3351 case CEPH_MDS_LEASE_RENEW:
3352 if (di->lease_session == session &&
3353 di->lease_gen == session->s_cap_gen &&
3354 di->lease_renew_from &&
3355 di->lease_renew_after == 0) {
3356 unsigned long duration =
3357 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
3359 di->lease_seq = seq;
3360 di->time = di->lease_renew_from + duration;
3361 di->lease_renew_after = di->lease_renew_from +
3363 di->lease_renew_from = 0;
3367 spin_unlock(&dentry->d_lock);
3374 /* let's just reuse the same message */
3375 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
3377 ceph_con_send(&session->s_con, msg);
3381 mutex_unlock(&session->s_mutex);
3385 pr_err("corrupt lease message\n");
3389 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
3390 struct inode *inode,
3391 struct dentry *dentry, char action,
3394 struct ceph_msg *msg;
3395 struct ceph_mds_lease *lease;
3396 int len = sizeof(*lease) + sizeof(u32);
3399 dout("lease_send_msg inode %p dentry %p %s to mds%d\n",
3400 inode, dentry, ceph_lease_op_name(action), session->s_mds);
3401 dnamelen = dentry->d_name.len;
3404 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
3407 lease = msg->front.iov_base;
3408 lease->action = action;
3409 lease->ino = cpu_to_le64(ceph_vino(inode).ino);
3410 lease->first = lease->last = cpu_to_le64(ceph_vino(inode).snap);
3411 lease->seq = cpu_to_le32(seq);
3412 put_unaligned_le32(dnamelen, lease + 1);
3413 memcpy((void *)(lease + 1) + 4, dentry->d_name.name, dnamelen);
3416 * if this is a preemptive lease RELEASE, no need to
3417 * flush request stream, since the actual request will
3420 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
3422 ceph_con_send(&session->s_con, msg);
3426 * drop all leases (and dentry refs) in preparation for umount
3428 static void drop_leases(struct ceph_mds_client *mdsc)
3432 dout("drop_leases\n");
3433 mutex_lock(&mdsc->mutex);
3434 for (i = 0; i < mdsc->max_sessions; i++) {
3435 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3438 mutex_unlock(&mdsc->mutex);
3439 mutex_lock(&s->s_mutex);
3440 mutex_unlock(&s->s_mutex);
3441 ceph_put_mds_session(s);
3442 mutex_lock(&mdsc->mutex);
3444 mutex_unlock(&mdsc->mutex);
3450 * delayed work -- periodically trim expired leases, renew caps with mds
3452 static void schedule_delayed(struct ceph_mds_client *mdsc)
3455 unsigned hz = round_jiffies_relative(HZ * delay);
3456 schedule_delayed_work(&mdsc->delayed_work, hz);
3459 static void delayed_work(struct work_struct *work)
3462 struct ceph_mds_client *mdsc =
3463 container_of(work, struct ceph_mds_client, delayed_work.work);
3467 dout("mdsc delayed_work\n");
3468 ceph_check_delayed_caps(mdsc);
3470 mutex_lock(&mdsc->mutex);
3471 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
3472 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
3473 mdsc->last_renew_caps);
3475 mdsc->last_renew_caps = jiffies;
3477 for (i = 0; i < mdsc->max_sessions; i++) {
3478 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
3481 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
3482 dout("resending session close request for mds%d\n",
3484 request_close_session(mdsc, s);
3485 ceph_put_mds_session(s);
3488 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
3489 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
3490 s->s_state = CEPH_MDS_SESSION_HUNG;
3491 pr_info("mds%d hung\n", s->s_mds);
3494 if (s->s_state < CEPH_MDS_SESSION_OPEN) {
3495 /* this mds is failed or recovering, just wait */
3496 ceph_put_mds_session(s);
3499 mutex_unlock(&mdsc->mutex);
3501 mutex_lock(&s->s_mutex);
3503 send_renew_caps(mdsc, s);
3505 ceph_con_keepalive(&s->s_con);
3506 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3507 s->s_state == CEPH_MDS_SESSION_HUNG)
3508 ceph_send_cap_releases(mdsc, s);
3509 mutex_unlock(&s->s_mutex);
3510 ceph_put_mds_session(s);
3512 mutex_lock(&mdsc->mutex);
3514 mutex_unlock(&mdsc->mutex);
3516 schedule_delayed(mdsc);
3519 int ceph_mdsc_init(struct ceph_fs_client *fsc)
3522 struct ceph_mds_client *mdsc;
3524 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
3529 mutex_init(&mdsc->mutex);
3530 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
3531 if (!mdsc->mdsmap) {
3536 init_completion(&mdsc->safe_umount_waiters);
3537 init_waitqueue_head(&mdsc->session_close_wq);
3538 INIT_LIST_HEAD(&mdsc->waiting_for_map);
3539 mdsc->sessions = NULL;
3540 atomic_set(&mdsc->num_sessions, 0);
3541 mdsc->max_sessions = 0;
3543 mdsc->last_snap_seq = 0;
3544 init_rwsem(&mdsc->snap_rwsem);
3545 mdsc->snap_realms = RB_ROOT;
3546 INIT_LIST_HEAD(&mdsc->snap_empty);
3547 spin_lock_init(&mdsc->snap_empty_lock);
3549 mdsc->oldest_tid = 0;
3550 mdsc->request_tree = RB_ROOT;
3551 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3552 mdsc->last_renew_caps = jiffies;
3553 INIT_LIST_HEAD(&mdsc->cap_delay_list);
3554 spin_lock_init(&mdsc->cap_delay_lock);
3555 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3556 spin_lock_init(&mdsc->snap_flush_lock);
3557 mdsc->last_cap_flush_tid = 1;
3558 INIT_LIST_HEAD(&mdsc->cap_flush_list);
3559 INIT_LIST_HEAD(&mdsc->cap_dirty);
3560 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3561 mdsc->num_cap_flushing = 0;
3562 spin_lock_init(&mdsc->cap_dirty_lock);
3563 init_waitqueue_head(&mdsc->cap_flushing_wq);
3564 spin_lock_init(&mdsc->dentry_lru_lock);
3565 INIT_LIST_HEAD(&mdsc->dentry_lru);
3567 ceph_caps_init(mdsc);
3568 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3570 init_rwsem(&mdsc->pool_perm_rwsem);
3571 mdsc->pool_perm_tree = RB_ROOT;
3573 strncpy(mdsc->nodename, utsname()->nodename,
3574 sizeof(mdsc->nodename) - 1);
3579 * Wait for safe replies on open mds requests. If we time out, drop
3580 * all requests from the tree to avoid dangling dentry refs.
3582 static void wait_requests(struct ceph_mds_client *mdsc)
3584 struct ceph_options *opts = mdsc->fsc->client->options;
3585 struct ceph_mds_request *req;
3587 mutex_lock(&mdsc->mutex);
3588 if (__get_oldest_req(mdsc)) {
3589 mutex_unlock(&mdsc->mutex);
3591 dout("wait_requests waiting for requests\n");
3592 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3593 ceph_timeout_jiffies(opts->mount_timeout));
3595 /* tear down remaining requests */
3596 mutex_lock(&mdsc->mutex);
3597 while ((req = __get_oldest_req(mdsc))) {
3598 dout("wait_requests timed out on tid %llu\n",
3600 __unregister_request(mdsc, req);
3603 mutex_unlock(&mdsc->mutex);
3604 dout("wait_requests done\n");
3608 * called before mount is ro, and before dentries are torn down.
3609 * (hmm, does this still race with new lookups?)
3611 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
3613 dout("pre_umount\n");
3617 ceph_flush_dirty_caps(mdsc);
3618 wait_requests(mdsc);
3621 * wait for reply handlers to drop their request refs and
3622 * their inode/dcache refs
3628 * wait for all write mds requests to flush.
3630 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
3632 struct ceph_mds_request *req = NULL, *nextreq;
3635 mutex_lock(&mdsc->mutex);
3636 dout("wait_unsafe_requests want %lld\n", want_tid);
3638 req = __get_oldest_req(mdsc);
3639 while (req && req->r_tid <= want_tid) {
3640 /* find next request */
3641 n = rb_next(&req->r_node);
3643 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3646 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3647 (req->r_op & CEPH_MDS_OP_WRITE)) {
3649 ceph_mdsc_get_request(req);
3651 ceph_mdsc_get_request(nextreq);
3652 mutex_unlock(&mdsc->mutex);
3653 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
3654 req->r_tid, want_tid);
3655 wait_for_completion(&req->r_safe_completion);
3656 mutex_lock(&mdsc->mutex);
3657 ceph_mdsc_put_request(req);
3659 break; /* next dne before, so we're done! */
3660 if (RB_EMPTY_NODE(&nextreq->r_node)) {
3661 /* next request was removed from tree */
3662 ceph_mdsc_put_request(nextreq);
3665 ceph_mdsc_put_request(nextreq); /* won't go away */
3669 mutex_unlock(&mdsc->mutex);
3670 dout("wait_unsafe_requests done\n");
3673 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3675 u64 want_tid, want_flush;
3677 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3681 mutex_lock(&mdsc->mutex);
3682 want_tid = mdsc->last_tid;
3683 mutex_unlock(&mdsc->mutex);
3685 ceph_flush_dirty_caps(mdsc);
3686 spin_lock(&mdsc->cap_dirty_lock);
3687 want_flush = mdsc->last_cap_flush_tid;
3688 if (!list_empty(&mdsc->cap_flush_list)) {
3689 struct ceph_cap_flush *cf =
3690 list_last_entry(&mdsc->cap_flush_list,
3691 struct ceph_cap_flush, g_list);
3694 spin_unlock(&mdsc->cap_dirty_lock);
3696 dout("sync want tid %lld flush_seq %lld\n",
3697 want_tid, want_flush);
3699 wait_unsafe_requests(mdsc, want_tid);
3700 wait_caps_flush(mdsc, want_flush);
3704 * true if all sessions are closed, or we force unmount
3706 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
3708 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
3710 return atomic_read(&mdsc->num_sessions) <= skipped;
3714 * called after sb is ro.
3716 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3718 struct ceph_options *opts = mdsc->fsc->client->options;
3719 struct ceph_mds_session *session;
3723 dout("close_sessions\n");
3725 /* close sessions */
3726 mutex_lock(&mdsc->mutex);
3727 for (i = 0; i < mdsc->max_sessions; i++) {
3728 session = __ceph_lookup_mds_session(mdsc, i);
3731 mutex_unlock(&mdsc->mutex);
3732 mutex_lock(&session->s_mutex);
3733 if (__close_session(mdsc, session) <= 0)
3735 mutex_unlock(&session->s_mutex);
3736 ceph_put_mds_session(session);
3737 mutex_lock(&mdsc->mutex);
3739 mutex_unlock(&mdsc->mutex);
3741 dout("waiting for sessions to close\n");
3742 wait_event_timeout(mdsc->session_close_wq,
3743 done_closing_sessions(mdsc, skipped),
3744 ceph_timeout_jiffies(opts->mount_timeout));
3746 /* tear down remaining sessions */
3747 mutex_lock(&mdsc->mutex);
3748 for (i = 0; i < mdsc->max_sessions; i++) {
3749 if (mdsc->sessions[i]) {
3750 session = get_session(mdsc->sessions[i]);
3751 __unregister_session(mdsc, session);
3752 mutex_unlock(&mdsc->mutex);
3753 mutex_lock(&session->s_mutex);
3754 remove_session_caps(session);
3755 mutex_unlock(&session->s_mutex);
3756 ceph_put_mds_session(session);
3757 mutex_lock(&mdsc->mutex);
3760 WARN_ON(!list_empty(&mdsc->cap_delay_list));
3761 mutex_unlock(&mdsc->mutex);
3763 ceph_cleanup_empty_realms(mdsc);
3765 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3770 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
3772 struct ceph_mds_session *session;
3775 dout("force umount\n");
3777 mutex_lock(&mdsc->mutex);
3778 for (mds = 0; mds < mdsc->max_sessions; mds++) {
3779 session = __ceph_lookup_mds_session(mdsc, mds);
3782 mutex_unlock(&mdsc->mutex);
3783 mutex_lock(&session->s_mutex);
3784 __close_session(mdsc, session);
3785 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
3786 cleanup_session_requests(mdsc, session);
3787 remove_session_caps(session);
3789 mutex_unlock(&session->s_mutex);
3790 ceph_put_mds_session(session);
3791 mutex_lock(&mdsc->mutex);
3792 kick_requests(mdsc, mds);
3794 __wake_requests(mdsc, &mdsc->waiting_for_map);
3795 mutex_unlock(&mdsc->mutex);
3798 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3801 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
3803 ceph_mdsmap_destroy(mdsc->mdsmap);
3804 kfree(mdsc->sessions);
3805 ceph_caps_finalize(mdsc);
3806 ceph_pool_perm_destroy(mdsc);
3809 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
3811 struct ceph_mds_client *mdsc = fsc->mdsc;
3812 dout("mdsc_destroy %p\n", mdsc);
3814 /* flush out any connection work with references to us */
3817 ceph_mdsc_stop(mdsc);
3821 dout("mdsc_destroy %p done\n", mdsc);
3824 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3826 struct ceph_fs_client *fsc = mdsc->fsc;
3827 const char *mds_namespace = fsc->mount_options->mds_namespace;
3828 void *p = msg->front.iov_base;
3829 void *end = p + msg->front.iov_len;
3833 u32 mount_fscid = (u32)-1;
3834 u8 struct_v, struct_cv;
3837 ceph_decode_need(&p, end, sizeof(u32), bad);
3838 epoch = ceph_decode_32(&p);
3840 dout("handle_fsmap epoch %u\n", epoch);
3842 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3843 struct_v = ceph_decode_8(&p);
3844 struct_cv = ceph_decode_8(&p);
3845 map_len = ceph_decode_32(&p);
3847 ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
3848 p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
3850 num_fs = ceph_decode_32(&p);
3851 while (num_fs-- > 0) {
3852 void *info_p, *info_end;
3857 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
3858 info_v = ceph_decode_8(&p);
3859 info_cv = ceph_decode_8(&p);
3860 info_len = ceph_decode_32(&p);
3861 ceph_decode_need(&p, end, info_len, bad);
3863 info_end = p + info_len;
3866 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
3867 fscid = ceph_decode_32(&info_p);
3868 namelen = ceph_decode_32(&info_p);
3869 ceph_decode_need(&info_p, info_end, namelen, bad);
3871 if (mds_namespace &&
3872 strlen(mds_namespace) == namelen &&
3873 !strncmp(mds_namespace, (char *)info_p, namelen)) {
3874 mount_fscid = fscid;
3879 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
3880 if (mount_fscid != (u32)-1) {
3881 fsc->client->monc.fs_cluster_id = mount_fscid;
3882 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
3884 ceph_monc_renew_subs(&fsc->client->monc);
3892 pr_err("error decoding fsmap\n");
3894 mutex_lock(&mdsc->mutex);
3895 mdsc->mdsmap_err = err;
3896 __wake_requests(mdsc, &mdsc->waiting_for_map);
3897 mutex_unlock(&mdsc->mutex);
3901 * handle mds map update.
3903 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
3907 void *p = msg->front.iov_base;
3908 void *end = p + msg->front.iov_len;
3909 struct ceph_mdsmap *newmap, *oldmap;
3910 struct ceph_fsid fsid;
3913 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
3914 ceph_decode_copy(&p, &fsid, sizeof(fsid));
3915 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
3917 epoch = ceph_decode_32(&p);
3918 maplen = ceph_decode_32(&p);
3919 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
3921 /* do we need it? */
3922 mutex_lock(&mdsc->mutex);
3923 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
3924 dout("handle_map epoch %u <= our %u\n",
3925 epoch, mdsc->mdsmap->m_epoch);
3926 mutex_unlock(&mdsc->mutex);
3930 newmap = ceph_mdsmap_decode(&p, end);
3931 if (IS_ERR(newmap)) {
3932 err = PTR_ERR(newmap);
3936 /* swap into place */
3938 oldmap = mdsc->mdsmap;
3939 mdsc->mdsmap = newmap;
3940 check_new_map(mdsc, newmap, oldmap);
3941 ceph_mdsmap_destroy(oldmap);
3943 mdsc->mdsmap = newmap; /* first mds map */
3945 mdsc->fsc->sb->s_maxbytes = mdsc->mdsmap->m_max_file_size;
3947 __wake_requests(mdsc, &mdsc->waiting_for_map);
3948 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
3949 mdsc->mdsmap->m_epoch);
3951 mutex_unlock(&mdsc->mutex);
3952 schedule_delayed(mdsc);
3956 mutex_unlock(&mdsc->mutex);
3958 pr_err("error decoding mdsmap %d\n", err);
3962 static struct ceph_connection *con_get(struct ceph_connection *con)
3964 struct ceph_mds_session *s = con->private;
3966 if (get_session(s)) {
3967 dout("mdsc con_get %p ok (%d)\n", s, refcount_read(&s->s_ref));
3970 dout("mdsc con_get %p FAIL\n", s);
3974 static void con_put(struct ceph_connection *con)
3976 struct ceph_mds_session *s = con->private;
3978 dout("mdsc con_put %p (%d)\n", s, refcount_read(&s->s_ref) - 1);
3979 ceph_put_mds_session(s);
3983 * if the client is unresponsive for long enough, the mds will kill
3984 * the session entirely.
3986 static void peer_reset(struct ceph_connection *con)
3988 struct ceph_mds_session *s = con->private;
3989 struct ceph_mds_client *mdsc = s->s_mdsc;
3991 pr_warn("mds%d closed our session\n", s->s_mds);
3992 send_mds_reconnect(mdsc, s);
3995 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
3997 struct ceph_mds_session *s = con->private;
3998 struct ceph_mds_client *mdsc = s->s_mdsc;
3999 int type = le16_to_cpu(msg->hdr.type);
4001 mutex_lock(&mdsc->mutex);
4002 if (__verify_registered_session(mdsc, s) < 0) {
4003 mutex_unlock(&mdsc->mutex);
4006 mutex_unlock(&mdsc->mutex);
4009 case CEPH_MSG_MDS_MAP:
4010 ceph_mdsc_handle_mdsmap(mdsc, msg);
4012 case CEPH_MSG_FS_MAP_USER:
4013 ceph_mdsc_handle_fsmap(mdsc, msg);
4015 case CEPH_MSG_CLIENT_SESSION:
4016 handle_session(s, msg);
4018 case CEPH_MSG_CLIENT_REPLY:
4019 handle_reply(s, msg);
4021 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4022 handle_forward(mdsc, s, msg);
4024 case CEPH_MSG_CLIENT_CAPS:
4025 ceph_handle_caps(s, msg);
4027 case CEPH_MSG_CLIENT_SNAP:
4028 ceph_handle_snap(mdsc, s, msg);
4030 case CEPH_MSG_CLIENT_LEASE:
4031 handle_lease(mdsc, s, msg);
4035 pr_err("received unknown message type %d %s\n", type,
4036 ceph_msg_type_name(type));
4047 * Note: returned pointer is the address of a structure that's
4048 * managed separately. Caller must *not* attempt to free it.
4050 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4051 int *proto, int force_new)
4053 struct ceph_mds_session *s = con->private;
4054 struct ceph_mds_client *mdsc = s->s_mdsc;
4055 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4056 struct ceph_auth_handshake *auth = &s->s_auth;
4058 if (force_new && auth->authorizer) {
4059 ceph_auth_destroy_authorizer(auth->authorizer);
4060 auth->authorizer = NULL;
4062 if (!auth->authorizer) {
4063 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4066 return ERR_PTR(ret);
4068 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4071 return ERR_PTR(ret);
4073 *proto = ac->protocol;
4079 static int verify_authorizer_reply(struct ceph_connection *con)
4081 struct ceph_mds_session *s = con->private;
4082 struct ceph_mds_client *mdsc = s->s_mdsc;
4083 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4085 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4088 static int invalidate_authorizer(struct ceph_connection *con)
4090 struct ceph_mds_session *s = con->private;
4091 struct ceph_mds_client *mdsc = s->s_mdsc;
4092 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4094 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4096 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4099 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4100 struct ceph_msg_header *hdr, int *skip)
4102 struct ceph_msg *msg;
4103 int type = (int) le16_to_cpu(hdr->type);
4104 int front_len = (int) le32_to_cpu(hdr->front_len);
4110 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4112 pr_err("unable to allocate msg type %d len %d\n",
4120 static int mds_sign_message(struct ceph_msg *msg)
4122 struct ceph_mds_session *s = msg->con->private;
4123 struct ceph_auth_handshake *auth = &s->s_auth;
4125 return ceph_auth_sign_message(auth, msg);
4128 static int mds_check_message_signature(struct ceph_msg *msg)
4130 struct ceph_mds_session *s = msg->con->private;
4131 struct ceph_auth_handshake *auth = &s->s_auth;
4133 return ceph_auth_check_message_signature(auth, msg);
4136 static const struct ceph_connection_operations mds_con_ops = {
4139 .dispatch = dispatch,
4140 .get_authorizer = get_authorizer,
4141 .verify_authorizer_reply = verify_authorizer_reply,
4142 .invalidate_authorizer = invalidate_authorizer,
4143 .peer_reset = peer_reset,
4144 .alloc_msg = mds_alloc_msg,
4145 .sign_message = mds_sign_message,
4146 .check_message_signature = mds_check_message_signature,