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>
12 #include <linux/bits.h>
15 #include "mds_client.h"
17 #include <linux/ceph/ceph_features.h>
18 #include <linux/ceph/messenger.h>
19 #include <linux/ceph/decode.h>
20 #include <linux/ceph/pagelist.h>
21 #include <linux/ceph/auth.h>
22 #include <linux/ceph/debugfs.h>
24 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
27 * A cluster of MDS (metadata server) daemons is responsible for
28 * managing the file system namespace (the directory hierarchy and
29 * inodes) and for coordinating shared access to storage. Metadata is
30 * partitioning hierarchically across a number of servers, and that
31 * partition varies over time as the cluster adjusts the distribution
32 * in order to balance load.
34 * The MDS client is primarily responsible to managing synchronous
35 * metadata requests for operations like open, unlink, and so forth.
36 * If there is a MDS failure, we find out about it when we (possibly
37 * request and) receive a new MDS map, and can resubmit affected
40 * For the most part, though, we take advantage of a lossless
41 * communications channel to the MDS, and do not need to worry about
42 * timing out or resubmitting requests.
44 * We maintain a stateful "session" with each MDS we interact with.
45 * Within each session, we sent periodic heartbeat messages to ensure
46 * any capabilities or leases we have been issues remain valid. If
47 * the session times out and goes stale, our leases and capabilities
48 * are no longer valid.
51 struct ceph_reconnect_state {
52 struct ceph_mds_session *session;
53 int nr_caps, nr_realms;
54 struct ceph_pagelist *pagelist;
59 static void __wake_requests(struct ceph_mds_client *mdsc,
60 struct list_head *head);
61 static void ceph_cap_release_work(struct work_struct *work);
62 static void ceph_cap_reclaim_work(struct work_struct *work);
64 static const struct ceph_connection_operations mds_con_ops;
71 static int parse_reply_info_quota(void **p, void *end,
72 struct ceph_mds_reply_info_in *info)
74 u8 struct_v, struct_compat;
77 ceph_decode_8_safe(p, end, struct_v, bad);
78 ceph_decode_8_safe(p, end, struct_compat, bad);
79 /* struct_v is expected to be >= 1. we only
80 * understand encoding with struct_compat == 1. */
81 if (!struct_v || struct_compat != 1)
83 ceph_decode_32_safe(p, end, struct_len, bad);
84 ceph_decode_need(p, end, struct_len, bad);
85 end = *p + struct_len;
86 ceph_decode_64_safe(p, end, info->max_bytes, bad);
87 ceph_decode_64_safe(p, end, info->max_files, bad);
95 * parse individual inode info
97 static int parse_reply_info_in(void **p, void *end,
98 struct ceph_mds_reply_info_in *info,
104 if (features == (u64)-1) {
107 ceph_decode_8_safe(p, end, struct_v, bad);
108 ceph_decode_8_safe(p, end, struct_compat, bad);
109 /* struct_v is expected to be >= 1. we only understand
110 * encoding with struct_compat == 1. */
111 if (!struct_v || struct_compat != 1)
113 ceph_decode_32_safe(p, end, struct_len, bad);
114 ceph_decode_need(p, end, struct_len, bad);
115 end = *p + struct_len;
118 ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
120 *p += sizeof(struct ceph_mds_reply_inode) +
121 sizeof(*info->in->fragtree.splits) *
122 le32_to_cpu(info->in->fragtree.nsplits);
124 ceph_decode_32_safe(p, end, info->symlink_len, bad);
125 ceph_decode_need(p, end, info->symlink_len, bad);
127 *p += info->symlink_len;
129 ceph_decode_copy_safe(p, end, &info->dir_layout,
130 sizeof(info->dir_layout), bad);
131 ceph_decode_32_safe(p, end, info->xattr_len, bad);
132 ceph_decode_need(p, end, info->xattr_len, bad);
133 info->xattr_data = *p;
134 *p += info->xattr_len;
136 if (features == (u64)-1) {
138 ceph_decode_64_safe(p, end, info->inline_version, bad);
139 ceph_decode_32_safe(p, end, info->inline_len, bad);
140 ceph_decode_need(p, end, info->inline_len, bad);
141 info->inline_data = *p;
142 *p += info->inline_len;
144 err = parse_reply_info_quota(p, end, info);
148 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
149 if (info->pool_ns_len > 0) {
150 ceph_decode_need(p, end, info->pool_ns_len, bad);
151 info->pool_ns_data = *p;
152 *p += info->pool_ns_len;
156 ceph_decode_need(p, end, sizeof(info->btime), bad);
157 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
159 /* change attribute */
160 ceph_decode_64_safe(p, end, info->change_attr, bad);
164 ceph_decode_32_safe(p, end, info->dir_pin, bad);
166 info->dir_pin = -ENODATA;
169 /* snapshot birth time, remains zero for v<=2 */
171 ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
172 ceph_decode_copy(p, &info->snap_btime,
173 sizeof(info->snap_btime));
175 memset(&info->snap_btime, 0, sizeof(info->snap_btime));
180 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
181 ceph_decode_64_safe(p, end, info->inline_version, bad);
182 ceph_decode_32_safe(p, end, info->inline_len, bad);
183 ceph_decode_need(p, end, info->inline_len, bad);
184 info->inline_data = *p;
185 *p += info->inline_len;
187 info->inline_version = CEPH_INLINE_NONE;
189 if (features & CEPH_FEATURE_MDS_QUOTA) {
190 err = parse_reply_info_quota(p, end, info);
198 info->pool_ns_len = 0;
199 info->pool_ns_data = NULL;
200 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
201 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
202 if (info->pool_ns_len > 0) {
203 ceph_decode_need(p, end, info->pool_ns_len, bad);
204 info->pool_ns_data = *p;
205 *p += info->pool_ns_len;
209 if (features & CEPH_FEATURE_FS_BTIME) {
210 ceph_decode_need(p, end, sizeof(info->btime), bad);
211 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
212 ceph_decode_64_safe(p, end, info->change_attr, bad);
215 info->dir_pin = -ENODATA;
216 /* info->snap_btime remains zero */
225 static int parse_reply_info_dir(void **p, void *end,
226 struct ceph_mds_reply_dirfrag **dirfrag,
229 if (features == (u64)-1) {
230 u8 struct_v, struct_compat;
232 ceph_decode_8_safe(p, end, struct_v, bad);
233 ceph_decode_8_safe(p, end, struct_compat, bad);
234 /* struct_v is expected to be >= 1. we only understand
235 * encoding whose struct_compat == 1. */
236 if (!struct_v || struct_compat != 1)
238 ceph_decode_32_safe(p, end, struct_len, bad);
239 ceph_decode_need(p, end, struct_len, bad);
240 end = *p + struct_len;
243 ceph_decode_need(p, end, sizeof(**dirfrag), bad);
245 *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
246 if (unlikely(*p > end))
248 if (features == (u64)-1)
255 static int parse_reply_info_lease(void **p, void *end,
256 struct ceph_mds_reply_lease **lease,
259 if (features == (u64)-1) {
260 u8 struct_v, struct_compat;
262 ceph_decode_8_safe(p, end, struct_v, bad);
263 ceph_decode_8_safe(p, end, struct_compat, bad);
264 /* struct_v is expected to be >= 1. we only understand
265 * encoding whose struct_compat == 1. */
266 if (!struct_v || struct_compat != 1)
268 ceph_decode_32_safe(p, end, struct_len, bad);
269 ceph_decode_need(p, end, struct_len, bad);
270 end = *p + struct_len;
273 ceph_decode_need(p, end, sizeof(**lease), bad);
275 *p += sizeof(**lease);
276 if (features == (u64)-1)
284 * parse a normal reply, which may contain a (dir+)dentry and/or a
287 static int parse_reply_info_trace(void **p, void *end,
288 struct ceph_mds_reply_info_parsed *info,
293 if (info->head->is_dentry) {
294 err = parse_reply_info_in(p, end, &info->diri, features);
298 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
302 ceph_decode_32_safe(p, end, info->dname_len, bad);
303 ceph_decode_need(p, end, info->dname_len, bad);
305 *p += info->dname_len;
307 err = parse_reply_info_lease(p, end, &info->dlease, features);
312 if (info->head->is_target) {
313 err = parse_reply_info_in(p, end, &info->targeti, features);
318 if (unlikely(*p != end))
325 pr_err("problem parsing mds trace %d\n", err);
330 * parse readdir results
332 static int parse_reply_info_readdir(void **p, void *end,
333 struct ceph_mds_reply_info_parsed *info,
339 err = parse_reply_info_dir(p, end, &info->dir_dir, features);
343 ceph_decode_need(p, end, sizeof(num) + 2, bad);
344 num = ceph_decode_32(p);
346 u16 flags = ceph_decode_16(p);
347 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
348 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
349 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
350 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
355 BUG_ON(!info->dir_entries);
356 if ((unsigned long)(info->dir_entries + num) >
357 (unsigned long)info->dir_entries + info->dir_buf_size) {
358 pr_err("dir contents are larger than expected\n");
365 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
367 ceph_decode_32_safe(p, end, rde->name_len, bad);
368 ceph_decode_need(p, end, rde->name_len, bad);
371 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
374 err = parse_reply_info_lease(p, end, &rde->lease, features);
378 err = parse_reply_info_in(p, end, &rde->inode, features);
381 /* ceph_readdir_prepopulate() will update it */
388 /* Skip over any unrecognized fields */
395 pr_err("problem parsing dir contents %d\n", err);
400 * parse fcntl F_GETLK results
402 static int parse_reply_info_filelock(void **p, void *end,
403 struct ceph_mds_reply_info_parsed *info,
406 if (*p + sizeof(*info->filelock_reply) > end)
409 info->filelock_reply = *p;
411 /* Skip over any unrecognized fields */
419 #if BITS_PER_LONG == 64
421 #define DELEGATED_INO_AVAILABLE xa_mk_value(1)
423 static int ceph_parse_deleg_inos(void **p, void *end,
424 struct ceph_mds_session *s)
428 ceph_decode_32_safe(p, end, sets, bad);
429 dout("got %u sets of delegated inodes\n", sets);
433 ceph_decode_64_safe(p, end, start, bad);
434 ceph_decode_64_safe(p, end, len, bad);
436 int err = xa_insert(&s->s_delegated_inos, ino = start++,
437 DELEGATED_INO_AVAILABLE,
440 dout("added delegated inode 0x%llx\n",
442 } else if (err == -EBUSY) {
443 pr_warn("ceph: MDS delegated inode 0x%llx more than once.\n",
455 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
460 xa_for_each(&s->s_delegated_inos, ino, val) {
461 val = xa_erase(&s->s_delegated_inos, ino);
462 if (val == DELEGATED_INO_AVAILABLE)
468 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
470 return xa_insert(&s->s_delegated_inos, ino, DELEGATED_INO_AVAILABLE,
473 #else /* BITS_PER_LONG == 64 */
475 * FIXME: xarrays can't handle 64-bit indexes on a 32-bit arch. For now, just
476 * ignore delegated_inos on 32 bit arch. Maybe eventually add xarrays for top
479 static int ceph_parse_deleg_inos(void **p, void *end,
480 struct ceph_mds_session *s)
484 ceph_decode_32_safe(p, end, sets, bad);
486 ceph_decode_skip_n(p, end, sets * 2 * sizeof(__le64), bad);
492 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
497 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
501 #endif /* BITS_PER_LONG == 64 */
504 * parse create results
506 static int parse_reply_info_create(void **p, void *end,
507 struct ceph_mds_reply_info_parsed *info,
508 u64 features, struct ceph_mds_session *s)
512 if (features == (u64)-1 ||
513 (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
515 /* Malformed reply? */
516 info->has_create_ino = false;
517 } else if (test_bit(CEPHFS_FEATURE_DELEG_INO, &s->s_features)) {
518 u8 struct_v, struct_compat;
521 info->has_create_ino = true;
522 ceph_decode_8_safe(p, end, struct_v, bad);
523 ceph_decode_8_safe(p, end, struct_compat, bad);
524 ceph_decode_32_safe(p, end, len, bad);
525 ceph_decode_64_safe(p, end, info->ino, bad);
526 ret = ceph_parse_deleg_inos(p, end, s);
531 ceph_decode_64_safe(p, end, info->ino, bad);
532 info->has_create_ino = true;
539 /* Skip over any unrecognized fields */
547 * parse extra results
549 static int parse_reply_info_extra(void **p, void *end,
550 struct ceph_mds_reply_info_parsed *info,
551 u64 features, struct ceph_mds_session *s)
553 u32 op = le32_to_cpu(info->head->op);
555 if (op == CEPH_MDS_OP_GETFILELOCK)
556 return parse_reply_info_filelock(p, end, info, features);
557 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
558 return parse_reply_info_readdir(p, end, info, features);
559 else if (op == CEPH_MDS_OP_CREATE)
560 return parse_reply_info_create(p, end, info, features, s);
566 * parse entire mds reply
568 static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
569 struct ceph_mds_reply_info_parsed *info,
576 info->head = msg->front.iov_base;
577 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
578 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
581 ceph_decode_32_safe(&p, end, len, bad);
583 ceph_decode_need(&p, end, len, bad);
584 err = parse_reply_info_trace(&p, p+len, info, features);
590 ceph_decode_32_safe(&p, end, len, bad);
592 ceph_decode_need(&p, end, len, bad);
593 err = parse_reply_info_extra(&p, p+len, info, features, s);
599 ceph_decode_32_safe(&p, end, len, bad);
600 info->snapblob_len = len;
611 pr_err("mds parse_reply err %d\n", err);
615 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
617 if (!info->dir_entries)
619 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
626 const char *ceph_session_state_name(int s)
629 case CEPH_MDS_SESSION_NEW: return "new";
630 case CEPH_MDS_SESSION_OPENING: return "opening";
631 case CEPH_MDS_SESSION_OPEN: return "open";
632 case CEPH_MDS_SESSION_HUNG: return "hung";
633 case CEPH_MDS_SESSION_CLOSING: return "closing";
634 case CEPH_MDS_SESSION_CLOSED: return "closed";
635 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
636 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
637 case CEPH_MDS_SESSION_REJECTED: return "rejected";
638 default: return "???";
642 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
644 if (refcount_inc_not_zero(&s->s_ref)) {
645 dout("mdsc get_session %p %d -> %d\n", s,
646 refcount_read(&s->s_ref)-1, refcount_read(&s->s_ref));
649 dout("mdsc get_session %p 0 -- FAIL\n", s);
654 void ceph_put_mds_session(struct ceph_mds_session *s)
656 dout("mdsc put_session %p %d -> %d\n", s,
657 refcount_read(&s->s_ref), refcount_read(&s->s_ref)-1);
658 if (refcount_dec_and_test(&s->s_ref)) {
659 if (s->s_auth.authorizer)
660 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
661 xa_destroy(&s->s_delegated_inos);
667 * called under mdsc->mutex
669 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
672 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
674 return ceph_get_mds_session(mdsc->sessions[mds]);
677 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
679 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
685 static int __verify_registered_session(struct ceph_mds_client *mdsc,
686 struct ceph_mds_session *s)
688 if (s->s_mds >= mdsc->max_sessions ||
689 mdsc->sessions[s->s_mds] != s)
695 * create+register a new session for given mds.
696 * called under mdsc->mutex.
698 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
701 struct ceph_mds_session *s;
703 if (mds >= mdsc->mdsmap->possible_max_rank)
704 return ERR_PTR(-EINVAL);
706 s = kzalloc(sizeof(*s), GFP_NOFS);
708 return ERR_PTR(-ENOMEM);
710 if (mds >= mdsc->max_sessions) {
711 int newmax = 1 << get_count_order(mds + 1);
712 struct ceph_mds_session **sa;
714 dout("%s: realloc to %d\n", __func__, newmax);
715 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
718 if (mdsc->sessions) {
719 memcpy(sa, mdsc->sessions,
720 mdsc->max_sessions * sizeof(void *));
721 kfree(mdsc->sessions);
724 mdsc->max_sessions = newmax;
727 dout("%s: mds%d\n", __func__, mds);
730 s->s_state = CEPH_MDS_SESSION_NEW;
733 mutex_init(&s->s_mutex);
735 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
737 spin_lock_init(&s->s_gen_ttl_lock);
739 s->s_cap_ttl = jiffies - 1;
741 spin_lock_init(&s->s_cap_lock);
742 s->s_renew_requested = 0;
744 INIT_LIST_HEAD(&s->s_caps);
746 refcount_set(&s->s_ref, 1);
747 INIT_LIST_HEAD(&s->s_waiting);
748 INIT_LIST_HEAD(&s->s_unsafe);
749 xa_init(&s->s_delegated_inos);
750 s->s_num_cap_releases = 0;
751 s->s_cap_reconnect = 0;
752 s->s_cap_iterator = NULL;
753 INIT_LIST_HEAD(&s->s_cap_releases);
754 INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
756 INIT_LIST_HEAD(&s->s_cap_flushing);
758 mdsc->sessions[mds] = s;
759 atomic_inc(&mdsc->num_sessions);
760 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
762 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
763 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
769 return ERR_PTR(-ENOMEM);
773 * called under mdsc->mutex
775 static void __unregister_session(struct ceph_mds_client *mdsc,
776 struct ceph_mds_session *s)
778 dout("__unregister_session mds%d %p\n", s->s_mds, s);
779 BUG_ON(mdsc->sessions[s->s_mds] != s);
780 mdsc->sessions[s->s_mds] = NULL;
781 ceph_con_close(&s->s_con);
782 ceph_put_mds_session(s);
783 atomic_dec(&mdsc->num_sessions);
787 * drop session refs in request.
789 * should be last request ref, or hold mdsc->mutex
791 static void put_request_session(struct ceph_mds_request *req)
793 if (req->r_session) {
794 ceph_put_mds_session(req->r_session);
795 req->r_session = NULL;
799 void ceph_mdsc_release_request(struct kref *kref)
801 struct ceph_mds_request *req = container_of(kref,
802 struct ceph_mds_request,
804 ceph_mdsc_release_dir_caps(req);
805 destroy_reply_info(&req->r_reply_info);
807 ceph_msg_put(req->r_request);
809 ceph_msg_put(req->r_reply);
811 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
812 /* avoid calling iput_final() in mds dispatch threads */
813 ceph_async_iput(req->r_inode);
816 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
817 ceph_async_iput(req->r_parent);
819 ceph_async_iput(req->r_target_inode);
822 if (req->r_old_dentry)
823 dput(req->r_old_dentry);
824 if (req->r_old_dentry_dir) {
826 * track (and drop pins for) r_old_dentry_dir
827 * separately, since r_old_dentry's d_parent may have
828 * changed between the dir mutex being dropped and
829 * this request being freed.
831 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
833 ceph_async_iput(req->r_old_dentry_dir);
838 ceph_pagelist_release(req->r_pagelist);
839 put_request_session(req);
840 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
841 WARN_ON_ONCE(!list_empty(&req->r_wait));
842 kmem_cache_free(ceph_mds_request_cachep, req);
845 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
848 * lookup session, bump ref if found.
850 * called under mdsc->mutex.
852 static struct ceph_mds_request *
853 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
855 struct ceph_mds_request *req;
857 req = lookup_request(&mdsc->request_tree, tid);
859 ceph_mdsc_get_request(req);
865 * Register an in-flight request, and assign a tid. Link to directory
866 * are modifying (if any).
868 * Called under mdsc->mutex.
870 static void __register_request(struct ceph_mds_client *mdsc,
871 struct ceph_mds_request *req,
876 req->r_tid = ++mdsc->last_tid;
877 if (req->r_num_caps) {
878 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
881 pr_err("__register_request %p "
882 "failed to reserve caps: %d\n", req, ret);
883 /* set req->r_err to fail early from __do_request */
888 dout("__register_request %p tid %lld\n", req, req->r_tid);
889 ceph_mdsc_get_request(req);
890 insert_request(&mdsc->request_tree, req);
892 req->r_uid = current_fsuid();
893 req->r_gid = current_fsgid();
895 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
896 mdsc->oldest_tid = req->r_tid;
899 struct ceph_inode_info *ci = ceph_inode(dir);
902 req->r_unsafe_dir = dir;
903 spin_lock(&ci->i_unsafe_lock);
904 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
905 spin_unlock(&ci->i_unsafe_lock);
909 static void __unregister_request(struct ceph_mds_client *mdsc,
910 struct ceph_mds_request *req)
912 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
914 /* Never leave an unregistered request on an unsafe list! */
915 list_del_init(&req->r_unsafe_item);
917 if (req->r_tid == mdsc->oldest_tid) {
918 struct rb_node *p = rb_next(&req->r_node);
919 mdsc->oldest_tid = 0;
921 struct ceph_mds_request *next_req =
922 rb_entry(p, struct ceph_mds_request, r_node);
923 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
924 mdsc->oldest_tid = next_req->r_tid;
931 erase_request(&mdsc->request_tree, req);
933 if (req->r_unsafe_dir) {
934 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
935 spin_lock(&ci->i_unsafe_lock);
936 list_del_init(&req->r_unsafe_dir_item);
937 spin_unlock(&ci->i_unsafe_lock);
939 if (req->r_target_inode &&
940 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
941 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
942 spin_lock(&ci->i_unsafe_lock);
943 list_del_init(&req->r_unsafe_target_item);
944 spin_unlock(&ci->i_unsafe_lock);
947 if (req->r_unsafe_dir) {
948 /* avoid calling iput_final() in mds dispatch threads */
949 ceph_async_iput(req->r_unsafe_dir);
950 req->r_unsafe_dir = NULL;
953 complete_all(&req->r_safe_completion);
955 ceph_mdsc_put_request(req);
959 * Walk back up the dentry tree until we hit a dentry representing a
960 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
961 * when calling this) to ensure that the objects won't disappear while we're
962 * working with them. Once we hit a candidate dentry, we attempt to take a
963 * reference to it, and return that as the result.
965 static struct inode *get_nonsnap_parent(struct dentry *dentry)
967 struct inode *inode = NULL;
969 while (dentry && !IS_ROOT(dentry)) {
970 inode = d_inode_rcu(dentry);
971 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
973 dentry = dentry->d_parent;
976 inode = igrab(inode);
981 * Choose mds to send request to next. If there is a hint set in the
982 * request (e.g., due to a prior forward hint from the mds), use that.
983 * Otherwise, consult frag tree and/or caps to identify the
984 * appropriate mds. If all else fails, choose randomly.
986 * Called under mdsc->mutex.
988 static int __choose_mds(struct ceph_mds_client *mdsc,
989 struct ceph_mds_request *req,
993 struct ceph_inode_info *ci;
994 struct ceph_cap *cap;
995 int mode = req->r_direct_mode;
997 u32 hash = req->r_direct_hash;
998 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
1004 * is there a specific mds we should try? ignore hint if we have
1005 * no session and the mds is not up (active or recovering).
1007 if (req->r_resend_mds >= 0 &&
1008 (__have_session(mdsc, req->r_resend_mds) ||
1009 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1010 dout("%s using resend_mds mds%d\n", __func__,
1012 return req->r_resend_mds;
1015 if (mode == USE_RANDOM_MDS)
1020 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1021 inode = req->r_inode;
1024 /* req->r_dentry is non-null for LSSNAP request */
1026 inode = get_nonsnap_parent(req->r_dentry);
1028 dout("%s using snapdir's parent %p\n", __func__, inode);
1030 } else if (req->r_dentry) {
1031 /* ignore race with rename; old or new d_parent is okay */
1032 struct dentry *parent;
1036 parent = READ_ONCE(req->r_dentry->d_parent);
1037 dir = req->r_parent ? : d_inode_rcu(parent);
1039 if (!dir || dir->i_sb != mdsc->fsc->sb) {
1040 /* not this fs or parent went negative */
1041 inode = d_inode(req->r_dentry);
1044 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
1045 /* direct snapped/virtual snapdir requests
1046 * based on parent dir inode */
1047 inode = get_nonsnap_parent(parent);
1048 dout("%s using nonsnap parent %p\n", __func__, inode);
1051 inode = d_inode(req->r_dentry);
1052 if (!inode || mode == USE_AUTH_MDS) {
1055 hash = ceph_dentry_hash(dir, req->r_dentry);
1064 dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
1068 ci = ceph_inode(inode);
1070 if (is_hash && S_ISDIR(inode->i_mode)) {
1071 struct ceph_inode_frag frag;
1074 ceph_choose_frag(ci, hash, &frag, &found);
1076 if (mode == USE_ANY_MDS && frag.ndist > 0) {
1079 /* choose a random replica */
1080 get_random_bytes(&r, 1);
1083 dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
1084 __func__, inode, ceph_vinop(inode),
1085 frag.frag, mds, (int)r, frag.ndist);
1086 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1087 CEPH_MDS_STATE_ACTIVE &&
1088 !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1092 /* since this file/dir wasn't known to be
1093 * replicated, then we want to look for the
1094 * authoritative mds. */
1095 if (frag.mds >= 0) {
1096 /* choose auth mds */
1098 dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
1099 __func__, inode, ceph_vinop(inode),
1101 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1102 CEPH_MDS_STATE_ACTIVE) {
1103 if (mode == USE_ANY_MDS &&
1104 !ceph_mdsmap_is_laggy(mdsc->mdsmap,
1109 mode = USE_AUTH_MDS;
1113 spin_lock(&ci->i_ceph_lock);
1115 if (mode == USE_AUTH_MDS)
1116 cap = ci->i_auth_cap;
1117 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1118 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1120 spin_unlock(&ci->i_ceph_lock);
1121 ceph_async_iput(inode);
1124 mds = cap->session->s_mds;
1125 dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1126 inode, ceph_vinop(inode), mds,
1127 cap == ci->i_auth_cap ? "auth " : "", cap);
1128 spin_unlock(&ci->i_ceph_lock);
1130 /* avoid calling iput_final() while holding mdsc->mutex or
1131 * in mds dispatch threads */
1132 ceph_async_iput(inode);
1139 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1140 dout("%s chose random mds%d\n", __func__, mds);
1148 static struct ceph_msg *create_session_msg(u32 op, u64 seq)
1150 struct ceph_msg *msg;
1151 struct ceph_mds_session_head *h;
1153 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1156 pr_err("create_session_msg ENOMEM creating msg\n");
1159 h = msg->front.iov_base;
1160 h->op = cpu_to_le32(op);
1161 h->seq = cpu_to_le64(seq);
1166 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1167 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1168 static void encode_supported_features(void **p, void *end)
1170 static const size_t count = ARRAY_SIZE(feature_bits);
1174 size_t size = FEATURE_BYTES(count);
1176 BUG_ON(*p + 4 + size > end);
1177 ceph_encode_32(p, size);
1178 memset(*p, 0, size);
1179 for (i = 0; i < count; i++)
1180 ((unsigned char*)(*p))[i / 8] |= BIT(feature_bits[i] % 8);
1183 BUG_ON(*p + 4 > end);
1184 ceph_encode_32(p, 0);
1189 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1190 * to include additional client metadata fields.
1192 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1194 struct ceph_msg *msg;
1195 struct ceph_mds_session_head *h;
1197 int extra_bytes = 0;
1198 int metadata_key_count = 0;
1199 struct ceph_options *opt = mdsc->fsc->client->options;
1200 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1204 const char* metadata[][2] = {
1205 {"hostname", mdsc->nodename},
1206 {"kernel_version", init_utsname()->release},
1207 {"entity_id", opt->name ? : ""},
1208 {"root", fsopt->server_path ? : "/"},
1212 /* Calculate serialized length of metadata */
1213 extra_bytes = 4; /* map length */
1214 for (i = 0; metadata[i][0]; ++i) {
1215 extra_bytes += 8 + strlen(metadata[i][0]) +
1216 strlen(metadata[i][1]);
1217 metadata_key_count++;
1220 /* supported feature */
1222 count = ARRAY_SIZE(feature_bits);
1224 size = FEATURE_BYTES(count);
1225 extra_bytes += 4 + size;
1227 /* Allocate the message */
1228 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1231 pr_err("create_session_msg ENOMEM creating msg\n");
1234 p = msg->front.iov_base;
1235 end = p + msg->front.iov_len;
1238 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1239 h->seq = cpu_to_le64(seq);
1242 * Serialize client metadata into waiting buffer space, using
1243 * the format that userspace expects for map<string, string>
1245 * ClientSession messages with metadata are v3
1247 msg->hdr.version = cpu_to_le16(3);
1248 msg->hdr.compat_version = cpu_to_le16(1);
1250 /* The write pointer, following the session_head structure */
1253 /* Number of entries in the map */
1254 ceph_encode_32(&p, metadata_key_count);
1256 /* Two length-prefixed strings for each entry in the map */
1257 for (i = 0; metadata[i][0]; ++i) {
1258 size_t const key_len = strlen(metadata[i][0]);
1259 size_t const val_len = strlen(metadata[i][1]);
1261 ceph_encode_32(&p, key_len);
1262 memcpy(p, metadata[i][0], key_len);
1264 ceph_encode_32(&p, val_len);
1265 memcpy(p, metadata[i][1], val_len);
1269 encode_supported_features(&p, end);
1270 msg->front.iov_len = p - msg->front.iov_base;
1271 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1277 * send session open request.
1279 * called under mdsc->mutex
1281 static int __open_session(struct ceph_mds_client *mdsc,
1282 struct ceph_mds_session *session)
1284 struct ceph_msg *msg;
1286 int mds = session->s_mds;
1288 /* wait for mds to go active? */
1289 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1290 dout("open_session to mds%d (%s)\n", mds,
1291 ceph_mds_state_name(mstate));
1292 session->s_state = CEPH_MDS_SESSION_OPENING;
1293 session->s_renew_requested = jiffies;
1295 /* send connect message */
1296 msg = create_session_open_msg(mdsc, session->s_seq);
1299 ceph_con_send(&session->s_con, msg);
1304 * open sessions for any export targets for the given mds
1306 * called under mdsc->mutex
1308 static struct ceph_mds_session *
1309 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1311 struct ceph_mds_session *session;
1313 session = __ceph_lookup_mds_session(mdsc, target);
1315 session = register_session(mdsc, target);
1316 if (IS_ERR(session))
1319 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1320 session->s_state == CEPH_MDS_SESSION_CLOSING)
1321 __open_session(mdsc, session);
1326 struct ceph_mds_session *
1327 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1329 struct ceph_mds_session *session;
1331 dout("open_export_target_session to mds%d\n", target);
1333 mutex_lock(&mdsc->mutex);
1334 session = __open_export_target_session(mdsc, target);
1335 mutex_unlock(&mdsc->mutex);
1340 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1341 struct ceph_mds_session *session)
1343 struct ceph_mds_info *mi;
1344 struct ceph_mds_session *ts;
1345 int i, mds = session->s_mds;
1347 if (mds >= mdsc->mdsmap->possible_max_rank)
1350 mi = &mdsc->mdsmap->m_info[mds];
1351 dout("open_export_target_sessions for mds%d (%d targets)\n",
1352 session->s_mds, mi->num_export_targets);
1354 for (i = 0; i < mi->num_export_targets; i++) {
1355 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1357 ceph_put_mds_session(ts);
1361 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1362 struct ceph_mds_session *session)
1364 mutex_lock(&mdsc->mutex);
1365 __open_export_target_sessions(mdsc, session);
1366 mutex_unlock(&mdsc->mutex);
1373 static void detach_cap_releases(struct ceph_mds_session *session,
1374 struct list_head *target)
1376 lockdep_assert_held(&session->s_cap_lock);
1378 list_splice_init(&session->s_cap_releases, target);
1379 session->s_num_cap_releases = 0;
1380 dout("dispose_cap_releases mds%d\n", session->s_mds);
1383 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1384 struct list_head *dispose)
1386 while (!list_empty(dispose)) {
1387 struct ceph_cap *cap;
1388 /* zero out the in-progress message */
1389 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1390 list_del(&cap->session_caps);
1391 ceph_put_cap(mdsc, cap);
1395 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1396 struct ceph_mds_session *session)
1398 struct ceph_mds_request *req;
1400 struct ceph_inode_info *ci;
1402 dout("cleanup_session_requests mds%d\n", session->s_mds);
1403 mutex_lock(&mdsc->mutex);
1404 while (!list_empty(&session->s_unsafe)) {
1405 req = list_first_entry(&session->s_unsafe,
1406 struct ceph_mds_request, r_unsafe_item);
1407 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1409 if (req->r_target_inode) {
1410 /* dropping unsafe change of inode's attributes */
1411 ci = ceph_inode(req->r_target_inode);
1412 errseq_set(&ci->i_meta_err, -EIO);
1414 if (req->r_unsafe_dir) {
1415 /* dropping unsafe directory operation */
1416 ci = ceph_inode(req->r_unsafe_dir);
1417 errseq_set(&ci->i_meta_err, -EIO);
1419 __unregister_request(mdsc, req);
1421 /* zero r_attempts, so kick_requests() will re-send requests */
1422 p = rb_first(&mdsc->request_tree);
1424 req = rb_entry(p, struct ceph_mds_request, r_node);
1426 if (req->r_session &&
1427 req->r_session->s_mds == session->s_mds)
1428 req->r_attempts = 0;
1430 mutex_unlock(&mdsc->mutex);
1434 * Helper to safely iterate over all caps associated with a session, with
1435 * special care taken to handle a racing __ceph_remove_cap().
1437 * Caller must hold session s_mutex.
1439 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1440 int (*cb)(struct inode *, struct ceph_cap *,
1443 struct list_head *p;
1444 struct ceph_cap *cap;
1445 struct inode *inode, *last_inode = NULL;
1446 struct ceph_cap *old_cap = NULL;
1449 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1450 spin_lock(&session->s_cap_lock);
1451 p = session->s_caps.next;
1452 while (p != &session->s_caps) {
1453 cap = list_entry(p, struct ceph_cap, session_caps);
1454 inode = igrab(&cap->ci->vfs_inode);
1459 session->s_cap_iterator = cap;
1460 spin_unlock(&session->s_cap_lock);
1463 /* avoid calling iput_final() while holding
1464 * s_mutex or in mds dispatch threads */
1465 ceph_async_iput(last_inode);
1469 ceph_put_cap(session->s_mdsc, old_cap);
1473 ret = cb(inode, cap, arg);
1476 spin_lock(&session->s_cap_lock);
1479 dout("iterate_session_caps finishing cap %p removal\n",
1481 BUG_ON(cap->session != session);
1482 cap->session = NULL;
1483 list_del_init(&cap->session_caps);
1484 session->s_nr_caps--;
1485 if (cap->queue_release)
1486 __ceph_queue_cap_release(session, cap);
1488 old_cap = cap; /* put_cap it w/o locks held */
1495 session->s_cap_iterator = NULL;
1496 spin_unlock(&session->s_cap_lock);
1498 ceph_async_iput(last_inode);
1500 ceph_put_cap(session->s_mdsc, old_cap);
1505 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1508 struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
1509 struct ceph_inode_info *ci = ceph_inode(inode);
1510 LIST_HEAD(to_remove);
1511 bool dirty_dropped = false;
1512 bool invalidate = false;
1514 dout("removing cap %p, ci is %p, inode is %p\n",
1515 cap, ci, &ci->vfs_inode);
1516 spin_lock(&ci->i_ceph_lock);
1517 __ceph_remove_cap(cap, false);
1518 if (!ci->i_auth_cap) {
1519 struct ceph_cap_flush *cf;
1520 struct ceph_mds_client *mdsc = fsc->mdsc;
1522 if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1523 if (inode->i_data.nrpages > 0)
1525 if (ci->i_wrbuffer_ref > 0)
1526 mapping_set_error(&inode->i_data, -EIO);
1529 while (!list_empty(&ci->i_cap_flush_list)) {
1530 cf = list_first_entry(&ci->i_cap_flush_list,
1531 struct ceph_cap_flush, i_list);
1532 list_move(&cf->i_list, &to_remove);
1535 spin_lock(&mdsc->cap_dirty_lock);
1537 list_for_each_entry(cf, &to_remove, i_list)
1538 list_del(&cf->g_list);
1540 if (!list_empty(&ci->i_dirty_item)) {
1541 pr_warn_ratelimited(
1542 " dropping dirty %s state for %p %lld\n",
1543 ceph_cap_string(ci->i_dirty_caps),
1544 inode, ceph_ino(inode));
1545 ci->i_dirty_caps = 0;
1546 list_del_init(&ci->i_dirty_item);
1547 dirty_dropped = true;
1549 if (!list_empty(&ci->i_flushing_item)) {
1550 pr_warn_ratelimited(
1551 " dropping dirty+flushing %s state for %p %lld\n",
1552 ceph_cap_string(ci->i_flushing_caps),
1553 inode, ceph_ino(inode));
1554 ci->i_flushing_caps = 0;
1555 list_del_init(&ci->i_flushing_item);
1556 mdsc->num_cap_flushing--;
1557 dirty_dropped = true;
1559 spin_unlock(&mdsc->cap_dirty_lock);
1561 if (dirty_dropped) {
1562 errseq_set(&ci->i_meta_err, -EIO);
1564 if (ci->i_wrbuffer_ref_head == 0 &&
1565 ci->i_wr_ref == 0 &&
1566 ci->i_dirty_caps == 0 &&
1567 ci->i_flushing_caps == 0) {
1568 ceph_put_snap_context(ci->i_head_snapc);
1569 ci->i_head_snapc = NULL;
1573 if (atomic_read(&ci->i_filelock_ref) > 0) {
1574 /* make further file lock syscall return -EIO */
1575 ci->i_ceph_flags |= CEPH_I_ERROR_FILELOCK;
1576 pr_warn_ratelimited(" dropping file locks for %p %lld\n",
1577 inode, ceph_ino(inode));
1580 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1581 list_add(&ci->i_prealloc_cap_flush->i_list, &to_remove);
1582 ci->i_prealloc_cap_flush = NULL;
1585 spin_unlock(&ci->i_ceph_lock);
1586 while (!list_empty(&to_remove)) {
1587 struct ceph_cap_flush *cf;
1588 cf = list_first_entry(&to_remove,
1589 struct ceph_cap_flush, i_list);
1590 list_del(&cf->i_list);
1591 ceph_free_cap_flush(cf);
1594 wake_up_all(&ci->i_cap_wq);
1596 ceph_queue_invalidate(inode);
1603 * caller must hold session s_mutex
1605 static void remove_session_caps(struct ceph_mds_session *session)
1607 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1608 struct super_block *sb = fsc->sb;
1611 dout("remove_session_caps on %p\n", session);
1612 ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1614 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1616 spin_lock(&session->s_cap_lock);
1617 if (session->s_nr_caps > 0) {
1618 struct inode *inode;
1619 struct ceph_cap *cap, *prev = NULL;
1620 struct ceph_vino vino;
1622 * iterate_session_caps() skips inodes that are being
1623 * deleted, we need to wait until deletions are complete.
1624 * __wait_on_freeing_inode() is designed for the job,
1625 * but it is not exported, so use lookup inode function
1628 while (!list_empty(&session->s_caps)) {
1629 cap = list_entry(session->s_caps.next,
1630 struct ceph_cap, session_caps);
1634 vino = cap->ci->i_vino;
1635 spin_unlock(&session->s_cap_lock);
1637 inode = ceph_find_inode(sb, vino);
1638 /* avoid calling iput_final() while holding s_mutex */
1639 ceph_async_iput(inode);
1641 spin_lock(&session->s_cap_lock);
1645 // drop cap expires and unlock s_cap_lock
1646 detach_cap_releases(session, &dispose);
1648 BUG_ON(session->s_nr_caps > 0);
1649 BUG_ON(!list_empty(&session->s_cap_flushing));
1650 spin_unlock(&session->s_cap_lock);
1651 dispose_cap_releases(session->s_mdsc, &dispose);
1661 * wake up any threads waiting on this session's caps. if the cap is
1662 * old (didn't get renewed on the client reconnect), remove it now.
1664 * caller must hold s_mutex.
1666 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1669 struct ceph_inode_info *ci = ceph_inode(inode);
1670 unsigned long ev = (unsigned long)arg;
1672 if (ev == RECONNECT) {
1673 spin_lock(&ci->i_ceph_lock);
1674 ci->i_wanted_max_size = 0;
1675 ci->i_requested_max_size = 0;
1676 spin_unlock(&ci->i_ceph_lock);
1677 } else if (ev == RENEWCAPS) {
1678 if (cap->cap_gen < cap->session->s_cap_gen) {
1679 /* mds did not re-issue stale cap */
1680 spin_lock(&ci->i_ceph_lock);
1681 cap->issued = cap->implemented = CEPH_CAP_PIN;
1682 spin_unlock(&ci->i_ceph_lock);
1684 } else if (ev == FORCE_RO) {
1686 wake_up_all(&ci->i_cap_wq);
1690 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1692 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1693 ceph_iterate_session_caps(session, wake_up_session_cb,
1694 (void *)(unsigned long)ev);
1698 * Send periodic message to MDS renewing all currently held caps. The
1699 * ack will reset the expiration for all caps from this session.
1701 * caller holds s_mutex
1703 static int send_renew_caps(struct ceph_mds_client *mdsc,
1704 struct ceph_mds_session *session)
1706 struct ceph_msg *msg;
1709 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1710 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1711 pr_info("mds%d caps stale\n", session->s_mds);
1712 session->s_renew_requested = jiffies;
1714 /* do not try to renew caps until a recovering mds has reconnected
1715 * with its clients. */
1716 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1717 if (state < CEPH_MDS_STATE_RECONNECT) {
1718 dout("send_renew_caps ignoring mds%d (%s)\n",
1719 session->s_mds, ceph_mds_state_name(state));
1723 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1724 ceph_mds_state_name(state));
1725 msg = create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1726 ++session->s_renew_seq);
1729 ceph_con_send(&session->s_con, msg);
1733 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1734 struct ceph_mds_session *session, u64 seq)
1736 struct ceph_msg *msg;
1738 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1739 session->s_mds, ceph_session_state_name(session->s_state), seq);
1740 msg = create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1743 ceph_con_send(&session->s_con, msg);
1749 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1751 * Called under session->s_mutex
1753 static void renewed_caps(struct ceph_mds_client *mdsc,
1754 struct ceph_mds_session *session, int is_renew)
1759 spin_lock(&session->s_cap_lock);
1760 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1762 session->s_cap_ttl = session->s_renew_requested +
1763 mdsc->mdsmap->m_session_timeout*HZ;
1766 if (time_before(jiffies, session->s_cap_ttl)) {
1767 pr_info("mds%d caps renewed\n", session->s_mds);
1770 pr_info("mds%d caps still stale\n", session->s_mds);
1773 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1774 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1775 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1776 spin_unlock(&session->s_cap_lock);
1779 wake_up_session_caps(session, RENEWCAPS);
1783 * send a session close request
1785 static int request_close_session(struct ceph_mds_client *mdsc,
1786 struct ceph_mds_session *session)
1788 struct ceph_msg *msg;
1790 dout("request_close_session mds%d state %s seq %lld\n",
1791 session->s_mds, ceph_session_state_name(session->s_state),
1793 msg = create_session_msg(CEPH_SESSION_REQUEST_CLOSE, session->s_seq);
1796 ceph_con_send(&session->s_con, msg);
1801 * Called with s_mutex held.
1803 static int __close_session(struct ceph_mds_client *mdsc,
1804 struct ceph_mds_session *session)
1806 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1808 session->s_state = CEPH_MDS_SESSION_CLOSING;
1809 return request_close_session(mdsc, session);
1812 static bool drop_negative_children(struct dentry *dentry)
1814 struct dentry *child;
1815 bool all_negative = true;
1817 if (!d_is_dir(dentry))
1820 spin_lock(&dentry->d_lock);
1821 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1822 if (d_really_is_positive(child)) {
1823 all_negative = false;
1827 spin_unlock(&dentry->d_lock);
1830 shrink_dcache_parent(dentry);
1832 return all_negative;
1836 * Trim old(er) caps.
1838 * Because we can't cache an inode without one or more caps, we do
1839 * this indirectly: if a cap is unused, we prune its aliases, at which
1840 * point the inode will hopefully get dropped to.
1842 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1843 * memory pressure from the MDS, though, so it needn't be perfect.
1845 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1847 int *remaining = arg;
1848 struct ceph_inode_info *ci = ceph_inode(inode);
1849 int used, wanted, oissued, mine;
1851 if (*remaining <= 0)
1854 spin_lock(&ci->i_ceph_lock);
1855 mine = cap->issued | cap->implemented;
1856 used = __ceph_caps_used(ci);
1857 wanted = __ceph_caps_file_wanted(ci);
1858 oissued = __ceph_caps_issued_other(ci, cap);
1860 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1861 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1862 ceph_cap_string(used), ceph_cap_string(wanted));
1863 if (cap == ci->i_auth_cap) {
1864 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1865 !list_empty(&ci->i_cap_snaps))
1867 if ((used | wanted) & CEPH_CAP_ANY_WR)
1869 /* Note: it's possible that i_filelock_ref becomes non-zero
1870 * after dropping auth caps. It doesn't hurt because reply
1871 * of lock mds request will re-add auth caps. */
1872 if (atomic_read(&ci->i_filelock_ref) > 0)
1875 /* The inode has cached pages, but it's no longer used.
1876 * we can safely drop it */
1877 if (S_ISREG(inode->i_mode) &&
1878 wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1879 !(oissued & CEPH_CAP_FILE_CACHE)) {
1883 if ((used | wanted) & ~oissued & mine)
1884 goto out; /* we need these caps */
1887 /* we aren't the only cap.. just remove us */
1888 __ceph_remove_cap(cap, true);
1891 struct dentry *dentry;
1892 /* try dropping referring dentries */
1893 spin_unlock(&ci->i_ceph_lock);
1894 dentry = d_find_any_alias(inode);
1895 if (dentry && drop_negative_children(dentry)) {
1898 d_prune_aliases(inode);
1899 count = atomic_read(&inode->i_count);
1902 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1911 spin_unlock(&ci->i_ceph_lock);
1916 * Trim session cap count down to some max number.
1918 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1919 struct ceph_mds_session *session,
1922 int trim_caps = session->s_nr_caps - max_caps;
1924 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1925 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1926 if (trim_caps > 0) {
1927 int remaining = trim_caps;
1929 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
1930 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1931 session->s_mds, session->s_nr_caps, max_caps,
1932 trim_caps - remaining);
1935 ceph_flush_cap_releases(mdsc, session);
1939 static int check_caps_flush(struct ceph_mds_client *mdsc,
1944 spin_lock(&mdsc->cap_dirty_lock);
1945 if (!list_empty(&mdsc->cap_flush_list)) {
1946 struct ceph_cap_flush *cf =
1947 list_first_entry(&mdsc->cap_flush_list,
1948 struct ceph_cap_flush, g_list);
1949 if (cf->tid <= want_flush_tid) {
1950 dout("check_caps_flush still flushing tid "
1951 "%llu <= %llu\n", cf->tid, want_flush_tid);
1955 spin_unlock(&mdsc->cap_dirty_lock);
1960 * flush all dirty inode data to disk.
1962 * returns true if we've flushed through want_flush_tid
1964 static void wait_caps_flush(struct ceph_mds_client *mdsc,
1967 dout("check_caps_flush want %llu\n", want_flush_tid);
1969 wait_event(mdsc->cap_flushing_wq,
1970 check_caps_flush(mdsc, want_flush_tid));
1972 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1976 * called under s_mutex
1978 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1979 struct ceph_mds_session *session)
1981 struct ceph_msg *msg = NULL;
1982 struct ceph_mds_cap_release *head;
1983 struct ceph_mds_cap_item *item;
1984 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
1985 struct ceph_cap *cap;
1986 LIST_HEAD(tmp_list);
1987 int num_cap_releases;
1988 __le32 barrier, *cap_barrier;
1990 down_read(&osdc->lock);
1991 barrier = cpu_to_le32(osdc->epoch_barrier);
1992 up_read(&osdc->lock);
1994 spin_lock(&session->s_cap_lock);
1996 list_splice_init(&session->s_cap_releases, &tmp_list);
1997 num_cap_releases = session->s_num_cap_releases;
1998 session->s_num_cap_releases = 0;
1999 spin_unlock(&session->s_cap_lock);
2001 while (!list_empty(&tmp_list)) {
2003 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2004 PAGE_SIZE, GFP_NOFS, false);
2007 head = msg->front.iov_base;
2008 head->num = cpu_to_le32(0);
2009 msg->front.iov_len = sizeof(*head);
2011 msg->hdr.version = cpu_to_le16(2);
2012 msg->hdr.compat_version = cpu_to_le16(1);
2015 cap = list_first_entry(&tmp_list, struct ceph_cap,
2017 list_del(&cap->session_caps);
2020 head = msg->front.iov_base;
2021 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2023 item = msg->front.iov_base + msg->front.iov_len;
2024 item->ino = cpu_to_le64(cap->cap_ino);
2025 item->cap_id = cpu_to_le64(cap->cap_id);
2026 item->migrate_seq = cpu_to_le32(cap->mseq);
2027 item->seq = cpu_to_le32(cap->issue_seq);
2028 msg->front.iov_len += sizeof(*item);
2030 ceph_put_cap(mdsc, cap);
2032 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2033 // Append cap_barrier field
2034 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2035 *cap_barrier = barrier;
2036 msg->front.iov_len += sizeof(*cap_barrier);
2038 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2039 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2040 ceph_con_send(&session->s_con, msg);
2045 BUG_ON(num_cap_releases != 0);
2047 spin_lock(&session->s_cap_lock);
2048 if (!list_empty(&session->s_cap_releases))
2050 spin_unlock(&session->s_cap_lock);
2053 // Append cap_barrier field
2054 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2055 *cap_barrier = barrier;
2056 msg->front.iov_len += sizeof(*cap_barrier);
2058 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2059 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2060 ceph_con_send(&session->s_con, msg);
2064 pr_err("send_cap_releases mds%d, failed to allocate message\n",
2066 spin_lock(&session->s_cap_lock);
2067 list_splice(&tmp_list, &session->s_cap_releases);
2068 session->s_num_cap_releases += num_cap_releases;
2069 spin_unlock(&session->s_cap_lock);
2072 static void ceph_cap_release_work(struct work_struct *work)
2074 struct ceph_mds_session *session =
2075 container_of(work, struct ceph_mds_session, s_cap_release_work);
2077 mutex_lock(&session->s_mutex);
2078 if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2079 session->s_state == CEPH_MDS_SESSION_HUNG)
2080 ceph_send_cap_releases(session->s_mdsc, session);
2081 mutex_unlock(&session->s_mutex);
2082 ceph_put_mds_session(session);
2085 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2086 struct ceph_mds_session *session)
2091 ceph_get_mds_session(session);
2092 if (queue_work(mdsc->fsc->cap_wq,
2093 &session->s_cap_release_work)) {
2094 dout("cap release work queued\n");
2096 ceph_put_mds_session(session);
2097 dout("failed to queue cap release work\n");
2102 * caller holds session->s_cap_lock
2104 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2105 struct ceph_cap *cap)
2107 list_add_tail(&cap->session_caps, &session->s_cap_releases);
2108 session->s_num_cap_releases++;
2110 if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2111 ceph_flush_cap_releases(session->s_mdsc, session);
2114 static void ceph_cap_reclaim_work(struct work_struct *work)
2116 struct ceph_mds_client *mdsc =
2117 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2118 int ret = ceph_trim_dentries(mdsc);
2120 ceph_queue_cap_reclaim_work(mdsc);
2123 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2128 if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2129 dout("caps reclaim work queued\n");
2131 dout("failed to queue caps release work\n");
2135 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2140 val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2141 if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2142 atomic_set(&mdsc->cap_reclaim_pending, 0);
2143 ceph_queue_cap_reclaim_work(mdsc);
2151 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2154 struct ceph_inode_info *ci = ceph_inode(dir);
2155 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2156 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2157 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2158 unsigned int num_entries;
2161 spin_lock(&ci->i_ceph_lock);
2162 num_entries = ci->i_files + ci->i_subdirs;
2163 spin_unlock(&ci->i_ceph_lock);
2164 num_entries = max(num_entries, 1U);
2165 num_entries = min(num_entries, opt->max_readdir);
2167 order = get_order(size * num_entries);
2168 while (order >= 0) {
2169 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2172 if (rinfo->dir_entries)
2176 if (!rinfo->dir_entries)
2179 num_entries = (PAGE_SIZE << order) / size;
2180 num_entries = min(num_entries, opt->max_readdir);
2182 rinfo->dir_buf_size = PAGE_SIZE << order;
2183 req->r_num_caps = num_entries + 1;
2184 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2185 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2190 * Create an mds request.
2192 struct ceph_mds_request *
2193 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2195 struct ceph_mds_request *req;
2197 req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2199 return ERR_PTR(-ENOMEM);
2201 mutex_init(&req->r_fill_mutex);
2203 req->r_started = jiffies;
2204 req->r_resend_mds = -1;
2205 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2206 INIT_LIST_HEAD(&req->r_unsafe_target_item);
2208 kref_init(&req->r_kref);
2209 RB_CLEAR_NODE(&req->r_node);
2210 INIT_LIST_HEAD(&req->r_wait);
2211 init_completion(&req->r_completion);
2212 init_completion(&req->r_safe_completion);
2213 INIT_LIST_HEAD(&req->r_unsafe_item);
2215 ktime_get_coarse_real_ts64(&req->r_stamp);
2218 req->r_direct_mode = mode;
2223 * return oldest (lowest) request, tid in request tree, 0 if none.
2225 * called under mdsc->mutex.
2227 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2229 if (RB_EMPTY_ROOT(&mdsc->request_tree))
2231 return rb_entry(rb_first(&mdsc->request_tree),
2232 struct ceph_mds_request, r_node);
2235 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2237 return mdsc->oldest_tid;
2241 * Build a dentry's path. Allocate on heap; caller must kfree. Based
2242 * on build_path_from_dentry in fs/cifs/dir.c.
2244 * If @stop_on_nosnap, generate path relative to the first non-snapped
2247 * Encode hidden .snap dirs as a double /, i.e.
2248 * foo/.snap/bar -> foo//bar
2250 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2253 struct dentry *temp;
2260 return ERR_PTR(-EINVAL);
2264 return ERR_PTR(-ENOMEM);
2269 seq = read_seqbegin(&rename_lock);
2273 struct inode *inode;
2275 spin_lock(&temp->d_lock);
2276 inode = d_inode(temp);
2277 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2278 dout("build_path path+%d: %p SNAPDIR\n",
2280 } else if (stop_on_nosnap && inode && dentry != temp &&
2281 ceph_snap(inode) == CEPH_NOSNAP) {
2282 spin_unlock(&temp->d_lock);
2283 pos++; /* get rid of any prepended '/' */
2286 pos -= temp->d_name.len;
2288 spin_unlock(&temp->d_lock);
2291 memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2293 spin_unlock(&temp->d_lock);
2294 temp = READ_ONCE(temp->d_parent);
2296 /* Are we at the root? */
2300 /* Are we out of buffer? */
2306 base = ceph_ino(d_inode(temp));
2309 if (read_seqretry(&rename_lock, seq))
2314 * A rename didn't occur, but somehow we didn't end up where
2315 * we thought we would. Throw a warning and try again.
2317 pr_warn("build_path did not end path lookup where "
2318 "expected, pos is %d\n", pos);
2323 *plen = PATH_MAX - 1 - pos;
2324 dout("build_path on %p %d built %llx '%.*s'\n",
2325 dentry, d_count(dentry), base, *plen, path + pos);
2329 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2330 const char **ppath, int *ppathlen, u64 *pino,
2331 bool *pfreepath, bool parent_locked)
2337 dir = d_inode_rcu(dentry->d_parent);
2338 if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2339 *pino = ceph_ino(dir);
2341 *ppath = dentry->d_name.name;
2342 *ppathlen = dentry->d_name.len;
2346 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2348 return PTR_ERR(path);
2354 static int build_inode_path(struct inode *inode,
2355 const char **ppath, int *ppathlen, u64 *pino,
2358 struct dentry *dentry;
2361 if (ceph_snap(inode) == CEPH_NOSNAP) {
2362 *pino = ceph_ino(inode);
2366 dentry = d_find_alias(inode);
2367 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2370 return PTR_ERR(path);
2377 * request arguments may be specified via an inode *, a dentry *, or
2378 * an explicit ino+path.
2380 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2381 struct inode *rdiri, const char *rpath,
2382 u64 rino, const char **ppath, int *pathlen,
2383 u64 *ino, bool *freepath, bool parent_locked)
2388 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2389 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2391 } else if (rdentry) {
2392 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2393 freepath, parent_locked);
2394 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2396 } else if (rpath || rino) {
2399 *pathlen = rpath ? strlen(rpath) : 0;
2400 dout(" path %.*s\n", *pathlen, rpath);
2407 * called under mdsc->mutex
2409 static struct ceph_msg *create_request_message(struct ceph_mds_client *mdsc,
2410 struct ceph_mds_request *req,
2411 int mds, bool drop_cap_releases)
2413 struct ceph_msg *msg;
2414 struct ceph_mds_request_head *head;
2415 const char *path1 = NULL;
2416 const char *path2 = NULL;
2417 u64 ino1 = 0, ino2 = 0;
2418 int pathlen1 = 0, pathlen2 = 0;
2419 bool freepath1 = false, freepath2 = false;
2425 ret = set_request_path_attr(req->r_inode, req->r_dentry,
2426 req->r_parent, req->r_path1, req->r_ino1.ino,
2427 &path1, &pathlen1, &ino1, &freepath1,
2428 test_bit(CEPH_MDS_R_PARENT_LOCKED,
2429 &req->r_req_flags));
2435 /* If r_old_dentry is set, then assume that its parent is locked */
2436 ret = set_request_path_attr(NULL, req->r_old_dentry,
2437 req->r_old_dentry_dir,
2438 req->r_path2, req->r_ino2.ino,
2439 &path2, &pathlen2, &ino2, &freepath2, true);
2445 len = sizeof(*head) +
2446 pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2447 sizeof(struct ceph_timespec);
2449 /* calculate (max) length for cap releases */
2450 len += sizeof(struct ceph_mds_request_release) *
2451 (!!req->r_inode_drop + !!req->r_dentry_drop +
2452 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2453 if (req->r_dentry_drop)
2455 if (req->r_old_dentry_drop)
2458 msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2460 msg = ERR_PTR(-ENOMEM);
2464 msg->hdr.version = cpu_to_le16(2);
2465 msg->hdr.tid = cpu_to_le64(req->r_tid);
2467 head = msg->front.iov_base;
2468 p = msg->front.iov_base + sizeof(*head);
2469 end = msg->front.iov_base + msg->front.iov_len;
2471 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2472 head->op = cpu_to_le32(req->r_op);
2473 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, req->r_uid));
2474 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, req->r_gid));
2475 head->ino = cpu_to_le64(req->r_deleg_ino);
2476 head->args = req->r_args;
2478 ceph_encode_filepath(&p, end, ino1, path1);
2479 ceph_encode_filepath(&p, end, ino2, path2);
2481 /* make note of release offset, in case we need to replay */
2482 req->r_request_release_offset = p - msg->front.iov_base;
2486 if (req->r_inode_drop)
2487 releases += ceph_encode_inode_release(&p,
2488 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2489 mds, req->r_inode_drop, req->r_inode_unless,
2490 req->r_op == CEPH_MDS_OP_READDIR);
2491 if (req->r_dentry_drop)
2492 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2493 req->r_parent, mds, req->r_dentry_drop,
2494 req->r_dentry_unless);
2495 if (req->r_old_dentry_drop)
2496 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2497 req->r_old_dentry_dir, mds,
2498 req->r_old_dentry_drop,
2499 req->r_old_dentry_unless);
2500 if (req->r_old_inode_drop)
2501 releases += ceph_encode_inode_release(&p,
2502 d_inode(req->r_old_dentry),
2503 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2505 if (drop_cap_releases) {
2507 p = msg->front.iov_base + req->r_request_release_offset;
2510 head->num_releases = cpu_to_le16(releases);
2514 struct ceph_timespec ts;
2515 ceph_encode_timespec64(&ts, &req->r_stamp);
2516 ceph_encode_copy(&p, &ts, sizeof(ts));
2520 msg->front.iov_len = p - msg->front.iov_base;
2521 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2523 if (req->r_pagelist) {
2524 struct ceph_pagelist *pagelist = req->r_pagelist;
2525 ceph_msg_data_add_pagelist(msg, pagelist);
2526 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2528 msg->hdr.data_len = 0;
2531 msg->hdr.data_off = cpu_to_le16(0);
2535 ceph_mdsc_free_path((char *)path2, pathlen2);
2538 ceph_mdsc_free_path((char *)path1, pathlen1);
2544 * called under mdsc->mutex if error, under no mutex if
2547 static void complete_request(struct ceph_mds_client *mdsc,
2548 struct ceph_mds_request *req)
2550 if (req->r_callback)
2551 req->r_callback(mdsc, req);
2552 complete_all(&req->r_completion);
2556 * called under mdsc->mutex
2558 static int __prepare_send_request(struct ceph_mds_client *mdsc,
2559 struct ceph_mds_request *req,
2560 int mds, bool drop_cap_releases)
2562 struct ceph_mds_request_head *rhead;
2563 struct ceph_msg *msg;
2568 struct ceph_cap *cap =
2569 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2572 req->r_sent_on_mseq = cap->mseq;
2574 req->r_sent_on_mseq = -1;
2576 dout("prepare_send_request %p tid %lld %s (attempt %d)\n", req,
2577 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2579 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2582 * Replay. Do not regenerate message (and rebuild
2583 * paths, etc.); just use the original message.
2584 * Rebuilding paths will break for renames because
2585 * d_move mangles the src name.
2587 msg = req->r_request;
2588 rhead = msg->front.iov_base;
2590 flags = le32_to_cpu(rhead->flags);
2591 flags |= CEPH_MDS_FLAG_REPLAY;
2592 rhead->flags = cpu_to_le32(flags);
2594 if (req->r_target_inode)
2595 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2597 rhead->num_retry = req->r_attempts - 1;
2599 /* remove cap/dentry releases from message */
2600 rhead->num_releases = 0;
2603 p = msg->front.iov_base + req->r_request_release_offset;
2605 struct ceph_timespec ts;
2606 ceph_encode_timespec64(&ts, &req->r_stamp);
2607 ceph_encode_copy(&p, &ts, sizeof(ts));
2610 msg->front.iov_len = p - msg->front.iov_base;
2611 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2615 if (req->r_request) {
2616 ceph_msg_put(req->r_request);
2617 req->r_request = NULL;
2619 msg = create_request_message(mdsc, req, mds, drop_cap_releases);
2621 req->r_err = PTR_ERR(msg);
2622 return PTR_ERR(msg);
2624 req->r_request = msg;
2626 rhead = msg->front.iov_base;
2627 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2628 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2629 flags |= CEPH_MDS_FLAG_REPLAY;
2630 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2631 flags |= CEPH_MDS_FLAG_ASYNC;
2633 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2634 rhead->flags = cpu_to_le32(flags);
2635 rhead->num_fwd = req->r_num_fwd;
2636 rhead->num_retry = req->r_attempts - 1;
2638 dout(" r_parent = %p\n", req->r_parent);
2643 * called under mdsc->mutex
2645 static int __send_request(struct ceph_mds_client *mdsc,
2646 struct ceph_mds_session *session,
2647 struct ceph_mds_request *req,
2648 bool drop_cap_releases)
2652 err = __prepare_send_request(mdsc, req, session->s_mds,
2655 ceph_msg_get(req->r_request);
2656 ceph_con_send(&session->s_con, req->r_request);
2663 * send request, or put it on the appropriate wait list.
2665 static void __do_request(struct ceph_mds_client *mdsc,
2666 struct ceph_mds_request *req)
2668 struct ceph_mds_session *session = NULL;
2673 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2674 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2675 __unregister_request(mdsc, req);
2679 if (req->r_timeout &&
2680 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2681 dout("do_request timed out\n");
2685 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2686 dout("do_request forced umount\n");
2690 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2691 if (mdsc->mdsmap_err) {
2692 err = mdsc->mdsmap_err;
2693 dout("do_request mdsmap err %d\n", err);
2696 if (mdsc->mdsmap->m_epoch == 0) {
2697 dout("do_request no mdsmap, waiting for map\n");
2698 list_add(&req->r_wait, &mdsc->waiting_for_map);
2701 if (!(mdsc->fsc->mount_options->flags &
2702 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2703 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2704 err = -EHOSTUNREACH;
2709 put_request_session(req);
2711 mds = __choose_mds(mdsc, req, &random);
2713 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2714 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2718 dout("do_request no mds or not active, waiting for map\n");
2719 list_add(&req->r_wait, &mdsc->waiting_for_map);
2723 /* get, open session */
2724 session = __ceph_lookup_mds_session(mdsc, mds);
2726 session = register_session(mdsc, mds);
2727 if (IS_ERR(session)) {
2728 err = PTR_ERR(session);
2732 req->r_session = ceph_get_mds_session(session);
2734 dout("do_request mds%d session %p state %s\n", mds, session,
2735 ceph_session_state_name(session->s_state));
2736 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2737 session->s_state != CEPH_MDS_SESSION_HUNG) {
2738 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2743 * We cannot queue async requests since the caps and delegated
2744 * inodes are bound to the session. Just return -EJUKEBOX and
2745 * let the caller retry a sync request in that case.
2747 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2751 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2752 session->s_state == CEPH_MDS_SESSION_CLOSING) {
2753 __open_session(mdsc, session);
2754 /* retry the same mds later */
2756 req->r_resend_mds = mds;
2758 list_add(&req->r_wait, &session->s_waiting);
2763 req->r_resend_mds = -1; /* forget any previous mds hint */
2765 if (req->r_request_started == 0) /* note request start time */
2766 req->r_request_started = jiffies;
2768 err = __send_request(mdsc, session, req, false);
2771 ceph_put_mds_session(session);
2774 dout("__do_request early error %d\n", err);
2776 complete_request(mdsc, req);
2777 __unregister_request(mdsc, req);
2783 * called under mdsc->mutex
2785 static void __wake_requests(struct ceph_mds_client *mdsc,
2786 struct list_head *head)
2788 struct ceph_mds_request *req;
2789 LIST_HEAD(tmp_list);
2791 list_splice_init(head, &tmp_list);
2793 while (!list_empty(&tmp_list)) {
2794 req = list_entry(tmp_list.next,
2795 struct ceph_mds_request, r_wait);
2796 list_del_init(&req->r_wait);
2797 dout(" wake request %p tid %llu\n", req, req->r_tid);
2798 __do_request(mdsc, req);
2803 * Wake up threads with requests pending for @mds, so that they can
2804 * resubmit their requests to a possibly different mds.
2806 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2808 struct ceph_mds_request *req;
2809 struct rb_node *p = rb_first(&mdsc->request_tree);
2811 dout("kick_requests mds%d\n", mds);
2813 req = rb_entry(p, struct ceph_mds_request, r_node);
2815 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2817 if (req->r_attempts > 0)
2818 continue; /* only new requests */
2819 if (req->r_session &&
2820 req->r_session->s_mds == mds) {
2821 dout(" kicking tid %llu\n", req->r_tid);
2822 list_del_init(&req->r_wait);
2823 __do_request(mdsc, req);
2828 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2829 struct ceph_mds_request *req)
2833 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2835 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2836 if (req->r_parent) {
2837 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
2838 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
2839 CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
2840 spin_lock(&ci->i_ceph_lock);
2841 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
2842 __ceph_touch_fmode(ci, mdsc, fmode);
2843 spin_unlock(&ci->i_ceph_lock);
2844 ihold(req->r_parent);
2846 if (req->r_old_dentry_dir)
2847 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2851 err = ceph_wait_on_async_create(req->r_inode);
2853 dout("%s: wait for async create returned: %d\n",
2859 if (!err && req->r_old_inode) {
2860 err = ceph_wait_on_async_create(req->r_old_inode);
2862 dout("%s: wait for async create returned: %d\n",
2868 dout("submit_request on %p for inode %p\n", req, dir);
2869 mutex_lock(&mdsc->mutex);
2870 __register_request(mdsc, req, dir);
2871 __do_request(mdsc, req);
2873 mutex_unlock(&mdsc->mutex);
2877 static int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2878 struct ceph_mds_request *req)
2883 dout("do_request waiting\n");
2884 if (!req->r_timeout && req->r_wait_for_completion) {
2885 err = req->r_wait_for_completion(mdsc, req);
2887 long timeleft = wait_for_completion_killable_timeout(
2889 ceph_timeout_jiffies(req->r_timeout));
2893 err = -ETIMEDOUT; /* timed out */
2895 err = timeleft; /* killed */
2897 dout("do_request waited, got %d\n", err);
2898 mutex_lock(&mdsc->mutex);
2900 /* only abort if we didn't race with a real reply */
2901 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2902 err = le32_to_cpu(req->r_reply_info.head->result);
2903 } else if (err < 0) {
2904 dout("aborted request %lld with %d\n", req->r_tid, err);
2907 * ensure we aren't running concurrently with
2908 * ceph_fill_trace or ceph_readdir_prepopulate, which
2909 * rely on locks (dir mutex) held by our caller.
2911 mutex_lock(&req->r_fill_mutex);
2913 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2914 mutex_unlock(&req->r_fill_mutex);
2916 if (req->r_parent &&
2917 (req->r_op & CEPH_MDS_OP_WRITE))
2918 ceph_invalidate_dir_request(req);
2923 mutex_unlock(&mdsc->mutex);
2928 * Synchrously perform an mds request. Take care of all of the
2929 * session setup, forwarding, retry details.
2931 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2933 struct ceph_mds_request *req)
2937 dout("do_request on %p\n", req);
2940 err = ceph_mdsc_submit_request(mdsc, dir, req);
2942 err = ceph_mdsc_wait_request(mdsc, req);
2943 dout("do_request %p done, result %d\n", req, err);
2948 * Invalidate dir's completeness, dentry lease state on an aborted MDS
2949 * namespace request.
2951 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
2953 struct inode *dir = req->r_parent;
2954 struct inode *old_dir = req->r_old_dentry_dir;
2956 dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
2958 ceph_dir_clear_complete(dir);
2960 ceph_dir_clear_complete(old_dir);
2962 ceph_invalidate_dentry_lease(req->r_dentry);
2963 if (req->r_old_dentry)
2964 ceph_invalidate_dentry_lease(req->r_old_dentry);
2970 * We take the session mutex and parse and process the reply immediately.
2971 * This preserves the logical ordering of replies, capabilities, etc., sent
2972 * by the MDS as they are applied to our local cache.
2974 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
2976 struct ceph_mds_client *mdsc = session->s_mdsc;
2977 struct ceph_mds_request *req;
2978 struct ceph_mds_reply_head *head = msg->front.iov_base;
2979 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
2980 struct ceph_snap_realm *realm;
2983 int mds = session->s_mds;
2985 if (msg->front.iov_len < sizeof(*head)) {
2986 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
2991 /* get request, session */
2992 tid = le64_to_cpu(msg->hdr.tid);
2993 mutex_lock(&mdsc->mutex);
2994 req = lookup_get_request(mdsc, tid);
2996 dout("handle_reply on unknown tid %llu\n", tid);
2997 mutex_unlock(&mdsc->mutex);
3000 dout("handle_reply %p\n", req);
3002 /* correct session? */
3003 if (req->r_session != session) {
3004 pr_err("mdsc_handle_reply got %llu on session mds%d"
3005 " not mds%d\n", tid, session->s_mds,
3006 req->r_session ? req->r_session->s_mds : -1);
3007 mutex_unlock(&mdsc->mutex);
3012 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3013 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3014 pr_warn("got a dup %s reply on %llu from mds%d\n",
3015 head->safe ? "safe" : "unsafe", tid, mds);
3016 mutex_unlock(&mdsc->mutex);
3019 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3020 pr_warn("got unsafe after safe on %llu from mds%d\n",
3022 mutex_unlock(&mdsc->mutex);
3026 result = le32_to_cpu(head->result);
3030 * if we're not talking to the authority, send to them
3031 * if the authority has changed while we weren't looking,
3032 * send to new authority
3033 * Otherwise we just have to return an ESTALE
3035 if (result == -ESTALE) {
3036 dout("got ESTALE on request %llu\n", req->r_tid);
3037 req->r_resend_mds = -1;
3038 if (req->r_direct_mode != USE_AUTH_MDS) {
3039 dout("not using auth, setting for that now\n");
3040 req->r_direct_mode = USE_AUTH_MDS;
3041 __do_request(mdsc, req);
3042 mutex_unlock(&mdsc->mutex);
3045 int mds = __choose_mds(mdsc, req, NULL);
3046 if (mds >= 0 && mds != req->r_session->s_mds) {
3047 dout("but auth changed, so resending\n");
3048 __do_request(mdsc, req);
3049 mutex_unlock(&mdsc->mutex);
3053 dout("have to return ESTALE on request %llu\n", req->r_tid);
3058 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3059 __unregister_request(mdsc, req);
3061 /* last request during umount? */
3062 if (mdsc->stopping && !__get_oldest_req(mdsc))
3063 complete_all(&mdsc->safe_umount_waiters);
3065 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3067 * We already handled the unsafe response, now do the
3068 * cleanup. No need to examine the response; the MDS
3069 * doesn't include any result info in the safe
3070 * response. And even if it did, there is nothing
3071 * useful we could do with a revised return value.
3073 dout("got safe reply %llu, mds%d\n", tid, mds);
3075 mutex_unlock(&mdsc->mutex);
3079 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3080 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3083 dout("handle_reply tid %lld result %d\n", tid, result);
3084 rinfo = &req->r_reply_info;
3085 if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3086 err = parse_reply_info(session, msg, rinfo, (u64)-1);
3088 err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3089 mutex_unlock(&mdsc->mutex);
3091 mutex_lock(&session->s_mutex);
3093 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3100 if (rinfo->snapblob_len) {
3101 down_write(&mdsc->snap_rwsem);
3102 ceph_update_snap_trace(mdsc, rinfo->snapblob,
3103 rinfo->snapblob + rinfo->snapblob_len,
3104 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3106 downgrade_write(&mdsc->snap_rwsem);
3108 down_read(&mdsc->snap_rwsem);
3111 /* insert trace into our cache */
3112 mutex_lock(&req->r_fill_mutex);
3113 current->journal_info = req;
3114 err = ceph_fill_trace(mdsc->fsc->sb, req);
3116 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3117 req->r_op == CEPH_MDS_OP_LSSNAP))
3118 ceph_readdir_prepopulate(req, req->r_session);
3120 current->journal_info = NULL;
3121 mutex_unlock(&req->r_fill_mutex);
3123 up_read(&mdsc->snap_rwsem);
3125 ceph_put_snap_realm(mdsc, realm);
3128 if (req->r_target_inode &&
3129 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3130 struct ceph_inode_info *ci =
3131 ceph_inode(req->r_target_inode);
3132 spin_lock(&ci->i_unsafe_lock);
3133 list_add_tail(&req->r_unsafe_target_item,
3134 &ci->i_unsafe_iops);
3135 spin_unlock(&ci->i_unsafe_lock);
3138 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3141 mutex_lock(&mdsc->mutex);
3142 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3146 req->r_reply = ceph_msg_get(msg);
3147 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3150 dout("reply arrived after request %lld was aborted\n", tid);
3152 mutex_unlock(&mdsc->mutex);
3154 mutex_unlock(&session->s_mutex);
3156 /* kick calling process */
3157 complete_request(mdsc, req);
3159 ceph_mdsc_put_request(req);
3166 * handle mds notification that our request has been forwarded.
3168 static void handle_forward(struct ceph_mds_client *mdsc,
3169 struct ceph_mds_session *session,
3170 struct ceph_msg *msg)
3172 struct ceph_mds_request *req;
3173 u64 tid = le64_to_cpu(msg->hdr.tid);
3177 void *p = msg->front.iov_base;
3178 void *end = p + msg->front.iov_len;
3180 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3181 next_mds = ceph_decode_32(&p);
3182 fwd_seq = ceph_decode_32(&p);
3184 mutex_lock(&mdsc->mutex);
3185 req = lookup_get_request(mdsc, tid);
3187 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3188 goto out; /* dup reply? */
3191 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3192 dout("forward tid %llu aborted, unregistering\n", tid);
3193 __unregister_request(mdsc, req);
3194 } else if (fwd_seq <= req->r_num_fwd) {
3195 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3196 tid, next_mds, req->r_num_fwd, fwd_seq);
3198 /* resend. forward race not possible; mds would drop */
3199 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3201 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3202 req->r_attempts = 0;
3203 req->r_num_fwd = fwd_seq;
3204 req->r_resend_mds = next_mds;
3205 put_request_session(req);
3206 __do_request(mdsc, req);
3208 ceph_mdsc_put_request(req);
3210 mutex_unlock(&mdsc->mutex);
3214 pr_err("mdsc_handle_forward decode error err=%d\n", err);
3217 static int __decode_session_metadata(void **p, void *end,
3220 /* map<string,string> */
3223 ceph_decode_32_safe(p, end, n, bad);
3226 ceph_decode_32_safe(p, end, len, bad);
3227 ceph_decode_need(p, end, len, bad);
3228 err_str = !strncmp(*p, "error_string", len);
3230 ceph_decode_32_safe(p, end, len, bad);
3231 ceph_decode_need(p, end, len, bad);
3232 if (err_str && strnstr(*p, "blacklisted", len))
3233 *blacklisted = true;
3242 * handle a mds session control message
3244 static void handle_session(struct ceph_mds_session *session,
3245 struct ceph_msg *msg)
3247 struct ceph_mds_client *mdsc = session->s_mdsc;
3248 int mds = session->s_mds;
3249 int msg_version = le16_to_cpu(msg->hdr.version);
3250 void *p = msg->front.iov_base;
3251 void *end = p + msg->front.iov_len;
3252 struct ceph_mds_session_head *h;
3254 u64 seq, features = 0;
3256 bool blacklisted = false;
3259 ceph_decode_need(&p, end, sizeof(*h), bad);
3263 op = le32_to_cpu(h->op);
3264 seq = le64_to_cpu(h->seq);
3266 if (msg_version >= 3) {
3268 /* version >= 2, metadata */
3269 if (__decode_session_metadata(&p, end, &blacklisted) < 0)
3271 /* version >= 3, feature bits */
3272 ceph_decode_32_safe(&p, end, len, bad);
3273 ceph_decode_64_safe(&p, end, features, bad);
3274 p += len - sizeof(features);
3277 mutex_lock(&mdsc->mutex);
3278 if (op == CEPH_SESSION_CLOSE) {
3279 ceph_get_mds_session(session);
3280 __unregister_session(mdsc, session);
3282 /* FIXME: this ttl calculation is generous */
3283 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3284 mutex_unlock(&mdsc->mutex);
3286 mutex_lock(&session->s_mutex);
3288 dout("handle_session mds%d %s %p state %s seq %llu\n",
3289 mds, ceph_session_op_name(op), session,
3290 ceph_session_state_name(session->s_state), seq);
3292 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3293 session->s_state = CEPH_MDS_SESSION_OPEN;
3294 pr_info("mds%d came back\n", session->s_mds);
3298 case CEPH_SESSION_OPEN:
3299 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3300 pr_info("mds%d reconnect success\n", session->s_mds);
3301 session->s_state = CEPH_MDS_SESSION_OPEN;
3302 session->s_features = features;
3303 renewed_caps(mdsc, session, 0);
3306 __close_session(mdsc, session);
3309 case CEPH_SESSION_RENEWCAPS:
3310 if (session->s_renew_seq == seq)
3311 renewed_caps(mdsc, session, 1);
3314 case CEPH_SESSION_CLOSE:
3315 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3316 pr_info("mds%d reconnect denied\n", session->s_mds);
3317 session->s_state = CEPH_MDS_SESSION_CLOSED;
3318 cleanup_session_requests(mdsc, session);
3319 remove_session_caps(session);
3320 wake = 2; /* for good measure */
3321 wake_up_all(&mdsc->session_close_wq);
3324 case CEPH_SESSION_STALE:
3325 pr_info("mds%d caps went stale, renewing\n",
3327 spin_lock(&session->s_gen_ttl_lock);
3328 session->s_cap_gen++;
3329 session->s_cap_ttl = jiffies - 1;
3330 spin_unlock(&session->s_gen_ttl_lock);
3331 send_renew_caps(mdsc, session);
3334 case CEPH_SESSION_RECALL_STATE:
3335 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3338 case CEPH_SESSION_FLUSHMSG:
3339 send_flushmsg_ack(mdsc, session, seq);
3342 case CEPH_SESSION_FORCE_RO:
3343 dout("force_session_readonly %p\n", session);
3344 spin_lock(&session->s_cap_lock);
3345 session->s_readonly = true;
3346 spin_unlock(&session->s_cap_lock);
3347 wake_up_session_caps(session, FORCE_RO);
3350 case CEPH_SESSION_REJECT:
3351 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3352 pr_info("mds%d rejected session\n", session->s_mds);
3353 session->s_state = CEPH_MDS_SESSION_REJECTED;
3354 cleanup_session_requests(mdsc, session);
3355 remove_session_caps(session);
3357 mdsc->fsc->blacklisted = true;
3358 wake = 2; /* for good measure */
3362 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3366 mutex_unlock(&session->s_mutex);
3368 mutex_lock(&mdsc->mutex);
3369 __wake_requests(mdsc, &session->s_waiting);
3371 kick_requests(mdsc, mds);
3372 mutex_unlock(&mdsc->mutex);
3374 if (op == CEPH_SESSION_CLOSE)
3375 ceph_put_mds_session(session);
3379 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3380 (int)msg->front.iov_len);
3385 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3389 dcaps = xchg(&req->r_dir_caps, 0);
3391 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3392 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3397 * called under session->mutex.
3399 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3400 struct ceph_mds_session *session)
3402 struct ceph_mds_request *req, *nreq;
3405 dout("replay_unsafe_requests mds%d\n", session->s_mds);
3407 mutex_lock(&mdsc->mutex);
3408 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3409 __send_request(mdsc, session, req, true);
3412 * also re-send old requests when MDS enters reconnect stage. So that MDS
3413 * can process completed request in clientreplay stage.
3415 p = rb_first(&mdsc->request_tree);
3417 req = rb_entry(p, struct ceph_mds_request, r_node);
3419 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3421 if (req->r_attempts == 0)
3422 continue; /* only old requests */
3423 if (!req->r_session)
3425 if (req->r_session->s_mds != session->s_mds)
3428 ceph_mdsc_release_dir_caps(req);
3430 __send_request(mdsc, session, req, true);
3432 mutex_unlock(&mdsc->mutex);
3435 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3437 struct ceph_msg *reply;
3438 struct ceph_pagelist *_pagelist;
3443 if (!recon_state->allow_multi)
3446 /* can't handle message that contains both caps and realm */
3447 BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3449 /* pre-allocate new pagelist */
3450 _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3454 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3458 /* placeholder for nr_caps */
3459 err = ceph_pagelist_encode_32(_pagelist, 0);
3463 if (recon_state->nr_caps) {
3464 /* currently encoding caps */
3465 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3469 /* placeholder for nr_realms (currently encoding relams) */
3470 err = ceph_pagelist_encode_32(_pagelist, 0);
3475 err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3479 page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3480 addr = kmap_atomic(page);
3481 if (recon_state->nr_caps) {
3482 /* currently encoding caps */
3483 *addr = cpu_to_le32(recon_state->nr_caps);
3485 /* currently encoding relams */
3486 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3488 kunmap_atomic(addr);
3490 reply->hdr.version = cpu_to_le16(5);
3491 reply->hdr.compat_version = cpu_to_le16(4);
3493 reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3494 ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3496 ceph_con_send(&recon_state->session->s_con, reply);
3497 ceph_pagelist_release(recon_state->pagelist);
3499 recon_state->pagelist = _pagelist;
3500 recon_state->nr_caps = 0;
3501 recon_state->nr_realms = 0;
3502 recon_state->msg_version = 5;
3505 ceph_msg_put(reply);
3507 ceph_pagelist_release(_pagelist);
3512 * Encode information about a cap for a reconnect with the MDS.
3514 static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3518 struct ceph_mds_cap_reconnect v2;
3519 struct ceph_mds_cap_reconnect_v1 v1;
3521 struct ceph_inode_info *ci = cap->ci;
3522 struct ceph_reconnect_state *recon_state = arg;
3523 struct ceph_pagelist *pagelist = recon_state->pagelist;
3527 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3528 inode, ceph_vinop(inode), cap, cap->cap_id,
3529 ceph_cap_string(cap->issued));
3531 spin_lock(&ci->i_ceph_lock);
3532 cap->seq = 0; /* reset cap seq */
3533 cap->issue_seq = 0; /* and issue_seq */
3534 cap->mseq = 0; /* and migrate_seq */
3535 cap->cap_gen = cap->session->s_cap_gen;
3537 /* These are lost when the session goes away */
3538 if (S_ISDIR(inode->i_mode)) {
3539 if (cap->issued & CEPH_CAP_DIR_CREATE) {
3540 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3541 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3543 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3546 if (recon_state->msg_version >= 2) {
3547 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3548 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3549 rec.v2.issued = cpu_to_le32(cap->issued);
3550 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3551 rec.v2.pathbase = 0;
3552 rec.v2.flock_len = (__force __le32)
3553 ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3555 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3556 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3557 rec.v1.issued = cpu_to_le32(cap->issued);
3558 rec.v1.size = cpu_to_le64(inode->i_size);
3559 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3560 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3561 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3562 rec.v1.pathbase = 0;
3565 if (list_empty(&ci->i_cap_snaps)) {
3566 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3568 struct ceph_cap_snap *capsnap =
3569 list_first_entry(&ci->i_cap_snaps,
3570 struct ceph_cap_snap, ci_item);
3571 snap_follows = capsnap->follows;
3573 spin_unlock(&ci->i_ceph_lock);
3575 if (recon_state->msg_version >= 2) {
3576 int num_fcntl_locks, num_flock_locks;
3577 struct ceph_filelock *flocks = NULL;
3578 size_t struct_len, total_len = sizeof(u64);
3582 if (rec.v2.flock_len) {
3583 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3585 num_fcntl_locks = 0;
3586 num_flock_locks = 0;
3588 if (num_fcntl_locks + num_flock_locks > 0) {
3589 flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3590 sizeof(struct ceph_filelock),
3596 err = ceph_encode_locks_to_buffer(inode, flocks,
3611 if (recon_state->msg_version >= 3) {
3612 /* version, compat_version and struct_len */
3613 total_len += 2 * sizeof(u8) + sizeof(u32);
3617 * number of encoded locks is stable, so copy to pagelist
3619 struct_len = 2 * sizeof(u32) +
3620 (num_fcntl_locks + num_flock_locks) *
3621 sizeof(struct ceph_filelock);
3622 rec.v2.flock_len = cpu_to_le32(struct_len);
3624 struct_len += sizeof(u32) + sizeof(rec.v2);
3627 struct_len += sizeof(u64); /* snap_follows */
3629 total_len += struct_len;
3631 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3632 err = send_reconnect_partial(recon_state);
3634 goto out_freeflocks;
3635 pagelist = recon_state->pagelist;
3638 err = ceph_pagelist_reserve(pagelist, total_len);
3640 goto out_freeflocks;
3642 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3643 if (recon_state->msg_version >= 3) {
3644 ceph_pagelist_encode_8(pagelist, struct_v);
3645 ceph_pagelist_encode_8(pagelist, 1);
3646 ceph_pagelist_encode_32(pagelist, struct_len);
3648 ceph_pagelist_encode_string(pagelist, NULL, 0);
3649 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3650 ceph_locks_to_pagelist(flocks, pagelist,
3651 num_fcntl_locks, num_flock_locks);
3653 ceph_pagelist_encode_64(pagelist, snap_follows);
3660 struct dentry *dentry;
3662 dentry = d_find_alias(inode);
3664 path = ceph_mdsc_build_path(dentry,
3665 &pathlen, &pathbase, 0);
3668 err = PTR_ERR(path);
3671 rec.v1.pathbase = cpu_to_le64(pathbase);
3674 err = ceph_pagelist_reserve(pagelist,
3675 sizeof(u64) + sizeof(u32) +
3676 pathlen + sizeof(rec.v1));
3681 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3682 ceph_pagelist_encode_string(pagelist, path, pathlen);
3683 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3685 ceph_mdsc_free_path(path, pathlen);
3690 recon_state->nr_caps++;
3694 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3695 struct ceph_reconnect_state *recon_state)
3698 struct ceph_pagelist *pagelist = recon_state->pagelist;
3701 if (recon_state->msg_version >= 4) {
3702 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3708 * snaprealms. we provide mds with the ino, seq (version), and
3709 * parent for all of our realms. If the mds has any newer info,
3712 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3713 struct ceph_snap_realm *realm =
3714 rb_entry(p, struct ceph_snap_realm, node);
3715 struct ceph_mds_snaprealm_reconnect sr_rec;
3717 if (recon_state->msg_version >= 4) {
3718 size_t need = sizeof(u8) * 2 + sizeof(u32) +
3721 if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3722 err = send_reconnect_partial(recon_state);
3725 pagelist = recon_state->pagelist;
3728 err = ceph_pagelist_reserve(pagelist, need);
3732 ceph_pagelist_encode_8(pagelist, 1);
3733 ceph_pagelist_encode_8(pagelist, 1);
3734 ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3737 dout(" adding snap realm %llx seq %lld parent %llx\n",
3738 realm->ino, realm->seq, realm->parent_ino);
3739 sr_rec.ino = cpu_to_le64(realm->ino);
3740 sr_rec.seq = cpu_to_le64(realm->seq);
3741 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3743 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3747 recon_state->nr_realms++;
3755 * If an MDS fails and recovers, clients need to reconnect in order to
3756 * reestablish shared state. This includes all caps issued through
3757 * this session _and_ the snap_realm hierarchy. Because it's not
3758 * clear which snap realms the mds cares about, we send everything we
3759 * know about.. that ensures we'll then get any new info the
3760 * recovering MDS might have.
3762 * This is a relatively heavyweight operation, but it's rare.
3764 * called with mdsc->mutex held.
3766 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3767 struct ceph_mds_session *session)
3769 struct ceph_msg *reply;
3770 int mds = session->s_mds;
3772 struct ceph_reconnect_state recon_state = {
3777 pr_info("mds%d reconnect start\n", mds);
3779 recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3780 if (!recon_state.pagelist)
3781 goto fail_nopagelist;
3783 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3787 xa_destroy(&session->s_delegated_inos);
3789 mutex_lock(&session->s_mutex);
3790 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
3793 dout("session %p state %s\n", session,
3794 ceph_session_state_name(session->s_state));
3796 spin_lock(&session->s_gen_ttl_lock);
3797 session->s_cap_gen++;
3798 spin_unlock(&session->s_gen_ttl_lock);
3800 spin_lock(&session->s_cap_lock);
3801 /* don't know if session is readonly */
3802 session->s_readonly = 0;
3804 * notify __ceph_remove_cap() that we are composing cap reconnect.
3805 * If a cap get released before being added to the cap reconnect,
3806 * __ceph_remove_cap() should skip queuing cap release.
3808 session->s_cap_reconnect = 1;
3809 /* drop old cap expires; we're about to reestablish that state */
3810 detach_cap_releases(session, &dispose);
3811 spin_unlock(&session->s_cap_lock);
3812 dispose_cap_releases(mdsc, &dispose);
3814 /* trim unused caps to reduce MDS's cache rejoin time */
3815 if (mdsc->fsc->sb->s_root)
3816 shrink_dcache_parent(mdsc->fsc->sb->s_root);
3818 ceph_con_close(&session->s_con);
3819 ceph_con_open(&session->s_con,
3820 CEPH_ENTITY_TYPE_MDS, mds,
3821 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
3823 /* replay unsafe requests */
3824 replay_unsafe_requests(mdsc, session);
3826 ceph_early_kick_flushing_caps(mdsc, session);
3828 down_read(&mdsc->snap_rwsem);
3830 /* placeholder for nr_caps */
3831 err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
3835 if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
3836 recon_state.msg_version = 3;
3837 recon_state.allow_multi = true;
3838 } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
3839 recon_state.msg_version = 3;
3841 recon_state.msg_version = 2;
3843 /* trsaverse this session's caps */
3844 err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
3846 spin_lock(&session->s_cap_lock);
3847 session->s_cap_reconnect = 0;
3848 spin_unlock(&session->s_cap_lock);
3853 /* check if all realms can be encoded into current message */
3854 if (mdsc->num_snap_realms) {
3856 recon_state.pagelist->length +
3857 mdsc->num_snap_realms *
3858 sizeof(struct ceph_mds_snaprealm_reconnect);
3859 if (recon_state.msg_version >= 4) {
3860 /* number of realms */
3861 total_len += sizeof(u32);
3862 /* version, compat_version and struct_len */
3863 total_len += mdsc->num_snap_realms *
3864 (2 * sizeof(u8) + sizeof(u32));
3866 if (total_len > RECONNECT_MAX_SIZE) {
3867 if (!recon_state.allow_multi) {
3871 if (recon_state.nr_caps) {
3872 err = send_reconnect_partial(&recon_state);
3876 recon_state.msg_version = 5;
3880 err = encode_snap_realms(mdsc, &recon_state);
3884 if (recon_state.msg_version >= 5) {
3885 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
3890 if (recon_state.nr_caps || recon_state.nr_realms) {
3892 list_first_entry(&recon_state.pagelist->head,
3894 __le32 *addr = kmap_atomic(page);
3895 if (recon_state.nr_caps) {
3896 WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
3897 *addr = cpu_to_le32(recon_state.nr_caps);
3898 } else if (recon_state.msg_version >= 4) {
3899 *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
3901 kunmap_atomic(addr);
3904 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
3905 if (recon_state.msg_version >= 4)
3906 reply->hdr.compat_version = cpu_to_le16(4);
3908 reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
3909 ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
3911 ceph_con_send(&session->s_con, reply);
3913 mutex_unlock(&session->s_mutex);
3915 mutex_lock(&mdsc->mutex);
3916 __wake_requests(mdsc, &session->s_waiting);
3917 mutex_unlock(&mdsc->mutex);
3919 up_read(&mdsc->snap_rwsem);
3920 ceph_pagelist_release(recon_state.pagelist);
3924 ceph_msg_put(reply);
3925 up_read(&mdsc->snap_rwsem);
3926 mutex_unlock(&session->s_mutex);
3928 ceph_pagelist_release(recon_state.pagelist);
3930 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
3936 * compare old and new mdsmaps, kicking requests
3937 * and closing out old connections as necessary
3939 * called under mdsc->mutex.
3941 static void check_new_map(struct ceph_mds_client *mdsc,
3942 struct ceph_mdsmap *newmap,
3943 struct ceph_mdsmap *oldmap)
3946 int oldstate, newstate;
3947 struct ceph_mds_session *s;
3949 dout("check_new_map new %u old %u\n",
3950 newmap->m_epoch, oldmap->m_epoch);
3952 for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
3953 if (!mdsc->sessions[i])
3955 s = mdsc->sessions[i];
3956 oldstate = ceph_mdsmap_get_state(oldmap, i);
3957 newstate = ceph_mdsmap_get_state(newmap, i);
3959 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
3960 i, ceph_mds_state_name(oldstate),
3961 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
3962 ceph_mds_state_name(newstate),
3963 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
3964 ceph_session_state_name(s->s_state));
3966 if (i >= newmap->possible_max_rank) {
3967 /* force close session for stopped mds */
3968 ceph_get_mds_session(s);
3969 __unregister_session(mdsc, s);
3970 __wake_requests(mdsc, &s->s_waiting);
3971 mutex_unlock(&mdsc->mutex);
3973 mutex_lock(&s->s_mutex);
3974 cleanup_session_requests(mdsc, s);
3975 remove_session_caps(s);
3976 mutex_unlock(&s->s_mutex);
3978 ceph_put_mds_session(s);
3980 mutex_lock(&mdsc->mutex);
3981 kick_requests(mdsc, i);
3985 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
3986 ceph_mdsmap_get_addr(newmap, i),
3987 sizeof(struct ceph_entity_addr))) {
3989 mutex_unlock(&mdsc->mutex);
3990 mutex_lock(&s->s_mutex);
3991 mutex_lock(&mdsc->mutex);
3992 ceph_con_close(&s->s_con);
3993 mutex_unlock(&s->s_mutex);
3994 s->s_state = CEPH_MDS_SESSION_RESTARTING;
3995 } else if (oldstate == newstate) {
3996 continue; /* nothing new with this mds */
4002 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4003 newstate >= CEPH_MDS_STATE_RECONNECT) {
4004 mutex_unlock(&mdsc->mutex);
4005 send_mds_reconnect(mdsc, s);
4006 mutex_lock(&mdsc->mutex);
4010 * kick request on any mds that has gone active.
4012 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4013 newstate >= CEPH_MDS_STATE_ACTIVE) {
4014 if (oldstate != CEPH_MDS_STATE_CREATING &&
4015 oldstate != CEPH_MDS_STATE_STARTING)
4016 pr_info("mds%d recovery completed\n", s->s_mds);
4017 kick_requests(mdsc, i);
4018 ceph_kick_flushing_caps(mdsc, s);
4019 wake_up_session_caps(s, RECONNECT);
4023 for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4024 s = mdsc->sessions[i];
4027 if (!ceph_mdsmap_is_laggy(newmap, i))
4029 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4030 s->s_state == CEPH_MDS_SESSION_HUNG ||
4031 s->s_state == CEPH_MDS_SESSION_CLOSING) {
4032 dout(" connecting to export targets of laggy mds%d\n",
4034 __open_export_target_sessions(mdsc, s);
4046 * caller must hold session s_mutex, dentry->d_lock
4048 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4050 struct ceph_dentry_info *di = ceph_dentry(dentry);
4052 ceph_put_mds_session(di->lease_session);
4053 di->lease_session = NULL;
4056 static void handle_lease(struct ceph_mds_client *mdsc,
4057 struct ceph_mds_session *session,
4058 struct ceph_msg *msg)
4060 struct super_block *sb = mdsc->fsc->sb;
4061 struct inode *inode;
4062 struct dentry *parent, *dentry;
4063 struct ceph_dentry_info *di;
4064 int mds = session->s_mds;
4065 struct ceph_mds_lease *h = msg->front.iov_base;
4067 struct ceph_vino vino;
4071 dout("handle_lease from mds%d\n", mds);
4074 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4076 vino.ino = le64_to_cpu(h->ino);
4077 vino.snap = CEPH_NOSNAP;
4078 seq = le32_to_cpu(h->seq);
4079 dname.len = get_unaligned_le32(h + 1);
4080 if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4082 dname.name = (void *)(h + 1) + sizeof(u32);
4085 inode = ceph_find_inode(sb, vino);
4086 dout("handle_lease %s, ino %llx %p %.*s\n",
4087 ceph_lease_op_name(h->action), vino.ino, inode,
4088 dname.len, dname.name);
4090 mutex_lock(&session->s_mutex);
4094 dout("handle_lease no inode %llx\n", vino.ino);
4099 parent = d_find_alias(inode);
4101 dout("no parent dentry on inode %p\n", inode);
4103 goto release; /* hrm... */
4105 dname.hash = full_name_hash(parent, dname.name, dname.len);
4106 dentry = d_lookup(parent, &dname);
4111 spin_lock(&dentry->d_lock);
4112 di = ceph_dentry(dentry);
4113 switch (h->action) {
4114 case CEPH_MDS_LEASE_REVOKE:
4115 if (di->lease_session == session) {
4116 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4117 h->seq = cpu_to_le32(di->lease_seq);
4118 __ceph_mdsc_drop_dentry_lease(dentry);
4123 case CEPH_MDS_LEASE_RENEW:
4124 if (di->lease_session == session &&
4125 di->lease_gen == session->s_cap_gen &&
4126 di->lease_renew_from &&
4127 di->lease_renew_after == 0) {
4128 unsigned long duration =
4129 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4131 di->lease_seq = seq;
4132 di->time = di->lease_renew_from + duration;
4133 di->lease_renew_after = di->lease_renew_from +
4135 di->lease_renew_from = 0;
4139 spin_unlock(&dentry->d_lock);
4146 /* let's just reuse the same message */
4147 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4149 ceph_con_send(&session->s_con, msg);
4152 mutex_unlock(&session->s_mutex);
4153 /* avoid calling iput_final() in mds dispatch threads */
4154 ceph_async_iput(inode);
4158 pr_err("corrupt lease message\n");
4162 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4163 struct dentry *dentry, char action,
4166 struct ceph_msg *msg;
4167 struct ceph_mds_lease *lease;
4169 int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4171 dout("lease_send_msg identry %p %s to mds%d\n",
4172 dentry, ceph_lease_op_name(action), session->s_mds);
4174 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4177 lease = msg->front.iov_base;
4178 lease->action = action;
4179 lease->seq = cpu_to_le32(seq);
4181 spin_lock(&dentry->d_lock);
4182 dir = d_inode(dentry->d_parent);
4183 lease->ino = cpu_to_le64(ceph_ino(dir));
4184 lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4186 put_unaligned_le32(dentry->d_name.len, lease + 1);
4187 memcpy((void *)(lease + 1) + 4,
4188 dentry->d_name.name, dentry->d_name.len);
4189 spin_unlock(&dentry->d_lock);
4191 * if this is a preemptive lease RELEASE, no need to
4192 * flush request stream, since the actual request will
4195 msg->more_to_follow = (action == CEPH_MDS_LEASE_RELEASE);
4197 ceph_con_send(&session->s_con, msg);
4201 * lock unlock sessions, to wait ongoing session activities
4203 static void lock_unlock_sessions(struct ceph_mds_client *mdsc)
4207 mutex_lock(&mdsc->mutex);
4208 for (i = 0; i < mdsc->max_sessions; i++) {
4209 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4212 mutex_unlock(&mdsc->mutex);
4213 mutex_lock(&s->s_mutex);
4214 mutex_unlock(&s->s_mutex);
4215 ceph_put_mds_session(s);
4216 mutex_lock(&mdsc->mutex);
4218 mutex_unlock(&mdsc->mutex);
4221 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4223 struct ceph_fs_client *fsc = mdsc->fsc;
4225 if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4228 if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4231 if (!READ_ONCE(fsc->blacklisted))
4234 if (fsc->last_auto_reconnect &&
4235 time_before(jiffies, fsc->last_auto_reconnect + HZ * 60 * 30))
4238 pr_info("auto reconnect after blacklisted\n");
4239 fsc->last_auto_reconnect = jiffies;
4240 ceph_force_reconnect(fsc->sb);
4244 * delayed work -- periodically trim expired leases, renew caps with mds
4246 static void schedule_delayed(struct ceph_mds_client *mdsc)
4249 unsigned hz = round_jiffies_relative(HZ * delay);
4250 schedule_delayed_work(&mdsc->delayed_work, hz);
4253 static void delayed_work(struct work_struct *work)
4256 struct ceph_mds_client *mdsc =
4257 container_of(work, struct ceph_mds_client, delayed_work.work);
4261 dout("mdsc delayed_work\n");
4263 mutex_lock(&mdsc->mutex);
4264 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4265 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4266 mdsc->last_renew_caps);
4268 mdsc->last_renew_caps = jiffies;
4270 for (i = 0; i < mdsc->max_sessions; i++) {
4271 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4274 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4275 dout("resending session close request for mds%d\n",
4277 request_close_session(mdsc, s);
4278 ceph_put_mds_session(s);
4281 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4282 if (s->s_state == CEPH_MDS_SESSION_OPEN) {
4283 s->s_state = CEPH_MDS_SESSION_HUNG;
4284 pr_info("mds%d hung\n", s->s_mds);
4287 if (s->s_state == CEPH_MDS_SESSION_NEW ||
4288 s->s_state == CEPH_MDS_SESSION_RESTARTING ||
4289 s->s_state == CEPH_MDS_SESSION_REJECTED) {
4290 /* this mds is failed or recovering, just wait */
4291 ceph_put_mds_session(s);
4294 mutex_unlock(&mdsc->mutex);
4296 mutex_lock(&s->s_mutex);
4298 send_renew_caps(mdsc, s);
4300 ceph_con_keepalive(&s->s_con);
4301 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4302 s->s_state == CEPH_MDS_SESSION_HUNG)
4303 ceph_send_cap_releases(mdsc, s);
4304 mutex_unlock(&s->s_mutex);
4305 ceph_put_mds_session(s);
4307 mutex_lock(&mdsc->mutex);
4309 mutex_unlock(&mdsc->mutex);
4311 ceph_check_delayed_caps(mdsc);
4313 ceph_queue_cap_reclaim_work(mdsc);
4315 ceph_trim_snapid_map(mdsc);
4317 maybe_recover_session(mdsc);
4319 schedule_delayed(mdsc);
4322 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4325 struct ceph_mds_client *mdsc;
4327 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4331 mutex_init(&mdsc->mutex);
4332 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4333 if (!mdsc->mdsmap) {
4339 init_completion(&mdsc->safe_umount_waiters);
4340 init_waitqueue_head(&mdsc->session_close_wq);
4341 INIT_LIST_HEAD(&mdsc->waiting_for_map);
4342 mdsc->sessions = NULL;
4343 atomic_set(&mdsc->num_sessions, 0);
4344 mdsc->max_sessions = 0;
4346 atomic64_set(&mdsc->quotarealms_count, 0);
4347 mdsc->quotarealms_inodes = RB_ROOT;
4348 mutex_init(&mdsc->quotarealms_inodes_mutex);
4349 mdsc->last_snap_seq = 0;
4350 init_rwsem(&mdsc->snap_rwsem);
4351 mdsc->snap_realms = RB_ROOT;
4352 INIT_LIST_HEAD(&mdsc->snap_empty);
4353 mdsc->num_snap_realms = 0;
4354 spin_lock_init(&mdsc->snap_empty_lock);
4356 mdsc->oldest_tid = 0;
4357 mdsc->request_tree = RB_ROOT;
4358 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4359 mdsc->last_renew_caps = jiffies;
4360 INIT_LIST_HEAD(&mdsc->cap_delay_list);
4361 INIT_LIST_HEAD(&mdsc->cap_wait_list);
4362 spin_lock_init(&mdsc->cap_delay_lock);
4363 INIT_LIST_HEAD(&mdsc->snap_flush_list);
4364 spin_lock_init(&mdsc->snap_flush_lock);
4365 mdsc->last_cap_flush_tid = 1;
4366 INIT_LIST_HEAD(&mdsc->cap_flush_list);
4367 INIT_LIST_HEAD(&mdsc->cap_dirty);
4368 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4369 mdsc->num_cap_flushing = 0;
4370 spin_lock_init(&mdsc->cap_dirty_lock);
4371 init_waitqueue_head(&mdsc->cap_flushing_wq);
4372 INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4373 atomic_set(&mdsc->cap_reclaim_pending, 0);
4375 spin_lock_init(&mdsc->dentry_list_lock);
4376 INIT_LIST_HEAD(&mdsc->dentry_leases);
4377 INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4379 ceph_caps_init(mdsc);
4380 ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4382 spin_lock_init(&mdsc->snapid_map_lock);
4383 mdsc->snapid_map_tree = RB_ROOT;
4384 INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4386 init_rwsem(&mdsc->pool_perm_rwsem);
4387 mdsc->pool_perm_tree = RB_ROOT;
4389 strscpy(mdsc->nodename, utsname()->nodename,
4390 sizeof(mdsc->nodename));
4395 * Wait for safe replies on open mds requests. If we time out, drop
4396 * all requests from the tree to avoid dangling dentry refs.
4398 static void wait_requests(struct ceph_mds_client *mdsc)
4400 struct ceph_options *opts = mdsc->fsc->client->options;
4401 struct ceph_mds_request *req;
4403 mutex_lock(&mdsc->mutex);
4404 if (__get_oldest_req(mdsc)) {
4405 mutex_unlock(&mdsc->mutex);
4407 dout("wait_requests waiting for requests\n");
4408 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4409 ceph_timeout_jiffies(opts->mount_timeout));
4411 /* tear down remaining requests */
4412 mutex_lock(&mdsc->mutex);
4413 while ((req = __get_oldest_req(mdsc))) {
4414 dout("wait_requests timed out on tid %llu\n",
4416 list_del_init(&req->r_wait);
4417 __unregister_request(mdsc, req);
4420 mutex_unlock(&mdsc->mutex);
4421 dout("wait_requests done\n");
4425 * called before mount is ro, and before dentries are torn down.
4426 * (hmm, does this still race with new lookups?)
4428 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4430 dout("pre_umount\n");
4433 lock_unlock_sessions(mdsc);
4434 ceph_flush_dirty_caps(mdsc);
4435 wait_requests(mdsc);
4438 * wait for reply handlers to drop their request refs and
4439 * their inode/dcache refs
4443 ceph_cleanup_quotarealms_inodes(mdsc);
4447 * wait for all write mds requests to flush.
4449 static void wait_unsafe_requests(struct ceph_mds_client *mdsc, u64 want_tid)
4451 struct ceph_mds_request *req = NULL, *nextreq;
4454 mutex_lock(&mdsc->mutex);
4455 dout("wait_unsafe_requests want %lld\n", want_tid);
4457 req = __get_oldest_req(mdsc);
4458 while (req && req->r_tid <= want_tid) {
4459 /* find next request */
4460 n = rb_next(&req->r_node);
4462 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4465 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4466 (req->r_op & CEPH_MDS_OP_WRITE)) {
4468 ceph_mdsc_get_request(req);
4470 ceph_mdsc_get_request(nextreq);
4471 mutex_unlock(&mdsc->mutex);
4472 dout("wait_unsafe_requests wait on %llu (want %llu)\n",
4473 req->r_tid, want_tid);
4474 wait_for_completion(&req->r_safe_completion);
4475 mutex_lock(&mdsc->mutex);
4476 ceph_mdsc_put_request(req);
4478 break; /* next dne before, so we're done! */
4479 if (RB_EMPTY_NODE(&nextreq->r_node)) {
4480 /* next request was removed from tree */
4481 ceph_mdsc_put_request(nextreq);
4484 ceph_mdsc_put_request(nextreq); /* won't go away */
4488 mutex_unlock(&mdsc->mutex);
4489 dout("wait_unsafe_requests done\n");
4492 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4494 u64 want_tid, want_flush;
4496 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4500 mutex_lock(&mdsc->mutex);
4501 want_tid = mdsc->last_tid;
4502 mutex_unlock(&mdsc->mutex);
4504 ceph_flush_dirty_caps(mdsc);
4505 spin_lock(&mdsc->cap_dirty_lock);
4506 want_flush = mdsc->last_cap_flush_tid;
4507 if (!list_empty(&mdsc->cap_flush_list)) {
4508 struct ceph_cap_flush *cf =
4509 list_last_entry(&mdsc->cap_flush_list,
4510 struct ceph_cap_flush, g_list);
4513 spin_unlock(&mdsc->cap_dirty_lock);
4515 dout("sync want tid %lld flush_seq %lld\n",
4516 want_tid, want_flush);
4518 wait_unsafe_requests(mdsc, want_tid);
4519 wait_caps_flush(mdsc, want_flush);
4523 * true if all sessions are closed, or we force unmount
4525 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4527 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4529 return atomic_read(&mdsc->num_sessions) <= skipped;
4533 * called after sb is ro.
4535 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4537 struct ceph_options *opts = mdsc->fsc->client->options;
4538 struct ceph_mds_session *session;
4542 dout("close_sessions\n");
4544 /* close sessions */
4545 mutex_lock(&mdsc->mutex);
4546 for (i = 0; i < mdsc->max_sessions; i++) {
4547 session = __ceph_lookup_mds_session(mdsc, i);
4550 mutex_unlock(&mdsc->mutex);
4551 mutex_lock(&session->s_mutex);
4552 if (__close_session(mdsc, session) <= 0)
4554 mutex_unlock(&session->s_mutex);
4555 ceph_put_mds_session(session);
4556 mutex_lock(&mdsc->mutex);
4558 mutex_unlock(&mdsc->mutex);
4560 dout("waiting for sessions to close\n");
4561 wait_event_timeout(mdsc->session_close_wq,
4562 done_closing_sessions(mdsc, skipped),
4563 ceph_timeout_jiffies(opts->mount_timeout));
4565 /* tear down remaining sessions */
4566 mutex_lock(&mdsc->mutex);
4567 for (i = 0; i < mdsc->max_sessions; i++) {
4568 if (mdsc->sessions[i]) {
4569 session = ceph_get_mds_session(mdsc->sessions[i]);
4570 __unregister_session(mdsc, session);
4571 mutex_unlock(&mdsc->mutex);
4572 mutex_lock(&session->s_mutex);
4573 remove_session_caps(session);
4574 mutex_unlock(&session->s_mutex);
4575 ceph_put_mds_session(session);
4576 mutex_lock(&mdsc->mutex);
4579 WARN_ON(!list_empty(&mdsc->cap_delay_list));
4580 mutex_unlock(&mdsc->mutex);
4582 ceph_cleanup_snapid_map(mdsc);
4583 ceph_cleanup_empty_realms(mdsc);
4585 cancel_work_sync(&mdsc->cap_reclaim_work);
4586 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4591 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4593 struct ceph_mds_session *session;
4596 dout("force umount\n");
4598 mutex_lock(&mdsc->mutex);
4599 for (mds = 0; mds < mdsc->max_sessions; mds++) {
4600 session = __ceph_lookup_mds_session(mdsc, mds);
4604 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4605 __unregister_session(mdsc, session);
4606 __wake_requests(mdsc, &session->s_waiting);
4607 mutex_unlock(&mdsc->mutex);
4609 mutex_lock(&session->s_mutex);
4610 __close_session(mdsc, session);
4611 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4612 cleanup_session_requests(mdsc, session);
4613 remove_session_caps(session);
4615 mutex_unlock(&session->s_mutex);
4616 ceph_put_mds_session(session);
4618 mutex_lock(&mdsc->mutex);
4619 kick_requests(mdsc, mds);
4621 __wake_requests(mdsc, &mdsc->waiting_for_map);
4622 mutex_unlock(&mdsc->mutex);
4625 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4628 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4630 ceph_mdsmap_destroy(mdsc->mdsmap);
4631 kfree(mdsc->sessions);
4632 ceph_caps_finalize(mdsc);
4633 ceph_pool_perm_destroy(mdsc);
4636 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4638 struct ceph_mds_client *mdsc = fsc->mdsc;
4639 dout("mdsc_destroy %p\n", mdsc);
4644 /* flush out any connection work with references to us */
4647 ceph_mdsc_stop(mdsc);
4651 dout("mdsc_destroy %p done\n", mdsc);
4654 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4656 struct ceph_fs_client *fsc = mdsc->fsc;
4657 const char *mds_namespace = fsc->mount_options->mds_namespace;
4658 void *p = msg->front.iov_base;
4659 void *end = p + msg->front.iov_len;
4663 u32 mount_fscid = (u32)-1;
4664 u8 struct_v, struct_cv;
4667 ceph_decode_need(&p, end, sizeof(u32), bad);
4668 epoch = ceph_decode_32(&p);
4670 dout("handle_fsmap epoch %u\n", epoch);
4672 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4673 struct_v = ceph_decode_8(&p);
4674 struct_cv = ceph_decode_8(&p);
4675 map_len = ceph_decode_32(&p);
4677 ceph_decode_need(&p, end, sizeof(u32) * 3, bad);
4678 p += sizeof(u32) * 2; /* skip epoch and legacy_client_fscid */
4680 num_fs = ceph_decode_32(&p);
4681 while (num_fs-- > 0) {
4682 void *info_p, *info_end;
4687 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
4688 info_v = ceph_decode_8(&p);
4689 info_cv = ceph_decode_8(&p);
4690 info_len = ceph_decode_32(&p);
4691 ceph_decode_need(&p, end, info_len, bad);
4693 info_end = p + info_len;
4696 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
4697 fscid = ceph_decode_32(&info_p);
4698 namelen = ceph_decode_32(&info_p);
4699 ceph_decode_need(&info_p, info_end, namelen, bad);
4701 if (mds_namespace &&
4702 strlen(mds_namespace) == namelen &&
4703 !strncmp(mds_namespace, (char *)info_p, namelen)) {
4704 mount_fscid = fscid;
4709 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
4710 if (mount_fscid != (u32)-1) {
4711 fsc->client->monc.fs_cluster_id = mount_fscid;
4712 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
4714 ceph_monc_renew_subs(&fsc->client->monc);
4722 pr_err("error decoding fsmap\n");
4724 mutex_lock(&mdsc->mutex);
4725 mdsc->mdsmap_err = err;
4726 __wake_requests(mdsc, &mdsc->waiting_for_map);
4727 mutex_unlock(&mdsc->mutex);
4731 * handle mds map update.
4733 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4737 void *p = msg->front.iov_base;
4738 void *end = p + msg->front.iov_len;
4739 struct ceph_mdsmap *newmap, *oldmap;
4740 struct ceph_fsid fsid;
4743 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
4744 ceph_decode_copy(&p, &fsid, sizeof(fsid));
4745 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
4747 epoch = ceph_decode_32(&p);
4748 maplen = ceph_decode_32(&p);
4749 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
4751 /* do we need it? */
4752 mutex_lock(&mdsc->mutex);
4753 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
4754 dout("handle_map epoch %u <= our %u\n",
4755 epoch, mdsc->mdsmap->m_epoch);
4756 mutex_unlock(&mdsc->mutex);
4760 newmap = ceph_mdsmap_decode(&p, end);
4761 if (IS_ERR(newmap)) {
4762 err = PTR_ERR(newmap);
4766 /* swap into place */
4768 oldmap = mdsc->mdsmap;
4769 mdsc->mdsmap = newmap;
4770 check_new_map(mdsc, newmap, oldmap);
4771 ceph_mdsmap_destroy(oldmap);
4773 mdsc->mdsmap = newmap; /* first mds map */
4775 mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
4778 __wake_requests(mdsc, &mdsc->waiting_for_map);
4779 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
4780 mdsc->mdsmap->m_epoch);
4782 mutex_unlock(&mdsc->mutex);
4783 schedule_delayed(mdsc);
4787 mutex_unlock(&mdsc->mutex);
4789 pr_err("error decoding mdsmap %d\n", err);
4793 static struct ceph_connection *con_get(struct ceph_connection *con)
4795 struct ceph_mds_session *s = con->private;
4797 if (ceph_get_mds_session(s))
4802 static void con_put(struct ceph_connection *con)
4804 struct ceph_mds_session *s = con->private;
4806 ceph_put_mds_session(s);
4810 * if the client is unresponsive for long enough, the mds will kill
4811 * the session entirely.
4813 static void peer_reset(struct ceph_connection *con)
4815 struct ceph_mds_session *s = con->private;
4816 struct ceph_mds_client *mdsc = s->s_mdsc;
4818 pr_warn("mds%d closed our session\n", s->s_mds);
4819 send_mds_reconnect(mdsc, s);
4822 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
4824 struct ceph_mds_session *s = con->private;
4825 struct ceph_mds_client *mdsc = s->s_mdsc;
4826 int type = le16_to_cpu(msg->hdr.type);
4828 mutex_lock(&mdsc->mutex);
4829 if (__verify_registered_session(mdsc, s) < 0) {
4830 mutex_unlock(&mdsc->mutex);
4833 mutex_unlock(&mdsc->mutex);
4836 case CEPH_MSG_MDS_MAP:
4837 ceph_mdsc_handle_mdsmap(mdsc, msg);
4839 case CEPH_MSG_FS_MAP_USER:
4840 ceph_mdsc_handle_fsmap(mdsc, msg);
4842 case CEPH_MSG_CLIENT_SESSION:
4843 handle_session(s, msg);
4845 case CEPH_MSG_CLIENT_REPLY:
4846 handle_reply(s, msg);
4848 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
4849 handle_forward(mdsc, s, msg);
4851 case CEPH_MSG_CLIENT_CAPS:
4852 ceph_handle_caps(s, msg);
4854 case CEPH_MSG_CLIENT_SNAP:
4855 ceph_handle_snap(mdsc, s, msg);
4857 case CEPH_MSG_CLIENT_LEASE:
4858 handle_lease(mdsc, s, msg);
4860 case CEPH_MSG_CLIENT_QUOTA:
4861 ceph_handle_quota(mdsc, s, msg);
4865 pr_err("received unknown message type %d %s\n", type,
4866 ceph_msg_type_name(type));
4877 * Note: returned pointer is the address of a structure that's
4878 * managed separately. Caller must *not* attempt to free it.
4880 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
4881 int *proto, int force_new)
4883 struct ceph_mds_session *s = con->private;
4884 struct ceph_mds_client *mdsc = s->s_mdsc;
4885 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4886 struct ceph_auth_handshake *auth = &s->s_auth;
4888 if (force_new && auth->authorizer) {
4889 ceph_auth_destroy_authorizer(auth->authorizer);
4890 auth->authorizer = NULL;
4892 if (!auth->authorizer) {
4893 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4896 return ERR_PTR(ret);
4898 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_MDS,
4901 return ERR_PTR(ret);
4903 *proto = ac->protocol;
4908 static int add_authorizer_challenge(struct ceph_connection *con,
4909 void *challenge_buf, int challenge_buf_len)
4911 struct ceph_mds_session *s = con->private;
4912 struct ceph_mds_client *mdsc = s->s_mdsc;
4913 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4915 return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
4916 challenge_buf, challenge_buf_len);
4919 static int verify_authorizer_reply(struct ceph_connection *con)
4921 struct ceph_mds_session *s = con->private;
4922 struct ceph_mds_client *mdsc = s->s_mdsc;
4923 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4925 return ceph_auth_verify_authorizer_reply(ac, s->s_auth.authorizer);
4928 static int invalidate_authorizer(struct ceph_connection *con)
4930 struct ceph_mds_session *s = con->private;
4931 struct ceph_mds_client *mdsc = s->s_mdsc;
4932 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
4934 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
4936 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
4939 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
4940 struct ceph_msg_header *hdr, int *skip)
4942 struct ceph_msg *msg;
4943 int type = (int) le16_to_cpu(hdr->type);
4944 int front_len = (int) le32_to_cpu(hdr->front_len);
4950 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
4952 pr_err("unable to allocate msg type %d len %d\n",
4960 static int mds_sign_message(struct ceph_msg *msg)
4962 struct ceph_mds_session *s = msg->con->private;
4963 struct ceph_auth_handshake *auth = &s->s_auth;
4965 return ceph_auth_sign_message(auth, msg);
4968 static int mds_check_message_signature(struct ceph_msg *msg)
4970 struct ceph_mds_session *s = msg->con->private;
4971 struct ceph_auth_handshake *auth = &s->s_auth;
4973 return ceph_auth_check_message_signature(auth, msg);
4976 static const struct ceph_connection_operations mds_con_ops = {
4979 .dispatch = dispatch,
4980 .get_authorizer = get_authorizer,
4981 .add_authorizer_challenge = add_authorizer_challenge,
4982 .verify_authorizer_reply = verify_authorizer_reply,
4983 .invalidate_authorizer = invalidate_authorizer,
4984 .peer_reset = peer_reset,
4985 .alloc_msg = mds_alloc_msg,
4986 .sign_message = mds_sign_message,
4987 .check_message_signature = mds_check_message_signature,