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>
13 #include <linux/ktime.h>
14 #include <linux/bitmap.h>
17 #include "mds_client.h"
19 #include <linux/ceph/ceph_features.h>
20 #include <linux/ceph/messenger.h>
21 #include <linux/ceph/decode.h>
22 #include <linux/ceph/pagelist.h>
23 #include <linux/ceph/auth.h>
24 #include <linux/ceph/debugfs.h>
26 #define RECONNECT_MAX_SIZE (INT_MAX - PAGE_SIZE)
29 * A cluster of MDS (metadata server) daemons is responsible for
30 * managing the file system namespace (the directory hierarchy and
31 * inodes) and for coordinating shared access to storage. Metadata is
32 * partitioning hierarchically across a number of servers, and that
33 * partition varies over time as the cluster adjusts the distribution
34 * in order to balance load.
36 * The MDS client is primarily responsible to managing synchronous
37 * metadata requests for operations like open, unlink, and so forth.
38 * If there is a MDS failure, we find out about it when we (possibly
39 * request and) receive a new MDS map, and can resubmit affected
42 * For the most part, though, we take advantage of a lossless
43 * communications channel to the MDS, and do not need to worry about
44 * timing out or resubmitting requests.
46 * We maintain a stateful "session" with each MDS we interact with.
47 * Within each session, we sent periodic heartbeat messages to ensure
48 * any capabilities or leases we have been issues remain valid. If
49 * the session times out and goes stale, our leases and capabilities
50 * are no longer valid.
53 struct ceph_reconnect_state {
54 struct ceph_mds_session *session;
55 int nr_caps, nr_realms;
56 struct ceph_pagelist *pagelist;
61 static void __wake_requests(struct ceph_mds_client *mdsc,
62 struct list_head *head);
63 static void ceph_cap_release_work(struct work_struct *work);
64 static void ceph_cap_reclaim_work(struct work_struct *work);
66 static const struct ceph_connection_operations mds_con_ops;
73 static int parse_reply_info_quota(void **p, void *end,
74 struct ceph_mds_reply_info_in *info)
76 u8 struct_v, struct_compat;
79 ceph_decode_8_safe(p, end, struct_v, bad);
80 ceph_decode_8_safe(p, end, struct_compat, bad);
81 /* struct_v is expected to be >= 1. we only
82 * understand encoding with struct_compat == 1. */
83 if (!struct_v || struct_compat != 1)
85 ceph_decode_32_safe(p, end, struct_len, bad);
86 ceph_decode_need(p, end, struct_len, bad);
87 end = *p + struct_len;
88 ceph_decode_64_safe(p, end, info->max_bytes, bad);
89 ceph_decode_64_safe(p, end, info->max_files, bad);
97 * parse individual inode info
99 static int parse_reply_info_in(void **p, void *end,
100 struct ceph_mds_reply_info_in *info,
106 if (features == (u64)-1) {
109 ceph_decode_8_safe(p, end, struct_v, bad);
110 ceph_decode_8_safe(p, end, struct_compat, bad);
111 /* struct_v is expected to be >= 1. we only understand
112 * encoding with struct_compat == 1. */
113 if (!struct_v || struct_compat != 1)
115 ceph_decode_32_safe(p, end, struct_len, bad);
116 ceph_decode_need(p, end, struct_len, bad);
117 end = *p + struct_len;
120 ceph_decode_need(p, end, sizeof(struct ceph_mds_reply_inode), bad);
122 *p += sizeof(struct ceph_mds_reply_inode) +
123 sizeof(*info->in->fragtree.splits) *
124 le32_to_cpu(info->in->fragtree.nsplits);
126 ceph_decode_32_safe(p, end, info->symlink_len, bad);
127 ceph_decode_need(p, end, info->symlink_len, bad);
129 *p += info->symlink_len;
131 ceph_decode_copy_safe(p, end, &info->dir_layout,
132 sizeof(info->dir_layout), bad);
133 ceph_decode_32_safe(p, end, info->xattr_len, bad);
134 ceph_decode_need(p, end, info->xattr_len, bad);
135 info->xattr_data = *p;
136 *p += info->xattr_len;
138 if (features == (u64)-1) {
140 ceph_decode_64_safe(p, end, info->inline_version, bad);
141 ceph_decode_32_safe(p, end, info->inline_len, bad);
142 ceph_decode_need(p, end, info->inline_len, bad);
143 info->inline_data = *p;
144 *p += info->inline_len;
146 err = parse_reply_info_quota(p, end, info);
150 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
151 if (info->pool_ns_len > 0) {
152 ceph_decode_need(p, end, info->pool_ns_len, bad);
153 info->pool_ns_data = *p;
154 *p += info->pool_ns_len;
158 ceph_decode_need(p, end, sizeof(info->btime), bad);
159 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
161 /* change attribute */
162 ceph_decode_64_safe(p, end, info->change_attr, bad);
166 ceph_decode_32_safe(p, end, info->dir_pin, bad);
168 info->dir_pin = -ENODATA;
171 /* snapshot birth time, remains zero for v<=2 */
173 ceph_decode_need(p, end, sizeof(info->snap_btime), bad);
174 ceph_decode_copy(p, &info->snap_btime,
175 sizeof(info->snap_btime));
177 memset(&info->snap_btime, 0, sizeof(info->snap_btime));
180 /* snapshot count, remains zero for v<=3 */
182 ceph_decode_64_safe(p, end, info->rsnaps, bad);
189 if (features & CEPH_FEATURE_MDS_INLINE_DATA) {
190 ceph_decode_64_safe(p, end, info->inline_version, bad);
191 ceph_decode_32_safe(p, end, info->inline_len, bad);
192 ceph_decode_need(p, end, info->inline_len, bad);
193 info->inline_data = *p;
194 *p += info->inline_len;
196 info->inline_version = CEPH_INLINE_NONE;
198 if (features & CEPH_FEATURE_MDS_QUOTA) {
199 err = parse_reply_info_quota(p, end, info);
207 info->pool_ns_len = 0;
208 info->pool_ns_data = NULL;
209 if (features & CEPH_FEATURE_FS_FILE_LAYOUT_V2) {
210 ceph_decode_32_safe(p, end, info->pool_ns_len, bad);
211 if (info->pool_ns_len > 0) {
212 ceph_decode_need(p, end, info->pool_ns_len, bad);
213 info->pool_ns_data = *p;
214 *p += info->pool_ns_len;
218 if (features & CEPH_FEATURE_FS_BTIME) {
219 ceph_decode_need(p, end, sizeof(info->btime), bad);
220 ceph_decode_copy(p, &info->btime, sizeof(info->btime));
221 ceph_decode_64_safe(p, end, info->change_attr, bad);
224 info->dir_pin = -ENODATA;
225 /* info->snap_btime and info->rsnaps remain zero */
234 static int parse_reply_info_dir(void **p, void *end,
235 struct ceph_mds_reply_dirfrag **dirfrag,
238 if (features == (u64)-1) {
239 u8 struct_v, struct_compat;
241 ceph_decode_8_safe(p, end, struct_v, bad);
242 ceph_decode_8_safe(p, end, struct_compat, bad);
243 /* struct_v is expected to be >= 1. we only understand
244 * encoding whose struct_compat == 1. */
245 if (!struct_v || struct_compat != 1)
247 ceph_decode_32_safe(p, end, struct_len, bad);
248 ceph_decode_need(p, end, struct_len, bad);
249 end = *p + struct_len;
252 ceph_decode_need(p, end, sizeof(**dirfrag), bad);
254 *p += sizeof(**dirfrag) + sizeof(u32) * le32_to_cpu((*dirfrag)->ndist);
255 if (unlikely(*p > end))
257 if (features == (u64)-1)
264 static int parse_reply_info_lease(void **p, void *end,
265 struct ceph_mds_reply_lease **lease,
268 if (features == (u64)-1) {
269 u8 struct_v, struct_compat;
271 ceph_decode_8_safe(p, end, struct_v, bad);
272 ceph_decode_8_safe(p, end, struct_compat, bad);
273 /* struct_v is expected to be >= 1. we only understand
274 * encoding whose struct_compat == 1. */
275 if (!struct_v || struct_compat != 1)
277 ceph_decode_32_safe(p, end, struct_len, bad);
278 ceph_decode_need(p, end, struct_len, bad);
279 end = *p + struct_len;
282 ceph_decode_need(p, end, sizeof(**lease), bad);
284 *p += sizeof(**lease);
285 if (features == (u64)-1)
293 * parse a normal reply, which may contain a (dir+)dentry and/or a
296 static int parse_reply_info_trace(void **p, void *end,
297 struct ceph_mds_reply_info_parsed *info,
302 if (info->head->is_dentry) {
303 err = parse_reply_info_in(p, end, &info->diri, features);
307 err = parse_reply_info_dir(p, end, &info->dirfrag, features);
311 ceph_decode_32_safe(p, end, info->dname_len, bad);
312 ceph_decode_need(p, end, info->dname_len, bad);
314 *p += info->dname_len;
316 err = parse_reply_info_lease(p, end, &info->dlease, features);
321 if (info->head->is_target) {
322 err = parse_reply_info_in(p, end, &info->targeti, features);
327 if (unlikely(*p != end))
334 pr_err("problem parsing mds trace %d\n", err);
339 * parse readdir results
341 static int parse_reply_info_readdir(void **p, void *end,
342 struct ceph_mds_reply_info_parsed *info,
348 err = parse_reply_info_dir(p, end, &info->dir_dir, features);
352 ceph_decode_need(p, end, sizeof(num) + 2, bad);
353 num = ceph_decode_32(p);
355 u16 flags = ceph_decode_16(p);
356 info->dir_end = !!(flags & CEPH_READDIR_FRAG_END);
357 info->dir_complete = !!(flags & CEPH_READDIR_FRAG_COMPLETE);
358 info->hash_order = !!(flags & CEPH_READDIR_HASH_ORDER);
359 info->offset_hash = !!(flags & CEPH_READDIR_OFFSET_HASH);
364 BUG_ON(!info->dir_entries);
365 if ((unsigned long)(info->dir_entries + num) >
366 (unsigned long)info->dir_entries + info->dir_buf_size) {
367 pr_err("dir contents are larger than expected\n");
374 struct ceph_mds_reply_dir_entry *rde = info->dir_entries + i;
376 ceph_decode_32_safe(p, end, rde->name_len, bad);
377 ceph_decode_need(p, end, rde->name_len, bad);
380 dout("parsed dir dname '%.*s'\n", rde->name_len, rde->name);
383 err = parse_reply_info_lease(p, end, &rde->lease, features);
387 err = parse_reply_info_in(p, end, &rde->inode, features);
390 /* ceph_readdir_prepopulate() will update it */
397 /* Skip over any unrecognized fields */
404 pr_err("problem parsing dir contents %d\n", err);
409 * parse fcntl F_GETLK results
411 static int parse_reply_info_filelock(void **p, void *end,
412 struct ceph_mds_reply_info_parsed *info,
415 if (*p + sizeof(*info->filelock_reply) > end)
418 info->filelock_reply = *p;
420 /* Skip over any unrecognized fields */
428 #if BITS_PER_LONG == 64
430 #define DELEGATED_INO_AVAILABLE xa_mk_value(1)
432 static int ceph_parse_deleg_inos(void **p, void *end,
433 struct ceph_mds_session *s)
437 ceph_decode_32_safe(p, end, sets, bad);
438 dout("got %u sets of delegated inodes\n", sets);
442 ceph_decode_64_safe(p, end, start, bad);
443 ceph_decode_64_safe(p, end, len, bad);
445 /* Don't accept a delegation of system inodes */
446 if (start < CEPH_INO_SYSTEM_BASE) {
447 pr_warn_ratelimited("ceph: ignoring reserved inode range delegation (start=0x%llx len=0x%llx)\n",
452 int err = xa_insert(&s->s_delegated_inos, start++,
453 DELEGATED_INO_AVAILABLE,
456 dout("added delegated inode 0x%llx\n",
458 } else if (err == -EBUSY) {
459 pr_warn("MDS delegated inode 0x%llx more than once.\n",
471 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
476 xa_for_each(&s->s_delegated_inos, ino, val) {
477 val = xa_erase(&s->s_delegated_inos, ino);
478 if (val == DELEGATED_INO_AVAILABLE)
484 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
486 return xa_insert(&s->s_delegated_inos, ino, DELEGATED_INO_AVAILABLE,
489 #else /* BITS_PER_LONG == 64 */
491 * FIXME: xarrays can't handle 64-bit indexes on a 32-bit arch. For now, just
492 * ignore delegated_inos on 32 bit arch. Maybe eventually add xarrays for top
495 static int ceph_parse_deleg_inos(void **p, void *end,
496 struct ceph_mds_session *s)
500 ceph_decode_32_safe(p, end, sets, bad);
502 ceph_decode_skip_n(p, end, sets * 2 * sizeof(__le64), bad);
508 u64 ceph_get_deleg_ino(struct ceph_mds_session *s)
513 int ceph_restore_deleg_ino(struct ceph_mds_session *s, u64 ino)
517 #endif /* BITS_PER_LONG == 64 */
520 * parse create results
522 static int parse_reply_info_create(void **p, void *end,
523 struct ceph_mds_reply_info_parsed *info,
524 u64 features, struct ceph_mds_session *s)
528 if (features == (u64)-1 ||
529 (features & CEPH_FEATURE_REPLY_CREATE_INODE)) {
531 /* Malformed reply? */
532 info->has_create_ino = false;
533 } else if (test_bit(CEPHFS_FEATURE_DELEG_INO, &s->s_features)) {
534 info->has_create_ino = true;
535 /* struct_v, struct_compat, and len */
536 ceph_decode_skip_n(p, end, 2 + sizeof(u32), bad);
537 ceph_decode_64_safe(p, end, info->ino, bad);
538 ret = ceph_parse_deleg_inos(p, end, s);
543 ceph_decode_64_safe(p, end, info->ino, bad);
544 info->has_create_ino = true;
551 /* Skip over any unrecognized fields */
558 static int parse_reply_info_getvxattr(void **p, void *end,
559 struct ceph_mds_reply_info_parsed *info,
564 ceph_decode_skip_8(p, end, bad); /* skip current version: 1 */
565 ceph_decode_skip_8(p, end, bad); /* skip first version: 1 */
566 ceph_decode_skip_32(p, end, bad); /* skip payload length */
568 ceph_decode_32_safe(p, end, value_len, bad);
570 if (value_len == end - *p) {
571 info->xattr_info.xattr_value = *p;
572 info->xattr_info.xattr_value_len = value_len;
581 * parse extra results
583 static int parse_reply_info_extra(void **p, void *end,
584 struct ceph_mds_reply_info_parsed *info,
585 u64 features, struct ceph_mds_session *s)
587 u32 op = le32_to_cpu(info->head->op);
589 if (op == CEPH_MDS_OP_GETFILELOCK)
590 return parse_reply_info_filelock(p, end, info, features);
591 else if (op == CEPH_MDS_OP_READDIR || op == CEPH_MDS_OP_LSSNAP)
592 return parse_reply_info_readdir(p, end, info, features);
593 else if (op == CEPH_MDS_OP_CREATE)
594 return parse_reply_info_create(p, end, info, features, s);
595 else if (op == CEPH_MDS_OP_GETVXATTR)
596 return parse_reply_info_getvxattr(p, end, info, features);
602 * parse entire mds reply
604 static int parse_reply_info(struct ceph_mds_session *s, struct ceph_msg *msg,
605 struct ceph_mds_reply_info_parsed *info,
612 info->head = msg->front.iov_base;
613 p = msg->front.iov_base + sizeof(struct ceph_mds_reply_head);
614 end = p + msg->front.iov_len - sizeof(struct ceph_mds_reply_head);
617 ceph_decode_32_safe(&p, end, len, bad);
619 ceph_decode_need(&p, end, len, bad);
620 err = parse_reply_info_trace(&p, p+len, info, features);
626 ceph_decode_32_safe(&p, end, len, bad);
628 ceph_decode_need(&p, end, len, bad);
629 err = parse_reply_info_extra(&p, p+len, info, features, s);
635 ceph_decode_32_safe(&p, end, len, bad);
636 info->snapblob_len = len;
647 pr_err("mds parse_reply err %d\n", err);
651 static void destroy_reply_info(struct ceph_mds_reply_info_parsed *info)
653 if (!info->dir_entries)
655 free_pages((unsigned long)info->dir_entries, get_order(info->dir_buf_size));
659 * In async unlink case the kclient won't wait for the first reply
660 * from MDS and just drop all the links and unhash the dentry and then
661 * succeeds immediately.
663 * For any new create/link/rename,etc requests followed by using the
664 * same file names we must wait for the first reply of the inflight
665 * unlink request, or the MDS possibly will fail these following
666 * requests with -EEXIST if the inflight async unlink request was
667 * delayed for some reasons.
669 * And the worst case is that for the none async openc request it will
670 * successfully open the file if the CDentry hasn't been unlinked yet,
671 * but later the previous delayed async unlink request will remove the
672 * CDenty. That means the just created file is possiblly deleted later
675 * We need to wait for the inflight async unlink requests to finish
676 * when creating new files/directories by using the same file names.
678 int ceph_wait_on_conflict_unlink(struct dentry *dentry)
680 struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
681 struct dentry *pdentry = dentry->d_parent;
682 struct dentry *udentry, *found = NULL;
683 struct ceph_dentry_info *di;
685 u32 hash = dentry->d_name.hash;
688 dname.name = dentry->d_name.name;
689 dname.len = dentry->d_name.len;
692 hash_for_each_possible_rcu(fsc->async_unlink_conflict, di,
694 udentry = di->dentry;
696 spin_lock(&udentry->d_lock);
697 if (udentry->d_name.hash != hash)
699 if (unlikely(udentry->d_parent != pdentry))
701 if (!hash_hashed(&di->hnode))
704 if (!test_bit(CEPH_DENTRY_ASYNC_UNLINK_BIT, &di->flags))
705 pr_warn("%s dentry %p:%pd async unlink bit is not set\n",
706 __func__, dentry, dentry);
708 if (!d_same_name(udentry, pdentry, &dname))
711 spin_unlock(&udentry->d_lock);
712 found = dget(udentry);
715 spin_unlock(&udentry->d_lock);
722 dout("%s dentry %p:%pd conflict with old %p:%pd\n", __func__,
723 dentry, dentry, found, found);
725 err = wait_on_bit(&di->flags, CEPH_DENTRY_ASYNC_UNLINK_BIT,
735 const char *ceph_session_state_name(int s)
738 case CEPH_MDS_SESSION_NEW: return "new";
739 case CEPH_MDS_SESSION_OPENING: return "opening";
740 case CEPH_MDS_SESSION_OPEN: return "open";
741 case CEPH_MDS_SESSION_HUNG: return "hung";
742 case CEPH_MDS_SESSION_CLOSING: return "closing";
743 case CEPH_MDS_SESSION_CLOSED: return "closed";
744 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
745 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
746 case CEPH_MDS_SESSION_REJECTED: return "rejected";
747 default: return "???";
751 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
753 if (refcount_inc_not_zero(&s->s_ref))
758 void ceph_put_mds_session(struct ceph_mds_session *s)
760 if (IS_ERR_OR_NULL(s))
763 if (refcount_dec_and_test(&s->s_ref)) {
764 if (s->s_auth.authorizer)
765 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
766 WARN_ON(mutex_is_locked(&s->s_mutex));
767 xa_destroy(&s->s_delegated_inos);
773 * called under mdsc->mutex
775 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
778 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
780 return ceph_get_mds_session(mdsc->sessions[mds]);
783 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
785 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
791 static int __verify_registered_session(struct ceph_mds_client *mdsc,
792 struct ceph_mds_session *s)
794 if (s->s_mds >= mdsc->max_sessions ||
795 mdsc->sessions[s->s_mds] != s)
801 * create+register a new session for given mds.
802 * called under mdsc->mutex.
804 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
807 struct ceph_mds_session *s;
809 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO)
810 return ERR_PTR(-EIO);
812 if (mds >= mdsc->mdsmap->possible_max_rank)
813 return ERR_PTR(-EINVAL);
815 s = kzalloc(sizeof(*s), GFP_NOFS);
817 return ERR_PTR(-ENOMEM);
819 if (mds >= mdsc->max_sessions) {
820 int newmax = 1 << get_count_order(mds + 1);
821 struct ceph_mds_session **sa;
823 dout("%s: realloc to %d\n", __func__, newmax);
824 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
827 if (mdsc->sessions) {
828 memcpy(sa, mdsc->sessions,
829 mdsc->max_sessions * sizeof(void *));
830 kfree(mdsc->sessions);
833 mdsc->max_sessions = newmax;
836 dout("%s: mds%d\n", __func__, mds);
839 s->s_state = CEPH_MDS_SESSION_NEW;
840 mutex_init(&s->s_mutex);
842 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
844 atomic_set(&s->s_cap_gen, 1);
845 s->s_cap_ttl = jiffies - 1;
847 spin_lock_init(&s->s_cap_lock);
848 INIT_LIST_HEAD(&s->s_caps);
849 refcount_set(&s->s_ref, 1);
850 INIT_LIST_HEAD(&s->s_waiting);
851 INIT_LIST_HEAD(&s->s_unsafe);
852 xa_init(&s->s_delegated_inos);
853 INIT_LIST_HEAD(&s->s_cap_releases);
854 INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
856 INIT_LIST_HEAD(&s->s_cap_dirty);
857 INIT_LIST_HEAD(&s->s_cap_flushing);
859 mdsc->sessions[mds] = s;
860 atomic_inc(&mdsc->num_sessions);
861 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
863 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
864 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
870 return ERR_PTR(-ENOMEM);
874 * called under mdsc->mutex
876 static void __unregister_session(struct ceph_mds_client *mdsc,
877 struct ceph_mds_session *s)
879 dout("__unregister_session mds%d %p\n", s->s_mds, s);
880 BUG_ON(mdsc->sessions[s->s_mds] != s);
881 mdsc->sessions[s->s_mds] = NULL;
882 ceph_con_close(&s->s_con);
883 ceph_put_mds_session(s);
884 atomic_dec(&mdsc->num_sessions);
888 * drop session refs in request.
890 * should be last request ref, or hold mdsc->mutex
892 static void put_request_session(struct ceph_mds_request *req)
894 if (req->r_session) {
895 ceph_put_mds_session(req->r_session);
896 req->r_session = NULL;
900 void ceph_mdsc_iterate_sessions(struct ceph_mds_client *mdsc,
901 void (*cb)(struct ceph_mds_session *),
906 mutex_lock(&mdsc->mutex);
907 for (mds = 0; mds < mdsc->max_sessions; ++mds) {
908 struct ceph_mds_session *s;
910 s = __ceph_lookup_mds_session(mdsc, mds);
914 if (check_state && !check_session_state(s)) {
915 ceph_put_mds_session(s);
919 mutex_unlock(&mdsc->mutex);
921 ceph_put_mds_session(s);
922 mutex_lock(&mdsc->mutex);
924 mutex_unlock(&mdsc->mutex);
927 void ceph_mdsc_release_request(struct kref *kref)
929 struct ceph_mds_request *req = container_of(kref,
930 struct ceph_mds_request,
932 ceph_mdsc_release_dir_caps_no_check(req);
933 destroy_reply_info(&req->r_reply_info);
935 ceph_msg_put(req->r_request);
937 ceph_msg_put(req->r_reply);
939 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
943 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
946 iput(req->r_target_inode);
949 if (req->r_old_dentry)
950 dput(req->r_old_dentry);
951 if (req->r_old_dentry_dir) {
953 * track (and drop pins for) r_old_dentry_dir
954 * separately, since r_old_dentry's d_parent may have
955 * changed between the dir mutex being dropped and
956 * this request being freed.
958 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
960 iput(req->r_old_dentry_dir);
964 put_cred(req->r_cred);
966 ceph_pagelist_release(req->r_pagelist);
967 put_request_session(req);
968 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
969 WARN_ON_ONCE(!list_empty(&req->r_wait));
970 kmem_cache_free(ceph_mds_request_cachep, req);
973 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
976 * lookup session, bump ref if found.
978 * called under mdsc->mutex.
980 static struct ceph_mds_request *
981 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
983 struct ceph_mds_request *req;
985 req = lookup_request(&mdsc->request_tree, tid);
987 ceph_mdsc_get_request(req);
993 * Register an in-flight request, and assign a tid. Link to directory
994 * are modifying (if any).
996 * Called under mdsc->mutex.
998 static void __register_request(struct ceph_mds_client *mdsc,
999 struct ceph_mds_request *req,
1004 req->r_tid = ++mdsc->last_tid;
1005 if (req->r_num_caps) {
1006 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
1009 pr_err("__register_request %p "
1010 "failed to reserve caps: %d\n", req, ret);
1011 /* set req->r_err to fail early from __do_request */
1016 dout("__register_request %p tid %lld\n", req, req->r_tid);
1017 ceph_mdsc_get_request(req);
1018 insert_request(&mdsc->request_tree, req);
1020 req->r_cred = get_current_cred();
1022 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
1023 mdsc->oldest_tid = req->r_tid;
1026 struct ceph_inode_info *ci = ceph_inode(dir);
1029 req->r_unsafe_dir = dir;
1030 spin_lock(&ci->i_unsafe_lock);
1031 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
1032 spin_unlock(&ci->i_unsafe_lock);
1036 static void __unregister_request(struct ceph_mds_client *mdsc,
1037 struct ceph_mds_request *req)
1039 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1041 /* Never leave an unregistered request on an unsafe list! */
1042 list_del_init(&req->r_unsafe_item);
1044 if (req->r_tid == mdsc->oldest_tid) {
1045 struct rb_node *p = rb_next(&req->r_node);
1046 mdsc->oldest_tid = 0;
1048 struct ceph_mds_request *next_req =
1049 rb_entry(p, struct ceph_mds_request, r_node);
1050 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
1051 mdsc->oldest_tid = next_req->r_tid;
1058 erase_request(&mdsc->request_tree, req);
1060 if (req->r_unsafe_dir) {
1061 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
1062 spin_lock(&ci->i_unsafe_lock);
1063 list_del_init(&req->r_unsafe_dir_item);
1064 spin_unlock(&ci->i_unsafe_lock);
1066 if (req->r_target_inode &&
1067 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
1068 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
1069 spin_lock(&ci->i_unsafe_lock);
1070 list_del_init(&req->r_unsafe_target_item);
1071 spin_unlock(&ci->i_unsafe_lock);
1074 if (req->r_unsafe_dir) {
1075 iput(req->r_unsafe_dir);
1076 req->r_unsafe_dir = NULL;
1079 complete_all(&req->r_safe_completion);
1081 ceph_mdsc_put_request(req);
1085 * Walk back up the dentry tree until we hit a dentry representing a
1086 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
1087 * when calling this) to ensure that the objects won't disappear while we're
1088 * working with them. Once we hit a candidate dentry, we attempt to take a
1089 * reference to it, and return that as the result.
1091 static struct inode *get_nonsnap_parent(struct dentry *dentry)
1093 struct inode *inode = NULL;
1095 while (dentry && !IS_ROOT(dentry)) {
1096 inode = d_inode_rcu(dentry);
1097 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
1099 dentry = dentry->d_parent;
1102 inode = igrab(inode);
1107 * Choose mds to send request to next. If there is a hint set in the
1108 * request (e.g., due to a prior forward hint from the mds), use that.
1109 * Otherwise, consult frag tree and/or caps to identify the
1110 * appropriate mds. If all else fails, choose randomly.
1112 * Called under mdsc->mutex.
1114 static int __choose_mds(struct ceph_mds_client *mdsc,
1115 struct ceph_mds_request *req,
1118 struct inode *inode;
1119 struct ceph_inode_info *ci;
1120 struct ceph_cap *cap;
1121 int mode = req->r_direct_mode;
1123 u32 hash = req->r_direct_hash;
1124 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
1130 * is there a specific mds we should try? ignore hint if we have
1131 * no session and the mds is not up (active or recovering).
1133 if (req->r_resend_mds >= 0 &&
1134 (__have_session(mdsc, req->r_resend_mds) ||
1135 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1136 dout("%s using resend_mds mds%d\n", __func__,
1138 return req->r_resend_mds;
1141 if (mode == USE_RANDOM_MDS)
1146 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1147 inode = req->r_inode;
1150 /* req->r_dentry is non-null for LSSNAP request */
1152 inode = get_nonsnap_parent(req->r_dentry);
1154 dout("%s using snapdir's parent %p\n", __func__, inode);
1156 } else if (req->r_dentry) {
1157 /* ignore race with rename; old or new d_parent is okay */
1158 struct dentry *parent;
1162 parent = READ_ONCE(req->r_dentry->d_parent);
1163 dir = req->r_parent ? : d_inode_rcu(parent);
1165 if (!dir || dir->i_sb != mdsc->fsc->sb) {
1166 /* not this fs or parent went negative */
1167 inode = d_inode(req->r_dentry);
1170 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
1171 /* direct snapped/virtual snapdir requests
1172 * based on parent dir inode */
1173 inode = get_nonsnap_parent(parent);
1174 dout("%s using nonsnap parent %p\n", __func__, inode);
1177 inode = d_inode(req->r_dentry);
1178 if (!inode || mode == USE_AUTH_MDS) {
1181 hash = ceph_dentry_hash(dir, req->r_dentry);
1190 dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
1194 ci = ceph_inode(inode);
1196 if (is_hash && S_ISDIR(inode->i_mode)) {
1197 struct ceph_inode_frag frag;
1200 ceph_choose_frag(ci, hash, &frag, &found);
1202 if (mode == USE_ANY_MDS && frag.ndist > 0) {
1205 /* choose a random replica */
1206 get_random_bytes(&r, 1);
1209 dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
1210 __func__, inode, ceph_vinop(inode),
1211 frag.frag, mds, (int)r, frag.ndist);
1212 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1213 CEPH_MDS_STATE_ACTIVE &&
1214 !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1218 /* since this file/dir wasn't known to be
1219 * replicated, then we want to look for the
1220 * authoritative mds. */
1221 if (frag.mds >= 0) {
1222 /* choose auth mds */
1224 dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
1225 __func__, inode, ceph_vinop(inode),
1227 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1228 CEPH_MDS_STATE_ACTIVE) {
1229 if (!ceph_mdsmap_is_laggy(mdsc->mdsmap,
1234 mode = USE_AUTH_MDS;
1238 spin_lock(&ci->i_ceph_lock);
1240 if (mode == USE_AUTH_MDS)
1241 cap = ci->i_auth_cap;
1242 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1243 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1245 spin_unlock(&ci->i_ceph_lock);
1249 mds = cap->session->s_mds;
1250 dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1251 inode, ceph_vinop(inode), mds,
1252 cap == ci->i_auth_cap ? "auth " : "", cap);
1253 spin_unlock(&ci->i_ceph_lock);
1262 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1263 dout("%s chose random mds%d\n", __func__, mds);
1271 struct ceph_msg *ceph_create_session_msg(u32 op, u64 seq)
1273 struct ceph_msg *msg;
1274 struct ceph_mds_session_head *h;
1276 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1279 pr_err("ENOMEM creating session %s msg\n",
1280 ceph_session_op_name(op));
1283 h = msg->front.iov_base;
1284 h->op = cpu_to_le32(op);
1285 h->seq = cpu_to_le64(seq);
1290 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1291 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1292 static int encode_supported_features(void **p, void *end)
1294 static const size_t count = ARRAY_SIZE(feature_bits);
1298 size_t size = FEATURE_BYTES(count);
1301 if (WARN_ON_ONCE(*p + 4 + size > end))
1304 ceph_encode_32(p, size);
1305 memset(*p, 0, size);
1306 for (i = 0; i < count; i++) {
1307 bit = feature_bits[i];
1308 ((unsigned char *)(*p))[bit / 8] |= BIT(bit % 8);
1312 if (WARN_ON_ONCE(*p + 4 > end))
1315 ceph_encode_32(p, 0);
1321 static const unsigned char metric_bits[] = CEPHFS_METRIC_SPEC_CLIENT_SUPPORTED;
1322 #define METRIC_BYTES(cnt) (DIV_ROUND_UP((size_t)metric_bits[cnt - 1] + 1, 64) * 8)
1323 static int encode_metric_spec(void **p, void *end)
1325 static const size_t count = ARRAY_SIZE(metric_bits);
1328 if (WARN_ON_ONCE(*p + 2 > end))
1331 ceph_encode_8(p, 1); /* version */
1332 ceph_encode_8(p, 1); /* compat */
1336 size_t size = METRIC_BYTES(count);
1338 if (WARN_ON_ONCE(*p + 4 + 4 + size > end))
1341 /* metric spec info length */
1342 ceph_encode_32(p, 4 + size);
1345 ceph_encode_32(p, size);
1346 memset(*p, 0, size);
1347 for (i = 0; i < count; i++)
1348 ((unsigned char *)(*p))[i / 8] |= BIT(metric_bits[i] % 8);
1351 if (WARN_ON_ONCE(*p + 4 + 4 > end))
1354 /* metric spec info length */
1355 ceph_encode_32(p, 4);
1357 ceph_encode_32(p, 0);
1364 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1365 * to include additional client metadata fields.
1367 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1369 struct ceph_msg *msg;
1370 struct ceph_mds_session_head *h;
1372 int extra_bytes = 0;
1373 int metadata_key_count = 0;
1374 struct ceph_options *opt = mdsc->fsc->client->options;
1375 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1380 const char* metadata[][2] = {
1381 {"hostname", mdsc->nodename},
1382 {"kernel_version", init_utsname()->release},
1383 {"entity_id", opt->name ? : ""},
1384 {"root", fsopt->server_path ? : "/"},
1388 /* Calculate serialized length of metadata */
1389 extra_bytes = 4; /* map length */
1390 for (i = 0; metadata[i][0]; ++i) {
1391 extra_bytes += 8 + strlen(metadata[i][0]) +
1392 strlen(metadata[i][1]);
1393 metadata_key_count++;
1396 /* supported feature */
1398 count = ARRAY_SIZE(feature_bits);
1400 size = FEATURE_BYTES(count);
1401 extra_bytes += 4 + size;
1405 count = ARRAY_SIZE(metric_bits);
1407 size = METRIC_BYTES(count);
1408 extra_bytes += 2 + 4 + 4 + size;
1410 /* Allocate the message */
1411 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1414 pr_err("ENOMEM creating session open msg\n");
1415 return ERR_PTR(-ENOMEM);
1417 p = msg->front.iov_base;
1418 end = p + msg->front.iov_len;
1421 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1422 h->seq = cpu_to_le64(seq);
1425 * Serialize client metadata into waiting buffer space, using
1426 * the format that userspace expects for map<string, string>
1428 * ClientSession messages with metadata are v4
1430 msg->hdr.version = cpu_to_le16(4);
1431 msg->hdr.compat_version = cpu_to_le16(1);
1433 /* The write pointer, following the session_head structure */
1436 /* Number of entries in the map */
1437 ceph_encode_32(&p, metadata_key_count);
1439 /* Two length-prefixed strings for each entry in the map */
1440 for (i = 0; metadata[i][0]; ++i) {
1441 size_t const key_len = strlen(metadata[i][0]);
1442 size_t const val_len = strlen(metadata[i][1]);
1444 ceph_encode_32(&p, key_len);
1445 memcpy(p, metadata[i][0], key_len);
1447 ceph_encode_32(&p, val_len);
1448 memcpy(p, metadata[i][1], val_len);
1452 ret = encode_supported_features(&p, end);
1454 pr_err("encode_supported_features failed!\n");
1456 return ERR_PTR(ret);
1459 ret = encode_metric_spec(&p, end);
1461 pr_err("encode_metric_spec failed!\n");
1463 return ERR_PTR(ret);
1466 msg->front.iov_len = p - msg->front.iov_base;
1467 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1473 * send session open request.
1475 * called under mdsc->mutex
1477 static int __open_session(struct ceph_mds_client *mdsc,
1478 struct ceph_mds_session *session)
1480 struct ceph_msg *msg;
1482 int mds = session->s_mds;
1484 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO)
1487 /* wait for mds to go active? */
1488 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1489 dout("open_session to mds%d (%s)\n", mds,
1490 ceph_mds_state_name(mstate));
1491 session->s_state = CEPH_MDS_SESSION_OPENING;
1492 session->s_renew_requested = jiffies;
1494 /* send connect message */
1495 msg = create_session_open_msg(mdsc, session->s_seq);
1497 return PTR_ERR(msg);
1498 ceph_con_send(&session->s_con, msg);
1503 * open sessions for any export targets for the given mds
1505 * called under mdsc->mutex
1507 static struct ceph_mds_session *
1508 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1510 struct ceph_mds_session *session;
1513 session = __ceph_lookup_mds_session(mdsc, target);
1515 session = register_session(mdsc, target);
1516 if (IS_ERR(session))
1519 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1520 session->s_state == CEPH_MDS_SESSION_CLOSING) {
1521 ret = __open_session(mdsc, session);
1523 return ERR_PTR(ret);
1529 struct ceph_mds_session *
1530 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1532 struct ceph_mds_session *session;
1534 dout("open_export_target_session to mds%d\n", target);
1536 mutex_lock(&mdsc->mutex);
1537 session = __open_export_target_session(mdsc, target);
1538 mutex_unlock(&mdsc->mutex);
1543 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1544 struct ceph_mds_session *session)
1546 struct ceph_mds_info *mi;
1547 struct ceph_mds_session *ts;
1548 int i, mds = session->s_mds;
1550 if (mds >= mdsc->mdsmap->possible_max_rank)
1553 mi = &mdsc->mdsmap->m_info[mds];
1554 dout("open_export_target_sessions for mds%d (%d targets)\n",
1555 session->s_mds, mi->num_export_targets);
1557 for (i = 0; i < mi->num_export_targets; i++) {
1558 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1559 ceph_put_mds_session(ts);
1563 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1564 struct ceph_mds_session *session)
1566 mutex_lock(&mdsc->mutex);
1567 __open_export_target_sessions(mdsc, session);
1568 mutex_unlock(&mdsc->mutex);
1575 static void detach_cap_releases(struct ceph_mds_session *session,
1576 struct list_head *target)
1578 lockdep_assert_held(&session->s_cap_lock);
1580 list_splice_init(&session->s_cap_releases, target);
1581 session->s_num_cap_releases = 0;
1582 dout("dispose_cap_releases mds%d\n", session->s_mds);
1585 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1586 struct list_head *dispose)
1588 while (!list_empty(dispose)) {
1589 struct ceph_cap *cap;
1590 /* zero out the in-progress message */
1591 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1592 list_del(&cap->session_caps);
1593 ceph_put_cap(mdsc, cap);
1597 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1598 struct ceph_mds_session *session)
1600 struct ceph_mds_request *req;
1603 dout("cleanup_session_requests mds%d\n", session->s_mds);
1604 mutex_lock(&mdsc->mutex);
1605 while (!list_empty(&session->s_unsafe)) {
1606 req = list_first_entry(&session->s_unsafe,
1607 struct ceph_mds_request, r_unsafe_item);
1608 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1610 if (req->r_target_inode)
1611 mapping_set_error(req->r_target_inode->i_mapping, -EIO);
1612 if (req->r_unsafe_dir)
1613 mapping_set_error(req->r_unsafe_dir->i_mapping, -EIO);
1614 __unregister_request(mdsc, req);
1616 /* zero r_attempts, so kick_requests() will re-send requests */
1617 p = rb_first(&mdsc->request_tree);
1619 req = rb_entry(p, struct ceph_mds_request, r_node);
1621 if (req->r_session &&
1622 req->r_session->s_mds == session->s_mds)
1623 req->r_attempts = 0;
1625 mutex_unlock(&mdsc->mutex);
1629 * Helper to safely iterate over all caps associated with a session, with
1630 * special care taken to handle a racing __ceph_remove_cap().
1632 * Caller must hold session s_mutex.
1634 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1635 int (*cb)(struct inode *, struct ceph_cap *,
1638 struct list_head *p;
1639 struct ceph_cap *cap;
1640 struct inode *inode, *last_inode = NULL;
1641 struct ceph_cap *old_cap = NULL;
1644 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1645 spin_lock(&session->s_cap_lock);
1646 p = session->s_caps.next;
1647 while (p != &session->s_caps) {
1648 cap = list_entry(p, struct ceph_cap, session_caps);
1649 inode = igrab(&cap->ci->netfs.inode);
1654 session->s_cap_iterator = cap;
1655 spin_unlock(&session->s_cap_lock);
1662 ceph_put_cap(session->s_mdsc, old_cap);
1666 ret = cb(inode, cap, arg);
1669 spin_lock(&session->s_cap_lock);
1672 dout("iterate_session_caps finishing cap %p removal\n",
1674 BUG_ON(cap->session != session);
1675 cap->session = NULL;
1676 list_del_init(&cap->session_caps);
1677 session->s_nr_caps--;
1678 atomic64_dec(&session->s_mdsc->metric.total_caps);
1679 if (cap->queue_release)
1680 __ceph_queue_cap_release(session, cap);
1682 old_cap = cap; /* put_cap it w/o locks held */
1689 session->s_cap_iterator = NULL;
1690 spin_unlock(&session->s_cap_lock);
1694 ceph_put_cap(session->s_mdsc, old_cap);
1699 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1702 struct ceph_inode_info *ci = ceph_inode(inode);
1703 bool invalidate = false;
1706 dout("removing cap %p, ci is %p, inode is %p\n",
1707 cap, ci, &ci->netfs.inode);
1708 spin_lock(&ci->i_ceph_lock);
1709 iputs = ceph_purge_inode_cap(inode, cap, &invalidate);
1710 spin_unlock(&ci->i_ceph_lock);
1712 wake_up_all(&ci->i_cap_wq);
1714 ceph_queue_invalidate(inode);
1721 * caller must hold session s_mutex
1723 static void remove_session_caps(struct ceph_mds_session *session)
1725 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1726 struct super_block *sb = fsc->sb;
1729 dout("remove_session_caps on %p\n", session);
1730 ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1732 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1734 spin_lock(&session->s_cap_lock);
1735 if (session->s_nr_caps > 0) {
1736 struct inode *inode;
1737 struct ceph_cap *cap, *prev = NULL;
1738 struct ceph_vino vino;
1740 * iterate_session_caps() skips inodes that are being
1741 * deleted, we need to wait until deletions are complete.
1742 * __wait_on_freeing_inode() is designed for the job,
1743 * but it is not exported, so use lookup inode function
1746 while (!list_empty(&session->s_caps)) {
1747 cap = list_entry(session->s_caps.next,
1748 struct ceph_cap, session_caps);
1752 vino = cap->ci->i_vino;
1753 spin_unlock(&session->s_cap_lock);
1755 inode = ceph_find_inode(sb, vino);
1758 spin_lock(&session->s_cap_lock);
1762 // drop cap expires and unlock s_cap_lock
1763 detach_cap_releases(session, &dispose);
1765 BUG_ON(session->s_nr_caps > 0);
1766 BUG_ON(!list_empty(&session->s_cap_flushing));
1767 spin_unlock(&session->s_cap_lock);
1768 dispose_cap_releases(session->s_mdsc, &dispose);
1778 * wake up any threads waiting on this session's caps. if the cap is
1779 * old (didn't get renewed on the client reconnect), remove it now.
1781 * caller must hold s_mutex.
1783 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1786 struct ceph_inode_info *ci = ceph_inode(inode);
1787 unsigned long ev = (unsigned long)arg;
1789 if (ev == RECONNECT) {
1790 spin_lock(&ci->i_ceph_lock);
1791 ci->i_wanted_max_size = 0;
1792 ci->i_requested_max_size = 0;
1793 spin_unlock(&ci->i_ceph_lock);
1794 } else if (ev == RENEWCAPS) {
1795 if (cap->cap_gen < atomic_read(&cap->session->s_cap_gen)) {
1796 /* mds did not re-issue stale cap */
1797 spin_lock(&ci->i_ceph_lock);
1798 cap->issued = cap->implemented = CEPH_CAP_PIN;
1799 spin_unlock(&ci->i_ceph_lock);
1801 } else if (ev == FORCE_RO) {
1803 wake_up_all(&ci->i_cap_wq);
1807 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1809 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1810 ceph_iterate_session_caps(session, wake_up_session_cb,
1811 (void *)(unsigned long)ev);
1815 * Send periodic message to MDS renewing all currently held caps. The
1816 * ack will reset the expiration for all caps from this session.
1818 * caller holds s_mutex
1820 static int send_renew_caps(struct ceph_mds_client *mdsc,
1821 struct ceph_mds_session *session)
1823 struct ceph_msg *msg;
1826 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1827 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1828 pr_info("mds%d caps stale\n", session->s_mds);
1829 session->s_renew_requested = jiffies;
1831 /* do not try to renew caps until a recovering mds has reconnected
1832 * with its clients. */
1833 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1834 if (state < CEPH_MDS_STATE_RECONNECT) {
1835 dout("send_renew_caps ignoring mds%d (%s)\n",
1836 session->s_mds, ceph_mds_state_name(state));
1840 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1841 ceph_mds_state_name(state));
1842 msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1843 ++session->s_renew_seq);
1846 ceph_con_send(&session->s_con, msg);
1850 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1851 struct ceph_mds_session *session, u64 seq)
1853 struct ceph_msg *msg;
1855 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1856 session->s_mds, ceph_session_state_name(session->s_state), seq);
1857 msg = ceph_create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1860 ceph_con_send(&session->s_con, msg);
1866 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1868 * Called under session->s_mutex
1870 static void renewed_caps(struct ceph_mds_client *mdsc,
1871 struct ceph_mds_session *session, int is_renew)
1876 spin_lock(&session->s_cap_lock);
1877 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1879 session->s_cap_ttl = session->s_renew_requested +
1880 mdsc->mdsmap->m_session_timeout*HZ;
1883 if (time_before(jiffies, session->s_cap_ttl)) {
1884 pr_info("mds%d caps renewed\n", session->s_mds);
1887 pr_info("mds%d caps still stale\n", session->s_mds);
1890 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1891 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1892 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1893 spin_unlock(&session->s_cap_lock);
1896 wake_up_session_caps(session, RENEWCAPS);
1900 * send a session close request
1902 static int request_close_session(struct ceph_mds_session *session)
1904 struct ceph_msg *msg;
1906 dout("request_close_session mds%d state %s seq %lld\n",
1907 session->s_mds, ceph_session_state_name(session->s_state),
1909 msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_CLOSE,
1913 ceph_con_send(&session->s_con, msg);
1918 * Called with s_mutex held.
1920 static int __close_session(struct ceph_mds_client *mdsc,
1921 struct ceph_mds_session *session)
1923 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1925 session->s_state = CEPH_MDS_SESSION_CLOSING;
1926 return request_close_session(session);
1929 static bool drop_negative_children(struct dentry *dentry)
1931 struct dentry *child;
1932 bool all_negative = true;
1934 if (!d_is_dir(dentry))
1937 spin_lock(&dentry->d_lock);
1938 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1939 if (d_really_is_positive(child)) {
1940 all_negative = false;
1944 spin_unlock(&dentry->d_lock);
1947 shrink_dcache_parent(dentry);
1949 return all_negative;
1953 * Trim old(er) caps.
1955 * Because we can't cache an inode without one or more caps, we do
1956 * this indirectly: if a cap is unused, we prune its aliases, at which
1957 * point the inode will hopefully get dropped to.
1959 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1960 * memory pressure from the MDS, though, so it needn't be perfect.
1962 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1964 int *remaining = arg;
1965 struct ceph_inode_info *ci = ceph_inode(inode);
1966 int used, wanted, oissued, mine;
1968 if (*remaining <= 0)
1971 spin_lock(&ci->i_ceph_lock);
1972 mine = cap->issued | cap->implemented;
1973 used = __ceph_caps_used(ci);
1974 wanted = __ceph_caps_file_wanted(ci);
1975 oissued = __ceph_caps_issued_other(ci, cap);
1977 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1978 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1979 ceph_cap_string(used), ceph_cap_string(wanted));
1980 if (cap == ci->i_auth_cap) {
1981 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1982 !list_empty(&ci->i_cap_snaps))
1984 if ((used | wanted) & CEPH_CAP_ANY_WR)
1986 /* Note: it's possible that i_filelock_ref becomes non-zero
1987 * after dropping auth caps. It doesn't hurt because reply
1988 * of lock mds request will re-add auth caps. */
1989 if (atomic_read(&ci->i_filelock_ref) > 0)
1992 /* The inode has cached pages, but it's no longer used.
1993 * we can safely drop it */
1994 if (S_ISREG(inode->i_mode) &&
1995 wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1996 !(oissued & CEPH_CAP_FILE_CACHE)) {
2000 if ((used | wanted) & ~oissued & mine)
2001 goto out; /* we need these caps */
2004 /* we aren't the only cap.. just remove us */
2005 ceph_remove_cap(cap, true);
2008 struct dentry *dentry;
2009 /* try dropping referring dentries */
2010 spin_unlock(&ci->i_ceph_lock);
2011 dentry = d_find_any_alias(inode);
2012 if (dentry && drop_negative_children(dentry)) {
2015 d_prune_aliases(inode);
2016 count = atomic_read(&inode->i_count);
2019 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
2028 spin_unlock(&ci->i_ceph_lock);
2033 * Trim session cap count down to some max number.
2035 int ceph_trim_caps(struct ceph_mds_client *mdsc,
2036 struct ceph_mds_session *session,
2039 int trim_caps = session->s_nr_caps - max_caps;
2041 dout("trim_caps mds%d start: %d / %d, trim %d\n",
2042 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
2043 if (trim_caps > 0) {
2044 int remaining = trim_caps;
2046 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
2047 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
2048 session->s_mds, session->s_nr_caps, max_caps,
2049 trim_caps - remaining);
2052 ceph_flush_cap_releases(mdsc, session);
2056 static int check_caps_flush(struct ceph_mds_client *mdsc,
2061 spin_lock(&mdsc->cap_dirty_lock);
2062 if (!list_empty(&mdsc->cap_flush_list)) {
2063 struct ceph_cap_flush *cf =
2064 list_first_entry(&mdsc->cap_flush_list,
2065 struct ceph_cap_flush, g_list);
2066 if (cf->tid <= want_flush_tid) {
2067 dout("check_caps_flush still flushing tid "
2068 "%llu <= %llu\n", cf->tid, want_flush_tid);
2072 spin_unlock(&mdsc->cap_dirty_lock);
2077 * flush all dirty inode data to disk.
2079 * returns true if we've flushed through want_flush_tid
2081 static void wait_caps_flush(struct ceph_mds_client *mdsc,
2084 dout("check_caps_flush want %llu\n", want_flush_tid);
2086 wait_event(mdsc->cap_flushing_wq,
2087 check_caps_flush(mdsc, want_flush_tid));
2089 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
2093 * called under s_mutex
2095 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2096 struct ceph_mds_session *session)
2098 struct ceph_msg *msg = NULL;
2099 struct ceph_mds_cap_release *head;
2100 struct ceph_mds_cap_item *item;
2101 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2102 struct ceph_cap *cap;
2103 LIST_HEAD(tmp_list);
2104 int num_cap_releases;
2105 __le32 barrier, *cap_barrier;
2107 down_read(&osdc->lock);
2108 barrier = cpu_to_le32(osdc->epoch_barrier);
2109 up_read(&osdc->lock);
2111 spin_lock(&session->s_cap_lock);
2113 list_splice_init(&session->s_cap_releases, &tmp_list);
2114 num_cap_releases = session->s_num_cap_releases;
2115 session->s_num_cap_releases = 0;
2116 spin_unlock(&session->s_cap_lock);
2118 while (!list_empty(&tmp_list)) {
2120 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2121 PAGE_SIZE, GFP_NOFS, false);
2124 head = msg->front.iov_base;
2125 head->num = cpu_to_le32(0);
2126 msg->front.iov_len = sizeof(*head);
2128 msg->hdr.version = cpu_to_le16(2);
2129 msg->hdr.compat_version = cpu_to_le16(1);
2132 cap = list_first_entry(&tmp_list, struct ceph_cap,
2134 list_del(&cap->session_caps);
2137 head = msg->front.iov_base;
2138 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2140 item = msg->front.iov_base + msg->front.iov_len;
2141 item->ino = cpu_to_le64(cap->cap_ino);
2142 item->cap_id = cpu_to_le64(cap->cap_id);
2143 item->migrate_seq = cpu_to_le32(cap->mseq);
2144 item->seq = cpu_to_le32(cap->issue_seq);
2145 msg->front.iov_len += sizeof(*item);
2147 ceph_put_cap(mdsc, cap);
2149 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2150 // Append cap_barrier field
2151 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2152 *cap_barrier = barrier;
2153 msg->front.iov_len += sizeof(*cap_barrier);
2155 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2156 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2157 ceph_con_send(&session->s_con, msg);
2162 BUG_ON(num_cap_releases != 0);
2164 spin_lock(&session->s_cap_lock);
2165 if (!list_empty(&session->s_cap_releases))
2167 spin_unlock(&session->s_cap_lock);
2170 // Append cap_barrier field
2171 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2172 *cap_barrier = barrier;
2173 msg->front.iov_len += sizeof(*cap_barrier);
2175 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2176 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2177 ceph_con_send(&session->s_con, msg);
2181 pr_err("send_cap_releases mds%d, failed to allocate message\n",
2183 spin_lock(&session->s_cap_lock);
2184 list_splice(&tmp_list, &session->s_cap_releases);
2185 session->s_num_cap_releases += num_cap_releases;
2186 spin_unlock(&session->s_cap_lock);
2189 static void ceph_cap_release_work(struct work_struct *work)
2191 struct ceph_mds_session *session =
2192 container_of(work, struct ceph_mds_session, s_cap_release_work);
2194 mutex_lock(&session->s_mutex);
2195 if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2196 session->s_state == CEPH_MDS_SESSION_HUNG)
2197 ceph_send_cap_releases(session->s_mdsc, session);
2198 mutex_unlock(&session->s_mutex);
2199 ceph_put_mds_session(session);
2202 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2203 struct ceph_mds_session *session)
2208 ceph_get_mds_session(session);
2209 if (queue_work(mdsc->fsc->cap_wq,
2210 &session->s_cap_release_work)) {
2211 dout("cap release work queued\n");
2213 ceph_put_mds_session(session);
2214 dout("failed to queue cap release work\n");
2219 * caller holds session->s_cap_lock
2221 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2222 struct ceph_cap *cap)
2224 list_add_tail(&cap->session_caps, &session->s_cap_releases);
2225 session->s_num_cap_releases++;
2227 if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2228 ceph_flush_cap_releases(session->s_mdsc, session);
2231 static void ceph_cap_reclaim_work(struct work_struct *work)
2233 struct ceph_mds_client *mdsc =
2234 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2235 int ret = ceph_trim_dentries(mdsc);
2237 ceph_queue_cap_reclaim_work(mdsc);
2240 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2245 if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2246 dout("caps reclaim work queued\n");
2248 dout("failed to queue caps release work\n");
2252 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2257 val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2258 if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2259 atomic_set(&mdsc->cap_reclaim_pending, 0);
2260 ceph_queue_cap_reclaim_work(mdsc);
2268 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2271 struct ceph_inode_info *ci = ceph_inode(dir);
2272 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2273 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2274 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2275 unsigned int num_entries;
2278 spin_lock(&ci->i_ceph_lock);
2279 num_entries = ci->i_files + ci->i_subdirs;
2280 spin_unlock(&ci->i_ceph_lock);
2281 num_entries = max(num_entries, 1U);
2282 num_entries = min(num_entries, opt->max_readdir);
2284 order = get_order(size * num_entries);
2285 while (order >= 0) {
2286 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2290 if (rinfo->dir_entries)
2294 if (!rinfo->dir_entries)
2297 num_entries = (PAGE_SIZE << order) / size;
2298 num_entries = min(num_entries, opt->max_readdir);
2300 rinfo->dir_buf_size = PAGE_SIZE << order;
2301 req->r_num_caps = num_entries + 1;
2302 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2303 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2308 * Create an mds request.
2310 struct ceph_mds_request *
2311 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2313 struct ceph_mds_request *req;
2315 req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2317 return ERR_PTR(-ENOMEM);
2319 mutex_init(&req->r_fill_mutex);
2321 req->r_started = jiffies;
2322 req->r_start_latency = ktime_get();
2323 req->r_resend_mds = -1;
2324 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2325 INIT_LIST_HEAD(&req->r_unsafe_target_item);
2327 req->r_feature_needed = -1;
2328 kref_init(&req->r_kref);
2329 RB_CLEAR_NODE(&req->r_node);
2330 INIT_LIST_HEAD(&req->r_wait);
2331 init_completion(&req->r_completion);
2332 init_completion(&req->r_safe_completion);
2333 INIT_LIST_HEAD(&req->r_unsafe_item);
2335 ktime_get_coarse_real_ts64(&req->r_stamp);
2338 req->r_direct_mode = mode;
2343 * return oldest (lowest) request, tid in request tree, 0 if none.
2345 * called under mdsc->mutex.
2347 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2349 if (RB_EMPTY_ROOT(&mdsc->request_tree))
2351 return rb_entry(rb_first(&mdsc->request_tree),
2352 struct ceph_mds_request, r_node);
2355 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2357 return mdsc->oldest_tid;
2361 * Build a dentry's path. Allocate on heap; caller must kfree. Based
2362 * on build_path_from_dentry in fs/cifs/dir.c.
2364 * If @stop_on_nosnap, generate path relative to the first non-snapped
2367 * Encode hidden .snap dirs as a double /, i.e.
2368 * foo/.snap/bar -> foo//bar
2370 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2373 struct dentry *temp;
2380 return ERR_PTR(-EINVAL);
2384 return ERR_PTR(-ENOMEM);
2389 seq = read_seqbegin(&rename_lock);
2393 struct inode *inode;
2395 spin_lock(&temp->d_lock);
2396 inode = d_inode(temp);
2397 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2398 dout("build_path path+%d: %p SNAPDIR\n",
2400 } else if (stop_on_nosnap && inode && dentry != temp &&
2401 ceph_snap(inode) == CEPH_NOSNAP) {
2402 spin_unlock(&temp->d_lock);
2403 pos++; /* get rid of any prepended '/' */
2406 pos -= temp->d_name.len;
2408 spin_unlock(&temp->d_lock);
2411 memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2413 spin_unlock(&temp->d_lock);
2414 temp = READ_ONCE(temp->d_parent);
2416 /* Are we at the root? */
2420 /* Are we out of buffer? */
2426 base = ceph_ino(d_inode(temp));
2429 if (read_seqretry(&rename_lock, seq))
2434 * A rename didn't occur, but somehow we didn't end up where
2435 * we thought we would. Throw a warning and try again.
2437 pr_warn("build_path did not end path lookup where "
2438 "expected, pos is %d\n", pos);
2443 *plen = PATH_MAX - 1 - pos;
2444 dout("build_path on %p %d built %llx '%.*s'\n",
2445 dentry, d_count(dentry), base, *plen, path + pos);
2449 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2450 const char **ppath, int *ppathlen, u64 *pino,
2451 bool *pfreepath, bool parent_locked)
2457 dir = d_inode_rcu(dentry->d_parent);
2458 if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2459 *pino = ceph_ino(dir);
2461 *ppath = dentry->d_name.name;
2462 *ppathlen = dentry->d_name.len;
2466 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2468 return PTR_ERR(path);
2474 static int build_inode_path(struct inode *inode,
2475 const char **ppath, int *ppathlen, u64 *pino,
2478 struct dentry *dentry;
2481 if (ceph_snap(inode) == CEPH_NOSNAP) {
2482 *pino = ceph_ino(inode);
2486 dentry = d_find_alias(inode);
2487 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2490 return PTR_ERR(path);
2497 * request arguments may be specified via an inode *, a dentry *, or
2498 * an explicit ino+path.
2500 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2501 struct inode *rdiri, const char *rpath,
2502 u64 rino, const char **ppath, int *pathlen,
2503 u64 *ino, bool *freepath, bool parent_locked)
2508 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2509 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2511 } else if (rdentry) {
2512 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2513 freepath, parent_locked);
2514 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2516 } else if (rpath || rino) {
2519 *pathlen = rpath ? strlen(rpath) : 0;
2520 dout(" path %.*s\n", *pathlen, rpath);
2526 static void encode_timestamp_and_gids(void **p,
2527 const struct ceph_mds_request *req)
2529 struct ceph_timespec ts;
2532 ceph_encode_timespec64(&ts, &req->r_stamp);
2533 ceph_encode_copy(p, &ts, sizeof(ts));
2536 ceph_encode_32(p, req->r_cred->group_info->ngroups);
2537 for (i = 0; i < req->r_cred->group_info->ngroups; i++)
2538 ceph_encode_64(p, from_kgid(&init_user_ns,
2539 req->r_cred->group_info->gid[i]));
2543 * called under mdsc->mutex
2545 static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
2546 struct ceph_mds_request *req,
2547 bool drop_cap_releases)
2549 int mds = session->s_mds;
2550 struct ceph_mds_client *mdsc = session->s_mdsc;
2551 struct ceph_msg *msg;
2552 struct ceph_mds_request_head_old *head;
2553 const char *path1 = NULL;
2554 const char *path2 = NULL;
2555 u64 ino1 = 0, ino2 = 0;
2556 int pathlen1 = 0, pathlen2 = 0;
2557 bool freepath1 = false, freepath2 = false;
2562 bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
2564 ret = set_request_path_attr(req->r_inode, req->r_dentry,
2565 req->r_parent, req->r_path1, req->r_ino1.ino,
2566 &path1, &pathlen1, &ino1, &freepath1,
2567 test_bit(CEPH_MDS_R_PARENT_LOCKED,
2568 &req->r_req_flags));
2574 /* If r_old_dentry is set, then assume that its parent is locked */
2575 ret = set_request_path_attr(NULL, req->r_old_dentry,
2576 req->r_old_dentry_dir,
2577 req->r_path2, req->r_ino2.ino,
2578 &path2, &pathlen2, &ino2, &freepath2, true);
2584 len = legacy ? sizeof(*head) : sizeof(struct ceph_mds_request_head);
2585 len += pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2586 sizeof(struct ceph_timespec);
2587 len += sizeof(u32) + (sizeof(u64) * req->r_cred->group_info->ngroups);
2589 /* calculate (max) length for cap releases */
2590 len += sizeof(struct ceph_mds_request_release) *
2591 (!!req->r_inode_drop + !!req->r_dentry_drop +
2592 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2594 if (req->r_dentry_drop)
2596 if (req->r_old_dentry_drop)
2599 msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2601 msg = ERR_PTR(-ENOMEM);
2605 msg->hdr.tid = cpu_to_le64(req->r_tid);
2608 * The old ceph_mds_request_head didn't contain a version field, and
2609 * one was added when we moved the message version from 3->4.
2612 msg->hdr.version = cpu_to_le16(3);
2613 head = msg->front.iov_base;
2614 p = msg->front.iov_base + sizeof(*head);
2616 struct ceph_mds_request_head *new_head = msg->front.iov_base;
2618 msg->hdr.version = cpu_to_le16(4);
2619 new_head->version = cpu_to_le16(CEPH_MDS_REQUEST_HEAD_VERSION);
2620 head = (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2621 p = msg->front.iov_base + sizeof(*new_head);
2624 end = msg->front.iov_base + msg->front.iov_len;
2626 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2627 head->op = cpu_to_le32(req->r_op);
2628 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
2629 req->r_cred->fsuid));
2630 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
2631 req->r_cred->fsgid));
2632 head->ino = cpu_to_le64(req->r_deleg_ino);
2633 head->args = req->r_args;
2635 ceph_encode_filepath(&p, end, ino1, path1);
2636 ceph_encode_filepath(&p, end, ino2, path2);
2638 /* make note of release offset, in case we need to replay */
2639 req->r_request_release_offset = p - msg->front.iov_base;
2643 if (req->r_inode_drop)
2644 releases += ceph_encode_inode_release(&p,
2645 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2646 mds, req->r_inode_drop, req->r_inode_unless,
2647 req->r_op == CEPH_MDS_OP_READDIR);
2648 if (req->r_dentry_drop)
2649 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2650 req->r_parent, mds, req->r_dentry_drop,
2651 req->r_dentry_unless);
2652 if (req->r_old_dentry_drop)
2653 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2654 req->r_old_dentry_dir, mds,
2655 req->r_old_dentry_drop,
2656 req->r_old_dentry_unless);
2657 if (req->r_old_inode_drop)
2658 releases += ceph_encode_inode_release(&p,
2659 d_inode(req->r_old_dentry),
2660 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2662 if (drop_cap_releases) {
2664 p = msg->front.iov_base + req->r_request_release_offset;
2667 head->num_releases = cpu_to_le16(releases);
2669 encode_timestamp_and_gids(&p, req);
2671 if (WARN_ON_ONCE(p > end)) {
2673 msg = ERR_PTR(-ERANGE);
2677 msg->front.iov_len = p - msg->front.iov_base;
2678 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2680 if (req->r_pagelist) {
2681 struct ceph_pagelist *pagelist = req->r_pagelist;
2682 ceph_msg_data_add_pagelist(msg, pagelist);
2683 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2685 msg->hdr.data_len = 0;
2688 msg->hdr.data_off = cpu_to_le16(0);
2692 ceph_mdsc_free_path((char *)path2, pathlen2);
2695 ceph_mdsc_free_path((char *)path1, pathlen1);
2701 * called under mdsc->mutex if error, under no mutex if
2704 static void complete_request(struct ceph_mds_client *mdsc,
2705 struct ceph_mds_request *req)
2707 req->r_end_latency = ktime_get();
2709 if (req->r_callback)
2710 req->r_callback(mdsc, req);
2711 complete_all(&req->r_completion);
2714 static struct ceph_mds_request_head_old *
2715 find_old_request_head(void *p, u64 features)
2717 bool legacy = !(features & CEPH_FEATURE_FS_BTIME);
2718 struct ceph_mds_request_head *new_head;
2721 return (struct ceph_mds_request_head_old *)p;
2722 new_head = (struct ceph_mds_request_head *)p;
2723 return (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2727 * called under mdsc->mutex
2729 static int __prepare_send_request(struct ceph_mds_session *session,
2730 struct ceph_mds_request *req,
2731 bool drop_cap_releases)
2733 int mds = session->s_mds;
2734 struct ceph_mds_client *mdsc = session->s_mdsc;
2735 struct ceph_mds_request_head_old *rhead;
2736 struct ceph_msg *msg;
2737 int flags = 0, max_retry;
2740 * The type of 'r_attempts' in kernel 'ceph_mds_request'
2741 * is 'int', while in 'ceph_mds_request_head' the type of
2742 * 'num_retry' is '__u8'. So in case the request retries
2743 * exceeding 256 times, the MDS will receive a incorrect
2746 * In this case it's ususally a bug in MDS and continue
2747 * retrying the request makes no sense.
2749 * In future this could be fixed in ceph code, so avoid
2750 * using the hardcode here.
2752 max_retry = sizeof_field(struct ceph_mds_request_head, num_retry);
2753 max_retry = 1 << (max_retry * BITS_PER_BYTE);
2754 if (req->r_attempts >= max_retry) {
2755 pr_warn_ratelimited("%s request tid %llu seq overflow\n",
2756 __func__, req->r_tid);
2762 struct ceph_cap *cap =
2763 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2766 req->r_sent_on_mseq = cap->mseq;
2768 req->r_sent_on_mseq = -1;
2770 dout("%s %p tid %lld %s (attempt %d)\n", __func__, req,
2771 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2773 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2777 * Replay. Do not regenerate message (and rebuild
2778 * paths, etc.); just use the original message.
2779 * Rebuilding paths will break for renames because
2780 * d_move mangles the src name.
2782 msg = req->r_request;
2783 rhead = find_old_request_head(msg->front.iov_base,
2784 session->s_con.peer_features);
2786 flags = le32_to_cpu(rhead->flags);
2787 flags |= CEPH_MDS_FLAG_REPLAY;
2788 rhead->flags = cpu_to_le32(flags);
2790 if (req->r_target_inode)
2791 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2793 rhead->num_retry = req->r_attempts - 1;
2795 /* remove cap/dentry releases from message */
2796 rhead->num_releases = 0;
2798 p = msg->front.iov_base + req->r_request_release_offset;
2799 encode_timestamp_and_gids(&p, req);
2801 msg->front.iov_len = p - msg->front.iov_base;
2802 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2806 if (req->r_request) {
2807 ceph_msg_put(req->r_request);
2808 req->r_request = NULL;
2810 msg = create_request_message(session, req, drop_cap_releases);
2812 req->r_err = PTR_ERR(msg);
2813 return PTR_ERR(msg);
2815 req->r_request = msg;
2817 rhead = find_old_request_head(msg->front.iov_base,
2818 session->s_con.peer_features);
2819 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2820 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2821 flags |= CEPH_MDS_FLAG_REPLAY;
2822 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2823 flags |= CEPH_MDS_FLAG_ASYNC;
2825 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2826 rhead->flags = cpu_to_le32(flags);
2827 rhead->num_fwd = req->r_num_fwd;
2828 rhead->num_retry = req->r_attempts - 1;
2830 dout(" r_parent = %p\n", req->r_parent);
2835 * called under mdsc->mutex
2837 static int __send_request(struct ceph_mds_session *session,
2838 struct ceph_mds_request *req,
2839 bool drop_cap_releases)
2843 err = __prepare_send_request(session, req, drop_cap_releases);
2845 ceph_msg_get(req->r_request);
2846 ceph_con_send(&session->s_con, req->r_request);
2853 * send request, or put it on the appropriate wait list.
2855 static void __do_request(struct ceph_mds_client *mdsc,
2856 struct ceph_mds_request *req)
2858 struct ceph_mds_session *session = NULL;
2863 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2864 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2865 __unregister_request(mdsc, req);
2869 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_FENCE_IO) {
2870 dout("do_request metadata corrupted\n");
2874 if (req->r_timeout &&
2875 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2876 dout("do_request timed out\n");
2880 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2881 dout("do_request forced umount\n");
2885 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2886 if (mdsc->mdsmap_err) {
2887 err = mdsc->mdsmap_err;
2888 dout("do_request mdsmap err %d\n", err);
2891 if (mdsc->mdsmap->m_epoch == 0) {
2892 dout("do_request no mdsmap, waiting for map\n");
2893 list_add(&req->r_wait, &mdsc->waiting_for_map);
2896 if (!(mdsc->fsc->mount_options->flags &
2897 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2898 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2899 err = -EHOSTUNREACH;
2904 put_request_session(req);
2906 mds = __choose_mds(mdsc, req, &random);
2908 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2909 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2913 dout("do_request no mds or not active, waiting for map\n");
2914 list_add(&req->r_wait, &mdsc->waiting_for_map);
2918 /* get, open session */
2919 session = __ceph_lookup_mds_session(mdsc, mds);
2921 session = register_session(mdsc, mds);
2922 if (IS_ERR(session)) {
2923 err = PTR_ERR(session);
2927 req->r_session = ceph_get_mds_session(session);
2929 dout("do_request mds%d session %p state %s\n", mds, session,
2930 ceph_session_state_name(session->s_state));
2933 * The old ceph will crash the MDSs when see unknown OPs
2935 if (req->r_feature_needed > 0 &&
2936 !test_bit(req->r_feature_needed, &session->s_features)) {
2941 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2942 session->s_state != CEPH_MDS_SESSION_HUNG) {
2944 * We cannot queue async requests since the caps and delegated
2945 * inodes are bound to the session. Just return -EJUKEBOX and
2946 * let the caller retry a sync request in that case.
2948 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2954 * If the session has been REJECTED, then return a hard error,
2955 * unless it's a CLEANRECOVER mount, in which case we'll queue
2956 * it to the mdsc queue.
2958 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2959 if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
2960 list_add(&req->r_wait, &mdsc->waiting_for_map);
2966 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2967 session->s_state == CEPH_MDS_SESSION_CLOSING) {
2968 err = __open_session(mdsc, session);
2971 /* retry the same mds later */
2973 req->r_resend_mds = mds;
2975 list_add(&req->r_wait, &session->s_waiting);
2980 req->r_resend_mds = -1; /* forget any previous mds hint */
2982 if (req->r_request_started == 0) /* note request start time */
2983 req->r_request_started = jiffies;
2986 * For async create we will choose the auth MDS of frag in parent
2987 * directory to send the request and ususally this works fine, but
2988 * if the migrated the dirtory to another MDS before it could handle
2989 * it the request will be forwarded.
2991 * And then the auth cap will be changed.
2993 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags) && req->r_num_fwd) {
2994 struct ceph_dentry_info *di = ceph_dentry(req->r_dentry);
2995 struct ceph_inode_info *ci;
2996 struct ceph_cap *cap;
2999 * The request maybe handled very fast and the new inode
3000 * hasn't been linked to the dentry yet. We need to wait
3001 * for the ceph_finish_async_create(), which shouldn't be
3002 * stuck too long or fail in thoery, to finish when forwarding
3005 if (!d_inode(req->r_dentry)) {
3006 err = wait_on_bit(&di->flags, CEPH_DENTRY_ASYNC_CREATE_BIT,
3009 mutex_lock(&req->r_fill_mutex);
3010 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3011 mutex_unlock(&req->r_fill_mutex);
3016 ci = ceph_inode(d_inode(req->r_dentry));
3018 spin_lock(&ci->i_ceph_lock);
3019 cap = ci->i_auth_cap;
3020 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE && mds != cap->mds) {
3021 dout("do_request session changed for auth cap %d -> %d\n",
3022 cap->session->s_mds, session->s_mds);
3024 /* Remove the auth cap from old session */
3025 spin_lock(&cap->session->s_cap_lock);
3026 cap->session->s_nr_caps--;
3027 list_del_init(&cap->session_caps);
3028 spin_unlock(&cap->session->s_cap_lock);
3030 /* Add the auth cap to the new session */
3032 cap->session = session;
3033 spin_lock(&session->s_cap_lock);
3034 session->s_nr_caps++;
3035 list_add_tail(&cap->session_caps, &session->s_caps);
3036 spin_unlock(&session->s_cap_lock);
3038 change_auth_cap_ses(ci, session);
3040 spin_unlock(&ci->i_ceph_lock);
3043 err = __send_request(session, req, false);
3046 ceph_put_mds_session(session);
3049 dout("__do_request early error %d\n", err);
3051 complete_request(mdsc, req);
3052 __unregister_request(mdsc, req);
3058 * called under mdsc->mutex
3060 static void __wake_requests(struct ceph_mds_client *mdsc,
3061 struct list_head *head)
3063 struct ceph_mds_request *req;
3064 LIST_HEAD(tmp_list);
3066 list_splice_init(head, &tmp_list);
3068 while (!list_empty(&tmp_list)) {
3069 req = list_entry(tmp_list.next,
3070 struct ceph_mds_request, r_wait);
3071 list_del_init(&req->r_wait);
3072 dout(" wake request %p tid %llu\n", req, req->r_tid);
3073 __do_request(mdsc, req);
3078 * Wake up threads with requests pending for @mds, so that they can
3079 * resubmit their requests to a possibly different mds.
3081 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
3083 struct ceph_mds_request *req;
3084 struct rb_node *p = rb_first(&mdsc->request_tree);
3086 dout("kick_requests mds%d\n", mds);
3088 req = rb_entry(p, struct ceph_mds_request, r_node);
3090 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3092 if (req->r_attempts > 0)
3093 continue; /* only new requests */
3094 if (req->r_session &&
3095 req->r_session->s_mds == mds) {
3096 dout(" kicking tid %llu\n", req->r_tid);
3097 list_del_init(&req->r_wait);
3098 __do_request(mdsc, req);
3103 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
3104 struct ceph_mds_request *req)
3108 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
3110 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
3111 if (req->r_parent) {
3112 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
3113 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
3114 CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
3115 spin_lock(&ci->i_ceph_lock);
3116 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
3117 __ceph_touch_fmode(ci, mdsc, fmode);
3118 spin_unlock(&ci->i_ceph_lock);
3120 if (req->r_old_dentry_dir)
3121 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
3125 err = ceph_wait_on_async_create(req->r_inode);
3127 dout("%s: wait for async create returned: %d\n",
3133 if (!err && req->r_old_inode) {
3134 err = ceph_wait_on_async_create(req->r_old_inode);
3136 dout("%s: wait for async create returned: %d\n",
3142 dout("submit_request on %p for inode %p\n", req, dir);
3143 mutex_lock(&mdsc->mutex);
3144 __register_request(mdsc, req, dir);
3145 __do_request(mdsc, req);
3147 mutex_unlock(&mdsc->mutex);
3151 int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
3152 struct ceph_mds_request *req,
3153 ceph_mds_request_wait_callback_t wait_func)
3158 dout("do_request waiting\n");
3160 err = wait_func(mdsc, req);
3162 long timeleft = wait_for_completion_killable_timeout(
3164 ceph_timeout_jiffies(req->r_timeout));
3168 err = -ETIMEDOUT; /* timed out */
3170 err = timeleft; /* killed */
3172 dout("do_request waited, got %d\n", err);
3173 mutex_lock(&mdsc->mutex);
3175 /* only abort if we didn't race with a real reply */
3176 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3177 err = le32_to_cpu(req->r_reply_info.head->result);
3178 } else if (err < 0) {
3179 dout("aborted request %lld with %d\n", req->r_tid, err);
3182 * ensure we aren't running concurrently with
3183 * ceph_fill_trace or ceph_readdir_prepopulate, which
3184 * rely on locks (dir mutex) held by our caller.
3186 mutex_lock(&req->r_fill_mutex);
3188 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3189 mutex_unlock(&req->r_fill_mutex);
3191 if (req->r_parent &&
3192 (req->r_op & CEPH_MDS_OP_WRITE))
3193 ceph_invalidate_dir_request(req);
3198 mutex_unlock(&mdsc->mutex);
3203 * Synchrously perform an mds request. Take care of all of the
3204 * session setup, forwarding, retry details.
3206 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3208 struct ceph_mds_request *req)
3212 dout("do_request on %p\n", req);
3215 err = ceph_mdsc_submit_request(mdsc, dir, req);
3217 err = ceph_mdsc_wait_request(mdsc, req, NULL);
3218 dout("do_request %p done, result %d\n", req, err);
3223 * Invalidate dir's completeness, dentry lease state on an aborted MDS
3224 * namespace request.
3226 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3228 struct inode *dir = req->r_parent;
3229 struct inode *old_dir = req->r_old_dentry_dir;
3231 dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
3233 ceph_dir_clear_complete(dir);
3235 ceph_dir_clear_complete(old_dir);
3237 ceph_invalidate_dentry_lease(req->r_dentry);
3238 if (req->r_old_dentry)
3239 ceph_invalidate_dentry_lease(req->r_old_dentry);
3245 * We take the session mutex and parse and process the reply immediately.
3246 * This preserves the logical ordering of replies, capabilities, etc., sent
3247 * by the MDS as they are applied to our local cache.
3249 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3251 struct ceph_mds_client *mdsc = session->s_mdsc;
3252 struct ceph_mds_request *req;
3253 struct ceph_mds_reply_head *head = msg->front.iov_base;
3254 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
3255 struct ceph_snap_realm *realm;
3258 int mds = session->s_mds;
3259 bool close_sessions = false;
3261 if (msg->front.iov_len < sizeof(*head)) {
3262 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
3267 /* get request, session */
3268 tid = le64_to_cpu(msg->hdr.tid);
3269 mutex_lock(&mdsc->mutex);
3270 req = lookup_get_request(mdsc, tid);
3272 dout("handle_reply on unknown tid %llu\n", tid);
3273 mutex_unlock(&mdsc->mutex);
3276 dout("handle_reply %p\n", req);
3278 /* correct session? */
3279 if (req->r_session != session) {
3280 pr_err("mdsc_handle_reply got %llu on session mds%d"
3281 " not mds%d\n", tid, session->s_mds,
3282 req->r_session ? req->r_session->s_mds : -1);
3283 mutex_unlock(&mdsc->mutex);
3288 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3289 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3290 pr_warn("got a dup %s reply on %llu from mds%d\n",
3291 head->safe ? "safe" : "unsafe", tid, mds);
3292 mutex_unlock(&mdsc->mutex);
3295 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3296 pr_warn("got unsafe after safe on %llu from mds%d\n",
3298 mutex_unlock(&mdsc->mutex);
3302 result = le32_to_cpu(head->result);
3305 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3306 __unregister_request(mdsc, req);
3308 /* last request during umount? */
3309 if (mdsc->stopping && !__get_oldest_req(mdsc))
3310 complete_all(&mdsc->safe_umount_waiters);
3312 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3314 * We already handled the unsafe response, now do the
3315 * cleanup. No need to examine the response; the MDS
3316 * doesn't include any result info in the safe
3317 * response. And even if it did, there is nothing
3318 * useful we could do with a revised return value.
3320 dout("got safe reply %llu, mds%d\n", tid, mds);
3322 mutex_unlock(&mdsc->mutex);
3326 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3327 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3330 dout("handle_reply tid %lld result %d\n", tid, result);
3331 rinfo = &req->r_reply_info;
3332 if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3333 err = parse_reply_info(session, msg, rinfo, (u64)-1);
3335 err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3336 mutex_unlock(&mdsc->mutex);
3338 /* Must find target inode outside of mutexes to avoid deadlocks */
3339 if ((err >= 0) && rinfo->head->is_target) {
3341 struct ceph_vino tvino = {
3342 .ino = le64_to_cpu(rinfo->targeti.in->ino),
3343 .snap = le64_to_cpu(rinfo->targeti.in->snapid)
3346 in = ceph_get_inode(mdsc->fsc->sb, tvino);
3349 mutex_lock(&session->s_mutex);
3352 req->r_target_inode = in;
3355 mutex_lock(&session->s_mutex);
3357 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3364 if (rinfo->snapblob_len) {
3365 down_write(&mdsc->snap_rwsem);
3366 err = ceph_update_snap_trace(mdsc, rinfo->snapblob,
3367 rinfo->snapblob + rinfo->snapblob_len,
3368 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3371 up_write(&mdsc->snap_rwsem);
3372 close_sessions = true;
3377 downgrade_write(&mdsc->snap_rwsem);
3379 down_read(&mdsc->snap_rwsem);
3382 /* insert trace into our cache */
3383 mutex_lock(&req->r_fill_mutex);
3384 current->journal_info = req;
3385 err = ceph_fill_trace(mdsc->fsc->sb, req);
3387 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3388 req->r_op == CEPH_MDS_OP_LSSNAP))
3389 ceph_readdir_prepopulate(req, req->r_session);
3391 current->journal_info = NULL;
3392 mutex_unlock(&req->r_fill_mutex);
3394 up_read(&mdsc->snap_rwsem);
3396 ceph_put_snap_realm(mdsc, realm);
3399 if (req->r_target_inode &&
3400 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3401 struct ceph_inode_info *ci =
3402 ceph_inode(req->r_target_inode);
3403 spin_lock(&ci->i_unsafe_lock);
3404 list_add_tail(&req->r_unsafe_target_item,
3405 &ci->i_unsafe_iops);
3406 spin_unlock(&ci->i_unsafe_lock);
3409 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3412 mutex_lock(&mdsc->mutex);
3413 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3417 req->r_reply = ceph_msg_get(msg);
3418 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3421 dout("reply arrived after request %lld was aborted\n", tid);
3423 mutex_unlock(&mdsc->mutex);
3425 mutex_unlock(&session->s_mutex);
3427 /* kick calling process */
3428 complete_request(mdsc, req);
3430 ceph_update_metadata_metrics(&mdsc->metric, req->r_start_latency,
3431 req->r_end_latency, err);
3433 ceph_mdsc_put_request(req);
3435 /* Defer closing the sessions after s_mutex lock being released */
3437 ceph_mdsc_close_sessions(mdsc);
3444 * handle mds notification that our request has been forwarded.
3446 static void handle_forward(struct ceph_mds_client *mdsc,
3447 struct ceph_mds_session *session,
3448 struct ceph_msg *msg)
3450 struct ceph_mds_request *req;
3451 u64 tid = le64_to_cpu(msg->hdr.tid);
3455 void *p = msg->front.iov_base;
3456 void *end = p + msg->front.iov_len;
3457 bool aborted = false;
3459 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3460 next_mds = ceph_decode_32(&p);
3461 fwd_seq = ceph_decode_32(&p);
3463 mutex_lock(&mdsc->mutex);
3464 req = lookup_get_request(mdsc, tid);
3466 mutex_unlock(&mdsc->mutex);
3467 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3468 return; /* dup reply? */
3471 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3472 dout("forward tid %llu aborted, unregistering\n", tid);
3473 __unregister_request(mdsc, req);
3474 } else if (fwd_seq <= req->r_num_fwd) {
3476 * The type of 'num_fwd' in ceph 'MClientRequestForward'
3477 * is 'int32_t', while in 'ceph_mds_request_head' the
3478 * type is '__u8'. So in case the request bounces between
3479 * MDSes exceeding 256 times, the client will get stuck.
3481 * In this case it's ususally a bug in MDS and continue
3482 * bouncing the request makes no sense.
3484 * In future this could be fixed in ceph code, so avoid
3485 * using the hardcode here.
3487 int max = sizeof_field(struct ceph_mds_request_head, num_fwd);
3488 max = 1 << (max * BITS_PER_BYTE);
3489 if (req->r_num_fwd >= max) {
3490 mutex_lock(&req->r_fill_mutex);
3491 req->r_err = -EMULTIHOP;
3492 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3493 mutex_unlock(&req->r_fill_mutex);
3495 pr_warn_ratelimited("forward tid %llu seq overflow\n",
3498 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3499 tid, next_mds, req->r_num_fwd, fwd_seq);
3502 /* resend. forward race not possible; mds would drop */
3503 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3505 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3506 req->r_attempts = 0;
3507 req->r_num_fwd = fwd_seq;
3508 req->r_resend_mds = next_mds;
3509 put_request_session(req);
3510 __do_request(mdsc, req);
3512 mutex_unlock(&mdsc->mutex);
3514 /* kick calling process */
3516 complete_request(mdsc, req);
3517 ceph_mdsc_put_request(req);
3521 pr_err("mdsc_handle_forward decode error err=%d\n", err);
3524 static int __decode_session_metadata(void **p, void *end,
3527 /* map<string,string> */
3530 ceph_decode_32_safe(p, end, n, bad);
3533 ceph_decode_32_safe(p, end, len, bad);
3534 ceph_decode_need(p, end, len, bad);
3535 err_str = !strncmp(*p, "error_string", len);
3537 ceph_decode_32_safe(p, end, len, bad);
3538 ceph_decode_need(p, end, len, bad);
3540 * Match "blocklisted (blacklisted)" from newer MDSes,
3541 * or "blacklisted" from older MDSes.
3543 if (err_str && strnstr(*p, "blacklisted", len))
3544 *blocklisted = true;
3553 * handle a mds session control message
3555 static void handle_session(struct ceph_mds_session *session,
3556 struct ceph_msg *msg)
3558 struct ceph_mds_client *mdsc = session->s_mdsc;
3559 int mds = session->s_mds;
3560 int msg_version = le16_to_cpu(msg->hdr.version);
3561 void *p = msg->front.iov_base;
3562 void *end = p + msg->front.iov_len;
3563 struct ceph_mds_session_head *h;
3565 u64 seq, features = 0;
3567 bool blocklisted = false;
3570 ceph_decode_need(&p, end, sizeof(*h), bad);
3574 op = le32_to_cpu(h->op);
3575 seq = le64_to_cpu(h->seq);
3577 if (msg_version >= 3) {
3579 /* version >= 2 and < 5, decode metadata, skip otherwise
3580 * as it's handled via flags.
3582 if (msg_version >= 5)
3583 ceph_decode_skip_map(&p, end, string, string, bad);
3584 else if (__decode_session_metadata(&p, end, &blocklisted) < 0)
3587 /* version >= 3, feature bits */
3588 ceph_decode_32_safe(&p, end, len, bad);
3590 ceph_decode_64_safe(&p, end, features, bad);
3591 p += len - sizeof(features);
3595 if (msg_version >= 5) {
3599 ceph_decode_skip_16(&p, end, bad); /* struct_v, struct_cv */
3600 ceph_decode_32_safe(&p, end, len, bad); /* len */
3601 ceph_decode_skip_n(&p, end, len, bad); /* metric_spec */
3603 /* version >= 5, flags */
3604 ceph_decode_32_safe(&p, end, flags, bad);
3605 if (flags & CEPH_SESSION_BLOCKLISTED) {
3606 pr_warn("mds%d session blocklisted\n", session->s_mds);
3611 mutex_lock(&mdsc->mutex);
3612 if (op == CEPH_SESSION_CLOSE) {
3613 ceph_get_mds_session(session);
3614 __unregister_session(mdsc, session);
3616 /* FIXME: this ttl calculation is generous */
3617 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3618 mutex_unlock(&mdsc->mutex);
3620 mutex_lock(&session->s_mutex);
3622 dout("handle_session mds%d %s %p state %s seq %llu\n",
3623 mds, ceph_session_op_name(op), session,
3624 ceph_session_state_name(session->s_state), seq);
3626 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3627 session->s_state = CEPH_MDS_SESSION_OPEN;
3628 pr_info("mds%d came back\n", session->s_mds);
3632 case CEPH_SESSION_OPEN:
3633 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3634 pr_info("mds%d reconnect success\n", session->s_mds);
3636 if (session->s_state == CEPH_MDS_SESSION_OPEN) {
3637 pr_notice("mds%d is already opened\n", session->s_mds);
3639 session->s_state = CEPH_MDS_SESSION_OPEN;
3640 session->s_features = features;
3641 renewed_caps(mdsc, session, 0);
3642 if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT,
3643 &session->s_features))
3644 metric_schedule_delayed(&mdsc->metric);
3648 * The connection maybe broken and the session in client
3649 * side has been reinitialized, need to update the seq
3652 if (!session->s_seq && seq)
3653 session->s_seq = seq;
3657 __close_session(mdsc, session);
3660 case CEPH_SESSION_RENEWCAPS:
3661 if (session->s_renew_seq == seq)
3662 renewed_caps(mdsc, session, 1);
3665 case CEPH_SESSION_CLOSE:
3666 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3667 pr_info("mds%d reconnect denied\n", session->s_mds);
3668 session->s_state = CEPH_MDS_SESSION_CLOSED;
3669 cleanup_session_requests(mdsc, session);
3670 remove_session_caps(session);
3671 wake = 2; /* for good measure */
3672 wake_up_all(&mdsc->session_close_wq);
3675 case CEPH_SESSION_STALE:
3676 pr_info("mds%d caps went stale, renewing\n",
3678 atomic_inc(&session->s_cap_gen);
3679 session->s_cap_ttl = jiffies - 1;
3680 send_renew_caps(mdsc, session);
3683 case CEPH_SESSION_RECALL_STATE:
3684 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3687 case CEPH_SESSION_FLUSHMSG:
3688 /* flush cap releases */
3689 spin_lock(&session->s_cap_lock);
3690 if (session->s_num_cap_releases)
3691 ceph_flush_cap_releases(mdsc, session);
3692 spin_unlock(&session->s_cap_lock);
3694 send_flushmsg_ack(mdsc, session, seq);
3697 case CEPH_SESSION_FORCE_RO:
3698 dout("force_session_readonly %p\n", session);
3699 spin_lock(&session->s_cap_lock);
3700 session->s_readonly = true;
3701 spin_unlock(&session->s_cap_lock);
3702 wake_up_session_caps(session, FORCE_RO);
3705 case CEPH_SESSION_REJECT:
3706 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3707 pr_info("mds%d rejected session\n", session->s_mds);
3708 session->s_state = CEPH_MDS_SESSION_REJECTED;
3709 cleanup_session_requests(mdsc, session);
3710 remove_session_caps(session);
3712 mdsc->fsc->blocklisted = true;
3713 wake = 2; /* for good measure */
3717 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3721 mutex_unlock(&session->s_mutex);
3723 mutex_lock(&mdsc->mutex);
3724 __wake_requests(mdsc, &session->s_waiting);
3726 kick_requests(mdsc, mds);
3727 mutex_unlock(&mdsc->mutex);
3729 if (op == CEPH_SESSION_CLOSE)
3730 ceph_put_mds_session(session);
3734 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3735 (int)msg->front.iov_len);
3740 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3744 dcaps = xchg(&req->r_dir_caps, 0);
3746 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3747 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3751 void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3755 dcaps = xchg(&req->r_dir_caps, 0);
3757 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3758 ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3764 * called under session->mutex.
3766 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3767 struct ceph_mds_session *session)
3769 struct ceph_mds_request *req, *nreq;
3772 dout("replay_unsafe_requests mds%d\n", session->s_mds);
3774 mutex_lock(&mdsc->mutex);
3775 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3776 __send_request(session, req, true);
3779 * also re-send old requests when MDS enters reconnect stage. So that MDS
3780 * can process completed request in clientreplay stage.
3782 p = rb_first(&mdsc->request_tree);
3784 req = rb_entry(p, struct ceph_mds_request, r_node);
3786 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3788 if (req->r_attempts == 0)
3789 continue; /* only old requests */
3790 if (!req->r_session)
3792 if (req->r_session->s_mds != session->s_mds)
3795 ceph_mdsc_release_dir_caps_no_check(req);
3797 __send_request(session, req, true);
3799 mutex_unlock(&mdsc->mutex);
3802 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3804 struct ceph_msg *reply;
3805 struct ceph_pagelist *_pagelist;
3810 if (!recon_state->allow_multi)
3813 /* can't handle message that contains both caps and realm */
3814 BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3816 /* pre-allocate new pagelist */
3817 _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3821 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3825 /* placeholder for nr_caps */
3826 err = ceph_pagelist_encode_32(_pagelist, 0);
3830 if (recon_state->nr_caps) {
3831 /* currently encoding caps */
3832 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3836 /* placeholder for nr_realms (currently encoding relams) */
3837 err = ceph_pagelist_encode_32(_pagelist, 0);
3842 err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3846 page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3847 addr = kmap_atomic(page);
3848 if (recon_state->nr_caps) {
3849 /* currently encoding caps */
3850 *addr = cpu_to_le32(recon_state->nr_caps);
3852 /* currently encoding relams */
3853 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3855 kunmap_atomic(addr);
3857 reply->hdr.version = cpu_to_le16(5);
3858 reply->hdr.compat_version = cpu_to_le16(4);
3860 reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3861 ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3863 ceph_con_send(&recon_state->session->s_con, reply);
3864 ceph_pagelist_release(recon_state->pagelist);
3866 recon_state->pagelist = _pagelist;
3867 recon_state->nr_caps = 0;
3868 recon_state->nr_realms = 0;
3869 recon_state->msg_version = 5;
3872 ceph_msg_put(reply);
3874 ceph_pagelist_release(_pagelist);
3878 static struct dentry* d_find_primary(struct inode *inode)
3880 struct dentry *alias, *dn = NULL;
3882 if (hlist_empty(&inode->i_dentry))
3885 spin_lock(&inode->i_lock);
3886 if (hlist_empty(&inode->i_dentry))
3889 if (S_ISDIR(inode->i_mode)) {
3890 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
3891 if (!IS_ROOT(alias))
3896 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
3897 spin_lock(&alias->d_lock);
3898 if (!d_unhashed(alias) &&
3899 (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
3900 dn = dget_dlock(alias);
3902 spin_unlock(&alias->d_lock);
3907 spin_unlock(&inode->i_lock);
3912 * Encode information about a cap for a reconnect with the MDS.
3914 static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3918 struct ceph_mds_cap_reconnect v2;
3919 struct ceph_mds_cap_reconnect_v1 v1;
3921 struct ceph_inode_info *ci = cap->ci;
3922 struct ceph_reconnect_state *recon_state = arg;
3923 struct ceph_pagelist *pagelist = recon_state->pagelist;
3924 struct dentry *dentry;
3926 int pathlen = 0, err;
3930 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3931 inode, ceph_vinop(inode), cap, cap->cap_id,
3932 ceph_cap_string(cap->issued));
3934 dentry = d_find_primary(inode);
3936 /* set pathbase to parent dir when msg_version >= 2 */
3937 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase,
3938 recon_state->msg_version >= 2);
3941 err = PTR_ERR(path);
3949 spin_lock(&ci->i_ceph_lock);
3950 cap->seq = 0; /* reset cap seq */
3951 cap->issue_seq = 0; /* and issue_seq */
3952 cap->mseq = 0; /* and migrate_seq */
3953 cap->cap_gen = atomic_read(&cap->session->s_cap_gen);
3955 /* These are lost when the session goes away */
3956 if (S_ISDIR(inode->i_mode)) {
3957 if (cap->issued & CEPH_CAP_DIR_CREATE) {
3958 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3959 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3961 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3964 if (recon_state->msg_version >= 2) {
3965 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3966 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3967 rec.v2.issued = cpu_to_le32(cap->issued);
3968 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3969 rec.v2.pathbase = cpu_to_le64(pathbase);
3970 rec.v2.flock_len = (__force __le32)
3971 ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3973 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3974 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3975 rec.v1.issued = cpu_to_le32(cap->issued);
3976 rec.v1.size = cpu_to_le64(i_size_read(inode));
3977 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3978 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3979 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3980 rec.v1.pathbase = cpu_to_le64(pathbase);
3983 if (list_empty(&ci->i_cap_snaps)) {
3984 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3986 struct ceph_cap_snap *capsnap =
3987 list_first_entry(&ci->i_cap_snaps,
3988 struct ceph_cap_snap, ci_item);
3989 snap_follows = capsnap->follows;
3991 spin_unlock(&ci->i_ceph_lock);
3993 if (recon_state->msg_version >= 2) {
3994 int num_fcntl_locks, num_flock_locks;
3995 struct ceph_filelock *flocks = NULL;
3996 size_t struct_len, total_len = sizeof(u64);
4000 if (rec.v2.flock_len) {
4001 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
4003 num_fcntl_locks = 0;
4004 num_flock_locks = 0;
4006 if (num_fcntl_locks + num_flock_locks > 0) {
4007 flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
4008 sizeof(struct ceph_filelock),
4014 err = ceph_encode_locks_to_buffer(inode, flocks,
4029 if (recon_state->msg_version >= 3) {
4030 /* version, compat_version and struct_len */
4031 total_len += 2 * sizeof(u8) + sizeof(u32);
4035 * number of encoded locks is stable, so copy to pagelist
4037 struct_len = 2 * sizeof(u32) +
4038 (num_fcntl_locks + num_flock_locks) *
4039 sizeof(struct ceph_filelock);
4040 rec.v2.flock_len = cpu_to_le32(struct_len);
4042 struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
4045 struct_len += sizeof(u64); /* snap_follows */
4047 total_len += struct_len;
4049 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
4050 err = send_reconnect_partial(recon_state);
4052 goto out_freeflocks;
4053 pagelist = recon_state->pagelist;
4056 err = ceph_pagelist_reserve(pagelist, total_len);
4058 goto out_freeflocks;
4060 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
4061 if (recon_state->msg_version >= 3) {
4062 ceph_pagelist_encode_8(pagelist, struct_v);
4063 ceph_pagelist_encode_8(pagelist, 1);
4064 ceph_pagelist_encode_32(pagelist, struct_len);
4066 ceph_pagelist_encode_string(pagelist, path, pathlen);
4067 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
4068 ceph_locks_to_pagelist(flocks, pagelist,
4069 num_fcntl_locks, num_flock_locks);
4071 ceph_pagelist_encode_64(pagelist, snap_follows);
4075 err = ceph_pagelist_reserve(pagelist,
4076 sizeof(u64) + sizeof(u32) +
4077 pathlen + sizeof(rec.v1));
4081 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
4082 ceph_pagelist_encode_string(pagelist, path, pathlen);
4083 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
4087 ceph_mdsc_free_path(path, pathlen);
4089 recon_state->nr_caps++;
4093 static int encode_snap_realms(struct ceph_mds_client *mdsc,
4094 struct ceph_reconnect_state *recon_state)
4097 struct ceph_pagelist *pagelist = recon_state->pagelist;
4100 if (recon_state->msg_version >= 4) {
4101 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
4107 * snaprealms. we provide mds with the ino, seq (version), and
4108 * parent for all of our realms. If the mds has any newer info,
4111 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
4112 struct ceph_snap_realm *realm =
4113 rb_entry(p, struct ceph_snap_realm, node);
4114 struct ceph_mds_snaprealm_reconnect sr_rec;
4116 if (recon_state->msg_version >= 4) {
4117 size_t need = sizeof(u8) * 2 + sizeof(u32) +
4120 if (pagelist->length + need > RECONNECT_MAX_SIZE) {
4121 err = send_reconnect_partial(recon_state);
4124 pagelist = recon_state->pagelist;
4127 err = ceph_pagelist_reserve(pagelist, need);
4131 ceph_pagelist_encode_8(pagelist, 1);
4132 ceph_pagelist_encode_8(pagelist, 1);
4133 ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
4136 dout(" adding snap realm %llx seq %lld parent %llx\n",
4137 realm->ino, realm->seq, realm->parent_ino);
4138 sr_rec.ino = cpu_to_le64(realm->ino);
4139 sr_rec.seq = cpu_to_le64(realm->seq);
4140 sr_rec.parent = cpu_to_le64(realm->parent_ino);
4142 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
4146 recon_state->nr_realms++;
4154 * If an MDS fails and recovers, clients need to reconnect in order to
4155 * reestablish shared state. This includes all caps issued through
4156 * this session _and_ the snap_realm hierarchy. Because it's not
4157 * clear which snap realms the mds cares about, we send everything we
4158 * know about.. that ensures we'll then get any new info the
4159 * recovering MDS might have.
4161 * This is a relatively heavyweight operation, but it's rare.
4163 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
4164 struct ceph_mds_session *session)
4166 struct ceph_msg *reply;
4167 int mds = session->s_mds;
4169 struct ceph_reconnect_state recon_state = {
4174 pr_info("mds%d reconnect start\n", mds);
4176 recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
4177 if (!recon_state.pagelist)
4178 goto fail_nopagelist;
4180 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4184 xa_destroy(&session->s_delegated_inos);
4186 mutex_lock(&session->s_mutex);
4187 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
4190 dout("session %p state %s\n", session,
4191 ceph_session_state_name(session->s_state));
4193 atomic_inc(&session->s_cap_gen);
4195 spin_lock(&session->s_cap_lock);
4196 /* don't know if session is readonly */
4197 session->s_readonly = 0;
4199 * notify __ceph_remove_cap() that we are composing cap reconnect.
4200 * If a cap get released before being added to the cap reconnect,
4201 * __ceph_remove_cap() should skip queuing cap release.
4203 session->s_cap_reconnect = 1;
4204 /* drop old cap expires; we're about to reestablish that state */
4205 detach_cap_releases(session, &dispose);
4206 spin_unlock(&session->s_cap_lock);
4207 dispose_cap_releases(mdsc, &dispose);
4209 /* trim unused caps to reduce MDS's cache rejoin time */
4210 if (mdsc->fsc->sb->s_root)
4211 shrink_dcache_parent(mdsc->fsc->sb->s_root);
4213 ceph_con_close(&session->s_con);
4214 ceph_con_open(&session->s_con,
4215 CEPH_ENTITY_TYPE_MDS, mds,
4216 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
4218 /* replay unsafe requests */
4219 replay_unsafe_requests(mdsc, session);
4221 ceph_early_kick_flushing_caps(mdsc, session);
4223 down_read(&mdsc->snap_rwsem);
4225 /* placeholder for nr_caps */
4226 err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
4230 if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4231 recon_state.msg_version = 3;
4232 recon_state.allow_multi = true;
4233 } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4234 recon_state.msg_version = 3;
4236 recon_state.msg_version = 2;
4238 /* trsaverse this session's caps */
4239 err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4241 spin_lock(&session->s_cap_lock);
4242 session->s_cap_reconnect = 0;
4243 spin_unlock(&session->s_cap_lock);
4248 /* check if all realms can be encoded into current message */
4249 if (mdsc->num_snap_realms) {
4251 recon_state.pagelist->length +
4252 mdsc->num_snap_realms *
4253 sizeof(struct ceph_mds_snaprealm_reconnect);
4254 if (recon_state.msg_version >= 4) {
4255 /* number of realms */
4256 total_len += sizeof(u32);
4257 /* version, compat_version and struct_len */
4258 total_len += mdsc->num_snap_realms *
4259 (2 * sizeof(u8) + sizeof(u32));
4261 if (total_len > RECONNECT_MAX_SIZE) {
4262 if (!recon_state.allow_multi) {
4266 if (recon_state.nr_caps) {
4267 err = send_reconnect_partial(&recon_state);
4271 recon_state.msg_version = 5;
4275 err = encode_snap_realms(mdsc, &recon_state);
4279 if (recon_state.msg_version >= 5) {
4280 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4285 if (recon_state.nr_caps || recon_state.nr_realms) {
4287 list_first_entry(&recon_state.pagelist->head,
4289 __le32 *addr = kmap_atomic(page);
4290 if (recon_state.nr_caps) {
4291 WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4292 *addr = cpu_to_le32(recon_state.nr_caps);
4293 } else if (recon_state.msg_version >= 4) {
4294 *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4296 kunmap_atomic(addr);
4299 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4300 if (recon_state.msg_version >= 4)
4301 reply->hdr.compat_version = cpu_to_le16(4);
4303 reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4304 ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4306 ceph_con_send(&session->s_con, reply);
4308 mutex_unlock(&session->s_mutex);
4310 mutex_lock(&mdsc->mutex);
4311 __wake_requests(mdsc, &session->s_waiting);
4312 mutex_unlock(&mdsc->mutex);
4314 up_read(&mdsc->snap_rwsem);
4315 ceph_pagelist_release(recon_state.pagelist);
4319 ceph_msg_put(reply);
4320 up_read(&mdsc->snap_rwsem);
4321 mutex_unlock(&session->s_mutex);
4323 ceph_pagelist_release(recon_state.pagelist);
4325 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
4331 * compare old and new mdsmaps, kicking requests
4332 * and closing out old connections as necessary
4334 * called under mdsc->mutex.
4336 static void check_new_map(struct ceph_mds_client *mdsc,
4337 struct ceph_mdsmap *newmap,
4338 struct ceph_mdsmap *oldmap)
4341 int oldstate, newstate;
4342 struct ceph_mds_session *s;
4343 unsigned long targets[DIV_ROUND_UP(CEPH_MAX_MDS, sizeof(unsigned long))] = {0};
4345 dout("check_new_map new %u old %u\n",
4346 newmap->m_epoch, oldmap->m_epoch);
4348 if (newmap->m_info) {
4349 for (i = 0; i < newmap->possible_max_rank; i++) {
4350 for (j = 0; j < newmap->m_info[i].num_export_targets; j++)
4351 set_bit(newmap->m_info[i].export_targets[j], targets);
4355 for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4356 if (!mdsc->sessions[i])
4358 s = mdsc->sessions[i];
4359 oldstate = ceph_mdsmap_get_state(oldmap, i);
4360 newstate = ceph_mdsmap_get_state(newmap, i);
4362 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
4363 i, ceph_mds_state_name(oldstate),
4364 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4365 ceph_mds_state_name(newstate),
4366 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4367 ceph_session_state_name(s->s_state));
4369 if (i >= newmap->possible_max_rank) {
4370 /* force close session for stopped mds */
4371 ceph_get_mds_session(s);
4372 __unregister_session(mdsc, s);
4373 __wake_requests(mdsc, &s->s_waiting);
4374 mutex_unlock(&mdsc->mutex);
4376 mutex_lock(&s->s_mutex);
4377 cleanup_session_requests(mdsc, s);
4378 remove_session_caps(s);
4379 mutex_unlock(&s->s_mutex);
4381 ceph_put_mds_session(s);
4383 mutex_lock(&mdsc->mutex);
4384 kick_requests(mdsc, i);
4388 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4389 ceph_mdsmap_get_addr(newmap, i),
4390 sizeof(struct ceph_entity_addr))) {
4392 mutex_unlock(&mdsc->mutex);
4393 mutex_lock(&s->s_mutex);
4394 mutex_lock(&mdsc->mutex);
4395 ceph_con_close(&s->s_con);
4396 mutex_unlock(&s->s_mutex);
4397 s->s_state = CEPH_MDS_SESSION_RESTARTING;
4398 } else if (oldstate == newstate) {
4399 continue; /* nothing new with this mds */
4405 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4406 newstate >= CEPH_MDS_STATE_RECONNECT) {
4407 mutex_unlock(&mdsc->mutex);
4408 clear_bit(i, targets);
4409 send_mds_reconnect(mdsc, s);
4410 mutex_lock(&mdsc->mutex);
4414 * kick request on any mds that has gone active.
4416 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4417 newstate >= CEPH_MDS_STATE_ACTIVE) {
4418 if (oldstate != CEPH_MDS_STATE_CREATING &&
4419 oldstate != CEPH_MDS_STATE_STARTING)
4420 pr_info("mds%d recovery completed\n", s->s_mds);
4421 kick_requests(mdsc, i);
4422 mutex_unlock(&mdsc->mutex);
4423 mutex_lock(&s->s_mutex);
4424 mutex_lock(&mdsc->mutex);
4425 ceph_kick_flushing_caps(mdsc, s);
4426 mutex_unlock(&s->s_mutex);
4427 wake_up_session_caps(s, RECONNECT);
4432 * Only open and reconnect sessions that don't exist yet.
4434 for (i = 0; i < newmap->possible_max_rank; i++) {
4436 * In case the import MDS is crashed just after
4437 * the EImportStart journal is flushed, so when
4438 * a standby MDS takes over it and is replaying
4439 * the EImportStart journal the new MDS daemon
4440 * will wait the client to reconnect it, but the
4441 * client may never register/open the session yet.
4443 * Will try to reconnect that MDS daemon if the
4444 * rank number is in the export targets array and
4445 * is the up:reconnect state.
4447 newstate = ceph_mdsmap_get_state(newmap, i);
4448 if (!test_bit(i, targets) || newstate != CEPH_MDS_STATE_RECONNECT)
4452 * The session maybe registered and opened by some
4453 * requests which were choosing random MDSes during
4454 * the mdsc->mutex's unlock/lock gap below in rare
4455 * case. But the related MDS daemon will just queue
4456 * that requests and be still waiting for the client's
4457 * reconnection request in up:reconnect state.
4459 s = __ceph_lookup_mds_session(mdsc, i);
4461 s = __open_export_target_session(mdsc, i);
4464 pr_err("failed to open export target session, err %d\n",
4469 dout("send reconnect to export target mds.%d\n", i);
4470 mutex_unlock(&mdsc->mutex);
4471 send_mds_reconnect(mdsc, s);
4472 ceph_put_mds_session(s);
4473 mutex_lock(&mdsc->mutex);
4476 for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4477 s = mdsc->sessions[i];
4480 if (!ceph_mdsmap_is_laggy(newmap, i))
4482 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4483 s->s_state == CEPH_MDS_SESSION_HUNG ||
4484 s->s_state == CEPH_MDS_SESSION_CLOSING) {
4485 dout(" connecting to export targets of laggy mds%d\n",
4487 __open_export_target_sessions(mdsc, s);
4499 * caller must hold session s_mutex, dentry->d_lock
4501 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4503 struct ceph_dentry_info *di = ceph_dentry(dentry);
4505 ceph_put_mds_session(di->lease_session);
4506 di->lease_session = NULL;
4509 static void handle_lease(struct ceph_mds_client *mdsc,
4510 struct ceph_mds_session *session,
4511 struct ceph_msg *msg)
4513 struct super_block *sb = mdsc->fsc->sb;
4514 struct inode *inode;
4515 struct dentry *parent, *dentry;
4516 struct ceph_dentry_info *di;
4517 int mds = session->s_mds;
4518 struct ceph_mds_lease *h = msg->front.iov_base;
4520 struct ceph_vino vino;
4524 dout("handle_lease from mds%d\n", mds);
4527 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4529 vino.ino = le64_to_cpu(h->ino);
4530 vino.snap = CEPH_NOSNAP;
4531 seq = le32_to_cpu(h->seq);
4532 dname.len = get_unaligned_le32(h + 1);
4533 if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4535 dname.name = (void *)(h + 1) + sizeof(u32);
4538 inode = ceph_find_inode(sb, vino);
4539 dout("handle_lease %s, ino %llx %p %.*s\n",
4540 ceph_lease_op_name(h->action), vino.ino, inode,
4541 dname.len, dname.name);
4543 mutex_lock(&session->s_mutex);
4544 inc_session_sequence(session);
4547 dout("handle_lease no inode %llx\n", vino.ino);
4552 parent = d_find_alias(inode);
4554 dout("no parent dentry on inode %p\n", inode);
4556 goto release; /* hrm... */
4558 dname.hash = full_name_hash(parent, dname.name, dname.len);
4559 dentry = d_lookup(parent, &dname);
4564 spin_lock(&dentry->d_lock);
4565 di = ceph_dentry(dentry);
4566 switch (h->action) {
4567 case CEPH_MDS_LEASE_REVOKE:
4568 if (di->lease_session == session) {
4569 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4570 h->seq = cpu_to_le32(di->lease_seq);
4571 __ceph_mdsc_drop_dentry_lease(dentry);
4576 case CEPH_MDS_LEASE_RENEW:
4577 if (di->lease_session == session &&
4578 di->lease_gen == atomic_read(&session->s_cap_gen) &&
4579 di->lease_renew_from &&
4580 di->lease_renew_after == 0) {
4581 unsigned long duration =
4582 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4584 di->lease_seq = seq;
4585 di->time = di->lease_renew_from + duration;
4586 di->lease_renew_after = di->lease_renew_from +
4588 di->lease_renew_from = 0;
4592 spin_unlock(&dentry->d_lock);
4599 /* let's just reuse the same message */
4600 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4602 ceph_con_send(&session->s_con, msg);
4605 mutex_unlock(&session->s_mutex);
4610 pr_err("corrupt lease message\n");
4614 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4615 struct dentry *dentry, char action,
4618 struct ceph_msg *msg;
4619 struct ceph_mds_lease *lease;
4621 int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4623 dout("lease_send_msg identry %p %s to mds%d\n",
4624 dentry, ceph_lease_op_name(action), session->s_mds);
4626 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4629 lease = msg->front.iov_base;
4630 lease->action = action;
4631 lease->seq = cpu_to_le32(seq);
4633 spin_lock(&dentry->d_lock);
4634 dir = d_inode(dentry->d_parent);
4635 lease->ino = cpu_to_le64(ceph_ino(dir));
4636 lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4638 put_unaligned_le32(dentry->d_name.len, lease + 1);
4639 memcpy((void *)(lease + 1) + 4,
4640 dentry->d_name.name, dentry->d_name.len);
4641 spin_unlock(&dentry->d_lock);
4643 ceph_con_send(&session->s_con, msg);
4647 * lock unlock the session, to wait ongoing session activities
4649 static void lock_unlock_session(struct ceph_mds_session *s)
4651 mutex_lock(&s->s_mutex);
4652 mutex_unlock(&s->s_mutex);
4655 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4657 struct ceph_fs_client *fsc = mdsc->fsc;
4659 if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4662 if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4665 if (!READ_ONCE(fsc->blocklisted))
4668 pr_info("auto reconnect after blocklisted\n");
4669 ceph_force_reconnect(fsc->sb);
4672 bool check_session_state(struct ceph_mds_session *s)
4674 switch (s->s_state) {
4675 case CEPH_MDS_SESSION_OPEN:
4676 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4677 s->s_state = CEPH_MDS_SESSION_HUNG;
4678 pr_info("mds%d hung\n", s->s_mds);
4681 case CEPH_MDS_SESSION_CLOSING:
4682 case CEPH_MDS_SESSION_NEW:
4683 case CEPH_MDS_SESSION_RESTARTING:
4684 case CEPH_MDS_SESSION_CLOSED:
4685 case CEPH_MDS_SESSION_REJECTED:
4693 * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
4694 * then we need to retransmit that request.
4696 void inc_session_sequence(struct ceph_mds_session *s)
4698 lockdep_assert_held(&s->s_mutex);
4702 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4705 dout("resending session close request for mds%d\n", s->s_mds);
4706 ret = request_close_session(s);
4708 pr_err("unable to close session to mds%d: %d\n",
4714 * delayed work -- periodically trim expired leases, renew caps with mds. If
4715 * the @delay parameter is set to 0 or if it's more than 5 secs, the default
4716 * workqueue delay value of 5 secs will be used.
4718 static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
4720 unsigned long max_delay = HZ * 5;
4722 /* 5 secs default delay */
4723 if (!delay || (delay > max_delay))
4725 schedule_delayed_work(&mdsc->delayed_work,
4726 round_jiffies_relative(delay));
4729 static void delayed_work(struct work_struct *work)
4731 struct ceph_mds_client *mdsc =
4732 container_of(work, struct ceph_mds_client, delayed_work.work);
4733 unsigned long delay;
4738 dout("mdsc delayed_work\n");
4743 mutex_lock(&mdsc->mutex);
4744 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4745 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4746 mdsc->last_renew_caps);
4748 mdsc->last_renew_caps = jiffies;
4750 for (i = 0; i < mdsc->max_sessions; i++) {
4751 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4755 if (!check_session_state(s)) {
4756 ceph_put_mds_session(s);
4759 mutex_unlock(&mdsc->mutex);
4761 mutex_lock(&s->s_mutex);
4763 send_renew_caps(mdsc, s);
4765 ceph_con_keepalive(&s->s_con);
4766 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4767 s->s_state == CEPH_MDS_SESSION_HUNG)
4768 ceph_send_cap_releases(mdsc, s);
4769 mutex_unlock(&s->s_mutex);
4770 ceph_put_mds_session(s);
4772 mutex_lock(&mdsc->mutex);
4774 mutex_unlock(&mdsc->mutex);
4776 delay = ceph_check_delayed_caps(mdsc);
4778 ceph_queue_cap_reclaim_work(mdsc);
4780 ceph_trim_snapid_map(mdsc);
4782 maybe_recover_session(mdsc);
4784 schedule_delayed(mdsc, delay);
4787 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4790 struct ceph_mds_client *mdsc;
4793 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4797 mutex_init(&mdsc->mutex);
4798 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4799 if (!mdsc->mdsmap) {
4804 init_completion(&mdsc->safe_umount_waiters);
4805 init_waitqueue_head(&mdsc->session_close_wq);
4806 INIT_LIST_HEAD(&mdsc->waiting_for_map);
4807 mdsc->quotarealms_inodes = RB_ROOT;
4808 mutex_init(&mdsc->quotarealms_inodes_mutex);
4809 init_rwsem(&mdsc->snap_rwsem);
4810 mdsc->snap_realms = RB_ROOT;
4811 INIT_LIST_HEAD(&mdsc->snap_empty);
4812 spin_lock_init(&mdsc->snap_empty_lock);
4813 mdsc->request_tree = RB_ROOT;
4814 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4815 mdsc->last_renew_caps = jiffies;
4816 INIT_LIST_HEAD(&mdsc->cap_delay_list);
4817 INIT_LIST_HEAD(&mdsc->cap_wait_list);
4818 spin_lock_init(&mdsc->cap_delay_lock);
4819 INIT_LIST_HEAD(&mdsc->snap_flush_list);
4820 spin_lock_init(&mdsc->snap_flush_lock);
4821 mdsc->last_cap_flush_tid = 1;
4822 INIT_LIST_HEAD(&mdsc->cap_flush_list);
4823 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4824 spin_lock_init(&mdsc->cap_dirty_lock);
4825 init_waitqueue_head(&mdsc->cap_flushing_wq);
4826 INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4827 err = ceph_metric_init(&mdsc->metric);
4831 spin_lock_init(&mdsc->dentry_list_lock);
4832 INIT_LIST_HEAD(&mdsc->dentry_leases);
4833 INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4835 ceph_caps_init(mdsc);
4836 ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4838 spin_lock_init(&mdsc->snapid_map_lock);
4839 mdsc->snapid_map_tree = RB_ROOT;
4840 INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4842 init_rwsem(&mdsc->pool_perm_rwsem);
4843 mdsc->pool_perm_tree = RB_ROOT;
4845 strscpy(mdsc->nodename, utsname()->nodename,
4846 sizeof(mdsc->nodename));
4852 kfree(mdsc->mdsmap);
4859 * Wait for safe replies on open mds requests. If we time out, drop
4860 * all requests from the tree to avoid dangling dentry refs.
4862 static void wait_requests(struct ceph_mds_client *mdsc)
4864 struct ceph_options *opts = mdsc->fsc->client->options;
4865 struct ceph_mds_request *req;
4867 mutex_lock(&mdsc->mutex);
4868 if (__get_oldest_req(mdsc)) {
4869 mutex_unlock(&mdsc->mutex);
4871 dout("wait_requests waiting for requests\n");
4872 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4873 ceph_timeout_jiffies(opts->mount_timeout));
4875 /* tear down remaining requests */
4876 mutex_lock(&mdsc->mutex);
4877 while ((req = __get_oldest_req(mdsc))) {
4878 dout("wait_requests timed out on tid %llu\n",
4880 list_del_init(&req->r_wait);
4881 __unregister_request(mdsc, req);
4884 mutex_unlock(&mdsc->mutex);
4885 dout("wait_requests done\n");
4888 void send_flush_mdlog(struct ceph_mds_session *s)
4890 struct ceph_msg *msg;
4893 * Pre-luminous MDS crashes when it sees an unknown session request
4895 if (!CEPH_HAVE_FEATURE(s->s_con.peer_features, SERVER_LUMINOUS))
4898 mutex_lock(&s->s_mutex);
4899 dout("request mdlog flush to mds%d (%s)s seq %lld\n", s->s_mds,
4900 ceph_session_state_name(s->s_state), s->s_seq);
4901 msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_FLUSH_MDLOG,
4904 pr_err("failed to request mdlog flush to mds%d (%s) seq %lld\n",
4905 s->s_mds, ceph_session_state_name(s->s_state), s->s_seq);
4907 ceph_con_send(&s->s_con, msg);
4909 mutex_unlock(&s->s_mutex);
4913 * called before mount is ro, and before dentries are torn down.
4914 * (hmm, does this still race with new lookups?)
4916 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4918 dout("pre_umount\n");
4921 ceph_mdsc_iterate_sessions(mdsc, send_flush_mdlog, true);
4922 ceph_mdsc_iterate_sessions(mdsc, lock_unlock_session, false);
4923 ceph_flush_dirty_caps(mdsc);
4924 wait_requests(mdsc);
4927 * wait for reply handlers to drop their request refs and
4928 * their inode/dcache refs
4932 ceph_cleanup_quotarealms_inodes(mdsc);
4936 * flush the mdlog and wait for all write mds requests to flush.
4938 static void flush_mdlog_and_wait_mdsc_unsafe_requests(struct ceph_mds_client *mdsc,
4941 struct ceph_mds_request *req = NULL, *nextreq;
4942 struct ceph_mds_session *last_session = NULL;
4945 mutex_lock(&mdsc->mutex);
4946 dout("%s want %lld\n", __func__, want_tid);
4948 req = __get_oldest_req(mdsc);
4949 while (req && req->r_tid <= want_tid) {
4950 /* find next request */
4951 n = rb_next(&req->r_node);
4953 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4956 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4957 (req->r_op & CEPH_MDS_OP_WRITE)) {
4958 struct ceph_mds_session *s = req->r_session;
4966 ceph_mdsc_get_request(req);
4968 ceph_mdsc_get_request(nextreq);
4969 s = ceph_get_mds_session(s);
4970 mutex_unlock(&mdsc->mutex);
4972 /* send flush mdlog request to MDS */
4973 if (last_session != s) {
4974 send_flush_mdlog(s);
4975 ceph_put_mds_session(last_session);
4978 ceph_put_mds_session(s);
4980 dout("%s wait on %llu (want %llu)\n", __func__,
4981 req->r_tid, want_tid);
4982 wait_for_completion(&req->r_safe_completion);
4984 mutex_lock(&mdsc->mutex);
4985 ceph_mdsc_put_request(req);
4987 break; /* next dne before, so we're done! */
4988 if (RB_EMPTY_NODE(&nextreq->r_node)) {
4989 /* next request was removed from tree */
4990 ceph_mdsc_put_request(nextreq);
4993 ceph_mdsc_put_request(nextreq); /* won't go away */
4997 mutex_unlock(&mdsc->mutex);
4998 ceph_put_mds_session(last_session);
4999 dout("%s done\n", __func__);
5002 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
5004 u64 want_tid, want_flush;
5006 if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
5010 mutex_lock(&mdsc->mutex);
5011 want_tid = mdsc->last_tid;
5012 mutex_unlock(&mdsc->mutex);
5014 ceph_flush_dirty_caps(mdsc);
5015 spin_lock(&mdsc->cap_dirty_lock);
5016 want_flush = mdsc->last_cap_flush_tid;
5017 if (!list_empty(&mdsc->cap_flush_list)) {
5018 struct ceph_cap_flush *cf =
5019 list_last_entry(&mdsc->cap_flush_list,
5020 struct ceph_cap_flush, g_list);
5023 spin_unlock(&mdsc->cap_dirty_lock);
5025 dout("sync want tid %lld flush_seq %lld\n",
5026 want_tid, want_flush);
5028 flush_mdlog_and_wait_mdsc_unsafe_requests(mdsc, want_tid);
5029 wait_caps_flush(mdsc, want_flush);
5033 * true if all sessions are closed, or we force unmount
5035 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
5037 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
5039 return atomic_read(&mdsc->num_sessions) <= skipped;
5043 * called after sb is ro or when metadata corrupted.
5045 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
5047 struct ceph_options *opts = mdsc->fsc->client->options;
5048 struct ceph_mds_session *session;
5052 dout("close_sessions\n");
5054 /* close sessions */
5055 mutex_lock(&mdsc->mutex);
5056 for (i = 0; i < mdsc->max_sessions; i++) {
5057 session = __ceph_lookup_mds_session(mdsc, i);
5060 mutex_unlock(&mdsc->mutex);
5061 mutex_lock(&session->s_mutex);
5062 if (__close_session(mdsc, session) <= 0)
5064 mutex_unlock(&session->s_mutex);
5065 ceph_put_mds_session(session);
5066 mutex_lock(&mdsc->mutex);
5068 mutex_unlock(&mdsc->mutex);
5070 dout("waiting for sessions to close\n");
5071 wait_event_timeout(mdsc->session_close_wq,
5072 done_closing_sessions(mdsc, skipped),
5073 ceph_timeout_jiffies(opts->mount_timeout));
5075 /* tear down remaining sessions */
5076 mutex_lock(&mdsc->mutex);
5077 for (i = 0; i < mdsc->max_sessions; i++) {
5078 if (mdsc->sessions[i]) {
5079 session = ceph_get_mds_session(mdsc->sessions[i]);
5080 __unregister_session(mdsc, session);
5081 mutex_unlock(&mdsc->mutex);
5082 mutex_lock(&session->s_mutex);
5083 remove_session_caps(session);
5084 mutex_unlock(&session->s_mutex);
5085 ceph_put_mds_session(session);
5086 mutex_lock(&mdsc->mutex);
5089 WARN_ON(!list_empty(&mdsc->cap_delay_list));
5090 mutex_unlock(&mdsc->mutex);
5092 ceph_cleanup_snapid_map(mdsc);
5093 ceph_cleanup_global_and_empty_realms(mdsc);
5095 cancel_work_sync(&mdsc->cap_reclaim_work);
5096 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
5101 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
5103 struct ceph_mds_session *session;
5106 dout("force umount\n");
5108 mutex_lock(&mdsc->mutex);
5109 for (mds = 0; mds < mdsc->max_sessions; mds++) {
5110 session = __ceph_lookup_mds_session(mdsc, mds);
5114 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
5115 __unregister_session(mdsc, session);
5116 __wake_requests(mdsc, &session->s_waiting);
5117 mutex_unlock(&mdsc->mutex);
5119 mutex_lock(&session->s_mutex);
5120 __close_session(mdsc, session);
5121 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
5122 cleanup_session_requests(mdsc, session);
5123 remove_session_caps(session);
5125 mutex_unlock(&session->s_mutex);
5126 ceph_put_mds_session(session);
5128 mutex_lock(&mdsc->mutex);
5129 kick_requests(mdsc, mds);
5131 __wake_requests(mdsc, &mdsc->waiting_for_map);
5132 mutex_unlock(&mdsc->mutex);
5135 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
5139 * Make sure the delayed work stopped before releasing
5142 * Because the cancel_delayed_work_sync() will only
5143 * guarantee that the work finishes executing. But the
5144 * delayed work will re-arm itself again after that.
5146 flush_delayed_work(&mdsc->delayed_work);
5149 ceph_mdsmap_destroy(mdsc->mdsmap);
5150 kfree(mdsc->sessions);
5151 ceph_caps_finalize(mdsc);
5152 ceph_pool_perm_destroy(mdsc);
5155 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
5157 struct ceph_mds_client *mdsc = fsc->mdsc;
5158 dout("mdsc_destroy %p\n", mdsc);
5163 /* flush out any connection work with references to us */
5166 ceph_mdsc_stop(mdsc);
5168 ceph_metric_destroy(&mdsc->metric);
5172 dout("mdsc_destroy %p done\n", mdsc);
5175 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5177 struct ceph_fs_client *fsc = mdsc->fsc;
5178 const char *mds_namespace = fsc->mount_options->mds_namespace;
5179 void *p = msg->front.iov_base;
5180 void *end = p + msg->front.iov_len;
5183 u32 mount_fscid = (u32)-1;
5186 ceph_decode_need(&p, end, sizeof(u32), bad);
5187 epoch = ceph_decode_32(&p);
5189 dout("handle_fsmap epoch %u\n", epoch);
5191 /* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
5192 ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
5194 ceph_decode_32_safe(&p, end, num_fs, bad);
5195 while (num_fs-- > 0) {
5196 void *info_p, *info_end;
5200 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
5201 p += 2; // info_v, info_cv
5202 info_len = ceph_decode_32(&p);
5203 ceph_decode_need(&p, end, info_len, bad);
5205 info_end = p + info_len;
5208 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
5209 fscid = ceph_decode_32(&info_p);
5210 namelen = ceph_decode_32(&info_p);
5211 ceph_decode_need(&info_p, info_end, namelen, bad);
5213 if (mds_namespace &&
5214 strlen(mds_namespace) == namelen &&
5215 !strncmp(mds_namespace, (char *)info_p, namelen)) {
5216 mount_fscid = fscid;
5221 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
5222 if (mount_fscid != (u32)-1) {
5223 fsc->client->monc.fs_cluster_id = mount_fscid;
5224 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
5226 ceph_monc_renew_subs(&fsc->client->monc);
5234 pr_err("error decoding fsmap %d. Shutting down mount.\n", err);
5235 ceph_umount_begin(mdsc->fsc->sb);
5237 mutex_lock(&mdsc->mutex);
5238 mdsc->mdsmap_err = err;
5239 __wake_requests(mdsc, &mdsc->waiting_for_map);
5240 mutex_unlock(&mdsc->mutex);
5244 * handle mds map update.
5246 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5250 void *p = msg->front.iov_base;
5251 void *end = p + msg->front.iov_len;
5252 struct ceph_mdsmap *newmap, *oldmap;
5253 struct ceph_fsid fsid;
5256 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
5257 ceph_decode_copy(&p, &fsid, sizeof(fsid));
5258 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
5260 epoch = ceph_decode_32(&p);
5261 maplen = ceph_decode_32(&p);
5262 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
5264 /* do we need it? */
5265 mutex_lock(&mdsc->mutex);
5266 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
5267 dout("handle_map epoch %u <= our %u\n",
5268 epoch, mdsc->mdsmap->m_epoch);
5269 mutex_unlock(&mdsc->mutex);
5273 newmap = ceph_mdsmap_decode(&p, end, ceph_msgr2(mdsc->fsc->client));
5274 if (IS_ERR(newmap)) {
5275 err = PTR_ERR(newmap);
5279 /* swap into place */
5281 oldmap = mdsc->mdsmap;
5282 mdsc->mdsmap = newmap;
5283 check_new_map(mdsc, newmap, oldmap);
5284 ceph_mdsmap_destroy(oldmap);
5286 mdsc->mdsmap = newmap; /* first mds map */
5288 mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
5291 __wake_requests(mdsc, &mdsc->waiting_for_map);
5292 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
5293 mdsc->mdsmap->m_epoch);
5295 mutex_unlock(&mdsc->mutex);
5296 schedule_delayed(mdsc, 0);
5300 mutex_unlock(&mdsc->mutex);
5302 pr_err("error decoding mdsmap %d. Shutting down mount.\n", err);
5303 ceph_umount_begin(mdsc->fsc->sb);
5307 static struct ceph_connection *mds_get_con(struct ceph_connection *con)
5309 struct ceph_mds_session *s = con->private;
5311 if (ceph_get_mds_session(s))
5316 static void mds_put_con(struct ceph_connection *con)
5318 struct ceph_mds_session *s = con->private;
5320 ceph_put_mds_session(s);
5324 * if the client is unresponsive for long enough, the mds will kill
5325 * the session entirely.
5327 static void mds_peer_reset(struct ceph_connection *con)
5329 struct ceph_mds_session *s = con->private;
5330 struct ceph_mds_client *mdsc = s->s_mdsc;
5332 pr_warn("mds%d closed our session\n", s->s_mds);
5333 if (READ_ONCE(mdsc->fsc->mount_state) != CEPH_MOUNT_FENCE_IO)
5334 send_mds_reconnect(mdsc, s);
5337 static void mds_dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5339 struct ceph_mds_session *s = con->private;
5340 struct ceph_mds_client *mdsc = s->s_mdsc;
5341 int type = le16_to_cpu(msg->hdr.type);
5343 mutex_lock(&mdsc->mutex);
5344 if (__verify_registered_session(mdsc, s) < 0) {
5345 mutex_unlock(&mdsc->mutex);
5348 mutex_unlock(&mdsc->mutex);
5351 case CEPH_MSG_MDS_MAP:
5352 ceph_mdsc_handle_mdsmap(mdsc, msg);
5354 case CEPH_MSG_FS_MAP_USER:
5355 ceph_mdsc_handle_fsmap(mdsc, msg);
5357 case CEPH_MSG_CLIENT_SESSION:
5358 handle_session(s, msg);
5360 case CEPH_MSG_CLIENT_REPLY:
5361 handle_reply(s, msg);
5363 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5364 handle_forward(mdsc, s, msg);
5366 case CEPH_MSG_CLIENT_CAPS:
5367 ceph_handle_caps(s, msg);
5369 case CEPH_MSG_CLIENT_SNAP:
5370 ceph_handle_snap(mdsc, s, msg);
5372 case CEPH_MSG_CLIENT_LEASE:
5373 handle_lease(mdsc, s, msg);
5375 case CEPH_MSG_CLIENT_QUOTA:
5376 ceph_handle_quota(mdsc, s, msg);
5380 pr_err("received unknown message type %d %s\n", type,
5381 ceph_msg_type_name(type));
5392 * Note: returned pointer is the address of a structure that's
5393 * managed separately. Caller must *not* attempt to free it.
5395 static struct ceph_auth_handshake *
5396 mds_get_authorizer(struct ceph_connection *con, int *proto, int force_new)
5398 struct ceph_mds_session *s = con->private;
5399 struct ceph_mds_client *mdsc = s->s_mdsc;
5400 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5401 struct ceph_auth_handshake *auth = &s->s_auth;
5404 ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5405 force_new, proto, NULL, NULL);
5407 return ERR_PTR(ret);
5412 static int mds_add_authorizer_challenge(struct ceph_connection *con,
5413 void *challenge_buf, int challenge_buf_len)
5415 struct ceph_mds_session *s = con->private;
5416 struct ceph_mds_client *mdsc = s->s_mdsc;
5417 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5419 return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
5420 challenge_buf, challenge_buf_len);
5423 static int mds_verify_authorizer_reply(struct ceph_connection *con)
5425 struct ceph_mds_session *s = con->private;
5426 struct ceph_mds_client *mdsc = s->s_mdsc;
5427 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5428 struct ceph_auth_handshake *auth = &s->s_auth;
5430 return ceph_auth_verify_authorizer_reply(ac, auth->authorizer,
5431 auth->authorizer_reply_buf, auth->authorizer_reply_buf_len,
5432 NULL, NULL, NULL, NULL);
5435 static int mds_invalidate_authorizer(struct ceph_connection *con)
5437 struct ceph_mds_session *s = con->private;
5438 struct ceph_mds_client *mdsc = s->s_mdsc;
5439 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5441 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5443 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5446 static int mds_get_auth_request(struct ceph_connection *con,
5447 void *buf, int *buf_len,
5448 void **authorizer, int *authorizer_len)
5450 struct ceph_mds_session *s = con->private;
5451 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5452 struct ceph_auth_handshake *auth = &s->s_auth;
5455 ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5460 *authorizer = auth->authorizer_buf;
5461 *authorizer_len = auth->authorizer_buf_len;
5465 static int mds_handle_auth_reply_more(struct ceph_connection *con,
5466 void *reply, int reply_len,
5467 void *buf, int *buf_len,
5468 void **authorizer, int *authorizer_len)
5470 struct ceph_mds_session *s = con->private;
5471 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5472 struct ceph_auth_handshake *auth = &s->s_auth;
5475 ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len,
5480 *authorizer = auth->authorizer_buf;
5481 *authorizer_len = auth->authorizer_buf_len;
5485 static int mds_handle_auth_done(struct ceph_connection *con,
5486 u64 global_id, void *reply, int reply_len,
5487 u8 *session_key, int *session_key_len,
5488 u8 *con_secret, int *con_secret_len)
5490 struct ceph_mds_session *s = con->private;
5491 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5492 struct ceph_auth_handshake *auth = &s->s_auth;
5494 return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len,
5495 session_key, session_key_len,
5496 con_secret, con_secret_len);
5499 static int mds_handle_auth_bad_method(struct ceph_connection *con,
5500 int used_proto, int result,
5501 const int *allowed_protos, int proto_cnt,
5502 const int *allowed_modes, int mode_cnt)
5504 struct ceph_mds_session *s = con->private;
5505 struct ceph_mon_client *monc = &s->s_mdsc->fsc->client->monc;
5508 if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_MDS,
5510 allowed_protos, proto_cnt,
5511 allowed_modes, mode_cnt)) {
5512 ret = ceph_monc_validate_auth(monc);
5520 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
5521 struct ceph_msg_header *hdr, int *skip)
5523 struct ceph_msg *msg;
5524 int type = (int) le16_to_cpu(hdr->type);
5525 int front_len = (int) le32_to_cpu(hdr->front_len);
5531 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
5533 pr_err("unable to allocate msg type %d len %d\n",
5541 static int mds_sign_message(struct ceph_msg *msg)
5543 struct ceph_mds_session *s = msg->con->private;
5544 struct ceph_auth_handshake *auth = &s->s_auth;
5546 return ceph_auth_sign_message(auth, msg);
5549 static int mds_check_message_signature(struct ceph_msg *msg)
5551 struct ceph_mds_session *s = msg->con->private;
5552 struct ceph_auth_handshake *auth = &s->s_auth;
5554 return ceph_auth_check_message_signature(auth, msg);
5557 static const struct ceph_connection_operations mds_con_ops = {
5560 .alloc_msg = mds_alloc_msg,
5561 .dispatch = mds_dispatch,
5562 .peer_reset = mds_peer_reset,
5563 .get_authorizer = mds_get_authorizer,
5564 .add_authorizer_challenge = mds_add_authorizer_challenge,
5565 .verify_authorizer_reply = mds_verify_authorizer_reply,
5566 .invalidate_authorizer = mds_invalidate_authorizer,
5567 .sign_message = mds_sign_message,
5568 .check_message_signature = mds_check_message_signature,
5569 .get_auth_request = mds_get_auth_request,
5570 .handle_auth_reply_more = mds_handle_auth_reply_more,
5571 .handle_auth_done = mds_handle_auth_done,
5572 .handle_auth_bad_method = mds_handle_auth_bad_method,