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50 * are no longer valid.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 (mds >= mdsc->mdsmap->possible_max_rank)
810 return ERR_PTR(-EINVAL);
812 s = kzalloc(sizeof(*s), GFP_NOFS);
814 return ERR_PTR(-ENOMEM);
816 if (mds >= mdsc->max_sessions) {
817 int newmax = 1 << get_count_order(mds + 1);
818 struct ceph_mds_session **sa;
820 dout("%s: realloc to %d\n", __func__, newmax);
821 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
824 if (mdsc->sessions) {
825 memcpy(sa, mdsc->sessions,
826 mdsc->max_sessions * sizeof(void *));
827 kfree(mdsc->sessions);
830 mdsc->max_sessions = newmax;
833 dout("%s: mds%d\n", __func__, mds);
836 s->s_state = CEPH_MDS_SESSION_NEW;
837 mutex_init(&s->s_mutex);
839 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
841 atomic_set(&s->s_cap_gen, 1);
842 s->s_cap_ttl = jiffies - 1;
844 spin_lock_init(&s->s_cap_lock);
845 INIT_LIST_HEAD(&s->s_caps);
846 refcount_set(&s->s_ref, 1);
847 INIT_LIST_HEAD(&s->s_waiting);
848 INIT_LIST_HEAD(&s->s_unsafe);
849 xa_init(&s->s_delegated_inos);
850 INIT_LIST_HEAD(&s->s_cap_releases);
851 INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
853 INIT_LIST_HEAD(&s->s_cap_dirty);
854 INIT_LIST_HEAD(&s->s_cap_flushing);
856 mdsc->sessions[mds] = s;
857 atomic_inc(&mdsc->num_sessions);
858 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
860 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
861 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
867 return ERR_PTR(-ENOMEM);
871 * called under mdsc->mutex
873 static void __unregister_session(struct ceph_mds_client *mdsc,
874 struct ceph_mds_session *s)
876 dout("__unregister_session mds%d %p\n", s->s_mds, s);
877 BUG_ON(mdsc->sessions[s->s_mds] != s);
878 mdsc->sessions[s->s_mds] = NULL;
879 ceph_con_close(&s->s_con);
880 ceph_put_mds_session(s);
881 atomic_dec(&mdsc->num_sessions);
885 * drop session refs in request.
887 * should be last request ref, or hold mdsc->mutex
889 static void put_request_session(struct ceph_mds_request *req)
891 if (req->r_session) {
892 ceph_put_mds_session(req->r_session);
893 req->r_session = NULL;
897 void ceph_mdsc_iterate_sessions(struct ceph_mds_client *mdsc,
898 void (*cb)(struct ceph_mds_session *),
903 mutex_lock(&mdsc->mutex);
904 for (mds = 0; mds < mdsc->max_sessions; ++mds) {
905 struct ceph_mds_session *s;
907 s = __ceph_lookup_mds_session(mdsc, mds);
911 if (check_state && !check_session_state(s)) {
912 ceph_put_mds_session(s);
916 mutex_unlock(&mdsc->mutex);
918 ceph_put_mds_session(s);
919 mutex_lock(&mdsc->mutex);
921 mutex_unlock(&mdsc->mutex);
924 void ceph_mdsc_release_request(struct kref *kref)
926 struct ceph_mds_request *req = container_of(kref,
927 struct ceph_mds_request,
929 ceph_mdsc_release_dir_caps_no_check(req);
930 destroy_reply_info(&req->r_reply_info);
932 ceph_msg_put(req->r_request);
934 ceph_msg_put(req->r_reply);
936 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
940 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
943 iput(req->r_target_inode);
946 if (req->r_old_dentry)
947 dput(req->r_old_dentry);
948 if (req->r_old_dentry_dir) {
950 * track (and drop pins for) r_old_dentry_dir
951 * separately, since r_old_dentry's d_parent may have
952 * changed between the dir mutex being dropped and
953 * this request being freed.
955 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
957 iput(req->r_old_dentry_dir);
961 put_cred(req->r_cred);
963 ceph_pagelist_release(req->r_pagelist);
964 put_request_session(req);
965 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
966 WARN_ON_ONCE(!list_empty(&req->r_wait));
967 kmem_cache_free(ceph_mds_request_cachep, req);
970 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
973 * lookup session, bump ref if found.
975 * called under mdsc->mutex.
977 static struct ceph_mds_request *
978 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
980 struct ceph_mds_request *req;
982 req = lookup_request(&mdsc->request_tree, tid);
984 ceph_mdsc_get_request(req);
990 * Register an in-flight request, and assign a tid. Link to directory
991 * are modifying (if any).
993 * Called under mdsc->mutex.
995 static void __register_request(struct ceph_mds_client *mdsc,
996 struct ceph_mds_request *req,
1001 req->r_tid = ++mdsc->last_tid;
1002 if (req->r_num_caps) {
1003 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
1006 pr_err("__register_request %p "
1007 "failed to reserve caps: %d\n", req, ret);
1008 /* set req->r_err to fail early from __do_request */
1013 dout("__register_request %p tid %lld\n", req, req->r_tid);
1014 ceph_mdsc_get_request(req);
1015 insert_request(&mdsc->request_tree, req);
1017 req->r_cred = get_current_cred();
1019 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
1020 mdsc->oldest_tid = req->r_tid;
1023 struct ceph_inode_info *ci = ceph_inode(dir);
1026 req->r_unsafe_dir = dir;
1027 spin_lock(&ci->i_unsafe_lock);
1028 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
1029 spin_unlock(&ci->i_unsafe_lock);
1033 static void __unregister_request(struct ceph_mds_client *mdsc,
1034 struct ceph_mds_request *req)
1036 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
1038 /* Never leave an unregistered request on an unsafe list! */
1039 list_del_init(&req->r_unsafe_item);
1041 if (req->r_tid == mdsc->oldest_tid) {
1042 struct rb_node *p = rb_next(&req->r_node);
1043 mdsc->oldest_tid = 0;
1045 struct ceph_mds_request *next_req =
1046 rb_entry(p, struct ceph_mds_request, r_node);
1047 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
1048 mdsc->oldest_tid = next_req->r_tid;
1055 erase_request(&mdsc->request_tree, req);
1057 if (req->r_unsafe_dir) {
1058 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
1059 spin_lock(&ci->i_unsafe_lock);
1060 list_del_init(&req->r_unsafe_dir_item);
1061 spin_unlock(&ci->i_unsafe_lock);
1063 if (req->r_target_inode &&
1064 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
1065 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
1066 spin_lock(&ci->i_unsafe_lock);
1067 list_del_init(&req->r_unsafe_target_item);
1068 spin_unlock(&ci->i_unsafe_lock);
1071 if (req->r_unsafe_dir) {
1072 iput(req->r_unsafe_dir);
1073 req->r_unsafe_dir = NULL;
1076 complete_all(&req->r_safe_completion);
1078 ceph_mdsc_put_request(req);
1082 * Walk back up the dentry tree until we hit a dentry representing a
1083 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
1084 * when calling this) to ensure that the objects won't disappear while we're
1085 * working with them. Once we hit a candidate dentry, we attempt to take a
1086 * reference to it, and return that as the result.
1088 static struct inode *get_nonsnap_parent(struct dentry *dentry)
1090 struct inode *inode = NULL;
1092 while (dentry && !IS_ROOT(dentry)) {
1093 inode = d_inode_rcu(dentry);
1094 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
1096 dentry = dentry->d_parent;
1099 inode = igrab(inode);
1104 * Choose mds to send request to next. If there is a hint set in the
1105 * request (e.g., due to a prior forward hint from the mds), use that.
1106 * Otherwise, consult frag tree and/or caps to identify the
1107 * appropriate mds. If all else fails, choose randomly.
1109 * Called under mdsc->mutex.
1111 static int __choose_mds(struct ceph_mds_client *mdsc,
1112 struct ceph_mds_request *req,
1115 struct inode *inode;
1116 struct ceph_inode_info *ci;
1117 struct ceph_cap *cap;
1118 int mode = req->r_direct_mode;
1120 u32 hash = req->r_direct_hash;
1121 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
1127 * is there a specific mds we should try? ignore hint if we have
1128 * no session and the mds is not up (active or recovering).
1130 if (req->r_resend_mds >= 0 &&
1131 (__have_session(mdsc, req->r_resend_mds) ||
1132 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1133 dout("%s using resend_mds mds%d\n", __func__,
1135 return req->r_resend_mds;
1138 if (mode == USE_RANDOM_MDS)
1143 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1144 inode = req->r_inode;
1147 /* req->r_dentry is non-null for LSSNAP request */
1149 inode = get_nonsnap_parent(req->r_dentry);
1151 dout("%s using snapdir's parent %p\n", __func__, inode);
1153 } else if (req->r_dentry) {
1154 /* ignore race with rename; old or new d_parent is okay */
1155 struct dentry *parent;
1159 parent = READ_ONCE(req->r_dentry->d_parent);
1160 dir = req->r_parent ? : d_inode_rcu(parent);
1162 if (!dir || dir->i_sb != mdsc->fsc->sb) {
1163 /* not this fs or parent went negative */
1164 inode = d_inode(req->r_dentry);
1167 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
1168 /* direct snapped/virtual snapdir requests
1169 * based on parent dir inode */
1170 inode = get_nonsnap_parent(parent);
1171 dout("%s using nonsnap parent %p\n", __func__, inode);
1174 inode = d_inode(req->r_dentry);
1175 if (!inode || mode == USE_AUTH_MDS) {
1178 hash = ceph_dentry_hash(dir, req->r_dentry);
1187 dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
1191 ci = ceph_inode(inode);
1193 if (is_hash && S_ISDIR(inode->i_mode)) {
1194 struct ceph_inode_frag frag;
1197 ceph_choose_frag(ci, hash, &frag, &found);
1199 if (mode == USE_ANY_MDS && frag.ndist > 0) {
1202 /* choose a random replica */
1203 get_random_bytes(&r, 1);
1206 dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
1207 __func__, inode, ceph_vinop(inode),
1208 frag.frag, mds, (int)r, frag.ndist);
1209 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1210 CEPH_MDS_STATE_ACTIVE &&
1211 !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1215 /* since this file/dir wasn't known to be
1216 * replicated, then we want to look for the
1217 * authoritative mds. */
1218 if (frag.mds >= 0) {
1219 /* choose auth mds */
1221 dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
1222 __func__, inode, ceph_vinop(inode),
1224 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1225 CEPH_MDS_STATE_ACTIVE) {
1226 if (!ceph_mdsmap_is_laggy(mdsc->mdsmap,
1231 mode = USE_AUTH_MDS;
1235 spin_lock(&ci->i_ceph_lock);
1237 if (mode == USE_AUTH_MDS)
1238 cap = ci->i_auth_cap;
1239 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1240 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1242 spin_unlock(&ci->i_ceph_lock);
1246 mds = cap->session->s_mds;
1247 dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1248 inode, ceph_vinop(inode), mds,
1249 cap == ci->i_auth_cap ? "auth " : "", cap);
1250 spin_unlock(&ci->i_ceph_lock);
1259 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1260 dout("%s chose random mds%d\n", __func__, mds);
1268 struct ceph_msg *ceph_create_session_msg(u32 op, u64 seq)
1270 struct ceph_msg *msg;
1271 struct ceph_mds_session_head *h;
1273 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1276 pr_err("ENOMEM creating session %s msg\n",
1277 ceph_session_op_name(op));
1280 h = msg->front.iov_base;
1281 h->op = cpu_to_le32(op);
1282 h->seq = cpu_to_le64(seq);
1287 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1288 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1289 static int encode_supported_features(void **p, void *end)
1291 static const size_t count = ARRAY_SIZE(feature_bits);
1295 size_t size = FEATURE_BYTES(count);
1298 if (WARN_ON_ONCE(*p + 4 + size > end))
1301 ceph_encode_32(p, size);
1302 memset(*p, 0, size);
1303 for (i = 0; i < count; i++) {
1304 bit = feature_bits[i];
1305 ((unsigned char *)(*p))[bit / 8] |= BIT(bit % 8);
1309 if (WARN_ON_ONCE(*p + 4 > end))
1312 ceph_encode_32(p, 0);
1318 static const unsigned char metric_bits[] = CEPHFS_METRIC_SPEC_CLIENT_SUPPORTED;
1319 #define METRIC_BYTES(cnt) (DIV_ROUND_UP((size_t)metric_bits[cnt - 1] + 1, 64) * 8)
1320 static int encode_metric_spec(void **p, void *end)
1322 static const size_t count = ARRAY_SIZE(metric_bits);
1325 if (WARN_ON_ONCE(*p + 2 > end))
1328 ceph_encode_8(p, 1); /* version */
1329 ceph_encode_8(p, 1); /* compat */
1333 size_t size = METRIC_BYTES(count);
1335 if (WARN_ON_ONCE(*p + 4 + 4 + size > end))
1338 /* metric spec info length */
1339 ceph_encode_32(p, 4 + size);
1342 ceph_encode_32(p, size);
1343 memset(*p, 0, size);
1344 for (i = 0; i < count; i++)
1345 ((unsigned char *)(*p))[i / 8] |= BIT(metric_bits[i] % 8);
1348 if (WARN_ON_ONCE(*p + 4 + 4 > end))
1351 /* metric spec info length */
1352 ceph_encode_32(p, 4);
1354 ceph_encode_32(p, 0);
1361 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1362 * to include additional client metadata fields.
1364 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1366 struct ceph_msg *msg;
1367 struct ceph_mds_session_head *h;
1369 int extra_bytes = 0;
1370 int metadata_key_count = 0;
1371 struct ceph_options *opt = mdsc->fsc->client->options;
1372 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1377 const char* metadata[][2] = {
1378 {"hostname", mdsc->nodename},
1379 {"kernel_version", init_utsname()->release},
1380 {"entity_id", opt->name ? : ""},
1381 {"root", fsopt->server_path ? : "/"},
1385 /* Calculate serialized length of metadata */
1386 extra_bytes = 4; /* map length */
1387 for (i = 0; metadata[i][0]; ++i) {
1388 extra_bytes += 8 + strlen(metadata[i][0]) +
1389 strlen(metadata[i][1]);
1390 metadata_key_count++;
1393 /* supported feature */
1395 count = ARRAY_SIZE(feature_bits);
1397 size = FEATURE_BYTES(count);
1398 extra_bytes += 4 + size;
1402 count = ARRAY_SIZE(metric_bits);
1404 size = METRIC_BYTES(count);
1405 extra_bytes += 2 + 4 + 4 + size;
1407 /* Allocate the message */
1408 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1411 pr_err("ENOMEM creating session open msg\n");
1412 return ERR_PTR(-ENOMEM);
1414 p = msg->front.iov_base;
1415 end = p + msg->front.iov_len;
1418 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1419 h->seq = cpu_to_le64(seq);
1422 * Serialize client metadata into waiting buffer space, using
1423 * the format that userspace expects for map<string, string>
1425 * ClientSession messages with metadata are v4
1427 msg->hdr.version = cpu_to_le16(4);
1428 msg->hdr.compat_version = cpu_to_le16(1);
1430 /* The write pointer, following the session_head structure */
1433 /* Number of entries in the map */
1434 ceph_encode_32(&p, metadata_key_count);
1436 /* Two length-prefixed strings for each entry in the map */
1437 for (i = 0; metadata[i][0]; ++i) {
1438 size_t const key_len = strlen(metadata[i][0]);
1439 size_t const val_len = strlen(metadata[i][1]);
1441 ceph_encode_32(&p, key_len);
1442 memcpy(p, metadata[i][0], key_len);
1444 ceph_encode_32(&p, val_len);
1445 memcpy(p, metadata[i][1], val_len);
1449 ret = encode_supported_features(&p, end);
1451 pr_err("encode_supported_features failed!\n");
1453 return ERR_PTR(ret);
1456 ret = encode_metric_spec(&p, end);
1458 pr_err("encode_metric_spec failed!\n");
1460 return ERR_PTR(ret);
1463 msg->front.iov_len = p - msg->front.iov_base;
1464 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1470 * send session open request.
1472 * called under mdsc->mutex
1474 static int __open_session(struct ceph_mds_client *mdsc,
1475 struct ceph_mds_session *session)
1477 struct ceph_msg *msg;
1479 int mds = session->s_mds;
1481 /* wait for mds to go active? */
1482 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1483 dout("open_session to mds%d (%s)\n", mds,
1484 ceph_mds_state_name(mstate));
1485 session->s_state = CEPH_MDS_SESSION_OPENING;
1486 session->s_renew_requested = jiffies;
1488 /* send connect message */
1489 msg = create_session_open_msg(mdsc, session->s_seq);
1491 return PTR_ERR(msg);
1492 ceph_con_send(&session->s_con, msg);
1497 * open sessions for any export targets for the given mds
1499 * called under mdsc->mutex
1501 static struct ceph_mds_session *
1502 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1504 struct ceph_mds_session *session;
1507 session = __ceph_lookup_mds_session(mdsc, target);
1509 session = register_session(mdsc, target);
1510 if (IS_ERR(session))
1513 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1514 session->s_state == CEPH_MDS_SESSION_CLOSING) {
1515 ret = __open_session(mdsc, session);
1517 return ERR_PTR(ret);
1523 struct ceph_mds_session *
1524 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1526 struct ceph_mds_session *session;
1528 dout("open_export_target_session to mds%d\n", target);
1530 mutex_lock(&mdsc->mutex);
1531 session = __open_export_target_session(mdsc, target);
1532 mutex_unlock(&mdsc->mutex);
1537 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1538 struct ceph_mds_session *session)
1540 struct ceph_mds_info *mi;
1541 struct ceph_mds_session *ts;
1542 int i, mds = session->s_mds;
1544 if (mds >= mdsc->mdsmap->possible_max_rank)
1547 mi = &mdsc->mdsmap->m_info[mds];
1548 dout("open_export_target_sessions for mds%d (%d targets)\n",
1549 session->s_mds, mi->num_export_targets);
1551 for (i = 0; i < mi->num_export_targets; i++) {
1552 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1553 ceph_put_mds_session(ts);
1557 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1558 struct ceph_mds_session *session)
1560 mutex_lock(&mdsc->mutex);
1561 __open_export_target_sessions(mdsc, session);
1562 mutex_unlock(&mdsc->mutex);
1569 static void detach_cap_releases(struct ceph_mds_session *session,
1570 struct list_head *target)
1572 lockdep_assert_held(&session->s_cap_lock);
1574 list_splice_init(&session->s_cap_releases, target);
1575 session->s_num_cap_releases = 0;
1576 dout("dispose_cap_releases mds%d\n", session->s_mds);
1579 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1580 struct list_head *dispose)
1582 while (!list_empty(dispose)) {
1583 struct ceph_cap *cap;
1584 /* zero out the in-progress message */
1585 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1586 list_del(&cap->session_caps);
1587 ceph_put_cap(mdsc, cap);
1591 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1592 struct ceph_mds_session *session)
1594 struct ceph_mds_request *req;
1597 dout("cleanup_session_requests mds%d\n", session->s_mds);
1598 mutex_lock(&mdsc->mutex);
1599 while (!list_empty(&session->s_unsafe)) {
1600 req = list_first_entry(&session->s_unsafe,
1601 struct ceph_mds_request, r_unsafe_item);
1602 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1604 if (req->r_target_inode)
1605 mapping_set_error(req->r_target_inode->i_mapping, -EIO);
1606 if (req->r_unsafe_dir)
1607 mapping_set_error(req->r_unsafe_dir->i_mapping, -EIO);
1608 __unregister_request(mdsc, req);
1610 /* zero r_attempts, so kick_requests() will re-send requests */
1611 p = rb_first(&mdsc->request_tree);
1613 req = rb_entry(p, struct ceph_mds_request, r_node);
1615 if (req->r_session &&
1616 req->r_session->s_mds == session->s_mds)
1617 req->r_attempts = 0;
1619 mutex_unlock(&mdsc->mutex);
1623 * Helper to safely iterate over all caps associated with a session, with
1624 * special care taken to handle a racing __ceph_remove_cap().
1626 * Caller must hold session s_mutex.
1628 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1629 int (*cb)(struct inode *, struct ceph_cap *,
1632 struct list_head *p;
1633 struct ceph_cap *cap;
1634 struct inode *inode, *last_inode = NULL;
1635 struct ceph_cap *old_cap = NULL;
1638 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1639 spin_lock(&session->s_cap_lock);
1640 p = session->s_caps.next;
1641 while (p != &session->s_caps) {
1642 cap = list_entry(p, struct ceph_cap, session_caps);
1643 inode = igrab(&cap->ci->netfs.inode);
1648 session->s_cap_iterator = cap;
1649 spin_unlock(&session->s_cap_lock);
1656 ceph_put_cap(session->s_mdsc, old_cap);
1660 ret = cb(inode, cap, arg);
1663 spin_lock(&session->s_cap_lock);
1666 dout("iterate_session_caps finishing cap %p removal\n",
1668 BUG_ON(cap->session != session);
1669 cap->session = NULL;
1670 list_del_init(&cap->session_caps);
1671 session->s_nr_caps--;
1672 atomic64_dec(&session->s_mdsc->metric.total_caps);
1673 if (cap->queue_release)
1674 __ceph_queue_cap_release(session, cap);
1676 old_cap = cap; /* put_cap it w/o locks held */
1683 session->s_cap_iterator = NULL;
1684 spin_unlock(&session->s_cap_lock);
1688 ceph_put_cap(session->s_mdsc, old_cap);
1693 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1696 struct ceph_inode_info *ci = ceph_inode(inode);
1697 bool invalidate = false;
1700 dout("removing cap %p, ci is %p, inode is %p\n",
1701 cap, ci, &ci->netfs.inode);
1702 spin_lock(&ci->i_ceph_lock);
1703 iputs = ceph_purge_inode_cap(inode, cap, &invalidate);
1704 spin_unlock(&ci->i_ceph_lock);
1706 wake_up_all(&ci->i_cap_wq);
1708 ceph_queue_invalidate(inode);
1715 * caller must hold session s_mutex
1717 static void remove_session_caps(struct ceph_mds_session *session)
1719 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1720 struct super_block *sb = fsc->sb;
1723 dout("remove_session_caps on %p\n", session);
1724 ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1726 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1728 spin_lock(&session->s_cap_lock);
1729 if (session->s_nr_caps > 0) {
1730 struct inode *inode;
1731 struct ceph_cap *cap, *prev = NULL;
1732 struct ceph_vino vino;
1734 * iterate_session_caps() skips inodes that are being
1735 * deleted, we need to wait until deletions are complete.
1736 * __wait_on_freeing_inode() is designed for the job,
1737 * but it is not exported, so use lookup inode function
1740 while (!list_empty(&session->s_caps)) {
1741 cap = list_entry(session->s_caps.next,
1742 struct ceph_cap, session_caps);
1746 vino = cap->ci->i_vino;
1747 spin_unlock(&session->s_cap_lock);
1749 inode = ceph_find_inode(sb, vino);
1752 spin_lock(&session->s_cap_lock);
1756 // drop cap expires and unlock s_cap_lock
1757 detach_cap_releases(session, &dispose);
1759 BUG_ON(session->s_nr_caps > 0);
1760 BUG_ON(!list_empty(&session->s_cap_flushing));
1761 spin_unlock(&session->s_cap_lock);
1762 dispose_cap_releases(session->s_mdsc, &dispose);
1772 * wake up any threads waiting on this session's caps. if the cap is
1773 * old (didn't get renewed on the client reconnect), remove it now.
1775 * caller must hold s_mutex.
1777 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1780 struct ceph_inode_info *ci = ceph_inode(inode);
1781 unsigned long ev = (unsigned long)arg;
1783 if (ev == RECONNECT) {
1784 spin_lock(&ci->i_ceph_lock);
1785 ci->i_wanted_max_size = 0;
1786 ci->i_requested_max_size = 0;
1787 spin_unlock(&ci->i_ceph_lock);
1788 } else if (ev == RENEWCAPS) {
1789 if (cap->cap_gen < atomic_read(&cap->session->s_cap_gen)) {
1790 /* mds did not re-issue stale cap */
1791 spin_lock(&ci->i_ceph_lock);
1792 cap->issued = cap->implemented = CEPH_CAP_PIN;
1793 spin_unlock(&ci->i_ceph_lock);
1795 } else if (ev == FORCE_RO) {
1797 wake_up_all(&ci->i_cap_wq);
1801 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1803 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1804 ceph_iterate_session_caps(session, wake_up_session_cb,
1805 (void *)(unsigned long)ev);
1809 * Send periodic message to MDS renewing all currently held caps. The
1810 * ack will reset the expiration for all caps from this session.
1812 * caller holds s_mutex
1814 static int send_renew_caps(struct ceph_mds_client *mdsc,
1815 struct ceph_mds_session *session)
1817 struct ceph_msg *msg;
1820 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1821 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1822 pr_info("mds%d caps stale\n", session->s_mds);
1823 session->s_renew_requested = jiffies;
1825 /* do not try to renew caps until a recovering mds has reconnected
1826 * with its clients. */
1827 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1828 if (state < CEPH_MDS_STATE_RECONNECT) {
1829 dout("send_renew_caps ignoring mds%d (%s)\n",
1830 session->s_mds, ceph_mds_state_name(state));
1834 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1835 ceph_mds_state_name(state));
1836 msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1837 ++session->s_renew_seq);
1840 ceph_con_send(&session->s_con, msg);
1844 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1845 struct ceph_mds_session *session, u64 seq)
1847 struct ceph_msg *msg;
1849 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1850 session->s_mds, ceph_session_state_name(session->s_state), seq);
1851 msg = ceph_create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1854 ceph_con_send(&session->s_con, msg);
1860 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1862 * Called under session->s_mutex
1864 static void renewed_caps(struct ceph_mds_client *mdsc,
1865 struct ceph_mds_session *session, int is_renew)
1870 spin_lock(&session->s_cap_lock);
1871 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1873 session->s_cap_ttl = session->s_renew_requested +
1874 mdsc->mdsmap->m_session_timeout*HZ;
1877 if (time_before(jiffies, session->s_cap_ttl)) {
1878 pr_info("mds%d caps renewed\n", session->s_mds);
1881 pr_info("mds%d caps still stale\n", session->s_mds);
1884 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1885 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1886 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1887 spin_unlock(&session->s_cap_lock);
1890 wake_up_session_caps(session, RENEWCAPS);
1894 * send a session close request
1896 static int request_close_session(struct ceph_mds_session *session)
1898 struct ceph_msg *msg;
1900 dout("request_close_session mds%d state %s seq %lld\n",
1901 session->s_mds, ceph_session_state_name(session->s_state),
1903 msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_CLOSE,
1907 ceph_con_send(&session->s_con, msg);
1912 * Called with s_mutex held.
1914 static int __close_session(struct ceph_mds_client *mdsc,
1915 struct ceph_mds_session *session)
1917 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1919 session->s_state = CEPH_MDS_SESSION_CLOSING;
1920 return request_close_session(session);
1923 static bool drop_negative_children(struct dentry *dentry)
1925 struct dentry *child;
1926 bool all_negative = true;
1928 if (!d_is_dir(dentry))
1931 spin_lock(&dentry->d_lock);
1932 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1933 if (d_really_is_positive(child)) {
1934 all_negative = false;
1938 spin_unlock(&dentry->d_lock);
1941 shrink_dcache_parent(dentry);
1943 return all_negative;
1947 * Trim old(er) caps.
1949 * Because we can't cache an inode without one or more caps, we do
1950 * this indirectly: if a cap is unused, we prune its aliases, at which
1951 * point the inode will hopefully get dropped to.
1953 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1954 * memory pressure from the MDS, though, so it needn't be perfect.
1956 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1958 int *remaining = arg;
1959 struct ceph_inode_info *ci = ceph_inode(inode);
1960 int used, wanted, oissued, mine;
1962 if (*remaining <= 0)
1965 spin_lock(&ci->i_ceph_lock);
1966 mine = cap->issued | cap->implemented;
1967 used = __ceph_caps_used(ci);
1968 wanted = __ceph_caps_file_wanted(ci);
1969 oissued = __ceph_caps_issued_other(ci, cap);
1971 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1972 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1973 ceph_cap_string(used), ceph_cap_string(wanted));
1974 if (cap == ci->i_auth_cap) {
1975 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1976 !list_empty(&ci->i_cap_snaps))
1978 if ((used | wanted) & CEPH_CAP_ANY_WR)
1980 /* Note: it's possible that i_filelock_ref becomes non-zero
1981 * after dropping auth caps. It doesn't hurt because reply
1982 * of lock mds request will re-add auth caps. */
1983 if (atomic_read(&ci->i_filelock_ref) > 0)
1986 /* The inode has cached pages, but it's no longer used.
1987 * we can safely drop it */
1988 if (S_ISREG(inode->i_mode) &&
1989 wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1990 !(oissued & CEPH_CAP_FILE_CACHE)) {
1994 if ((used | wanted) & ~oissued & mine)
1995 goto out; /* we need these caps */
1998 /* we aren't the only cap.. just remove us */
1999 ceph_remove_cap(cap, true);
2002 struct dentry *dentry;
2003 /* try dropping referring dentries */
2004 spin_unlock(&ci->i_ceph_lock);
2005 dentry = d_find_any_alias(inode);
2006 if (dentry && drop_negative_children(dentry)) {
2009 d_prune_aliases(inode);
2010 count = atomic_read(&inode->i_count);
2013 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
2022 spin_unlock(&ci->i_ceph_lock);
2027 * Trim session cap count down to some max number.
2029 int ceph_trim_caps(struct ceph_mds_client *mdsc,
2030 struct ceph_mds_session *session,
2033 int trim_caps = session->s_nr_caps - max_caps;
2035 dout("trim_caps mds%d start: %d / %d, trim %d\n",
2036 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
2037 if (trim_caps > 0) {
2038 int remaining = trim_caps;
2040 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
2041 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
2042 session->s_mds, session->s_nr_caps, max_caps,
2043 trim_caps - remaining);
2046 ceph_flush_cap_releases(mdsc, session);
2050 static int check_caps_flush(struct ceph_mds_client *mdsc,
2055 spin_lock(&mdsc->cap_dirty_lock);
2056 if (!list_empty(&mdsc->cap_flush_list)) {
2057 struct ceph_cap_flush *cf =
2058 list_first_entry(&mdsc->cap_flush_list,
2059 struct ceph_cap_flush, g_list);
2060 if (cf->tid <= want_flush_tid) {
2061 dout("check_caps_flush still flushing tid "
2062 "%llu <= %llu\n", cf->tid, want_flush_tid);
2066 spin_unlock(&mdsc->cap_dirty_lock);
2071 * flush all dirty inode data to disk.
2073 * returns true if we've flushed through want_flush_tid
2075 static void wait_caps_flush(struct ceph_mds_client *mdsc,
2078 dout("check_caps_flush want %llu\n", want_flush_tid);
2080 wait_event(mdsc->cap_flushing_wq,
2081 check_caps_flush(mdsc, want_flush_tid));
2083 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
2087 * called under s_mutex
2089 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2090 struct ceph_mds_session *session)
2092 struct ceph_msg *msg = NULL;
2093 struct ceph_mds_cap_release *head;
2094 struct ceph_mds_cap_item *item;
2095 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2096 struct ceph_cap *cap;
2097 LIST_HEAD(tmp_list);
2098 int num_cap_releases;
2099 __le32 barrier, *cap_barrier;
2101 down_read(&osdc->lock);
2102 barrier = cpu_to_le32(osdc->epoch_barrier);
2103 up_read(&osdc->lock);
2105 spin_lock(&session->s_cap_lock);
2107 list_splice_init(&session->s_cap_releases, &tmp_list);
2108 num_cap_releases = session->s_num_cap_releases;
2109 session->s_num_cap_releases = 0;
2110 spin_unlock(&session->s_cap_lock);
2112 while (!list_empty(&tmp_list)) {
2114 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2115 PAGE_SIZE, GFP_NOFS, false);
2118 head = msg->front.iov_base;
2119 head->num = cpu_to_le32(0);
2120 msg->front.iov_len = sizeof(*head);
2122 msg->hdr.version = cpu_to_le16(2);
2123 msg->hdr.compat_version = cpu_to_le16(1);
2126 cap = list_first_entry(&tmp_list, struct ceph_cap,
2128 list_del(&cap->session_caps);
2131 head = msg->front.iov_base;
2132 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2134 item = msg->front.iov_base + msg->front.iov_len;
2135 item->ino = cpu_to_le64(cap->cap_ino);
2136 item->cap_id = cpu_to_le64(cap->cap_id);
2137 item->migrate_seq = cpu_to_le32(cap->mseq);
2138 item->seq = cpu_to_le32(cap->issue_seq);
2139 msg->front.iov_len += sizeof(*item);
2141 ceph_put_cap(mdsc, cap);
2143 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2144 // Append cap_barrier field
2145 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2146 *cap_barrier = barrier;
2147 msg->front.iov_len += sizeof(*cap_barrier);
2149 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2150 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2151 ceph_con_send(&session->s_con, msg);
2156 BUG_ON(num_cap_releases != 0);
2158 spin_lock(&session->s_cap_lock);
2159 if (!list_empty(&session->s_cap_releases))
2161 spin_unlock(&session->s_cap_lock);
2164 // Append cap_barrier field
2165 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2166 *cap_barrier = barrier;
2167 msg->front.iov_len += sizeof(*cap_barrier);
2169 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2170 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2171 ceph_con_send(&session->s_con, msg);
2175 pr_err("send_cap_releases mds%d, failed to allocate message\n",
2177 spin_lock(&session->s_cap_lock);
2178 list_splice(&tmp_list, &session->s_cap_releases);
2179 session->s_num_cap_releases += num_cap_releases;
2180 spin_unlock(&session->s_cap_lock);
2183 static void ceph_cap_release_work(struct work_struct *work)
2185 struct ceph_mds_session *session =
2186 container_of(work, struct ceph_mds_session, s_cap_release_work);
2188 mutex_lock(&session->s_mutex);
2189 if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2190 session->s_state == CEPH_MDS_SESSION_HUNG)
2191 ceph_send_cap_releases(session->s_mdsc, session);
2192 mutex_unlock(&session->s_mutex);
2193 ceph_put_mds_session(session);
2196 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2197 struct ceph_mds_session *session)
2202 ceph_get_mds_session(session);
2203 if (queue_work(mdsc->fsc->cap_wq,
2204 &session->s_cap_release_work)) {
2205 dout("cap release work queued\n");
2207 ceph_put_mds_session(session);
2208 dout("failed to queue cap release work\n");
2213 * caller holds session->s_cap_lock
2215 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2216 struct ceph_cap *cap)
2218 list_add_tail(&cap->session_caps, &session->s_cap_releases);
2219 session->s_num_cap_releases++;
2221 if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2222 ceph_flush_cap_releases(session->s_mdsc, session);
2225 static void ceph_cap_reclaim_work(struct work_struct *work)
2227 struct ceph_mds_client *mdsc =
2228 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2229 int ret = ceph_trim_dentries(mdsc);
2231 ceph_queue_cap_reclaim_work(mdsc);
2234 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2239 if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2240 dout("caps reclaim work queued\n");
2242 dout("failed to queue caps release work\n");
2246 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2251 val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2252 if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2253 atomic_set(&mdsc->cap_reclaim_pending, 0);
2254 ceph_queue_cap_reclaim_work(mdsc);
2262 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2265 struct ceph_inode_info *ci = ceph_inode(dir);
2266 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2267 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2268 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2269 unsigned int num_entries;
2272 spin_lock(&ci->i_ceph_lock);
2273 num_entries = ci->i_files + ci->i_subdirs;
2274 spin_unlock(&ci->i_ceph_lock);
2275 num_entries = max(num_entries, 1U);
2276 num_entries = min(num_entries, opt->max_readdir);
2278 order = get_order(size * num_entries);
2279 while (order >= 0) {
2280 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2284 if (rinfo->dir_entries)
2288 if (!rinfo->dir_entries)
2291 num_entries = (PAGE_SIZE << order) / size;
2292 num_entries = min(num_entries, opt->max_readdir);
2294 rinfo->dir_buf_size = PAGE_SIZE << order;
2295 req->r_num_caps = num_entries + 1;
2296 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2297 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2302 * Create an mds request.
2304 struct ceph_mds_request *
2305 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2307 struct ceph_mds_request *req;
2309 req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2311 return ERR_PTR(-ENOMEM);
2313 mutex_init(&req->r_fill_mutex);
2315 req->r_started = jiffies;
2316 req->r_start_latency = ktime_get();
2317 req->r_resend_mds = -1;
2318 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2319 INIT_LIST_HEAD(&req->r_unsafe_target_item);
2321 kref_init(&req->r_kref);
2322 RB_CLEAR_NODE(&req->r_node);
2323 INIT_LIST_HEAD(&req->r_wait);
2324 init_completion(&req->r_completion);
2325 init_completion(&req->r_safe_completion);
2326 INIT_LIST_HEAD(&req->r_unsafe_item);
2328 ktime_get_coarse_real_ts64(&req->r_stamp);
2331 req->r_direct_mode = mode;
2336 * return oldest (lowest) request, tid in request tree, 0 if none.
2338 * called under mdsc->mutex.
2340 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2342 if (RB_EMPTY_ROOT(&mdsc->request_tree))
2344 return rb_entry(rb_first(&mdsc->request_tree),
2345 struct ceph_mds_request, r_node);
2348 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2350 return mdsc->oldest_tid;
2354 * Build a dentry's path. Allocate on heap; caller must kfree. Based
2355 * on build_path_from_dentry in fs/cifs/dir.c.
2357 * If @stop_on_nosnap, generate path relative to the first non-snapped
2360 * Encode hidden .snap dirs as a double /, i.e.
2361 * foo/.snap/bar -> foo//bar
2363 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2366 struct dentry *temp;
2373 return ERR_PTR(-EINVAL);
2377 return ERR_PTR(-ENOMEM);
2382 seq = read_seqbegin(&rename_lock);
2386 struct inode *inode;
2388 spin_lock(&temp->d_lock);
2389 inode = d_inode(temp);
2390 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2391 dout("build_path path+%d: %p SNAPDIR\n",
2393 } else if (stop_on_nosnap && inode && dentry != temp &&
2394 ceph_snap(inode) == CEPH_NOSNAP) {
2395 spin_unlock(&temp->d_lock);
2396 pos++; /* get rid of any prepended '/' */
2399 pos -= temp->d_name.len;
2401 spin_unlock(&temp->d_lock);
2404 memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2406 spin_unlock(&temp->d_lock);
2407 temp = READ_ONCE(temp->d_parent);
2409 /* Are we at the root? */
2413 /* Are we out of buffer? */
2419 base = ceph_ino(d_inode(temp));
2422 if (read_seqretry(&rename_lock, seq))
2427 * A rename didn't occur, but somehow we didn't end up where
2428 * we thought we would. Throw a warning and try again.
2430 pr_warn("build_path did not end path lookup where "
2431 "expected, pos is %d\n", pos);
2436 *plen = PATH_MAX - 1 - pos;
2437 dout("build_path on %p %d built %llx '%.*s'\n",
2438 dentry, d_count(dentry), base, *plen, path + pos);
2442 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2443 const char **ppath, int *ppathlen, u64 *pino,
2444 bool *pfreepath, bool parent_locked)
2450 dir = d_inode_rcu(dentry->d_parent);
2451 if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2452 *pino = ceph_ino(dir);
2454 *ppath = dentry->d_name.name;
2455 *ppathlen = dentry->d_name.len;
2459 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2461 return PTR_ERR(path);
2467 static int build_inode_path(struct inode *inode,
2468 const char **ppath, int *ppathlen, u64 *pino,
2471 struct dentry *dentry;
2474 if (ceph_snap(inode) == CEPH_NOSNAP) {
2475 *pino = ceph_ino(inode);
2479 dentry = d_find_alias(inode);
2480 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2483 return PTR_ERR(path);
2490 * request arguments may be specified via an inode *, a dentry *, or
2491 * an explicit ino+path.
2493 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2494 struct inode *rdiri, const char *rpath,
2495 u64 rino, const char **ppath, int *pathlen,
2496 u64 *ino, bool *freepath, bool parent_locked)
2501 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2502 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2504 } else if (rdentry) {
2505 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2506 freepath, parent_locked);
2507 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2509 } else if (rpath || rino) {
2512 *pathlen = rpath ? strlen(rpath) : 0;
2513 dout(" path %.*s\n", *pathlen, rpath);
2519 static void encode_timestamp_and_gids(void **p,
2520 const struct ceph_mds_request *req)
2522 struct ceph_timespec ts;
2525 ceph_encode_timespec64(&ts, &req->r_stamp);
2526 ceph_encode_copy(p, &ts, sizeof(ts));
2529 ceph_encode_32(p, req->r_cred->group_info->ngroups);
2530 for (i = 0; i < req->r_cred->group_info->ngroups; i++)
2531 ceph_encode_64(p, from_kgid(&init_user_ns,
2532 req->r_cred->group_info->gid[i]));
2536 * called under mdsc->mutex
2538 static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
2539 struct ceph_mds_request *req,
2540 bool drop_cap_releases)
2542 int mds = session->s_mds;
2543 struct ceph_mds_client *mdsc = session->s_mdsc;
2544 struct ceph_msg *msg;
2545 struct ceph_mds_request_head_old *head;
2546 const char *path1 = NULL;
2547 const char *path2 = NULL;
2548 u64 ino1 = 0, ino2 = 0;
2549 int pathlen1 = 0, pathlen2 = 0;
2550 bool freepath1 = false, freepath2 = false;
2555 bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
2557 ret = set_request_path_attr(req->r_inode, req->r_dentry,
2558 req->r_parent, req->r_path1, req->r_ino1.ino,
2559 &path1, &pathlen1, &ino1, &freepath1,
2560 test_bit(CEPH_MDS_R_PARENT_LOCKED,
2561 &req->r_req_flags));
2567 /* If r_old_dentry is set, then assume that its parent is locked */
2568 ret = set_request_path_attr(NULL, req->r_old_dentry,
2569 req->r_old_dentry_dir,
2570 req->r_path2, req->r_ino2.ino,
2571 &path2, &pathlen2, &ino2, &freepath2, true);
2577 len = legacy ? sizeof(*head) : sizeof(struct ceph_mds_request_head);
2578 len += pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2579 sizeof(struct ceph_timespec);
2580 len += sizeof(u32) + (sizeof(u64) * req->r_cred->group_info->ngroups);
2582 /* calculate (max) length for cap releases */
2583 len += sizeof(struct ceph_mds_request_release) *
2584 (!!req->r_inode_drop + !!req->r_dentry_drop +
2585 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2587 if (req->r_dentry_drop)
2589 if (req->r_old_dentry_drop)
2592 msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2594 msg = ERR_PTR(-ENOMEM);
2598 msg->hdr.tid = cpu_to_le64(req->r_tid);
2601 * The old ceph_mds_request_head didn't contain a version field, and
2602 * one was added when we moved the message version from 3->4.
2605 msg->hdr.version = cpu_to_le16(3);
2606 head = msg->front.iov_base;
2607 p = msg->front.iov_base + sizeof(*head);
2609 struct ceph_mds_request_head *new_head = msg->front.iov_base;
2611 msg->hdr.version = cpu_to_le16(4);
2612 new_head->version = cpu_to_le16(CEPH_MDS_REQUEST_HEAD_VERSION);
2613 head = (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2614 p = msg->front.iov_base + sizeof(*new_head);
2617 end = msg->front.iov_base + msg->front.iov_len;
2619 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2620 head->op = cpu_to_le32(req->r_op);
2621 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
2622 req->r_cred->fsuid));
2623 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
2624 req->r_cred->fsgid));
2625 head->ino = cpu_to_le64(req->r_deleg_ino);
2626 head->args = req->r_args;
2628 ceph_encode_filepath(&p, end, ino1, path1);
2629 ceph_encode_filepath(&p, end, ino2, path2);
2631 /* make note of release offset, in case we need to replay */
2632 req->r_request_release_offset = p - msg->front.iov_base;
2636 if (req->r_inode_drop)
2637 releases += ceph_encode_inode_release(&p,
2638 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2639 mds, req->r_inode_drop, req->r_inode_unless,
2640 req->r_op == CEPH_MDS_OP_READDIR);
2641 if (req->r_dentry_drop)
2642 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2643 req->r_parent, mds, req->r_dentry_drop,
2644 req->r_dentry_unless);
2645 if (req->r_old_dentry_drop)
2646 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2647 req->r_old_dentry_dir, mds,
2648 req->r_old_dentry_drop,
2649 req->r_old_dentry_unless);
2650 if (req->r_old_inode_drop)
2651 releases += ceph_encode_inode_release(&p,
2652 d_inode(req->r_old_dentry),
2653 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2655 if (drop_cap_releases) {
2657 p = msg->front.iov_base + req->r_request_release_offset;
2660 head->num_releases = cpu_to_le16(releases);
2662 encode_timestamp_and_gids(&p, req);
2664 if (WARN_ON_ONCE(p > end)) {
2666 msg = ERR_PTR(-ERANGE);
2670 msg->front.iov_len = p - msg->front.iov_base;
2671 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2673 if (req->r_pagelist) {
2674 struct ceph_pagelist *pagelist = req->r_pagelist;
2675 ceph_msg_data_add_pagelist(msg, pagelist);
2676 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2678 msg->hdr.data_len = 0;
2681 msg->hdr.data_off = cpu_to_le16(0);
2685 ceph_mdsc_free_path((char *)path2, pathlen2);
2688 ceph_mdsc_free_path((char *)path1, pathlen1);
2694 * called under mdsc->mutex if error, under no mutex if
2697 static void complete_request(struct ceph_mds_client *mdsc,
2698 struct ceph_mds_request *req)
2700 req->r_end_latency = ktime_get();
2702 if (req->r_callback)
2703 req->r_callback(mdsc, req);
2704 complete_all(&req->r_completion);
2707 static struct ceph_mds_request_head_old *
2708 find_old_request_head(void *p, u64 features)
2710 bool legacy = !(features & CEPH_FEATURE_FS_BTIME);
2711 struct ceph_mds_request_head *new_head;
2714 return (struct ceph_mds_request_head_old *)p;
2715 new_head = (struct ceph_mds_request_head *)p;
2716 return (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2720 * called under mdsc->mutex
2722 static int __prepare_send_request(struct ceph_mds_session *session,
2723 struct ceph_mds_request *req,
2724 bool drop_cap_releases)
2726 int mds = session->s_mds;
2727 struct ceph_mds_client *mdsc = session->s_mdsc;
2728 struct ceph_mds_request_head_old *rhead;
2729 struct ceph_msg *msg;
2730 int flags = 0, max_retry;
2733 * The type of 'r_attempts' in kernel 'ceph_mds_request'
2734 * is 'int', while in 'ceph_mds_request_head' the type of
2735 * 'num_retry' is '__u8'. So in case the request retries
2736 * exceeding 256 times, the MDS will receive a incorrect
2739 * In this case it's ususally a bug in MDS and continue
2740 * retrying the request makes no sense.
2742 * In future this could be fixed in ceph code, so avoid
2743 * using the hardcode here.
2745 max_retry = sizeof_field(struct ceph_mds_request_head, num_retry);
2746 max_retry = 1 << (max_retry * BITS_PER_BYTE);
2747 if (req->r_attempts >= max_retry) {
2748 pr_warn_ratelimited("%s request tid %llu seq overflow\n",
2749 __func__, req->r_tid);
2755 struct ceph_cap *cap =
2756 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2759 req->r_sent_on_mseq = cap->mseq;
2761 req->r_sent_on_mseq = -1;
2763 dout("%s %p tid %lld %s (attempt %d)\n", __func__, req,
2764 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2766 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2770 * Replay. Do not regenerate message (and rebuild
2771 * paths, etc.); just use the original message.
2772 * Rebuilding paths will break for renames because
2773 * d_move mangles the src name.
2775 msg = req->r_request;
2776 rhead = find_old_request_head(msg->front.iov_base,
2777 session->s_con.peer_features);
2779 flags = le32_to_cpu(rhead->flags);
2780 flags |= CEPH_MDS_FLAG_REPLAY;
2781 rhead->flags = cpu_to_le32(flags);
2783 if (req->r_target_inode)
2784 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2786 rhead->num_retry = req->r_attempts - 1;
2788 /* remove cap/dentry releases from message */
2789 rhead->num_releases = 0;
2791 p = msg->front.iov_base + req->r_request_release_offset;
2792 encode_timestamp_and_gids(&p, req);
2794 msg->front.iov_len = p - msg->front.iov_base;
2795 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2799 if (req->r_request) {
2800 ceph_msg_put(req->r_request);
2801 req->r_request = NULL;
2803 msg = create_request_message(session, req, drop_cap_releases);
2805 req->r_err = PTR_ERR(msg);
2806 return PTR_ERR(msg);
2808 req->r_request = msg;
2810 rhead = find_old_request_head(msg->front.iov_base,
2811 session->s_con.peer_features);
2812 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2813 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2814 flags |= CEPH_MDS_FLAG_REPLAY;
2815 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2816 flags |= CEPH_MDS_FLAG_ASYNC;
2818 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2819 rhead->flags = cpu_to_le32(flags);
2820 rhead->num_fwd = req->r_num_fwd;
2821 rhead->num_retry = req->r_attempts - 1;
2823 dout(" r_parent = %p\n", req->r_parent);
2828 * called under mdsc->mutex
2830 static int __send_request(struct ceph_mds_session *session,
2831 struct ceph_mds_request *req,
2832 bool drop_cap_releases)
2836 err = __prepare_send_request(session, req, drop_cap_releases);
2838 ceph_msg_get(req->r_request);
2839 ceph_con_send(&session->s_con, req->r_request);
2846 * send request, or put it on the appropriate wait list.
2848 static void __do_request(struct ceph_mds_client *mdsc,
2849 struct ceph_mds_request *req)
2851 struct ceph_mds_session *session = NULL;
2856 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2857 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2858 __unregister_request(mdsc, req);
2862 if (req->r_timeout &&
2863 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2864 dout("do_request timed out\n");
2868 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2869 dout("do_request forced umount\n");
2873 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2874 if (mdsc->mdsmap_err) {
2875 err = mdsc->mdsmap_err;
2876 dout("do_request mdsmap err %d\n", err);
2879 if (mdsc->mdsmap->m_epoch == 0) {
2880 dout("do_request no mdsmap, waiting for map\n");
2881 list_add(&req->r_wait, &mdsc->waiting_for_map);
2884 if (!(mdsc->fsc->mount_options->flags &
2885 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2886 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2887 err = -EHOSTUNREACH;
2892 put_request_session(req);
2894 mds = __choose_mds(mdsc, req, &random);
2896 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2897 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2901 dout("do_request no mds or not active, waiting for map\n");
2902 list_add(&req->r_wait, &mdsc->waiting_for_map);
2906 /* get, open session */
2907 session = __ceph_lookup_mds_session(mdsc, mds);
2909 session = register_session(mdsc, mds);
2910 if (IS_ERR(session)) {
2911 err = PTR_ERR(session);
2915 req->r_session = ceph_get_mds_session(session);
2917 dout("do_request mds%d session %p state %s\n", mds, session,
2918 ceph_session_state_name(session->s_state));
2919 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2920 session->s_state != CEPH_MDS_SESSION_HUNG) {
2922 * We cannot queue async requests since the caps and delegated
2923 * inodes are bound to the session. Just return -EJUKEBOX and
2924 * let the caller retry a sync request in that case.
2926 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2932 * If the session has been REJECTED, then return a hard error,
2933 * unless it's a CLEANRECOVER mount, in which case we'll queue
2934 * it to the mdsc queue.
2936 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2937 if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
2938 list_add(&req->r_wait, &mdsc->waiting_for_map);
2944 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2945 session->s_state == CEPH_MDS_SESSION_CLOSING) {
2946 err = __open_session(mdsc, session);
2949 /* retry the same mds later */
2951 req->r_resend_mds = mds;
2953 list_add(&req->r_wait, &session->s_waiting);
2958 req->r_resend_mds = -1; /* forget any previous mds hint */
2960 if (req->r_request_started == 0) /* note request start time */
2961 req->r_request_started = jiffies;
2964 * For async create we will choose the auth MDS of frag in parent
2965 * directory to send the request and ususally this works fine, but
2966 * if the migrated the dirtory to another MDS before it could handle
2967 * it the request will be forwarded.
2969 * And then the auth cap will be changed.
2971 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags) && req->r_num_fwd) {
2972 struct ceph_dentry_info *di = ceph_dentry(req->r_dentry);
2973 struct ceph_inode_info *ci;
2974 struct ceph_cap *cap;
2977 * The request maybe handled very fast and the new inode
2978 * hasn't been linked to the dentry yet. We need to wait
2979 * for the ceph_finish_async_create(), which shouldn't be
2980 * stuck too long or fail in thoery, to finish when forwarding
2983 if (!d_inode(req->r_dentry)) {
2984 err = wait_on_bit(&di->flags, CEPH_DENTRY_ASYNC_CREATE_BIT,
2987 mutex_lock(&req->r_fill_mutex);
2988 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
2989 mutex_unlock(&req->r_fill_mutex);
2994 ci = ceph_inode(d_inode(req->r_dentry));
2996 spin_lock(&ci->i_ceph_lock);
2997 cap = ci->i_auth_cap;
2998 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE && mds != cap->mds) {
2999 dout("do_request session changed for auth cap %d -> %d\n",
3000 cap->session->s_mds, session->s_mds);
3002 /* Remove the auth cap from old session */
3003 spin_lock(&cap->session->s_cap_lock);
3004 cap->session->s_nr_caps--;
3005 list_del_init(&cap->session_caps);
3006 spin_unlock(&cap->session->s_cap_lock);
3008 /* Add the auth cap to the new session */
3010 cap->session = session;
3011 spin_lock(&session->s_cap_lock);
3012 session->s_nr_caps++;
3013 list_add_tail(&cap->session_caps, &session->s_caps);
3014 spin_unlock(&session->s_cap_lock);
3016 change_auth_cap_ses(ci, session);
3018 spin_unlock(&ci->i_ceph_lock);
3021 err = __send_request(session, req, false);
3024 ceph_put_mds_session(session);
3027 dout("__do_request early error %d\n", err);
3029 complete_request(mdsc, req);
3030 __unregister_request(mdsc, req);
3036 * called under mdsc->mutex
3038 static void __wake_requests(struct ceph_mds_client *mdsc,
3039 struct list_head *head)
3041 struct ceph_mds_request *req;
3042 LIST_HEAD(tmp_list);
3044 list_splice_init(head, &tmp_list);
3046 while (!list_empty(&tmp_list)) {
3047 req = list_entry(tmp_list.next,
3048 struct ceph_mds_request, r_wait);
3049 list_del_init(&req->r_wait);
3050 dout(" wake request %p tid %llu\n", req, req->r_tid);
3051 __do_request(mdsc, req);
3056 * Wake up threads with requests pending for @mds, so that they can
3057 * resubmit their requests to a possibly different mds.
3059 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
3061 struct ceph_mds_request *req;
3062 struct rb_node *p = rb_first(&mdsc->request_tree);
3064 dout("kick_requests mds%d\n", mds);
3066 req = rb_entry(p, struct ceph_mds_request, r_node);
3068 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3070 if (req->r_attempts > 0)
3071 continue; /* only new requests */
3072 if (req->r_session &&
3073 req->r_session->s_mds == mds) {
3074 dout(" kicking tid %llu\n", req->r_tid);
3075 list_del_init(&req->r_wait);
3076 __do_request(mdsc, req);
3081 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
3082 struct ceph_mds_request *req)
3086 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
3088 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
3089 if (req->r_parent) {
3090 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
3091 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
3092 CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
3093 spin_lock(&ci->i_ceph_lock);
3094 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
3095 __ceph_touch_fmode(ci, mdsc, fmode);
3096 spin_unlock(&ci->i_ceph_lock);
3098 if (req->r_old_dentry_dir)
3099 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
3103 err = ceph_wait_on_async_create(req->r_inode);
3105 dout("%s: wait for async create returned: %d\n",
3111 if (!err && req->r_old_inode) {
3112 err = ceph_wait_on_async_create(req->r_old_inode);
3114 dout("%s: wait for async create returned: %d\n",
3120 dout("submit_request on %p for inode %p\n", req, dir);
3121 mutex_lock(&mdsc->mutex);
3122 __register_request(mdsc, req, dir);
3123 __do_request(mdsc, req);
3125 mutex_unlock(&mdsc->mutex);
3129 int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
3130 struct ceph_mds_request *req,
3131 ceph_mds_request_wait_callback_t wait_func)
3136 dout("do_request waiting\n");
3138 err = wait_func(mdsc, req);
3140 long timeleft = wait_for_completion_killable_timeout(
3142 ceph_timeout_jiffies(req->r_timeout));
3146 err = -ETIMEDOUT; /* timed out */
3148 err = timeleft; /* killed */
3150 dout("do_request waited, got %d\n", err);
3151 mutex_lock(&mdsc->mutex);
3153 /* only abort if we didn't race with a real reply */
3154 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3155 err = le32_to_cpu(req->r_reply_info.head->result);
3156 } else if (err < 0) {
3157 dout("aborted request %lld with %d\n", req->r_tid, err);
3160 * ensure we aren't running concurrently with
3161 * ceph_fill_trace or ceph_readdir_prepopulate, which
3162 * rely on locks (dir mutex) held by our caller.
3164 mutex_lock(&req->r_fill_mutex);
3166 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3167 mutex_unlock(&req->r_fill_mutex);
3169 if (req->r_parent &&
3170 (req->r_op & CEPH_MDS_OP_WRITE))
3171 ceph_invalidate_dir_request(req);
3176 mutex_unlock(&mdsc->mutex);
3181 * Synchrously perform an mds request. Take care of all of the
3182 * session setup, forwarding, retry details.
3184 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3186 struct ceph_mds_request *req)
3190 dout("do_request on %p\n", req);
3193 err = ceph_mdsc_submit_request(mdsc, dir, req);
3195 err = ceph_mdsc_wait_request(mdsc, req, NULL);
3196 dout("do_request %p done, result %d\n", req, err);
3201 * Invalidate dir's completeness, dentry lease state on an aborted MDS
3202 * namespace request.
3204 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3206 struct inode *dir = req->r_parent;
3207 struct inode *old_dir = req->r_old_dentry_dir;
3209 dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
3211 ceph_dir_clear_complete(dir);
3213 ceph_dir_clear_complete(old_dir);
3215 ceph_invalidate_dentry_lease(req->r_dentry);
3216 if (req->r_old_dentry)
3217 ceph_invalidate_dentry_lease(req->r_old_dentry);
3223 * We take the session mutex and parse and process the reply immediately.
3224 * This preserves the logical ordering of replies, capabilities, etc., sent
3225 * by the MDS as they are applied to our local cache.
3227 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3229 struct ceph_mds_client *mdsc = session->s_mdsc;
3230 struct ceph_mds_request *req;
3231 struct ceph_mds_reply_head *head = msg->front.iov_base;
3232 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
3233 struct ceph_snap_realm *realm;
3236 int mds = session->s_mds;
3238 if (msg->front.iov_len < sizeof(*head)) {
3239 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
3244 /* get request, session */
3245 tid = le64_to_cpu(msg->hdr.tid);
3246 mutex_lock(&mdsc->mutex);
3247 req = lookup_get_request(mdsc, tid);
3249 dout("handle_reply on unknown tid %llu\n", tid);
3250 mutex_unlock(&mdsc->mutex);
3253 dout("handle_reply %p\n", req);
3255 /* correct session? */
3256 if (req->r_session != session) {
3257 pr_err("mdsc_handle_reply got %llu on session mds%d"
3258 " not mds%d\n", tid, session->s_mds,
3259 req->r_session ? req->r_session->s_mds : -1);
3260 mutex_unlock(&mdsc->mutex);
3265 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3266 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3267 pr_warn("got a dup %s reply on %llu from mds%d\n",
3268 head->safe ? "safe" : "unsafe", tid, mds);
3269 mutex_unlock(&mdsc->mutex);
3272 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3273 pr_warn("got unsafe after safe on %llu from mds%d\n",
3275 mutex_unlock(&mdsc->mutex);
3279 result = le32_to_cpu(head->result);
3282 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3283 __unregister_request(mdsc, req);
3285 /* last request during umount? */
3286 if (mdsc->stopping && !__get_oldest_req(mdsc))
3287 complete_all(&mdsc->safe_umount_waiters);
3289 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3291 * We already handled the unsafe response, now do the
3292 * cleanup. No need to examine the response; the MDS
3293 * doesn't include any result info in the safe
3294 * response. And even if it did, there is nothing
3295 * useful we could do with a revised return value.
3297 dout("got safe reply %llu, mds%d\n", tid, mds);
3299 mutex_unlock(&mdsc->mutex);
3303 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3304 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3307 dout("handle_reply tid %lld result %d\n", tid, result);
3308 rinfo = &req->r_reply_info;
3309 if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3310 err = parse_reply_info(session, msg, rinfo, (u64)-1);
3312 err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3313 mutex_unlock(&mdsc->mutex);
3315 /* Must find target inode outside of mutexes to avoid deadlocks */
3316 if ((err >= 0) && rinfo->head->is_target) {
3318 struct ceph_vino tvino = {
3319 .ino = le64_to_cpu(rinfo->targeti.in->ino),
3320 .snap = le64_to_cpu(rinfo->targeti.in->snapid)
3323 in = ceph_get_inode(mdsc->fsc->sb, tvino);
3326 mutex_lock(&session->s_mutex);
3329 req->r_target_inode = in;
3332 mutex_lock(&session->s_mutex);
3334 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3341 if (rinfo->snapblob_len) {
3342 down_write(&mdsc->snap_rwsem);
3343 ceph_update_snap_trace(mdsc, rinfo->snapblob,
3344 rinfo->snapblob + rinfo->snapblob_len,
3345 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3347 downgrade_write(&mdsc->snap_rwsem);
3349 down_read(&mdsc->snap_rwsem);
3352 /* insert trace into our cache */
3353 mutex_lock(&req->r_fill_mutex);
3354 current->journal_info = req;
3355 err = ceph_fill_trace(mdsc->fsc->sb, req);
3357 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3358 req->r_op == CEPH_MDS_OP_LSSNAP))
3359 ceph_readdir_prepopulate(req, req->r_session);
3361 current->journal_info = NULL;
3362 mutex_unlock(&req->r_fill_mutex);
3364 up_read(&mdsc->snap_rwsem);
3366 ceph_put_snap_realm(mdsc, realm);
3369 if (req->r_target_inode &&
3370 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3371 struct ceph_inode_info *ci =
3372 ceph_inode(req->r_target_inode);
3373 spin_lock(&ci->i_unsafe_lock);
3374 list_add_tail(&req->r_unsafe_target_item,
3375 &ci->i_unsafe_iops);
3376 spin_unlock(&ci->i_unsafe_lock);
3379 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3382 mutex_lock(&mdsc->mutex);
3383 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3387 req->r_reply = ceph_msg_get(msg);
3388 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3391 dout("reply arrived after request %lld was aborted\n", tid);
3393 mutex_unlock(&mdsc->mutex);
3395 mutex_unlock(&session->s_mutex);
3397 /* kick calling process */
3398 complete_request(mdsc, req);
3400 ceph_update_metadata_metrics(&mdsc->metric, req->r_start_latency,
3401 req->r_end_latency, err);
3403 ceph_mdsc_put_request(req);
3410 * handle mds notification that our request has been forwarded.
3412 static void handle_forward(struct ceph_mds_client *mdsc,
3413 struct ceph_mds_session *session,
3414 struct ceph_msg *msg)
3416 struct ceph_mds_request *req;
3417 u64 tid = le64_to_cpu(msg->hdr.tid);
3421 void *p = msg->front.iov_base;
3422 void *end = p + msg->front.iov_len;
3423 bool aborted = false;
3425 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3426 next_mds = ceph_decode_32(&p);
3427 fwd_seq = ceph_decode_32(&p);
3429 mutex_lock(&mdsc->mutex);
3430 req = lookup_get_request(mdsc, tid);
3432 mutex_unlock(&mdsc->mutex);
3433 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3434 return; /* dup reply? */
3437 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3438 dout("forward tid %llu aborted, unregistering\n", tid);
3439 __unregister_request(mdsc, req);
3440 } else if (fwd_seq <= req->r_num_fwd) {
3442 * The type of 'num_fwd' in ceph 'MClientRequestForward'
3443 * is 'int32_t', while in 'ceph_mds_request_head' the
3444 * type is '__u8'. So in case the request bounces between
3445 * MDSes exceeding 256 times, the client will get stuck.
3447 * In this case it's ususally a bug in MDS and continue
3448 * bouncing the request makes no sense.
3450 * In future this could be fixed in ceph code, so avoid
3451 * using the hardcode here.
3453 int max = sizeof_field(struct ceph_mds_request_head, num_fwd);
3454 max = 1 << (max * BITS_PER_BYTE);
3455 if (req->r_num_fwd >= max) {
3456 mutex_lock(&req->r_fill_mutex);
3457 req->r_err = -EMULTIHOP;
3458 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3459 mutex_unlock(&req->r_fill_mutex);
3461 pr_warn_ratelimited("forward tid %llu seq overflow\n",
3464 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3465 tid, next_mds, req->r_num_fwd, fwd_seq);
3468 /* resend. forward race not possible; mds would drop */
3469 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3471 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3472 req->r_attempts = 0;
3473 req->r_num_fwd = fwd_seq;
3474 req->r_resend_mds = next_mds;
3475 put_request_session(req);
3476 __do_request(mdsc, req);
3478 mutex_unlock(&mdsc->mutex);
3480 /* kick calling process */
3482 complete_request(mdsc, req);
3483 ceph_mdsc_put_request(req);
3487 pr_err("mdsc_handle_forward decode error err=%d\n", err);
3490 static int __decode_session_metadata(void **p, void *end,
3493 /* map<string,string> */
3496 ceph_decode_32_safe(p, end, n, bad);
3499 ceph_decode_32_safe(p, end, len, bad);
3500 ceph_decode_need(p, end, len, bad);
3501 err_str = !strncmp(*p, "error_string", len);
3503 ceph_decode_32_safe(p, end, len, bad);
3504 ceph_decode_need(p, end, len, bad);
3506 * Match "blocklisted (blacklisted)" from newer MDSes,
3507 * or "blacklisted" from older MDSes.
3509 if (err_str && strnstr(*p, "blacklisted", len))
3510 *blocklisted = true;
3519 * handle a mds session control message
3521 static void handle_session(struct ceph_mds_session *session,
3522 struct ceph_msg *msg)
3524 struct ceph_mds_client *mdsc = session->s_mdsc;
3525 int mds = session->s_mds;
3526 int msg_version = le16_to_cpu(msg->hdr.version);
3527 void *p = msg->front.iov_base;
3528 void *end = p + msg->front.iov_len;
3529 struct ceph_mds_session_head *h;
3531 u64 seq, features = 0;
3533 bool blocklisted = false;
3536 ceph_decode_need(&p, end, sizeof(*h), bad);
3540 op = le32_to_cpu(h->op);
3541 seq = le64_to_cpu(h->seq);
3543 if (msg_version >= 3) {
3545 /* version >= 2 and < 5, decode metadata, skip otherwise
3546 * as it's handled via flags.
3548 if (msg_version >= 5)
3549 ceph_decode_skip_map(&p, end, string, string, bad);
3550 else if (__decode_session_metadata(&p, end, &blocklisted) < 0)
3553 /* version >= 3, feature bits */
3554 ceph_decode_32_safe(&p, end, len, bad);
3556 ceph_decode_64_safe(&p, end, features, bad);
3557 p += len - sizeof(features);
3561 if (msg_version >= 5) {
3565 ceph_decode_skip_16(&p, end, bad); /* struct_v, struct_cv */
3566 ceph_decode_32_safe(&p, end, len, bad); /* len */
3567 ceph_decode_skip_n(&p, end, len, bad); /* metric_spec */
3569 /* version >= 5, flags */
3570 ceph_decode_32_safe(&p, end, flags, bad);
3571 if (flags & CEPH_SESSION_BLOCKLISTED) {
3572 pr_warn("mds%d session blocklisted\n", session->s_mds);
3577 mutex_lock(&mdsc->mutex);
3578 if (op == CEPH_SESSION_CLOSE) {
3579 ceph_get_mds_session(session);
3580 __unregister_session(mdsc, session);
3582 /* FIXME: this ttl calculation is generous */
3583 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3584 mutex_unlock(&mdsc->mutex);
3586 mutex_lock(&session->s_mutex);
3588 dout("handle_session mds%d %s %p state %s seq %llu\n",
3589 mds, ceph_session_op_name(op), session,
3590 ceph_session_state_name(session->s_state), seq);
3592 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3593 session->s_state = CEPH_MDS_SESSION_OPEN;
3594 pr_info("mds%d came back\n", session->s_mds);
3598 case CEPH_SESSION_OPEN:
3599 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3600 pr_info("mds%d reconnect success\n", session->s_mds);
3602 if (session->s_state == CEPH_MDS_SESSION_OPEN) {
3603 pr_notice("mds%d is already opened\n", session->s_mds);
3605 session->s_state = CEPH_MDS_SESSION_OPEN;
3606 session->s_features = features;
3607 renewed_caps(mdsc, session, 0);
3608 if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT,
3609 &session->s_features))
3610 metric_schedule_delayed(&mdsc->metric);
3614 * The connection maybe broken and the session in client
3615 * side has been reinitialized, need to update the seq
3618 if (!session->s_seq && seq)
3619 session->s_seq = seq;
3623 __close_session(mdsc, session);
3626 case CEPH_SESSION_RENEWCAPS:
3627 if (session->s_renew_seq == seq)
3628 renewed_caps(mdsc, session, 1);
3631 case CEPH_SESSION_CLOSE:
3632 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3633 pr_info("mds%d reconnect denied\n", session->s_mds);
3634 session->s_state = CEPH_MDS_SESSION_CLOSED;
3635 cleanup_session_requests(mdsc, session);
3636 remove_session_caps(session);
3637 wake = 2; /* for good measure */
3638 wake_up_all(&mdsc->session_close_wq);
3641 case CEPH_SESSION_STALE:
3642 pr_info("mds%d caps went stale, renewing\n",
3644 atomic_inc(&session->s_cap_gen);
3645 session->s_cap_ttl = jiffies - 1;
3646 send_renew_caps(mdsc, session);
3649 case CEPH_SESSION_RECALL_STATE:
3650 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3653 case CEPH_SESSION_FLUSHMSG:
3654 send_flushmsg_ack(mdsc, session, seq);
3657 case CEPH_SESSION_FORCE_RO:
3658 dout("force_session_readonly %p\n", session);
3659 spin_lock(&session->s_cap_lock);
3660 session->s_readonly = true;
3661 spin_unlock(&session->s_cap_lock);
3662 wake_up_session_caps(session, FORCE_RO);
3665 case CEPH_SESSION_REJECT:
3666 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3667 pr_info("mds%d rejected session\n", session->s_mds);
3668 session->s_state = CEPH_MDS_SESSION_REJECTED;
3669 cleanup_session_requests(mdsc, session);
3670 remove_session_caps(session);
3672 mdsc->fsc->blocklisted = true;
3673 wake = 2; /* for good measure */
3677 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3681 mutex_unlock(&session->s_mutex);
3683 mutex_lock(&mdsc->mutex);
3684 __wake_requests(mdsc, &session->s_waiting);
3686 kick_requests(mdsc, mds);
3687 mutex_unlock(&mdsc->mutex);
3689 if (op == CEPH_SESSION_CLOSE)
3690 ceph_put_mds_session(session);
3694 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3695 (int)msg->front.iov_len);
3700 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3704 dcaps = xchg(&req->r_dir_caps, 0);
3706 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3707 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3711 void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3715 dcaps = xchg(&req->r_dir_caps, 0);
3717 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3718 ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3724 * called under session->mutex.
3726 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3727 struct ceph_mds_session *session)
3729 struct ceph_mds_request *req, *nreq;
3732 dout("replay_unsafe_requests mds%d\n", session->s_mds);
3734 mutex_lock(&mdsc->mutex);
3735 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3736 __send_request(session, req, true);
3739 * also re-send old requests when MDS enters reconnect stage. So that MDS
3740 * can process completed request in clientreplay stage.
3742 p = rb_first(&mdsc->request_tree);
3744 req = rb_entry(p, struct ceph_mds_request, r_node);
3746 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3748 if (req->r_attempts == 0)
3749 continue; /* only old requests */
3750 if (!req->r_session)
3752 if (req->r_session->s_mds != session->s_mds)
3755 ceph_mdsc_release_dir_caps_no_check(req);
3757 __send_request(session, req, true);
3759 mutex_unlock(&mdsc->mutex);
3762 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3764 struct ceph_msg *reply;
3765 struct ceph_pagelist *_pagelist;
3770 if (!recon_state->allow_multi)
3773 /* can't handle message that contains both caps and realm */
3774 BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3776 /* pre-allocate new pagelist */
3777 _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3781 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3785 /* placeholder for nr_caps */
3786 err = ceph_pagelist_encode_32(_pagelist, 0);
3790 if (recon_state->nr_caps) {
3791 /* currently encoding caps */
3792 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3796 /* placeholder for nr_realms (currently encoding relams) */
3797 err = ceph_pagelist_encode_32(_pagelist, 0);
3802 err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3806 page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3807 addr = kmap_atomic(page);
3808 if (recon_state->nr_caps) {
3809 /* currently encoding caps */
3810 *addr = cpu_to_le32(recon_state->nr_caps);
3812 /* currently encoding relams */
3813 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3815 kunmap_atomic(addr);
3817 reply->hdr.version = cpu_to_le16(5);
3818 reply->hdr.compat_version = cpu_to_le16(4);
3820 reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3821 ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3823 ceph_con_send(&recon_state->session->s_con, reply);
3824 ceph_pagelist_release(recon_state->pagelist);
3826 recon_state->pagelist = _pagelist;
3827 recon_state->nr_caps = 0;
3828 recon_state->nr_realms = 0;
3829 recon_state->msg_version = 5;
3832 ceph_msg_put(reply);
3834 ceph_pagelist_release(_pagelist);
3838 static struct dentry* d_find_primary(struct inode *inode)
3840 struct dentry *alias, *dn = NULL;
3842 if (hlist_empty(&inode->i_dentry))
3845 spin_lock(&inode->i_lock);
3846 if (hlist_empty(&inode->i_dentry))
3849 if (S_ISDIR(inode->i_mode)) {
3850 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
3851 if (!IS_ROOT(alias))
3856 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
3857 spin_lock(&alias->d_lock);
3858 if (!d_unhashed(alias) &&
3859 (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
3860 dn = dget_dlock(alias);
3862 spin_unlock(&alias->d_lock);
3867 spin_unlock(&inode->i_lock);
3872 * Encode information about a cap for a reconnect with the MDS.
3874 static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3878 struct ceph_mds_cap_reconnect v2;
3879 struct ceph_mds_cap_reconnect_v1 v1;
3881 struct ceph_inode_info *ci = cap->ci;
3882 struct ceph_reconnect_state *recon_state = arg;
3883 struct ceph_pagelist *pagelist = recon_state->pagelist;
3884 struct dentry *dentry;
3886 int pathlen = 0, err;
3890 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3891 inode, ceph_vinop(inode), cap, cap->cap_id,
3892 ceph_cap_string(cap->issued));
3894 dentry = d_find_primary(inode);
3896 /* set pathbase to parent dir when msg_version >= 2 */
3897 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase,
3898 recon_state->msg_version >= 2);
3901 err = PTR_ERR(path);
3909 spin_lock(&ci->i_ceph_lock);
3910 cap->seq = 0; /* reset cap seq */
3911 cap->issue_seq = 0; /* and issue_seq */
3912 cap->mseq = 0; /* and migrate_seq */
3913 cap->cap_gen = atomic_read(&cap->session->s_cap_gen);
3915 /* These are lost when the session goes away */
3916 if (S_ISDIR(inode->i_mode)) {
3917 if (cap->issued & CEPH_CAP_DIR_CREATE) {
3918 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3919 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3921 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3924 if (recon_state->msg_version >= 2) {
3925 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3926 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3927 rec.v2.issued = cpu_to_le32(cap->issued);
3928 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3929 rec.v2.pathbase = cpu_to_le64(pathbase);
3930 rec.v2.flock_len = (__force __le32)
3931 ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3933 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3934 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3935 rec.v1.issued = cpu_to_le32(cap->issued);
3936 rec.v1.size = cpu_to_le64(i_size_read(inode));
3937 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3938 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3939 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3940 rec.v1.pathbase = cpu_to_le64(pathbase);
3943 if (list_empty(&ci->i_cap_snaps)) {
3944 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3946 struct ceph_cap_snap *capsnap =
3947 list_first_entry(&ci->i_cap_snaps,
3948 struct ceph_cap_snap, ci_item);
3949 snap_follows = capsnap->follows;
3951 spin_unlock(&ci->i_ceph_lock);
3953 if (recon_state->msg_version >= 2) {
3954 int num_fcntl_locks, num_flock_locks;
3955 struct ceph_filelock *flocks = NULL;
3956 size_t struct_len, total_len = sizeof(u64);
3960 if (rec.v2.flock_len) {
3961 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3963 num_fcntl_locks = 0;
3964 num_flock_locks = 0;
3966 if (num_fcntl_locks + num_flock_locks > 0) {
3967 flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3968 sizeof(struct ceph_filelock),
3974 err = ceph_encode_locks_to_buffer(inode, flocks,
3989 if (recon_state->msg_version >= 3) {
3990 /* version, compat_version and struct_len */
3991 total_len += 2 * sizeof(u8) + sizeof(u32);
3995 * number of encoded locks is stable, so copy to pagelist
3997 struct_len = 2 * sizeof(u32) +
3998 (num_fcntl_locks + num_flock_locks) *
3999 sizeof(struct ceph_filelock);
4000 rec.v2.flock_len = cpu_to_le32(struct_len);
4002 struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
4005 struct_len += sizeof(u64); /* snap_follows */
4007 total_len += struct_len;
4009 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
4010 err = send_reconnect_partial(recon_state);
4012 goto out_freeflocks;
4013 pagelist = recon_state->pagelist;
4016 err = ceph_pagelist_reserve(pagelist, total_len);
4018 goto out_freeflocks;
4020 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
4021 if (recon_state->msg_version >= 3) {
4022 ceph_pagelist_encode_8(pagelist, struct_v);
4023 ceph_pagelist_encode_8(pagelist, 1);
4024 ceph_pagelist_encode_32(pagelist, struct_len);
4026 ceph_pagelist_encode_string(pagelist, path, pathlen);
4027 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
4028 ceph_locks_to_pagelist(flocks, pagelist,
4029 num_fcntl_locks, num_flock_locks);
4031 ceph_pagelist_encode_64(pagelist, snap_follows);
4035 err = ceph_pagelist_reserve(pagelist,
4036 sizeof(u64) + sizeof(u32) +
4037 pathlen + sizeof(rec.v1));
4041 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
4042 ceph_pagelist_encode_string(pagelist, path, pathlen);
4043 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
4047 ceph_mdsc_free_path(path, pathlen);
4049 recon_state->nr_caps++;
4053 static int encode_snap_realms(struct ceph_mds_client *mdsc,
4054 struct ceph_reconnect_state *recon_state)
4057 struct ceph_pagelist *pagelist = recon_state->pagelist;
4060 if (recon_state->msg_version >= 4) {
4061 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
4067 * snaprealms. we provide mds with the ino, seq (version), and
4068 * parent for all of our realms. If the mds has any newer info,
4071 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
4072 struct ceph_snap_realm *realm =
4073 rb_entry(p, struct ceph_snap_realm, node);
4074 struct ceph_mds_snaprealm_reconnect sr_rec;
4076 if (recon_state->msg_version >= 4) {
4077 size_t need = sizeof(u8) * 2 + sizeof(u32) +
4080 if (pagelist->length + need > RECONNECT_MAX_SIZE) {
4081 err = send_reconnect_partial(recon_state);
4084 pagelist = recon_state->pagelist;
4087 err = ceph_pagelist_reserve(pagelist, need);
4091 ceph_pagelist_encode_8(pagelist, 1);
4092 ceph_pagelist_encode_8(pagelist, 1);
4093 ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
4096 dout(" adding snap realm %llx seq %lld parent %llx\n",
4097 realm->ino, realm->seq, realm->parent_ino);
4098 sr_rec.ino = cpu_to_le64(realm->ino);
4099 sr_rec.seq = cpu_to_le64(realm->seq);
4100 sr_rec.parent = cpu_to_le64(realm->parent_ino);
4102 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
4106 recon_state->nr_realms++;
4114 * If an MDS fails and recovers, clients need to reconnect in order to
4115 * reestablish shared state. This includes all caps issued through
4116 * this session _and_ the snap_realm hierarchy. Because it's not
4117 * clear which snap realms the mds cares about, we send everything we
4118 * know about.. that ensures we'll then get any new info the
4119 * recovering MDS might have.
4121 * This is a relatively heavyweight operation, but it's rare.
4123 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
4124 struct ceph_mds_session *session)
4126 struct ceph_msg *reply;
4127 int mds = session->s_mds;
4129 struct ceph_reconnect_state recon_state = {
4134 pr_info("mds%d reconnect start\n", mds);
4136 recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
4137 if (!recon_state.pagelist)
4138 goto fail_nopagelist;
4140 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
4144 xa_destroy(&session->s_delegated_inos);
4146 mutex_lock(&session->s_mutex);
4147 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
4150 dout("session %p state %s\n", session,
4151 ceph_session_state_name(session->s_state));
4153 atomic_inc(&session->s_cap_gen);
4155 spin_lock(&session->s_cap_lock);
4156 /* don't know if session is readonly */
4157 session->s_readonly = 0;
4159 * notify __ceph_remove_cap() that we are composing cap reconnect.
4160 * If a cap get released before being added to the cap reconnect,
4161 * __ceph_remove_cap() should skip queuing cap release.
4163 session->s_cap_reconnect = 1;
4164 /* drop old cap expires; we're about to reestablish that state */
4165 detach_cap_releases(session, &dispose);
4166 spin_unlock(&session->s_cap_lock);
4167 dispose_cap_releases(mdsc, &dispose);
4169 /* trim unused caps to reduce MDS's cache rejoin time */
4170 if (mdsc->fsc->sb->s_root)
4171 shrink_dcache_parent(mdsc->fsc->sb->s_root);
4173 ceph_con_close(&session->s_con);
4174 ceph_con_open(&session->s_con,
4175 CEPH_ENTITY_TYPE_MDS, mds,
4176 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
4178 /* replay unsafe requests */
4179 replay_unsafe_requests(mdsc, session);
4181 ceph_early_kick_flushing_caps(mdsc, session);
4183 down_read(&mdsc->snap_rwsem);
4185 /* placeholder for nr_caps */
4186 err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
4190 if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4191 recon_state.msg_version = 3;
4192 recon_state.allow_multi = true;
4193 } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4194 recon_state.msg_version = 3;
4196 recon_state.msg_version = 2;
4198 /* trsaverse this session's caps */
4199 err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4201 spin_lock(&session->s_cap_lock);
4202 session->s_cap_reconnect = 0;
4203 spin_unlock(&session->s_cap_lock);
4208 /* check if all realms can be encoded into current message */
4209 if (mdsc->num_snap_realms) {
4211 recon_state.pagelist->length +
4212 mdsc->num_snap_realms *
4213 sizeof(struct ceph_mds_snaprealm_reconnect);
4214 if (recon_state.msg_version >= 4) {
4215 /* number of realms */
4216 total_len += sizeof(u32);
4217 /* version, compat_version and struct_len */
4218 total_len += mdsc->num_snap_realms *
4219 (2 * sizeof(u8) + sizeof(u32));
4221 if (total_len > RECONNECT_MAX_SIZE) {
4222 if (!recon_state.allow_multi) {
4226 if (recon_state.nr_caps) {
4227 err = send_reconnect_partial(&recon_state);
4231 recon_state.msg_version = 5;
4235 err = encode_snap_realms(mdsc, &recon_state);
4239 if (recon_state.msg_version >= 5) {
4240 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4245 if (recon_state.nr_caps || recon_state.nr_realms) {
4247 list_first_entry(&recon_state.pagelist->head,
4249 __le32 *addr = kmap_atomic(page);
4250 if (recon_state.nr_caps) {
4251 WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4252 *addr = cpu_to_le32(recon_state.nr_caps);
4253 } else if (recon_state.msg_version >= 4) {
4254 *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4256 kunmap_atomic(addr);
4259 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4260 if (recon_state.msg_version >= 4)
4261 reply->hdr.compat_version = cpu_to_le16(4);
4263 reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4264 ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4266 ceph_con_send(&session->s_con, reply);
4268 mutex_unlock(&session->s_mutex);
4270 mutex_lock(&mdsc->mutex);
4271 __wake_requests(mdsc, &session->s_waiting);
4272 mutex_unlock(&mdsc->mutex);
4274 up_read(&mdsc->snap_rwsem);
4275 ceph_pagelist_release(recon_state.pagelist);
4279 ceph_msg_put(reply);
4280 up_read(&mdsc->snap_rwsem);
4281 mutex_unlock(&session->s_mutex);
4283 ceph_pagelist_release(recon_state.pagelist);
4285 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
4291 * compare old and new mdsmaps, kicking requests
4292 * and closing out old connections as necessary
4294 * called under mdsc->mutex.
4296 static void check_new_map(struct ceph_mds_client *mdsc,
4297 struct ceph_mdsmap *newmap,
4298 struct ceph_mdsmap *oldmap)
4301 int oldstate, newstate;
4302 struct ceph_mds_session *s;
4303 unsigned long targets[DIV_ROUND_UP(CEPH_MAX_MDS, sizeof(unsigned long))] = {0};
4305 dout("check_new_map new %u old %u\n",
4306 newmap->m_epoch, oldmap->m_epoch);
4308 if (newmap->m_info) {
4309 for (i = 0; i < newmap->possible_max_rank; i++) {
4310 for (j = 0; j < newmap->m_info[i].num_export_targets; j++)
4311 set_bit(newmap->m_info[i].export_targets[j], targets);
4315 for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4316 if (!mdsc->sessions[i])
4318 s = mdsc->sessions[i];
4319 oldstate = ceph_mdsmap_get_state(oldmap, i);
4320 newstate = ceph_mdsmap_get_state(newmap, i);
4322 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
4323 i, ceph_mds_state_name(oldstate),
4324 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4325 ceph_mds_state_name(newstate),
4326 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4327 ceph_session_state_name(s->s_state));
4329 if (i >= newmap->possible_max_rank) {
4330 /* force close session for stopped mds */
4331 ceph_get_mds_session(s);
4332 __unregister_session(mdsc, s);
4333 __wake_requests(mdsc, &s->s_waiting);
4334 mutex_unlock(&mdsc->mutex);
4336 mutex_lock(&s->s_mutex);
4337 cleanup_session_requests(mdsc, s);
4338 remove_session_caps(s);
4339 mutex_unlock(&s->s_mutex);
4341 ceph_put_mds_session(s);
4343 mutex_lock(&mdsc->mutex);
4344 kick_requests(mdsc, i);
4348 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4349 ceph_mdsmap_get_addr(newmap, i),
4350 sizeof(struct ceph_entity_addr))) {
4352 mutex_unlock(&mdsc->mutex);
4353 mutex_lock(&s->s_mutex);
4354 mutex_lock(&mdsc->mutex);
4355 ceph_con_close(&s->s_con);
4356 mutex_unlock(&s->s_mutex);
4357 s->s_state = CEPH_MDS_SESSION_RESTARTING;
4358 } else if (oldstate == newstate) {
4359 continue; /* nothing new with this mds */
4365 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4366 newstate >= CEPH_MDS_STATE_RECONNECT) {
4367 mutex_unlock(&mdsc->mutex);
4368 clear_bit(i, targets);
4369 send_mds_reconnect(mdsc, s);
4370 mutex_lock(&mdsc->mutex);
4374 * kick request on any mds that has gone active.
4376 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4377 newstate >= CEPH_MDS_STATE_ACTIVE) {
4378 if (oldstate != CEPH_MDS_STATE_CREATING &&
4379 oldstate != CEPH_MDS_STATE_STARTING)
4380 pr_info("mds%d recovery completed\n", s->s_mds);
4381 kick_requests(mdsc, i);
4382 mutex_unlock(&mdsc->mutex);
4383 mutex_lock(&s->s_mutex);
4384 mutex_lock(&mdsc->mutex);
4385 ceph_kick_flushing_caps(mdsc, s);
4386 mutex_unlock(&s->s_mutex);
4387 wake_up_session_caps(s, RECONNECT);
4392 * Only open and reconnect sessions that don't exist yet.
4394 for (i = 0; i < newmap->possible_max_rank; i++) {
4396 * In case the import MDS is crashed just after
4397 * the EImportStart journal is flushed, so when
4398 * a standby MDS takes over it and is replaying
4399 * the EImportStart journal the new MDS daemon
4400 * will wait the client to reconnect it, but the
4401 * client may never register/open the session yet.
4403 * Will try to reconnect that MDS daemon if the
4404 * rank number is in the export targets array and
4405 * is the up:reconnect state.
4407 newstate = ceph_mdsmap_get_state(newmap, i);
4408 if (!test_bit(i, targets) || newstate != CEPH_MDS_STATE_RECONNECT)
4412 * The session maybe registered and opened by some
4413 * requests which were choosing random MDSes during
4414 * the mdsc->mutex's unlock/lock gap below in rare
4415 * case. But the related MDS daemon will just queue
4416 * that requests and be still waiting for the client's
4417 * reconnection request in up:reconnect state.
4419 s = __ceph_lookup_mds_session(mdsc, i);
4421 s = __open_export_target_session(mdsc, i);
4424 pr_err("failed to open export target session, err %d\n",
4429 dout("send reconnect to export target mds.%d\n", i);
4430 mutex_unlock(&mdsc->mutex);
4431 send_mds_reconnect(mdsc, s);
4432 ceph_put_mds_session(s);
4433 mutex_lock(&mdsc->mutex);
4436 for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4437 s = mdsc->sessions[i];
4440 if (!ceph_mdsmap_is_laggy(newmap, i))
4442 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4443 s->s_state == CEPH_MDS_SESSION_HUNG ||
4444 s->s_state == CEPH_MDS_SESSION_CLOSING) {
4445 dout(" connecting to export targets of laggy mds%d\n",
4447 __open_export_target_sessions(mdsc, s);
4459 * caller must hold session s_mutex, dentry->d_lock
4461 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4463 struct ceph_dentry_info *di = ceph_dentry(dentry);
4465 ceph_put_mds_session(di->lease_session);
4466 di->lease_session = NULL;
4469 static void handle_lease(struct ceph_mds_client *mdsc,
4470 struct ceph_mds_session *session,
4471 struct ceph_msg *msg)
4473 struct super_block *sb = mdsc->fsc->sb;
4474 struct inode *inode;
4475 struct dentry *parent, *dentry;
4476 struct ceph_dentry_info *di;
4477 int mds = session->s_mds;
4478 struct ceph_mds_lease *h = msg->front.iov_base;
4480 struct ceph_vino vino;
4484 dout("handle_lease from mds%d\n", mds);
4487 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4489 vino.ino = le64_to_cpu(h->ino);
4490 vino.snap = CEPH_NOSNAP;
4491 seq = le32_to_cpu(h->seq);
4492 dname.len = get_unaligned_le32(h + 1);
4493 if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4495 dname.name = (void *)(h + 1) + sizeof(u32);
4498 inode = ceph_find_inode(sb, vino);
4499 dout("handle_lease %s, ino %llx %p %.*s\n",
4500 ceph_lease_op_name(h->action), vino.ino, inode,
4501 dname.len, dname.name);
4503 mutex_lock(&session->s_mutex);
4504 inc_session_sequence(session);
4507 dout("handle_lease no inode %llx\n", vino.ino);
4512 parent = d_find_alias(inode);
4514 dout("no parent dentry on inode %p\n", inode);
4516 goto release; /* hrm... */
4518 dname.hash = full_name_hash(parent, dname.name, dname.len);
4519 dentry = d_lookup(parent, &dname);
4524 spin_lock(&dentry->d_lock);
4525 di = ceph_dentry(dentry);
4526 switch (h->action) {
4527 case CEPH_MDS_LEASE_REVOKE:
4528 if (di->lease_session == session) {
4529 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4530 h->seq = cpu_to_le32(di->lease_seq);
4531 __ceph_mdsc_drop_dentry_lease(dentry);
4536 case CEPH_MDS_LEASE_RENEW:
4537 if (di->lease_session == session &&
4538 di->lease_gen == atomic_read(&session->s_cap_gen) &&
4539 di->lease_renew_from &&
4540 di->lease_renew_after == 0) {
4541 unsigned long duration =
4542 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4544 di->lease_seq = seq;
4545 di->time = di->lease_renew_from + duration;
4546 di->lease_renew_after = di->lease_renew_from +
4548 di->lease_renew_from = 0;
4552 spin_unlock(&dentry->d_lock);
4559 /* let's just reuse the same message */
4560 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4562 ceph_con_send(&session->s_con, msg);
4565 mutex_unlock(&session->s_mutex);
4570 pr_err("corrupt lease message\n");
4574 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4575 struct dentry *dentry, char action,
4578 struct ceph_msg *msg;
4579 struct ceph_mds_lease *lease;
4581 int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4583 dout("lease_send_msg identry %p %s to mds%d\n",
4584 dentry, ceph_lease_op_name(action), session->s_mds);
4586 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4589 lease = msg->front.iov_base;
4590 lease->action = action;
4591 lease->seq = cpu_to_le32(seq);
4593 spin_lock(&dentry->d_lock);
4594 dir = d_inode(dentry->d_parent);
4595 lease->ino = cpu_to_le64(ceph_ino(dir));
4596 lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4598 put_unaligned_le32(dentry->d_name.len, lease + 1);
4599 memcpy((void *)(lease + 1) + 4,
4600 dentry->d_name.name, dentry->d_name.len);
4601 spin_unlock(&dentry->d_lock);
4603 ceph_con_send(&session->s_con, msg);
4607 * lock unlock the session, to wait ongoing session activities
4609 static void lock_unlock_session(struct ceph_mds_session *s)
4611 mutex_lock(&s->s_mutex);
4612 mutex_unlock(&s->s_mutex);
4615 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4617 struct ceph_fs_client *fsc = mdsc->fsc;
4619 if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4622 if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4625 if (!READ_ONCE(fsc->blocklisted))
4628 pr_info("auto reconnect after blocklisted\n");
4629 ceph_force_reconnect(fsc->sb);
4632 bool check_session_state(struct ceph_mds_session *s)
4634 switch (s->s_state) {
4635 case CEPH_MDS_SESSION_OPEN:
4636 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4637 s->s_state = CEPH_MDS_SESSION_HUNG;
4638 pr_info("mds%d hung\n", s->s_mds);
4641 case CEPH_MDS_SESSION_CLOSING:
4642 case CEPH_MDS_SESSION_NEW:
4643 case CEPH_MDS_SESSION_RESTARTING:
4644 case CEPH_MDS_SESSION_CLOSED:
4645 case CEPH_MDS_SESSION_REJECTED:
4653 * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
4654 * then we need to retransmit that request.
4656 void inc_session_sequence(struct ceph_mds_session *s)
4658 lockdep_assert_held(&s->s_mutex);
4662 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4665 dout("resending session close request for mds%d\n", s->s_mds);
4666 ret = request_close_session(s);
4668 pr_err("unable to close session to mds%d: %d\n",
4674 * delayed work -- periodically trim expired leases, renew caps with mds. If
4675 * the @delay parameter is set to 0 or if it's more than 5 secs, the default
4676 * workqueue delay value of 5 secs will be used.
4678 static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
4680 unsigned long max_delay = HZ * 5;
4682 /* 5 secs default delay */
4683 if (!delay || (delay > max_delay))
4685 schedule_delayed_work(&mdsc->delayed_work,
4686 round_jiffies_relative(delay));
4689 static void delayed_work(struct work_struct *work)
4691 struct ceph_mds_client *mdsc =
4692 container_of(work, struct ceph_mds_client, delayed_work.work);
4693 unsigned long delay;
4698 dout("mdsc delayed_work\n");
4703 mutex_lock(&mdsc->mutex);
4704 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4705 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4706 mdsc->last_renew_caps);
4708 mdsc->last_renew_caps = jiffies;
4710 for (i = 0; i < mdsc->max_sessions; i++) {
4711 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4715 if (!check_session_state(s)) {
4716 ceph_put_mds_session(s);
4719 mutex_unlock(&mdsc->mutex);
4721 mutex_lock(&s->s_mutex);
4723 send_renew_caps(mdsc, s);
4725 ceph_con_keepalive(&s->s_con);
4726 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4727 s->s_state == CEPH_MDS_SESSION_HUNG)
4728 ceph_send_cap_releases(mdsc, s);
4729 mutex_unlock(&s->s_mutex);
4730 ceph_put_mds_session(s);
4732 mutex_lock(&mdsc->mutex);
4734 mutex_unlock(&mdsc->mutex);
4736 delay = ceph_check_delayed_caps(mdsc);
4738 ceph_queue_cap_reclaim_work(mdsc);
4740 ceph_trim_snapid_map(mdsc);
4742 maybe_recover_session(mdsc);
4744 schedule_delayed(mdsc, delay);
4747 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4750 struct ceph_mds_client *mdsc;
4753 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4757 mutex_init(&mdsc->mutex);
4758 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4759 if (!mdsc->mdsmap) {
4764 init_completion(&mdsc->safe_umount_waiters);
4765 init_waitqueue_head(&mdsc->session_close_wq);
4766 INIT_LIST_HEAD(&mdsc->waiting_for_map);
4767 mdsc->quotarealms_inodes = RB_ROOT;
4768 mutex_init(&mdsc->quotarealms_inodes_mutex);
4769 init_rwsem(&mdsc->snap_rwsem);
4770 mdsc->snap_realms = RB_ROOT;
4771 INIT_LIST_HEAD(&mdsc->snap_empty);
4772 spin_lock_init(&mdsc->snap_empty_lock);
4773 mdsc->request_tree = RB_ROOT;
4774 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4775 mdsc->last_renew_caps = jiffies;
4776 INIT_LIST_HEAD(&mdsc->cap_delay_list);
4777 INIT_LIST_HEAD(&mdsc->cap_wait_list);
4778 spin_lock_init(&mdsc->cap_delay_lock);
4779 INIT_LIST_HEAD(&mdsc->snap_flush_list);
4780 spin_lock_init(&mdsc->snap_flush_lock);
4781 mdsc->last_cap_flush_tid = 1;
4782 INIT_LIST_HEAD(&mdsc->cap_flush_list);
4783 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4784 spin_lock_init(&mdsc->cap_dirty_lock);
4785 init_waitqueue_head(&mdsc->cap_flushing_wq);
4786 INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4787 err = ceph_metric_init(&mdsc->metric);
4791 spin_lock_init(&mdsc->dentry_list_lock);
4792 INIT_LIST_HEAD(&mdsc->dentry_leases);
4793 INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4795 ceph_caps_init(mdsc);
4796 ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4798 spin_lock_init(&mdsc->snapid_map_lock);
4799 mdsc->snapid_map_tree = RB_ROOT;
4800 INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4802 init_rwsem(&mdsc->pool_perm_rwsem);
4803 mdsc->pool_perm_tree = RB_ROOT;
4805 strscpy(mdsc->nodename, utsname()->nodename,
4806 sizeof(mdsc->nodename));
4812 kfree(mdsc->mdsmap);
4819 * Wait for safe replies on open mds requests. If we time out, drop
4820 * all requests from the tree to avoid dangling dentry refs.
4822 static void wait_requests(struct ceph_mds_client *mdsc)
4824 struct ceph_options *opts = mdsc->fsc->client->options;
4825 struct ceph_mds_request *req;
4827 mutex_lock(&mdsc->mutex);
4828 if (__get_oldest_req(mdsc)) {
4829 mutex_unlock(&mdsc->mutex);
4831 dout("wait_requests waiting for requests\n");
4832 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4833 ceph_timeout_jiffies(opts->mount_timeout));
4835 /* tear down remaining requests */
4836 mutex_lock(&mdsc->mutex);
4837 while ((req = __get_oldest_req(mdsc))) {
4838 dout("wait_requests timed out on tid %llu\n",
4840 list_del_init(&req->r_wait);
4841 __unregister_request(mdsc, req);
4844 mutex_unlock(&mdsc->mutex);
4845 dout("wait_requests done\n");
4848 void send_flush_mdlog(struct ceph_mds_session *s)
4850 struct ceph_msg *msg;
4853 * Pre-luminous MDS crashes when it sees an unknown session request
4855 if (!CEPH_HAVE_FEATURE(s->s_con.peer_features, SERVER_LUMINOUS))
4858 mutex_lock(&s->s_mutex);
4859 dout("request mdlog flush to mds%d (%s)s seq %lld\n", s->s_mds,
4860 ceph_session_state_name(s->s_state), s->s_seq);
4861 msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_FLUSH_MDLOG,
4864 pr_err("failed to request mdlog flush to mds%d (%s) seq %lld\n",
4865 s->s_mds, ceph_session_state_name(s->s_state), s->s_seq);
4867 ceph_con_send(&s->s_con, msg);
4869 mutex_unlock(&s->s_mutex);
4873 * called before mount is ro, and before dentries are torn down.
4874 * (hmm, does this still race with new lookups?)
4876 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4878 dout("pre_umount\n");
4881 ceph_mdsc_iterate_sessions(mdsc, send_flush_mdlog, true);
4882 ceph_mdsc_iterate_sessions(mdsc, lock_unlock_session, false);
4883 ceph_flush_dirty_caps(mdsc);
4884 wait_requests(mdsc);
4887 * wait for reply handlers to drop their request refs and
4888 * their inode/dcache refs
4892 ceph_cleanup_quotarealms_inodes(mdsc);
4896 * flush the mdlog and wait for all write mds requests to flush.
4898 static void flush_mdlog_and_wait_mdsc_unsafe_requests(struct ceph_mds_client *mdsc,
4901 struct ceph_mds_request *req = NULL, *nextreq;
4902 struct ceph_mds_session *last_session = NULL;
4905 mutex_lock(&mdsc->mutex);
4906 dout("%s want %lld\n", __func__, want_tid);
4908 req = __get_oldest_req(mdsc);
4909 while (req && req->r_tid <= want_tid) {
4910 /* find next request */
4911 n = rb_next(&req->r_node);
4913 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4916 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4917 (req->r_op & CEPH_MDS_OP_WRITE)) {
4918 struct ceph_mds_session *s = req->r_session;
4926 ceph_mdsc_get_request(req);
4928 ceph_mdsc_get_request(nextreq);
4929 s = ceph_get_mds_session(s);
4930 mutex_unlock(&mdsc->mutex);
4932 /* send flush mdlog request to MDS */
4933 if (last_session != s) {
4934 send_flush_mdlog(s);
4935 ceph_put_mds_session(last_session);
4938 ceph_put_mds_session(s);
4940 dout("%s wait on %llu (want %llu)\n", __func__,
4941 req->r_tid, want_tid);
4942 wait_for_completion(&req->r_safe_completion);
4944 mutex_lock(&mdsc->mutex);
4945 ceph_mdsc_put_request(req);
4947 break; /* next dne before, so we're done! */
4948 if (RB_EMPTY_NODE(&nextreq->r_node)) {
4949 /* next request was removed from tree */
4950 ceph_mdsc_put_request(nextreq);
4953 ceph_mdsc_put_request(nextreq); /* won't go away */
4957 mutex_unlock(&mdsc->mutex);
4958 ceph_put_mds_session(last_session);
4959 dout("%s done\n", __func__);
4962 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4964 u64 want_tid, want_flush;
4966 if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
4970 mutex_lock(&mdsc->mutex);
4971 want_tid = mdsc->last_tid;
4972 mutex_unlock(&mdsc->mutex);
4974 ceph_flush_dirty_caps(mdsc);
4975 spin_lock(&mdsc->cap_dirty_lock);
4976 want_flush = mdsc->last_cap_flush_tid;
4977 if (!list_empty(&mdsc->cap_flush_list)) {
4978 struct ceph_cap_flush *cf =
4979 list_last_entry(&mdsc->cap_flush_list,
4980 struct ceph_cap_flush, g_list);
4983 spin_unlock(&mdsc->cap_dirty_lock);
4985 dout("sync want tid %lld flush_seq %lld\n",
4986 want_tid, want_flush);
4988 flush_mdlog_and_wait_mdsc_unsafe_requests(mdsc, want_tid);
4989 wait_caps_flush(mdsc, want_flush);
4993 * true if all sessions are closed, or we force unmount
4995 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4997 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4999 return atomic_read(&mdsc->num_sessions) <= skipped;
5003 * called after sb is ro.
5005 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
5007 struct ceph_options *opts = mdsc->fsc->client->options;
5008 struct ceph_mds_session *session;
5012 dout("close_sessions\n");
5014 /* close sessions */
5015 mutex_lock(&mdsc->mutex);
5016 for (i = 0; i < mdsc->max_sessions; i++) {
5017 session = __ceph_lookup_mds_session(mdsc, i);
5020 mutex_unlock(&mdsc->mutex);
5021 mutex_lock(&session->s_mutex);
5022 if (__close_session(mdsc, session) <= 0)
5024 mutex_unlock(&session->s_mutex);
5025 ceph_put_mds_session(session);
5026 mutex_lock(&mdsc->mutex);
5028 mutex_unlock(&mdsc->mutex);
5030 dout("waiting for sessions to close\n");
5031 wait_event_timeout(mdsc->session_close_wq,
5032 done_closing_sessions(mdsc, skipped),
5033 ceph_timeout_jiffies(opts->mount_timeout));
5035 /* tear down remaining sessions */
5036 mutex_lock(&mdsc->mutex);
5037 for (i = 0; i < mdsc->max_sessions; i++) {
5038 if (mdsc->sessions[i]) {
5039 session = ceph_get_mds_session(mdsc->sessions[i]);
5040 __unregister_session(mdsc, session);
5041 mutex_unlock(&mdsc->mutex);
5042 mutex_lock(&session->s_mutex);
5043 remove_session_caps(session);
5044 mutex_unlock(&session->s_mutex);
5045 ceph_put_mds_session(session);
5046 mutex_lock(&mdsc->mutex);
5049 WARN_ON(!list_empty(&mdsc->cap_delay_list));
5050 mutex_unlock(&mdsc->mutex);
5052 ceph_cleanup_snapid_map(mdsc);
5053 ceph_cleanup_global_and_empty_realms(mdsc);
5055 cancel_work_sync(&mdsc->cap_reclaim_work);
5056 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
5061 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
5063 struct ceph_mds_session *session;
5066 dout("force umount\n");
5068 mutex_lock(&mdsc->mutex);
5069 for (mds = 0; mds < mdsc->max_sessions; mds++) {
5070 session = __ceph_lookup_mds_session(mdsc, mds);
5074 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
5075 __unregister_session(mdsc, session);
5076 __wake_requests(mdsc, &session->s_waiting);
5077 mutex_unlock(&mdsc->mutex);
5079 mutex_lock(&session->s_mutex);
5080 __close_session(mdsc, session);
5081 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
5082 cleanup_session_requests(mdsc, session);
5083 remove_session_caps(session);
5085 mutex_unlock(&session->s_mutex);
5086 ceph_put_mds_session(session);
5088 mutex_lock(&mdsc->mutex);
5089 kick_requests(mdsc, mds);
5091 __wake_requests(mdsc, &mdsc->waiting_for_map);
5092 mutex_unlock(&mdsc->mutex);
5095 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
5099 * Make sure the delayed work stopped before releasing
5102 * Because the cancel_delayed_work_sync() will only
5103 * guarantee that the work finishes executing. But the
5104 * delayed work will re-arm itself again after that.
5106 flush_delayed_work(&mdsc->delayed_work);
5109 ceph_mdsmap_destroy(mdsc->mdsmap);
5110 kfree(mdsc->sessions);
5111 ceph_caps_finalize(mdsc);
5112 ceph_pool_perm_destroy(mdsc);
5115 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
5117 struct ceph_mds_client *mdsc = fsc->mdsc;
5118 dout("mdsc_destroy %p\n", mdsc);
5123 /* flush out any connection work with references to us */
5126 ceph_mdsc_stop(mdsc);
5128 ceph_metric_destroy(&mdsc->metric);
5132 dout("mdsc_destroy %p done\n", mdsc);
5135 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5137 struct ceph_fs_client *fsc = mdsc->fsc;
5138 const char *mds_namespace = fsc->mount_options->mds_namespace;
5139 void *p = msg->front.iov_base;
5140 void *end = p + msg->front.iov_len;
5143 u32 mount_fscid = (u32)-1;
5146 ceph_decode_need(&p, end, sizeof(u32), bad);
5147 epoch = ceph_decode_32(&p);
5149 dout("handle_fsmap epoch %u\n", epoch);
5151 /* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
5152 ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
5154 ceph_decode_32_safe(&p, end, num_fs, bad);
5155 while (num_fs-- > 0) {
5156 void *info_p, *info_end;
5160 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
5161 p += 2; // info_v, info_cv
5162 info_len = ceph_decode_32(&p);
5163 ceph_decode_need(&p, end, info_len, bad);
5165 info_end = p + info_len;
5168 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
5169 fscid = ceph_decode_32(&info_p);
5170 namelen = ceph_decode_32(&info_p);
5171 ceph_decode_need(&info_p, info_end, namelen, bad);
5173 if (mds_namespace &&
5174 strlen(mds_namespace) == namelen &&
5175 !strncmp(mds_namespace, (char *)info_p, namelen)) {
5176 mount_fscid = fscid;
5181 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
5182 if (mount_fscid != (u32)-1) {
5183 fsc->client->monc.fs_cluster_id = mount_fscid;
5184 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
5186 ceph_monc_renew_subs(&fsc->client->monc);
5194 pr_err("error decoding fsmap %d. Shutting down mount.\n", err);
5195 ceph_umount_begin(mdsc->fsc->sb);
5197 mutex_lock(&mdsc->mutex);
5198 mdsc->mdsmap_err = err;
5199 __wake_requests(mdsc, &mdsc->waiting_for_map);
5200 mutex_unlock(&mdsc->mutex);
5204 * handle mds map update.
5206 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5210 void *p = msg->front.iov_base;
5211 void *end = p + msg->front.iov_len;
5212 struct ceph_mdsmap *newmap, *oldmap;
5213 struct ceph_fsid fsid;
5216 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
5217 ceph_decode_copy(&p, &fsid, sizeof(fsid));
5218 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
5220 epoch = ceph_decode_32(&p);
5221 maplen = ceph_decode_32(&p);
5222 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
5224 /* do we need it? */
5225 mutex_lock(&mdsc->mutex);
5226 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
5227 dout("handle_map epoch %u <= our %u\n",
5228 epoch, mdsc->mdsmap->m_epoch);
5229 mutex_unlock(&mdsc->mutex);
5233 newmap = ceph_mdsmap_decode(&p, end, ceph_msgr2(mdsc->fsc->client));
5234 if (IS_ERR(newmap)) {
5235 err = PTR_ERR(newmap);
5239 /* swap into place */
5241 oldmap = mdsc->mdsmap;
5242 mdsc->mdsmap = newmap;
5243 check_new_map(mdsc, newmap, oldmap);
5244 ceph_mdsmap_destroy(oldmap);
5246 mdsc->mdsmap = newmap; /* first mds map */
5248 mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
5251 __wake_requests(mdsc, &mdsc->waiting_for_map);
5252 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
5253 mdsc->mdsmap->m_epoch);
5255 mutex_unlock(&mdsc->mutex);
5256 schedule_delayed(mdsc, 0);
5260 mutex_unlock(&mdsc->mutex);
5262 pr_err("error decoding mdsmap %d. Shutting down mount.\n", err);
5263 ceph_umount_begin(mdsc->fsc->sb);
5267 static struct ceph_connection *mds_get_con(struct ceph_connection *con)
5269 struct ceph_mds_session *s = con->private;
5271 if (ceph_get_mds_session(s))
5276 static void mds_put_con(struct ceph_connection *con)
5278 struct ceph_mds_session *s = con->private;
5280 ceph_put_mds_session(s);
5284 * if the client is unresponsive for long enough, the mds will kill
5285 * the session entirely.
5287 static void mds_peer_reset(struct ceph_connection *con)
5289 struct ceph_mds_session *s = con->private;
5290 struct ceph_mds_client *mdsc = s->s_mdsc;
5292 pr_warn("mds%d closed our session\n", s->s_mds);
5293 send_mds_reconnect(mdsc, s);
5296 static void mds_dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5298 struct ceph_mds_session *s = con->private;
5299 struct ceph_mds_client *mdsc = s->s_mdsc;
5300 int type = le16_to_cpu(msg->hdr.type);
5302 mutex_lock(&mdsc->mutex);
5303 if (__verify_registered_session(mdsc, s) < 0) {
5304 mutex_unlock(&mdsc->mutex);
5307 mutex_unlock(&mdsc->mutex);
5310 case CEPH_MSG_MDS_MAP:
5311 ceph_mdsc_handle_mdsmap(mdsc, msg);
5313 case CEPH_MSG_FS_MAP_USER:
5314 ceph_mdsc_handle_fsmap(mdsc, msg);
5316 case CEPH_MSG_CLIENT_SESSION:
5317 handle_session(s, msg);
5319 case CEPH_MSG_CLIENT_REPLY:
5320 handle_reply(s, msg);
5322 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5323 handle_forward(mdsc, s, msg);
5325 case CEPH_MSG_CLIENT_CAPS:
5326 ceph_handle_caps(s, msg);
5328 case CEPH_MSG_CLIENT_SNAP:
5329 ceph_handle_snap(mdsc, s, msg);
5331 case CEPH_MSG_CLIENT_LEASE:
5332 handle_lease(mdsc, s, msg);
5334 case CEPH_MSG_CLIENT_QUOTA:
5335 ceph_handle_quota(mdsc, s, msg);
5339 pr_err("received unknown message type %d %s\n", type,
5340 ceph_msg_type_name(type));
5351 * Note: returned pointer is the address of a structure that's
5352 * managed separately. Caller must *not* attempt to free it.
5354 static struct ceph_auth_handshake *
5355 mds_get_authorizer(struct ceph_connection *con, int *proto, int force_new)
5357 struct ceph_mds_session *s = con->private;
5358 struct ceph_mds_client *mdsc = s->s_mdsc;
5359 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5360 struct ceph_auth_handshake *auth = &s->s_auth;
5363 ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5364 force_new, proto, NULL, NULL);
5366 return ERR_PTR(ret);
5371 static int mds_add_authorizer_challenge(struct ceph_connection *con,
5372 void *challenge_buf, int challenge_buf_len)
5374 struct ceph_mds_session *s = con->private;
5375 struct ceph_mds_client *mdsc = s->s_mdsc;
5376 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5378 return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
5379 challenge_buf, challenge_buf_len);
5382 static int mds_verify_authorizer_reply(struct ceph_connection *con)
5384 struct ceph_mds_session *s = con->private;
5385 struct ceph_mds_client *mdsc = s->s_mdsc;
5386 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5387 struct ceph_auth_handshake *auth = &s->s_auth;
5389 return ceph_auth_verify_authorizer_reply(ac, auth->authorizer,
5390 auth->authorizer_reply_buf, auth->authorizer_reply_buf_len,
5391 NULL, NULL, NULL, NULL);
5394 static int mds_invalidate_authorizer(struct ceph_connection *con)
5396 struct ceph_mds_session *s = con->private;
5397 struct ceph_mds_client *mdsc = s->s_mdsc;
5398 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5400 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5402 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5405 static int mds_get_auth_request(struct ceph_connection *con,
5406 void *buf, int *buf_len,
5407 void **authorizer, int *authorizer_len)
5409 struct ceph_mds_session *s = con->private;
5410 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5411 struct ceph_auth_handshake *auth = &s->s_auth;
5414 ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5419 *authorizer = auth->authorizer_buf;
5420 *authorizer_len = auth->authorizer_buf_len;
5424 static int mds_handle_auth_reply_more(struct ceph_connection *con,
5425 void *reply, int reply_len,
5426 void *buf, int *buf_len,
5427 void **authorizer, int *authorizer_len)
5429 struct ceph_mds_session *s = con->private;
5430 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5431 struct ceph_auth_handshake *auth = &s->s_auth;
5434 ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len,
5439 *authorizer = auth->authorizer_buf;
5440 *authorizer_len = auth->authorizer_buf_len;
5444 static int mds_handle_auth_done(struct ceph_connection *con,
5445 u64 global_id, void *reply, int reply_len,
5446 u8 *session_key, int *session_key_len,
5447 u8 *con_secret, int *con_secret_len)
5449 struct ceph_mds_session *s = con->private;
5450 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5451 struct ceph_auth_handshake *auth = &s->s_auth;
5453 return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len,
5454 session_key, session_key_len,
5455 con_secret, con_secret_len);
5458 static int mds_handle_auth_bad_method(struct ceph_connection *con,
5459 int used_proto, int result,
5460 const int *allowed_protos, int proto_cnt,
5461 const int *allowed_modes, int mode_cnt)
5463 struct ceph_mds_session *s = con->private;
5464 struct ceph_mon_client *monc = &s->s_mdsc->fsc->client->monc;
5467 if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_MDS,
5469 allowed_protos, proto_cnt,
5470 allowed_modes, mode_cnt)) {
5471 ret = ceph_monc_validate_auth(monc);
5479 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
5480 struct ceph_msg_header *hdr, int *skip)
5482 struct ceph_msg *msg;
5483 int type = (int) le16_to_cpu(hdr->type);
5484 int front_len = (int) le32_to_cpu(hdr->front_len);
5490 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
5492 pr_err("unable to allocate msg type %d len %d\n",
5500 static int mds_sign_message(struct ceph_msg *msg)
5502 struct ceph_mds_session *s = msg->con->private;
5503 struct ceph_auth_handshake *auth = &s->s_auth;
5505 return ceph_auth_sign_message(auth, msg);
5508 static int mds_check_message_signature(struct ceph_msg *msg)
5510 struct ceph_mds_session *s = msg->con->private;
5511 struct ceph_auth_handshake *auth = &s->s_auth;
5513 return ceph_auth_check_message_signature(auth, msg);
5516 static const struct ceph_connection_operations mds_con_ops = {
5519 .alloc_msg = mds_alloc_msg,
5520 .dispatch = mds_dispatch,
5521 .peer_reset = mds_peer_reset,
5522 .get_authorizer = mds_get_authorizer,
5523 .add_authorizer_challenge = mds_add_authorizer_challenge,
5524 .verify_authorizer_reply = mds_verify_authorizer_reply,
5525 .invalidate_authorizer = mds_invalidate_authorizer,
5526 .sign_message = mds_sign_message,
5527 .check_message_signature = mds_check_message_signature,
5528 .get_auth_request = mds_get_auth_request,
5529 .handle_auth_reply_more = mds_handle_auth_reply_more,
5530 .handle_auth_done = mds_handle_auth_done,
5531 .handle_auth_bad_method = mds_handle_auth_bad_method,