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50 * are no longer valid.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("ceph: 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));
662 const char *ceph_session_state_name(int s)
665 case CEPH_MDS_SESSION_NEW: return "new";
666 case CEPH_MDS_SESSION_OPENING: return "opening";
667 case CEPH_MDS_SESSION_OPEN: return "open";
668 case CEPH_MDS_SESSION_HUNG: return "hung";
669 case CEPH_MDS_SESSION_CLOSING: return "closing";
670 case CEPH_MDS_SESSION_CLOSED: return "closed";
671 case CEPH_MDS_SESSION_RESTARTING: return "restarting";
672 case CEPH_MDS_SESSION_RECONNECTING: return "reconnecting";
673 case CEPH_MDS_SESSION_REJECTED: return "rejected";
674 default: return "???";
678 struct ceph_mds_session *ceph_get_mds_session(struct ceph_mds_session *s)
680 if (refcount_inc_not_zero(&s->s_ref))
685 void ceph_put_mds_session(struct ceph_mds_session *s)
687 if (IS_ERR_OR_NULL(s))
690 if (refcount_dec_and_test(&s->s_ref)) {
691 if (s->s_auth.authorizer)
692 ceph_auth_destroy_authorizer(s->s_auth.authorizer);
693 WARN_ON(mutex_is_locked(&s->s_mutex));
694 xa_destroy(&s->s_delegated_inos);
700 * called under mdsc->mutex
702 struct ceph_mds_session *__ceph_lookup_mds_session(struct ceph_mds_client *mdsc,
705 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
707 return ceph_get_mds_session(mdsc->sessions[mds]);
710 static bool __have_session(struct ceph_mds_client *mdsc, int mds)
712 if (mds >= mdsc->max_sessions || !mdsc->sessions[mds])
718 static int __verify_registered_session(struct ceph_mds_client *mdsc,
719 struct ceph_mds_session *s)
721 if (s->s_mds >= mdsc->max_sessions ||
722 mdsc->sessions[s->s_mds] != s)
728 * create+register a new session for given mds.
729 * called under mdsc->mutex.
731 static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
734 struct ceph_mds_session *s;
736 if (mds >= mdsc->mdsmap->possible_max_rank)
737 return ERR_PTR(-EINVAL);
739 s = kzalloc(sizeof(*s), GFP_NOFS);
741 return ERR_PTR(-ENOMEM);
743 if (mds >= mdsc->max_sessions) {
744 int newmax = 1 << get_count_order(mds + 1);
745 struct ceph_mds_session **sa;
747 dout("%s: realloc to %d\n", __func__, newmax);
748 sa = kcalloc(newmax, sizeof(void *), GFP_NOFS);
751 if (mdsc->sessions) {
752 memcpy(sa, mdsc->sessions,
753 mdsc->max_sessions * sizeof(void *));
754 kfree(mdsc->sessions);
757 mdsc->max_sessions = newmax;
760 dout("%s: mds%d\n", __func__, mds);
763 s->s_state = CEPH_MDS_SESSION_NEW;
764 mutex_init(&s->s_mutex);
766 ceph_con_init(&s->s_con, s, &mds_con_ops, &mdsc->fsc->client->msgr);
768 atomic_set(&s->s_cap_gen, 1);
769 s->s_cap_ttl = jiffies - 1;
771 spin_lock_init(&s->s_cap_lock);
772 INIT_LIST_HEAD(&s->s_caps);
773 refcount_set(&s->s_ref, 1);
774 INIT_LIST_HEAD(&s->s_waiting);
775 INIT_LIST_HEAD(&s->s_unsafe);
776 xa_init(&s->s_delegated_inos);
777 INIT_LIST_HEAD(&s->s_cap_releases);
778 INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
780 INIT_LIST_HEAD(&s->s_cap_dirty);
781 INIT_LIST_HEAD(&s->s_cap_flushing);
783 mdsc->sessions[mds] = s;
784 atomic_inc(&mdsc->num_sessions);
785 refcount_inc(&s->s_ref); /* one ref to sessions[], one to caller */
787 ceph_con_open(&s->s_con, CEPH_ENTITY_TYPE_MDS, mds,
788 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
794 return ERR_PTR(-ENOMEM);
798 * called under mdsc->mutex
800 static void __unregister_session(struct ceph_mds_client *mdsc,
801 struct ceph_mds_session *s)
803 dout("__unregister_session mds%d %p\n", s->s_mds, s);
804 BUG_ON(mdsc->sessions[s->s_mds] != s);
805 mdsc->sessions[s->s_mds] = NULL;
806 ceph_con_close(&s->s_con);
807 ceph_put_mds_session(s);
808 atomic_dec(&mdsc->num_sessions);
812 * drop session refs in request.
814 * should be last request ref, or hold mdsc->mutex
816 static void put_request_session(struct ceph_mds_request *req)
818 if (req->r_session) {
819 ceph_put_mds_session(req->r_session);
820 req->r_session = NULL;
824 void ceph_mdsc_iterate_sessions(struct ceph_mds_client *mdsc,
825 void (*cb)(struct ceph_mds_session *),
830 mutex_lock(&mdsc->mutex);
831 for (mds = 0; mds < mdsc->max_sessions; ++mds) {
832 struct ceph_mds_session *s;
834 s = __ceph_lookup_mds_session(mdsc, mds);
838 if (check_state && !check_session_state(s)) {
839 ceph_put_mds_session(s);
843 mutex_unlock(&mdsc->mutex);
845 ceph_put_mds_session(s);
846 mutex_lock(&mdsc->mutex);
848 mutex_unlock(&mdsc->mutex);
851 void ceph_mdsc_release_request(struct kref *kref)
853 struct ceph_mds_request *req = container_of(kref,
854 struct ceph_mds_request,
856 ceph_mdsc_release_dir_caps_no_check(req);
857 destroy_reply_info(&req->r_reply_info);
859 ceph_msg_put(req->r_request);
861 ceph_msg_put(req->r_reply);
863 ceph_put_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
867 ceph_put_cap_refs(ceph_inode(req->r_parent), CEPH_CAP_PIN);
870 iput(req->r_target_inode);
873 if (req->r_old_dentry)
874 dput(req->r_old_dentry);
875 if (req->r_old_dentry_dir) {
877 * track (and drop pins for) r_old_dentry_dir
878 * separately, since r_old_dentry's d_parent may have
879 * changed between the dir mutex being dropped and
880 * this request being freed.
882 ceph_put_cap_refs(ceph_inode(req->r_old_dentry_dir),
884 iput(req->r_old_dentry_dir);
888 put_cred(req->r_cred);
890 ceph_pagelist_release(req->r_pagelist);
891 put_request_session(req);
892 ceph_unreserve_caps(req->r_mdsc, &req->r_caps_reservation);
893 WARN_ON_ONCE(!list_empty(&req->r_wait));
894 kmem_cache_free(ceph_mds_request_cachep, req);
897 DEFINE_RB_FUNCS(request, struct ceph_mds_request, r_tid, r_node)
900 * lookup session, bump ref if found.
902 * called under mdsc->mutex.
904 static struct ceph_mds_request *
905 lookup_get_request(struct ceph_mds_client *mdsc, u64 tid)
907 struct ceph_mds_request *req;
909 req = lookup_request(&mdsc->request_tree, tid);
911 ceph_mdsc_get_request(req);
917 * Register an in-flight request, and assign a tid. Link to directory
918 * are modifying (if any).
920 * Called under mdsc->mutex.
922 static void __register_request(struct ceph_mds_client *mdsc,
923 struct ceph_mds_request *req,
928 req->r_tid = ++mdsc->last_tid;
929 if (req->r_num_caps) {
930 ret = ceph_reserve_caps(mdsc, &req->r_caps_reservation,
933 pr_err("__register_request %p "
934 "failed to reserve caps: %d\n", req, ret);
935 /* set req->r_err to fail early from __do_request */
940 dout("__register_request %p tid %lld\n", req, req->r_tid);
941 ceph_mdsc_get_request(req);
942 insert_request(&mdsc->request_tree, req);
944 req->r_cred = get_current_cred();
946 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
947 mdsc->oldest_tid = req->r_tid;
950 struct ceph_inode_info *ci = ceph_inode(dir);
953 req->r_unsafe_dir = dir;
954 spin_lock(&ci->i_unsafe_lock);
955 list_add_tail(&req->r_unsafe_dir_item, &ci->i_unsafe_dirops);
956 spin_unlock(&ci->i_unsafe_lock);
960 static void __unregister_request(struct ceph_mds_client *mdsc,
961 struct ceph_mds_request *req)
963 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
965 /* Never leave an unregistered request on an unsafe list! */
966 list_del_init(&req->r_unsafe_item);
968 if (req->r_tid == mdsc->oldest_tid) {
969 struct rb_node *p = rb_next(&req->r_node);
970 mdsc->oldest_tid = 0;
972 struct ceph_mds_request *next_req =
973 rb_entry(p, struct ceph_mds_request, r_node);
974 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
975 mdsc->oldest_tid = next_req->r_tid;
982 erase_request(&mdsc->request_tree, req);
984 if (req->r_unsafe_dir) {
985 struct ceph_inode_info *ci = ceph_inode(req->r_unsafe_dir);
986 spin_lock(&ci->i_unsafe_lock);
987 list_del_init(&req->r_unsafe_dir_item);
988 spin_unlock(&ci->i_unsafe_lock);
990 if (req->r_target_inode &&
991 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
992 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
993 spin_lock(&ci->i_unsafe_lock);
994 list_del_init(&req->r_unsafe_target_item);
995 spin_unlock(&ci->i_unsafe_lock);
998 if (req->r_unsafe_dir) {
999 iput(req->r_unsafe_dir);
1000 req->r_unsafe_dir = NULL;
1003 complete_all(&req->r_safe_completion);
1005 ceph_mdsc_put_request(req);
1009 * Walk back up the dentry tree until we hit a dentry representing a
1010 * non-snapshot inode. We do this using the rcu_read_lock (which must be held
1011 * when calling this) to ensure that the objects won't disappear while we're
1012 * working with them. Once we hit a candidate dentry, we attempt to take a
1013 * reference to it, and return that as the result.
1015 static struct inode *get_nonsnap_parent(struct dentry *dentry)
1017 struct inode *inode = NULL;
1019 while (dentry && !IS_ROOT(dentry)) {
1020 inode = d_inode_rcu(dentry);
1021 if (!inode || ceph_snap(inode) == CEPH_NOSNAP)
1023 dentry = dentry->d_parent;
1026 inode = igrab(inode);
1031 * Choose mds to send request to next. If there is a hint set in the
1032 * request (e.g., due to a prior forward hint from the mds), use that.
1033 * Otherwise, consult frag tree and/or caps to identify the
1034 * appropriate mds. If all else fails, choose randomly.
1036 * Called under mdsc->mutex.
1038 static int __choose_mds(struct ceph_mds_client *mdsc,
1039 struct ceph_mds_request *req,
1042 struct inode *inode;
1043 struct ceph_inode_info *ci;
1044 struct ceph_cap *cap;
1045 int mode = req->r_direct_mode;
1047 u32 hash = req->r_direct_hash;
1048 bool is_hash = test_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
1054 * is there a specific mds we should try? ignore hint if we have
1055 * no session and the mds is not up (active or recovering).
1057 if (req->r_resend_mds >= 0 &&
1058 (__have_session(mdsc, req->r_resend_mds) ||
1059 ceph_mdsmap_get_state(mdsc->mdsmap, req->r_resend_mds) > 0)) {
1060 dout("%s using resend_mds mds%d\n", __func__,
1062 return req->r_resend_mds;
1065 if (mode == USE_RANDOM_MDS)
1070 if (ceph_snap(req->r_inode) != CEPH_SNAPDIR) {
1071 inode = req->r_inode;
1074 /* req->r_dentry is non-null for LSSNAP request */
1076 inode = get_nonsnap_parent(req->r_dentry);
1078 dout("%s using snapdir's parent %p\n", __func__, inode);
1080 } else if (req->r_dentry) {
1081 /* ignore race with rename; old or new d_parent is okay */
1082 struct dentry *parent;
1086 parent = READ_ONCE(req->r_dentry->d_parent);
1087 dir = req->r_parent ? : d_inode_rcu(parent);
1089 if (!dir || dir->i_sb != mdsc->fsc->sb) {
1090 /* not this fs or parent went negative */
1091 inode = d_inode(req->r_dentry);
1094 } else if (ceph_snap(dir) != CEPH_NOSNAP) {
1095 /* direct snapped/virtual snapdir requests
1096 * based on parent dir inode */
1097 inode = get_nonsnap_parent(parent);
1098 dout("%s using nonsnap parent %p\n", __func__, inode);
1101 inode = d_inode(req->r_dentry);
1102 if (!inode || mode == USE_AUTH_MDS) {
1105 hash = ceph_dentry_hash(dir, req->r_dentry);
1114 dout("%s %p is_hash=%d (0x%x) mode %d\n", __func__, inode, (int)is_hash,
1118 ci = ceph_inode(inode);
1120 if (is_hash && S_ISDIR(inode->i_mode)) {
1121 struct ceph_inode_frag frag;
1124 ceph_choose_frag(ci, hash, &frag, &found);
1126 if (mode == USE_ANY_MDS && frag.ndist > 0) {
1129 /* choose a random replica */
1130 get_random_bytes(&r, 1);
1133 dout("%s %p %llx.%llx frag %u mds%d (%d/%d)\n",
1134 __func__, inode, ceph_vinop(inode),
1135 frag.frag, mds, (int)r, frag.ndist);
1136 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1137 CEPH_MDS_STATE_ACTIVE &&
1138 !ceph_mdsmap_is_laggy(mdsc->mdsmap, mds))
1142 /* since this file/dir wasn't known to be
1143 * replicated, then we want to look for the
1144 * authoritative mds. */
1145 if (frag.mds >= 0) {
1146 /* choose auth mds */
1148 dout("%s %p %llx.%llx frag %u mds%d (auth)\n",
1149 __func__, inode, ceph_vinop(inode),
1151 if (ceph_mdsmap_get_state(mdsc->mdsmap, mds) >=
1152 CEPH_MDS_STATE_ACTIVE) {
1153 if (!ceph_mdsmap_is_laggy(mdsc->mdsmap,
1158 mode = USE_AUTH_MDS;
1162 spin_lock(&ci->i_ceph_lock);
1164 if (mode == USE_AUTH_MDS)
1165 cap = ci->i_auth_cap;
1166 if (!cap && !RB_EMPTY_ROOT(&ci->i_caps))
1167 cap = rb_entry(rb_first(&ci->i_caps), struct ceph_cap, ci_node);
1169 spin_unlock(&ci->i_ceph_lock);
1173 mds = cap->session->s_mds;
1174 dout("%s %p %llx.%llx mds%d (%scap %p)\n", __func__,
1175 inode, ceph_vinop(inode), mds,
1176 cap == ci->i_auth_cap ? "auth " : "", cap);
1177 spin_unlock(&ci->i_ceph_lock);
1186 mds = ceph_mdsmap_get_random_mds(mdsc->mdsmap);
1187 dout("%s chose random mds%d\n", __func__, mds);
1195 struct ceph_msg *ceph_create_session_msg(u32 op, u64 seq)
1197 struct ceph_msg *msg;
1198 struct ceph_mds_session_head *h;
1200 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h), GFP_NOFS,
1203 pr_err("ENOMEM creating session %s msg\n",
1204 ceph_session_op_name(op));
1207 h = msg->front.iov_base;
1208 h->op = cpu_to_le32(op);
1209 h->seq = cpu_to_le64(seq);
1214 static const unsigned char feature_bits[] = CEPHFS_FEATURES_CLIENT_SUPPORTED;
1215 #define FEATURE_BYTES(c) (DIV_ROUND_UP((size_t)feature_bits[c - 1] + 1, 64) * 8)
1216 static int encode_supported_features(void **p, void *end)
1218 static const size_t count = ARRAY_SIZE(feature_bits);
1222 size_t size = FEATURE_BYTES(count);
1224 if (WARN_ON_ONCE(*p + 4 + size > end))
1227 ceph_encode_32(p, size);
1228 memset(*p, 0, size);
1229 for (i = 0; i < count; i++)
1230 ((unsigned char*)(*p))[i / 8] |= BIT(feature_bits[i] % 8);
1233 if (WARN_ON_ONCE(*p + 4 > end))
1236 ceph_encode_32(p, 0);
1242 static const unsigned char metric_bits[] = CEPHFS_METRIC_SPEC_CLIENT_SUPPORTED;
1243 #define METRIC_BYTES(cnt) (DIV_ROUND_UP((size_t)metric_bits[cnt - 1] + 1, 64) * 8)
1244 static int encode_metric_spec(void **p, void *end)
1246 static const size_t count = ARRAY_SIZE(metric_bits);
1249 if (WARN_ON_ONCE(*p + 2 > end))
1252 ceph_encode_8(p, 1); /* version */
1253 ceph_encode_8(p, 1); /* compat */
1257 size_t size = METRIC_BYTES(count);
1259 if (WARN_ON_ONCE(*p + 4 + 4 + size > end))
1262 /* metric spec info length */
1263 ceph_encode_32(p, 4 + size);
1266 ceph_encode_32(p, size);
1267 memset(*p, 0, size);
1268 for (i = 0; i < count; i++)
1269 ((unsigned char *)(*p))[i / 8] |= BIT(metric_bits[i] % 8);
1272 if (WARN_ON_ONCE(*p + 4 + 4 > end))
1275 /* metric spec info length */
1276 ceph_encode_32(p, 4);
1278 ceph_encode_32(p, 0);
1285 * session message, specialization for CEPH_SESSION_REQUEST_OPEN
1286 * to include additional client metadata fields.
1288 static struct ceph_msg *create_session_open_msg(struct ceph_mds_client *mdsc, u64 seq)
1290 struct ceph_msg *msg;
1291 struct ceph_mds_session_head *h;
1293 int extra_bytes = 0;
1294 int metadata_key_count = 0;
1295 struct ceph_options *opt = mdsc->fsc->client->options;
1296 struct ceph_mount_options *fsopt = mdsc->fsc->mount_options;
1301 const char* metadata[][2] = {
1302 {"hostname", mdsc->nodename},
1303 {"kernel_version", init_utsname()->release},
1304 {"entity_id", opt->name ? : ""},
1305 {"root", fsopt->server_path ? : "/"},
1309 /* Calculate serialized length of metadata */
1310 extra_bytes = 4; /* map length */
1311 for (i = 0; metadata[i][0]; ++i) {
1312 extra_bytes += 8 + strlen(metadata[i][0]) +
1313 strlen(metadata[i][1]);
1314 metadata_key_count++;
1317 /* supported feature */
1319 count = ARRAY_SIZE(feature_bits);
1321 size = FEATURE_BYTES(count);
1322 extra_bytes += 4 + size;
1326 count = ARRAY_SIZE(metric_bits);
1328 size = METRIC_BYTES(count);
1329 extra_bytes += 2 + 4 + 4 + size;
1331 /* Allocate the message */
1332 msg = ceph_msg_new(CEPH_MSG_CLIENT_SESSION, sizeof(*h) + extra_bytes,
1335 pr_err("ENOMEM creating session open msg\n");
1336 return ERR_PTR(-ENOMEM);
1338 p = msg->front.iov_base;
1339 end = p + msg->front.iov_len;
1342 h->op = cpu_to_le32(CEPH_SESSION_REQUEST_OPEN);
1343 h->seq = cpu_to_le64(seq);
1346 * Serialize client metadata into waiting buffer space, using
1347 * the format that userspace expects for map<string, string>
1349 * ClientSession messages with metadata are v4
1351 msg->hdr.version = cpu_to_le16(4);
1352 msg->hdr.compat_version = cpu_to_le16(1);
1354 /* The write pointer, following the session_head structure */
1357 /* Number of entries in the map */
1358 ceph_encode_32(&p, metadata_key_count);
1360 /* Two length-prefixed strings for each entry in the map */
1361 for (i = 0; metadata[i][0]; ++i) {
1362 size_t const key_len = strlen(metadata[i][0]);
1363 size_t const val_len = strlen(metadata[i][1]);
1365 ceph_encode_32(&p, key_len);
1366 memcpy(p, metadata[i][0], key_len);
1368 ceph_encode_32(&p, val_len);
1369 memcpy(p, metadata[i][1], val_len);
1373 ret = encode_supported_features(&p, end);
1375 pr_err("encode_supported_features failed!\n");
1377 return ERR_PTR(ret);
1380 ret = encode_metric_spec(&p, end);
1382 pr_err("encode_metric_spec failed!\n");
1384 return ERR_PTR(ret);
1387 msg->front.iov_len = p - msg->front.iov_base;
1388 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1394 * send session open request.
1396 * called under mdsc->mutex
1398 static int __open_session(struct ceph_mds_client *mdsc,
1399 struct ceph_mds_session *session)
1401 struct ceph_msg *msg;
1403 int mds = session->s_mds;
1405 /* wait for mds to go active? */
1406 mstate = ceph_mdsmap_get_state(mdsc->mdsmap, mds);
1407 dout("open_session to mds%d (%s)\n", mds,
1408 ceph_mds_state_name(mstate));
1409 session->s_state = CEPH_MDS_SESSION_OPENING;
1410 session->s_renew_requested = jiffies;
1412 /* send connect message */
1413 msg = create_session_open_msg(mdsc, session->s_seq);
1415 return PTR_ERR(msg);
1416 ceph_con_send(&session->s_con, msg);
1421 * open sessions for any export targets for the given mds
1423 * called under mdsc->mutex
1425 static struct ceph_mds_session *
1426 __open_export_target_session(struct ceph_mds_client *mdsc, int target)
1428 struct ceph_mds_session *session;
1431 session = __ceph_lookup_mds_session(mdsc, target);
1433 session = register_session(mdsc, target);
1434 if (IS_ERR(session))
1437 if (session->s_state == CEPH_MDS_SESSION_NEW ||
1438 session->s_state == CEPH_MDS_SESSION_CLOSING) {
1439 ret = __open_session(mdsc, session);
1441 return ERR_PTR(ret);
1447 struct ceph_mds_session *
1448 ceph_mdsc_open_export_target_session(struct ceph_mds_client *mdsc, int target)
1450 struct ceph_mds_session *session;
1452 dout("open_export_target_session to mds%d\n", target);
1454 mutex_lock(&mdsc->mutex);
1455 session = __open_export_target_session(mdsc, target);
1456 mutex_unlock(&mdsc->mutex);
1461 static void __open_export_target_sessions(struct ceph_mds_client *mdsc,
1462 struct ceph_mds_session *session)
1464 struct ceph_mds_info *mi;
1465 struct ceph_mds_session *ts;
1466 int i, mds = session->s_mds;
1468 if (mds >= mdsc->mdsmap->possible_max_rank)
1471 mi = &mdsc->mdsmap->m_info[mds];
1472 dout("open_export_target_sessions for mds%d (%d targets)\n",
1473 session->s_mds, mi->num_export_targets);
1475 for (i = 0; i < mi->num_export_targets; i++) {
1476 ts = __open_export_target_session(mdsc, mi->export_targets[i]);
1477 ceph_put_mds_session(ts);
1481 void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
1482 struct ceph_mds_session *session)
1484 mutex_lock(&mdsc->mutex);
1485 __open_export_target_sessions(mdsc, session);
1486 mutex_unlock(&mdsc->mutex);
1493 static void detach_cap_releases(struct ceph_mds_session *session,
1494 struct list_head *target)
1496 lockdep_assert_held(&session->s_cap_lock);
1498 list_splice_init(&session->s_cap_releases, target);
1499 session->s_num_cap_releases = 0;
1500 dout("dispose_cap_releases mds%d\n", session->s_mds);
1503 static void dispose_cap_releases(struct ceph_mds_client *mdsc,
1504 struct list_head *dispose)
1506 while (!list_empty(dispose)) {
1507 struct ceph_cap *cap;
1508 /* zero out the in-progress message */
1509 cap = list_first_entry(dispose, struct ceph_cap, session_caps);
1510 list_del(&cap->session_caps);
1511 ceph_put_cap(mdsc, cap);
1515 static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1516 struct ceph_mds_session *session)
1518 struct ceph_mds_request *req;
1521 dout("cleanup_session_requests mds%d\n", session->s_mds);
1522 mutex_lock(&mdsc->mutex);
1523 while (!list_empty(&session->s_unsafe)) {
1524 req = list_first_entry(&session->s_unsafe,
1525 struct ceph_mds_request, r_unsafe_item);
1526 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1528 if (req->r_target_inode)
1529 mapping_set_error(req->r_target_inode->i_mapping, -EIO);
1530 if (req->r_unsafe_dir)
1531 mapping_set_error(req->r_unsafe_dir->i_mapping, -EIO);
1532 __unregister_request(mdsc, req);
1534 /* zero r_attempts, so kick_requests() will re-send requests */
1535 p = rb_first(&mdsc->request_tree);
1537 req = rb_entry(p, struct ceph_mds_request, r_node);
1539 if (req->r_session &&
1540 req->r_session->s_mds == session->s_mds)
1541 req->r_attempts = 0;
1543 mutex_unlock(&mdsc->mutex);
1547 * Helper to safely iterate over all caps associated with a session, with
1548 * special care taken to handle a racing __ceph_remove_cap().
1550 * Caller must hold session s_mutex.
1552 int ceph_iterate_session_caps(struct ceph_mds_session *session,
1553 int (*cb)(struct inode *, struct ceph_cap *,
1556 struct list_head *p;
1557 struct ceph_cap *cap;
1558 struct inode *inode, *last_inode = NULL;
1559 struct ceph_cap *old_cap = NULL;
1562 dout("iterate_session_caps %p mds%d\n", session, session->s_mds);
1563 spin_lock(&session->s_cap_lock);
1564 p = session->s_caps.next;
1565 while (p != &session->s_caps) {
1566 cap = list_entry(p, struct ceph_cap, session_caps);
1567 inode = igrab(&cap->ci->netfs.inode);
1572 session->s_cap_iterator = cap;
1573 spin_unlock(&session->s_cap_lock);
1580 ceph_put_cap(session->s_mdsc, old_cap);
1584 ret = cb(inode, cap, arg);
1587 spin_lock(&session->s_cap_lock);
1590 dout("iterate_session_caps finishing cap %p removal\n",
1592 BUG_ON(cap->session != session);
1593 cap->session = NULL;
1594 list_del_init(&cap->session_caps);
1595 session->s_nr_caps--;
1596 atomic64_dec(&session->s_mdsc->metric.total_caps);
1597 if (cap->queue_release)
1598 __ceph_queue_cap_release(session, cap);
1600 old_cap = cap; /* put_cap it w/o locks held */
1607 session->s_cap_iterator = NULL;
1608 spin_unlock(&session->s_cap_lock);
1612 ceph_put_cap(session->s_mdsc, old_cap);
1617 static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1620 struct ceph_inode_info *ci = ceph_inode(inode);
1621 bool invalidate = false;
1624 dout("removing cap %p, ci is %p, inode is %p\n",
1625 cap, ci, &ci->netfs.inode);
1626 spin_lock(&ci->i_ceph_lock);
1627 iputs = ceph_purge_inode_cap(inode, cap, &invalidate);
1628 spin_unlock(&ci->i_ceph_lock);
1630 wake_up_all(&ci->i_cap_wq);
1632 ceph_queue_invalidate(inode);
1639 * caller must hold session s_mutex
1641 static void remove_session_caps(struct ceph_mds_session *session)
1643 struct ceph_fs_client *fsc = session->s_mdsc->fsc;
1644 struct super_block *sb = fsc->sb;
1647 dout("remove_session_caps on %p\n", session);
1648 ceph_iterate_session_caps(session, remove_session_caps_cb, fsc);
1650 wake_up_all(&fsc->mdsc->cap_flushing_wq);
1652 spin_lock(&session->s_cap_lock);
1653 if (session->s_nr_caps > 0) {
1654 struct inode *inode;
1655 struct ceph_cap *cap, *prev = NULL;
1656 struct ceph_vino vino;
1658 * iterate_session_caps() skips inodes that are being
1659 * deleted, we need to wait until deletions are complete.
1660 * __wait_on_freeing_inode() is designed for the job,
1661 * but it is not exported, so use lookup inode function
1664 while (!list_empty(&session->s_caps)) {
1665 cap = list_entry(session->s_caps.next,
1666 struct ceph_cap, session_caps);
1670 vino = cap->ci->i_vino;
1671 spin_unlock(&session->s_cap_lock);
1673 inode = ceph_find_inode(sb, vino);
1676 spin_lock(&session->s_cap_lock);
1680 // drop cap expires and unlock s_cap_lock
1681 detach_cap_releases(session, &dispose);
1683 BUG_ON(session->s_nr_caps > 0);
1684 BUG_ON(!list_empty(&session->s_cap_flushing));
1685 spin_unlock(&session->s_cap_lock);
1686 dispose_cap_releases(session->s_mdsc, &dispose);
1696 * wake up any threads waiting on this session's caps. if the cap is
1697 * old (didn't get renewed on the client reconnect), remove it now.
1699 * caller must hold s_mutex.
1701 static int wake_up_session_cb(struct inode *inode, struct ceph_cap *cap,
1704 struct ceph_inode_info *ci = ceph_inode(inode);
1705 unsigned long ev = (unsigned long)arg;
1707 if (ev == RECONNECT) {
1708 spin_lock(&ci->i_ceph_lock);
1709 ci->i_wanted_max_size = 0;
1710 ci->i_requested_max_size = 0;
1711 spin_unlock(&ci->i_ceph_lock);
1712 } else if (ev == RENEWCAPS) {
1713 if (cap->cap_gen < atomic_read(&cap->session->s_cap_gen)) {
1714 /* mds did not re-issue stale cap */
1715 spin_lock(&ci->i_ceph_lock);
1716 cap->issued = cap->implemented = CEPH_CAP_PIN;
1717 spin_unlock(&ci->i_ceph_lock);
1719 } else if (ev == FORCE_RO) {
1721 wake_up_all(&ci->i_cap_wq);
1725 static void wake_up_session_caps(struct ceph_mds_session *session, int ev)
1727 dout("wake_up_session_caps %p mds%d\n", session, session->s_mds);
1728 ceph_iterate_session_caps(session, wake_up_session_cb,
1729 (void *)(unsigned long)ev);
1733 * Send periodic message to MDS renewing all currently held caps. The
1734 * ack will reset the expiration for all caps from this session.
1736 * caller holds s_mutex
1738 static int send_renew_caps(struct ceph_mds_client *mdsc,
1739 struct ceph_mds_session *session)
1741 struct ceph_msg *msg;
1744 if (time_after_eq(jiffies, session->s_cap_ttl) &&
1745 time_after_eq(session->s_cap_ttl, session->s_renew_requested))
1746 pr_info("mds%d caps stale\n", session->s_mds);
1747 session->s_renew_requested = jiffies;
1749 /* do not try to renew caps until a recovering mds has reconnected
1750 * with its clients. */
1751 state = ceph_mdsmap_get_state(mdsc->mdsmap, session->s_mds);
1752 if (state < CEPH_MDS_STATE_RECONNECT) {
1753 dout("send_renew_caps ignoring mds%d (%s)\n",
1754 session->s_mds, ceph_mds_state_name(state));
1758 dout("send_renew_caps to mds%d (%s)\n", session->s_mds,
1759 ceph_mds_state_name(state));
1760 msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_RENEWCAPS,
1761 ++session->s_renew_seq);
1764 ceph_con_send(&session->s_con, msg);
1768 static int send_flushmsg_ack(struct ceph_mds_client *mdsc,
1769 struct ceph_mds_session *session, u64 seq)
1771 struct ceph_msg *msg;
1773 dout("send_flushmsg_ack to mds%d (%s)s seq %lld\n",
1774 session->s_mds, ceph_session_state_name(session->s_state), seq);
1775 msg = ceph_create_session_msg(CEPH_SESSION_FLUSHMSG_ACK, seq);
1778 ceph_con_send(&session->s_con, msg);
1784 * Note new cap ttl, and any transition from stale -> not stale (fresh?).
1786 * Called under session->s_mutex
1788 static void renewed_caps(struct ceph_mds_client *mdsc,
1789 struct ceph_mds_session *session, int is_renew)
1794 spin_lock(&session->s_cap_lock);
1795 was_stale = is_renew && time_after_eq(jiffies, session->s_cap_ttl);
1797 session->s_cap_ttl = session->s_renew_requested +
1798 mdsc->mdsmap->m_session_timeout*HZ;
1801 if (time_before(jiffies, session->s_cap_ttl)) {
1802 pr_info("mds%d caps renewed\n", session->s_mds);
1805 pr_info("mds%d caps still stale\n", session->s_mds);
1808 dout("renewed_caps mds%d ttl now %lu, was %s, now %s\n",
1809 session->s_mds, session->s_cap_ttl, was_stale ? "stale" : "fresh",
1810 time_before(jiffies, session->s_cap_ttl) ? "stale" : "fresh");
1811 spin_unlock(&session->s_cap_lock);
1814 wake_up_session_caps(session, RENEWCAPS);
1818 * send a session close request
1820 static int request_close_session(struct ceph_mds_session *session)
1822 struct ceph_msg *msg;
1824 dout("request_close_session mds%d state %s seq %lld\n",
1825 session->s_mds, ceph_session_state_name(session->s_state),
1827 msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_CLOSE,
1831 ceph_con_send(&session->s_con, msg);
1836 * Called with s_mutex held.
1838 static int __close_session(struct ceph_mds_client *mdsc,
1839 struct ceph_mds_session *session)
1841 if (session->s_state >= CEPH_MDS_SESSION_CLOSING)
1843 session->s_state = CEPH_MDS_SESSION_CLOSING;
1844 return request_close_session(session);
1847 static bool drop_negative_children(struct dentry *dentry)
1849 struct dentry *child;
1850 bool all_negative = true;
1852 if (!d_is_dir(dentry))
1855 spin_lock(&dentry->d_lock);
1856 list_for_each_entry(child, &dentry->d_subdirs, d_child) {
1857 if (d_really_is_positive(child)) {
1858 all_negative = false;
1862 spin_unlock(&dentry->d_lock);
1865 shrink_dcache_parent(dentry);
1867 return all_negative;
1871 * Trim old(er) caps.
1873 * Because we can't cache an inode without one or more caps, we do
1874 * this indirectly: if a cap is unused, we prune its aliases, at which
1875 * point the inode will hopefully get dropped to.
1877 * Yes, this is a bit sloppy. Our only real goal here is to respond to
1878 * memory pressure from the MDS, though, so it needn't be perfect.
1880 static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1882 int *remaining = arg;
1883 struct ceph_inode_info *ci = ceph_inode(inode);
1884 int used, wanted, oissued, mine;
1886 if (*remaining <= 0)
1889 spin_lock(&ci->i_ceph_lock);
1890 mine = cap->issued | cap->implemented;
1891 used = __ceph_caps_used(ci);
1892 wanted = __ceph_caps_file_wanted(ci);
1893 oissued = __ceph_caps_issued_other(ci, cap);
1895 dout("trim_caps_cb %p cap %p mine %s oissued %s used %s wanted %s\n",
1896 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1897 ceph_cap_string(used), ceph_cap_string(wanted));
1898 if (cap == ci->i_auth_cap) {
1899 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1900 !list_empty(&ci->i_cap_snaps))
1902 if ((used | wanted) & CEPH_CAP_ANY_WR)
1904 /* Note: it's possible that i_filelock_ref becomes non-zero
1905 * after dropping auth caps. It doesn't hurt because reply
1906 * of lock mds request will re-add auth caps. */
1907 if (atomic_read(&ci->i_filelock_ref) > 0)
1910 /* The inode has cached pages, but it's no longer used.
1911 * we can safely drop it */
1912 if (S_ISREG(inode->i_mode) &&
1913 wanted == 0 && used == CEPH_CAP_FILE_CACHE &&
1914 !(oissued & CEPH_CAP_FILE_CACHE)) {
1918 if ((used | wanted) & ~oissued & mine)
1919 goto out; /* we need these caps */
1922 /* we aren't the only cap.. just remove us */
1923 ceph_remove_cap(cap, true);
1926 struct dentry *dentry;
1927 /* try dropping referring dentries */
1928 spin_unlock(&ci->i_ceph_lock);
1929 dentry = d_find_any_alias(inode);
1930 if (dentry && drop_negative_children(dentry)) {
1933 d_prune_aliases(inode);
1934 count = atomic_read(&inode->i_count);
1937 dout("trim_caps_cb %p cap %p pruned, count now %d\n",
1946 spin_unlock(&ci->i_ceph_lock);
1951 * Trim session cap count down to some max number.
1953 int ceph_trim_caps(struct ceph_mds_client *mdsc,
1954 struct ceph_mds_session *session,
1957 int trim_caps = session->s_nr_caps - max_caps;
1959 dout("trim_caps mds%d start: %d / %d, trim %d\n",
1960 session->s_mds, session->s_nr_caps, max_caps, trim_caps);
1961 if (trim_caps > 0) {
1962 int remaining = trim_caps;
1964 ceph_iterate_session_caps(session, trim_caps_cb, &remaining);
1965 dout("trim_caps mds%d done: %d / %d, trimmed %d\n",
1966 session->s_mds, session->s_nr_caps, max_caps,
1967 trim_caps - remaining);
1970 ceph_flush_cap_releases(mdsc, session);
1974 static int check_caps_flush(struct ceph_mds_client *mdsc,
1979 spin_lock(&mdsc->cap_dirty_lock);
1980 if (!list_empty(&mdsc->cap_flush_list)) {
1981 struct ceph_cap_flush *cf =
1982 list_first_entry(&mdsc->cap_flush_list,
1983 struct ceph_cap_flush, g_list);
1984 if (cf->tid <= want_flush_tid) {
1985 dout("check_caps_flush still flushing tid "
1986 "%llu <= %llu\n", cf->tid, want_flush_tid);
1990 spin_unlock(&mdsc->cap_dirty_lock);
1995 * flush all dirty inode data to disk.
1997 * returns true if we've flushed through want_flush_tid
1999 static void wait_caps_flush(struct ceph_mds_client *mdsc,
2002 dout("check_caps_flush want %llu\n", want_flush_tid);
2004 wait_event(mdsc->cap_flushing_wq,
2005 check_caps_flush(mdsc, want_flush_tid));
2007 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
2011 * called under s_mutex
2013 static void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
2014 struct ceph_mds_session *session)
2016 struct ceph_msg *msg = NULL;
2017 struct ceph_mds_cap_release *head;
2018 struct ceph_mds_cap_item *item;
2019 struct ceph_osd_client *osdc = &mdsc->fsc->client->osdc;
2020 struct ceph_cap *cap;
2021 LIST_HEAD(tmp_list);
2022 int num_cap_releases;
2023 __le32 barrier, *cap_barrier;
2025 down_read(&osdc->lock);
2026 barrier = cpu_to_le32(osdc->epoch_barrier);
2027 up_read(&osdc->lock);
2029 spin_lock(&session->s_cap_lock);
2031 list_splice_init(&session->s_cap_releases, &tmp_list);
2032 num_cap_releases = session->s_num_cap_releases;
2033 session->s_num_cap_releases = 0;
2034 spin_unlock(&session->s_cap_lock);
2036 while (!list_empty(&tmp_list)) {
2038 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
2039 PAGE_SIZE, GFP_NOFS, false);
2042 head = msg->front.iov_base;
2043 head->num = cpu_to_le32(0);
2044 msg->front.iov_len = sizeof(*head);
2046 msg->hdr.version = cpu_to_le16(2);
2047 msg->hdr.compat_version = cpu_to_le16(1);
2050 cap = list_first_entry(&tmp_list, struct ceph_cap,
2052 list_del(&cap->session_caps);
2055 head = msg->front.iov_base;
2056 put_unaligned_le32(get_unaligned_le32(&head->num) + 1,
2058 item = msg->front.iov_base + msg->front.iov_len;
2059 item->ino = cpu_to_le64(cap->cap_ino);
2060 item->cap_id = cpu_to_le64(cap->cap_id);
2061 item->migrate_seq = cpu_to_le32(cap->mseq);
2062 item->seq = cpu_to_le32(cap->issue_seq);
2063 msg->front.iov_len += sizeof(*item);
2065 ceph_put_cap(mdsc, cap);
2067 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
2068 // Append cap_barrier field
2069 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2070 *cap_barrier = barrier;
2071 msg->front.iov_len += sizeof(*cap_barrier);
2073 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2074 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2075 ceph_con_send(&session->s_con, msg);
2080 BUG_ON(num_cap_releases != 0);
2082 spin_lock(&session->s_cap_lock);
2083 if (!list_empty(&session->s_cap_releases))
2085 spin_unlock(&session->s_cap_lock);
2088 // Append cap_barrier field
2089 cap_barrier = msg->front.iov_base + msg->front.iov_len;
2090 *cap_barrier = barrier;
2091 msg->front.iov_len += sizeof(*cap_barrier);
2093 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2094 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
2095 ceph_con_send(&session->s_con, msg);
2099 pr_err("send_cap_releases mds%d, failed to allocate message\n",
2101 spin_lock(&session->s_cap_lock);
2102 list_splice(&tmp_list, &session->s_cap_releases);
2103 session->s_num_cap_releases += num_cap_releases;
2104 spin_unlock(&session->s_cap_lock);
2107 static void ceph_cap_release_work(struct work_struct *work)
2109 struct ceph_mds_session *session =
2110 container_of(work, struct ceph_mds_session, s_cap_release_work);
2112 mutex_lock(&session->s_mutex);
2113 if (session->s_state == CEPH_MDS_SESSION_OPEN ||
2114 session->s_state == CEPH_MDS_SESSION_HUNG)
2115 ceph_send_cap_releases(session->s_mdsc, session);
2116 mutex_unlock(&session->s_mutex);
2117 ceph_put_mds_session(session);
2120 void ceph_flush_cap_releases(struct ceph_mds_client *mdsc,
2121 struct ceph_mds_session *session)
2126 ceph_get_mds_session(session);
2127 if (queue_work(mdsc->fsc->cap_wq,
2128 &session->s_cap_release_work)) {
2129 dout("cap release work queued\n");
2131 ceph_put_mds_session(session);
2132 dout("failed to queue cap release work\n");
2137 * caller holds session->s_cap_lock
2139 void __ceph_queue_cap_release(struct ceph_mds_session *session,
2140 struct ceph_cap *cap)
2142 list_add_tail(&cap->session_caps, &session->s_cap_releases);
2143 session->s_num_cap_releases++;
2145 if (!(session->s_num_cap_releases % CEPH_CAPS_PER_RELEASE))
2146 ceph_flush_cap_releases(session->s_mdsc, session);
2149 static void ceph_cap_reclaim_work(struct work_struct *work)
2151 struct ceph_mds_client *mdsc =
2152 container_of(work, struct ceph_mds_client, cap_reclaim_work);
2153 int ret = ceph_trim_dentries(mdsc);
2155 ceph_queue_cap_reclaim_work(mdsc);
2158 void ceph_queue_cap_reclaim_work(struct ceph_mds_client *mdsc)
2163 if (queue_work(mdsc->fsc->cap_wq, &mdsc->cap_reclaim_work)) {
2164 dout("caps reclaim work queued\n");
2166 dout("failed to queue caps release work\n");
2170 void ceph_reclaim_caps_nr(struct ceph_mds_client *mdsc, int nr)
2175 val = atomic_add_return(nr, &mdsc->cap_reclaim_pending);
2176 if ((val % CEPH_CAPS_PER_RELEASE) < nr) {
2177 atomic_set(&mdsc->cap_reclaim_pending, 0);
2178 ceph_queue_cap_reclaim_work(mdsc);
2186 int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
2189 struct ceph_inode_info *ci = ceph_inode(dir);
2190 struct ceph_mds_reply_info_parsed *rinfo = &req->r_reply_info;
2191 struct ceph_mount_options *opt = req->r_mdsc->fsc->mount_options;
2192 size_t size = sizeof(struct ceph_mds_reply_dir_entry);
2193 unsigned int num_entries;
2196 spin_lock(&ci->i_ceph_lock);
2197 num_entries = ci->i_files + ci->i_subdirs;
2198 spin_unlock(&ci->i_ceph_lock);
2199 num_entries = max(num_entries, 1U);
2200 num_entries = min(num_entries, opt->max_readdir);
2202 order = get_order(size * num_entries);
2203 while (order >= 0) {
2204 rinfo->dir_entries = (void*)__get_free_pages(GFP_KERNEL |
2208 if (rinfo->dir_entries)
2212 if (!rinfo->dir_entries)
2215 num_entries = (PAGE_SIZE << order) / size;
2216 num_entries = min(num_entries, opt->max_readdir);
2218 rinfo->dir_buf_size = PAGE_SIZE << order;
2219 req->r_num_caps = num_entries + 1;
2220 req->r_args.readdir.max_entries = cpu_to_le32(num_entries);
2221 req->r_args.readdir.max_bytes = cpu_to_le32(opt->max_readdir_bytes);
2226 * Create an mds request.
2228 struct ceph_mds_request *
2229 ceph_mdsc_create_request(struct ceph_mds_client *mdsc, int op, int mode)
2231 struct ceph_mds_request *req;
2233 req = kmem_cache_zalloc(ceph_mds_request_cachep, GFP_NOFS);
2235 return ERR_PTR(-ENOMEM);
2237 mutex_init(&req->r_fill_mutex);
2239 req->r_started = jiffies;
2240 req->r_start_latency = ktime_get();
2241 req->r_resend_mds = -1;
2242 INIT_LIST_HEAD(&req->r_unsafe_dir_item);
2243 INIT_LIST_HEAD(&req->r_unsafe_target_item);
2245 kref_init(&req->r_kref);
2246 RB_CLEAR_NODE(&req->r_node);
2247 INIT_LIST_HEAD(&req->r_wait);
2248 init_completion(&req->r_completion);
2249 init_completion(&req->r_safe_completion);
2250 INIT_LIST_HEAD(&req->r_unsafe_item);
2252 ktime_get_coarse_real_ts64(&req->r_stamp);
2255 req->r_direct_mode = mode;
2260 * return oldest (lowest) request, tid in request tree, 0 if none.
2262 * called under mdsc->mutex.
2264 static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
2266 if (RB_EMPTY_ROOT(&mdsc->request_tree))
2268 return rb_entry(rb_first(&mdsc->request_tree),
2269 struct ceph_mds_request, r_node);
2272 static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
2274 return mdsc->oldest_tid;
2278 * Build a dentry's path. Allocate on heap; caller must kfree. Based
2279 * on build_path_from_dentry in fs/cifs/dir.c.
2281 * If @stop_on_nosnap, generate path relative to the first non-snapped
2284 * Encode hidden .snap dirs as a double /, i.e.
2285 * foo/.snap/bar -> foo//bar
2287 char *ceph_mdsc_build_path(struct dentry *dentry, int *plen, u64 *pbase,
2290 struct dentry *temp;
2297 return ERR_PTR(-EINVAL);
2301 return ERR_PTR(-ENOMEM);
2306 seq = read_seqbegin(&rename_lock);
2310 struct inode *inode;
2312 spin_lock(&temp->d_lock);
2313 inode = d_inode(temp);
2314 if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
2315 dout("build_path path+%d: %p SNAPDIR\n",
2317 } else if (stop_on_nosnap && inode && dentry != temp &&
2318 ceph_snap(inode) == CEPH_NOSNAP) {
2319 spin_unlock(&temp->d_lock);
2320 pos++; /* get rid of any prepended '/' */
2323 pos -= temp->d_name.len;
2325 spin_unlock(&temp->d_lock);
2328 memcpy(path + pos, temp->d_name.name, temp->d_name.len);
2330 spin_unlock(&temp->d_lock);
2331 temp = READ_ONCE(temp->d_parent);
2333 /* Are we at the root? */
2337 /* Are we out of buffer? */
2343 base = ceph_ino(d_inode(temp));
2346 if (read_seqretry(&rename_lock, seq))
2351 * A rename didn't occur, but somehow we didn't end up where
2352 * we thought we would. Throw a warning and try again.
2354 pr_warn("build_path did not end path lookup where "
2355 "expected, pos is %d\n", pos);
2360 *plen = PATH_MAX - 1 - pos;
2361 dout("build_path on %p %d built %llx '%.*s'\n",
2362 dentry, d_count(dentry), base, *plen, path + pos);
2366 static int build_dentry_path(struct dentry *dentry, struct inode *dir,
2367 const char **ppath, int *ppathlen, u64 *pino,
2368 bool *pfreepath, bool parent_locked)
2374 dir = d_inode_rcu(dentry->d_parent);
2375 if (dir && parent_locked && ceph_snap(dir) == CEPH_NOSNAP) {
2376 *pino = ceph_ino(dir);
2378 *ppath = dentry->d_name.name;
2379 *ppathlen = dentry->d_name.len;
2383 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2385 return PTR_ERR(path);
2391 static int build_inode_path(struct inode *inode,
2392 const char **ppath, int *ppathlen, u64 *pino,
2395 struct dentry *dentry;
2398 if (ceph_snap(inode) == CEPH_NOSNAP) {
2399 *pino = ceph_ino(inode);
2403 dentry = d_find_alias(inode);
2404 path = ceph_mdsc_build_path(dentry, ppathlen, pino, 1);
2407 return PTR_ERR(path);
2414 * request arguments may be specified via an inode *, a dentry *, or
2415 * an explicit ino+path.
2417 static int set_request_path_attr(struct inode *rinode, struct dentry *rdentry,
2418 struct inode *rdiri, const char *rpath,
2419 u64 rino, const char **ppath, int *pathlen,
2420 u64 *ino, bool *freepath, bool parent_locked)
2425 r = build_inode_path(rinode, ppath, pathlen, ino, freepath);
2426 dout(" inode %p %llx.%llx\n", rinode, ceph_ino(rinode),
2428 } else if (rdentry) {
2429 r = build_dentry_path(rdentry, rdiri, ppath, pathlen, ino,
2430 freepath, parent_locked);
2431 dout(" dentry %p %llx/%.*s\n", rdentry, *ino, *pathlen,
2433 } else if (rpath || rino) {
2436 *pathlen = rpath ? strlen(rpath) : 0;
2437 dout(" path %.*s\n", *pathlen, rpath);
2443 static void encode_timestamp_and_gids(void **p,
2444 const struct ceph_mds_request *req)
2446 struct ceph_timespec ts;
2449 ceph_encode_timespec64(&ts, &req->r_stamp);
2450 ceph_encode_copy(p, &ts, sizeof(ts));
2453 ceph_encode_32(p, req->r_cred->group_info->ngroups);
2454 for (i = 0; i < req->r_cred->group_info->ngroups; i++)
2455 ceph_encode_64(p, from_kgid(&init_user_ns,
2456 req->r_cred->group_info->gid[i]));
2460 * called under mdsc->mutex
2462 static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
2463 struct ceph_mds_request *req,
2464 bool drop_cap_releases)
2466 int mds = session->s_mds;
2467 struct ceph_mds_client *mdsc = session->s_mdsc;
2468 struct ceph_msg *msg;
2469 struct ceph_mds_request_head_old *head;
2470 const char *path1 = NULL;
2471 const char *path2 = NULL;
2472 u64 ino1 = 0, ino2 = 0;
2473 int pathlen1 = 0, pathlen2 = 0;
2474 bool freepath1 = false, freepath2 = false;
2479 bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
2481 ret = set_request_path_attr(req->r_inode, req->r_dentry,
2482 req->r_parent, req->r_path1, req->r_ino1.ino,
2483 &path1, &pathlen1, &ino1, &freepath1,
2484 test_bit(CEPH_MDS_R_PARENT_LOCKED,
2485 &req->r_req_flags));
2491 /* If r_old_dentry is set, then assume that its parent is locked */
2492 ret = set_request_path_attr(NULL, req->r_old_dentry,
2493 req->r_old_dentry_dir,
2494 req->r_path2, req->r_ino2.ino,
2495 &path2, &pathlen2, &ino2, &freepath2, true);
2501 len = legacy ? sizeof(*head) : sizeof(struct ceph_mds_request_head);
2502 len += pathlen1 + pathlen2 + 2*(1 + sizeof(u32) + sizeof(u64)) +
2503 sizeof(struct ceph_timespec);
2504 len += sizeof(u32) + (sizeof(u64) * req->r_cred->group_info->ngroups);
2506 /* calculate (max) length for cap releases */
2507 len += sizeof(struct ceph_mds_request_release) *
2508 (!!req->r_inode_drop + !!req->r_dentry_drop +
2509 !!req->r_old_inode_drop + !!req->r_old_dentry_drop);
2511 if (req->r_dentry_drop)
2513 if (req->r_old_dentry_drop)
2516 msg = ceph_msg_new2(CEPH_MSG_CLIENT_REQUEST, len, 1, GFP_NOFS, false);
2518 msg = ERR_PTR(-ENOMEM);
2522 msg->hdr.tid = cpu_to_le64(req->r_tid);
2525 * The old ceph_mds_request_head didn't contain a version field, and
2526 * one was added when we moved the message version from 3->4.
2529 msg->hdr.version = cpu_to_le16(3);
2530 head = msg->front.iov_base;
2531 p = msg->front.iov_base + sizeof(*head);
2533 struct ceph_mds_request_head *new_head = msg->front.iov_base;
2535 msg->hdr.version = cpu_to_le16(4);
2536 new_head->version = cpu_to_le16(CEPH_MDS_REQUEST_HEAD_VERSION);
2537 head = (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2538 p = msg->front.iov_base + sizeof(*new_head);
2541 end = msg->front.iov_base + msg->front.iov_len;
2543 head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
2544 head->op = cpu_to_le32(req->r_op);
2545 head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
2546 req->r_cred->fsuid));
2547 head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
2548 req->r_cred->fsgid));
2549 head->ino = cpu_to_le64(req->r_deleg_ino);
2550 head->args = req->r_args;
2552 ceph_encode_filepath(&p, end, ino1, path1);
2553 ceph_encode_filepath(&p, end, ino2, path2);
2555 /* make note of release offset, in case we need to replay */
2556 req->r_request_release_offset = p - msg->front.iov_base;
2560 if (req->r_inode_drop)
2561 releases += ceph_encode_inode_release(&p,
2562 req->r_inode ? req->r_inode : d_inode(req->r_dentry),
2563 mds, req->r_inode_drop, req->r_inode_unless,
2564 req->r_op == CEPH_MDS_OP_READDIR);
2565 if (req->r_dentry_drop)
2566 releases += ceph_encode_dentry_release(&p, req->r_dentry,
2567 req->r_parent, mds, req->r_dentry_drop,
2568 req->r_dentry_unless);
2569 if (req->r_old_dentry_drop)
2570 releases += ceph_encode_dentry_release(&p, req->r_old_dentry,
2571 req->r_old_dentry_dir, mds,
2572 req->r_old_dentry_drop,
2573 req->r_old_dentry_unless);
2574 if (req->r_old_inode_drop)
2575 releases += ceph_encode_inode_release(&p,
2576 d_inode(req->r_old_dentry),
2577 mds, req->r_old_inode_drop, req->r_old_inode_unless, 0);
2579 if (drop_cap_releases) {
2581 p = msg->front.iov_base + req->r_request_release_offset;
2584 head->num_releases = cpu_to_le16(releases);
2586 encode_timestamp_and_gids(&p, req);
2588 if (WARN_ON_ONCE(p > end)) {
2590 msg = ERR_PTR(-ERANGE);
2594 msg->front.iov_len = p - msg->front.iov_base;
2595 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2597 if (req->r_pagelist) {
2598 struct ceph_pagelist *pagelist = req->r_pagelist;
2599 ceph_msg_data_add_pagelist(msg, pagelist);
2600 msg->hdr.data_len = cpu_to_le32(pagelist->length);
2602 msg->hdr.data_len = 0;
2605 msg->hdr.data_off = cpu_to_le16(0);
2609 ceph_mdsc_free_path((char *)path2, pathlen2);
2612 ceph_mdsc_free_path((char *)path1, pathlen1);
2618 * called under mdsc->mutex if error, under no mutex if
2621 static void complete_request(struct ceph_mds_client *mdsc,
2622 struct ceph_mds_request *req)
2624 req->r_end_latency = ktime_get();
2626 if (req->r_callback)
2627 req->r_callback(mdsc, req);
2628 complete_all(&req->r_completion);
2631 static struct ceph_mds_request_head_old *
2632 find_old_request_head(void *p, u64 features)
2634 bool legacy = !(features & CEPH_FEATURE_FS_BTIME);
2635 struct ceph_mds_request_head *new_head;
2638 return (struct ceph_mds_request_head_old *)p;
2639 new_head = (struct ceph_mds_request_head *)p;
2640 return (struct ceph_mds_request_head_old *)&new_head->oldest_client_tid;
2644 * called under mdsc->mutex
2646 static int __prepare_send_request(struct ceph_mds_session *session,
2647 struct ceph_mds_request *req,
2648 bool drop_cap_releases)
2650 int mds = session->s_mds;
2651 struct ceph_mds_client *mdsc = session->s_mdsc;
2652 struct ceph_mds_request_head_old *rhead;
2653 struct ceph_msg *msg;
2654 int flags = 0, max_retry;
2657 * The type of 'r_attempts' in kernel 'ceph_mds_request'
2658 * is 'int', while in 'ceph_mds_request_head' the type of
2659 * 'num_retry' is '__u8'. So in case the request retries
2660 * exceeding 256 times, the MDS will receive a incorrect
2663 * In this case it's ususally a bug in MDS and continue
2664 * retrying the request makes no sense.
2666 * In future this could be fixed in ceph code, so avoid
2667 * using the hardcode here.
2669 max_retry = sizeof_field(struct ceph_mds_request_head, num_retry);
2670 max_retry = 1 << (max_retry * BITS_PER_BYTE);
2671 if (req->r_attempts >= max_retry) {
2672 pr_warn_ratelimited("%s request tid %llu seq overflow\n",
2673 __func__, req->r_tid);
2679 struct ceph_cap *cap =
2680 ceph_get_cap_for_mds(ceph_inode(req->r_inode), mds);
2683 req->r_sent_on_mseq = cap->mseq;
2685 req->r_sent_on_mseq = -1;
2687 dout("%s %p tid %lld %s (attempt %d)\n", __func__, req,
2688 req->r_tid, ceph_mds_op_name(req->r_op), req->r_attempts);
2690 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
2694 * Replay. Do not regenerate message (and rebuild
2695 * paths, etc.); just use the original message.
2696 * Rebuilding paths will break for renames because
2697 * d_move mangles the src name.
2699 msg = req->r_request;
2700 rhead = find_old_request_head(msg->front.iov_base,
2701 session->s_con.peer_features);
2703 flags = le32_to_cpu(rhead->flags);
2704 flags |= CEPH_MDS_FLAG_REPLAY;
2705 rhead->flags = cpu_to_le32(flags);
2707 if (req->r_target_inode)
2708 rhead->ino = cpu_to_le64(ceph_ino(req->r_target_inode));
2710 rhead->num_retry = req->r_attempts - 1;
2712 /* remove cap/dentry releases from message */
2713 rhead->num_releases = 0;
2715 p = msg->front.iov_base + req->r_request_release_offset;
2716 encode_timestamp_and_gids(&p, req);
2718 msg->front.iov_len = p - msg->front.iov_base;
2719 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2723 if (req->r_request) {
2724 ceph_msg_put(req->r_request);
2725 req->r_request = NULL;
2727 msg = create_request_message(session, req, drop_cap_releases);
2729 req->r_err = PTR_ERR(msg);
2730 return PTR_ERR(msg);
2732 req->r_request = msg;
2734 rhead = find_old_request_head(msg->front.iov_base,
2735 session->s_con.peer_features);
2736 rhead->oldest_client_tid = cpu_to_le64(__get_oldest_tid(mdsc));
2737 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2738 flags |= CEPH_MDS_FLAG_REPLAY;
2739 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags))
2740 flags |= CEPH_MDS_FLAG_ASYNC;
2742 flags |= CEPH_MDS_FLAG_WANT_DENTRY;
2743 rhead->flags = cpu_to_le32(flags);
2744 rhead->num_fwd = req->r_num_fwd;
2745 rhead->num_retry = req->r_attempts - 1;
2747 dout(" r_parent = %p\n", req->r_parent);
2752 * called under mdsc->mutex
2754 static int __send_request(struct ceph_mds_session *session,
2755 struct ceph_mds_request *req,
2756 bool drop_cap_releases)
2760 err = __prepare_send_request(session, req, drop_cap_releases);
2762 ceph_msg_get(req->r_request);
2763 ceph_con_send(&session->s_con, req->r_request);
2770 * send request, or put it on the appropriate wait list.
2772 static void __do_request(struct ceph_mds_client *mdsc,
2773 struct ceph_mds_request *req)
2775 struct ceph_mds_session *session = NULL;
2780 if (req->r_err || test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
2781 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags))
2782 __unregister_request(mdsc, req);
2786 if (req->r_timeout &&
2787 time_after_eq(jiffies, req->r_started + req->r_timeout)) {
2788 dout("do_request timed out\n");
2792 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
2793 dout("do_request forced umount\n");
2797 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_MOUNTING) {
2798 if (mdsc->mdsmap_err) {
2799 err = mdsc->mdsmap_err;
2800 dout("do_request mdsmap err %d\n", err);
2803 if (mdsc->mdsmap->m_epoch == 0) {
2804 dout("do_request no mdsmap, waiting for map\n");
2805 list_add(&req->r_wait, &mdsc->waiting_for_map);
2808 if (!(mdsc->fsc->mount_options->flags &
2809 CEPH_MOUNT_OPT_MOUNTWAIT) &&
2810 !ceph_mdsmap_is_cluster_available(mdsc->mdsmap)) {
2811 err = -EHOSTUNREACH;
2816 put_request_session(req);
2818 mds = __choose_mds(mdsc, req, &random);
2820 ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
2821 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2825 dout("do_request no mds or not active, waiting for map\n");
2826 list_add(&req->r_wait, &mdsc->waiting_for_map);
2830 /* get, open session */
2831 session = __ceph_lookup_mds_session(mdsc, mds);
2833 session = register_session(mdsc, mds);
2834 if (IS_ERR(session)) {
2835 err = PTR_ERR(session);
2839 req->r_session = ceph_get_mds_session(session);
2841 dout("do_request mds%d session %p state %s\n", mds, session,
2842 ceph_session_state_name(session->s_state));
2843 if (session->s_state != CEPH_MDS_SESSION_OPEN &&
2844 session->s_state != CEPH_MDS_SESSION_HUNG) {
2846 * We cannot queue async requests since the caps and delegated
2847 * inodes are bound to the session. Just return -EJUKEBOX and
2848 * let the caller retry a sync request in that case.
2850 if (test_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags)) {
2856 * If the session has been REJECTED, then return a hard error,
2857 * unless it's a CLEANRECOVER mount, in which case we'll queue
2858 * it to the mdsc queue.
2860 if (session->s_state == CEPH_MDS_SESSION_REJECTED) {
2861 if (ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER))
2862 list_add(&req->r_wait, &mdsc->waiting_for_map);
2868 if (session->s_state == CEPH_MDS_SESSION_NEW ||
2869 session->s_state == CEPH_MDS_SESSION_CLOSING) {
2870 err = __open_session(mdsc, session);
2873 /* retry the same mds later */
2875 req->r_resend_mds = mds;
2877 list_add(&req->r_wait, &session->s_waiting);
2882 req->r_resend_mds = -1; /* forget any previous mds hint */
2884 if (req->r_request_started == 0) /* note request start time */
2885 req->r_request_started = jiffies;
2887 err = __send_request(session, req, false);
2890 ceph_put_mds_session(session);
2893 dout("__do_request early error %d\n", err);
2895 complete_request(mdsc, req);
2896 __unregister_request(mdsc, req);
2902 * called under mdsc->mutex
2904 static void __wake_requests(struct ceph_mds_client *mdsc,
2905 struct list_head *head)
2907 struct ceph_mds_request *req;
2908 LIST_HEAD(tmp_list);
2910 list_splice_init(head, &tmp_list);
2912 while (!list_empty(&tmp_list)) {
2913 req = list_entry(tmp_list.next,
2914 struct ceph_mds_request, r_wait);
2915 list_del_init(&req->r_wait);
2916 dout(" wake request %p tid %llu\n", req, req->r_tid);
2917 __do_request(mdsc, req);
2922 * Wake up threads with requests pending for @mds, so that they can
2923 * resubmit their requests to a possibly different mds.
2925 static void kick_requests(struct ceph_mds_client *mdsc, int mds)
2927 struct ceph_mds_request *req;
2928 struct rb_node *p = rb_first(&mdsc->request_tree);
2930 dout("kick_requests mds%d\n", mds);
2932 req = rb_entry(p, struct ceph_mds_request, r_node);
2934 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
2936 if (req->r_attempts > 0)
2937 continue; /* only new requests */
2938 if (req->r_session &&
2939 req->r_session->s_mds == mds) {
2940 dout(" kicking tid %llu\n", req->r_tid);
2941 list_del_init(&req->r_wait);
2942 __do_request(mdsc, req);
2947 int ceph_mdsc_submit_request(struct ceph_mds_client *mdsc, struct inode *dir,
2948 struct ceph_mds_request *req)
2952 /* take CAP_PIN refs for r_inode, r_parent, r_old_dentry */
2954 ceph_get_cap_refs(ceph_inode(req->r_inode), CEPH_CAP_PIN);
2955 if (req->r_parent) {
2956 struct ceph_inode_info *ci = ceph_inode(req->r_parent);
2957 int fmode = (req->r_op & CEPH_MDS_OP_WRITE) ?
2958 CEPH_FILE_MODE_WR : CEPH_FILE_MODE_RD;
2959 spin_lock(&ci->i_ceph_lock);
2960 ceph_take_cap_refs(ci, CEPH_CAP_PIN, false);
2961 __ceph_touch_fmode(ci, mdsc, fmode);
2962 spin_unlock(&ci->i_ceph_lock);
2964 if (req->r_old_dentry_dir)
2965 ceph_get_cap_refs(ceph_inode(req->r_old_dentry_dir),
2969 err = ceph_wait_on_async_create(req->r_inode);
2971 dout("%s: wait for async create returned: %d\n",
2977 if (!err && req->r_old_inode) {
2978 err = ceph_wait_on_async_create(req->r_old_inode);
2980 dout("%s: wait for async create returned: %d\n",
2986 dout("submit_request on %p for inode %p\n", req, dir);
2987 mutex_lock(&mdsc->mutex);
2988 __register_request(mdsc, req, dir);
2989 __do_request(mdsc, req);
2991 mutex_unlock(&mdsc->mutex);
2995 int ceph_mdsc_wait_request(struct ceph_mds_client *mdsc,
2996 struct ceph_mds_request *req,
2997 ceph_mds_request_wait_callback_t wait_func)
3002 dout("do_request waiting\n");
3004 err = wait_func(mdsc, req);
3006 long timeleft = wait_for_completion_killable_timeout(
3008 ceph_timeout_jiffies(req->r_timeout));
3012 err = -ETIMEDOUT; /* timed out */
3014 err = timeleft; /* killed */
3016 dout("do_request waited, got %d\n", err);
3017 mutex_lock(&mdsc->mutex);
3019 /* only abort if we didn't race with a real reply */
3020 if (test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags)) {
3021 err = le32_to_cpu(req->r_reply_info.head->result);
3022 } else if (err < 0) {
3023 dout("aborted request %lld with %d\n", req->r_tid, err);
3026 * ensure we aren't running concurrently with
3027 * ceph_fill_trace or ceph_readdir_prepopulate, which
3028 * rely on locks (dir mutex) held by our caller.
3030 mutex_lock(&req->r_fill_mutex);
3032 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3033 mutex_unlock(&req->r_fill_mutex);
3035 if (req->r_parent &&
3036 (req->r_op & CEPH_MDS_OP_WRITE))
3037 ceph_invalidate_dir_request(req);
3042 mutex_unlock(&mdsc->mutex);
3047 * Synchrously perform an mds request. Take care of all of the
3048 * session setup, forwarding, retry details.
3050 int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
3052 struct ceph_mds_request *req)
3056 dout("do_request on %p\n", req);
3059 err = ceph_mdsc_submit_request(mdsc, dir, req);
3061 err = ceph_mdsc_wait_request(mdsc, req, NULL);
3062 dout("do_request %p done, result %d\n", req, err);
3067 * Invalidate dir's completeness, dentry lease state on an aborted MDS
3068 * namespace request.
3070 void ceph_invalidate_dir_request(struct ceph_mds_request *req)
3072 struct inode *dir = req->r_parent;
3073 struct inode *old_dir = req->r_old_dentry_dir;
3075 dout("invalidate_dir_request %p %p (complete, lease(s))\n", dir, old_dir);
3077 ceph_dir_clear_complete(dir);
3079 ceph_dir_clear_complete(old_dir);
3081 ceph_invalidate_dentry_lease(req->r_dentry);
3082 if (req->r_old_dentry)
3083 ceph_invalidate_dentry_lease(req->r_old_dentry);
3089 * We take the session mutex and parse and process the reply immediately.
3090 * This preserves the logical ordering of replies, capabilities, etc., sent
3091 * by the MDS as they are applied to our local cache.
3093 static void handle_reply(struct ceph_mds_session *session, struct ceph_msg *msg)
3095 struct ceph_mds_client *mdsc = session->s_mdsc;
3096 struct ceph_mds_request *req;
3097 struct ceph_mds_reply_head *head = msg->front.iov_base;
3098 struct ceph_mds_reply_info_parsed *rinfo; /* parsed reply info */
3099 struct ceph_snap_realm *realm;
3102 int mds = session->s_mds;
3104 if (msg->front.iov_len < sizeof(*head)) {
3105 pr_err("mdsc_handle_reply got corrupt (short) reply\n");
3110 /* get request, session */
3111 tid = le64_to_cpu(msg->hdr.tid);
3112 mutex_lock(&mdsc->mutex);
3113 req = lookup_get_request(mdsc, tid);
3115 dout("handle_reply on unknown tid %llu\n", tid);
3116 mutex_unlock(&mdsc->mutex);
3119 dout("handle_reply %p\n", req);
3121 /* correct session? */
3122 if (req->r_session != session) {
3123 pr_err("mdsc_handle_reply got %llu on session mds%d"
3124 " not mds%d\n", tid, session->s_mds,
3125 req->r_session ? req->r_session->s_mds : -1);
3126 mutex_unlock(&mdsc->mutex);
3131 if ((test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags) && !head->safe) ||
3132 (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags) && head->safe)) {
3133 pr_warn("got a dup %s reply on %llu from mds%d\n",
3134 head->safe ? "safe" : "unsafe", tid, mds);
3135 mutex_unlock(&mdsc->mutex);
3138 if (test_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags)) {
3139 pr_warn("got unsafe after safe on %llu from mds%d\n",
3141 mutex_unlock(&mdsc->mutex);
3145 result = le32_to_cpu(head->result);
3148 set_bit(CEPH_MDS_R_GOT_SAFE, &req->r_req_flags);
3149 __unregister_request(mdsc, req);
3151 /* last request during umount? */
3152 if (mdsc->stopping && !__get_oldest_req(mdsc))
3153 complete_all(&mdsc->safe_umount_waiters);
3155 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3157 * We already handled the unsafe response, now do the
3158 * cleanup. No need to examine the response; the MDS
3159 * doesn't include any result info in the safe
3160 * response. And even if it did, there is nothing
3161 * useful we could do with a revised return value.
3163 dout("got safe reply %llu, mds%d\n", tid, mds);
3165 mutex_unlock(&mdsc->mutex);
3169 set_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags);
3170 list_add_tail(&req->r_unsafe_item, &req->r_session->s_unsafe);
3173 dout("handle_reply tid %lld result %d\n", tid, result);
3174 rinfo = &req->r_reply_info;
3175 if (test_bit(CEPHFS_FEATURE_REPLY_ENCODING, &session->s_features))
3176 err = parse_reply_info(session, msg, rinfo, (u64)-1);
3178 err = parse_reply_info(session, msg, rinfo, session->s_con.peer_features);
3179 mutex_unlock(&mdsc->mutex);
3181 /* Must find target inode outside of mutexes to avoid deadlocks */
3182 if ((err >= 0) && rinfo->head->is_target) {
3184 struct ceph_vino tvino = {
3185 .ino = le64_to_cpu(rinfo->targeti.in->ino),
3186 .snap = le64_to_cpu(rinfo->targeti.in->snapid)
3189 in = ceph_get_inode(mdsc->fsc->sb, tvino);
3192 mutex_lock(&session->s_mutex);
3195 req->r_target_inode = in;
3198 mutex_lock(&session->s_mutex);
3200 pr_err("mdsc_handle_reply got corrupt reply mds%d(tid:%lld)\n", mds, tid);
3207 if (rinfo->snapblob_len) {
3208 down_write(&mdsc->snap_rwsem);
3209 ceph_update_snap_trace(mdsc, rinfo->snapblob,
3210 rinfo->snapblob + rinfo->snapblob_len,
3211 le32_to_cpu(head->op) == CEPH_MDS_OP_RMSNAP,
3213 downgrade_write(&mdsc->snap_rwsem);
3215 down_read(&mdsc->snap_rwsem);
3218 /* insert trace into our cache */
3219 mutex_lock(&req->r_fill_mutex);
3220 current->journal_info = req;
3221 err = ceph_fill_trace(mdsc->fsc->sb, req);
3223 if (result == 0 && (req->r_op == CEPH_MDS_OP_READDIR ||
3224 req->r_op == CEPH_MDS_OP_LSSNAP))
3225 ceph_readdir_prepopulate(req, req->r_session);
3227 current->journal_info = NULL;
3228 mutex_unlock(&req->r_fill_mutex);
3230 up_read(&mdsc->snap_rwsem);
3232 ceph_put_snap_realm(mdsc, realm);
3235 if (req->r_target_inode &&
3236 test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags)) {
3237 struct ceph_inode_info *ci =
3238 ceph_inode(req->r_target_inode);
3239 spin_lock(&ci->i_unsafe_lock);
3240 list_add_tail(&req->r_unsafe_target_item,
3241 &ci->i_unsafe_iops);
3242 spin_unlock(&ci->i_unsafe_lock);
3245 ceph_unreserve_caps(mdsc, &req->r_caps_reservation);
3248 mutex_lock(&mdsc->mutex);
3249 if (!test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3253 req->r_reply = ceph_msg_get(msg);
3254 set_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags);
3257 dout("reply arrived after request %lld was aborted\n", tid);
3259 mutex_unlock(&mdsc->mutex);
3261 mutex_unlock(&session->s_mutex);
3263 /* kick calling process */
3264 complete_request(mdsc, req);
3266 ceph_update_metadata_metrics(&mdsc->metric, req->r_start_latency,
3267 req->r_end_latency, err);
3269 ceph_mdsc_put_request(req);
3276 * handle mds notification that our request has been forwarded.
3278 static void handle_forward(struct ceph_mds_client *mdsc,
3279 struct ceph_mds_session *session,
3280 struct ceph_msg *msg)
3282 struct ceph_mds_request *req;
3283 u64 tid = le64_to_cpu(msg->hdr.tid);
3287 void *p = msg->front.iov_base;
3288 void *end = p + msg->front.iov_len;
3289 bool aborted = false;
3291 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
3292 next_mds = ceph_decode_32(&p);
3293 fwd_seq = ceph_decode_32(&p);
3295 mutex_lock(&mdsc->mutex);
3296 req = lookup_get_request(mdsc, tid);
3298 mutex_unlock(&mdsc->mutex);
3299 dout("forward tid %llu to mds%d - req dne\n", tid, next_mds);
3300 return; /* dup reply? */
3303 if (test_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags)) {
3304 dout("forward tid %llu aborted, unregistering\n", tid);
3305 __unregister_request(mdsc, req);
3306 } else if (fwd_seq <= req->r_num_fwd) {
3308 * The type of 'num_fwd' in ceph 'MClientRequestForward'
3309 * is 'int32_t', while in 'ceph_mds_request_head' the
3310 * type is '__u8'. So in case the request bounces between
3311 * MDSes exceeding 256 times, the client will get stuck.
3313 * In this case it's ususally a bug in MDS and continue
3314 * bouncing the request makes no sense.
3316 * In future this could be fixed in ceph code, so avoid
3317 * using the hardcode here.
3319 int max = sizeof_field(struct ceph_mds_request_head, num_fwd);
3320 max = 1 << (max * BITS_PER_BYTE);
3321 if (req->r_num_fwd >= max) {
3322 mutex_lock(&req->r_fill_mutex);
3323 req->r_err = -EMULTIHOP;
3324 set_bit(CEPH_MDS_R_ABORTED, &req->r_req_flags);
3325 mutex_unlock(&req->r_fill_mutex);
3327 pr_warn_ratelimited("forward tid %llu seq overflow\n",
3330 dout("forward tid %llu to mds%d - old seq %d <= %d\n",
3331 tid, next_mds, req->r_num_fwd, fwd_seq);
3334 /* resend. forward race not possible; mds would drop */
3335 dout("forward tid %llu to mds%d (we resend)\n", tid, next_mds);
3337 BUG_ON(test_bit(CEPH_MDS_R_GOT_RESULT, &req->r_req_flags));
3338 req->r_attempts = 0;
3339 req->r_num_fwd = fwd_seq;
3340 req->r_resend_mds = next_mds;
3341 put_request_session(req);
3342 __do_request(mdsc, req);
3344 mutex_unlock(&mdsc->mutex);
3346 /* kick calling process */
3348 complete_request(mdsc, req);
3349 ceph_mdsc_put_request(req);
3353 pr_err("mdsc_handle_forward decode error err=%d\n", err);
3356 static int __decode_session_metadata(void **p, void *end,
3359 /* map<string,string> */
3362 ceph_decode_32_safe(p, end, n, bad);
3365 ceph_decode_32_safe(p, end, len, bad);
3366 ceph_decode_need(p, end, len, bad);
3367 err_str = !strncmp(*p, "error_string", len);
3369 ceph_decode_32_safe(p, end, len, bad);
3370 ceph_decode_need(p, end, len, bad);
3372 * Match "blocklisted (blacklisted)" from newer MDSes,
3373 * or "blacklisted" from older MDSes.
3375 if (err_str && strnstr(*p, "blacklisted", len))
3376 *blocklisted = true;
3385 * handle a mds session control message
3387 static void handle_session(struct ceph_mds_session *session,
3388 struct ceph_msg *msg)
3390 struct ceph_mds_client *mdsc = session->s_mdsc;
3391 int mds = session->s_mds;
3392 int msg_version = le16_to_cpu(msg->hdr.version);
3393 void *p = msg->front.iov_base;
3394 void *end = p + msg->front.iov_len;
3395 struct ceph_mds_session_head *h;
3397 u64 seq, features = 0;
3399 bool blocklisted = false;
3402 ceph_decode_need(&p, end, sizeof(*h), bad);
3406 op = le32_to_cpu(h->op);
3407 seq = le64_to_cpu(h->seq);
3409 if (msg_version >= 3) {
3411 /* version >= 2 and < 5, decode metadata, skip otherwise
3412 * as it's handled via flags.
3414 if (msg_version >= 5)
3415 ceph_decode_skip_map(&p, end, string, string, bad);
3416 else if (__decode_session_metadata(&p, end, &blocklisted) < 0)
3419 /* version >= 3, feature bits */
3420 ceph_decode_32_safe(&p, end, len, bad);
3422 ceph_decode_64_safe(&p, end, features, bad);
3423 p += len - sizeof(features);
3427 if (msg_version >= 5) {
3431 ceph_decode_skip_16(&p, end, bad); /* struct_v, struct_cv */
3432 ceph_decode_32_safe(&p, end, len, bad); /* len */
3433 ceph_decode_skip_n(&p, end, len, bad); /* metric_spec */
3435 /* version >= 5, flags */
3436 ceph_decode_32_safe(&p, end, flags, bad);
3437 if (flags & CEPH_SESSION_BLOCKLISTED) {
3438 pr_warn("mds%d session blocklisted\n", session->s_mds);
3443 mutex_lock(&mdsc->mutex);
3444 if (op == CEPH_SESSION_CLOSE) {
3445 ceph_get_mds_session(session);
3446 __unregister_session(mdsc, session);
3448 /* FIXME: this ttl calculation is generous */
3449 session->s_ttl = jiffies + HZ*mdsc->mdsmap->m_session_autoclose;
3450 mutex_unlock(&mdsc->mutex);
3452 mutex_lock(&session->s_mutex);
3454 dout("handle_session mds%d %s %p state %s seq %llu\n",
3455 mds, ceph_session_op_name(op), session,
3456 ceph_session_state_name(session->s_state), seq);
3458 if (session->s_state == CEPH_MDS_SESSION_HUNG) {
3459 session->s_state = CEPH_MDS_SESSION_OPEN;
3460 pr_info("mds%d came back\n", session->s_mds);
3464 case CEPH_SESSION_OPEN:
3465 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3466 pr_info("mds%d reconnect success\n", session->s_mds);
3467 session->s_state = CEPH_MDS_SESSION_OPEN;
3468 session->s_features = features;
3469 renewed_caps(mdsc, session, 0);
3470 if (test_bit(CEPHFS_FEATURE_METRIC_COLLECT, &session->s_features))
3471 metric_schedule_delayed(&mdsc->metric);
3474 __close_session(mdsc, session);
3477 case CEPH_SESSION_RENEWCAPS:
3478 if (session->s_renew_seq == seq)
3479 renewed_caps(mdsc, session, 1);
3482 case CEPH_SESSION_CLOSE:
3483 if (session->s_state == CEPH_MDS_SESSION_RECONNECTING)
3484 pr_info("mds%d reconnect denied\n", session->s_mds);
3485 session->s_state = CEPH_MDS_SESSION_CLOSED;
3486 cleanup_session_requests(mdsc, session);
3487 remove_session_caps(session);
3488 wake = 2; /* for good measure */
3489 wake_up_all(&mdsc->session_close_wq);
3492 case CEPH_SESSION_STALE:
3493 pr_info("mds%d caps went stale, renewing\n",
3495 atomic_inc(&session->s_cap_gen);
3496 session->s_cap_ttl = jiffies - 1;
3497 send_renew_caps(mdsc, session);
3500 case CEPH_SESSION_RECALL_STATE:
3501 ceph_trim_caps(mdsc, session, le32_to_cpu(h->max_caps));
3504 case CEPH_SESSION_FLUSHMSG:
3505 send_flushmsg_ack(mdsc, session, seq);
3508 case CEPH_SESSION_FORCE_RO:
3509 dout("force_session_readonly %p\n", session);
3510 spin_lock(&session->s_cap_lock);
3511 session->s_readonly = true;
3512 spin_unlock(&session->s_cap_lock);
3513 wake_up_session_caps(session, FORCE_RO);
3516 case CEPH_SESSION_REJECT:
3517 WARN_ON(session->s_state != CEPH_MDS_SESSION_OPENING);
3518 pr_info("mds%d rejected session\n", session->s_mds);
3519 session->s_state = CEPH_MDS_SESSION_REJECTED;
3520 cleanup_session_requests(mdsc, session);
3521 remove_session_caps(session);
3523 mdsc->fsc->blocklisted = true;
3524 wake = 2; /* for good measure */
3528 pr_err("mdsc_handle_session bad op %d mds%d\n", op, mds);
3532 mutex_unlock(&session->s_mutex);
3534 mutex_lock(&mdsc->mutex);
3535 __wake_requests(mdsc, &session->s_waiting);
3537 kick_requests(mdsc, mds);
3538 mutex_unlock(&mdsc->mutex);
3540 if (op == CEPH_SESSION_CLOSE)
3541 ceph_put_mds_session(session);
3545 pr_err("mdsc_handle_session corrupt message mds%d len %d\n", mds,
3546 (int)msg->front.iov_len);
3551 void ceph_mdsc_release_dir_caps(struct ceph_mds_request *req)
3555 dcaps = xchg(&req->r_dir_caps, 0);
3557 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3558 ceph_put_cap_refs(ceph_inode(req->r_parent), dcaps);
3562 void ceph_mdsc_release_dir_caps_no_check(struct ceph_mds_request *req)
3566 dcaps = xchg(&req->r_dir_caps, 0);
3568 dout("releasing r_dir_caps=%s\n", ceph_cap_string(dcaps));
3569 ceph_put_cap_refs_no_check_caps(ceph_inode(req->r_parent),
3575 * called under session->mutex.
3577 static void replay_unsafe_requests(struct ceph_mds_client *mdsc,
3578 struct ceph_mds_session *session)
3580 struct ceph_mds_request *req, *nreq;
3583 dout("replay_unsafe_requests mds%d\n", session->s_mds);
3585 mutex_lock(&mdsc->mutex);
3586 list_for_each_entry_safe(req, nreq, &session->s_unsafe, r_unsafe_item)
3587 __send_request(session, req, true);
3590 * also re-send old requests when MDS enters reconnect stage. So that MDS
3591 * can process completed request in clientreplay stage.
3593 p = rb_first(&mdsc->request_tree);
3595 req = rb_entry(p, struct ceph_mds_request, r_node);
3597 if (test_bit(CEPH_MDS_R_GOT_UNSAFE, &req->r_req_flags))
3599 if (req->r_attempts == 0)
3600 continue; /* only old requests */
3601 if (!req->r_session)
3603 if (req->r_session->s_mds != session->s_mds)
3606 ceph_mdsc_release_dir_caps_no_check(req);
3608 __send_request(session, req, true);
3610 mutex_unlock(&mdsc->mutex);
3613 static int send_reconnect_partial(struct ceph_reconnect_state *recon_state)
3615 struct ceph_msg *reply;
3616 struct ceph_pagelist *_pagelist;
3621 if (!recon_state->allow_multi)
3624 /* can't handle message that contains both caps and realm */
3625 BUG_ON(!recon_state->nr_caps == !recon_state->nr_realms);
3627 /* pre-allocate new pagelist */
3628 _pagelist = ceph_pagelist_alloc(GFP_NOFS);
3632 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3636 /* placeholder for nr_caps */
3637 err = ceph_pagelist_encode_32(_pagelist, 0);
3641 if (recon_state->nr_caps) {
3642 /* currently encoding caps */
3643 err = ceph_pagelist_encode_32(recon_state->pagelist, 0);
3647 /* placeholder for nr_realms (currently encoding relams) */
3648 err = ceph_pagelist_encode_32(_pagelist, 0);
3653 err = ceph_pagelist_encode_8(recon_state->pagelist, 1);
3657 page = list_first_entry(&recon_state->pagelist->head, struct page, lru);
3658 addr = kmap_atomic(page);
3659 if (recon_state->nr_caps) {
3660 /* currently encoding caps */
3661 *addr = cpu_to_le32(recon_state->nr_caps);
3663 /* currently encoding relams */
3664 *(addr + 1) = cpu_to_le32(recon_state->nr_realms);
3666 kunmap_atomic(addr);
3668 reply->hdr.version = cpu_to_le16(5);
3669 reply->hdr.compat_version = cpu_to_le16(4);
3671 reply->hdr.data_len = cpu_to_le32(recon_state->pagelist->length);
3672 ceph_msg_data_add_pagelist(reply, recon_state->pagelist);
3674 ceph_con_send(&recon_state->session->s_con, reply);
3675 ceph_pagelist_release(recon_state->pagelist);
3677 recon_state->pagelist = _pagelist;
3678 recon_state->nr_caps = 0;
3679 recon_state->nr_realms = 0;
3680 recon_state->msg_version = 5;
3683 ceph_msg_put(reply);
3685 ceph_pagelist_release(_pagelist);
3689 static struct dentry* d_find_primary(struct inode *inode)
3691 struct dentry *alias, *dn = NULL;
3693 if (hlist_empty(&inode->i_dentry))
3696 spin_lock(&inode->i_lock);
3697 if (hlist_empty(&inode->i_dentry))
3700 if (S_ISDIR(inode->i_mode)) {
3701 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
3702 if (!IS_ROOT(alias))
3707 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
3708 spin_lock(&alias->d_lock);
3709 if (!d_unhashed(alias) &&
3710 (ceph_dentry(alias)->flags & CEPH_DENTRY_PRIMARY_LINK)) {
3711 dn = dget_dlock(alias);
3713 spin_unlock(&alias->d_lock);
3718 spin_unlock(&inode->i_lock);
3723 * Encode information about a cap for a reconnect with the MDS.
3725 static int reconnect_caps_cb(struct inode *inode, struct ceph_cap *cap,
3729 struct ceph_mds_cap_reconnect v2;
3730 struct ceph_mds_cap_reconnect_v1 v1;
3732 struct ceph_inode_info *ci = cap->ci;
3733 struct ceph_reconnect_state *recon_state = arg;
3734 struct ceph_pagelist *pagelist = recon_state->pagelist;
3735 struct dentry *dentry;
3737 int pathlen = 0, err;
3741 dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
3742 inode, ceph_vinop(inode), cap, cap->cap_id,
3743 ceph_cap_string(cap->issued));
3745 dentry = d_find_primary(inode);
3747 /* set pathbase to parent dir when msg_version >= 2 */
3748 path = ceph_mdsc_build_path(dentry, &pathlen, &pathbase,
3749 recon_state->msg_version >= 2);
3752 err = PTR_ERR(path);
3760 spin_lock(&ci->i_ceph_lock);
3761 cap->seq = 0; /* reset cap seq */
3762 cap->issue_seq = 0; /* and issue_seq */
3763 cap->mseq = 0; /* and migrate_seq */
3764 cap->cap_gen = atomic_read(&cap->session->s_cap_gen);
3766 /* These are lost when the session goes away */
3767 if (S_ISDIR(inode->i_mode)) {
3768 if (cap->issued & CEPH_CAP_DIR_CREATE) {
3769 ceph_put_string(rcu_dereference_raw(ci->i_cached_layout.pool_ns));
3770 memset(&ci->i_cached_layout, 0, sizeof(ci->i_cached_layout));
3772 cap->issued &= ~CEPH_CAP_ANY_DIR_OPS;
3775 if (recon_state->msg_version >= 2) {
3776 rec.v2.cap_id = cpu_to_le64(cap->cap_id);
3777 rec.v2.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3778 rec.v2.issued = cpu_to_le32(cap->issued);
3779 rec.v2.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3780 rec.v2.pathbase = cpu_to_le64(pathbase);
3781 rec.v2.flock_len = (__force __le32)
3782 ((ci->i_ceph_flags & CEPH_I_ERROR_FILELOCK) ? 0 : 1);
3784 rec.v1.cap_id = cpu_to_le64(cap->cap_id);
3785 rec.v1.wanted = cpu_to_le32(__ceph_caps_wanted(ci));
3786 rec.v1.issued = cpu_to_le32(cap->issued);
3787 rec.v1.size = cpu_to_le64(i_size_read(inode));
3788 ceph_encode_timespec64(&rec.v1.mtime, &inode->i_mtime);
3789 ceph_encode_timespec64(&rec.v1.atime, &inode->i_atime);
3790 rec.v1.snaprealm = cpu_to_le64(ci->i_snap_realm->ino);
3791 rec.v1.pathbase = cpu_to_le64(pathbase);
3794 if (list_empty(&ci->i_cap_snaps)) {
3795 snap_follows = ci->i_head_snapc ? ci->i_head_snapc->seq : 0;
3797 struct ceph_cap_snap *capsnap =
3798 list_first_entry(&ci->i_cap_snaps,
3799 struct ceph_cap_snap, ci_item);
3800 snap_follows = capsnap->follows;
3802 spin_unlock(&ci->i_ceph_lock);
3804 if (recon_state->msg_version >= 2) {
3805 int num_fcntl_locks, num_flock_locks;
3806 struct ceph_filelock *flocks = NULL;
3807 size_t struct_len, total_len = sizeof(u64);
3811 if (rec.v2.flock_len) {
3812 ceph_count_locks(inode, &num_fcntl_locks, &num_flock_locks);
3814 num_fcntl_locks = 0;
3815 num_flock_locks = 0;
3817 if (num_fcntl_locks + num_flock_locks > 0) {
3818 flocks = kmalloc_array(num_fcntl_locks + num_flock_locks,
3819 sizeof(struct ceph_filelock),
3825 err = ceph_encode_locks_to_buffer(inode, flocks,
3840 if (recon_state->msg_version >= 3) {
3841 /* version, compat_version and struct_len */
3842 total_len += 2 * sizeof(u8) + sizeof(u32);
3846 * number of encoded locks is stable, so copy to pagelist
3848 struct_len = 2 * sizeof(u32) +
3849 (num_fcntl_locks + num_flock_locks) *
3850 sizeof(struct ceph_filelock);
3851 rec.v2.flock_len = cpu_to_le32(struct_len);
3853 struct_len += sizeof(u32) + pathlen + sizeof(rec.v2);
3856 struct_len += sizeof(u64); /* snap_follows */
3858 total_len += struct_len;
3860 if (pagelist->length + total_len > RECONNECT_MAX_SIZE) {
3861 err = send_reconnect_partial(recon_state);
3863 goto out_freeflocks;
3864 pagelist = recon_state->pagelist;
3867 err = ceph_pagelist_reserve(pagelist, total_len);
3869 goto out_freeflocks;
3871 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3872 if (recon_state->msg_version >= 3) {
3873 ceph_pagelist_encode_8(pagelist, struct_v);
3874 ceph_pagelist_encode_8(pagelist, 1);
3875 ceph_pagelist_encode_32(pagelist, struct_len);
3877 ceph_pagelist_encode_string(pagelist, path, pathlen);
3878 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v2));
3879 ceph_locks_to_pagelist(flocks, pagelist,
3880 num_fcntl_locks, num_flock_locks);
3882 ceph_pagelist_encode_64(pagelist, snap_follows);
3886 err = ceph_pagelist_reserve(pagelist,
3887 sizeof(u64) + sizeof(u32) +
3888 pathlen + sizeof(rec.v1));
3892 ceph_pagelist_encode_64(pagelist, ceph_ino(inode));
3893 ceph_pagelist_encode_string(pagelist, path, pathlen);
3894 ceph_pagelist_append(pagelist, &rec, sizeof(rec.v1));
3898 ceph_mdsc_free_path(path, pathlen);
3900 recon_state->nr_caps++;
3904 static int encode_snap_realms(struct ceph_mds_client *mdsc,
3905 struct ceph_reconnect_state *recon_state)
3908 struct ceph_pagelist *pagelist = recon_state->pagelist;
3911 if (recon_state->msg_version >= 4) {
3912 err = ceph_pagelist_encode_32(pagelist, mdsc->num_snap_realms);
3918 * snaprealms. we provide mds with the ino, seq (version), and
3919 * parent for all of our realms. If the mds has any newer info,
3922 for (p = rb_first(&mdsc->snap_realms); p; p = rb_next(p)) {
3923 struct ceph_snap_realm *realm =
3924 rb_entry(p, struct ceph_snap_realm, node);
3925 struct ceph_mds_snaprealm_reconnect sr_rec;
3927 if (recon_state->msg_version >= 4) {
3928 size_t need = sizeof(u8) * 2 + sizeof(u32) +
3931 if (pagelist->length + need > RECONNECT_MAX_SIZE) {
3932 err = send_reconnect_partial(recon_state);
3935 pagelist = recon_state->pagelist;
3938 err = ceph_pagelist_reserve(pagelist, need);
3942 ceph_pagelist_encode_8(pagelist, 1);
3943 ceph_pagelist_encode_8(pagelist, 1);
3944 ceph_pagelist_encode_32(pagelist, sizeof(sr_rec));
3947 dout(" adding snap realm %llx seq %lld parent %llx\n",
3948 realm->ino, realm->seq, realm->parent_ino);
3949 sr_rec.ino = cpu_to_le64(realm->ino);
3950 sr_rec.seq = cpu_to_le64(realm->seq);
3951 sr_rec.parent = cpu_to_le64(realm->parent_ino);
3953 err = ceph_pagelist_append(pagelist, &sr_rec, sizeof(sr_rec));
3957 recon_state->nr_realms++;
3965 * If an MDS fails and recovers, clients need to reconnect in order to
3966 * reestablish shared state. This includes all caps issued through
3967 * this session _and_ the snap_realm hierarchy. Because it's not
3968 * clear which snap realms the mds cares about, we send everything we
3969 * know about.. that ensures we'll then get any new info the
3970 * recovering MDS might have.
3972 * This is a relatively heavyweight operation, but it's rare.
3974 static void send_mds_reconnect(struct ceph_mds_client *mdsc,
3975 struct ceph_mds_session *session)
3977 struct ceph_msg *reply;
3978 int mds = session->s_mds;
3980 struct ceph_reconnect_state recon_state = {
3985 pr_info("mds%d reconnect start\n", mds);
3987 recon_state.pagelist = ceph_pagelist_alloc(GFP_NOFS);
3988 if (!recon_state.pagelist)
3989 goto fail_nopagelist;
3991 reply = ceph_msg_new2(CEPH_MSG_CLIENT_RECONNECT, 0, 1, GFP_NOFS, false);
3995 xa_destroy(&session->s_delegated_inos);
3997 mutex_lock(&session->s_mutex);
3998 session->s_state = CEPH_MDS_SESSION_RECONNECTING;
4001 dout("session %p state %s\n", session,
4002 ceph_session_state_name(session->s_state));
4004 atomic_inc(&session->s_cap_gen);
4006 spin_lock(&session->s_cap_lock);
4007 /* don't know if session is readonly */
4008 session->s_readonly = 0;
4010 * notify __ceph_remove_cap() that we are composing cap reconnect.
4011 * If a cap get released before being added to the cap reconnect,
4012 * __ceph_remove_cap() should skip queuing cap release.
4014 session->s_cap_reconnect = 1;
4015 /* drop old cap expires; we're about to reestablish that state */
4016 detach_cap_releases(session, &dispose);
4017 spin_unlock(&session->s_cap_lock);
4018 dispose_cap_releases(mdsc, &dispose);
4020 /* trim unused caps to reduce MDS's cache rejoin time */
4021 if (mdsc->fsc->sb->s_root)
4022 shrink_dcache_parent(mdsc->fsc->sb->s_root);
4024 ceph_con_close(&session->s_con);
4025 ceph_con_open(&session->s_con,
4026 CEPH_ENTITY_TYPE_MDS, mds,
4027 ceph_mdsmap_get_addr(mdsc->mdsmap, mds));
4029 /* replay unsafe requests */
4030 replay_unsafe_requests(mdsc, session);
4032 ceph_early_kick_flushing_caps(mdsc, session);
4034 down_read(&mdsc->snap_rwsem);
4036 /* placeholder for nr_caps */
4037 err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
4041 if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
4042 recon_state.msg_version = 3;
4043 recon_state.allow_multi = true;
4044 } else if (session->s_con.peer_features & CEPH_FEATURE_MDSENC) {
4045 recon_state.msg_version = 3;
4047 recon_state.msg_version = 2;
4049 /* trsaverse this session's caps */
4050 err = ceph_iterate_session_caps(session, reconnect_caps_cb, &recon_state);
4052 spin_lock(&session->s_cap_lock);
4053 session->s_cap_reconnect = 0;
4054 spin_unlock(&session->s_cap_lock);
4059 /* check if all realms can be encoded into current message */
4060 if (mdsc->num_snap_realms) {
4062 recon_state.pagelist->length +
4063 mdsc->num_snap_realms *
4064 sizeof(struct ceph_mds_snaprealm_reconnect);
4065 if (recon_state.msg_version >= 4) {
4066 /* number of realms */
4067 total_len += sizeof(u32);
4068 /* version, compat_version and struct_len */
4069 total_len += mdsc->num_snap_realms *
4070 (2 * sizeof(u8) + sizeof(u32));
4072 if (total_len > RECONNECT_MAX_SIZE) {
4073 if (!recon_state.allow_multi) {
4077 if (recon_state.nr_caps) {
4078 err = send_reconnect_partial(&recon_state);
4082 recon_state.msg_version = 5;
4086 err = encode_snap_realms(mdsc, &recon_state);
4090 if (recon_state.msg_version >= 5) {
4091 err = ceph_pagelist_encode_8(recon_state.pagelist, 0);
4096 if (recon_state.nr_caps || recon_state.nr_realms) {
4098 list_first_entry(&recon_state.pagelist->head,
4100 __le32 *addr = kmap_atomic(page);
4101 if (recon_state.nr_caps) {
4102 WARN_ON(recon_state.nr_realms != mdsc->num_snap_realms);
4103 *addr = cpu_to_le32(recon_state.nr_caps);
4104 } else if (recon_state.msg_version >= 4) {
4105 *(addr + 1) = cpu_to_le32(recon_state.nr_realms);
4107 kunmap_atomic(addr);
4110 reply->hdr.version = cpu_to_le16(recon_state.msg_version);
4111 if (recon_state.msg_version >= 4)
4112 reply->hdr.compat_version = cpu_to_le16(4);
4114 reply->hdr.data_len = cpu_to_le32(recon_state.pagelist->length);
4115 ceph_msg_data_add_pagelist(reply, recon_state.pagelist);
4117 ceph_con_send(&session->s_con, reply);
4119 mutex_unlock(&session->s_mutex);
4121 mutex_lock(&mdsc->mutex);
4122 __wake_requests(mdsc, &session->s_waiting);
4123 mutex_unlock(&mdsc->mutex);
4125 up_read(&mdsc->snap_rwsem);
4126 ceph_pagelist_release(recon_state.pagelist);
4130 ceph_msg_put(reply);
4131 up_read(&mdsc->snap_rwsem);
4132 mutex_unlock(&session->s_mutex);
4134 ceph_pagelist_release(recon_state.pagelist);
4136 pr_err("error %d preparing reconnect for mds%d\n", err, mds);
4142 * compare old and new mdsmaps, kicking requests
4143 * and closing out old connections as necessary
4145 * called under mdsc->mutex.
4147 static void check_new_map(struct ceph_mds_client *mdsc,
4148 struct ceph_mdsmap *newmap,
4149 struct ceph_mdsmap *oldmap)
4152 int oldstate, newstate;
4153 struct ceph_mds_session *s;
4154 unsigned long targets[DIV_ROUND_UP(CEPH_MAX_MDS, sizeof(unsigned long))] = {0};
4156 dout("check_new_map new %u old %u\n",
4157 newmap->m_epoch, oldmap->m_epoch);
4159 if (newmap->m_info) {
4160 for (i = 0; i < newmap->possible_max_rank; i++) {
4161 for (j = 0; j < newmap->m_info[i].num_export_targets; j++)
4162 set_bit(newmap->m_info[i].export_targets[j], targets);
4166 for (i = 0; i < oldmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4167 if (!mdsc->sessions[i])
4169 s = mdsc->sessions[i];
4170 oldstate = ceph_mdsmap_get_state(oldmap, i);
4171 newstate = ceph_mdsmap_get_state(newmap, i);
4173 dout("check_new_map mds%d state %s%s -> %s%s (session %s)\n",
4174 i, ceph_mds_state_name(oldstate),
4175 ceph_mdsmap_is_laggy(oldmap, i) ? " (laggy)" : "",
4176 ceph_mds_state_name(newstate),
4177 ceph_mdsmap_is_laggy(newmap, i) ? " (laggy)" : "",
4178 ceph_session_state_name(s->s_state));
4180 if (i >= newmap->possible_max_rank) {
4181 /* force close session for stopped mds */
4182 ceph_get_mds_session(s);
4183 __unregister_session(mdsc, s);
4184 __wake_requests(mdsc, &s->s_waiting);
4185 mutex_unlock(&mdsc->mutex);
4187 mutex_lock(&s->s_mutex);
4188 cleanup_session_requests(mdsc, s);
4189 remove_session_caps(s);
4190 mutex_unlock(&s->s_mutex);
4192 ceph_put_mds_session(s);
4194 mutex_lock(&mdsc->mutex);
4195 kick_requests(mdsc, i);
4199 if (memcmp(ceph_mdsmap_get_addr(oldmap, i),
4200 ceph_mdsmap_get_addr(newmap, i),
4201 sizeof(struct ceph_entity_addr))) {
4203 mutex_unlock(&mdsc->mutex);
4204 mutex_lock(&s->s_mutex);
4205 mutex_lock(&mdsc->mutex);
4206 ceph_con_close(&s->s_con);
4207 mutex_unlock(&s->s_mutex);
4208 s->s_state = CEPH_MDS_SESSION_RESTARTING;
4209 } else if (oldstate == newstate) {
4210 continue; /* nothing new with this mds */
4216 if (s->s_state == CEPH_MDS_SESSION_RESTARTING &&
4217 newstate >= CEPH_MDS_STATE_RECONNECT) {
4218 mutex_unlock(&mdsc->mutex);
4219 clear_bit(i, targets);
4220 send_mds_reconnect(mdsc, s);
4221 mutex_lock(&mdsc->mutex);
4225 * kick request on any mds that has gone active.
4227 if (oldstate < CEPH_MDS_STATE_ACTIVE &&
4228 newstate >= CEPH_MDS_STATE_ACTIVE) {
4229 if (oldstate != CEPH_MDS_STATE_CREATING &&
4230 oldstate != CEPH_MDS_STATE_STARTING)
4231 pr_info("mds%d recovery completed\n", s->s_mds);
4232 kick_requests(mdsc, i);
4233 mutex_unlock(&mdsc->mutex);
4234 mutex_lock(&s->s_mutex);
4235 mutex_lock(&mdsc->mutex);
4236 ceph_kick_flushing_caps(mdsc, s);
4237 mutex_unlock(&s->s_mutex);
4238 wake_up_session_caps(s, RECONNECT);
4243 * Only open and reconnect sessions that don't exist yet.
4245 for (i = 0; i < newmap->possible_max_rank; i++) {
4247 * In case the import MDS is crashed just after
4248 * the EImportStart journal is flushed, so when
4249 * a standby MDS takes over it and is replaying
4250 * the EImportStart journal the new MDS daemon
4251 * will wait the client to reconnect it, but the
4252 * client may never register/open the session yet.
4254 * Will try to reconnect that MDS daemon if the
4255 * rank number is in the export targets array and
4256 * is the up:reconnect state.
4258 newstate = ceph_mdsmap_get_state(newmap, i);
4259 if (!test_bit(i, targets) || newstate != CEPH_MDS_STATE_RECONNECT)
4263 * The session maybe registered and opened by some
4264 * requests which were choosing random MDSes during
4265 * the mdsc->mutex's unlock/lock gap below in rare
4266 * case. But the related MDS daemon will just queue
4267 * that requests and be still waiting for the client's
4268 * reconnection request in up:reconnect state.
4270 s = __ceph_lookup_mds_session(mdsc, i);
4272 s = __open_export_target_session(mdsc, i);
4275 pr_err("failed to open export target session, err %d\n",
4280 dout("send reconnect to export target mds.%d\n", i);
4281 mutex_unlock(&mdsc->mutex);
4282 send_mds_reconnect(mdsc, s);
4283 ceph_put_mds_session(s);
4284 mutex_lock(&mdsc->mutex);
4287 for (i = 0; i < newmap->possible_max_rank && i < mdsc->max_sessions; i++) {
4288 s = mdsc->sessions[i];
4291 if (!ceph_mdsmap_is_laggy(newmap, i))
4293 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4294 s->s_state == CEPH_MDS_SESSION_HUNG ||
4295 s->s_state == CEPH_MDS_SESSION_CLOSING) {
4296 dout(" connecting to export targets of laggy mds%d\n",
4298 __open_export_target_sessions(mdsc, s);
4310 * caller must hold session s_mutex, dentry->d_lock
4312 void __ceph_mdsc_drop_dentry_lease(struct dentry *dentry)
4314 struct ceph_dentry_info *di = ceph_dentry(dentry);
4316 ceph_put_mds_session(di->lease_session);
4317 di->lease_session = NULL;
4320 static void handle_lease(struct ceph_mds_client *mdsc,
4321 struct ceph_mds_session *session,
4322 struct ceph_msg *msg)
4324 struct super_block *sb = mdsc->fsc->sb;
4325 struct inode *inode;
4326 struct dentry *parent, *dentry;
4327 struct ceph_dentry_info *di;
4328 int mds = session->s_mds;
4329 struct ceph_mds_lease *h = msg->front.iov_base;
4331 struct ceph_vino vino;
4335 dout("handle_lease from mds%d\n", mds);
4338 if (msg->front.iov_len < sizeof(*h) + sizeof(u32))
4340 vino.ino = le64_to_cpu(h->ino);
4341 vino.snap = CEPH_NOSNAP;
4342 seq = le32_to_cpu(h->seq);
4343 dname.len = get_unaligned_le32(h + 1);
4344 if (msg->front.iov_len < sizeof(*h) + sizeof(u32) + dname.len)
4346 dname.name = (void *)(h + 1) + sizeof(u32);
4349 inode = ceph_find_inode(sb, vino);
4350 dout("handle_lease %s, ino %llx %p %.*s\n",
4351 ceph_lease_op_name(h->action), vino.ino, inode,
4352 dname.len, dname.name);
4354 mutex_lock(&session->s_mutex);
4355 inc_session_sequence(session);
4358 dout("handle_lease no inode %llx\n", vino.ino);
4363 parent = d_find_alias(inode);
4365 dout("no parent dentry on inode %p\n", inode);
4367 goto release; /* hrm... */
4369 dname.hash = full_name_hash(parent, dname.name, dname.len);
4370 dentry = d_lookup(parent, &dname);
4375 spin_lock(&dentry->d_lock);
4376 di = ceph_dentry(dentry);
4377 switch (h->action) {
4378 case CEPH_MDS_LEASE_REVOKE:
4379 if (di->lease_session == session) {
4380 if (ceph_seq_cmp(di->lease_seq, seq) > 0)
4381 h->seq = cpu_to_le32(di->lease_seq);
4382 __ceph_mdsc_drop_dentry_lease(dentry);
4387 case CEPH_MDS_LEASE_RENEW:
4388 if (di->lease_session == session &&
4389 di->lease_gen == atomic_read(&session->s_cap_gen) &&
4390 di->lease_renew_from &&
4391 di->lease_renew_after == 0) {
4392 unsigned long duration =
4393 msecs_to_jiffies(le32_to_cpu(h->duration_ms));
4395 di->lease_seq = seq;
4396 di->time = di->lease_renew_from + duration;
4397 di->lease_renew_after = di->lease_renew_from +
4399 di->lease_renew_from = 0;
4403 spin_unlock(&dentry->d_lock);
4410 /* let's just reuse the same message */
4411 h->action = CEPH_MDS_LEASE_REVOKE_ACK;
4413 ceph_con_send(&session->s_con, msg);
4416 mutex_unlock(&session->s_mutex);
4421 pr_err("corrupt lease message\n");
4425 void ceph_mdsc_lease_send_msg(struct ceph_mds_session *session,
4426 struct dentry *dentry, char action,
4429 struct ceph_msg *msg;
4430 struct ceph_mds_lease *lease;
4432 int len = sizeof(*lease) + sizeof(u32) + NAME_MAX;
4434 dout("lease_send_msg identry %p %s to mds%d\n",
4435 dentry, ceph_lease_op_name(action), session->s_mds);
4437 msg = ceph_msg_new(CEPH_MSG_CLIENT_LEASE, len, GFP_NOFS, false);
4440 lease = msg->front.iov_base;
4441 lease->action = action;
4442 lease->seq = cpu_to_le32(seq);
4444 spin_lock(&dentry->d_lock);
4445 dir = d_inode(dentry->d_parent);
4446 lease->ino = cpu_to_le64(ceph_ino(dir));
4447 lease->first = lease->last = cpu_to_le64(ceph_snap(dir));
4449 put_unaligned_le32(dentry->d_name.len, lease + 1);
4450 memcpy((void *)(lease + 1) + 4,
4451 dentry->d_name.name, dentry->d_name.len);
4452 spin_unlock(&dentry->d_lock);
4454 ceph_con_send(&session->s_con, msg);
4458 * lock unlock the session, to wait ongoing session activities
4460 static void lock_unlock_session(struct ceph_mds_session *s)
4462 mutex_lock(&s->s_mutex);
4463 mutex_unlock(&s->s_mutex);
4466 static void maybe_recover_session(struct ceph_mds_client *mdsc)
4468 struct ceph_fs_client *fsc = mdsc->fsc;
4470 if (!ceph_test_mount_opt(fsc, CLEANRECOVER))
4473 if (READ_ONCE(fsc->mount_state) != CEPH_MOUNT_MOUNTED)
4476 if (!READ_ONCE(fsc->blocklisted))
4479 pr_info("auto reconnect after blocklisted\n");
4480 ceph_force_reconnect(fsc->sb);
4483 bool check_session_state(struct ceph_mds_session *s)
4485 switch (s->s_state) {
4486 case CEPH_MDS_SESSION_OPEN:
4487 if (s->s_ttl && time_after(jiffies, s->s_ttl)) {
4488 s->s_state = CEPH_MDS_SESSION_HUNG;
4489 pr_info("mds%d hung\n", s->s_mds);
4492 case CEPH_MDS_SESSION_CLOSING:
4493 case CEPH_MDS_SESSION_NEW:
4494 case CEPH_MDS_SESSION_RESTARTING:
4495 case CEPH_MDS_SESSION_CLOSED:
4496 case CEPH_MDS_SESSION_REJECTED:
4504 * If the sequence is incremented while we're waiting on a REQUEST_CLOSE reply,
4505 * then we need to retransmit that request.
4507 void inc_session_sequence(struct ceph_mds_session *s)
4509 lockdep_assert_held(&s->s_mutex);
4513 if (s->s_state == CEPH_MDS_SESSION_CLOSING) {
4516 dout("resending session close request for mds%d\n", s->s_mds);
4517 ret = request_close_session(s);
4519 pr_err("unable to close session to mds%d: %d\n",
4525 * delayed work -- periodically trim expired leases, renew caps with mds. If
4526 * the @delay parameter is set to 0 or if it's more than 5 secs, the default
4527 * workqueue delay value of 5 secs will be used.
4529 static void schedule_delayed(struct ceph_mds_client *mdsc, unsigned long delay)
4531 unsigned long max_delay = HZ * 5;
4533 /* 5 secs default delay */
4534 if (!delay || (delay > max_delay))
4536 schedule_delayed_work(&mdsc->delayed_work,
4537 round_jiffies_relative(delay));
4540 static void delayed_work(struct work_struct *work)
4542 struct ceph_mds_client *mdsc =
4543 container_of(work, struct ceph_mds_client, delayed_work.work);
4544 unsigned long delay;
4549 dout("mdsc delayed_work\n");
4554 mutex_lock(&mdsc->mutex);
4555 renew_interval = mdsc->mdsmap->m_session_timeout >> 2;
4556 renew_caps = time_after_eq(jiffies, HZ*renew_interval +
4557 mdsc->last_renew_caps);
4559 mdsc->last_renew_caps = jiffies;
4561 for (i = 0; i < mdsc->max_sessions; i++) {
4562 struct ceph_mds_session *s = __ceph_lookup_mds_session(mdsc, i);
4566 if (!check_session_state(s)) {
4567 ceph_put_mds_session(s);
4570 mutex_unlock(&mdsc->mutex);
4572 mutex_lock(&s->s_mutex);
4574 send_renew_caps(mdsc, s);
4576 ceph_con_keepalive(&s->s_con);
4577 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
4578 s->s_state == CEPH_MDS_SESSION_HUNG)
4579 ceph_send_cap_releases(mdsc, s);
4580 mutex_unlock(&s->s_mutex);
4581 ceph_put_mds_session(s);
4583 mutex_lock(&mdsc->mutex);
4585 mutex_unlock(&mdsc->mutex);
4587 delay = ceph_check_delayed_caps(mdsc);
4589 ceph_queue_cap_reclaim_work(mdsc);
4591 ceph_trim_snapid_map(mdsc);
4593 maybe_recover_session(mdsc);
4595 schedule_delayed(mdsc, delay);
4598 int ceph_mdsc_init(struct ceph_fs_client *fsc)
4601 struct ceph_mds_client *mdsc;
4604 mdsc = kzalloc(sizeof(struct ceph_mds_client), GFP_NOFS);
4608 mutex_init(&mdsc->mutex);
4609 mdsc->mdsmap = kzalloc(sizeof(*mdsc->mdsmap), GFP_NOFS);
4610 if (!mdsc->mdsmap) {
4615 init_completion(&mdsc->safe_umount_waiters);
4616 init_waitqueue_head(&mdsc->session_close_wq);
4617 INIT_LIST_HEAD(&mdsc->waiting_for_map);
4618 mdsc->quotarealms_inodes = RB_ROOT;
4619 mutex_init(&mdsc->quotarealms_inodes_mutex);
4620 init_rwsem(&mdsc->snap_rwsem);
4621 mdsc->snap_realms = RB_ROOT;
4622 INIT_LIST_HEAD(&mdsc->snap_empty);
4623 spin_lock_init(&mdsc->snap_empty_lock);
4624 mdsc->request_tree = RB_ROOT;
4625 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
4626 mdsc->last_renew_caps = jiffies;
4627 INIT_LIST_HEAD(&mdsc->cap_delay_list);
4628 INIT_LIST_HEAD(&mdsc->cap_wait_list);
4629 spin_lock_init(&mdsc->cap_delay_lock);
4630 INIT_LIST_HEAD(&mdsc->snap_flush_list);
4631 spin_lock_init(&mdsc->snap_flush_lock);
4632 mdsc->last_cap_flush_tid = 1;
4633 INIT_LIST_HEAD(&mdsc->cap_flush_list);
4634 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
4635 spin_lock_init(&mdsc->cap_dirty_lock);
4636 init_waitqueue_head(&mdsc->cap_flushing_wq);
4637 INIT_WORK(&mdsc->cap_reclaim_work, ceph_cap_reclaim_work);
4638 err = ceph_metric_init(&mdsc->metric);
4642 spin_lock_init(&mdsc->dentry_list_lock);
4643 INIT_LIST_HEAD(&mdsc->dentry_leases);
4644 INIT_LIST_HEAD(&mdsc->dentry_dir_leases);
4646 ceph_caps_init(mdsc);
4647 ceph_adjust_caps_max_min(mdsc, fsc->mount_options);
4649 spin_lock_init(&mdsc->snapid_map_lock);
4650 mdsc->snapid_map_tree = RB_ROOT;
4651 INIT_LIST_HEAD(&mdsc->snapid_map_lru);
4653 init_rwsem(&mdsc->pool_perm_rwsem);
4654 mdsc->pool_perm_tree = RB_ROOT;
4656 strscpy(mdsc->nodename, utsname()->nodename,
4657 sizeof(mdsc->nodename));
4663 kfree(mdsc->mdsmap);
4670 * Wait for safe replies on open mds requests. If we time out, drop
4671 * all requests from the tree to avoid dangling dentry refs.
4673 static void wait_requests(struct ceph_mds_client *mdsc)
4675 struct ceph_options *opts = mdsc->fsc->client->options;
4676 struct ceph_mds_request *req;
4678 mutex_lock(&mdsc->mutex);
4679 if (__get_oldest_req(mdsc)) {
4680 mutex_unlock(&mdsc->mutex);
4682 dout("wait_requests waiting for requests\n");
4683 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
4684 ceph_timeout_jiffies(opts->mount_timeout));
4686 /* tear down remaining requests */
4687 mutex_lock(&mdsc->mutex);
4688 while ((req = __get_oldest_req(mdsc))) {
4689 dout("wait_requests timed out on tid %llu\n",
4691 list_del_init(&req->r_wait);
4692 __unregister_request(mdsc, req);
4695 mutex_unlock(&mdsc->mutex);
4696 dout("wait_requests done\n");
4699 void send_flush_mdlog(struct ceph_mds_session *s)
4701 struct ceph_msg *msg;
4704 * Pre-luminous MDS crashes when it sees an unknown session request
4706 if (!CEPH_HAVE_FEATURE(s->s_con.peer_features, SERVER_LUMINOUS))
4709 mutex_lock(&s->s_mutex);
4710 dout("request mdlog flush to mds%d (%s)s seq %lld\n", s->s_mds,
4711 ceph_session_state_name(s->s_state), s->s_seq);
4712 msg = ceph_create_session_msg(CEPH_SESSION_REQUEST_FLUSH_MDLOG,
4715 pr_err("failed to request mdlog flush to mds%d (%s) seq %lld\n",
4716 s->s_mds, ceph_session_state_name(s->s_state), s->s_seq);
4718 ceph_con_send(&s->s_con, msg);
4720 mutex_unlock(&s->s_mutex);
4724 * called before mount is ro, and before dentries are torn down.
4725 * (hmm, does this still race with new lookups?)
4727 void ceph_mdsc_pre_umount(struct ceph_mds_client *mdsc)
4729 dout("pre_umount\n");
4732 ceph_mdsc_iterate_sessions(mdsc, send_flush_mdlog, true);
4733 ceph_mdsc_iterate_sessions(mdsc, lock_unlock_session, false);
4734 ceph_flush_dirty_caps(mdsc);
4735 wait_requests(mdsc);
4738 * wait for reply handlers to drop their request refs and
4739 * their inode/dcache refs
4743 ceph_cleanup_quotarealms_inodes(mdsc);
4747 * flush the mdlog and wait for all write mds requests to flush.
4749 static void flush_mdlog_and_wait_mdsc_unsafe_requests(struct ceph_mds_client *mdsc,
4752 struct ceph_mds_request *req = NULL, *nextreq;
4753 struct ceph_mds_session *last_session = NULL;
4756 mutex_lock(&mdsc->mutex);
4757 dout("%s want %lld\n", __func__, want_tid);
4759 req = __get_oldest_req(mdsc);
4760 while (req && req->r_tid <= want_tid) {
4761 /* find next request */
4762 n = rb_next(&req->r_node);
4764 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
4767 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
4768 (req->r_op & CEPH_MDS_OP_WRITE)) {
4769 struct ceph_mds_session *s = req->r_session;
4777 ceph_mdsc_get_request(req);
4779 ceph_mdsc_get_request(nextreq);
4780 s = ceph_get_mds_session(s);
4781 mutex_unlock(&mdsc->mutex);
4783 /* send flush mdlog request to MDS */
4784 if (last_session != s) {
4785 send_flush_mdlog(s);
4786 ceph_put_mds_session(last_session);
4789 ceph_put_mds_session(s);
4791 dout("%s wait on %llu (want %llu)\n", __func__,
4792 req->r_tid, want_tid);
4793 wait_for_completion(&req->r_safe_completion);
4795 mutex_lock(&mdsc->mutex);
4796 ceph_mdsc_put_request(req);
4798 break; /* next dne before, so we're done! */
4799 if (RB_EMPTY_NODE(&nextreq->r_node)) {
4800 /* next request was removed from tree */
4801 ceph_mdsc_put_request(nextreq);
4804 ceph_mdsc_put_request(nextreq); /* won't go away */
4808 mutex_unlock(&mdsc->mutex);
4809 ceph_put_mds_session(last_session);
4810 dout("%s done\n", __func__);
4813 void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
4815 u64 want_tid, want_flush;
4817 if (READ_ONCE(mdsc->fsc->mount_state) >= CEPH_MOUNT_SHUTDOWN)
4821 mutex_lock(&mdsc->mutex);
4822 want_tid = mdsc->last_tid;
4823 mutex_unlock(&mdsc->mutex);
4825 ceph_flush_dirty_caps(mdsc);
4826 spin_lock(&mdsc->cap_dirty_lock);
4827 want_flush = mdsc->last_cap_flush_tid;
4828 if (!list_empty(&mdsc->cap_flush_list)) {
4829 struct ceph_cap_flush *cf =
4830 list_last_entry(&mdsc->cap_flush_list,
4831 struct ceph_cap_flush, g_list);
4834 spin_unlock(&mdsc->cap_dirty_lock);
4836 dout("sync want tid %lld flush_seq %lld\n",
4837 want_tid, want_flush);
4839 flush_mdlog_and_wait_mdsc_unsafe_requests(mdsc, want_tid);
4840 wait_caps_flush(mdsc, want_flush);
4844 * true if all sessions are closed, or we force unmount
4846 static bool done_closing_sessions(struct ceph_mds_client *mdsc, int skipped)
4848 if (READ_ONCE(mdsc->fsc->mount_state) == CEPH_MOUNT_SHUTDOWN)
4850 return atomic_read(&mdsc->num_sessions) <= skipped;
4854 * called after sb is ro.
4856 void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
4858 struct ceph_options *opts = mdsc->fsc->client->options;
4859 struct ceph_mds_session *session;
4863 dout("close_sessions\n");
4865 /* close sessions */
4866 mutex_lock(&mdsc->mutex);
4867 for (i = 0; i < mdsc->max_sessions; i++) {
4868 session = __ceph_lookup_mds_session(mdsc, i);
4871 mutex_unlock(&mdsc->mutex);
4872 mutex_lock(&session->s_mutex);
4873 if (__close_session(mdsc, session) <= 0)
4875 mutex_unlock(&session->s_mutex);
4876 ceph_put_mds_session(session);
4877 mutex_lock(&mdsc->mutex);
4879 mutex_unlock(&mdsc->mutex);
4881 dout("waiting for sessions to close\n");
4882 wait_event_timeout(mdsc->session_close_wq,
4883 done_closing_sessions(mdsc, skipped),
4884 ceph_timeout_jiffies(opts->mount_timeout));
4886 /* tear down remaining sessions */
4887 mutex_lock(&mdsc->mutex);
4888 for (i = 0; i < mdsc->max_sessions; i++) {
4889 if (mdsc->sessions[i]) {
4890 session = ceph_get_mds_session(mdsc->sessions[i]);
4891 __unregister_session(mdsc, session);
4892 mutex_unlock(&mdsc->mutex);
4893 mutex_lock(&session->s_mutex);
4894 remove_session_caps(session);
4895 mutex_unlock(&session->s_mutex);
4896 ceph_put_mds_session(session);
4897 mutex_lock(&mdsc->mutex);
4900 WARN_ON(!list_empty(&mdsc->cap_delay_list));
4901 mutex_unlock(&mdsc->mutex);
4903 ceph_cleanup_snapid_map(mdsc);
4904 ceph_cleanup_global_and_empty_realms(mdsc);
4906 cancel_work_sync(&mdsc->cap_reclaim_work);
4907 cancel_delayed_work_sync(&mdsc->delayed_work); /* cancel timer */
4912 void ceph_mdsc_force_umount(struct ceph_mds_client *mdsc)
4914 struct ceph_mds_session *session;
4917 dout("force umount\n");
4919 mutex_lock(&mdsc->mutex);
4920 for (mds = 0; mds < mdsc->max_sessions; mds++) {
4921 session = __ceph_lookup_mds_session(mdsc, mds);
4925 if (session->s_state == CEPH_MDS_SESSION_REJECTED)
4926 __unregister_session(mdsc, session);
4927 __wake_requests(mdsc, &session->s_waiting);
4928 mutex_unlock(&mdsc->mutex);
4930 mutex_lock(&session->s_mutex);
4931 __close_session(mdsc, session);
4932 if (session->s_state == CEPH_MDS_SESSION_CLOSING) {
4933 cleanup_session_requests(mdsc, session);
4934 remove_session_caps(session);
4936 mutex_unlock(&session->s_mutex);
4937 ceph_put_mds_session(session);
4939 mutex_lock(&mdsc->mutex);
4940 kick_requests(mdsc, mds);
4942 __wake_requests(mdsc, &mdsc->waiting_for_map);
4943 mutex_unlock(&mdsc->mutex);
4946 static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
4950 * Make sure the delayed work stopped before releasing
4953 * Because the cancel_delayed_work_sync() will only
4954 * guarantee that the work finishes executing. But the
4955 * delayed work will re-arm itself again after that.
4957 flush_delayed_work(&mdsc->delayed_work);
4960 ceph_mdsmap_destroy(mdsc->mdsmap);
4961 kfree(mdsc->sessions);
4962 ceph_caps_finalize(mdsc);
4963 ceph_pool_perm_destroy(mdsc);
4966 void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
4968 struct ceph_mds_client *mdsc = fsc->mdsc;
4969 dout("mdsc_destroy %p\n", mdsc);
4974 /* flush out any connection work with references to us */
4977 ceph_mdsc_stop(mdsc);
4979 ceph_metric_destroy(&mdsc->metric);
4983 dout("mdsc_destroy %p done\n", mdsc);
4986 void ceph_mdsc_handle_fsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
4988 struct ceph_fs_client *fsc = mdsc->fsc;
4989 const char *mds_namespace = fsc->mount_options->mds_namespace;
4990 void *p = msg->front.iov_base;
4991 void *end = p + msg->front.iov_len;
4994 u32 mount_fscid = (u32)-1;
4997 ceph_decode_need(&p, end, sizeof(u32), bad);
4998 epoch = ceph_decode_32(&p);
5000 dout("handle_fsmap epoch %u\n", epoch);
5002 /* struct_v, struct_cv, map_len, epoch, legacy_client_fscid */
5003 ceph_decode_skip_n(&p, end, 2 + sizeof(u32) * 3, bad);
5005 ceph_decode_32_safe(&p, end, num_fs, bad);
5006 while (num_fs-- > 0) {
5007 void *info_p, *info_end;
5011 ceph_decode_need(&p, end, 2 + sizeof(u32), bad);
5012 p += 2; // info_v, info_cv
5013 info_len = ceph_decode_32(&p);
5014 ceph_decode_need(&p, end, info_len, bad);
5016 info_end = p + info_len;
5019 ceph_decode_need(&info_p, info_end, sizeof(u32) * 2, bad);
5020 fscid = ceph_decode_32(&info_p);
5021 namelen = ceph_decode_32(&info_p);
5022 ceph_decode_need(&info_p, info_end, namelen, bad);
5024 if (mds_namespace &&
5025 strlen(mds_namespace) == namelen &&
5026 !strncmp(mds_namespace, (char *)info_p, namelen)) {
5027 mount_fscid = fscid;
5032 ceph_monc_got_map(&fsc->client->monc, CEPH_SUB_FSMAP, epoch);
5033 if (mount_fscid != (u32)-1) {
5034 fsc->client->monc.fs_cluster_id = mount_fscid;
5035 ceph_monc_want_map(&fsc->client->monc, CEPH_SUB_MDSMAP,
5037 ceph_monc_renew_subs(&fsc->client->monc);
5045 pr_err("error decoding fsmap %d. Shutting down mount.\n", err);
5046 ceph_umount_begin(mdsc->fsc->sb);
5048 mutex_lock(&mdsc->mutex);
5049 mdsc->mdsmap_err = err;
5050 __wake_requests(mdsc, &mdsc->waiting_for_map);
5051 mutex_unlock(&mdsc->mutex);
5055 * handle mds map update.
5057 void ceph_mdsc_handle_mdsmap(struct ceph_mds_client *mdsc, struct ceph_msg *msg)
5061 void *p = msg->front.iov_base;
5062 void *end = p + msg->front.iov_len;
5063 struct ceph_mdsmap *newmap, *oldmap;
5064 struct ceph_fsid fsid;
5067 ceph_decode_need(&p, end, sizeof(fsid)+2*sizeof(u32), bad);
5068 ceph_decode_copy(&p, &fsid, sizeof(fsid));
5069 if (ceph_check_fsid(mdsc->fsc->client, &fsid) < 0)
5071 epoch = ceph_decode_32(&p);
5072 maplen = ceph_decode_32(&p);
5073 dout("handle_map epoch %u len %d\n", epoch, (int)maplen);
5075 /* do we need it? */
5076 mutex_lock(&mdsc->mutex);
5077 if (mdsc->mdsmap && epoch <= mdsc->mdsmap->m_epoch) {
5078 dout("handle_map epoch %u <= our %u\n",
5079 epoch, mdsc->mdsmap->m_epoch);
5080 mutex_unlock(&mdsc->mutex);
5084 newmap = ceph_mdsmap_decode(&p, end, ceph_msgr2(mdsc->fsc->client));
5085 if (IS_ERR(newmap)) {
5086 err = PTR_ERR(newmap);
5090 /* swap into place */
5092 oldmap = mdsc->mdsmap;
5093 mdsc->mdsmap = newmap;
5094 check_new_map(mdsc, newmap, oldmap);
5095 ceph_mdsmap_destroy(oldmap);
5097 mdsc->mdsmap = newmap; /* first mds map */
5099 mdsc->fsc->max_file_size = min((loff_t)mdsc->mdsmap->m_max_file_size,
5102 __wake_requests(mdsc, &mdsc->waiting_for_map);
5103 ceph_monc_got_map(&mdsc->fsc->client->monc, CEPH_SUB_MDSMAP,
5104 mdsc->mdsmap->m_epoch);
5106 mutex_unlock(&mdsc->mutex);
5107 schedule_delayed(mdsc, 0);
5111 mutex_unlock(&mdsc->mutex);
5113 pr_err("error decoding mdsmap %d. Shutting down mount.\n", err);
5114 ceph_umount_begin(mdsc->fsc->sb);
5118 static struct ceph_connection *mds_get_con(struct ceph_connection *con)
5120 struct ceph_mds_session *s = con->private;
5122 if (ceph_get_mds_session(s))
5127 static void mds_put_con(struct ceph_connection *con)
5129 struct ceph_mds_session *s = con->private;
5131 ceph_put_mds_session(s);
5135 * if the client is unresponsive for long enough, the mds will kill
5136 * the session entirely.
5138 static void mds_peer_reset(struct ceph_connection *con)
5140 struct ceph_mds_session *s = con->private;
5141 struct ceph_mds_client *mdsc = s->s_mdsc;
5143 pr_warn("mds%d closed our session\n", s->s_mds);
5144 send_mds_reconnect(mdsc, s);
5147 static void mds_dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5149 struct ceph_mds_session *s = con->private;
5150 struct ceph_mds_client *mdsc = s->s_mdsc;
5151 int type = le16_to_cpu(msg->hdr.type);
5153 mutex_lock(&mdsc->mutex);
5154 if (__verify_registered_session(mdsc, s) < 0) {
5155 mutex_unlock(&mdsc->mutex);
5158 mutex_unlock(&mdsc->mutex);
5161 case CEPH_MSG_MDS_MAP:
5162 ceph_mdsc_handle_mdsmap(mdsc, msg);
5164 case CEPH_MSG_FS_MAP_USER:
5165 ceph_mdsc_handle_fsmap(mdsc, msg);
5167 case CEPH_MSG_CLIENT_SESSION:
5168 handle_session(s, msg);
5170 case CEPH_MSG_CLIENT_REPLY:
5171 handle_reply(s, msg);
5173 case CEPH_MSG_CLIENT_REQUEST_FORWARD:
5174 handle_forward(mdsc, s, msg);
5176 case CEPH_MSG_CLIENT_CAPS:
5177 ceph_handle_caps(s, msg);
5179 case CEPH_MSG_CLIENT_SNAP:
5180 ceph_handle_snap(mdsc, s, msg);
5182 case CEPH_MSG_CLIENT_LEASE:
5183 handle_lease(mdsc, s, msg);
5185 case CEPH_MSG_CLIENT_QUOTA:
5186 ceph_handle_quota(mdsc, s, msg);
5190 pr_err("received unknown message type %d %s\n", type,
5191 ceph_msg_type_name(type));
5202 * Note: returned pointer is the address of a structure that's
5203 * managed separately. Caller must *not* attempt to free it.
5205 static struct ceph_auth_handshake *
5206 mds_get_authorizer(struct ceph_connection *con, int *proto, int force_new)
5208 struct ceph_mds_session *s = con->private;
5209 struct ceph_mds_client *mdsc = s->s_mdsc;
5210 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5211 struct ceph_auth_handshake *auth = &s->s_auth;
5214 ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5215 force_new, proto, NULL, NULL);
5217 return ERR_PTR(ret);
5222 static int mds_add_authorizer_challenge(struct ceph_connection *con,
5223 void *challenge_buf, int challenge_buf_len)
5225 struct ceph_mds_session *s = con->private;
5226 struct ceph_mds_client *mdsc = s->s_mdsc;
5227 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5229 return ceph_auth_add_authorizer_challenge(ac, s->s_auth.authorizer,
5230 challenge_buf, challenge_buf_len);
5233 static int mds_verify_authorizer_reply(struct ceph_connection *con)
5235 struct ceph_mds_session *s = con->private;
5236 struct ceph_mds_client *mdsc = s->s_mdsc;
5237 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5238 struct ceph_auth_handshake *auth = &s->s_auth;
5240 return ceph_auth_verify_authorizer_reply(ac, auth->authorizer,
5241 auth->authorizer_reply_buf, auth->authorizer_reply_buf_len,
5242 NULL, NULL, NULL, NULL);
5245 static int mds_invalidate_authorizer(struct ceph_connection *con)
5247 struct ceph_mds_session *s = con->private;
5248 struct ceph_mds_client *mdsc = s->s_mdsc;
5249 struct ceph_auth_client *ac = mdsc->fsc->client->monc.auth;
5251 ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_MDS);
5253 return ceph_monc_validate_auth(&mdsc->fsc->client->monc);
5256 static int mds_get_auth_request(struct ceph_connection *con,
5257 void *buf, int *buf_len,
5258 void **authorizer, int *authorizer_len)
5260 struct ceph_mds_session *s = con->private;
5261 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5262 struct ceph_auth_handshake *auth = &s->s_auth;
5265 ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_MDS,
5270 *authorizer = auth->authorizer_buf;
5271 *authorizer_len = auth->authorizer_buf_len;
5275 static int mds_handle_auth_reply_more(struct ceph_connection *con,
5276 void *reply, int reply_len,
5277 void *buf, int *buf_len,
5278 void **authorizer, int *authorizer_len)
5280 struct ceph_mds_session *s = con->private;
5281 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5282 struct ceph_auth_handshake *auth = &s->s_auth;
5285 ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len,
5290 *authorizer = auth->authorizer_buf;
5291 *authorizer_len = auth->authorizer_buf_len;
5295 static int mds_handle_auth_done(struct ceph_connection *con,
5296 u64 global_id, void *reply, int reply_len,
5297 u8 *session_key, int *session_key_len,
5298 u8 *con_secret, int *con_secret_len)
5300 struct ceph_mds_session *s = con->private;
5301 struct ceph_auth_client *ac = s->s_mdsc->fsc->client->monc.auth;
5302 struct ceph_auth_handshake *auth = &s->s_auth;
5304 return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len,
5305 session_key, session_key_len,
5306 con_secret, con_secret_len);
5309 static int mds_handle_auth_bad_method(struct ceph_connection *con,
5310 int used_proto, int result,
5311 const int *allowed_protos, int proto_cnt,
5312 const int *allowed_modes, int mode_cnt)
5314 struct ceph_mds_session *s = con->private;
5315 struct ceph_mon_client *monc = &s->s_mdsc->fsc->client->monc;
5318 if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_MDS,
5320 allowed_protos, proto_cnt,
5321 allowed_modes, mode_cnt)) {
5322 ret = ceph_monc_validate_auth(monc);
5330 static struct ceph_msg *mds_alloc_msg(struct ceph_connection *con,
5331 struct ceph_msg_header *hdr, int *skip)
5333 struct ceph_msg *msg;
5334 int type = (int) le16_to_cpu(hdr->type);
5335 int front_len = (int) le32_to_cpu(hdr->front_len);
5341 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
5343 pr_err("unable to allocate msg type %d len %d\n",
5351 static int mds_sign_message(struct ceph_msg *msg)
5353 struct ceph_mds_session *s = msg->con->private;
5354 struct ceph_auth_handshake *auth = &s->s_auth;
5356 return ceph_auth_sign_message(auth, msg);
5359 static int mds_check_message_signature(struct ceph_msg *msg)
5361 struct ceph_mds_session *s = msg->con->private;
5362 struct ceph_auth_handshake *auth = &s->s_auth;
5364 return ceph_auth_check_message_signature(auth, msg);
5367 static const struct ceph_connection_operations mds_con_ops = {
5370 .alloc_msg = mds_alloc_msg,
5371 .dispatch = mds_dispatch,
5372 .peer_reset = mds_peer_reset,
5373 .get_authorizer = mds_get_authorizer,
5374 .add_authorizer_challenge = mds_add_authorizer_challenge,
5375 .verify_authorizer_reply = mds_verify_authorizer_reply,
5376 .invalidate_authorizer = mds_invalidate_authorizer,
5377 .sign_message = mds_sign_message,
5378 .check_message_signature = mds_check_message_signature,
5379 .get_auth_request = mds_get_auth_request,
5380 .handle_auth_reply_more = mds_handle_auth_reply_more,
5381 .handle_auth_done = mds_handle_auth_done,
5382 .handle_auth_bad_method = mds_handle_auth_bad_method,